1 // SPDX-License-Identifier: GPL-2.0
3 * Kernel timekeeping code and accessor functions. Based on code from
4 * timer.c, moved in commit 8524070b7982.
6 #include <linux/timekeeper_internal.h>
7 #include <linux/module.h>
8 #include <linux/interrupt.h>
9 #include <linux/percpu.h>
10 #include <linux/init.h>
12 #include <linux/nmi.h>
13 #include <linux/sched.h>
14 #include <linux/sched/loadavg.h>
15 #include <linux/sched/clock.h>
16 #include <linux/syscore_ops.h>
17 #include <linux/clocksource.h>
18 #include <linux/jiffies.h>
19 #include <linux/time.h>
20 #include <linux/timex.h>
21 #include <linux/tick.h>
22 #include <linux/stop_machine.h>
23 #include <linux/pvclock_gtod.h>
24 #include <linux/compiler.h>
25 #include <linux/audit.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)
35 enum timekeeping_adv_mode {
36 /* Update timekeeper when a tick has passed */
39 /* Update timekeeper on a direct frequency change */
43 DEFINE_RAW_SPINLOCK(timekeeper_lock);
46 * The most important data for readout fits into a single 64 byte
50 seqcount_raw_spinlock_t seq;
51 struct timekeeper timekeeper;
52 } tk_core ____cacheline_aligned = {
53 .seq = SEQCNT_RAW_SPINLOCK_ZERO(tk_core.seq, &timekeeper_lock),
56 static struct timekeeper shadow_timekeeper;
58 /* flag for if timekeeping is suspended */
59 int __read_mostly timekeeping_suspended;
62 * struct tk_fast - NMI safe timekeeper
63 * @seq: Sequence counter for protecting updates. The lowest bit
64 * is the index for the tk_read_base array
65 * @base: tk_read_base array. Access is indexed by the lowest bit of
68 * See @update_fast_timekeeper() below.
72 struct tk_read_base base[2];
75 /* Suspend-time cycles value for halted fast timekeeper. */
76 static u64 cycles_at_suspend;
78 static u64 dummy_clock_read(struct clocksource *cs)
80 if (timekeeping_suspended)
81 return cycles_at_suspend;
85 static struct clocksource dummy_clock = {
86 .read = dummy_clock_read,
90 * Boot time initialization which allows local_clock() to be utilized
91 * during early boot when clocksources are not available. local_clock()
92 * returns nanoseconds already so no conversion is required, hence mult=1
93 * and shift=0. When the first proper clocksource is installed then
94 * the fast time keepers are updated with the correct values.
96 #define FAST_TK_INIT \
98 .clock = &dummy_clock, \
99 .mask = CLOCKSOURCE_MASK(64), \
104 static struct tk_fast tk_fast_mono ____cacheline_aligned = {
105 .seq = SEQCNT_LATCH_ZERO(tk_fast_mono.seq),
106 .base[0] = FAST_TK_INIT,
107 .base[1] = FAST_TK_INIT,
110 static struct tk_fast tk_fast_raw ____cacheline_aligned = {
111 .seq = SEQCNT_LATCH_ZERO(tk_fast_raw.seq),
112 .base[0] = FAST_TK_INIT,
113 .base[1] = FAST_TK_INIT,
116 static inline void tk_normalize_xtime(struct timekeeper *tk)
118 while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) {
119 tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
122 while (tk->tkr_raw.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_raw.shift)) {
123 tk->tkr_raw.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
128 static inline struct timespec64 tk_xtime(const struct timekeeper *tk)
130 struct timespec64 ts;
132 ts.tv_sec = tk->xtime_sec;
133 ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
137 static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
139 tk->xtime_sec = ts->tv_sec;
140 tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift;
143 static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
145 tk->xtime_sec += ts->tv_sec;
146 tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift;
147 tk_normalize_xtime(tk);
150 static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
152 struct timespec64 tmp;
155 * Verify consistency of: offset_real = -wall_to_monotonic
156 * before modifying anything
158 set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
159 -tk->wall_to_monotonic.tv_nsec);
160 WARN_ON_ONCE(tk->offs_real != timespec64_to_ktime(tmp));
161 tk->wall_to_monotonic = wtm;
162 set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
163 tk->offs_real = timespec64_to_ktime(tmp);
164 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0));
167 static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
169 tk->offs_boot = ktime_add(tk->offs_boot, delta);
171 * Timespec representation for VDSO update to avoid 64bit division
174 tk->monotonic_to_boot = ktime_to_timespec64(tk->offs_boot);
178 * tk_clock_read - atomic clocksource read() helper
180 * This helper is necessary to use in the read paths because, while the
181 * seqcount ensures we don't return a bad value while structures are updated,
182 * it doesn't protect from potential crashes. There is the possibility that
183 * the tkr's clocksource may change between the read reference, and the
184 * clock reference passed to the read function. This can cause crashes if
185 * the wrong clocksource is passed to the wrong read function.
186 * This isn't necessary to use when holding the timekeeper_lock or doing
187 * a read of the fast-timekeeper tkrs (which is protected by its own locking
190 static inline u64 tk_clock_read(const struct tk_read_base *tkr)
192 struct clocksource *clock = READ_ONCE(tkr->clock);
194 return clock->read(clock);
197 #ifdef CONFIG_DEBUG_TIMEKEEPING
198 #define WARNING_FREQ (HZ*300) /* 5 minute rate-limiting */
200 static void timekeeping_check_update(struct timekeeper *tk, u64 offset)
203 u64 max_cycles = tk->tkr_mono.clock->max_cycles;
204 const char *name = tk->tkr_mono.clock->name;
206 if (offset > max_cycles) {
207 printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n",
208 offset, name, max_cycles);
209 printk_deferred(" timekeeping: Your kernel is sick, but tries to cope by capping time updates\n");
211 if (offset > (max_cycles >> 1)) {
212 printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the '%s' clock's 50%% safety margin (%lld)\n",
213 offset, name, max_cycles >> 1);
214 printk_deferred(" timekeeping: Your kernel is still fine, but is feeling a bit nervous\n");
218 if (tk->underflow_seen) {
219 if (jiffies - tk->last_warning > WARNING_FREQ) {
220 printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name);
221 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
222 printk_deferred(" Your kernel is probably still fine.\n");
223 tk->last_warning = jiffies;
225 tk->underflow_seen = 0;
228 if (tk->overflow_seen) {
229 if (jiffies - tk->last_warning > WARNING_FREQ) {
230 printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name);
231 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
232 printk_deferred(" Your kernel is probably still fine.\n");
233 tk->last_warning = jiffies;
235 tk->overflow_seen = 0;
239 static inline u64 timekeeping_get_delta(const struct tk_read_base *tkr)
241 struct timekeeper *tk = &tk_core.timekeeper;
242 u64 now, last, mask, max, delta;
246 * Since we're called holding a seqcount, the data may shift
247 * under us while we're doing the calculation. This can cause
248 * false positives, since we'd note a problem but throw the
249 * results away. So nest another seqcount here to atomically
250 * grab the points we are checking with.
253 seq = read_seqcount_begin(&tk_core.seq);
254 now = tk_clock_read(tkr);
255 last = tkr->cycle_last;
257 max = tkr->clock->max_cycles;
258 } while (read_seqcount_retry(&tk_core.seq, seq));
260 delta = clocksource_delta(now, last, mask);
263 * Try to catch underflows by checking if we are seeing small
264 * mask-relative negative values.
266 if (unlikely((~delta & mask) < (mask >> 3))) {
267 tk->underflow_seen = 1;
271 /* Cap delta value to the max_cycles values to avoid mult overflows */
272 if (unlikely(delta > max)) {
273 tk->overflow_seen = 1;
274 delta = tkr->clock->max_cycles;
280 static inline void timekeeping_check_update(struct timekeeper *tk, u64 offset)
283 static inline u64 timekeeping_get_delta(const struct tk_read_base *tkr)
285 u64 cycle_now, delta;
287 /* read clocksource */
288 cycle_now = tk_clock_read(tkr);
290 /* calculate the delta since the last update_wall_time */
291 delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
298 * tk_setup_internals - Set up internals to use clocksource clock.
300 * @tk: The target timekeeper to setup.
301 * @clock: Pointer to clocksource.
303 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
304 * pair and interval request.
306 * Unless you're the timekeeping code, you should not be using this!
308 static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
311 u64 tmp, ntpinterval;
312 struct clocksource *old_clock;
314 ++tk->cs_was_changed_seq;
315 old_clock = tk->tkr_mono.clock;
316 tk->tkr_mono.clock = clock;
317 tk->tkr_mono.mask = clock->mask;
318 tk->tkr_mono.cycle_last = tk_clock_read(&tk->tkr_mono);
320 tk->tkr_raw.clock = clock;
321 tk->tkr_raw.mask = clock->mask;
322 tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
324 /* Do the ns -> cycle conversion first, using original mult */
325 tmp = NTP_INTERVAL_LENGTH;
326 tmp <<= clock->shift;
328 tmp += clock->mult/2;
329 do_div(tmp, clock->mult);
333 interval = (u64) tmp;
334 tk->cycle_interval = interval;
336 /* Go back from cycles -> shifted ns */
337 tk->xtime_interval = interval * clock->mult;
338 tk->xtime_remainder = ntpinterval - tk->xtime_interval;
339 tk->raw_interval = interval * clock->mult;
341 /* if changing clocks, convert xtime_nsec shift units */
343 int shift_change = clock->shift - old_clock->shift;
344 if (shift_change < 0) {
345 tk->tkr_mono.xtime_nsec >>= -shift_change;
346 tk->tkr_raw.xtime_nsec >>= -shift_change;
348 tk->tkr_mono.xtime_nsec <<= shift_change;
349 tk->tkr_raw.xtime_nsec <<= shift_change;
353 tk->tkr_mono.shift = clock->shift;
354 tk->tkr_raw.shift = clock->shift;
357 tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
358 tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
361 * The timekeeper keeps its own mult values for the currently
362 * active clocksource. These value will be adjusted via NTP
363 * to counteract clock drifting.
365 tk->tkr_mono.mult = clock->mult;
366 tk->tkr_raw.mult = clock->mult;
367 tk->ntp_err_mult = 0;
368 tk->skip_second_overflow = 0;
371 /* Timekeeper helper functions. */
373 static inline u64 timekeeping_delta_to_ns(const struct tk_read_base *tkr, u64 delta)
377 nsec = delta * tkr->mult + tkr->xtime_nsec;
383 static inline u64 timekeeping_get_ns(const struct tk_read_base *tkr)
387 delta = timekeeping_get_delta(tkr);
388 return timekeeping_delta_to_ns(tkr, delta);
391 static inline u64 timekeeping_cycles_to_ns(const struct tk_read_base *tkr, u64 cycles)
395 /* calculate the delta since the last update_wall_time */
396 delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask);
397 return timekeeping_delta_to_ns(tkr, delta);
401 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
402 * @tkr: Timekeeping readout base from which we take the update
403 * @tkf: Pointer to NMI safe timekeeper
405 * We want to use this from any context including NMI and tracing /
406 * instrumenting the timekeeping code itself.
408 * Employ the latch technique; see @raw_write_seqcount_latch.
410 * So if a NMI hits the update of base[0] then it will use base[1]
411 * which is still consistent. In the worst case this can result is a
412 * slightly wrong timestamp (a few nanoseconds). See
413 * @ktime_get_mono_fast_ns.
415 static void update_fast_timekeeper(const struct tk_read_base *tkr,
418 struct tk_read_base *base = tkf->base;
420 /* Force readers off to base[1] */
421 raw_write_seqcount_latch(&tkf->seq);
424 memcpy(base, tkr, sizeof(*base));
426 /* Force readers back to base[0] */
427 raw_write_seqcount_latch(&tkf->seq);
430 memcpy(base + 1, base, sizeof(*base));
433 static __always_inline u64 fast_tk_get_delta_ns(struct tk_read_base *tkr)
435 u64 delta, cycles = tk_clock_read(tkr);
437 delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask);
438 return timekeeping_delta_to_ns(tkr, delta);
441 static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
443 struct tk_read_base *tkr;
448 seq = raw_read_seqcount_latch(&tkf->seq);
449 tkr = tkf->base + (seq & 0x01);
450 now = ktime_to_ns(tkr->base);
451 now += fast_tk_get_delta_ns(tkr);
452 } while (read_seqcount_latch_retry(&tkf->seq, seq));
458 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
460 * This timestamp is not guaranteed to be monotonic across an update.
461 * The timestamp is calculated by:
463 * now = base_mono + clock_delta * slope
465 * So if the update lowers the slope, readers who are forced to the
466 * not yet updated second array are still using the old steeper slope.
475 * |12345678---> reader order
481 * So reader 6 will observe time going backwards versus reader 5.
483 * While other CPUs are likely to be able to observe that, the only way
484 * for a CPU local observation is when an NMI hits in the middle of
485 * the update. Timestamps taken from that NMI context might be ahead
486 * of the following timestamps. Callers need to be aware of that and
489 u64 notrace ktime_get_mono_fast_ns(void)
491 return __ktime_get_fast_ns(&tk_fast_mono);
493 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
496 * ktime_get_raw_fast_ns - Fast NMI safe access to clock monotonic raw
498 * Contrary to ktime_get_mono_fast_ns() this is always correct because the
499 * conversion factor is not affected by NTP/PTP correction.
501 u64 notrace ktime_get_raw_fast_ns(void)
503 return __ktime_get_fast_ns(&tk_fast_raw);
505 EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
508 * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock.
510 * To keep it NMI safe since we're accessing from tracing, we're not using a
511 * separate timekeeper with updates to monotonic clock and boot offset
512 * protected with seqcounts. This has the following minor side effects:
514 * (1) Its possible that a timestamp be taken after the boot offset is updated
515 * but before the timekeeper is updated. If this happens, the new boot offset
516 * is added to the old timekeeping making the clock appear to update slightly
519 * timekeeping_inject_sleeptime64()
520 * __timekeeping_inject_sleeptime(tk, delta);
522 * timekeeping_update(tk, TK_CLEAR_NTP...);
524 * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be
525 * partially updated. Since the tk->offs_boot update is a rare event, this
526 * should be a rare occurrence which postprocessing should be able to handle.
528 * The caveats vs. timestamp ordering as documented for ktime_get_fast_ns()
531 u64 notrace ktime_get_boot_fast_ns(void)
533 struct timekeeper *tk = &tk_core.timekeeper;
535 return (ktime_get_mono_fast_ns() + ktime_to_ns(data_race(tk->offs_boot)));
537 EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns);
540 * ktime_get_tai_fast_ns - NMI safe and fast access to tai clock.
542 * The same limitations as described for ktime_get_boot_fast_ns() apply. The
543 * mono time and the TAI offset are not read atomically which may yield wrong
544 * readouts. However, an update of the TAI offset is an rare event e.g., caused
545 * by settime or adjtimex with an offset. The user of this function has to deal
546 * with the possibility of wrong timestamps in post processing.
548 u64 notrace ktime_get_tai_fast_ns(void)
550 struct timekeeper *tk = &tk_core.timekeeper;
552 return (ktime_get_mono_fast_ns() + ktime_to_ns(data_race(tk->offs_tai)));
554 EXPORT_SYMBOL_GPL(ktime_get_tai_fast_ns);
556 static __always_inline u64 __ktime_get_real_fast(struct tk_fast *tkf, u64 *mono)
558 struct tk_read_base *tkr;
559 u64 basem, baser, delta;
563 seq = raw_read_seqcount_latch(&tkf->seq);
564 tkr = tkf->base + (seq & 0x01);
565 basem = ktime_to_ns(tkr->base);
566 baser = ktime_to_ns(tkr->base_real);
567 delta = fast_tk_get_delta_ns(tkr);
568 } while (read_seqcount_latch_retry(&tkf->seq, seq));
571 *mono = basem + delta;
572 return baser + delta;
576 * ktime_get_real_fast_ns: - NMI safe and fast access to clock realtime.
578 * See ktime_get_fast_ns() for documentation of the time stamp ordering.
580 u64 ktime_get_real_fast_ns(void)
582 return __ktime_get_real_fast(&tk_fast_mono, NULL);
584 EXPORT_SYMBOL_GPL(ktime_get_real_fast_ns);
587 * ktime_get_fast_timestamps: - NMI safe timestamps
588 * @snapshot: Pointer to timestamp storage
590 * Stores clock monotonic, boottime and realtime timestamps.
592 * Boot time is a racy access on 32bit systems if the sleep time injection
593 * happens late during resume and not in timekeeping_resume(). That could
594 * be avoided by expanding struct tk_read_base with boot offset for 32bit
595 * and adding more overhead to the update. As this is a hard to observe
596 * once per resume event which can be filtered with reasonable effort using
597 * the accurate mono/real timestamps, it's probably not worth the trouble.
599 * Aside of that it might be possible on 32 and 64 bit to observe the
600 * following when the sleep time injection happens late:
603 * timekeeping_resume()
604 * ktime_get_fast_timestamps()
605 * mono, real = __ktime_get_real_fast()
606 * inject_sleep_time()
608 * boot = mono + bootoffset;
610 * That means that boot time already has the sleep time adjustment, but
611 * real time does not. On the next readout both are in sync again.
613 * Preventing this for 64bit is not really feasible without destroying the
614 * careful cache layout of the timekeeper because the sequence count and
615 * struct tk_read_base would then need two cache lines instead of one.
617 * Access to the time keeper clock source is disabled across the innermost
618 * steps of suspend/resume. The accessors still work, but the timestamps
619 * are frozen until time keeping is resumed which happens very early.
621 * For regular suspend/resume there is no observable difference vs. sched
622 * clock, but it might affect some of the nasty low level debug printks.
624 * OTOH, access to sched clock is not guaranteed across suspend/resume on
625 * all systems either so it depends on the hardware in use.
627 * If that turns out to be a real problem then this could be mitigated by
628 * using sched clock in a similar way as during early boot. But it's not as
629 * trivial as on early boot because it needs some careful protection
630 * against the clock monotonic timestamp jumping backwards on resume.
632 void ktime_get_fast_timestamps(struct ktime_timestamps *snapshot)
634 struct timekeeper *tk = &tk_core.timekeeper;
636 snapshot->real = __ktime_get_real_fast(&tk_fast_mono, &snapshot->mono);
637 snapshot->boot = snapshot->mono + ktime_to_ns(data_race(tk->offs_boot));
641 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
642 * @tk: Timekeeper to snapshot.
644 * It generally is unsafe to access the clocksource after timekeeping has been
645 * suspended, so take a snapshot of the readout base of @tk and use it as the
646 * fast timekeeper's readout base while suspended. It will return the same
647 * number of cycles every time until timekeeping is resumed at which time the
648 * proper readout base for the fast timekeeper will be restored automatically.
650 static void halt_fast_timekeeper(const struct timekeeper *tk)
652 static struct tk_read_base tkr_dummy;
653 const struct tk_read_base *tkr = &tk->tkr_mono;
655 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
656 cycles_at_suspend = tk_clock_read(tkr);
657 tkr_dummy.clock = &dummy_clock;
658 tkr_dummy.base_real = tkr->base + tk->offs_real;
659 update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
662 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
663 tkr_dummy.clock = &dummy_clock;
664 update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
667 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
669 static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
671 raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
675 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
676 * @nb: Pointer to the notifier block to register
678 int pvclock_gtod_register_notifier(struct notifier_block *nb)
680 struct timekeeper *tk = &tk_core.timekeeper;
684 raw_spin_lock_irqsave(&timekeeper_lock, flags);
685 ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
686 update_pvclock_gtod(tk, true);
687 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
691 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
694 * pvclock_gtod_unregister_notifier - unregister a pvclock
695 * timedata update listener
696 * @nb: Pointer to the notifier block to unregister
698 int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
703 raw_spin_lock_irqsave(&timekeeper_lock, flags);
704 ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
705 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
709 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
712 * tk_update_leap_state - helper to update the next_leap_ktime
714 static inline void tk_update_leap_state(struct timekeeper *tk)
716 tk->next_leap_ktime = ntp_get_next_leap();
717 if (tk->next_leap_ktime != KTIME_MAX)
718 /* Convert to monotonic time */
719 tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
723 * Update the ktime_t based scalar nsec members of the timekeeper
725 static inline void tk_update_ktime_data(struct timekeeper *tk)
731 * The xtime based monotonic readout is:
732 * nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
733 * The ktime based monotonic readout is:
734 * nsec = base_mono + now();
735 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
737 seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
738 nsec = (u32) tk->wall_to_monotonic.tv_nsec;
739 tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
742 * The sum of the nanoseconds portions of xtime and
743 * wall_to_monotonic can be greater/equal one second. Take
744 * this into account before updating tk->ktime_sec.
746 nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
747 if (nsec >= NSEC_PER_SEC)
749 tk->ktime_sec = seconds;
751 /* Update the monotonic raw base */
752 tk->tkr_raw.base = ns_to_ktime(tk->raw_sec * NSEC_PER_SEC);
755 /* must hold timekeeper_lock */
756 static void timekeeping_update(struct timekeeper *tk, unsigned int action)
758 if (action & TK_CLEAR_NTP) {
763 tk_update_leap_state(tk);
764 tk_update_ktime_data(tk);
767 update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
769 tk->tkr_mono.base_real = tk->tkr_mono.base + tk->offs_real;
770 update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
771 update_fast_timekeeper(&tk->tkr_raw, &tk_fast_raw);
773 if (action & TK_CLOCK_WAS_SET)
774 tk->clock_was_set_seq++;
776 * The mirroring of the data to the shadow-timekeeper needs
777 * to happen last here to ensure we don't over-write the
778 * timekeeper structure on the next update with stale data
780 if (action & TK_MIRROR)
781 memcpy(&shadow_timekeeper, &tk_core.timekeeper,
782 sizeof(tk_core.timekeeper));
786 * timekeeping_forward_now - update clock to the current time
787 * @tk: Pointer to the timekeeper to update
789 * Forward the current clock to update its state since the last call to
790 * update_wall_time(). This is useful before significant clock changes,
791 * as it avoids having to deal with this time offset explicitly.
793 static void timekeeping_forward_now(struct timekeeper *tk)
795 u64 cycle_now, delta;
797 cycle_now = tk_clock_read(&tk->tkr_mono);
798 delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
799 tk->tkr_mono.cycle_last = cycle_now;
800 tk->tkr_raw.cycle_last = cycle_now;
802 tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult;
803 tk->tkr_raw.xtime_nsec += delta * tk->tkr_raw.mult;
805 tk_normalize_xtime(tk);
809 * ktime_get_real_ts64 - Returns the time of day in a timespec64.
810 * @ts: pointer to the timespec to be set
812 * Returns the time of day in a timespec64 (WARN if suspended).
814 void ktime_get_real_ts64(struct timespec64 *ts)
816 struct timekeeper *tk = &tk_core.timekeeper;
820 WARN_ON(timekeeping_suspended);
823 seq = read_seqcount_begin(&tk_core.seq);
825 ts->tv_sec = tk->xtime_sec;
826 nsecs = timekeeping_get_ns(&tk->tkr_mono);
828 } while (read_seqcount_retry(&tk_core.seq, seq));
831 timespec64_add_ns(ts, nsecs);
833 EXPORT_SYMBOL(ktime_get_real_ts64);
835 ktime_t ktime_get(void)
837 struct timekeeper *tk = &tk_core.timekeeper;
842 WARN_ON(timekeeping_suspended);
845 seq = read_seqcount_begin(&tk_core.seq);
846 base = tk->tkr_mono.base;
847 nsecs = timekeeping_get_ns(&tk->tkr_mono);
849 } while (read_seqcount_retry(&tk_core.seq, seq));
851 return ktime_add_ns(base, nsecs);
853 EXPORT_SYMBOL_GPL(ktime_get);
855 u32 ktime_get_resolution_ns(void)
857 struct timekeeper *tk = &tk_core.timekeeper;
861 WARN_ON(timekeeping_suspended);
864 seq = read_seqcount_begin(&tk_core.seq);
865 nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
866 } while (read_seqcount_retry(&tk_core.seq, seq));
870 EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
872 static ktime_t *offsets[TK_OFFS_MAX] = {
873 [TK_OFFS_REAL] = &tk_core.timekeeper.offs_real,
874 [TK_OFFS_BOOT] = &tk_core.timekeeper.offs_boot,
875 [TK_OFFS_TAI] = &tk_core.timekeeper.offs_tai,
878 ktime_t ktime_get_with_offset(enum tk_offsets offs)
880 struct timekeeper *tk = &tk_core.timekeeper;
882 ktime_t base, *offset = offsets[offs];
885 WARN_ON(timekeeping_suspended);
888 seq = read_seqcount_begin(&tk_core.seq);
889 base = ktime_add(tk->tkr_mono.base, *offset);
890 nsecs = timekeeping_get_ns(&tk->tkr_mono);
892 } while (read_seqcount_retry(&tk_core.seq, seq));
894 return ktime_add_ns(base, nsecs);
897 EXPORT_SYMBOL_GPL(ktime_get_with_offset);
899 ktime_t ktime_get_coarse_with_offset(enum tk_offsets offs)
901 struct timekeeper *tk = &tk_core.timekeeper;
903 ktime_t base, *offset = offsets[offs];
906 WARN_ON(timekeeping_suspended);
909 seq = read_seqcount_begin(&tk_core.seq);
910 base = ktime_add(tk->tkr_mono.base, *offset);
911 nsecs = tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift;
913 } while (read_seqcount_retry(&tk_core.seq, seq));
915 return ktime_add_ns(base, nsecs);
917 EXPORT_SYMBOL_GPL(ktime_get_coarse_with_offset);
920 * ktime_mono_to_any() - convert monotonic time to any other time
921 * @tmono: time to convert.
922 * @offs: which offset to use
924 ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
926 ktime_t *offset = offsets[offs];
931 seq = read_seqcount_begin(&tk_core.seq);
932 tconv = ktime_add(tmono, *offset);
933 } while (read_seqcount_retry(&tk_core.seq, seq));
937 EXPORT_SYMBOL_GPL(ktime_mono_to_any);
940 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
942 ktime_t ktime_get_raw(void)
944 struct timekeeper *tk = &tk_core.timekeeper;
950 seq = read_seqcount_begin(&tk_core.seq);
951 base = tk->tkr_raw.base;
952 nsecs = timekeeping_get_ns(&tk->tkr_raw);
954 } while (read_seqcount_retry(&tk_core.seq, seq));
956 return ktime_add_ns(base, nsecs);
958 EXPORT_SYMBOL_GPL(ktime_get_raw);
961 * ktime_get_ts64 - get the monotonic clock in timespec64 format
962 * @ts: pointer to timespec variable
964 * The function calculates the monotonic clock from the realtime
965 * clock and the wall_to_monotonic offset and stores the result
966 * in normalized timespec64 format in the variable pointed to by @ts.
968 void ktime_get_ts64(struct timespec64 *ts)
970 struct timekeeper *tk = &tk_core.timekeeper;
971 struct timespec64 tomono;
975 WARN_ON(timekeeping_suspended);
978 seq = read_seqcount_begin(&tk_core.seq);
979 ts->tv_sec = tk->xtime_sec;
980 nsec = timekeeping_get_ns(&tk->tkr_mono);
981 tomono = tk->wall_to_monotonic;
983 } while (read_seqcount_retry(&tk_core.seq, seq));
985 ts->tv_sec += tomono.tv_sec;
987 timespec64_add_ns(ts, nsec + tomono.tv_nsec);
989 EXPORT_SYMBOL_GPL(ktime_get_ts64);
992 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
994 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
995 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
996 * works on both 32 and 64 bit systems. On 32 bit systems the readout
997 * covers ~136 years of uptime which should be enough to prevent
998 * premature wrap arounds.
1000 time64_t ktime_get_seconds(void)
1002 struct timekeeper *tk = &tk_core.timekeeper;
1004 WARN_ON(timekeeping_suspended);
1005 return tk->ktime_sec;
1007 EXPORT_SYMBOL_GPL(ktime_get_seconds);
1010 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
1012 * Returns the wall clock seconds since 1970.
1014 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
1015 * 32bit systems the access must be protected with the sequence
1016 * counter to provide "atomic" access to the 64bit tk->xtime_sec
1019 time64_t ktime_get_real_seconds(void)
1021 struct timekeeper *tk = &tk_core.timekeeper;
1025 if (IS_ENABLED(CONFIG_64BIT))
1026 return tk->xtime_sec;
1029 seq = read_seqcount_begin(&tk_core.seq);
1030 seconds = tk->xtime_sec;
1032 } while (read_seqcount_retry(&tk_core.seq, seq));
1036 EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
1039 * __ktime_get_real_seconds - The same as ktime_get_real_seconds
1040 * but without the sequence counter protect. This internal function
1041 * is called just when timekeeping lock is already held.
1043 noinstr time64_t __ktime_get_real_seconds(void)
1045 struct timekeeper *tk = &tk_core.timekeeper;
1047 return tk->xtime_sec;
1051 * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter
1052 * @systime_snapshot: pointer to struct receiving the system time snapshot
1054 void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot)
1056 struct timekeeper *tk = &tk_core.timekeeper;
1064 WARN_ON_ONCE(timekeeping_suspended);
1067 seq = read_seqcount_begin(&tk_core.seq);
1068 now = tk_clock_read(&tk->tkr_mono);
1069 systime_snapshot->cs_id = tk->tkr_mono.clock->id;
1070 systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq;
1071 systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq;
1072 base_real = ktime_add(tk->tkr_mono.base,
1073 tk_core.timekeeper.offs_real);
1074 base_raw = tk->tkr_raw.base;
1075 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now);
1076 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw, now);
1077 } while (read_seqcount_retry(&tk_core.seq, seq));
1079 systime_snapshot->cycles = now;
1080 systime_snapshot->real = ktime_add_ns(base_real, nsec_real);
1081 systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw);
1083 EXPORT_SYMBOL_GPL(ktime_get_snapshot);
1085 /* Scale base by mult/div checking for overflow */
1086 static int scale64_check_overflow(u64 mult, u64 div, u64 *base)
1090 tmp = div64_u64_rem(*base, div, &rem);
1092 if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) ||
1093 ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem)))
1097 rem = div64_u64(rem * mult, div);
1103 * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval
1104 * @history: Snapshot representing start of history
1105 * @partial_history_cycles: Cycle offset into history (fractional part)
1106 * @total_history_cycles: Total history length in cycles
1107 * @discontinuity: True indicates clock was set on history period
1108 * @ts: Cross timestamp that should be adjusted using
1109 * partial/total ratio
1111 * Helper function used by get_device_system_crosststamp() to correct the
1112 * crosstimestamp corresponding to the start of the current interval to the
1113 * system counter value (timestamp point) provided by the driver. The
1114 * total_history_* quantities are the total history starting at the provided
1115 * reference point and ending at the start of the current interval. The cycle
1116 * count between the driver timestamp point and the start of the current
1117 * interval is partial_history_cycles.
1119 static int adjust_historical_crosststamp(struct system_time_snapshot *history,
1120 u64 partial_history_cycles,
1121 u64 total_history_cycles,
1123 struct system_device_crosststamp *ts)
1125 struct timekeeper *tk = &tk_core.timekeeper;
1126 u64 corr_raw, corr_real;
1127 bool interp_forward;
1130 if (total_history_cycles == 0 || partial_history_cycles == 0)
1133 /* Interpolate shortest distance from beginning or end of history */
1134 interp_forward = partial_history_cycles > total_history_cycles / 2;
1135 partial_history_cycles = interp_forward ?
1136 total_history_cycles - partial_history_cycles :
1137 partial_history_cycles;
1140 * Scale the monotonic raw time delta by:
1141 * partial_history_cycles / total_history_cycles
1143 corr_raw = (u64)ktime_to_ns(
1144 ktime_sub(ts->sys_monoraw, history->raw));
1145 ret = scale64_check_overflow(partial_history_cycles,
1146 total_history_cycles, &corr_raw);
1151 * If there is a discontinuity in the history, scale monotonic raw
1153 * mult(real)/mult(raw) yielding the realtime correction
1154 * Otherwise, calculate the realtime correction similar to monotonic
1157 if (discontinuity) {
1158 corr_real = mul_u64_u32_div
1159 (corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult);
1161 corr_real = (u64)ktime_to_ns(
1162 ktime_sub(ts->sys_realtime, history->real));
1163 ret = scale64_check_overflow(partial_history_cycles,
1164 total_history_cycles, &corr_real);
1169 /* Fixup monotonic raw and real time time values */
1170 if (interp_forward) {
1171 ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw);
1172 ts->sys_realtime = ktime_add_ns(history->real, corr_real);
1174 ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw);
1175 ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real);
1182 * cycle_between - true if test occurs chronologically between before and after
1184 static bool cycle_between(u64 before, u64 test, u64 after)
1186 if (test > before && test < after)
1188 if (test < before && before > after)
1194 * get_device_system_crosststamp - Synchronously capture system/device timestamp
1195 * @get_time_fn: Callback to get simultaneous device time and
1196 * system counter from the device driver
1197 * @ctx: Context passed to get_time_fn()
1198 * @history_begin: Historical reference point used to interpolate system
1199 * time when counter provided by the driver is before the current interval
1200 * @xtstamp: Receives simultaneously captured system and device time
1202 * Reads a timestamp from a device and correlates it to system time
1204 int get_device_system_crosststamp(int (*get_time_fn)
1205 (ktime_t *device_time,
1206 struct system_counterval_t *sys_counterval,
1209 struct system_time_snapshot *history_begin,
1210 struct system_device_crosststamp *xtstamp)
1212 struct system_counterval_t system_counterval;
1213 struct timekeeper *tk = &tk_core.timekeeper;
1214 u64 cycles, now, interval_start;
1215 unsigned int clock_was_set_seq = 0;
1216 ktime_t base_real, base_raw;
1217 u64 nsec_real, nsec_raw;
1218 u8 cs_was_changed_seq;
1224 seq = read_seqcount_begin(&tk_core.seq);
1226 * Try to synchronously capture device time and a system
1227 * counter value calling back into the device driver
1229 ret = get_time_fn(&xtstamp->device, &system_counterval, ctx);
1234 * Verify that the clocksource associated with the captured
1235 * system counter value is the same as the currently installed
1236 * timekeeper clocksource
1238 if (tk->tkr_mono.clock != system_counterval.cs)
1240 cycles = system_counterval.cycles;
1243 * Check whether the system counter value provided by the
1244 * device driver is on the current timekeeping interval.
1246 now = tk_clock_read(&tk->tkr_mono);
1247 interval_start = tk->tkr_mono.cycle_last;
1248 if (!cycle_between(interval_start, cycles, now)) {
1249 clock_was_set_seq = tk->clock_was_set_seq;
1250 cs_was_changed_seq = tk->cs_was_changed_seq;
1251 cycles = interval_start;
1257 base_real = ktime_add(tk->tkr_mono.base,
1258 tk_core.timekeeper.offs_real);
1259 base_raw = tk->tkr_raw.base;
1261 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono,
1262 system_counterval.cycles);
1263 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw,
1264 system_counterval.cycles);
1265 } while (read_seqcount_retry(&tk_core.seq, seq));
1267 xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real);
1268 xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);
1271 * Interpolate if necessary, adjusting back from the start of the
1275 u64 partial_history_cycles, total_history_cycles;
1279 * Check that the counter value occurs after the provided
1280 * history reference and that the history doesn't cross a
1281 * clocksource change
1283 if (!history_begin ||
1284 !cycle_between(history_begin->cycles,
1285 system_counterval.cycles, cycles) ||
1286 history_begin->cs_was_changed_seq != cs_was_changed_seq)
1288 partial_history_cycles = cycles - system_counterval.cycles;
1289 total_history_cycles = cycles - history_begin->cycles;
1291 history_begin->clock_was_set_seq != clock_was_set_seq;
1293 ret = adjust_historical_crosststamp(history_begin,
1294 partial_history_cycles,
1295 total_history_cycles,
1296 discontinuity, xtstamp);
1303 EXPORT_SYMBOL_GPL(get_device_system_crosststamp);
1306 * do_settimeofday64 - Sets the time of day.
1307 * @ts: pointer to the timespec64 variable containing the new time
1309 * Sets the time of day to the new time and update NTP and notify hrtimers
1311 int do_settimeofday64(const struct timespec64 *ts)
1313 struct timekeeper *tk = &tk_core.timekeeper;
1314 struct timespec64 ts_delta, xt;
1315 unsigned long flags;
1318 if (!timespec64_valid_settod(ts))
1321 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1322 write_seqcount_begin(&tk_core.seq);
1324 timekeeping_forward_now(tk);
1327 ts_delta = timespec64_sub(*ts, xt);
1329 if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) {
1334 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
1336 tk_set_xtime(tk, ts);
1338 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1340 write_seqcount_end(&tk_core.seq);
1341 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1343 /* Signal hrtimers about time change */
1344 clock_was_set(CLOCK_SET_WALL);
1347 audit_tk_injoffset(ts_delta);
1351 EXPORT_SYMBOL(do_settimeofday64);
1354 * timekeeping_inject_offset - Adds or subtracts from the current time.
1355 * @ts: Pointer to the timespec variable containing the offset
1357 * Adds or subtracts an offset value from the current time.
1359 static int timekeeping_inject_offset(const struct timespec64 *ts)
1361 struct timekeeper *tk = &tk_core.timekeeper;
1362 unsigned long flags;
1363 struct timespec64 tmp;
1366 if (ts->tv_nsec < 0 || ts->tv_nsec >= NSEC_PER_SEC)
1369 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1370 write_seqcount_begin(&tk_core.seq);
1372 timekeeping_forward_now(tk);
1374 /* Make sure the proposed value is valid */
1375 tmp = timespec64_add(tk_xtime(tk), *ts);
1376 if (timespec64_compare(&tk->wall_to_monotonic, ts) > 0 ||
1377 !timespec64_valid_settod(&tmp)) {
1382 tk_xtime_add(tk, ts);
1383 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *ts));
1385 error: /* even if we error out, we forwarded the time, so call update */
1386 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1388 write_seqcount_end(&tk_core.seq);
1389 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1391 /* Signal hrtimers about time change */
1392 clock_was_set(CLOCK_SET_WALL);
1398 * Indicates if there is an offset between the system clock and the hardware
1399 * clock/persistent clock/rtc.
1401 int persistent_clock_is_local;
1404 * Adjust the time obtained from the CMOS to be UTC time instead of
1407 * This is ugly, but preferable to the alternatives. Otherwise we
1408 * would either need to write a program to do it in /etc/rc (and risk
1409 * confusion if the program gets run more than once; it would also be
1410 * hard to make the program warp the clock precisely n hours) or
1411 * compile in the timezone information into the kernel. Bad, bad....
1415 * The best thing to do is to keep the CMOS clock in universal time (UTC)
1416 * as real UNIX machines always do it. This avoids all headaches about
1417 * daylight saving times and warping kernel clocks.
1419 void timekeeping_warp_clock(void)
1421 if (sys_tz.tz_minuteswest != 0) {
1422 struct timespec64 adjust;
1424 persistent_clock_is_local = 1;
1425 adjust.tv_sec = sys_tz.tz_minuteswest * 60;
1427 timekeeping_inject_offset(&adjust);
1432 * __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic
1434 static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1436 tk->tai_offset = tai_offset;
1437 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1441 * change_clocksource - Swaps clocksources if a new one is available
1443 * Accumulates current time interval and initializes new clocksource
1445 static int change_clocksource(void *data)
1447 struct timekeeper *tk = &tk_core.timekeeper;
1448 struct clocksource *new, *old = NULL;
1449 unsigned long flags;
1450 bool change = false;
1452 new = (struct clocksource *) data;
1455 * If the cs is in module, get a module reference. Succeeds
1456 * for built-in code (owner == NULL) as well.
1458 if (try_module_get(new->owner)) {
1459 if (!new->enable || new->enable(new) == 0)
1462 module_put(new->owner);
1465 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1466 write_seqcount_begin(&tk_core.seq);
1468 timekeeping_forward_now(tk);
1471 old = tk->tkr_mono.clock;
1472 tk_setup_internals(tk, new);
1475 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1477 write_seqcount_end(&tk_core.seq);
1478 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1484 module_put(old->owner);
1491 * timekeeping_notify - Install a new clock source
1492 * @clock: pointer to the clock source
1494 * This function is called from clocksource.c after a new, better clock
1495 * source has been registered. The caller holds the clocksource_mutex.
1497 int timekeeping_notify(struct clocksource *clock)
1499 struct timekeeper *tk = &tk_core.timekeeper;
1501 if (tk->tkr_mono.clock == clock)
1503 stop_machine(change_clocksource, clock, NULL);
1504 tick_clock_notify();
1505 return tk->tkr_mono.clock == clock ? 0 : -1;
1509 * ktime_get_raw_ts64 - Returns the raw monotonic time in a timespec
1510 * @ts: pointer to the timespec64 to be set
1512 * Returns the raw monotonic time (completely un-modified by ntp)
1514 void ktime_get_raw_ts64(struct timespec64 *ts)
1516 struct timekeeper *tk = &tk_core.timekeeper;
1521 seq = read_seqcount_begin(&tk_core.seq);
1522 ts->tv_sec = tk->raw_sec;
1523 nsecs = timekeeping_get_ns(&tk->tkr_raw);
1525 } while (read_seqcount_retry(&tk_core.seq, seq));
1528 timespec64_add_ns(ts, nsecs);
1530 EXPORT_SYMBOL(ktime_get_raw_ts64);
1534 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1536 int timekeeping_valid_for_hres(void)
1538 struct timekeeper *tk = &tk_core.timekeeper;
1543 seq = read_seqcount_begin(&tk_core.seq);
1545 ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1547 } while (read_seqcount_retry(&tk_core.seq, seq));
1553 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1555 u64 timekeeping_max_deferment(void)
1557 struct timekeeper *tk = &tk_core.timekeeper;
1562 seq = read_seqcount_begin(&tk_core.seq);
1564 ret = tk->tkr_mono.clock->max_idle_ns;
1566 } while (read_seqcount_retry(&tk_core.seq, seq));
1572 * read_persistent_clock64 - Return time from the persistent clock.
1573 * @ts: Pointer to the storage for the readout value
1575 * Weak dummy function for arches that do not yet support it.
1576 * Reads the time from the battery backed persistent clock.
1577 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1579 * XXX - Do be sure to remove it once all arches implement it.
1581 void __weak read_persistent_clock64(struct timespec64 *ts)
1588 * read_persistent_wall_and_boot_offset - Read persistent clock, and also offset
1591 * Weak dummy function for arches that do not yet support it.
1592 * @wall_time: - current time as returned by persistent clock
1593 * @boot_offset: - offset that is defined as wall_time - boot_time
1595 * The default function calculates offset based on the current value of
1596 * local_clock(). This way architectures that support sched_clock() but don't
1597 * support dedicated boot time clock will provide the best estimate of the
1601 read_persistent_wall_and_boot_offset(struct timespec64 *wall_time,
1602 struct timespec64 *boot_offset)
1604 read_persistent_clock64(wall_time);
1605 *boot_offset = ns_to_timespec64(local_clock());
1609 * Flag reflecting whether timekeeping_resume() has injected sleeptime.
1611 * The flag starts of false and is only set when a suspend reaches
1612 * timekeeping_suspend(), timekeeping_resume() sets it to false when the
1613 * timekeeper clocksource is not stopping across suspend and has been
1614 * used to update sleep time. If the timekeeper clocksource has stopped
1615 * then the flag stays true and is used by the RTC resume code to decide
1616 * whether sleeptime must be injected and if so the flag gets false then.
1618 * If a suspend fails before reaching timekeeping_resume() then the flag
1619 * stays false and prevents erroneous sleeptime injection.
1621 static bool suspend_timing_needed;
1623 /* Flag for if there is a persistent clock on this platform */
1624 static bool persistent_clock_exists;
1627 * timekeeping_init - Initializes the clocksource and common timekeeping values
1629 void __init timekeeping_init(void)
1631 struct timespec64 wall_time, boot_offset, wall_to_mono;
1632 struct timekeeper *tk = &tk_core.timekeeper;
1633 struct clocksource *clock;
1634 unsigned long flags;
1636 read_persistent_wall_and_boot_offset(&wall_time, &boot_offset);
1637 if (timespec64_valid_settod(&wall_time) &&
1638 timespec64_to_ns(&wall_time) > 0) {
1639 persistent_clock_exists = true;
1640 } else if (timespec64_to_ns(&wall_time) != 0) {
1641 pr_warn("Persistent clock returned invalid value");
1642 wall_time = (struct timespec64){0};
1645 if (timespec64_compare(&wall_time, &boot_offset) < 0)
1646 boot_offset = (struct timespec64){0};
1649 * We want set wall_to_mono, so the following is true:
1650 * wall time + wall_to_mono = boot time
1652 wall_to_mono = timespec64_sub(boot_offset, wall_time);
1654 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1655 write_seqcount_begin(&tk_core.seq);
1658 clock = clocksource_default_clock();
1660 clock->enable(clock);
1661 tk_setup_internals(tk, clock);
1663 tk_set_xtime(tk, &wall_time);
1666 tk_set_wall_to_mono(tk, wall_to_mono);
1668 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1670 write_seqcount_end(&tk_core.seq);
1671 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1674 /* time in seconds when suspend began for persistent clock */
1675 static struct timespec64 timekeeping_suspend_time;
1678 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1679 * @tk: Pointer to the timekeeper to be updated
1680 * @delta: Pointer to the delta value in timespec64 format
1682 * Takes a timespec offset measuring a suspend interval and properly
1683 * adds the sleep offset to the timekeeping variables.
1685 static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1686 const struct timespec64 *delta)
1688 if (!timespec64_valid_strict(delta)) {
1689 printk_deferred(KERN_WARNING
1690 "__timekeeping_inject_sleeptime: Invalid "
1691 "sleep delta value!\n");
1694 tk_xtime_add(tk, delta);
1695 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1696 tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1697 tk_debug_account_sleep_time(delta);
1700 #if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1702 * We have three kinds of time sources to use for sleep time
1703 * injection, the preference order is:
1704 * 1) non-stop clocksource
1705 * 2) persistent clock (ie: RTC accessible when irqs are off)
1708 * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1709 * If system has neither 1) nor 2), 3) will be used finally.
1712 * If timekeeping has injected sleeptime via either 1) or 2),
1713 * 3) becomes needless, so in this case we don't need to call
1714 * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1717 bool timekeeping_rtc_skipresume(void)
1719 return !suspend_timing_needed;
1723 * 1) can be determined whether to use or not only when doing
1724 * timekeeping_resume() which is invoked after rtc_suspend(),
1725 * so we can't skip rtc_suspend() surely if system has 1).
1727 * But if system has 2), 2) will definitely be used, so in this
1728 * case we don't need to call rtc_suspend(), and this is what
1729 * timekeeping_rtc_skipsuspend() means.
1731 bool timekeeping_rtc_skipsuspend(void)
1733 return persistent_clock_exists;
1737 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1738 * @delta: pointer to a timespec64 delta value
1740 * This hook is for architectures that cannot support read_persistent_clock64
1741 * because their RTC/persistent clock is only accessible when irqs are enabled.
1742 * and also don't have an effective nonstop clocksource.
1744 * This function should only be called by rtc_resume(), and allows
1745 * a suspend offset to be injected into the timekeeping values.
1747 void timekeeping_inject_sleeptime64(const struct timespec64 *delta)
1749 struct timekeeper *tk = &tk_core.timekeeper;
1750 unsigned long flags;
1752 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1753 write_seqcount_begin(&tk_core.seq);
1755 suspend_timing_needed = false;
1757 timekeeping_forward_now(tk);
1759 __timekeeping_inject_sleeptime(tk, delta);
1761 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1763 write_seqcount_end(&tk_core.seq);
1764 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1766 /* Signal hrtimers about time change */
1767 clock_was_set(CLOCK_SET_WALL | CLOCK_SET_BOOT);
1772 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1774 void timekeeping_resume(void)
1776 struct timekeeper *tk = &tk_core.timekeeper;
1777 struct clocksource *clock = tk->tkr_mono.clock;
1778 unsigned long flags;
1779 struct timespec64 ts_new, ts_delta;
1780 u64 cycle_now, nsec;
1781 bool inject_sleeptime = false;
1783 read_persistent_clock64(&ts_new);
1785 clockevents_resume();
1786 clocksource_resume();
1788 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1789 write_seqcount_begin(&tk_core.seq);
1792 * After system resumes, we need to calculate the suspended time and
1793 * compensate it for the OS time. There are 3 sources that could be
1794 * used: Nonstop clocksource during suspend, persistent clock and rtc
1797 * One specific platform may have 1 or 2 or all of them, and the
1798 * preference will be:
1799 * suspend-nonstop clocksource -> persistent clock -> rtc
1800 * The less preferred source will only be tried if there is no better
1801 * usable source. The rtc part is handled separately in rtc core code.
1803 cycle_now = tk_clock_read(&tk->tkr_mono);
1804 nsec = clocksource_stop_suspend_timing(clock, cycle_now);
1806 ts_delta = ns_to_timespec64(nsec);
1807 inject_sleeptime = true;
1808 } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1809 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1810 inject_sleeptime = true;
1813 if (inject_sleeptime) {
1814 suspend_timing_needed = false;
1815 __timekeeping_inject_sleeptime(tk, &ts_delta);
1818 /* Re-base the last cycle value */
1819 tk->tkr_mono.cycle_last = cycle_now;
1820 tk->tkr_raw.cycle_last = cycle_now;
1823 timekeeping_suspended = 0;
1824 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1825 write_seqcount_end(&tk_core.seq);
1826 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1828 touch_softlockup_watchdog();
1830 /* Resume the clockevent device(s) and hrtimers */
1832 /* Notify timerfd as resume is equivalent to clock_was_set() */
1836 int timekeeping_suspend(void)
1838 struct timekeeper *tk = &tk_core.timekeeper;
1839 unsigned long flags;
1840 struct timespec64 delta, delta_delta;
1841 static struct timespec64 old_delta;
1842 struct clocksource *curr_clock;
1845 read_persistent_clock64(&timekeeping_suspend_time);
1848 * On some systems the persistent_clock can not be detected at
1849 * timekeeping_init by its return value, so if we see a valid
1850 * value returned, update the persistent_clock_exists flag.
1852 if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1853 persistent_clock_exists = true;
1855 suspend_timing_needed = true;
1857 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1858 write_seqcount_begin(&tk_core.seq);
1859 timekeeping_forward_now(tk);
1860 timekeeping_suspended = 1;
1863 * Since we've called forward_now, cycle_last stores the value
1864 * just read from the current clocksource. Save this to potentially
1865 * use in suspend timing.
1867 curr_clock = tk->tkr_mono.clock;
1868 cycle_now = tk->tkr_mono.cycle_last;
1869 clocksource_start_suspend_timing(curr_clock, cycle_now);
1871 if (persistent_clock_exists) {
1873 * To avoid drift caused by repeated suspend/resumes,
1874 * which each can add ~1 second drift error,
1875 * try to compensate so the difference in system time
1876 * and persistent_clock time stays close to constant.
1878 delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1879 delta_delta = timespec64_sub(delta, old_delta);
1880 if (abs(delta_delta.tv_sec) >= 2) {
1882 * if delta_delta is too large, assume time correction
1883 * has occurred and set old_delta to the current delta.
1887 /* Otherwise try to adjust old_system to compensate */
1888 timekeeping_suspend_time =
1889 timespec64_add(timekeeping_suspend_time, delta_delta);
1893 timekeeping_update(tk, TK_MIRROR);
1894 halt_fast_timekeeper(tk);
1895 write_seqcount_end(&tk_core.seq);
1896 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1899 clocksource_suspend();
1900 clockevents_suspend();
1905 /* sysfs resume/suspend bits for timekeeping */
1906 static struct syscore_ops timekeeping_syscore_ops = {
1907 .resume = timekeeping_resume,
1908 .suspend = timekeeping_suspend,
1911 static int __init timekeeping_init_ops(void)
1913 register_syscore_ops(&timekeeping_syscore_ops);
1916 device_initcall(timekeeping_init_ops);
1919 * Apply a multiplier adjustment to the timekeeper
1921 static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1925 s64 interval = tk->cycle_interval;
1927 if (mult_adj == 0) {
1929 } else if (mult_adj == -1) {
1930 interval = -interval;
1932 } else if (mult_adj != 1) {
1933 interval *= mult_adj;
1938 * So the following can be confusing.
1940 * To keep things simple, lets assume mult_adj == 1 for now.
1942 * When mult_adj != 1, remember that the interval and offset values
1943 * have been appropriately scaled so the math is the same.
1945 * The basic idea here is that we're increasing the multiplier
1946 * by one, this causes the xtime_interval to be incremented by
1947 * one cycle_interval. This is because:
1948 * xtime_interval = cycle_interval * mult
1949 * So if mult is being incremented by one:
1950 * xtime_interval = cycle_interval * (mult + 1)
1952 * xtime_interval = (cycle_interval * mult) + cycle_interval
1953 * Which can be shortened to:
1954 * xtime_interval += cycle_interval
1956 * So offset stores the non-accumulated cycles. Thus the current
1957 * time (in shifted nanoseconds) is:
1958 * now = (offset * adj) + xtime_nsec
1959 * Now, even though we're adjusting the clock frequency, we have
1960 * to keep time consistent. In other words, we can't jump back
1961 * in time, and we also want to avoid jumping forward in time.
1963 * So given the same offset value, we need the time to be the same
1964 * both before and after the freq adjustment.
1965 * now = (offset * adj_1) + xtime_nsec_1
1966 * now = (offset * adj_2) + xtime_nsec_2
1968 * (offset * adj_1) + xtime_nsec_1 =
1969 * (offset * adj_2) + xtime_nsec_2
1973 * (offset * adj_1) + xtime_nsec_1 =
1974 * (offset * (adj_1+1)) + xtime_nsec_2
1975 * (offset * adj_1) + xtime_nsec_1 =
1976 * (offset * adj_1) + offset + xtime_nsec_2
1977 * Canceling the sides:
1978 * xtime_nsec_1 = offset + xtime_nsec_2
1980 * xtime_nsec_2 = xtime_nsec_1 - offset
1981 * Which simplifies to:
1982 * xtime_nsec -= offset
1984 if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1985 /* NTP adjustment caused clocksource mult overflow */
1990 tk->tkr_mono.mult += mult_adj;
1991 tk->xtime_interval += interval;
1992 tk->tkr_mono.xtime_nsec -= offset;
1996 * Adjust the timekeeper's multiplier to the correct frequency
1997 * and also to reduce the accumulated error value.
1999 static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
2004 * Determine the multiplier from the current NTP tick length.
2005 * Avoid expensive division when the tick length doesn't change.
2007 if (likely(tk->ntp_tick == ntp_tick_length())) {
2008 mult = tk->tkr_mono.mult - tk->ntp_err_mult;
2010 tk->ntp_tick = ntp_tick_length();
2011 mult = div64_u64((tk->ntp_tick >> tk->ntp_error_shift) -
2012 tk->xtime_remainder, tk->cycle_interval);
2016 * If the clock is behind the NTP time, increase the multiplier by 1
2017 * to catch up with it. If it's ahead and there was a remainder in the
2018 * tick division, the clock will slow down. Otherwise it will stay
2019 * ahead until the tick length changes to a non-divisible value.
2021 tk->ntp_err_mult = tk->ntp_error > 0 ? 1 : 0;
2022 mult += tk->ntp_err_mult;
2024 timekeeping_apply_adjustment(tk, offset, mult - tk->tkr_mono.mult);
2026 if (unlikely(tk->tkr_mono.clock->maxadj &&
2027 (abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
2028 > tk->tkr_mono.clock->maxadj))) {
2029 printk_once(KERN_WARNING
2030 "Adjusting %s more than 11%% (%ld vs %ld)\n",
2031 tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
2032 (long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
2036 * It may be possible that when we entered this function, xtime_nsec
2037 * was very small. Further, if we're slightly speeding the clocksource
2038 * in the code above, its possible the required corrective factor to
2039 * xtime_nsec could cause it to underflow.
2041 * Now, since we have already accumulated the second and the NTP
2042 * subsystem has been notified via second_overflow(), we need to skip
2045 if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
2046 tk->tkr_mono.xtime_nsec += (u64)NSEC_PER_SEC <<
2049 tk->skip_second_overflow = 1;
2054 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
2056 * Helper function that accumulates the nsecs greater than a second
2057 * from the xtime_nsec field to the xtime_secs field.
2058 * It also calls into the NTP code to handle leapsecond processing.
2060 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
2062 u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
2063 unsigned int clock_set = 0;
2065 while (tk->tkr_mono.xtime_nsec >= nsecps) {
2068 tk->tkr_mono.xtime_nsec -= nsecps;
2072 * Skip NTP update if this second was accumulated before,
2073 * i.e. xtime_nsec underflowed in timekeeping_adjust()
2075 if (unlikely(tk->skip_second_overflow)) {
2076 tk->skip_second_overflow = 0;
2080 /* Figure out if its a leap sec and apply if needed */
2081 leap = second_overflow(tk->xtime_sec);
2082 if (unlikely(leap)) {
2083 struct timespec64 ts;
2085 tk->xtime_sec += leap;
2089 tk_set_wall_to_mono(tk,
2090 timespec64_sub(tk->wall_to_monotonic, ts));
2092 __timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
2094 clock_set = TK_CLOCK_WAS_SET;
2101 * logarithmic_accumulation - shifted accumulation of cycles
2103 * This functions accumulates a shifted interval of cycles into
2104 * a shifted interval nanoseconds. Allows for O(log) accumulation
2107 * Returns the unconsumed cycles.
2109 static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset,
2110 u32 shift, unsigned int *clock_set)
2112 u64 interval = tk->cycle_interval << shift;
2115 /* If the offset is smaller than a shifted interval, do nothing */
2116 if (offset < interval)
2119 /* Accumulate one shifted interval */
2121 tk->tkr_mono.cycle_last += interval;
2122 tk->tkr_raw.cycle_last += interval;
2124 tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
2125 *clock_set |= accumulate_nsecs_to_secs(tk);
2127 /* Accumulate raw time */
2128 tk->tkr_raw.xtime_nsec += tk->raw_interval << shift;
2129 snsec_per_sec = (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
2130 while (tk->tkr_raw.xtime_nsec >= snsec_per_sec) {
2131 tk->tkr_raw.xtime_nsec -= snsec_per_sec;
2135 /* Accumulate error between NTP and clock interval */
2136 tk->ntp_error += tk->ntp_tick << shift;
2137 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
2138 (tk->ntp_error_shift + shift);
2144 * timekeeping_advance - Updates the timekeeper to the current time and
2145 * current NTP tick length
2147 static bool timekeeping_advance(enum timekeeping_adv_mode mode)
2149 struct timekeeper *real_tk = &tk_core.timekeeper;
2150 struct timekeeper *tk = &shadow_timekeeper;
2152 int shift = 0, maxshift;
2153 unsigned int clock_set = 0;
2154 unsigned long flags;
2156 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2158 /* Make sure we're fully resumed: */
2159 if (unlikely(timekeeping_suspended))
2162 offset = clocksource_delta(tk_clock_read(&tk->tkr_mono),
2163 tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
2165 /* Check if there's really nothing to do */
2166 if (offset < real_tk->cycle_interval && mode == TK_ADV_TICK)
2169 /* Do some additional sanity checking */
2170 timekeeping_check_update(tk, offset);
2173 * With NO_HZ we may have to accumulate many cycle_intervals
2174 * (think "ticks") worth of time at once. To do this efficiently,
2175 * we calculate the largest doubling multiple of cycle_intervals
2176 * that is smaller than the offset. We then accumulate that
2177 * chunk in one go, and then try to consume the next smaller
2180 shift = ilog2(offset) - ilog2(tk->cycle_interval);
2181 shift = max(0, shift);
2182 /* Bound shift to one less than what overflows tick_length */
2183 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
2184 shift = min(shift, maxshift);
2185 while (offset >= tk->cycle_interval) {
2186 offset = logarithmic_accumulation(tk, offset, shift,
2188 if (offset < tk->cycle_interval<<shift)
2192 /* Adjust the multiplier to correct NTP error */
2193 timekeeping_adjust(tk, offset);
2196 * Finally, make sure that after the rounding
2197 * xtime_nsec isn't larger than NSEC_PER_SEC
2199 clock_set |= accumulate_nsecs_to_secs(tk);
2201 write_seqcount_begin(&tk_core.seq);
2203 * Update the real timekeeper.
2205 * We could avoid this memcpy by switching pointers, but that
2206 * requires changes to all other timekeeper usage sites as
2207 * well, i.e. move the timekeeper pointer getter into the
2208 * spinlocked/seqcount protected sections. And we trade this
2209 * memcpy under the tk_core.seq against one before we start
2212 timekeeping_update(tk, clock_set);
2213 memcpy(real_tk, tk, sizeof(*tk));
2214 /* The memcpy must come last. Do not put anything here! */
2215 write_seqcount_end(&tk_core.seq);
2217 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2223 * update_wall_time - Uses the current clocksource to increment the wall time
2226 void update_wall_time(void)
2228 if (timekeeping_advance(TK_ADV_TICK))
2229 clock_was_set_delayed();
2233 * getboottime64 - Return the real time of system boot.
2234 * @ts: pointer to the timespec64 to be set
2236 * Returns the wall-time of boot in a timespec64.
2238 * This is based on the wall_to_monotonic offset and the total suspend
2239 * time. Calls to settimeofday will affect the value returned (which
2240 * basically means that however wrong your real time clock is at boot time,
2241 * you get the right time here).
2243 void getboottime64(struct timespec64 *ts)
2245 struct timekeeper *tk = &tk_core.timekeeper;
2246 ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
2248 *ts = ktime_to_timespec64(t);
2250 EXPORT_SYMBOL_GPL(getboottime64);
2252 void ktime_get_coarse_real_ts64(struct timespec64 *ts)
2254 struct timekeeper *tk = &tk_core.timekeeper;
2258 seq = read_seqcount_begin(&tk_core.seq);
2261 } while (read_seqcount_retry(&tk_core.seq, seq));
2263 EXPORT_SYMBOL(ktime_get_coarse_real_ts64);
2265 void ktime_get_coarse_ts64(struct timespec64 *ts)
2267 struct timekeeper *tk = &tk_core.timekeeper;
2268 struct timespec64 now, mono;
2272 seq = read_seqcount_begin(&tk_core.seq);
2275 mono = tk->wall_to_monotonic;
2276 } while (read_seqcount_retry(&tk_core.seq, seq));
2278 set_normalized_timespec64(ts, now.tv_sec + mono.tv_sec,
2279 now.tv_nsec + mono.tv_nsec);
2281 EXPORT_SYMBOL(ktime_get_coarse_ts64);
2284 * Must hold jiffies_lock
2286 void do_timer(unsigned long ticks)
2288 jiffies_64 += ticks;
2293 * ktime_get_update_offsets_now - hrtimer helper
2294 * @cwsseq: pointer to check and store the clock was set sequence number
2295 * @offs_real: pointer to storage for monotonic -> realtime offset
2296 * @offs_boot: pointer to storage for monotonic -> boottime offset
2297 * @offs_tai: pointer to storage for monotonic -> clock tai offset
2299 * Returns current monotonic time and updates the offsets if the
2300 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
2303 * Called from hrtimer_interrupt() or retrigger_next_event()
2305 ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
2306 ktime_t *offs_boot, ktime_t *offs_tai)
2308 struct timekeeper *tk = &tk_core.timekeeper;
2314 seq = read_seqcount_begin(&tk_core.seq);
2316 base = tk->tkr_mono.base;
2317 nsecs = timekeeping_get_ns(&tk->tkr_mono);
2318 base = ktime_add_ns(base, nsecs);
2320 if (*cwsseq != tk->clock_was_set_seq) {
2321 *cwsseq = tk->clock_was_set_seq;
2322 *offs_real = tk->offs_real;
2323 *offs_boot = tk->offs_boot;
2324 *offs_tai = tk->offs_tai;
2327 /* Handle leapsecond insertion adjustments */
2328 if (unlikely(base >= tk->next_leap_ktime))
2329 *offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
2331 } while (read_seqcount_retry(&tk_core.seq, seq));
2337 * timekeeping_validate_timex - Ensures the timex is ok for use in do_adjtimex
2339 static int timekeeping_validate_timex(const struct __kernel_timex *txc)
2341 if (txc->modes & ADJ_ADJTIME) {
2342 /* singleshot must not be used with any other mode bits */
2343 if (!(txc->modes & ADJ_OFFSET_SINGLESHOT))
2345 if (!(txc->modes & ADJ_OFFSET_READONLY) &&
2346 !capable(CAP_SYS_TIME))
2349 /* In order to modify anything, you gotta be super-user! */
2350 if (txc->modes && !capable(CAP_SYS_TIME))
2353 * if the quartz is off by more than 10% then
2354 * something is VERY wrong!
2356 if (txc->modes & ADJ_TICK &&
2357 (txc->tick < 900000/USER_HZ ||
2358 txc->tick > 1100000/USER_HZ))
2362 if (txc->modes & ADJ_SETOFFSET) {
2363 /* In order to inject time, you gotta be super-user! */
2364 if (!capable(CAP_SYS_TIME))
2368 * Validate if a timespec/timeval used to inject a time
2369 * offset is valid. Offsets can be positive or negative, so
2370 * we don't check tv_sec. The value of the timeval/timespec
2371 * is the sum of its fields,but *NOTE*:
2372 * The field tv_usec/tv_nsec must always be non-negative and
2373 * we can't have more nanoseconds/microseconds than a second.
2375 if (txc->time.tv_usec < 0)
2378 if (txc->modes & ADJ_NANO) {
2379 if (txc->time.tv_usec >= NSEC_PER_SEC)
2382 if (txc->time.tv_usec >= USEC_PER_SEC)
2388 * Check for potential multiplication overflows that can
2389 * only happen on 64-bit systems:
2391 if ((txc->modes & ADJ_FREQUENCY) && (BITS_PER_LONG == 64)) {
2392 if (LLONG_MIN / PPM_SCALE > txc->freq)
2394 if (LLONG_MAX / PPM_SCALE < txc->freq)
2402 * random_get_entropy_fallback - Returns the raw clock source value,
2403 * used by random.c for platforms with no valid random_get_entropy().
2405 unsigned long random_get_entropy_fallback(void)
2407 struct tk_read_base *tkr = &tk_core.timekeeper.tkr_mono;
2408 struct clocksource *clock = READ_ONCE(tkr->clock);
2410 if (unlikely(timekeeping_suspended || !clock))
2412 return clock->read(clock);
2414 EXPORT_SYMBOL_GPL(random_get_entropy_fallback);
2417 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2419 int do_adjtimex(struct __kernel_timex *txc)
2421 struct timekeeper *tk = &tk_core.timekeeper;
2422 struct audit_ntp_data ad;
2423 bool clock_set = false;
2424 struct timespec64 ts;
2425 unsigned long flags;
2429 /* Validate the data before disabling interrupts */
2430 ret = timekeeping_validate_timex(txc);
2434 if (txc->modes & ADJ_SETOFFSET) {
2435 struct timespec64 delta;
2436 delta.tv_sec = txc->time.tv_sec;
2437 delta.tv_nsec = txc->time.tv_usec;
2438 if (!(txc->modes & ADJ_NANO))
2439 delta.tv_nsec *= 1000;
2440 ret = timekeeping_inject_offset(&delta);
2444 audit_tk_injoffset(delta);
2447 audit_ntp_init(&ad);
2449 ktime_get_real_ts64(&ts);
2451 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2452 write_seqcount_begin(&tk_core.seq);
2454 orig_tai = tai = tk->tai_offset;
2455 ret = __do_adjtimex(txc, &ts, &tai, &ad);
2457 if (tai != orig_tai) {
2458 __timekeeping_set_tai_offset(tk, tai);
2459 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
2462 tk_update_leap_state(tk);
2464 write_seqcount_end(&tk_core.seq);
2465 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2469 /* Update the multiplier immediately if frequency was set directly */
2470 if (txc->modes & (ADJ_FREQUENCY | ADJ_TICK))
2471 clock_set |= timekeeping_advance(TK_ADV_FREQ);
2474 clock_was_set(CLOCK_REALTIME);
2476 ntp_notify_cmos_timer();
2481 #ifdef CONFIG_NTP_PPS
2483 * hardpps() - Accessor function to NTP __hardpps function
2485 void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2487 unsigned long flags;
2489 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2490 write_seqcount_begin(&tk_core.seq);
2492 __hardpps(phase_ts, raw_ts);
2494 write_seqcount_end(&tk_core.seq);
2495 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2497 EXPORT_SYMBOL(hardpps);
2498 #endif /* CONFIG_NTP_PPS */