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>
26 #include <linux/random.h>
28 #include "tick-internal.h"
29 #include "ntp_internal.h"
30 #include "timekeeping_internal.h"
32 #define TK_CLEAR_NTP (1 << 0)
33 #define TK_MIRROR (1 << 1)
34 #define TK_CLOCK_WAS_SET (1 << 2)
36 enum timekeeping_adv_mode {
37 /* Update timekeeper when a tick has passed */
40 /* Update timekeeper on a direct frequency change */
44 DEFINE_RAW_SPINLOCK(timekeeper_lock);
47 * The most important data for readout fits into a single 64 byte
51 seqcount_raw_spinlock_t seq;
52 struct timekeeper timekeeper;
53 } tk_core ____cacheline_aligned = {
54 .seq = SEQCNT_RAW_SPINLOCK_ZERO(tk_core.seq, &timekeeper_lock),
57 static struct timekeeper shadow_timekeeper;
59 /* flag for if timekeeping is suspended */
60 int __read_mostly timekeeping_suspended;
63 * struct tk_fast - NMI safe timekeeper
64 * @seq: Sequence counter for protecting updates. The lowest bit
65 * is the index for the tk_read_base array
66 * @base: tk_read_base array. Access is indexed by the lowest bit of
69 * See @update_fast_timekeeper() below.
73 struct tk_read_base base[2];
76 /* Suspend-time cycles value for halted fast timekeeper. */
77 static u64 cycles_at_suspend;
79 static u64 dummy_clock_read(struct clocksource *cs)
81 if (timekeeping_suspended)
82 return cycles_at_suspend;
86 static struct clocksource dummy_clock = {
87 .read = dummy_clock_read,
91 * Boot time initialization which allows local_clock() to be utilized
92 * during early boot when clocksources are not available. local_clock()
93 * returns nanoseconds already so no conversion is required, hence mult=1
94 * and shift=0. When the first proper clocksource is installed then
95 * the fast time keepers are updated with the correct values.
97 #define FAST_TK_INIT \
99 .clock = &dummy_clock, \
100 .mask = CLOCKSOURCE_MASK(64), \
105 static struct tk_fast tk_fast_mono ____cacheline_aligned = {
106 .seq = SEQCNT_LATCH_ZERO(tk_fast_mono.seq),
107 .base[0] = FAST_TK_INIT,
108 .base[1] = FAST_TK_INIT,
111 static struct tk_fast tk_fast_raw ____cacheline_aligned = {
112 .seq = SEQCNT_LATCH_ZERO(tk_fast_raw.seq),
113 .base[0] = FAST_TK_INIT,
114 .base[1] = FAST_TK_INIT,
117 static inline void tk_normalize_xtime(struct timekeeper *tk)
119 while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) {
120 tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
123 while (tk->tkr_raw.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_raw.shift)) {
124 tk->tkr_raw.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
129 static inline struct timespec64 tk_xtime(const struct timekeeper *tk)
131 struct timespec64 ts;
133 ts.tv_sec = tk->xtime_sec;
134 ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
138 static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
140 tk->xtime_sec = ts->tv_sec;
141 tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift;
144 static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
146 tk->xtime_sec += ts->tv_sec;
147 tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift;
148 tk_normalize_xtime(tk);
151 static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
153 struct timespec64 tmp;
156 * Verify consistency of: offset_real = -wall_to_monotonic
157 * before modifying anything
159 set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
160 -tk->wall_to_monotonic.tv_nsec);
161 WARN_ON_ONCE(tk->offs_real != timespec64_to_ktime(tmp));
162 tk->wall_to_monotonic = wtm;
163 set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
164 tk->offs_real = timespec64_to_ktime(tmp);
165 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0));
168 static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
170 tk->offs_boot = ktime_add(tk->offs_boot, delta);
172 * Timespec representation for VDSO update to avoid 64bit division
175 tk->monotonic_to_boot = ktime_to_timespec64(tk->offs_boot);
179 * tk_clock_read - atomic clocksource read() helper
181 * This helper is necessary to use in the read paths because, while the
182 * seqcount ensures we don't return a bad value while structures are updated,
183 * it doesn't protect from potential crashes. There is the possibility that
184 * the tkr's clocksource may change between the read reference, and the
185 * clock reference passed to the read function. This can cause crashes if
186 * the wrong clocksource is passed to the wrong read function.
187 * This isn't necessary to use when holding the timekeeper_lock or doing
188 * a read of the fast-timekeeper tkrs (which is protected by its own locking
191 static inline u64 tk_clock_read(const struct tk_read_base *tkr)
193 struct clocksource *clock = READ_ONCE(tkr->clock);
195 return clock->read(clock);
198 #ifdef CONFIG_DEBUG_TIMEKEEPING
199 #define WARNING_FREQ (HZ*300) /* 5 minute rate-limiting */
201 static void timekeeping_check_update(struct timekeeper *tk, u64 offset)
204 u64 max_cycles = tk->tkr_mono.clock->max_cycles;
205 const char *name = tk->tkr_mono.clock->name;
207 if (offset > max_cycles) {
208 printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n",
209 offset, name, max_cycles);
210 printk_deferred(" timekeeping: Your kernel is sick, but tries to cope by capping time updates\n");
212 if (offset > (max_cycles >> 1)) {
213 printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the '%s' clock's 50%% safety margin (%lld)\n",
214 offset, name, max_cycles >> 1);
215 printk_deferred(" timekeeping: Your kernel is still fine, but is feeling a bit nervous\n");
219 if (tk->underflow_seen) {
220 if (jiffies - tk->last_warning > WARNING_FREQ) {
221 printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name);
222 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
223 printk_deferred(" Your kernel is probably still fine.\n");
224 tk->last_warning = jiffies;
226 tk->underflow_seen = 0;
229 if (tk->overflow_seen) {
230 if (jiffies - tk->last_warning > WARNING_FREQ) {
231 printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name);
232 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
233 printk_deferred(" Your kernel is probably still fine.\n");
234 tk->last_warning = jiffies;
236 tk->overflow_seen = 0;
240 static inline u64 timekeeping_get_delta(const struct tk_read_base *tkr)
242 struct timekeeper *tk = &tk_core.timekeeper;
243 u64 now, last, mask, max, delta;
247 * Since we're called holding a seqcount, the data may shift
248 * under us while we're doing the calculation. This can cause
249 * false positives, since we'd note a problem but throw the
250 * results away. So nest another seqcount here to atomically
251 * grab the points we are checking with.
254 seq = read_seqcount_begin(&tk_core.seq);
255 now = tk_clock_read(tkr);
256 last = tkr->cycle_last;
258 max = tkr->clock->max_cycles;
259 } while (read_seqcount_retry(&tk_core.seq, seq));
261 delta = clocksource_delta(now, last, mask);
264 * Try to catch underflows by checking if we are seeing small
265 * mask-relative negative values.
267 if (unlikely((~delta & mask) < (mask >> 3))) {
268 tk->underflow_seen = 1;
272 /* Cap delta value to the max_cycles values to avoid mult overflows */
273 if (unlikely(delta > max)) {
274 tk->overflow_seen = 1;
275 delta = tkr->clock->max_cycles;
281 static inline void timekeeping_check_update(struct timekeeper *tk, u64 offset)
284 static inline u64 timekeeping_get_delta(const struct tk_read_base *tkr)
286 u64 cycle_now, delta;
288 /* read clocksource */
289 cycle_now = tk_clock_read(tkr);
291 /* calculate the delta since the last update_wall_time */
292 delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
299 * tk_setup_internals - Set up internals to use clocksource clock.
301 * @tk: The target timekeeper to setup.
302 * @clock: Pointer to clocksource.
304 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
305 * pair and interval request.
307 * Unless you're the timekeeping code, you should not be using this!
309 static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
312 u64 tmp, ntpinterval;
313 struct clocksource *old_clock;
315 ++tk->cs_was_changed_seq;
316 old_clock = tk->tkr_mono.clock;
317 tk->tkr_mono.clock = clock;
318 tk->tkr_mono.mask = clock->mask;
319 tk->tkr_mono.cycle_last = tk_clock_read(&tk->tkr_mono);
321 tk->tkr_raw.clock = clock;
322 tk->tkr_raw.mask = clock->mask;
323 tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
325 /* Do the ns -> cycle conversion first, using original mult */
326 tmp = NTP_INTERVAL_LENGTH;
327 tmp <<= clock->shift;
329 tmp += clock->mult/2;
330 do_div(tmp, clock->mult);
334 interval = (u64) tmp;
335 tk->cycle_interval = interval;
337 /* Go back from cycles -> shifted ns */
338 tk->xtime_interval = interval * clock->mult;
339 tk->xtime_remainder = ntpinterval - tk->xtime_interval;
340 tk->raw_interval = interval * clock->mult;
342 /* if changing clocks, convert xtime_nsec shift units */
344 int shift_change = clock->shift - old_clock->shift;
345 if (shift_change < 0) {
346 tk->tkr_mono.xtime_nsec >>= -shift_change;
347 tk->tkr_raw.xtime_nsec >>= -shift_change;
349 tk->tkr_mono.xtime_nsec <<= shift_change;
350 tk->tkr_raw.xtime_nsec <<= shift_change;
354 tk->tkr_mono.shift = clock->shift;
355 tk->tkr_raw.shift = clock->shift;
358 tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
359 tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
362 * The timekeeper keeps its own mult values for the currently
363 * active clocksource. These value will be adjusted via NTP
364 * to counteract clock drifting.
366 tk->tkr_mono.mult = clock->mult;
367 tk->tkr_raw.mult = clock->mult;
368 tk->ntp_err_mult = 0;
369 tk->skip_second_overflow = 0;
372 /* Timekeeper helper functions. */
374 static inline u64 timekeeping_delta_to_ns(const struct tk_read_base *tkr, u64 delta)
378 nsec = delta * tkr->mult + tkr->xtime_nsec;
384 static inline u64 timekeeping_get_ns(const struct tk_read_base *tkr)
388 delta = timekeeping_get_delta(tkr);
389 return timekeeping_delta_to_ns(tkr, delta);
392 static inline u64 timekeeping_cycles_to_ns(const struct tk_read_base *tkr, u64 cycles)
396 /* calculate the delta since the last update_wall_time */
397 delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask);
398 return timekeeping_delta_to_ns(tkr, delta);
402 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
403 * @tkr: Timekeeping readout base from which we take the update
404 * @tkf: Pointer to NMI safe timekeeper
406 * We want to use this from any context including NMI and tracing /
407 * instrumenting the timekeeping code itself.
409 * Employ the latch technique; see @raw_write_seqcount_latch.
411 * So if a NMI hits the update of base[0] then it will use base[1]
412 * which is still consistent. In the worst case this can result is a
413 * slightly wrong timestamp (a few nanoseconds). See
414 * @ktime_get_mono_fast_ns.
416 static void update_fast_timekeeper(const struct tk_read_base *tkr,
419 struct tk_read_base *base = tkf->base;
421 /* Force readers off to base[1] */
422 raw_write_seqcount_latch(&tkf->seq);
425 memcpy(base, tkr, sizeof(*base));
427 /* Force readers back to base[0] */
428 raw_write_seqcount_latch(&tkf->seq);
431 memcpy(base + 1, base, sizeof(*base));
434 static __always_inline u64 fast_tk_get_delta_ns(struct tk_read_base *tkr)
436 u64 delta, cycles = tk_clock_read(tkr);
438 delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask);
439 return timekeeping_delta_to_ns(tkr, delta);
442 static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
444 struct tk_read_base *tkr;
449 seq = raw_read_seqcount_latch(&tkf->seq);
450 tkr = tkf->base + (seq & 0x01);
451 now = ktime_to_ns(tkr->base);
452 now += fast_tk_get_delta_ns(tkr);
453 } while (raw_read_seqcount_latch_retry(&tkf->seq, seq));
459 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
461 * This timestamp is not guaranteed to be monotonic across an update.
462 * The timestamp is calculated by:
464 * now = base_mono + clock_delta * slope
466 * So if the update lowers the slope, readers who are forced to the
467 * not yet updated second array are still using the old steeper slope.
476 * |12345678---> reader order
482 * So reader 6 will observe time going backwards versus reader 5.
484 * While other CPUs are likely to be able to observe that, the only way
485 * for a CPU local observation is when an NMI hits in the middle of
486 * the update. Timestamps taken from that NMI context might be ahead
487 * of the following timestamps. Callers need to be aware of that and
490 u64 notrace ktime_get_mono_fast_ns(void)
492 return __ktime_get_fast_ns(&tk_fast_mono);
494 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
497 * ktime_get_raw_fast_ns - Fast NMI safe access to clock monotonic raw
499 * Contrary to ktime_get_mono_fast_ns() this is always correct because the
500 * conversion factor is not affected by NTP/PTP correction.
502 u64 notrace ktime_get_raw_fast_ns(void)
504 return __ktime_get_fast_ns(&tk_fast_raw);
506 EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
509 * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock.
511 * To keep it NMI safe since we're accessing from tracing, we're not using a
512 * separate timekeeper with updates to monotonic clock and boot offset
513 * protected with seqcounts. This has the following minor side effects:
515 * (1) Its possible that a timestamp be taken after the boot offset is updated
516 * but before the timekeeper is updated. If this happens, the new boot offset
517 * is added to the old timekeeping making the clock appear to update slightly
520 * timekeeping_inject_sleeptime64()
521 * __timekeeping_inject_sleeptime(tk, delta);
523 * timekeeping_update(tk, TK_CLEAR_NTP...);
525 * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be
526 * partially updated. Since the tk->offs_boot update is a rare event, this
527 * should be a rare occurrence which postprocessing should be able to handle.
529 * The caveats vs. timestamp ordering as documented for ktime_get_mono_fast_ns()
532 u64 notrace ktime_get_boot_fast_ns(void)
534 struct timekeeper *tk = &tk_core.timekeeper;
536 return (ktime_get_mono_fast_ns() + ktime_to_ns(data_race(tk->offs_boot)));
538 EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns);
541 * ktime_get_tai_fast_ns - NMI safe and fast access to tai clock.
543 * The same limitations as described for ktime_get_boot_fast_ns() apply. The
544 * mono time and the TAI offset are not read atomically which may yield wrong
545 * readouts. However, an update of the TAI offset is an rare event e.g., caused
546 * by settime or adjtimex with an offset. The user of this function has to deal
547 * with the possibility of wrong timestamps in post processing.
549 u64 notrace ktime_get_tai_fast_ns(void)
551 struct timekeeper *tk = &tk_core.timekeeper;
553 return (ktime_get_mono_fast_ns() + ktime_to_ns(data_race(tk->offs_tai)));
555 EXPORT_SYMBOL_GPL(ktime_get_tai_fast_ns);
557 static __always_inline u64 __ktime_get_real_fast(struct tk_fast *tkf, u64 *mono)
559 struct tk_read_base *tkr;
560 u64 basem, baser, delta;
564 seq = raw_read_seqcount_latch(&tkf->seq);
565 tkr = tkf->base + (seq & 0x01);
566 basem = ktime_to_ns(tkr->base);
567 baser = ktime_to_ns(tkr->base_real);
568 delta = fast_tk_get_delta_ns(tkr);
569 } while (raw_read_seqcount_latch_retry(&tkf->seq, seq));
572 *mono = basem + delta;
573 return baser + delta;
577 * ktime_get_real_fast_ns: - NMI safe and fast access to clock realtime.
579 * See ktime_get_mono_fast_ns() for documentation of the time stamp ordering.
581 u64 ktime_get_real_fast_ns(void)
583 return __ktime_get_real_fast(&tk_fast_mono, NULL);
585 EXPORT_SYMBOL_GPL(ktime_get_real_fast_ns);
588 * ktime_get_fast_timestamps: - NMI safe timestamps
589 * @snapshot: Pointer to timestamp storage
591 * Stores clock monotonic, boottime and realtime timestamps.
593 * Boot time is a racy access on 32bit systems if the sleep time injection
594 * happens late during resume and not in timekeeping_resume(). That could
595 * be avoided by expanding struct tk_read_base with boot offset for 32bit
596 * and adding more overhead to the update. As this is a hard to observe
597 * once per resume event which can be filtered with reasonable effort using
598 * the accurate mono/real timestamps, it's probably not worth the trouble.
600 * Aside of that it might be possible on 32 and 64 bit to observe the
601 * following when the sleep time injection happens late:
604 * timekeeping_resume()
605 * ktime_get_fast_timestamps()
606 * mono, real = __ktime_get_real_fast()
607 * inject_sleep_time()
609 * boot = mono + bootoffset;
611 * That means that boot time already has the sleep time adjustment, but
612 * real time does not. On the next readout both are in sync again.
614 * Preventing this for 64bit is not really feasible without destroying the
615 * careful cache layout of the timekeeper because the sequence count and
616 * struct tk_read_base would then need two cache lines instead of one.
618 * Access to the time keeper clock source is disabled across the innermost
619 * steps of suspend/resume. The accessors still work, but the timestamps
620 * are frozen until time keeping is resumed which happens very early.
622 * For regular suspend/resume there is no observable difference vs. sched
623 * clock, but it might affect some of the nasty low level debug printks.
625 * OTOH, access to sched clock is not guaranteed across suspend/resume on
626 * all systems either so it depends on the hardware in use.
628 * If that turns out to be a real problem then this could be mitigated by
629 * using sched clock in a similar way as during early boot. But it's not as
630 * trivial as on early boot because it needs some careful protection
631 * against the clock monotonic timestamp jumping backwards on resume.
633 void ktime_get_fast_timestamps(struct ktime_timestamps *snapshot)
635 struct timekeeper *tk = &tk_core.timekeeper;
637 snapshot->real = __ktime_get_real_fast(&tk_fast_mono, &snapshot->mono);
638 snapshot->boot = snapshot->mono + ktime_to_ns(data_race(tk->offs_boot));
642 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
643 * @tk: Timekeeper to snapshot.
645 * It generally is unsafe to access the clocksource after timekeeping has been
646 * suspended, so take a snapshot of the readout base of @tk and use it as the
647 * fast timekeeper's readout base while suspended. It will return the same
648 * number of cycles every time until timekeeping is resumed at which time the
649 * proper readout base for the fast timekeeper will be restored automatically.
651 static void halt_fast_timekeeper(const struct timekeeper *tk)
653 static struct tk_read_base tkr_dummy;
654 const struct tk_read_base *tkr = &tk->tkr_mono;
656 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
657 cycles_at_suspend = tk_clock_read(tkr);
658 tkr_dummy.clock = &dummy_clock;
659 tkr_dummy.base_real = tkr->base + tk->offs_real;
660 update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
663 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
664 tkr_dummy.clock = &dummy_clock;
665 update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
668 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
670 static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
672 raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
676 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
677 * @nb: Pointer to the notifier block to register
679 int pvclock_gtod_register_notifier(struct notifier_block *nb)
681 struct timekeeper *tk = &tk_core.timekeeper;
685 raw_spin_lock_irqsave(&timekeeper_lock, flags);
686 ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
687 update_pvclock_gtod(tk, true);
688 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
692 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
695 * pvclock_gtod_unregister_notifier - unregister a pvclock
696 * timedata update listener
697 * @nb: Pointer to the notifier block to unregister
699 int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
704 raw_spin_lock_irqsave(&timekeeper_lock, flags);
705 ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
706 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
710 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
713 * tk_update_leap_state - helper to update the next_leap_ktime
715 static inline void tk_update_leap_state(struct timekeeper *tk)
717 tk->next_leap_ktime = ntp_get_next_leap();
718 if (tk->next_leap_ktime != KTIME_MAX)
719 /* Convert to monotonic time */
720 tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
724 * Update the ktime_t based scalar nsec members of the timekeeper
726 static inline void tk_update_ktime_data(struct timekeeper *tk)
732 * The xtime based monotonic readout is:
733 * nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
734 * The ktime based monotonic readout is:
735 * nsec = base_mono + now();
736 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
738 seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
739 nsec = (u32) tk->wall_to_monotonic.tv_nsec;
740 tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
743 * The sum of the nanoseconds portions of xtime and
744 * wall_to_monotonic can be greater/equal one second. Take
745 * this into account before updating tk->ktime_sec.
747 nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
748 if (nsec >= NSEC_PER_SEC)
750 tk->ktime_sec = seconds;
752 /* Update the monotonic raw base */
753 tk->tkr_raw.base = ns_to_ktime(tk->raw_sec * NSEC_PER_SEC);
756 /* must hold timekeeper_lock */
757 static void timekeeping_update(struct timekeeper *tk, unsigned int action)
759 if (action & TK_CLEAR_NTP) {
764 tk_update_leap_state(tk);
765 tk_update_ktime_data(tk);
768 update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
770 tk->tkr_mono.base_real = tk->tkr_mono.base + tk->offs_real;
771 update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
772 update_fast_timekeeper(&tk->tkr_raw, &tk_fast_raw);
774 if (action & TK_CLOCK_WAS_SET)
775 tk->clock_was_set_seq++;
777 * The mirroring of the data to the shadow-timekeeper needs
778 * to happen last here to ensure we don't over-write the
779 * timekeeper structure on the next update with stale data
781 if (action & TK_MIRROR)
782 memcpy(&shadow_timekeeper, &tk_core.timekeeper,
783 sizeof(tk_core.timekeeper));
787 * timekeeping_forward_now - update clock to the current time
788 * @tk: Pointer to the timekeeper to update
790 * Forward the current clock to update its state since the last call to
791 * update_wall_time(). This is useful before significant clock changes,
792 * as it avoids having to deal with this time offset explicitly.
794 static void timekeeping_forward_now(struct timekeeper *tk)
796 u64 cycle_now, delta;
798 cycle_now = tk_clock_read(&tk->tkr_mono);
799 delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
800 tk->tkr_mono.cycle_last = cycle_now;
801 tk->tkr_raw.cycle_last = cycle_now;
803 tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult;
804 tk->tkr_raw.xtime_nsec += delta * tk->tkr_raw.mult;
806 tk_normalize_xtime(tk);
810 * ktime_get_real_ts64 - Returns the time of day in a timespec64.
811 * @ts: pointer to the timespec to be set
813 * Returns the time of day in a timespec64 (WARN if suspended).
815 void ktime_get_real_ts64(struct timespec64 *ts)
817 struct timekeeper *tk = &tk_core.timekeeper;
821 WARN_ON(timekeeping_suspended);
824 seq = read_seqcount_begin(&tk_core.seq);
826 ts->tv_sec = tk->xtime_sec;
827 nsecs = timekeeping_get_ns(&tk->tkr_mono);
829 } while (read_seqcount_retry(&tk_core.seq, seq));
832 timespec64_add_ns(ts, nsecs);
834 EXPORT_SYMBOL(ktime_get_real_ts64);
836 ktime_t ktime_get(void)
838 struct timekeeper *tk = &tk_core.timekeeper;
843 WARN_ON(timekeeping_suspended);
846 seq = read_seqcount_begin(&tk_core.seq);
847 base = tk->tkr_mono.base;
848 nsecs = timekeeping_get_ns(&tk->tkr_mono);
850 } while (read_seqcount_retry(&tk_core.seq, seq));
852 return ktime_add_ns(base, nsecs);
854 EXPORT_SYMBOL_GPL(ktime_get);
856 u32 ktime_get_resolution_ns(void)
858 struct timekeeper *tk = &tk_core.timekeeper;
862 WARN_ON(timekeeping_suspended);
865 seq = read_seqcount_begin(&tk_core.seq);
866 nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
867 } while (read_seqcount_retry(&tk_core.seq, seq));
871 EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
873 static ktime_t *offsets[TK_OFFS_MAX] = {
874 [TK_OFFS_REAL] = &tk_core.timekeeper.offs_real,
875 [TK_OFFS_BOOT] = &tk_core.timekeeper.offs_boot,
876 [TK_OFFS_TAI] = &tk_core.timekeeper.offs_tai,
879 ktime_t ktime_get_with_offset(enum tk_offsets offs)
881 struct timekeeper *tk = &tk_core.timekeeper;
883 ktime_t base, *offset = offsets[offs];
886 WARN_ON(timekeeping_suspended);
889 seq = read_seqcount_begin(&tk_core.seq);
890 base = ktime_add(tk->tkr_mono.base, *offset);
891 nsecs = timekeeping_get_ns(&tk->tkr_mono);
893 } while (read_seqcount_retry(&tk_core.seq, seq));
895 return ktime_add_ns(base, nsecs);
898 EXPORT_SYMBOL_GPL(ktime_get_with_offset);
900 ktime_t ktime_get_coarse_with_offset(enum tk_offsets offs)
902 struct timekeeper *tk = &tk_core.timekeeper;
904 ktime_t base, *offset = offsets[offs];
907 WARN_ON(timekeeping_suspended);
910 seq = read_seqcount_begin(&tk_core.seq);
911 base = ktime_add(tk->tkr_mono.base, *offset);
912 nsecs = tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift;
914 } while (read_seqcount_retry(&tk_core.seq, seq));
916 return ktime_add_ns(base, nsecs);
918 EXPORT_SYMBOL_GPL(ktime_get_coarse_with_offset);
921 * ktime_mono_to_any() - convert monotonic time to any other time
922 * @tmono: time to convert.
923 * @offs: which offset to use
925 ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
927 ktime_t *offset = offsets[offs];
932 seq = read_seqcount_begin(&tk_core.seq);
933 tconv = ktime_add(tmono, *offset);
934 } while (read_seqcount_retry(&tk_core.seq, seq));
938 EXPORT_SYMBOL_GPL(ktime_mono_to_any);
941 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
943 ktime_t ktime_get_raw(void)
945 struct timekeeper *tk = &tk_core.timekeeper;
951 seq = read_seqcount_begin(&tk_core.seq);
952 base = tk->tkr_raw.base;
953 nsecs = timekeeping_get_ns(&tk->tkr_raw);
955 } while (read_seqcount_retry(&tk_core.seq, seq));
957 return ktime_add_ns(base, nsecs);
959 EXPORT_SYMBOL_GPL(ktime_get_raw);
962 * ktime_get_ts64 - get the monotonic clock in timespec64 format
963 * @ts: pointer to timespec variable
965 * The function calculates the monotonic clock from the realtime
966 * clock and the wall_to_monotonic offset and stores the result
967 * in normalized timespec64 format in the variable pointed to by @ts.
969 void ktime_get_ts64(struct timespec64 *ts)
971 struct timekeeper *tk = &tk_core.timekeeper;
972 struct timespec64 tomono;
976 WARN_ON(timekeeping_suspended);
979 seq = read_seqcount_begin(&tk_core.seq);
980 ts->tv_sec = tk->xtime_sec;
981 nsec = timekeeping_get_ns(&tk->tkr_mono);
982 tomono = tk->wall_to_monotonic;
984 } while (read_seqcount_retry(&tk_core.seq, seq));
986 ts->tv_sec += tomono.tv_sec;
988 timespec64_add_ns(ts, nsec + tomono.tv_nsec);
990 EXPORT_SYMBOL_GPL(ktime_get_ts64);
993 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
995 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
996 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
997 * works on both 32 and 64 bit systems. On 32 bit systems the readout
998 * covers ~136 years of uptime which should be enough to prevent
999 * premature wrap arounds.
1001 time64_t ktime_get_seconds(void)
1003 struct timekeeper *tk = &tk_core.timekeeper;
1005 WARN_ON(timekeeping_suspended);
1006 return tk->ktime_sec;
1008 EXPORT_SYMBOL_GPL(ktime_get_seconds);
1011 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
1013 * Returns the wall clock seconds since 1970.
1015 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
1016 * 32bit systems the access must be protected with the sequence
1017 * counter to provide "atomic" access to the 64bit tk->xtime_sec
1020 time64_t ktime_get_real_seconds(void)
1022 struct timekeeper *tk = &tk_core.timekeeper;
1026 if (IS_ENABLED(CONFIG_64BIT))
1027 return tk->xtime_sec;
1030 seq = read_seqcount_begin(&tk_core.seq);
1031 seconds = tk->xtime_sec;
1033 } while (read_seqcount_retry(&tk_core.seq, seq));
1037 EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
1040 * __ktime_get_real_seconds - The same as ktime_get_real_seconds
1041 * but without the sequence counter protect. This internal function
1042 * is called just when timekeeping lock is already held.
1044 noinstr time64_t __ktime_get_real_seconds(void)
1046 struct timekeeper *tk = &tk_core.timekeeper;
1048 return tk->xtime_sec;
1052 * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter
1053 * @systime_snapshot: pointer to struct receiving the system time snapshot
1055 void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot)
1057 struct timekeeper *tk = &tk_core.timekeeper;
1065 WARN_ON_ONCE(timekeeping_suspended);
1068 seq = read_seqcount_begin(&tk_core.seq);
1069 now = tk_clock_read(&tk->tkr_mono);
1070 systime_snapshot->cs_id = tk->tkr_mono.clock->id;
1071 systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq;
1072 systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq;
1073 base_real = ktime_add(tk->tkr_mono.base,
1074 tk_core.timekeeper.offs_real);
1075 base_raw = tk->tkr_raw.base;
1076 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now);
1077 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw, now);
1078 } while (read_seqcount_retry(&tk_core.seq, seq));
1080 systime_snapshot->cycles = now;
1081 systime_snapshot->real = ktime_add_ns(base_real, nsec_real);
1082 systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw);
1084 EXPORT_SYMBOL_GPL(ktime_get_snapshot);
1086 /* Scale base by mult/div checking for overflow */
1087 static int scale64_check_overflow(u64 mult, u64 div, u64 *base)
1091 tmp = div64_u64_rem(*base, div, &rem);
1093 if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) ||
1094 ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem)))
1098 rem = div64_u64(rem * mult, div);
1104 * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval
1105 * @history: Snapshot representing start of history
1106 * @partial_history_cycles: Cycle offset into history (fractional part)
1107 * @total_history_cycles: Total history length in cycles
1108 * @discontinuity: True indicates clock was set on history period
1109 * @ts: Cross timestamp that should be adjusted using
1110 * partial/total ratio
1112 * Helper function used by get_device_system_crosststamp() to correct the
1113 * crosstimestamp corresponding to the start of the current interval to the
1114 * system counter value (timestamp point) provided by the driver. The
1115 * total_history_* quantities are the total history starting at the provided
1116 * reference point and ending at the start of the current interval. The cycle
1117 * count between the driver timestamp point and the start of the current
1118 * interval is partial_history_cycles.
1120 static int adjust_historical_crosststamp(struct system_time_snapshot *history,
1121 u64 partial_history_cycles,
1122 u64 total_history_cycles,
1124 struct system_device_crosststamp *ts)
1126 struct timekeeper *tk = &tk_core.timekeeper;
1127 u64 corr_raw, corr_real;
1128 bool interp_forward;
1131 if (total_history_cycles == 0 || partial_history_cycles == 0)
1134 /* Interpolate shortest distance from beginning or end of history */
1135 interp_forward = partial_history_cycles > total_history_cycles / 2;
1136 partial_history_cycles = interp_forward ?
1137 total_history_cycles - partial_history_cycles :
1138 partial_history_cycles;
1141 * Scale the monotonic raw time delta by:
1142 * partial_history_cycles / total_history_cycles
1144 corr_raw = (u64)ktime_to_ns(
1145 ktime_sub(ts->sys_monoraw, history->raw));
1146 ret = scale64_check_overflow(partial_history_cycles,
1147 total_history_cycles, &corr_raw);
1152 * If there is a discontinuity in the history, scale monotonic raw
1154 * mult(real)/mult(raw) yielding the realtime correction
1155 * Otherwise, calculate the realtime correction similar to monotonic
1158 if (discontinuity) {
1159 corr_real = mul_u64_u32_div
1160 (corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult);
1162 corr_real = (u64)ktime_to_ns(
1163 ktime_sub(ts->sys_realtime, history->real));
1164 ret = scale64_check_overflow(partial_history_cycles,
1165 total_history_cycles, &corr_real);
1170 /* Fixup monotonic raw and real time time values */
1171 if (interp_forward) {
1172 ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw);
1173 ts->sys_realtime = ktime_add_ns(history->real, corr_real);
1175 ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw);
1176 ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real);
1183 * timestamp_in_interval - true if ts is chronologically in [start, end]
1185 * True if ts occurs chronologically at or after start, and before or at end.
1187 static bool timestamp_in_interval(u64 start, u64 end, u64 ts)
1189 if (ts >= start && ts <= end)
1191 if (start > end && (ts >= start || ts <= end))
1197 * get_device_system_crosststamp - Synchronously capture system/device timestamp
1198 * @get_time_fn: Callback to get simultaneous device time and
1199 * system counter from the device driver
1200 * @ctx: Context passed to get_time_fn()
1201 * @history_begin: Historical reference point used to interpolate system
1202 * time when counter provided by the driver is before the current interval
1203 * @xtstamp: Receives simultaneously captured system and device time
1205 * Reads a timestamp from a device and correlates it to system time
1207 int get_device_system_crosststamp(int (*get_time_fn)
1208 (ktime_t *device_time,
1209 struct system_counterval_t *sys_counterval,
1212 struct system_time_snapshot *history_begin,
1213 struct system_device_crosststamp *xtstamp)
1215 struct system_counterval_t system_counterval;
1216 struct timekeeper *tk = &tk_core.timekeeper;
1217 u64 cycles, now, interval_start;
1218 unsigned int clock_was_set_seq = 0;
1219 ktime_t base_real, base_raw;
1220 u64 nsec_real, nsec_raw;
1221 u8 cs_was_changed_seq;
1227 seq = read_seqcount_begin(&tk_core.seq);
1229 * Try to synchronously capture device time and a system
1230 * counter value calling back into the device driver
1232 ret = get_time_fn(&xtstamp->device, &system_counterval, ctx);
1237 * Verify that the clocksource ID associated with the captured
1238 * system counter value is the same as for the currently
1239 * installed timekeeper clocksource
1241 if (system_counterval.cs_id == CSID_GENERIC ||
1242 tk->tkr_mono.clock->id != system_counterval.cs_id)
1244 cycles = system_counterval.cycles;
1247 * Check whether the system counter value provided by the
1248 * device driver is on the current timekeeping interval.
1250 now = tk_clock_read(&tk->tkr_mono);
1251 interval_start = tk->tkr_mono.cycle_last;
1252 if (!timestamp_in_interval(interval_start, now, cycles)) {
1253 clock_was_set_seq = tk->clock_was_set_seq;
1254 cs_was_changed_seq = tk->cs_was_changed_seq;
1255 cycles = interval_start;
1261 base_real = ktime_add(tk->tkr_mono.base,
1262 tk_core.timekeeper.offs_real);
1263 base_raw = tk->tkr_raw.base;
1265 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, cycles);
1266 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw, cycles);
1267 } while (read_seqcount_retry(&tk_core.seq, seq));
1269 xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real);
1270 xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);
1273 * Interpolate if necessary, adjusting back from the start of the
1277 u64 partial_history_cycles, total_history_cycles;
1281 * Check that the counter value is not before the provided
1282 * history reference and that the history doesn't cross a
1283 * clocksource change
1285 if (!history_begin ||
1286 !timestamp_in_interval(history_begin->cycles,
1287 cycles, system_counterval.cycles) ||
1288 history_begin->cs_was_changed_seq != cs_was_changed_seq)
1290 partial_history_cycles = cycles - system_counterval.cycles;
1291 total_history_cycles = cycles - history_begin->cycles;
1293 history_begin->clock_was_set_seq != clock_was_set_seq;
1295 ret = adjust_historical_crosststamp(history_begin,
1296 partial_history_cycles,
1297 total_history_cycles,
1298 discontinuity, xtstamp);
1305 EXPORT_SYMBOL_GPL(get_device_system_crosststamp);
1308 * do_settimeofday64 - Sets the time of day.
1309 * @ts: pointer to the timespec64 variable containing the new time
1311 * Sets the time of day to the new time and update NTP and notify hrtimers
1313 int do_settimeofday64(const struct timespec64 *ts)
1315 struct timekeeper *tk = &tk_core.timekeeper;
1316 struct timespec64 ts_delta, xt;
1317 unsigned long flags;
1320 if (!timespec64_valid_settod(ts))
1323 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1324 write_seqcount_begin(&tk_core.seq);
1326 timekeeping_forward_now(tk);
1329 ts_delta = timespec64_sub(*ts, xt);
1331 if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) {
1336 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
1338 tk_set_xtime(tk, ts);
1340 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1342 write_seqcount_end(&tk_core.seq);
1343 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1345 /* Signal hrtimers about time change */
1346 clock_was_set(CLOCK_SET_WALL);
1349 audit_tk_injoffset(ts_delta);
1350 add_device_randomness(ts, sizeof(*ts));
1355 EXPORT_SYMBOL(do_settimeofday64);
1358 * timekeeping_inject_offset - Adds or subtracts from the current time.
1359 * @ts: Pointer to the timespec variable containing the offset
1361 * Adds or subtracts an offset value from the current time.
1363 static int timekeeping_inject_offset(const struct timespec64 *ts)
1365 struct timekeeper *tk = &tk_core.timekeeper;
1366 unsigned long flags;
1367 struct timespec64 tmp;
1370 if (ts->tv_nsec < 0 || ts->tv_nsec >= NSEC_PER_SEC)
1373 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1374 write_seqcount_begin(&tk_core.seq);
1376 timekeeping_forward_now(tk);
1378 /* Make sure the proposed value is valid */
1379 tmp = timespec64_add(tk_xtime(tk), *ts);
1380 if (timespec64_compare(&tk->wall_to_monotonic, ts) > 0 ||
1381 !timespec64_valid_settod(&tmp)) {
1386 tk_xtime_add(tk, ts);
1387 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *ts));
1389 error: /* even if we error out, we forwarded the time, so call update */
1390 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1392 write_seqcount_end(&tk_core.seq);
1393 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1395 /* Signal hrtimers about time change */
1396 clock_was_set(CLOCK_SET_WALL);
1402 * Indicates if there is an offset between the system clock and the hardware
1403 * clock/persistent clock/rtc.
1405 int persistent_clock_is_local;
1408 * Adjust the time obtained from the CMOS to be UTC time instead of
1411 * This is ugly, but preferable to the alternatives. Otherwise we
1412 * would either need to write a program to do it in /etc/rc (and risk
1413 * confusion if the program gets run more than once; it would also be
1414 * hard to make the program warp the clock precisely n hours) or
1415 * compile in the timezone information into the kernel. Bad, bad....
1419 * The best thing to do is to keep the CMOS clock in universal time (UTC)
1420 * as real UNIX machines always do it. This avoids all headaches about
1421 * daylight saving times and warping kernel clocks.
1423 void timekeeping_warp_clock(void)
1425 if (sys_tz.tz_minuteswest != 0) {
1426 struct timespec64 adjust;
1428 persistent_clock_is_local = 1;
1429 adjust.tv_sec = sys_tz.tz_minuteswest * 60;
1431 timekeeping_inject_offset(&adjust);
1436 * __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic
1438 static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1440 tk->tai_offset = tai_offset;
1441 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1445 * change_clocksource - Swaps clocksources if a new one is available
1447 * Accumulates current time interval and initializes new clocksource
1449 static int change_clocksource(void *data)
1451 struct timekeeper *tk = &tk_core.timekeeper;
1452 struct clocksource *new, *old = NULL;
1453 unsigned long flags;
1454 bool change = false;
1456 new = (struct clocksource *) data;
1459 * If the cs is in module, get a module reference. Succeeds
1460 * for built-in code (owner == NULL) as well.
1462 if (try_module_get(new->owner)) {
1463 if (!new->enable || new->enable(new) == 0)
1466 module_put(new->owner);
1469 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1470 write_seqcount_begin(&tk_core.seq);
1472 timekeeping_forward_now(tk);
1475 old = tk->tkr_mono.clock;
1476 tk_setup_internals(tk, new);
1479 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1481 write_seqcount_end(&tk_core.seq);
1482 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1488 module_put(old->owner);
1495 * timekeeping_notify - Install a new clock source
1496 * @clock: pointer to the clock source
1498 * This function is called from clocksource.c after a new, better clock
1499 * source has been registered. The caller holds the clocksource_mutex.
1501 int timekeeping_notify(struct clocksource *clock)
1503 struct timekeeper *tk = &tk_core.timekeeper;
1505 if (tk->tkr_mono.clock == clock)
1507 stop_machine(change_clocksource, clock, NULL);
1508 tick_clock_notify();
1509 return tk->tkr_mono.clock == clock ? 0 : -1;
1513 * ktime_get_raw_ts64 - Returns the raw monotonic time in a timespec
1514 * @ts: pointer to the timespec64 to be set
1516 * Returns the raw monotonic time (completely un-modified by ntp)
1518 void ktime_get_raw_ts64(struct timespec64 *ts)
1520 struct timekeeper *tk = &tk_core.timekeeper;
1525 seq = read_seqcount_begin(&tk_core.seq);
1526 ts->tv_sec = tk->raw_sec;
1527 nsecs = timekeeping_get_ns(&tk->tkr_raw);
1529 } while (read_seqcount_retry(&tk_core.seq, seq));
1532 timespec64_add_ns(ts, nsecs);
1534 EXPORT_SYMBOL(ktime_get_raw_ts64);
1538 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1540 int timekeeping_valid_for_hres(void)
1542 struct timekeeper *tk = &tk_core.timekeeper;
1547 seq = read_seqcount_begin(&tk_core.seq);
1549 ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1551 } while (read_seqcount_retry(&tk_core.seq, seq));
1557 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1559 u64 timekeeping_max_deferment(void)
1561 struct timekeeper *tk = &tk_core.timekeeper;
1566 seq = read_seqcount_begin(&tk_core.seq);
1568 ret = tk->tkr_mono.clock->max_idle_ns;
1570 } while (read_seqcount_retry(&tk_core.seq, seq));
1576 * read_persistent_clock64 - Return time from the persistent clock.
1577 * @ts: Pointer to the storage for the readout value
1579 * Weak dummy function for arches that do not yet support it.
1580 * Reads the time from the battery backed persistent clock.
1581 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1583 * XXX - Do be sure to remove it once all arches implement it.
1585 void __weak read_persistent_clock64(struct timespec64 *ts)
1592 * read_persistent_wall_and_boot_offset - Read persistent clock, and also offset
1594 * @wall_time: current time as returned by persistent clock
1595 * @boot_offset: offset that is defined as wall_time - boot_time
1597 * Weak dummy function for arches that do not yet support it.
1599 * The default function calculates offset based on the current value of
1600 * local_clock(). This way architectures that support sched_clock() but don't
1601 * support dedicated boot time clock will provide the best estimate of the
1605 read_persistent_wall_and_boot_offset(struct timespec64 *wall_time,
1606 struct timespec64 *boot_offset)
1608 read_persistent_clock64(wall_time);
1609 *boot_offset = ns_to_timespec64(local_clock());
1613 * Flag reflecting whether timekeeping_resume() has injected sleeptime.
1615 * The flag starts of false and is only set when a suspend reaches
1616 * timekeeping_suspend(), timekeeping_resume() sets it to false when the
1617 * timekeeper clocksource is not stopping across suspend and has been
1618 * used to update sleep time. If the timekeeper clocksource has stopped
1619 * then the flag stays true and is used by the RTC resume code to decide
1620 * whether sleeptime must be injected and if so the flag gets false then.
1622 * If a suspend fails before reaching timekeeping_resume() then the flag
1623 * stays false and prevents erroneous sleeptime injection.
1625 static bool suspend_timing_needed;
1627 /* Flag for if there is a persistent clock on this platform */
1628 static bool persistent_clock_exists;
1631 * timekeeping_init - Initializes the clocksource and common timekeeping values
1633 void __init timekeeping_init(void)
1635 struct timespec64 wall_time, boot_offset, wall_to_mono;
1636 struct timekeeper *tk = &tk_core.timekeeper;
1637 struct clocksource *clock;
1638 unsigned long flags;
1640 read_persistent_wall_and_boot_offset(&wall_time, &boot_offset);
1641 if (timespec64_valid_settod(&wall_time) &&
1642 timespec64_to_ns(&wall_time) > 0) {
1643 persistent_clock_exists = true;
1644 } else if (timespec64_to_ns(&wall_time) != 0) {
1645 pr_warn("Persistent clock returned invalid value");
1646 wall_time = (struct timespec64){0};
1649 if (timespec64_compare(&wall_time, &boot_offset) < 0)
1650 boot_offset = (struct timespec64){0};
1653 * We want set wall_to_mono, so the following is true:
1654 * wall time + wall_to_mono = boot time
1656 wall_to_mono = timespec64_sub(boot_offset, wall_time);
1658 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1659 write_seqcount_begin(&tk_core.seq);
1662 clock = clocksource_default_clock();
1664 clock->enable(clock);
1665 tk_setup_internals(tk, clock);
1667 tk_set_xtime(tk, &wall_time);
1670 tk_set_wall_to_mono(tk, wall_to_mono);
1672 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1674 write_seqcount_end(&tk_core.seq);
1675 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1678 /* time in seconds when suspend began for persistent clock */
1679 static struct timespec64 timekeeping_suspend_time;
1682 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1683 * @tk: Pointer to the timekeeper to be updated
1684 * @delta: Pointer to the delta value in timespec64 format
1686 * Takes a timespec offset measuring a suspend interval and properly
1687 * adds the sleep offset to the timekeeping variables.
1689 static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1690 const struct timespec64 *delta)
1692 if (!timespec64_valid_strict(delta)) {
1693 printk_deferred(KERN_WARNING
1694 "__timekeeping_inject_sleeptime: Invalid "
1695 "sleep delta value!\n");
1698 tk_xtime_add(tk, delta);
1699 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1700 tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1701 tk_debug_account_sleep_time(delta);
1704 #if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1706 * We have three kinds of time sources to use for sleep time
1707 * injection, the preference order is:
1708 * 1) non-stop clocksource
1709 * 2) persistent clock (ie: RTC accessible when irqs are off)
1712 * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1713 * If system has neither 1) nor 2), 3) will be used finally.
1716 * If timekeeping has injected sleeptime via either 1) or 2),
1717 * 3) becomes needless, so in this case we don't need to call
1718 * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1721 bool timekeeping_rtc_skipresume(void)
1723 return !suspend_timing_needed;
1727 * 1) can be determined whether to use or not only when doing
1728 * timekeeping_resume() which is invoked after rtc_suspend(),
1729 * so we can't skip rtc_suspend() surely if system has 1).
1731 * But if system has 2), 2) will definitely be used, so in this
1732 * case we don't need to call rtc_suspend(), and this is what
1733 * timekeeping_rtc_skipsuspend() means.
1735 bool timekeeping_rtc_skipsuspend(void)
1737 return persistent_clock_exists;
1741 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1742 * @delta: pointer to a timespec64 delta value
1744 * This hook is for architectures that cannot support read_persistent_clock64
1745 * because their RTC/persistent clock is only accessible when irqs are enabled.
1746 * and also don't have an effective nonstop clocksource.
1748 * This function should only be called by rtc_resume(), and allows
1749 * a suspend offset to be injected into the timekeeping values.
1751 void timekeeping_inject_sleeptime64(const struct timespec64 *delta)
1753 struct timekeeper *tk = &tk_core.timekeeper;
1754 unsigned long flags;
1756 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1757 write_seqcount_begin(&tk_core.seq);
1759 suspend_timing_needed = false;
1761 timekeeping_forward_now(tk);
1763 __timekeeping_inject_sleeptime(tk, delta);
1765 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1767 write_seqcount_end(&tk_core.seq);
1768 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1770 /* Signal hrtimers about time change */
1771 clock_was_set(CLOCK_SET_WALL | CLOCK_SET_BOOT);
1776 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1778 void timekeeping_resume(void)
1780 struct timekeeper *tk = &tk_core.timekeeper;
1781 struct clocksource *clock = tk->tkr_mono.clock;
1782 unsigned long flags;
1783 struct timespec64 ts_new, ts_delta;
1784 u64 cycle_now, nsec;
1785 bool inject_sleeptime = false;
1787 read_persistent_clock64(&ts_new);
1789 clockevents_resume();
1790 clocksource_resume();
1792 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1793 write_seqcount_begin(&tk_core.seq);
1796 * After system resumes, we need to calculate the suspended time and
1797 * compensate it for the OS time. There are 3 sources that could be
1798 * used: Nonstop clocksource during suspend, persistent clock and rtc
1801 * One specific platform may have 1 or 2 or all of them, and the
1802 * preference will be:
1803 * suspend-nonstop clocksource -> persistent clock -> rtc
1804 * The less preferred source will only be tried if there is no better
1805 * usable source. The rtc part is handled separately in rtc core code.
1807 cycle_now = tk_clock_read(&tk->tkr_mono);
1808 nsec = clocksource_stop_suspend_timing(clock, cycle_now);
1810 ts_delta = ns_to_timespec64(nsec);
1811 inject_sleeptime = true;
1812 } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1813 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1814 inject_sleeptime = true;
1817 if (inject_sleeptime) {
1818 suspend_timing_needed = false;
1819 __timekeeping_inject_sleeptime(tk, &ts_delta);
1822 /* Re-base the last cycle value */
1823 tk->tkr_mono.cycle_last = cycle_now;
1824 tk->tkr_raw.cycle_last = cycle_now;
1827 timekeeping_suspended = 0;
1828 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1829 write_seqcount_end(&tk_core.seq);
1830 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1832 touch_softlockup_watchdog();
1834 /* Resume the clockevent device(s) and hrtimers */
1836 /* Notify timerfd as resume is equivalent to clock_was_set() */
1840 int timekeeping_suspend(void)
1842 struct timekeeper *tk = &tk_core.timekeeper;
1843 unsigned long flags;
1844 struct timespec64 delta, delta_delta;
1845 static struct timespec64 old_delta;
1846 struct clocksource *curr_clock;
1849 read_persistent_clock64(&timekeeping_suspend_time);
1852 * On some systems the persistent_clock can not be detected at
1853 * timekeeping_init by its return value, so if we see a valid
1854 * value returned, update the persistent_clock_exists flag.
1856 if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1857 persistent_clock_exists = true;
1859 suspend_timing_needed = true;
1861 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1862 write_seqcount_begin(&tk_core.seq);
1863 timekeeping_forward_now(tk);
1864 timekeeping_suspended = 1;
1867 * Since we've called forward_now, cycle_last stores the value
1868 * just read from the current clocksource. Save this to potentially
1869 * use in suspend timing.
1871 curr_clock = tk->tkr_mono.clock;
1872 cycle_now = tk->tkr_mono.cycle_last;
1873 clocksource_start_suspend_timing(curr_clock, cycle_now);
1875 if (persistent_clock_exists) {
1877 * To avoid drift caused by repeated suspend/resumes,
1878 * which each can add ~1 second drift error,
1879 * try to compensate so the difference in system time
1880 * and persistent_clock time stays close to constant.
1882 delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1883 delta_delta = timespec64_sub(delta, old_delta);
1884 if (abs(delta_delta.tv_sec) >= 2) {
1886 * if delta_delta is too large, assume time correction
1887 * has occurred and set old_delta to the current delta.
1891 /* Otherwise try to adjust old_system to compensate */
1892 timekeeping_suspend_time =
1893 timespec64_add(timekeeping_suspend_time, delta_delta);
1897 timekeeping_update(tk, TK_MIRROR);
1898 halt_fast_timekeeper(tk);
1899 write_seqcount_end(&tk_core.seq);
1900 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1903 clocksource_suspend();
1904 clockevents_suspend();
1909 /* sysfs resume/suspend bits for timekeeping */
1910 static struct syscore_ops timekeeping_syscore_ops = {
1911 .resume = timekeeping_resume,
1912 .suspend = timekeeping_suspend,
1915 static int __init timekeeping_init_ops(void)
1917 register_syscore_ops(&timekeeping_syscore_ops);
1920 device_initcall(timekeeping_init_ops);
1923 * Apply a multiplier adjustment to the timekeeper
1925 static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1929 s64 interval = tk->cycle_interval;
1931 if (mult_adj == 0) {
1933 } else if (mult_adj == -1) {
1934 interval = -interval;
1936 } else if (mult_adj != 1) {
1937 interval *= mult_adj;
1942 * So the following can be confusing.
1944 * To keep things simple, lets assume mult_adj == 1 for now.
1946 * When mult_adj != 1, remember that the interval and offset values
1947 * have been appropriately scaled so the math is the same.
1949 * The basic idea here is that we're increasing the multiplier
1950 * by one, this causes the xtime_interval to be incremented by
1951 * one cycle_interval. This is because:
1952 * xtime_interval = cycle_interval * mult
1953 * So if mult is being incremented by one:
1954 * xtime_interval = cycle_interval * (mult + 1)
1956 * xtime_interval = (cycle_interval * mult) + cycle_interval
1957 * Which can be shortened to:
1958 * xtime_interval += cycle_interval
1960 * So offset stores the non-accumulated cycles. Thus the current
1961 * time (in shifted nanoseconds) is:
1962 * now = (offset * adj) + xtime_nsec
1963 * Now, even though we're adjusting the clock frequency, we have
1964 * to keep time consistent. In other words, we can't jump back
1965 * in time, and we also want to avoid jumping forward in time.
1967 * So given the same offset value, we need the time to be the same
1968 * both before and after the freq adjustment.
1969 * now = (offset * adj_1) + xtime_nsec_1
1970 * now = (offset * adj_2) + xtime_nsec_2
1972 * (offset * adj_1) + xtime_nsec_1 =
1973 * (offset * adj_2) + xtime_nsec_2
1977 * (offset * adj_1) + xtime_nsec_1 =
1978 * (offset * (adj_1+1)) + xtime_nsec_2
1979 * (offset * adj_1) + xtime_nsec_1 =
1980 * (offset * adj_1) + offset + xtime_nsec_2
1981 * Canceling the sides:
1982 * xtime_nsec_1 = offset + xtime_nsec_2
1984 * xtime_nsec_2 = xtime_nsec_1 - offset
1985 * Which simplifies to:
1986 * xtime_nsec -= offset
1988 if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1989 /* NTP adjustment caused clocksource mult overflow */
1994 tk->tkr_mono.mult += mult_adj;
1995 tk->xtime_interval += interval;
1996 tk->tkr_mono.xtime_nsec -= offset;
2000 * Adjust the timekeeper's multiplier to the correct frequency
2001 * and also to reduce the accumulated error value.
2003 static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
2008 * Determine the multiplier from the current NTP tick length.
2009 * Avoid expensive division when the tick length doesn't change.
2011 if (likely(tk->ntp_tick == ntp_tick_length())) {
2012 mult = tk->tkr_mono.mult - tk->ntp_err_mult;
2014 tk->ntp_tick = ntp_tick_length();
2015 mult = div64_u64((tk->ntp_tick >> tk->ntp_error_shift) -
2016 tk->xtime_remainder, tk->cycle_interval);
2020 * If the clock is behind the NTP time, increase the multiplier by 1
2021 * to catch up with it. If it's ahead and there was a remainder in the
2022 * tick division, the clock will slow down. Otherwise it will stay
2023 * ahead until the tick length changes to a non-divisible value.
2025 tk->ntp_err_mult = tk->ntp_error > 0 ? 1 : 0;
2026 mult += tk->ntp_err_mult;
2028 timekeeping_apply_adjustment(tk, offset, mult - tk->tkr_mono.mult);
2030 if (unlikely(tk->tkr_mono.clock->maxadj &&
2031 (abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
2032 > tk->tkr_mono.clock->maxadj))) {
2033 printk_once(KERN_WARNING
2034 "Adjusting %s more than 11%% (%ld vs %ld)\n",
2035 tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
2036 (long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
2040 * It may be possible that when we entered this function, xtime_nsec
2041 * was very small. Further, if we're slightly speeding the clocksource
2042 * in the code above, its possible the required corrective factor to
2043 * xtime_nsec could cause it to underflow.
2045 * Now, since we have already accumulated the second and the NTP
2046 * subsystem has been notified via second_overflow(), we need to skip
2049 if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
2050 tk->tkr_mono.xtime_nsec += (u64)NSEC_PER_SEC <<
2053 tk->skip_second_overflow = 1;
2058 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
2060 * Helper function that accumulates the nsecs greater than a second
2061 * from the xtime_nsec field to the xtime_secs field.
2062 * It also calls into the NTP code to handle leapsecond processing.
2064 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
2066 u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
2067 unsigned int clock_set = 0;
2069 while (tk->tkr_mono.xtime_nsec >= nsecps) {
2072 tk->tkr_mono.xtime_nsec -= nsecps;
2076 * Skip NTP update if this second was accumulated before,
2077 * i.e. xtime_nsec underflowed in timekeeping_adjust()
2079 if (unlikely(tk->skip_second_overflow)) {
2080 tk->skip_second_overflow = 0;
2084 /* Figure out if its a leap sec and apply if needed */
2085 leap = second_overflow(tk->xtime_sec);
2086 if (unlikely(leap)) {
2087 struct timespec64 ts;
2089 tk->xtime_sec += leap;
2093 tk_set_wall_to_mono(tk,
2094 timespec64_sub(tk->wall_to_monotonic, ts));
2096 __timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
2098 clock_set = TK_CLOCK_WAS_SET;
2105 * logarithmic_accumulation - shifted accumulation of cycles
2107 * This functions accumulates a shifted interval of cycles into
2108 * a shifted interval nanoseconds. Allows for O(log) accumulation
2111 * Returns the unconsumed cycles.
2113 static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset,
2114 u32 shift, unsigned int *clock_set)
2116 u64 interval = tk->cycle_interval << shift;
2119 /* If the offset is smaller than a shifted interval, do nothing */
2120 if (offset < interval)
2123 /* Accumulate one shifted interval */
2125 tk->tkr_mono.cycle_last += interval;
2126 tk->tkr_raw.cycle_last += interval;
2128 tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
2129 *clock_set |= accumulate_nsecs_to_secs(tk);
2131 /* Accumulate raw time */
2132 tk->tkr_raw.xtime_nsec += tk->raw_interval << shift;
2133 snsec_per_sec = (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
2134 while (tk->tkr_raw.xtime_nsec >= snsec_per_sec) {
2135 tk->tkr_raw.xtime_nsec -= snsec_per_sec;
2139 /* Accumulate error between NTP and clock interval */
2140 tk->ntp_error += tk->ntp_tick << shift;
2141 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
2142 (tk->ntp_error_shift + shift);
2148 * timekeeping_advance - Updates the timekeeper to the current time and
2149 * current NTP tick length
2151 static bool timekeeping_advance(enum timekeeping_adv_mode mode)
2153 struct timekeeper *real_tk = &tk_core.timekeeper;
2154 struct timekeeper *tk = &shadow_timekeeper;
2156 int shift = 0, maxshift;
2157 unsigned int clock_set = 0;
2158 unsigned long flags;
2160 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2162 /* Make sure we're fully resumed: */
2163 if (unlikely(timekeeping_suspended))
2166 offset = clocksource_delta(tk_clock_read(&tk->tkr_mono),
2167 tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
2169 /* Check if there's really nothing to do */
2170 if (offset < real_tk->cycle_interval && mode == TK_ADV_TICK)
2173 /* Do some additional sanity checking */
2174 timekeeping_check_update(tk, offset);
2177 * With NO_HZ we may have to accumulate many cycle_intervals
2178 * (think "ticks") worth of time at once. To do this efficiently,
2179 * we calculate the largest doubling multiple of cycle_intervals
2180 * that is smaller than the offset. We then accumulate that
2181 * chunk in one go, and then try to consume the next smaller
2184 shift = ilog2(offset) - ilog2(tk->cycle_interval);
2185 shift = max(0, shift);
2186 /* Bound shift to one less than what overflows tick_length */
2187 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
2188 shift = min(shift, maxshift);
2189 while (offset >= tk->cycle_interval) {
2190 offset = logarithmic_accumulation(tk, offset, shift,
2192 if (offset < tk->cycle_interval<<shift)
2196 /* Adjust the multiplier to correct NTP error */
2197 timekeeping_adjust(tk, offset);
2200 * Finally, make sure that after the rounding
2201 * xtime_nsec isn't larger than NSEC_PER_SEC
2203 clock_set |= accumulate_nsecs_to_secs(tk);
2205 write_seqcount_begin(&tk_core.seq);
2207 * Update the real timekeeper.
2209 * We could avoid this memcpy by switching pointers, but that
2210 * requires changes to all other timekeeper usage sites as
2211 * well, i.e. move the timekeeper pointer getter into the
2212 * spinlocked/seqcount protected sections. And we trade this
2213 * memcpy under the tk_core.seq against one before we start
2216 timekeeping_update(tk, clock_set);
2217 memcpy(real_tk, tk, sizeof(*tk));
2218 /* The memcpy must come last. Do not put anything here! */
2219 write_seqcount_end(&tk_core.seq);
2221 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2227 * update_wall_time - Uses the current clocksource to increment the wall time
2230 void update_wall_time(void)
2232 if (timekeeping_advance(TK_ADV_TICK))
2233 clock_was_set_delayed();
2237 * getboottime64 - Return the real time of system boot.
2238 * @ts: pointer to the timespec64 to be set
2240 * Returns the wall-time of boot in a timespec64.
2242 * This is based on the wall_to_monotonic offset and the total suspend
2243 * time. Calls to settimeofday will affect the value returned (which
2244 * basically means that however wrong your real time clock is at boot time,
2245 * you get the right time here).
2247 void getboottime64(struct timespec64 *ts)
2249 struct timekeeper *tk = &tk_core.timekeeper;
2250 ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
2252 *ts = ktime_to_timespec64(t);
2254 EXPORT_SYMBOL_GPL(getboottime64);
2256 void ktime_get_coarse_real_ts64(struct timespec64 *ts)
2258 struct timekeeper *tk = &tk_core.timekeeper;
2262 seq = read_seqcount_begin(&tk_core.seq);
2265 } while (read_seqcount_retry(&tk_core.seq, seq));
2267 EXPORT_SYMBOL(ktime_get_coarse_real_ts64);
2269 void ktime_get_coarse_ts64(struct timespec64 *ts)
2271 struct timekeeper *tk = &tk_core.timekeeper;
2272 struct timespec64 now, mono;
2276 seq = read_seqcount_begin(&tk_core.seq);
2279 mono = tk->wall_to_monotonic;
2280 } while (read_seqcount_retry(&tk_core.seq, seq));
2282 set_normalized_timespec64(ts, now.tv_sec + mono.tv_sec,
2283 now.tv_nsec + mono.tv_nsec);
2285 EXPORT_SYMBOL(ktime_get_coarse_ts64);
2288 * Must hold jiffies_lock
2290 void do_timer(unsigned long ticks)
2292 jiffies_64 += ticks;
2297 * ktime_get_update_offsets_now - hrtimer helper
2298 * @cwsseq: pointer to check and store the clock was set sequence number
2299 * @offs_real: pointer to storage for monotonic -> realtime offset
2300 * @offs_boot: pointer to storage for monotonic -> boottime offset
2301 * @offs_tai: pointer to storage for monotonic -> clock tai offset
2303 * Returns current monotonic time and updates the offsets if the
2304 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
2307 * Called from hrtimer_interrupt() or retrigger_next_event()
2309 ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
2310 ktime_t *offs_boot, ktime_t *offs_tai)
2312 struct timekeeper *tk = &tk_core.timekeeper;
2318 seq = read_seqcount_begin(&tk_core.seq);
2320 base = tk->tkr_mono.base;
2321 nsecs = timekeeping_get_ns(&tk->tkr_mono);
2322 base = ktime_add_ns(base, nsecs);
2324 if (*cwsseq != tk->clock_was_set_seq) {
2325 *cwsseq = tk->clock_was_set_seq;
2326 *offs_real = tk->offs_real;
2327 *offs_boot = tk->offs_boot;
2328 *offs_tai = tk->offs_tai;
2331 /* Handle leapsecond insertion adjustments */
2332 if (unlikely(base >= tk->next_leap_ktime))
2333 *offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
2335 } while (read_seqcount_retry(&tk_core.seq, seq));
2341 * timekeeping_validate_timex - Ensures the timex is ok for use in do_adjtimex
2343 static int timekeeping_validate_timex(const struct __kernel_timex *txc)
2345 if (txc->modes & ADJ_ADJTIME) {
2346 /* singleshot must not be used with any other mode bits */
2347 if (!(txc->modes & ADJ_OFFSET_SINGLESHOT))
2349 if (!(txc->modes & ADJ_OFFSET_READONLY) &&
2350 !capable(CAP_SYS_TIME))
2353 /* In order to modify anything, you gotta be super-user! */
2354 if (txc->modes && !capable(CAP_SYS_TIME))
2357 * if the quartz is off by more than 10% then
2358 * something is VERY wrong!
2360 if (txc->modes & ADJ_TICK &&
2361 (txc->tick < 900000/USER_HZ ||
2362 txc->tick > 1100000/USER_HZ))
2366 if (txc->modes & ADJ_SETOFFSET) {
2367 /* In order to inject time, you gotta be super-user! */
2368 if (!capable(CAP_SYS_TIME))
2372 * Validate if a timespec/timeval used to inject a time
2373 * offset is valid. Offsets can be positive or negative, so
2374 * we don't check tv_sec. The value of the timeval/timespec
2375 * is the sum of its fields,but *NOTE*:
2376 * The field tv_usec/tv_nsec must always be non-negative and
2377 * we can't have more nanoseconds/microseconds than a second.
2379 if (txc->time.tv_usec < 0)
2382 if (txc->modes & ADJ_NANO) {
2383 if (txc->time.tv_usec >= NSEC_PER_SEC)
2386 if (txc->time.tv_usec >= USEC_PER_SEC)
2392 * Check for potential multiplication overflows that can
2393 * only happen on 64-bit systems:
2395 if ((txc->modes & ADJ_FREQUENCY) && (BITS_PER_LONG == 64)) {
2396 if (LLONG_MIN / PPM_SCALE > txc->freq)
2398 if (LLONG_MAX / PPM_SCALE < txc->freq)
2406 * random_get_entropy_fallback - Returns the raw clock source value,
2407 * used by random.c for platforms with no valid random_get_entropy().
2409 unsigned long random_get_entropy_fallback(void)
2411 struct tk_read_base *tkr = &tk_core.timekeeper.tkr_mono;
2412 struct clocksource *clock = READ_ONCE(tkr->clock);
2414 if (unlikely(timekeeping_suspended || !clock))
2416 return clock->read(clock);
2418 EXPORT_SYMBOL_GPL(random_get_entropy_fallback);
2421 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2423 int do_adjtimex(struct __kernel_timex *txc)
2425 struct timekeeper *tk = &tk_core.timekeeper;
2426 struct audit_ntp_data ad;
2427 bool clock_set = false;
2428 struct timespec64 ts;
2429 unsigned long flags;
2433 /* Validate the data before disabling interrupts */
2434 ret = timekeeping_validate_timex(txc);
2437 add_device_randomness(txc, sizeof(*txc));
2439 if (txc->modes & ADJ_SETOFFSET) {
2440 struct timespec64 delta;
2441 delta.tv_sec = txc->time.tv_sec;
2442 delta.tv_nsec = txc->time.tv_usec;
2443 if (!(txc->modes & ADJ_NANO))
2444 delta.tv_nsec *= 1000;
2445 ret = timekeeping_inject_offset(&delta);
2449 audit_tk_injoffset(delta);
2452 audit_ntp_init(&ad);
2454 ktime_get_real_ts64(&ts);
2455 add_device_randomness(&ts, sizeof(ts));
2457 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2458 write_seqcount_begin(&tk_core.seq);
2460 orig_tai = tai = tk->tai_offset;
2461 ret = __do_adjtimex(txc, &ts, &tai, &ad);
2463 if (tai != orig_tai) {
2464 __timekeeping_set_tai_offset(tk, tai);
2465 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
2468 tk_update_leap_state(tk);
2470 write_seqcount_end(&tk_core.seq);
2471 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2475 /* Update the multiplier immediately if frequency was set directly */
2476 if (txc->modes & (ADJ_FREQUENCY | ADJ_TICK))
2477 clock_set |= timekeeping_advance(TK_ADV_FREQ);
2480 clock_was_set(CLOCK_REALTIME);
2482 ntp_notify_cmos_timer();
2487 #ifdef CONFIG_NTP_PPS
2489 * hardpps() - Accessor function to NTP __hardpps function
2491 void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2493 unsigned long flags;
2495 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2496 write_seqcount_begin(&tk_core.seq);
2498 __hardpps(phase_ts, raw_ts);
2500 write_seqcount_end(&tk_core.seq);
2501 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2503 EXPORT_SYMBOL(hardpps);
2504 #endif /* CONFIG_NTP_PPS */