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 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
375 static u32 default_arch_gettimeoffset(void) { return 0; }
376 u32 (*arch_gettimeoffset)(void) = default_arch_gettimeoffset;
378 static inline u32 arch_gettimeoffset(void) { return 0; }
381 static inline u64 timekeeping_delta_to_ns(const struct tk_read_base *tkr, u64 delta)
385 nsec = delta * tkr->mult + tkr->xtime_nsec;
388 /* If arch requires, add in get_arch_timeoffset() */
389 return nsec + arch_gettimeoffset();
392 static inline u64 timekeeping_get_ns(const struct tk_read_base *tkr)
396 delta = timekeeping_get_delta(tkr);
397 return timekeeping_delta_to_ns(tkr, delta);
400 static inline u64 timekeeping_cycles_to_ns(const struct tk_read_base *tkr, u64 cycles)
404 /* calculate the delta since the last update_wall_time */
405 delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask);
406 return timekeeping_delta_to_ns(tkr, delta);
410 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
411 * @tkr: Timekeeping readout base from which we take the update
413 * We want to use this from any context including NMI and tracing /
414 * instrumenting the timekeeping code itself.
416 * Employ the latch technique; see @raw_write_seqcount_latch.
418 * So if a NMI hits the update of base[0] then it will use base[1]
419 * which is still consistent. In the worst case this can result is a
420 * slightly wrong timestamp (a few nanoseconds). See
421 * @ktime_get_mono_fast_ns.
423 static void update_fast_timekeeper(const struct tk_read_base *tkr,
426 struct tk_read_base *base = tkf->base;
428 /* Force readers off to base[1] */
429 raw_write_seqcount_latch(&tkf->seq);
432 memcpy(base, tkr, sizeof(*base));
434 /* Force readers back to base[0] */
435 raw_write_seqcount_latch(&tkf->seq);
438 memcpy(base + 1, base, sizeof(*base));
442 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
444 * This timestamp is not guaranteed to be monotonic across an update.
445 * The timestamp is calculated by:
447 * now = base_mono + clock_delta * slope
449 * So if the update lowers the slope, readers who are forced to the
450 * not yet updated second array are still using the old steeper slope.
459 * |12345678---> reader order
465 * So reader 6 will observe time going backwards versus reader 5.
467 * While other CPUs are likely to be able observe that, the only way
468 * for a CPU local observation is when an NMI hits in the middle of
469 * the update. Timestamps taken from that NMI context might be ahead
470 * of the following timestamps. Callers need to be aware of that and
473 static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
475 struct tk_read_base *tkr;
480 seq = raw_read_seqcount_latch(&tkf->seq);
481 tkr = tkf->base + (seq & 0x01);
482 now = ktime_to_ns(tkr->base);
484 now += timekeeping_delta_to_ns(tkr,
489 } while (read_seqcount_latch_retry(&tkf->seq, seq));
494 u64 ktime_get_mono_fast_ns(void)
496 return __ktime_get_fast_ns(&tk_fast_mono);
498 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
500 u64 ktime_get_raw_fast_ns(void)
502 return __ktime_get_fast_ns(&tk_fast_raw);
504 EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
507 * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock.
509 * To keep it NMI safe since we're accessing from tracing, we're not using a
510 * separate timekeeper with updates to monotonic clock and boot offset
511 * protected with seqcounts. This has the following minor side effects:
513 * (1) Its possible that a timestamp be taken after the boot offset is updated
514 * but before the timekeeper is updated. If this happens, the new boot offset
515 * is added to the old timekeeping making the clock appear to update slightly
518 * timekeeping_inject_sleeptime64()
519 * __timekeeping_inject_sleeptime(tk, delta);
521 * timekeeping_update(tk, TK_CLEAR_NTP...);
523 * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be
524 * partially updated. Since the tk->offs_boot update is a rare event, this
525 * should be a rare occurrence which postprocessing should be able to handle.
527 u64 notrace ktime_get_boot_fast_ns(void)
529 struct timekeeper *tk = &tk_core.timekeeper;
531 return (ktime_get_mono_fast_ns() + ktime_to_ns(tk->offs_boot));
533 EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns);
536 * See comment for __ktime_get_fast_ns() vs. timestamp ordering
538 static __always_inline u64 __ktime_get_real_fast(struct tk_fast *tkf, u64 *mono)
540 struct tk_read_base *tkr;
541 u64 basem, baser, delta;
545 seq = raw_read_seqcount_latch(&tkf->seq);
546 tkr = tkf->base + (seq & 0x01);
547 basem = ktime_to_ns(tkr->base);
548 baser = ktime_to_ns(tkr->base_real);
550 delta = timekeeping_delta_to_ns(tkr,
551 clocksource_delta(tk_clock_read(tkr),
552 tkr->cycle_last, tkr->mask));
553 } while (read_seqcount_latch_retry(&tkf->seq, seq));
556 *mono = basem + delta;
557 return baser + delta;
561 * ktime_get_real_fast_ns: - NMI safe and fast access to clock realtime.
563 u64 ktime_get_real_fast_ns(void)
565 return __ktime_get_real_fast(&tk_fast_mono, NULL);
567 EXPORT_SYMBOL_GPL(ktime_get_real_fast_ns);
570 * ktime_get_fast_timestamps: - NMI safe timestamps
571 * @snapshot: Pointer to timestamp storage
573 * Stores clock monotonic, boottime and realtime timestamps.
575 * Boot time is a racy access on 32bit systems if the sleep time injection
576 * happens late during resume and not in timekeeping_resume(). That could
577 * be avoided by expanding struct tk_read_base with boot offset for 32bit
578 * and adding more overhead to the update. As this is a hard to observe
579 * once per resume event which can be filtered with reasonable effort using
580 * the accurate mono/real timestamps, it's probably not worth the trouble.
582 * Aside of that it might be possible on 32 and 64 bit to observe the
583 * following when the sleep time injection happens late:
586 * timekeeping_resume()
587 * ktime_get_fast_timestamps()
588 * mono, real = __ktime_get_real_fast()
589 * inject_sleep_time()
591 * boot = mono + bootoffset;
593 * That means that boot time already has the sleep time adjustment, but
594 * real time does not. On the next readout both are in sync again.
596 * Preventing this for 64bit is not really feasible without destroying the
597 * careful cache layout of the timekeeper because the sequence count and
598 * struct tk_read_base would then need two cache lines instead of one.
600 * Access to the time keeper clock source is disabled accross the innermost
601 * steps of suspend/resume. The accessors still work, but the timestamps
602 * are frozen until time keeping is resumed which happens very early.
604 * For regular suspend/resume there is no observable difference vs. sched
605 * clock, but it might affect some of the nasty low level debug printks.
607 * OTOH, access to sched clock is not guaranteed accross suspend/resume on
608 * all systems either so it depends on the hardware in use.
610 * If that turns out to be a real problem then this could be mitigated by
611 * using sched clock in a similar way as during early boot. But it's not as
612 * trivial as on early boot because it needs some careful protection
613 * against the clock monotonic timestamp jumping backwards on resume.
615 void ktime_get_fast_timestamps(struct ktime_timestamps *snapshot)
617 struct timekeeper *tk = &tk_core.timekeeper;
619 snapshot->real = __ktime_get_real_fast(&tk_fast_mono, &snapshot->mono);
620 snapshot->boot = snapshot->mono + ktime_to_ns(data_race(tk->offs_boot));
624 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
625 * @tk: Timekeeper to snapshot.
627 * It generally is unsafe to access the clocksource after timekeeping has been
628 * suspended, so take a snapshot of the readout base of @tk and use it as the
629 * fast timekeeper's readout base while suspended. It will return the same
630 * number of cycles every time until timekeeping is resumed at which time the
631 * proper readout base for the fast timekeeper will be restored automatically.
633 static void halt_fast_timekeeper(const struct timekeeper *tk)
635 static struct tk_read_base tkr_dummy;
636 const struct tk_read_base *tkr = &tk->tkr_mono;
638 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
639 cycles_at_suspend = tk_clock_read(tkr);
640 tkr_dummy.clock = &dummy_clock;
641 tkr_dummy.base_real = tkr->base + tk->offs_real;
642 update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
645 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
646 tkr_dummy.clock = &dummy_clock;
647 update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
650 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
652 static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
654 raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
658 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
660 int pvclock_gtod_register_notifier(struct notifier_block *nb)
662 struct timekeeper *tk = &tk_core.timekeeper;
666 raw_spin_lock_irqsave(&timekeeper_lock, flags);
667 ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
668 update_pvclock_gtod(tk, true);
669 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
673 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
676 * pvclock_gtod_unregister_notifier - unregister a pvclock
677 * timedata update listener
679 int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
684 raw_spin_lock_irqsave(&timekeeper_lock, flags);
685 ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
686 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
690 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
693 * tk_update_leap_state - helper to update the next_leap_ktime
695 static inline void tk_update_leap_state(struct timekeeper *tk)
697 tk->next_leap_ktime = ntp_get_next_leap();
698 if (tk->next_leap_ktime != KTIME_MAX)
699 /* Convert to monotonic time */
700 tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
704 * Update the ktime_t based scalar nsec members of the timekeeper
706 static inline void tk_update_ktime_data(struct timekeeper *tk)
712 * The xtime based monotonic readout is:
713 * nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
714 * The ktime based monotonic readout is:
715 * nsec = base_mono + now();
716 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
718 seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
719 nsec = (u32) tk->wall_to_monotonic.tv_nsec;
720 tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
723 * The sum of the nanoseconds portions of xtime and
724 * wall_to_monotonic can be greater/equal one second. Take
725 * this into account before updating tk->ktime_sec.
727 nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
728 if (nsec >= NSEC_PER_SEC)
730 tk->ktime_sec = seconds;
732 /* Update the monotonic raw base */
733 tk->tkr_raw.base = ns_to_ktime(tk->raw_sec * NSEC_PER_SEC);
736 /* must hold timekeeper_lock */
737 static void timekeeping_update(struct timekeeper *tk, unsigned int action)
739 if (action & TK_CLEAR_NTP) {
744 tk_update_leap_state(tk);
745 tk_update_ktime_data(tk);
748 update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
750 tk->tkr_mono.base_real = tk->tkr_mono.base + tk->offs_real;
751 update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
752 update_fast_timekeeper(&tk->tkr_raw, &tk_fast_raw);
754 if (action & TK_CLOCK_WAS_SET)
755 tk->clock_was_set_seq++;
757 * The mirroring of the data to the shadow-timekeeper needs
758 * to happen last here to ensure we don't over-write the
759 * timekeeper structure on the next update with stale data
761 if (action & TK_MIRROR)
762 memcpy(&shadow_timekeeper, &tk_core.timekeeper,
763 sizeof(tk_core.timekeeper));
767 * timekeeping_forward_now - update clock to the current time
769 * Forward the current clock to update its state since the last call to
770 * update_wall_time(). This is useful before significant clock changes,
771 * as it avoids having to deal with this time offset explicitly.
773 static void timekeeping_forward_now(struct timekeeper *tk)
775 u64 cycle_now, delta;
777 cycle_now = tk_clock_read(&tk->tkr_mono);
778 delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
779 tk->tkr_mono.cycle_last = cycle_now;
780 tk->tkr_raw.cycle_last = cycle_now;
782 tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult;
784 /* If arch requires, add in get_arch_timeoffset() */
785 tk->tkr_mono.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_mono.shift;
788 tk->tkr_raw.xtime_nsec += delta * tk->tkr_raw.mult;
790 /* If arch requires, add in get_arch_timeoffset() */
791 tk->tkr_raw.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_raw.shift;
793 tk_normalize_xtime(tk);
797 * ktime_get_real_ts64 - Returns the time of day in a timespec64.
798 * @ts: pointer to the timespec to be set
800 * Returns the time of day in a timespec64 (WARN if suspended).
802 void ktime_get_real_ts64(struct timespec64 *ts)
804 struct timekeeper *tk = &tk_core.timekeeper;
808 WARN_ON(timekeeping_suspended);
811 seq = read_seqcount_begin(&tk_core.seq);
813 ts->tv_sec = tk->xtime_sec;
814 nsecs = timekeeping_get_ns(&tk->tkr_mono);
816 } while (read_seqcount_retry(&tk_core.seq, seq));
819 timespec64_add_ns(ts, nsecs);
821 EXPORT_SYMBOL(ktime_get_real_ts64);
823 ktime_t ktime_get(void)
825 struct timekeeper *tk = &tk_core.timekeeper;
830 WARN_ON(timekeeping_suspended);
833 seq = read_seqcount_begin(&tk_core.seq);
834 base = tk->tkr_mono.base;
835 nsecs = timekeeping_get_ns(&tk->tkr_mono);
837 } while (read_seqcount_retry(&tk_core.seq, seq));
839 return ktime_add_ns(base, nsecs);
841 EXPORT_SYMBOL_GPL(ktime_get);
843 u32 ktime_get_resolution_ns(void)
845 struct timekeeper *tk = &tk_core.timekeeper;
849 WARN_ON(timekeeping_suspended);
852 seq = read_seqcount_begin(&tk_core.seq);
853 nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
854 } while (read_seqcount_retry(&tk_core.seq, seq));
858 EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
860 static ktime_t *offsets[TK_OFFS_MAX] = {
861 [TK_OFFS_REAL] = &tk_core.timekeeper.offs_real,
862 [TK_OFFS_BOOT] = &tk_core.timekeeper.offs_boot,
863 [TK_OFFS_TAI] = &tk_core.timekeeper.offs_tai,
866 ktime_t ktime_get_with_offset(enum tk_offsets offs)
868 struct timekeeper *tk = &tk_core.timekeeper;
870 ktime_t base, *offset = offsets[offs];
873 WARN_ON(timekeeping_suspended);
876 seq = read_seqcount_begin(&tk_core.seq);
877 base = ktime_add(tk->tkr_mono.base, *offset);
878 nsecs = timekeeping_get_ns(&tk->tkr_mono);
880 } while (read_seqcount_retry(&tk_core.seq, seq));
882 return ktime_add_ns(base, nsecs);
885 EXPORT_SYMBOL_GPL(ktime_get_with_offset);
887 ktime_t ktime_get_coarse_with_offset(enum tk_offsets offs)
889 struct timekeeper *tk = &tk_core.timekeeper;
891 ktime_t base, *offset = offsets[offs];
894 WARN_ON(timekeeping_suspended);
897 seq = read_seqcount_begin(&tk_core.seq);
898 base = ktime_add(tk->tkr_mono.base, *offset);
899 nsecs = tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift;
901 } while (read_seqcount_retry(&tk_core.seq, seq));
903 return ktime_add_ns(base, nsecs);
905 EXPORT_SYMBOL_GPL(ktime_get_coarse_with_offset);
908 * ktime_mono_to_any() - convert mononotic time to any other time
909 * @tmono: time to convert.
910 * @offs: which offset to use
912 ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
914 ktime_t *offset = offsets[offs];
919 seq = read_seqcount_begin(&tk_core.seq);
920 tconv = ktime_add(tmono, *offset);
921 } while (read_seqcount_retry(&tk_core.seq, seq));
925 EXPORT_SYMBOL_GPL(ktime_mono_to_any);
928 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
930 ktime_t ktime_get_raw(void)
932 struct timekeeper *tk = &tk_core.timekeeper;
938 seq = read_seqcount_begin(&tk_core.seq);
939 base = tk->tkr_raw.base;
940 nsecs = timekeeping_get_ns(&tk->tkr_raw);
942 } while (read_seqcount_retry(&tk_core.seq, seq));
944 return ktime_add_ns(base, nsecs);
946 EXPORT_SYMBOL_GPL(ktime_get_raw);
949 * ktime_get_ts64 - get the monotonic clock in timespec64 format
950 * @ts: pointer to timespec variable
952 * The function calculates the monotonic clock from the realtime
953 * clock and the wall_to_monotonic offset and stores the result
954 * in normalized timespec64 format in the variable pointed to by @ts.
956 void ktime_get_ts64(struct timespec64 *ts)
958 struct timekeeper *tk = &tk_core.timekeeper;
959 struct timespec64 tomono;
963 WARN_ON(timekeeping_suspended);
966 seq = read_seqcount_begin(&tk_core.seq);
967 ts->tv_sec = tk->xtime_sec;
968 nsec = timekeeping_get_ns(&tk->tkr_mono);
969 tomono = tk->wall_to_monotonic;
971 } while (read_seqcount_retry(&tk_core.seq, seq));
973 ts->tv_sec += tomono.tv_sec;
975 timespec64_add_ns(ts, nsec + tomono.tv_nsec);
977 EXPORT_SYMBOL_GPL(ktime_get_ts64);
980 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
982 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
983 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
984 * works on both 32 and 64 bit systems. On 32 bit systems the readout
985 * covers ~136 years of uptime which should be enough to prevent
986 * premature wrap arounds.
988 time64_t ktime_get_seconds(void)
990 struct timekeeper *tk = &tk_core.timekeeper;
992 WARN_ON(timekeeping_suspended);
993 return tk->ktime_sec;
995 EXPORT_SYMBOL_GPL(ktime_get_seconds);
998 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
1000 * Returns the wall clock seconds since 1970. This replaces the
1001 * get_seconds() interface which is not y2038 safe on 32bit systems.
1003 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
1004 * 32bit systems the access must be protected with the sequence
1005 * counter to provide "atomic" access to the 64bit tk->xtime_sec
1008 time64_t ktime_get_real_seconds(void)
1010 struct timekeeper *tk = &tk_core.timekeeper;
1014 if (IS_ENABLED(CONFIG_64BIT))
1015 return tk->xtime_sec;
1018 seq = read_seqcount_begin(&tk_core.seq);
1019 seconds = tk->xtime_sec;
1021 } while (read_seqcount_retry(&tk_core.seq, seq));
1025 EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
1028 * __ktime_get_real_seconds - The same as ktime_get_real_seconds
1029 * but without the sequence counter protect. This internal function
1030 * is called just when timekeeping lock is already held.
1032 noinstr time64_t __ktime_get_real_seconds(void)
1034 struct timekeeper *tk = &tk_core.timekeeper;
1036 return tk->xtime_sec;
1040 * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter
1041 * @systime_snapshot: pointer to struct receiving the system time snapshot
1043 void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot)
1045 struct timekeeper *tk = &tk_core.timekeeper;
1053 WARN_ON_ONCE(timekeeping_suspended);
1056 seq = read_seqcount_begin(&tk_core.seq);
1057 now = tk_clock_read(&tk->tkr_mono);
1058 systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq;
1059 systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq;
1060 base_real = ktime_add(tk->tkr_mono.base,
1061 tk_core.timekeeper.offs_real);
1062 base_raw = tk->tkr_raw.base;
1063 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now);
1064 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw, now);
1065 } while (read_seqcount_retry(&tk_core.seq, seq));
1067 systime_snapshot->cycles = now;
1068 systime_snapshot->real = ktime_add_ns(base_real, nsec_real);
1069 systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw);
1071 EXPORT_SYMBOL_GPL(ktime_get_snapshot);
1073 /* Scale base by mult/div checking for overflow */
1074 static int scale64_check_overflow(u64 mult, u64 div, u64 *base)
1078 tmp = div64_u64_rem(*base, div, &rem);
1080 if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) ||
1081 ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem)))
1085 rem = div64_u64(rem * mult, div);
1091 * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval
1092 * @history: Snapshot representing start of history
1093 * @partial_history_cycles: Cycle offset into history (fractional part)
1094 * @total_history_cycles: Total history length in cycles
1095 * @discontinuity: True indicates clock was set on history period
1096 * @ts: Cross timestamp that should be adjusted using
1097 * partial/total ratio
1099 * Helper function used by get_device_system_crosststamp() to correct the
1100 * crosstimestamp corresponding to the start of the current interval to the
1101 * system counter value (timestamp point) provided by the driver. The
1102 * total_history_* quantities are the total history starting at the provided
1103 * reference point and ending at the start of the current interval. The cycle
1104 * count between the driver timestamp point and the start of the current
1105 * interval is partial_history_cycles.
1107 static int adjust_historical_crosststamp(struct system_time_snapshot *history,
1108 u64 partial_history_cycles,
1109 u64 total_history_cycles,
1111 struct system_device_crosststamp *ts)
1113 struct timekeeper *tk = &tk_core.timekeeper;
1114 u64 corr_raw, corr_real;
1115 bool interp_forward;
1118 if (total_history_cycles == 0 || partial_history_cycles == 0)
1121 /* Interpolate shortest distance from beginning or end of history */
1122 interp_forward = partial_history_cycles > total_history_cycles / 2;
1123 partial_history_cycles = interp_forward ?
1124 total_history_cycles - partial_history_cycles :
1125 partial_history_cycles;
1128 * Scale the monotonic raw time delta by:
1129 * partial_history_cycles / total_history_cycles
1131 corr_raw = (u64)ktime_to_ns(
1132 ktime_sub(ts->sys_monoraw, history->raw));
1133 ret = scale64_check_overflow(partial_history_cycles,
1134 total_history_cycles, &corr_raw);
1139 * If there is a discontinuity in the history, scale monotonic raw
1141 * mult(real)/mult(raw) yielding the realtime correction
1142 * Otherwise, calculate the realtime correction similar to monotonic
1145 if (discontinuity) {
1146 corr_real = mul_u64_u32_div
1147 (corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult);
1149 corr_real = (u64)ktime_to_ns(
1150 ktime_sub(ts->sys_realtime, history->real));
1151 ret = scale64_check_overflow(partial_history_cycles,
1152 total_history_cycles, &corr_real);
1157 /* Fixup monotonic raw and real time time values */
1158 if (interp_forward) {
1159 ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw);
1160 ts->sys_realtime = ktime_add_ns(history->real, corr_real);
1162 ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw);
1163 ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real);
1170 * timestamp_in_interval - true if ts is chronologically in [start, end]
1172 * True if ts occurs chronologically at or after start, and before or at end.
1174 static bool timestamp_in_interval(u64 start, u64 end, u64 ts)
1176 if (ts >= start && ts <= end)
1178 if (start > end && (ts >= start || ts <= end))
1184 * get_device_system_crosststamp - Synchronously capture system/device timestamp
1185 * @get_time_fn: Callback to get simultaneous device time and
1186 * system counter from the device driver
1187 * @ctx: Context passed to get_time_fn()
1188 * @history_begin: Historical reference point used to interpolate system
1189 * time when counter provided by the driver is before the current interval
1190 * @xtstamp: Receives simultaneously captured system and device time
1192 * Reads a timestamp from a device and correlates it to system time
1194 int get_device_system_crosststamp(int (*get_time_fn)
1195 (ktime_t *device_time,
1196 struct system_counterval_t *sys_counterval,
1199 struct system_time_snapshot *history_begin,
1200 struct system_device_crosststamp *xtstamp)
1202 struct system_counterval_t system_counterval;
1203 struct timekeeper *tk = &tk_core.timekeeper;
1204 u64 cycles, now, interval_start;
1205 unsigned int clock_was_set_seq = 0;
1206 ktime_t base_real, base_raw;
1207 u64 nsec_real, nsec_raw;
1208 u8 cs_was_changed_seq;
1214 seq = read_seqcount_begin(&tk_core.seq);
1216 * Try to synchronously capture device time and a system
1217 * counter value calling back into the device driver
1219 ret = get_time_fn(&xtstamp->device, &system_counterval, ctx);
1224 * Verify that the clocksource associated with the captured
1225 * system counter value is the same as the currently installed
1226 * timekeeper clocksource
1228 if (tk->tkr_mono.clock != system_counterval.cs)
1230 cycles = system_counterval.cycles;
1233 * Check whether the system counter value provided by the
1234 * device driver is on the current timekeeping interval.
1236 now = tk_clock_read(&tk->tkr_mono);
1237 interval_start = tk->tkr_mono.cycle_last;
1238 if (!timestamp_in_interval(interval_start, now, cycles)) {
1239 clock_was_set_seq = tk->clock_was_set_seq;
1240 cs_was_changed_seq = tk->cs_was_changed_seq;
1241 cycles = interval_start;
1247 base_real = ktime_add(tk->tkr_mono.base,
1248 tk_core.timekeeper.offs_real);
1249 base_raw = tk->tkr_raw.base;
1251 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, cycles);
1252 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw, cycles);
1253 } while (read_seqcount_retry(&tk_core.seq, seq));
1255 xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real);
1256 xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);
1259 * Interpolate if necessary, adjusting back from the start of the
1263 u64 partial_history_cycles, total_history_cycles;
1267 * Check that the counter value is not before the provided
1268 * history reference and that the history doesn't cross a
1269 * clocksource change
1271 if (!history_begin ||
1272 !timestamp_in_interval(history_begin->cycles,
1273 cycles, system_counterval.cycles) ||
1274 history_begin->cs_was_changed_seq != cs_was_changed_seq)
1276 partial_history_cycles = cycles - system_counterval.cycles;
1277 total_history_cycles = cycles - history_begin->cycles;
1279 history_begin->clock_was_set_seq != clock_was_set_seq;
1281 ret = adjust_historical_crosststamp(history_begin,
1282 partial_history_cycles,
1283 total_history_cycles,
1284 discontinuity, xtstamp);
1291 EXPORT_SYMBOL_GPL(get_device_system_crosststamp);
1294 * do_settimeofday64 - Sets the time of day.
1295 * @ts: pointer to the timespec64 variable containing the new time
1297 * Sets the time of day to the new time and update NTP and notify hrtimers
1299 int do_settimeofday64(const struct timespec64 *ts)
1301 struct timekeeper *tk = &tk_core.timekeeper;
1302 struct timespec64 ts_delta, xt;
1303 unsigned long flags;
1306 if (!timespec64_valid_settod(ts))
1309 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1310 write_seqcount_begin(&tk_core.seq);
1312 timekeeping_forward_now(tk);
1315 ts_delta = timespec64_sub(*ts, xt);
1317 if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) {
1322 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
1324 tk_set_xtime(tk, ts);
1326 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1328 write_seqcount_end(&tk_core.seq);
1329 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1331 /* signal hrtimers about time change */
1335 audit_tk_injoffset(ts_delta);
1336 add_device_randomness(ts, sizeof(*ts));
1341 EXPORT_SYMBOL(do_settimeofday64);
1344 * timekeeping_inject_offset - Adds or subtracts from the current time.
1345 * @tv: pointer to the timespec variable containing the offset
1347 * Adds or subtracts an offset value from the current time.
1349 static int timekeeping_inject_offset(const struct timespec64 *ts)
1351 struct timekeeper *tk = &tk_core.timekeeper;
1352 unsigned long flags;
1353 struct timespec64 tmp;
1356 if (ts->tv_nsec < 0 || ts->tv_nsec >= NSEC_PER_SEC)
1359 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1360 write_seqcount_begin(&tk_core.seq);
1362 timekeeping_forward_now(tk);
1364 /* Make sure the proposed value is valid */
1365 tmp = timespec64_add(tk_xtime(tk), *ts);
1366 if (timespec64_compare(&tk->wall_to_monotonic, ts) > 0 ||
1367 !timespec64_valid_settod(&tmp)) {
1372 tk_xtime_add(tk, ts);
1373 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *ts));
1375 error: /* even if we error out, we forwarded the time, so call update */
1376 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1378 write_seqcount_end(&tk_core.seq);
1379 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1381 /* signal hrtimers about time change */
1388 * Indicates if there is an offset between the system clock and the hardware
1389 * clock/persistent clock/rtc.
1391 int persistent_clock_is_local;
1394 * Adjust the time obtained from the CMOS to be UTC time instead of
1397 * This is ugly, but preferable to the alternatives. Otherwise we
1398 * would either need to write a program to do it in /etc/rc (and risk
1399 * confusion if the program gets run more than once; it would also be
1400 * hard to make the program warp the clock precisely n hours) or
1401 * compile in the timezone information into the kernel. Bad, bad....
1405 * The best thing to do is to keep the CMOS clock in universal time (UTC)
1406 * as real UNIX machines always do it. This avoids all headaches about
1407 * daylight saving times and warping kernel clocks.
1409 void timekeeping_warp_clock(void)
1411 if (sys_tz.tz_minuteswest != 0) {
1412 struct timespec64 adjust;
1414 persistent_clock_is_local = 1;
1415 adjust.tv_sec = sys_tz.tz_minuteswest * 60;
1417 timekeeping_inject_offset(&adjust);
1422 * __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic
1425 static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1427 tk->tai_offset = tai_offset;
1428 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1432 * change_clocksource - Swaps clocksources if a new one is available
1434 * Accumulates current time interval and initializes new clocksource
1436 static int change_clocksource(void *data)
1438 struct timekeeper *tk = &tk_core.timekeeper;
1439 struct clocksource *new, *old;
1440 unsigned long flags;
1442 new = (struct clocksource *) data;
1444 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1445 write_seqcount_begin(&tk_core.seq);
1447 timekeeping_forward_now(tk);
1449 * If the cs is in module, get a module reference. Succeeds
1450 * for built-in code (owner == NULL) as well.
1452 if (try_module_get(new->owner)) {
1453 if (!new->enable || new->enable(new) == 0) {
1454 old = tk->tkr_mono.clock;
1455 tk_setup_internals(tk, new);
1458 module_put(old->owner);
1460 module_put(new->owner);
1463 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1465 write_seqcount_end(&tk_core.seq);
1466 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1472 * timekeeping_notify - Install a new clock source
1473 * @clock: pointer to the clock source
1475 * This function is called from clocksource.c after a new, better clock
1476 * source has been registered. The caller holds the clocksource_mutex.
1478 int timekeeping_notify(struct clocksource *clock)
1480 struct timekeeper *tk = &tk_core.timekeeper;
1482 if (tk->tkr_mono.clock == clock)
1484 stop_machine(change_clocksource, clock, NULL);
1485 tick_clock_notify();
1486 return tk->tkr_mono.clock == clock ? 0 : -1;
1490 * ktime_get_raw_ts64 - Returns the raw monotonic time in a timespec
1491 * @ts: pointer to the timespec64 to be set
1493 * Returns the raw monotonic time (completely un-modified by ntp)
1495 void ktime_get_raw_ts64(struct timespec64 *ts)
1497 struct timekeeper *tk = &tk_core.timekeeper;
1502 seq = read_seqcount_begin(&tk_core.seq);
1503 ts->tv_sec = tk->raw_sec;
1504 nsecs = timekeeping_get_ns(&tk->tkr_raw);
1506 } while (read_seqcount_retry(&tk_core.seq, seq));
1509 timespec64_add_ns(ts, nsecs);
1511 EXPORT_SYMBOL(ktime_get_raw_ts64);
1515 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1517 int timekeeping_valid_for_hres(void)
1519 struct timekeeper *tk = &tk_core.timekeeper;
1524 seq = read_seqcount_begin(&tk_core.seq);
1526 ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1528 } while (read_seqcount_retry(&tk_core.seq, seq));
1534 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1536 u64 timekeeping_max_deferment(void)
1538 struct timekeeper *tk = &tk_core.timekeeper;
1543 seq = read_seqcount_begin(&tk_core.seq);
1545 ret = tk->tkr_mono.clock->max_idle_ns;
1547 } while (read_seqcount_retry(&tk_core.seq, seq));
1553 * read_persistent_clock64 - Return time from the persistent clock.
1555 * Weak dummy function for arches that do not yet support it.
1556 * Reads the time from the battery backed persistent clock.
1557 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1559 * XXX - Do be sure to remove it once all arches implement it.
1561 void __weak read_persistent_clock64(struct timespec64 *ts)
1568 * read_persistent_wall_and_boot_offset - Read persistent clock, and also offset
1571 * Weak dummy function for arches that do not yet support it.
1572 * wall_time - current time as returned by persistent clock
1573 * boot_offset - offset that is defined as wall_time - boot_time
1574 * The default function calculates offset based on the current value of
1575 * local_clock(). This way architectures that support sched_clock() but don't
1576 * support dedicated boot time clock will provide the best estimate of the
1580 read_persistent_wall_and_boot_offset(struct timespec64 *wall_time,
1581 struct timespec64 *boot_offset)
1583 read_persistent_clock64(wall_time);
1584 *boot_offset = ns_to_timespec64(local_clock());
1588 * Flag reflecting whether timekeeping_resume() has injected sleeptime.
1590 * The flag starts of false and is only set when a suspend reaches
1591 * timekeeping_suspend(), timekeeping_resume() sets it to false when the
1592 * timekeeper clocksource is not stopping across suspend and has been
1593 * used to update sleep time. If the timekeeper clocksource has stopped
1594 * then the flag stays true and is used by the RTC resume code to decide
1595 * whether sleeptime must be injected and if so the flag gets false then.
1597 * If a suspend fails before reaching timekeeping_resume() then the flag
1598 * stays false and prevents erroneous sleeptime injection.
1600 static bool suspend_timing_needed;
1602 /* Flag for if there is a persistent clock on this platform */
1603 static bool persistent_clock_exists;
1606 * timekeeping_init - Initializes the clocksource and common timekeeping values
1608 void __init timekeeping_init(void)
1610 struct timespec64 wall_time, boot_offset, wall_to_mono;
1611 struct timekeeper *tk = &tk_core.timekeeper;
1612 struct clocksource *clock;
1613 unsigned long flags;
1615 read_persistent_wall_and_boot_offset(&wall_time, &boot_offset);
1616 if (timespec64_valid_settod(&wall_time) &&
1617 timespec64_to_ns(&wall_time) > 0) {
1618 persistent_clock_exists = true;
1619 } else if (timespec64_to_ns(&wall_time) != 0) {
1620 pr_warn("Persistent clock returned invalid value");
1621 wall_time = (struct timespec64){0};
1624 if (timespec64_compare(&wall_time, &boot_offset) < 0)
1625 boot_offset = (struct timespec64){0};
1628 * We want set wall_to_mono, so the following is true:
1629 * wall time + wall_to_mono = boot time
1631 wall_to_mono = timespec64_sub(boot_offset, wall_time);
1633 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1634 write_seqcount_begin(&tk_core.seq);
1637 clock = clocksource_default_clock();
1639 clock->enable(clock);
1640 tk_setup_internals(tk, clock);
1642 tk_set_xtime(tk, &wall_time);
1645 tk_set_wall_to_mono(tk, wall_to_mono);
1647 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1649 write_seqcount_end(&tk_core.seq);
1650 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1653 /* time in seconds when suspend began for persistent clock */
1654 static struct timespec64 timekeeping_suspend_time;
1657 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1658 * @delta: pointer to a timespec delta value
1660 * Takes a timespec offset measuring a suspend interval and properly
1661 * adds the sleep offset to the timekeeping variables.
1663 static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1664 const struct timespec64 *delta)
1666 if (!timespec64_valid_strict(delta)) {
1667 printk_deferred(KERN_WARNING
1668 "__timekeeping_inject_sleeptime: Invalid "
1669 "sleep delta value!\n");
1672 tk_xtime_add(tk, delta);
1673 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1674 tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1675 tk_debug_account_sleep_time(delta);
1678 #if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1680 * We have three kinds of time sources to use for sleep time
1681 * injection, the preference order is:
1682 * 1) non-stop clocksource
1683 * 2) persistent clock (ie: RTC accessible when irqs are off)
1686 * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1687 * If system has neither 1) nor 2), 3) will be used finally.
1690 * If timekeeping has injected sleeptime via either 1) or 2),
1691 * 3) becomes needless, so in this case we don't need to call
1692 * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1695 bool timekeeping_rtc_skipresume(void)
1697 return !suspend_timing_needed;
1701 * 1) can be determined whether to use or not only when doing
1702 * timekeeping_resume() which is invoked after rtc_suspend(),
1703 * so we can't skip rtc_suspend() surely if system has 1).
1705 * But if system has 2), 2) will definitely be used, so in this
1706 * case we don't need to call rtc_suspend(), and this is what
1707 * timekeeping_rtc_skipsuspend() means.
1709 bool timekeeping_rtc_skipsuspend(void)
1711 return persistent_clock_exists;
1715 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1716 * @delta: pointer to a timespec64 delta value
1718 * This hook is for architectures that cannot support read_persistent_clock64
1719 * because their RTC/persistent clock is only accessible when irqs are enabled.
1720 * and also don't have an effective nonstop clocksource.
1722 * This function should only be called by rtc_resume(), and allows
1723 * a suspend offset to be injected into the timekeeping values.
1725 void timekeeping_inject_sleeptime64(const struct timespec64 *delta)
1727 struct timekeeper *tk = &tk_core.timekeeper;
1728 unsigned long flags;
1730 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1731 write_seqcount_begin(&tk_core.seq);
1733 suspend_timing_needed = false;
1735 timekeeping_forward_now(tk);
1737 __timekeeping_inject_sleeptime(tk, delta);
1739 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1741 write_seqcount_end(&tk_core.seq);
1742 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1744 /* signal hrtimers about time change */
1750 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1752 void timekeeping_resume(void)
1754 struct timekeeper *tk = &tk_core.timekeeper;
1755 struct clocksource *clock = tk->tkr_mono.clock;
1756 unsigned long flags;
1757 struct timespec64 ts_new, ts_delta;
1758 u64 cycle_now, nsec;
1759 bool inject_sleeptime = false;
1761 read_persistent_clock64(&ts_new);
1763 clockevents_resume();
1764 clocksource_resume();
1766 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1767 write_seqcount_begin(&tk_core.seq);
1770 * After system resumes, we need to calculate the suspended time and
1771 * compensate it for the OS time. There are 3 sources that could be
1772 * used: Nonstop clocksource during suspend, persistent clock and rtc
1775 * One specific platform may have 1 or 2 or all of them, and the
1776 * preference will be:
1777 * suspend-nonstop clocksource -> persistent clock -> rtc
1778 * The less preferred source will only be tried if there is no better
1779 * usable source. The rtc part is handled separately in rtc core code.
1781 cycle_now = tk_clock_read(&tk->tkr_mono);
1782 nsec = clocksource_stop_suspend_timing(clock, cycle_now);
1784 ts_delta = ns_to_timespec64(nsec);
1785 inject_sleeptime = true;
1786 } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1787 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1788 inject_sleeptime = true;
1791 if (inject_sleeptime) {
1792 suspend_timing_needed = false;
1793 __timekeeping_inject_sleeptime(tk, &ts_delta);
1796 /* Re-base the last cycle value */
1797 tk->tkr_mono.cycle_last = cycle_now;
1798 tk->tkr_raw.cycle_last = cycle_now;
1801 timekeeping_suspended = 0;
1802 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1803 write_seqcount_end(&tk_core.seq);
1804 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1806 touch_softlockup_watchdog();
1812 int timekeeping_suspend(void)
1814 struct timekeeper *tk = &tk_core.timekeeper;
1815 unsigned long flags;
1816 struct timespec64 delta, delta_delta;
1817 static struct timespec64 old_delta;
1818 struct clocksource *curr_clock;
1821 read_persistent_clock64(&timekeeping_suspend_time);
1824 * On some systems the persistent_clock can not be detected at
1825 * timekeeping_init by its return value, so if we see a valid
1826 * value returned, update the persistent_clock_exists flag.
1828 if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1829 persistent_clock_exists = true;
1831 suspend_timing_needed = true;
1833 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1834 write_seqcount_begin(&tk_core.seq);
1835 timekeeping_forward_now(tk);
1836 timekeeping_suspended = 1;
1839 * Since we've called forward_now, cycle_last stores the value
1840 * just read from the current clocksource. Save this to potentially
1841 * use in suspend timing.
1843 curr_clock = tk->tkr_mono.clock;
1844 cycle_now = tk->tkr_mono.cycle_last;
1845 clocksource_start_suspend_timing(curr_clock, cycle_now);
1847 if (persistent_clock_exists) {
1849 * To avoid drift caused by repeated suspend/resumes,
1850 * which each can add ~1 second drift error,
1851 * try to compensate so the difference in system time
1852 * and persistent_clock time stays close to constant.
1854 delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1855 delta_delta = timespec64_sub(delta, old_delta);
1856 if (abs(delta_delta.tv_sec) >= 2) {
1858 * if delta_delta is too large, assume time correction
1859 * has occurred and set old_delta to the current delta.
1863 /* Otherwise try to adjust old_system to compensate */
1864 timekeeping_suspend_time =
1865 timespec64_add(timekeeping_suspend_time, delta_delta);
1869 timekeeping_update(tk, TK_MIRROR);
1870 halt_fast_timekeeper(tk);
1871 write_seqcount_end(&tk_core.seq);
1872 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1875 clocksource_suspend();
1876 clockevents_suspend();
1881 /* sysfs resume/suspend bits for timekeeping */
1882 static struct syscore_ops timekeeping_syscore_ops = {
1883 .resume = timekeeping_resume,
1884 .suspend = timekeeping_suspend,
1887 static int __init timekeeping_init_ops(void)
1889 register_syscore_ops(&timekeeping_syscore_ops);
1892 device_initcall(timekeeping_init_ops);
1895 * Apply a multiplier adjustment to the timekeeper
1897 static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1901 s64 interval = tk->cycle_interval;
1903 if (mult_adj == 0) {
1905 } else if (mult_adj == -1) {
1906 interval = -interval;
1908 } else if (mult_adj != 1) {
1909 interval *= mult_adj;
1914 * So the following can be confusing.
1916 * To keep things simple, lets assume mult_adj == 1 for now.
1918 * When mult_adj != 1, remember that the interval and offset values
1919 * have been appropriately scaled so the math is the same.
1921 * The basic idea here is that we're increasing the multiplier
1922 * by one, this causes the xtime_interval to be incremented by
1923 * one cycle_interval. This is because:
1924 * xtime_interval = cycle_interval * mult
1925 * So if mult is being incremented by one:
1926 * xtime_interval = cycle_interval * (mult + 1)
1928 * xtime_interval = (cycle_interval * mult) + cycle_interval
1929 * Which can be shortened to:
1930 * xtime_interval += cycle_interval
1932 * So offset stores the non-accumulated cycles. Thus the current
1933 * time (in shifted nanoseconds) is:
1934 * now = (offset * adj) + xtime_nsec
1935 * Now, even though we're adjusting the clock frequency, we have
1936 * to keep time consistent. In other words, we can't jump back
1937 * in time, and we also want to avoid jumping forward in time.
1939 * So given the same offset value, we need the time to be the same
1940 * both before and after the freq adjustment.
1941 * now = (offset * adj_1) + xtime_nsec_1
1942 * now = (offset * adj_2) + xtime_nsec_2
1944 * (offset * adj_1) + xtime_nsec_1 =
1945 * (offset * adj_2) + xtime_nsec_2
1949 * (offset * adj_1) + xtime_nsec_1 =
1950 * (offset * (adj_1+1)) + xtime_nsec_2
1951 * (offset * adj_1) + xtime_nsec_1 =
1952 * (offset * adj_1) + offset + xtime_nsec_2
1953 * Canceling the sides:
1954 * xtime_nsec_1 = offset + xtime_nsec_2
1956 * xtime_nsec_2 = xtime_nsec_1 - offset
1957 * Which simplfies to:
1958 * xtime_nsec -= offset
1960 if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1961 /* NTP adjustment caused clocksource mult overflow */
1966 tk->tkr_mono.mult += mult_adj;
1967 tk->xtime_interval += interval;
1968 tk->tkr_mono.xtime_nsec -= offset;
1972 * Adjust the timekeeper's multiplier to the correct frequency
1973 * and also to reduce the accumulated error value.
1975 static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
1980 * Determine the multiplier from the current NTP tick length.
1981 * Avoid expensive division when the tick length doesn't change.
1983 if (likely(tk->ntp_tick == ntp_tick_length())) {
1984 mult = tk->tkr_mono.mult - tk->ntp_err_mult;
1986 tk->ntp_tick = ntp_tick_length();
1987 mult = div64_u64((tk->ntp_tick >> tk->ntp_error_shift) -
1988 tk->xtime_remainder, tk->cycle_interval);
1992 * If the clock is behind the NTP time, increase the multiplier by 1
1993 * to catch up with it. If it's ahead and there was a remainder in the
1994 * tick division, the clock will slow down. Otherwise it will stay
1995 * ahead until the tick length changes to a non-divisible value.
1997 tk->ntp_err_mult = tk->ntp_error > 0 ? 1 : 0;
1998 mult += tk->ntp_err_mult;
2000 timekeeping_apply_adjustment(tk, offset, mult - tk->tkr_mono.mult);
2002 if (unlikely(tk->tkr_mono.clock->maxadj &&
2003 (abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
2004 > tk->tkr_mono.clock->maxadj))) {
2005 printk_once(KERN_WARNING
2006 "Adjusting %s more than 11%% (%ld vs %ld)\n",
2007 tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
2008 (long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
2012 * It may be possible that when we entered this function, xtime_nsec
2013 * was very small. Further, if we're slightly speeding the clocksource
2014 * in the code above, its possible the required corrective factor to
2015 * xtime_nsec could cause it to underflow.
2017 * Now, since we have already accumulated the second and the NTP
2018 * subsystem has been notified via second_overflow(), we need to skip
2021 if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
2022 tk->tkr_mono.xtime_nsec += (u64)NSEC_PER_SEC <<
2025 tk->skip_second_overflow = 1;
2030 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
2032 * Helper function that accumulates the nsecs greater than a second
2033 * from the xtime_nsec field to the xtime_secs field.
2034 * It also calls into the NTP code to handle leapsecond processing.
2037 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
2039 u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
2040 unsigned int clock_set = 0;
2042 while (tk->tkr_mono.xtime_nsec >= nsecps) {
2045 tk->tkr_mono.xtime_nsec -= nsecps;
2049 * Skip NTP update if this second was accumulated before,
2050 * i.e. xtime_nsec underflowed in timekeeping_adjust()
2052 if (unlikely(tk->skip_second_overflow)) {
2053 tk->skip_second_overflow = 0;
2057 /* Figure out if its a leap sec and apply if needed */
2058 leap = second_overflow(tk->xtime_sec);
2059 if (unlikely(leap)) {
2060 struct timespec64 ts;
2062 tk->xtime_sec += leap;
2066 tk_set_wall_to_mono(tk,
2067 timespec64_sub(tk->wall_to_monotonic, ts));
2069 __timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
2071 clock_set = TK_CLOCK_WAS_SET;
2078 * logarithmic_accumulation - shifted accumulation of cycles
2080 * This functions accumulates a shifted interval of cycles into
2081 * a shifted interval nanoseconds. Allows for O(log) accumulation
2084 * Returns the unconsumed cycles.
2086 static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset,
2087 u32 shift, unsigned int *clock_set)
2089 u64 interval = tk->cycle_interval << shift;
2092 /* If the offset is smaller than a shifted interval, do nothing */
2093 if (offset < interval)
2096 /* Accumulate one shifted interval */
2098 tk->tkr_mono.cycle_last += interval;
2099 tk->tkr_raw.cycle_last += interval;
2101 tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
2102 *clock_set |= accumulate_nsecs_to_secs(tk);
2104 /* Accumulate raw time */
2105 tk->tkr_raw.xtime_nsec += tk->raw_interval << shift;
2106 snsec_per_sec = (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
2107 while (tk->tkr_raw.xtime_nsec >= snsec_per_sec) {
2108 tk->tkr_raw.xtime_nsec -= snsec_per_sec;
2112 /* Accumulate error between NTP and clock interval */
2113 tk->ntp_error += tk->ntp_tick << shift;
2114 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
2115 (tk->ntp_error_shift + shift);
2121 * timekeeping_advance - Updates the timekeeper to the current time and
2122 * current NTP tick length
2124 static void timekeeping_advance(enum timekeeping_adv_mode mode)
2126 struct timekeeper *real_tk = &tk_core.timekeeper;
2127 struct timekeeper *tk = &shadow_timekeeper;
2129 int shift = 0, maxshift;
2130 unsigned int clock_set = 0;
2131 unsigned long flags;
2133 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2135 /* Make sure we're fully resumed: */
2136 if (unlikely(timekeeping_suspended))
2139 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
2140 offset = real_tk->cycle_interval;
2142 if (mode != TK_ADV_TICK)
2145 offset = clocksource_delta(tk_clock_read(&tk->tkr_mono),
2146 tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
2148 /* Check if there's really nothing to do */
2149 if (offset < real_tk->cycle_interval && mode == TK_ADV_TICK)
2153 /* Do some additional sanity checking */
2154 timekeeping_check_update(tk, offset);
2157 * With NO_HZ we may have to accumulate many cycle_intervals
2158 * (think "ticks") worth of time at once. To do this efficiently,
2159 * we calculate the largest doubling multiple of cycle_intervals
2160 * that is smaller than the offset. We then accumulate that
2161 * chunk in one go, and then try to consume the next smaller
2164 shift = ilog2(offset) - ilog2(tk->cycle_interval);
2165 shift = max(0, shift);
2166 /* Bound shift to one less than what overflows tick_length */
2167 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
2168 shift = min(shift, maxshift);
2169 while (offset >= tk->cycle_interval) {
2170 offset = logarithmic_accumulation(tk, offset, shift,
2172 if (offset < tk->cycle_interval<<shift)
2176 /* Adjust the multiplier to correct NTP error */
2177 timekeeping_adjust(tk, offset);
2180 * Finally, make sure that after the rounding
2181 * xtime_nsec isn't larger than NSEC_PER_SEC
2183 clock_set |= accumulate_nsecs_to_secs(tk);
2185 write_seqcount_begin(&tk_core.seq);
2187 * Update the real timekeeper.
2189 * We could avoid this memcpy by switching pointers, but that
2190 * requires changes to all other timekeeper usage sites as
2191 * well, i.e. move the timekeeper pointer getter into the
2192 * spinlocked/seqcount protected sections. And we trade this
2193 * memcpy under the tk_core.seq against one before we start
2196 timekeeping_update(tk, clock_set);
2197 memcpy(real_tk, tk, sizeof(*tk));
2198 /* The memcpy must come last. Do not put anything here! */
2199 write_seqcount_end(&tk_core.seq);
2201 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2203 /* Have to call _delayed version, since in irq context*/
2204 clock_was_set_delayed();
2208 * update_wall_time - Uses the current clocksource to increment the wall time
2211 void update_wall_time(void)
2213 timekeeping_advance(TK_ADV_TICK);
2217 * getboottime64 - Return the real time of system boot.
2218 * @ts: pointer to the timespec64 to be set
2220 * Returns the wall-time of boot in a timespec64.
2222 * This is based on the wall_to_monotonic offset and the total suspend
2223 * time. Calls to settimeofday will affect the value returned (which
2224 * basically means that however wrong your real time clock is at boot time,
2225 * you get the right time here).
2227 void getboottime64(struct timespec64 *ts)
2229 struct timekeeper *tk = &tk_core.timekeeper;
2230 ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
2232 *ts = ktime_to_timespec64(t);
2234 EXPORT_SYMBOL_GPL(getboottime64);
2236 void ktime_get_coarse_real_ts64(struct timespec64 *ts)
2238 struct timekeeper *tk = &tk_core.timekeeper;
2242 seq = read_seqcount_begin(&tk_core.seq);
2245 } while (read_seqcount_retry(&tk_core.seq, seq));
2247 EXPORT_SYMBOL(ktime_get_coarse_real_ts64);
2249 void ktime_get_coarse_ts64(struct timespec64 *ts)
2251 struct timekeeper *tk = &tk_core.timekeeper;
2252 struct timespec64 now, mono;
2256 seq = read_seqcount_begin(&tk_core.seq);
2259 mono = tk->wall_to_monotonic;
2260 } while (read_seqcount_retry(&tk_core.seq, seq));
2262 set_normalized_timespec64(ts, now.tv_sec + mono.tv_sec,
2263 now.tv_nsec + mono.tv_nsec);
2265 EXPORT_SYMBOL(ktime_get_coarse_ts64);
2268 * Must hold jiffies_lock
2270 void do_timer(unsigned long ticks)
2272 jiffies_64 += ticks;
2277 * ktime_get_update_offsets_now - hrtimer helper
2278 * @cwsseq: pointer to check and store the clock was set sequence number
2279 * @offs_real: pointer to storage for monotonic -> realtime offset
2280 * @offs_boot: pointer to storage for monotonic -> boottime offset
2281 * @offs_tai: pointer to storage for monotonic -> clock tai offset
2283 * Returns current monotonic time and updates the offsets if the
2284 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
2287 * Called from hrtimer_interrupt() or retrigger_next_event()
2289 ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
2290 ktime_t *offs_boot, ktime_t *offs_tai)
2292 struct timekeeper *tk = &tk_core.timekeeper;
2298 seq = read_seqcount_begin(&tk_core.seq);
2300 base = tk->tkr_mono.base;
2301 nsecs = timekeeping_get_ns(&tk->tkr_mono);
2302 base = ktime_add_ns(base, nsecs);
2304 if (*cwsseq != tk->clock_was_set_seq) {
2305 *cwsseq = tk->clock_was_set_seq;
2306 *offs_real = tk->offs_real;
2307 *offs_boot = tk->offs_boot;
2308 *offs_tai = tk->offs_tai;
2311 /* Handle leapsecond insertion adjustments */
2312 if (unlikely(base >= tk->next_leap_ktime))
2313 *offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
2315 } while (read_seqcount_retry(&tk_core.seq, seq));
2321 * timekeeping_validate_timex - Ensures the timex is ok for use in do_adjtimex
2323 static int timekeeping_validate_timex(const struct __kernel_timex *txc)
2325 if (txc->modes & ADJ_ADJTIME) {
2326 /* singleshot must not be used with any other mode bits */
2327 if (!(txc->modes & ADJ_OFFSET_SINGLESHOT))
2329 if (!(txc->modes & ADJ_OFFSET_READONLY) &&
2330 !capable(CAP_SYS_TIME))
2333 /* In order to modify anything, you gotta be super-user! */
2334 if (txc->modes && !capable(CAP_SYS_TIME))
2337 * if the quartz is off by more than 10% then
2338 * something is VERY wrong!
2340 if (txc->modes & ADJ_TICK &&
2341 (txc->tick < 900000/USER_HZ ||
2342 txc->tick > 1100000/USER_HZ))
2346 if (txc->modes & ADJ_SETOFFSET) {
2347 /* In order to inject time, you gotta be super-user! */
2348 if (!capable(CAP_SYS_TIME))
2352 * Validate if a timespec/timeval used to inject a time
2353 * offset is valid. Offsets can be postive or negative, so
2354 * we don't check tv_sec. The value of the timeval/timespec
2355 * is the sum of its fields,but *NOTE*:
2356 * The field tv_usec/tv_nsec must always be non-negative and
2357 * we can't have more nanoseconds/microseconds than a second.
2359 if (txc->time.tv_usec < 0)
2362 if (txc->modes & ADJ_NANO) {
2363 if (txc->time.tv_usec >= NSEC_PER_SEC)
2366 if (txc->time.tv_usec >= USEC_PER_SEC)
2372 * Check for potential multiplication overflows that can
2373 * only happen on 64-bit systems:
2375 if ((txc->modes & ADJ_FREQUENCY) && (BITS_PER_LONG == 64)) {
2376 if (LLONG_MIN / PPM_SCALE > txc->freq)
2378 if (LLONG_MAX / PPM_SCALE < txc->freq)
2386 * random_get_entropy_fallback - Returns the raw clock source value,
2387 * used by random.c for platforms with no valid random_get_entropy().
2389 unsigned long random_get_entropy_fallback(void)
2391 struct tk_read_base *tkr = &tk_core.timekeeper.tkr_mono;
2392 struct clocksource *clock = READ_ONCE(tkr->clock);
2394 if (unlikely(timekeeping_suspended || !clock))
2396 return clock->read(clock);
2398 EXPORT_SYMBOL_GPL(random_get_entropy_fallback);
2401 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2403 int do_adjtimex(struct __kernel_timex *txc)
2405 struct timekeeper *tk = &tk_core.timekeeper;
2406 struct audit_ntp_data ad;
2407 unsigned long flags;
2408 struct timespec64 ts;
2412 /* Validate the data before disabling interrupts */
2413 ret = timekeeping_validate_timex(txc);
2416 add_device_randomness(txc, sizeof(*txc));
2418 if (txc->modes & ADJ_SETOFFSET) {
2419 struct timespec64 delta;
2420 delta.tv_sec = txc->time.tv_sec;
2421 delta.tv_nsec = txc->time.tv_usec;
2422 if (!(txc->modes & ADJ_NANO))
2423 delta.tv_nsec *= 1000;
2424 ret = timekeeping_inject_offset(&delta);
2428 audit_tk_injoffset(delta);
2431 audit_ntp_init(&ad);
2433 ktime_get_real_ts64(&ts);
2434 add_device_randomness(&ts, sizeof(ts));
2436 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2437 write_seqcount_begin(&tk_core.seq);
2439 orig_tai = tai = tk->tai_offset;
2440 ret = __do_adjtimex(txc, &ts, &tai, &ad);
2442 if (tai != orig_tai) {
2443 __timekeeping_set_tai_offset(tk, tai);
2444 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
2446 tk_update_leap_state(tk);
2448 write_seqcount_end(&tk_core.seq);
2449 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2453 /* Update the multiplier immediately if frequency was set directly */
2454 if (txc->modes & (ADJ_FREQUENCY | ADJ_TICK))
2455 timekeeping_advance(TK_ADV_FREQ);
2457 if (tai != orig_tai)
2460 ntp_notify_cmos_timer();
2465 #ifdef CONFIG_NTP_PPS
2467 * hardpps() - Accessor function to NTP __hardpps function
2469 void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2471 unsigned long flags;
2473 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2474 write_seqcount_begin(&tk_core.seq);
2476 __hardpps(phase_ts, raw_ts);
2478 write_seqcount_end(&tk_core.seq);
2479 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2481 EXPORT_SYMBOL(hardpps);
2482 #endif /* CONFIG_NTP_PPS */
2485 * xtime_update() - advances the timekeeping infrastructure
2486 * @ticks: number of ticks, that have elapsed since the last call.
2488 * Must be called with interrupts disabled.
2490 void xtime_update(unsigned long ticks)
2492 raw_spin_lock(&jiffies_lock);
2493 write_seqcount_begin(&jiffies_seq);
2495 write_seqcount_end(&jiffies_seq);
2496 raw_spin_unlock(&jiffies_lock);