2 * linux/kernel/time/timekeeping.c
4 * Kernel timekeeping code and accessor functions
6 * This code was moved from linux/kernel/timer.c.
7 * Please see that file for copyright and history logs.
11 #include <linux/timekeeper_internal.h>
12 #include <linux/module.h>
13 #include <linux/interrupt.h>
14 #include <linux/percpu.h>
15 #include <linux/init.h>
17 #include <linux/sched.h>
18 #include <linux/syscore_ops.h>
19 #include <linux/clocksource.h>
20 #include <linux/jiffies.h>
21 #include <linux/time.h>
22 #include <linux/tick.h>
23 #include <linux/stop_machine.h>
24 #include <linux/pvclock_gtod.h>
25 #include <linux/compiler.h>
27 #include "tick-internal.h"
28 #include "ntp_internal.h"
29 #include "timekeeping_internal.h"
31 #define TK_CLEAR_NTP (1 << 0)
32 #define TK_MIRROR (1 << 1)
33 #define TK_CLOCK_WAS_SET (1 << 2)
36 * The most important data for readout fits into a single 64 byte
41 struct timekeeper timekeeper;
42 } tk_core ____cacheline_aligned;
44 static DEFINE_RAW_SPINLOCK(timekeeper_lock);
45 static struct timekeeper shadow_timekeeper;
48 * struct tk_fast - NMI safe timekeeper
49 * @seq: Sequence counter for protecting updates. The lowest bit
50 * is the index for the tk_read_base array
51 * @base: tk_read_base array. Access is indexed by the lowest bit of
54 * See @update_fast_timekeeper() below.
58 struct tk_read_base base[2];
61 static struct tk_fast tk_fast_mono ____cacheline_aligned;
62 static struct tk_fast tk_fast_raw ____cacheline_aligned;
64 /* flag for if timekeeping is suspended */
65 int __read_mostly timekeeping_suspended;
67 static inline void tk_normalize_xtime(struct timekeeper *tk)
69 while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) {
70 tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
75 static inline struct timespec64 tk_xtime(struct timekeeper *tk)
79 ts.tv_sec = tk->xtime_sec;
80 ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
84 static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
86 tk->xtime_sec = ts->tv_sec;
87 tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift;
90 static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
92 tk->xtime_sec += ts->tv_sec;
93 tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift;
94 tk_normalize_xtime(tk);
97 static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
99 struct timespec64 tmp;
102 * Verify consistency of: offset_real = -wall_to_monotonic
103 * before modifying anything
105 set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
106 -tk->wall_to_monotonic.tv_nsec);
107 WARN_ON_ONCE(tk->offs_real.tv64 != timespec64_to_ktime(tmp).tv64);
108 tk->wall_to_monotonic = wtm;
109 set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
110 tk->offs_real = timespec64_to_ktime(tmp);
111 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0));
114 static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
116 tk->offs_boot = ktime_add(tk->offs_boot, delta);
119 #ifdef CONFIG_DEBUG_TIMEKEEPING
120 #define WARNING_FREQ (HZ*300) /* 5 minute rate-limiting */
122 static void timekeeping_check_update(struct timekeeper *tk, cycle_t offset)
125 cycle_t max_cycles = tk->tkr_mono.clock->max_cycles;
126 const char *name = tk->tkr_mono.clock->name;
128 if (offset > max_cycles) {
129 printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n",
130 offset, name, max_cycles);
131 printk_deferred(" timekeeping: Your kernel is sick, but tries to cope by capping time updates\n");
133 if (offset > (max_cycles >> 1)) {
134 printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the the '%s' clock's 50%% safety margin (%lld)\n",
135 offset, name, max_cycles >> 1);
136 printk_deferred(" timekeeping: Your kernel is still fine, but is feeling a bit nervous\n");
140 if (tk->underflow_seen) {
141 if (jiffies - tk->last_warning > WARNING_FREQ) {
142 printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name);
143 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
144 printk_deferred(" Your kernel is probably still fine.\n");
145 tk->last_warning = jiffies;
147 tk->underflow_seen = 0;
150 if (tk->overflow_seen) {
151 if (jiffies - tk->last_warning > WARNING_FREQ) {
152 printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name);
153 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
154 printk_deferred(" Your kernel is probably still fine.\n");
155 tk->last_warning = jiffies;
157 tk->overflow_seen = 0;
161 static inline cycle_t timekeeping_get_delta(struct tk_read_base *tkr)
163 struct timekeeper *tk = &tk_core.timekeeper;
164 cycle_t now, last, mask, max, delta;
168 * Since we're called holding a seqlock, the data may shift
169 * under us while we're doing the calculation. This can cause
170 * false positives, since we'd note a problem but throw the
171 * results away. So nest another seqlock here to atomically
172 * grab the points we are checking with.
175 seq = read_seqcount_begin(&tk_core.seq);
176 now = tkr->read(tkr->clock);
177 last = tkr->cycle_last;
179 max = tkr->clock->max_cycles;
180 } while (read_seqcount_retry(&tk_core.seq, seq));
182 delta = clocksource_delta(now, last, mask);
185 * Try to catch underflows by checking if we are seeing small
186 * mask-relative negative values.
188 if (unlikely((~delta & mask) < (mask >> 3))) {
189 tk->underflow_seen = 1;
193 /* Cap delta value to the max_cycles values to avoid mult overflows */
194 if (unlikely(delta > max)) {
195 tk->overflow_seen = 1;
196 delta = tkr->clock->max_cycles;
202 static inline void timekeeping_check_update(struct timekeeper *tk, cycle_t offset)
205 static inline cycle_t timekeeping_get_delta(struct tk_read_base *tkr)
207 cycle_t cycle_now, delta;
209 /* read clocksource */
210 cycle_now = tkr->read(tkr->clock);
212 /* calculate the delta since the last update_wall_time */
213 delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
220 * tk_setup_internals - Set up internals to use clocksource clock.
222 * @tk: The target timekeeper to setup.
223 * @clock: Pointer to clocksource.
225 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
226 * pair and interval request.
228 * Unless you're the timekeeping code, you should not be using this!
230 static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
233 u64 tmp, ntpinterval;
234 struct clocksource *old_clock;
236 old_clock = tk->tkr_mono.clock;
237 tk->tkr_mono.clock = clock;
238 tk->tkr_mono.read = clock->read;
239 tk->tkr_mono.mask = clock->mask;
240 tk->tkr_mono.cycle_last = tk->tkr_mono.read(clock);
242 tk->tkr_raw.clock = clock;
243 tk->tkr_raw.read = clock->read;
244 tk->tkr_raw.mask = clock->mask;
245 tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
247 /* Do the ns -> cycle conversion first, using original mult */
248 tmp = NTP_INTERVAL_LENGTH;
249 tmp <<= clock->shift;
251 tmp += clock->mult/2;
252 do_div(tmp, clock->mult);
256 interval = (cycle_t) tmp;
257 tk->cycle_interval = interval;
259 /* Go back from cycles -> shifted ns */
260 tk->xtime_interval = (u64) interval * clock->mult;
261 tk->xtime_remainder = ntpinterval - tk->xtime_interval;
263 ((u64) interval * clock->mult) >> clock->shift;
265 /* if changing clocks, convert xtime_nsec shift units */
267 int shift_change = clock->shift - old_clock->shift;
268 if (shift_change < 0)
269 tk->tkr_mono.xtime_nsec >>= -shift_change;
271 tk->tkr_mono.xtime_nsec <<= shift_change;
273 tk->tkr_raw.xtime_nsec = 0;
275 tk->tkr_mono.shift = clock->shift;
276 tk->tkr_raw.shift = clock->shift;
279 tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
280 tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
283 * The timekeeper keeps its own mult values for the currently
284 * active clocksource. These value will be adjusted via NTP
285 * to counteract clock drifting.
287 tk->tkr_mono.mult = clock->mult;
288 tk->tkr_raw.mult = clock->mult;
289 tk->ntp_err_mult = 0;
292 /* Timekeeper helper functions. */
294 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
295 static u32 default_arch_gettimeoffset(void) { return 0; }
296 u32 (*arch_gettimeoffset)(void) = default_arch_gettimeoffset;
298 static inline u32 arch_gettimeoffset(void) { return 0; }
301 static inline s64 timekeeping_get_ns(struct tk_read_base *tkr)
306 delta = timekeeping_get_delta(tkr);
308 nsec = (delta * tkr->mult + tkr->xtime_nsec) >> tkr->shift;
310 /* If arch requires, add in get_arch_timeoffset() */
311 return nsec + arch_gettimeoffset();
315 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
316 * @tkr: Timekeeping readout base from which we take the update
318 * We want to use this from any context including NMI and tracing /
319 * instrumenting the timekeeping code itself.
321 * Employ the latch technique; see @raw_write_seqcount_latch.
323 * So if a NMI hits the update of base[0] then it will use base[1]
324 * which is still consistent. In the worst case this can result is a
325 * slightly wrong timestamp (a few nanoseconds). See
326 * @ktime_get_mono_fast_ns.
328 static void update_fast_timekeeper(struct tk_read_base *tkr, struct tk_fast *tkf)
330 struct tk_read_base *base = tkf->base;
332 /* Force readers off to base[1] */
333 raw_write_seqcount_latch(&tkf->seq);
336 memcpy(base, tkr, sizeof(*base));
338 /* Force readers back to base[0] */
339 raw_write_seqcount_latch(&tkf->seq);
342 memcpy(base + 1, base, sizeof(*base));
346 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
348 * This timestamp is not guaranteed to be monotonic across an update.
349 * The timestamp is calculated by:
351 * now = base_mono + clock_delta * slope
353 * So if the update lowers the slope, readers who are forced to the
354 * not yet updated second array are still using the old steeper slope.
363 * |12345678---> reader order
369 * So reader 6 will observe time going backwards versus reader 5.
371 * While other CPUs are likely to be able observe that, the only way
372 * for a CPU local observation is when an NMI hits in the middle of
373 * the update. Timestamps taken from that NMI context might be ahead
374 * of the following timestamps. Callers need to be aware of that and
377 static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
379 struct tk_read_base *tkr;
384 seq = raw_read_seqcount_latch(&tkf->seq);
385 tkr = tkf->base + (seq & 0x01);
386 now = ktime_to_ns(tkr->base);
388 now += clocksource_delta(tkr->read(tkr->clock),
389 tkr->cycle_last, tkr->mask);
390 } while (read_seqcount_retry(&tkf->seq, seq));
395 u64 ktime_get_mono_fast_ns(void)
397 return __ktime_get_fast_ns(&tk_fast_mono);
399 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
401 u64 ktime_get_raw_fast_ns(void)
403 return __ktime_get_fast_ns(&tk_fast_raw);
405 EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
407 /* Suspend-time cycles value for halted fast timekeeper. */
408 static cycle_t cycles_at_suspend;
410 static cycle_t dummy_clock_read(struct clocksource *cs)
412 return cycles_at_suspend;
416 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
417 * @tk: Timekeeper to snapshot.
419 * It generally is unsafe to access the clocksource after timekeeping has been
420 * suspended, so take a snapshot of the readout base of @tk and use it as the
421 * fast timekeeper's readout base while suspended. It will return the same
422 * number of cycles every time until timekeeping is resumed at which time the
423 * proper readout base for the fast timekeeper will be restored automatically.
425 static void halt_fast_timekeeper(struct timekeeper *tk)
427 static struct tk_read_base tkr_dummy;
428 struct tk_read_base *tkr = &tk->tkr_mono;
430 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
431 cycles_at_suspend = tkr->read(tkr->clock);
432 tkr_dummy.read = dummy_clock_read;
433 update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
436 memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
437 tkr_dummy.read = dummy_clock_read;
438 update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
441 #ifdef CONFIG_GENERIC_TIME_VSYSCALL_OLD
443 static inline void update_vsyscall(struct timekeeper *tk)
445 struct timespec xt, wm;
447 xt = timespec64_to_timespec(tk_xtime(tk));
448 wm = timespec64_to_timespec(tk->wall_to_monotonic);
449 update_vsyscall_old(&xt, &wm, tk->tkr_mono.clock, tk->tkr_mono.mult,
450 tk->tkr_mono.cycle_last);
453 static inline void old_vsyscall_fixup(struct timekeeper *tk)
458 * Store only full nanoseconds into xtime_nsec after rounding
459 * it up and add the remainder to the error difference.
460 * XXX - This is necessary to avoid small 1ns inconsistnecies caused
461 * by truncating the remainder in vsyscalls. However, it causes
462 * additional work to be done in timekeeping_adjust(). Once
463 * the vsyscall implementations are converted to use xtime_nsec
464 * (shifted nanoseconds), and CONFIG_GENERIC_TIME_VSYSCALL_OLD
465 * users are removed, this can be killed.
467 remainder = tk->tkr_mono.xtime_nsec & ((1ULL << tk->tkr_mono.shift) - 1);
468 tk->tkr_mono.xtime_nsec -= remainder;
469 tk->tkr_mono.xtime_nsec += 1ULL << tk->tkr_mono.shift;
470 tk->ntp_error += remainder << tk->ntp_error_shift;
471 tk->ntp_error -= (1ULL << tk->tkr_mono.shift) << tk->ntp_error_shift;
474 #define old_vsyscall_fixup(tk)
477 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
479 static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
481 raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
485 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
487 int pvclock_gtod_register_notifier(struct notifier_block *nb)
489 struct timekeeper *tk = &tk_core.timekeeper;
493 raw_spin_lock_irqsave(&timekeeper_lock, flags);
494 ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
495 update_pvclock_gtod(tk, true);
496 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
500 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
503 * pvclock_gtod_unregister_notifier - unregister a pvclock
504 * timedata update listener
506 int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
511 raw_spin_lock_irqsave(&timekeeper_lock, flags);
512 ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
513 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
517 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
520 * tk_update_leap_state - helper to update the next_leap_ktime
522 static inline void tk_update_leap_state(struct timekeeper *tk)
524 tk->next_leap_ktime = ntp_get_next_leap();
525 if (tk->next_leap_ktime.tv64 != KTIME_MAX)
526 /* Convert to monotonic time */
527 tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
531 * Update the ktime_t based scalar nsec members of the timekeeper
533 static inline void tk_update_ktime_data(struct timekeeper *tk)
539 * The xtime based monotonic readout is:
540 * nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
541 * The ktime based monotonic readout is:
542 * nsec = base_mono + now();
543 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
545 seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
546 nsec = (u32) tk->wall_to_monotonic.tv_nsec;
547 tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
549 /* Update the monotonic raw base */
550 tk->tkr_raw.base = timespec64_to_ktime(tk->raw_time);
553 * The sum of the nanoseconds portions of xtime and
554 * wall_to_monotonic can be greater/equal one second. Take
555 * this into account before updating tk->ktime_sec.
557 nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
558 if (nsec >= NSEC_PER_SEC)
560 tk->ktime_sec = seconds;
563 /* must hold timekeeper_lock */
564 static void timekeeping_update(struct timekeeper *tk, unsigned int action)
566 if (action & TK_CLEAR_NTP) {
571 tk_update_leap_state(tk);
572 tk_update_ktime_data(tk);
575 update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
577 update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
578 update_fast_timekeeper(&tk->tkr_raw, &tk_fast_raw);
580 if (action & TK_CLOCK_WAS_SET)
581 tk->clock_was_set_seq++;
583 * The mirroring of the data to the shadow-timekeeper needs
584 * to happen last here to ensure we don't over-write the
585 * timekeeper structure on the next update with stale data
587 if (action & TK_MIRROR)
588 memcpy(&shadow_timekeeper, &tk_core.timekeeper,
589 sizeof(tk_core.timekeeper));
593 * timekeeping_forward_now - update clock to the current time
595 * Forward the current clock to update its state since the last call to
596 * update_wall_time(). This is useful before significant clock changes,
597 * as it avoids having to deal with this time offset explicitly.
599 static void timekeeping_forward_now(struct timekeeper *tk)
601 struct clocksource *clock = tk->tkr_mono.clock;
602 cycle_t cycle_now, delta;
605 cycle_now = tk->tkr_mono.read(clock);
606 delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
607 tk->tkr_mono.cycle_last = cycle_now;
608 tk->tkr_raw.cycle_last = cycle_now;
610 tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult;
612 /* If arch requires, add in get_arch_timeoffset() */
613 tk->tkr_mono.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_mono.shift;
615 tk_normalize_xtime(tk);
617 nsec = clocksource_cyc2ns(delta, tk->tkr_raw.mult, tk->tkr_raw.shift);
618 timespec64_add_ns(&tk->raw_time, nsec);
622 * __getnstimeofday64 - Returns the time of day in a timespec64.
623 * @ts: pointer to the timespec to be set
625 * Updates the time of day in the timespec.
626 * Returns 0 on success, or -ve when suspended (timespec will be undefined).
628 int __getnstimeofday64(struct timespec64 *ts)
630 struct timekeeper *tk = &tk_core.timekeeper;
635 seq = read_seqcount_begin(&tk_core.seq);
637 ts->tv_sec = tk->xtime_sec;
638 nsecs = timekeeping_get_ns(&tk->tkr_mono);
640 } while (read_seqcount_retry(&tk_core.seq, seq));
643 timespec64_add_ns(ts, nsecs);
646 * Do not bail out early, in case there were callers still using
647 * the value, even in the face of the WARN_ON.
649 if (unlikely(timekeeping_suspended))
653 EXPORT_SYMBOL(__getnstimeofday64);
656 * getnstimeofday64 - Returns the time of day in a timespec64.
657 * @ts: pointer to the timespec64 to be set
659 * Returns the time of day in a timespec64 (WARN if suspended).
661 void getnstimeofday64(struct timespec64 *ts)
663 WARN_ON(__getnstimeofday64(ts));
665 EXPORT_SYMBOL(getnstimeofday64);
667 ktime_t ktime_get(void)
669 struct timekeeper *tk = &tk_core.timekeeper;
674 WARN_ON(timekeeping_suspended);
677 seq = read_seqcount_begin(&tk_core.seq);
678 base = tk->tkr_mono.base;
679 nsecs = timekeeping_get_ns(&tk->tkr_mono);
681 } while (read_seqcount_retry(&tk_core.seq, seq));
683 return ktime_add_ns(base, nsecs);
685 EXPORT_SYMBOL_GPL(ktime_get);
687 u32 ktime_get_resolution_ns(void)
689 struct timekeeper *tk = &tk_core.timekeeper;
693 WARN_ON(timekeeping_suspended);
696 seq = read_seqcount_begin(&tk_core.seq);
697 nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
698 } while (read_seqcount_retry(&tk_core.seq, seq));
702 EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
704 static ktime_t *offsets[TK_OFFS_MAX] = {
705 [TK_OFFS_REAL] = &tk_core.timekeeper.offs_real,
706 [TK_OFFS_BOOT] = &tk_core.timekeeper.offs_boot,
707 [TK_OFFS_TAI] = &tk_core.timekeeper.offs_tai,
710 ktime_t ktime_get_with_offset(enum tk_offsets offs)
712 struct timekeeper *tk = &tk_core.timekeeper;
714 ktime_t base, *offset = offsets[offs];
717 WARN_ON(timekeeping_suspended);
720 seq = read_seqcount_begin(&tk_core.seq);
721 base = ktime_add(tk->tkr_mono.base, *offset);
722 nsecs = timekeeping_get_ns(&tk->tkr_mono);
724 } while (read_seqcount_retry(&tk_core.seq, seq));
726 return ktime_add_ns(base, nsecs);
729 EXPORT_SYMBOL_GPL(ktime_get_with_offset);
732 * ktime_mono_to_any() - convert mononotic time to any other time
733 * @tmono: time to convert.
734 * @offs: which offset to use
736 ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
738 ktime_t *offset = offsets[offs];
743 seq = read_seqcount_begin(&tk_core.seq);
744 tconv = ktime_add(tmono, *offset);
745 } while (read_seqcount_retry(&tk_core.seq, seq));
749 EXPORT_SYMBOL_GPL(ktime_mono_to_any);
752 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
754 ktime_t ktime_get_raw(void)
756 struct timekeeper *tk = &tk_core.timekeeper;
762 seq = read_seqcount_begin(&tk_core.seq);
763 base = tk->tkr_raw.base;
764 nsecs = timekeeping_get_ns(&tk->tkr_raw);
766 } while (read_seqcount_retry(&tk_core.seq, seq));
768 return ktime_add_ns(base, nsecs);
770 EXPORT_SYMBOL_GPL(ktime_get_raw);
773 * ktime_get_ts64 - get the monotonic clock in timespec64 format
774 * @ts: pointer to timespec variable
776 * The function calculates the monotonic clock from the realtime
777 * clock and the wall_to_monotonic offset and stores the result
778 * in normalized timespec64 format in the variable pointed to by @ts.
780 void ktime_get_ts64(struct timespec64 *ts)
782 struct timekeeper *tk = &tk_core.timekeeper;
783 struct timespec64 tomono;
787 WARN_ON(timekeeping_suspended);
790 seq = read_seqcount_begin(&tk_core.seq);
791 ts->tv_sec = tk->xtime_sec;
792 nsec = timekeeping_get_ns(&tk->tkr_mono);
793 tomono = tk->wall_to_monotonic;
795 } while (read_seqcount_retry(&tk_core.seq, seq));
797 ts->tv_sec += tomono.tv_sec;
799 timespec64_add_ns(ts, nsec + tomono.tv_nsec);
801 EXPORT_SYMBOL_GPL(ktime_get_ts64);
804 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
806 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
807 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
808 * works on both 32 and 64 bit systems. On 32 bit systems the readout
809 * covers ~136 years of uptime which should be enough to prevent
810 * premature wrap arounds.
812 time64_t ktime_get_seconds(void)
814 struct timekeeper *tk = &tk_core.timekeeper;
816 WARN_ON(timekeeping_suspended);
817 return tk->ktime_sec;
819 EXPORT_SYMBOL_GPL(ktime_get_seconds);
822 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
824 * Returns the wall clock seconds since 1970. This replaces the
825 * get_seconds() interface which is not y2038 safe on 32bit systems.
827 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
828 * 32bit systems the access must be protected with the sequence
829 * counter to provide "atomic" access to the 64bit tk->xtime_sec
832 time64_t ktime_get_real_seconds(void)
834 struct timekeeper *tk = &tk_core.timekeeper;
838 if (IS_ENABLED(CONFIG_64BIT))
839 return tk->xtime_sec;
842 seq = read_seqcount_begin(&tk_core.seq);
843 seconds = tk->xtime_sec;
845 } while (read_seqcount_retry(&tk_core.seq, seq));
849 EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
851 #ifdef CONFIG_NTP_PPS
854 * ktime_get_raw_and_real_ts64 - get day and raw monotonic time in timespec format
855 * @ts_raw: pointer to the timespec to be set to raw monotonic time
856 * @ts_real: pointer to the timespec to be set to the time of day
858 * This function reads both the time of day and raw monotonic time at the
859 * same time atomically and stores the resulting timestamps in timespec
862 void ktime_get_raw_and_real_ts64(struct timespec64 *ts_raw, struct timespec64 *ts_real)
864 struct timekeeper *tk = &tk_core.timekeeper;
866 s64 nsecs_raw, nsecs_real;
868 WARN_ON_ONCE(timekeeping_suspended);
871 seq = read_seqcount_begin(&tk_core.seq);
873 *ts_raw = tk->raw_time;
874 ts_real->tv_sec = tk->xtime_sec;
875 ts_real->tv_nsec = 0;
877 nsecs_raw = timekeeping_get_ns(&tk->tkr_raw);
878 nsecs_real = timekeeping_get_ns(&tk->tkr_mono);
880 } while (read_seqcount_retry(&tk_core.seq, seq));
882 timespec64_add_ns(ts_raw, nsecs_raw);
883 timespec64_add_ns(ts_real, nsecs_real);
885 EXPORT_SYMBOL(ktime_get_raw_and_real_ts64);
887 #endif /* CONFIG_NTP_PPS */
890 * do_gettimeofday - Returns the time of day in a timeval
891 * @tv: pointer to the timeval to be set
893 * NOTE: Users should be converted to using getnstimeofday()
895 void do_gettimeofday(struct timeval *tv)
897 struct timespec64 now;
899 getnstimeofday64(&now);
900 tv->tv_sec = now.tv_sec;
901 tv->tv_usec = now.tv_nsec/1000;
903 EXPORT_SYMBOL(do_gettimeofday);
906 * do_settimeofday64 - Sets the time of day.
907 * @ts: pointer to the timespec64 variable containing the new time
909 * Sets the time of day to the new time and update NTP and notify hrtimers
911 int do_settimeofday64(const struct timespec64 *ts)
913 struct timekeeper *tk = &tk_core.timekeeper;
914 struct timespec64 ts_delta, xt;
918 if (!timespec64_valid_strict(ts))
921 raw_spin_lock_irqsave(&timekeeper_lock, flags);
922 write_seqcount_begin(&tk_core.seq);
924 timekeeping_forward_now(tk);
927 ts_delta.tv_sec = ts->tv_sec - xt.tv_sec;
928 ts_delta.tv_nsec = ts->tv_nsec - xt.tv_nsec;
930 if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) {
935 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
937 tk_set_xtime(tk, ts);
939 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
941 write_seqcount_end(&tk_core.seq);
942 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
944 /* signal hrtimers about time change */
949 EXPORT_SYMBOL(do_settimeofday64);
952 * timekeeping_inject_offset - Adds or subtracts from the current time.
953 * @tv: pointer to the timespec variable containing the offset
955 * Adds or subtracts an offset value from the current time.
957 int timekeeping_inject_offset(struct timespec *ts)
959 struct timekeeper *tk = &tk_core.timekeeper;
961 struct timespec64 ts64, tmp;
964 if (!timespec_inject_offset_valid(ts))
967 ts64 = timespec_to_timespec64(*ts);
969 raw_spin_lock_irqsave(&timekeeper_lock, flags);
970 write_seqcount_begin(&tk_core.seq);
972 timekeeping_forward_now(tk);
974 /* Make sure the proposed value is valid */
975 tmp = timespec64_add(tk_xtime(tk), ts64);
976 if (timespec64_compare(&tk->wall_to_monotonic, &ts64) > 0 ||
977 !timespec64_valid_strict(&tmp)) {
982 tk_xtime_add(tk, &ts64);
983 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts64));
985 error: /* even if we error out, we forwarded the time, so call update */
986 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
988 write_seqcount_end(&tk_core.seq);
989 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
991 /* signal hrtimers about time change */
996 EXPORT_SYMBOL(timekeeping_inject_offset);
1000 * timekeeping_get_tai_offset - Returns current TAI offset from UTC
1003 s32 timekeeping_get_tai_offset(void)
1005 struct timekeeper *tk = &tk_core.timekeeper;
1010 seq = read_seqcount_begin(&tk_core.seq);
1011 ret = tk->tai_offset;
1012 } while (read_seqcount_retry(&tk_core.seq, seq));
1018 * __timekeeping_set_tai_offset - Lock free worker function
1021 static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1023 tk->tai_offset = tai_offset;
1024 tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1028 * timekeeping_set_tai_offset - Sets the current TAI offset from UTC
1031 void timekeeping_set_tai_offset(s32 tai_offset)
1033 struct timekeeper *tk = &tk_core.timekeeper;
1034 unsigned long flags;
1036 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1037 write_seqcount_begin(&tk_core.seq);
1038 __timekeeping_set_tai_offset(tk, tai_offset);
1039 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1040 write_seqcount_end(&tk_core.seq);
1041 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1046 * change_clocksource - Swaps clocksources if a new one is available
1048 * Accumulates current time interval and initializes new clocksource
1050 static int change_clocksource(void *data)
1052 struct timekeeper *tk = &tk_core.timekeeper;
1053 struct clocksource *new, *old;
1054 unsigned long flags;
1056 new = (struct clocksource *) data;
1058 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1059 write_seqcount_begin(&tk_core.seq);
1061 timekeeping_forward_now(tk);
1063 * If the cs is in module, get a module reference. Succeeds
1064 * for built-in code (owner == NULL) as well.
1066 if (try_module_get(new->owner)) {
1067 if (!new->enable || new->enable(new) == 0) {
1068 old = tk->tkr_mono.clock;
1069 tk_setup_internals(tk, new);
1072 module_put(old->owner);
1074 module_put(new->owner);
1077 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1079 write_seqcount_end(&tk_core.seq);
1080 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1086 * timekeeping_notify - Install a new clock source
1087 * @clock: pointer to the clock source
1089 * This function is called from clocksource.c after a new, better clock
1090 * source has been registered. The caller holds the clocksource_mutex.
1092 int timekeeping_notify(struct clocksource *clock)
1094 struct timekeeper *tk = &tk_core.timekeeper;
1096 if (tk->tkr_mono.clock == clock)
1098 stop_machine(change_clocksource, clock, NULL);
1099 tick_clock_notify();
1100 return tk->tkr_mono.clock == clock ? 0 : -1;
1104 * getrawmonotonic64 - Returns the raw monotonic time in a timespec
1105 * @ts: pointer to the timespec64 to be set
1107 * Returns the raw monotonic time (completely un-modified by ntp)
1109 void getrawmonotonic64(struct timespec64 *ts)
1111 struct timekeeper *tk = &tk_core.timekeeper;
1112 struct timespec64 ts64;
1117 seq = read_seqcount_begin(&tk_core.seq);
1118 nsecs = timekeeping_get_ns(&tk->tkr_raw);
1119 ts64 = tk->raw_time;
1121 } while (read_seqcount_retry(&tk_core.seq, seq));
1123 timespec64_add_ns(&ts64, nsecs);
1126 EXPORT_SYMBOL(getrawmonotonic64);
1130 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1132 int timekeeping_valid_for_hres(void)
1134 struct timekeeper *tk = &tk_core.timekeeper;
1139 seq = read_seqcount_begin(&tk_core.seq);
1141 ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1143 } while (read_seqcount_retry(&tk_core.seq, seq));
1149 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1151 u64 timekeeping_max_deferment(void)
1153 struct timekeeper *tk = &tk_core.timekeeper;
1158 seq = read_seqcount_begin(&tk_core.seq);
1160 ret = tk->tkr_mono.clock->max_idle_ns;
1162 } while (read_seqcount_retry(&tk_core.seq, seq));
1168 * read_persistent_clock - Return time from the persistent clock.
1170 * Weak dummy function for arches that do not yet support it.
1171 * Reads the time from the battery backed persistent clock.
1172 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1174 * XXX - Do be sure to remove it once all arches implement it.
1176 void __weak read_persistent_clock(struct timespec *ts)
1182 void __weak read_persistent_clock64(struct timespec64 *ts64)
1186 read_persistent_clock(&ts);
1187 *ts64 = timespec_to_timespec64(ts);
1191 * read_boot_clock64 - Return time of the system start.
1193 * Weak dummy function for arches that do not yet support it.
1194 * Function to read the exact time the system has been started.
1195 * Returns a timespec64 with tv_sec=0 and tv_nsec=0 if unsupported.
1197 * XXX - Do be sure to remove it once all arches implement it.
1199 void __weak read_boot_clock64(struct timespec64 *ts)
1205 /* Flag for if timekeeping_resume() has injected sleeptime */
1206 static bool sleeptime_injected;
1208 /* Flag for if there is a persistent clock on this platform */
1209 static bool persistent_clock_exists;
1212 * timekeeping_init - Initializes the clocksource and common timekeeping values
1214 void __init timekeeping_init(void)
1216 struct timekeeper *tk = &tk_core.timekeeper;
1217 struct clocksource *clock;
1218 unsigned long flags;
1219 struct timespec64 now, boot, tmp;
1221 read_persistent_clock64(&now);
1222 if (!timespec64_valid_strict(&now)) {
1223 pr_warn("WARNING: Persistent clock returned invalid value!\n"
1224 " Check your CMOS/BIOS settings.\n");
1227 } else if (now.tv_sec || now.tv_nsec)
1228 persistent_clock_exists = true;
1230 read_boot_clock64(&boot);
1231 if (!timespec64_valid_strict(&boot)) {
1232 pr_warn("WARNING: Boot clock returned invalid value!\n"
1233 " Check your CMOS/BIOS settings.\n");
1238 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1239 write_seqcount_begin(&tk_core.seq);
1242 clock = clocksource_default_clock();
1244 clock->enable(clock);
1245 tk_setup_internals(tk, clock);
1247 tk_set_xtime(tk, &now);
1248 tk->raw_time.tv_sec = 0;
1249 tk->raw_time.tv_nsec = 0;
1250 if (boot.tv_sec == 0 && boot.tv_nsec == 0)
1251 boot = tk_xtime(tk);
1253 set_normalized_timespec64(&tmp, -boot.tv_sec, -boot.tv_nsec);
1254 tk_set_wall_to_mono(tk, tmp);
1256 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1258 write_seqcount_end(&tk_core.seq);
1259 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1262 /* time in seconds when suspend began for persistent clock */
1263 static struct timespec64 timekeeping_suspend_time;
1266 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1267 * @delta: pointer to a timespec delta value
1269 * Takes a timespec offset measuring a suspend interval and properly
1270 * adds the sleep offset to the timekeeping variables.
1272 static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1273 struct timespec64 *delta)
1275 if (!timespec64_valid_strict(delta)) {
1276 printk_deferred(KERN_WARNING
1277 "__timekeeping_inject_sleeptime: Invalid "
1278 "sleep delta value!\n");
1281 tk_xtime_add(tk, delta);
1282 tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1283 tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1284 tk_debug_account_sleep_time(delta);
1287 #if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1289 * We have three kinds of time sources to use for sleep time
1290 * injection, the preference order is:
1291 * 1) non-stop clocksource
1292 * 2) persistent clock (ie: RTC accessible when irqs are off)
1295 * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1296 * If system has neither 1) nor 2), 3) will be used finally.
1299 * If timekeeping has injected sleeptime via either 1) or 2),
1300 * 3) becomes needless, so in this case we don't need to call
1301 * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1304 bool timekeeping_rtc_skipresume(void)
1306 return sleeptime_injected;
1310 * 1) can be determined whether to use or not only when doing
1311 * timekeeping_resume() which is invoked after rtc_suspend(),
1312 * so we can't skip rtc_suspend() surely if system has 1).
1314 * But if system has 2), 2) will definitely be used, so in this
1315 * case we don't need to call rtc_suspend(), and this is what
1316 * timekeeping_rtc_skipsuspend() means.
1318 bool timekeeping_rtc_skipsuspend(void)
1320 return persistent_clock_exists;
1324 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1325 * @delta: pointer to a timespec64 delta value
1327 * This hook is for architectures that cannot support read_persistent_clock64
1328 * because their RTC/persistent clock is only accessible when irqs are enabled.
1329 * and also don't have an effective nonstop clocksource.
1331 * This function should only be called by rtc_resume(), and allows
1332 * a suspend offset to be injected into the timekeeping values.
1334 void timekeeping_inject_sleeptime64(struct timespec64 *delta)
1336 struct timekeeper *tk = &tk_core.timekeeper;
1337 unsigned long flags;
1339 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1340 write_seqcount_begin(&tk_core.seq);
1342 timekeeping_forward_now(tk);
1344 __timekeeping_inject_sleeptime(tk, delta);
1346 timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1348 write_seqcount_end(&tk_core.seq);
1349 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1351 /* signal hrtimers about time change */
1357 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1359 void timekeeping_resume(void)
1361 struct timekeeper *tk = &tk_core.timekeeper;
1362 struct clocksource *clock = tk->tkr_mono.clock;
1363 unsigned long flags;
1364 struct timespec64 ts_new, ts_delta;
1365 cycle_t cycle_now, cycle_delta;
1367 sleeptime_injected = false;
1368 read_persistent_clock64(&ts_new);
1370 clockevents_resume();
1371 clocksource_resume();
1373 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1374 write_seqcount_begin(&tk_core.seq);
1377 * After system resumes, we need to calculate the suspended time and
1378 * compensate it for the OS time. There are 3 sources that could be
1379 * used: Nonstop clocksource during suspend, persistent clock and rtc
1382 * One specific platform may have 1 or 2 or all of them, and the
1383 * preference will be:
1384 * suspend-nonstop clocksource -> persistent clock -> rtc
1385 * The less preferred source will only be tried if there is no better
1386 * usable source. The rtc part is handled separately in rtc core code.
1388 cycle_now = tk->tkr_mono.read(clock);
1389 if ((clock->flags & CLOCK_SOURCE_SUSPEND_NONSTOP) &&
1390 cycle_now > tk->tkr_mono.cycle_last) {
1391 u64 num, max = ULLONG_MAX;
1392 u32 mult = clock->mult;
1393 u32 shift = clock->shift;
1396 cycle_delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last,
1400 * "cycle_delta * mutl" may cause 64 bits overflow, if the
1401 * suspended time is too long. In that case we need do the
1402 * 64 bits math carefully
1405 if (cycle_delta > max) {
1406 num = div64_u64(cycle_delta, max);
1407 nsec = (((u64) max * mult) >> shift) * num;
1408 cycle_delta -= num * max;
1410 nsec += ((u64) cycle_delta * mult) >> shift;
1412 ts_delta = ns_to_timespec64(nsec);
1413 sleeptime_injected = true;
1414 } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1415 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1416 sleeptime_injected = true;
1419 if (sleeptime_injected)
1420 __timekeeping_inject_sleeptime(tk, &ts_delta);
1422 /* Re-base the last cycle value */
1423 tk->tkr_mono.cycle_last = cycle_now;
1424 tk->tkr_raw.cycle_last = cycle_now;
1427 timekeeping_suspended = 0;
1428 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1429 write_seqcount_end(&tk_core.seq);
1430 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1432 touch_softlockup_watchdog();
1438 int timekeeping_suspend(void)
1440 struct timekeeper *tk = &tk_core.timekeeper;
1441 unsigned long flags;
1442 struct timespec64 delta, delta_delta;
1443 static struct timespec64 old_delta;
1445 read_persistent_clock64(&timekeeping_suspend_time);
1448 * On some systems the persistent_clock can not be detected at
1449 * timekeeping_init by its return value, so if we see a valid
1450 * value returned, update the persistent_clock_exists flag.
1452 if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1453 persistent_clock_exists = true;
1455 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1456 write_seqcount_begin(&tk_core.seq);
1457 timekeeping_forward_now(tk);
1458 timekeeping_suspended = 1;
1460 if (persistent_clock_exists) {
1462 * To avoid drift caused by repeated suspend/resumes,
1463 * which each can add ~1 second drift error,
1464 * try to compensate so the difference in system time
1465 * and persistent_clock time stays close to constant.
1467 delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1468 delta_delta = timespec64_sub(delta, old_delta);
1469 if (abs(delta_delta.tv_sec) >= 2) {
1471 * if delta_delta is too large, assume time correction
1472 * has occurred and set old_delta to the current delta.
1476 /* Otherwise try to adjust old_system to compensate */
1477 timekeeping_suspend_time =
1478 timespec64_add(timekeeping_suspend_time, delta_delta);
1482 timekeeping_update(tk, TK_MIRROR);
1483 halt_fast_timekeeper(tk);
1484 write_seqcount_end(&tk_core.seq);
1485 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1488 clocksource_suspend();
1489 clockevents_suspend();
1494 /* sysfs resume/suspend bits for timekeeping */
1495 static struct syscore_ops timekeeping_syscore_ops = {
1496 .resume = timekeeping_resume,
1497 .suspend = timekeeping_suspend,
1500 static int __init timekeeping_init_ops(void)
1502 register_syscore_ops(&timekeeping_syscore_ops);
1505 device_initcall(timekeeping_init_ops);
1508 * Apply a multiplier adjustment to the timekeeper
1510 static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1515 s64 interval = tk->cycle_interval;
1519 mult_adj = -mult_adj;
1520 interval = -interval;
1523 mult_adj <<= adj_scale;
1524 interval <<= adj_scale;
1525 offset <<= adj_scale;
1528 * So the following can be confusing.
1530 * To keep things simple, lets assume mult_adj == 1 for now.
1532 * When mult_adj != 1, remember that the interval and offset values
1533 * have been appropriately scaled so the math is the same.
1535 * The basic idea here is that we're increasing the multiplier
1536 * by one, this causes the xtime_interval to be incremented by
1537 * one cycle_interval. This is because:
1538 * xtime_interval = cycle_interval * mult
1539 * So if mult is being incremented by one:
1540 * xtime_interval = cycle_interval * (mult + 1)
1542 * xtime_interval = (cycle_interval * mult) + cycle_interval
1543 * Which can be shortened to:
1544 * xtime_interval += cycle_interval
1546 * So offset stores the non-accumulated cycles. Thus the current
1547 * time (in shifted nanoseconds) is:
1548 * now = (offset * adj) + xtime_nsec
1549 * Now, even though we're adjusting the clock frequency, we have
1550 * to keep time consistent. In other words, we can't jump back
1551 * in time, and we also want to avoid jumping forward in time.
1553 * So given the same offset value, we need the time to be the same
1554 * both before and after the freq adjustment.
1555 * now = (offset * adj_1) + xtime_nsec_1
1556 * now = (offset * adj_2) + xtime_nsec_2
1558 * (offset * adj_1) + xtime_nsec_1 =
1559 * (offset * adj_2) + xtime_nsec_2
1563 * (offset * adj_1) + xtime_nsec_1 =
1564 * (offset * (adj_1+1)) + xtime_nsec_2
1565 * (offset * adj_1) + xtime_nsec_1 =
1566 * (offset * adj_1) + offset + xtime_nsec_2
1567 * Canceling the sides:
1568 * xtime_nsec_1 = offset + xtime_nsec_2
1570 * xtime_nsec_2 = xtime_nsec_1 - offset
1571 * Which simplfies to:
1572 * xtime_nsec -= offset
1574 * XXX - TODO: Doc ntp_error calculation.
1576 if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1577 /* NTP adjustment caused clocksource mult overflow */
1582 tk->tkr_mono.mult += mult_adj;
1583 tk->xtime_interval += interval;
1584 tk->tkr_mono.xtime_nsec -= offset;
1585 tk->ntp_error -= (interval - offset) << tk->ntp_error_shift;
1589 * Calculate the multiplier adjustment needed to match the frequency
1592 static __always_inline void timekeeping_freqadjust(struct timekeeper *tk,
1595 s64 interval = tk->cycle_interval;
1596 s64 xinterval = tk->xtime_interval;
1601 /* Remove any current error adj from freq calculation */
1602 if (tk->ntp_err_mult)
1603 xinterval -= tk->cycle_interval;
1605 tk->ntp_tick = ntp_tick_length();
1607 /* Calculate current error per tick */
1608 tick_error = ntp_tick_length() >> tk->ntp_error_shift;
1609 tick_error -= (xinterval + tk->xtime_remainder);
1611 /* Don't worry about correcting it if its small */
1612 if (likely((tick_error >= 0) && (tick_error <= interval)))
1615 /* preserve the direction of correction */
1616 negative = (tick_error < 0);
1618 /* Sort out the magnitude of the correction */
1619 tick_error = abs(tick_error);
1620 for (adj = 0; tick_error > interval; adj++)
1623 /* scale the corrections */
1624 timekeeping_apply_adjustment(tk, offset, negative, adj);
1628 * Adjust the timekeeper's multiplier to the correct frequency
1629 * and also to reduce the accumulated error value.
1631 static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
1633 /* Correct for the current frequency error */
1634 timekeeping_freqadjust(tk, offset);
1636 /* Next make a small adjustment to fix any cumulative error */
1637 if (!tk->ntp_err_mult && (tk->ntp_error > 0)) {
1638 tk->ntp_err_mult = 1;
1639 timekeeping_apply_adjustment(tk, offset, 0, 0);
1640 } else if (tk->ntp_err_mult && (tk->ntp_error <= 0)) {
1641 /* Undo any existing error adjustment */
1642 timekeeping_apply_adjustment(tk, offset, 1, 0);
1643 tk->ntp_err_mult = 0;
1646 if (unlikely(tk->tkr_mono.clock->maxadj &&
1647 (abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
1648 > tk->tkr_mono.clock->maxadj))) {
1649 printk_once(KERN_WARNING
1650 "Adjusting %s more than 11%% (%ld vs %ld)\n",
1651 tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
1652 (long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
1656 * It may be possible that when we entered this function, xtime_nsec
1657 * was very small. Further, if we're slightly speeding the clocksource
1658 * in the code above, its possible the required corrective factor to
1659 * xtime_nsec could cause it to underflow.
1661 * Now, since we already accumulated the second, cannot simply roll
1662 * the accumulated second back, since the NTP subsystem has been
1663 * notified via second_overflow. So instead we push xtime_nsec forward
1664 * by the amount we underflowed, and add that amount into the error.
1666 * We'll correct this error next time through this function, when
1667 * xtime_nsec is not as small.
1669 if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
1670 s64 neg = -(s64)tk->tkr_mono.xtime_nsec;
1671 tk->tkr_mono.xtime_nsec = 0;
1672 tk->ntp_error += neg << tk->ntp_error_shift;
1677 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1679 * Helper function that accumulates the nsecs greater than a second
1680 * from the xtime_nsec field to the xtime_secs field.
1681 * It also calls into the NTP code to handle leapsecond processing.
1684 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
1686 u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
1687 unsigned int clock_set = 0;
1689 while (tk->tkr_mono.xtime_nsec >= nsecps) {
1692 tk->tkr_mono.xtime_nsec -= nsecps;
1695 /* Figure out if its a leap sec and apply if needed */
1696 leap = second_overflow(tk->xtime_sec);
1697 if (unlikely(leap)) {
1698 struct timespec64 ts;
1700 tk->xtime_sec += leap;
1704 tk_set_wall_to_mono(tk,
1705 timespec64_sub(tk->wall_to_monotonic, ts));
1707 __timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
1709 clock_set = TK_CLOCK_WAS_SET;
1716 * logarithmic_accumulation - shifted accumulation of cycles
1718 * This functions accumulates a shifted interval of cycles into
1719 * into a shifted interval nanoseconds. Allows for O(log) accumulation
1722 * Returns the unconsumed cycles.
1724 static cycle_t logarithmic_accumulation(struct timekeeper *tk, cycle_t offset,
1726 unsigned int *clock_set)
1728 cycle_t interval = tk->cycle_interval << shift;
1731 /* If the offset is smaller than a shifted interval, do nothing */
1732 if (offset < interval)
1735 /* Accumulate one shifted interval */
1737 tk->tkr_mono.cycle_last += interval;
1738 tk->tkr_raw.cycle_last += interval;
1740 tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
1741 *clock_set |= accumulate_nsecs_to_secs(tk);
1743 /* Accumulate raw time */
1744 raw_nsecs = (u64)tk->raw_interval << shift;
1745 raw_nsecs += tk->raw_time.tv_nsec;
1746 if (raw_nsecs >= NSEC_PER_SEC) {
1747 u64 raw_secs = raw_nsecs;
1748 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
1749 tk->raw_time.tv_sec += raw_secs;
1751 tk->raw_time.tv_nsec = raw_nsecs;
1753 /* Accumulate error between NTP and clock interval */
1754 tk->ntp_error += tk->ntp_tick << shift;
1755 tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
1756 (tk->ntp_error_shift + shift);
1762 * update_wall_time - Uses the current clocksource to increment the wall time
1765 void update_wall_time(void)
1767 struct timekeeper *real_tk = &tk_core.timekeeper;
1768 struct timekeeper *tk = &shadow_timekeeper;
1770 int shift = 0, maxshift;
1771 unsigned int clock_set = 0;
1772 unsigned long flags;
1774 raw_spin_lock_irqsave(&timekeeper_lock, flags);
1776 /* Make sure we're fully resumed: */
1777 if (unlikely(timekeeping_suspended))
1780 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
1781 offset = real_tk->cycle_interval;
1783 offset = clocksource_delta(tk->tkr_mono.read(tk->tkr_mono.clock),
1784 tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
1787 /* Check if there's really nothing to do */
1788 if (offset < real_tk->cycle_interval)
1791 /* Do some additional sanity checking */
1792 timekeeping_check_update(real_tk, offset);
1795 * With NO_HZ we may have to accumulate many cycle_intervals
1796 * (think "ticks") worth of time at once. To do this efficiently,
1797 * we calculate the largest doubling multiple of cycle_intervals
1798 * that is smaller than the offset. We then accumulate that
1799 * chunk in one go, and then try to consume the next smaller
1802 shift = ilog2(offset) - ilog2(tk->cycle_interval);
1803 shift = max(0, shift);
1804 /* Bound shift to one less than what overflows tick_length */
1805 maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
1806 shift = min(shift, maxshift);
1807 while (offset >= tk->cycle_interval) {
1808 offset = logarithmic_accumulation(tk, offset, shift,
1810 if (offset < tk->cycle_interval<<shift)
1814 /* correct the clock when NTP error is too big */
1815 timekeeping_adjust(tk, offset);
1818 * XXX This can be killed once everyone converts
1819 * to the new update_vsyscall.
1821 old_vsyscall_fixup(tk);
1824 * Finally, make sure that after the rounding
1825 * xtime_nsec isn't larger than NSEC_PER_SEC
1827 clock_set |= accumulate_nsecs_to_secs(tk);
1829 write_seqcount_begin(&tk_core.seq);
1831 * Update the real timekeeper.
1833 * We could avoid this memcpy by switching pointers, but that
1834 * requires changes to all other timekeeper usage sites as
1835 * well, i.e. move the timekeeper pointer getter into the
1836 * spinlocked/seqcount protected sections. And we trade this
1837 * memcpy under the tk_core.seq against one before we start
1840 timekeeping_update(tk, clock_set);
1841 memcpy(real_tk, tk, sizeof(*tk));
1842 /* The memcpy must come last. Do not put anything here! */
1843 write_seqcount_end(&tk_core.seq);
1845 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1847 /* Have to call _delayed version, since in irq context*/
1848 clock_was_set_delayed();
1852 * getboottime64 - Return the real time of system boot.
1853 * @ts: pointer to the timespec64 to be set
1855 * Returns the wall-time of boot in a timespec64.
1857 * This is based on the wall_to_monotonic offset and the total suspend
1858 * time. Calls to settimeofday will affect the value returned (which
1859 * basically means that however wrong your real time clock is at boot time,
1860 * you get the right time here).
1862 void getboottime64(struct timespec64 *ts)
1864 struct timekeeper *tk = &tk_core.timekeeper;
1865 ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
1867 *ts = ktime_to_timespec64(t);
1869 EXPORT_SYMBOL_GPL(getboottime64);
1871 unsigned long get_seconds(void)
1873 struct timekeeper *tk = &tk_core.timekeeper;
1875 return tk->xtime_sec;
1877 EXPORT_SYMBOL(get_seconds);
1879 struct timespec __current_kernel_time(void)
1881 struct timekeeper *tk = &tk_core.timekeeper;
1883 return timespec64_to_timespec(tk_xtime(tk));
1886 struct timespec64 current_kernel_time64(void)
1888 struct timekeeper *tk = &tk_core.timekeeper;
1889 struct timespec64 now;
1893 seq = read_seqcount_begin(&tk_core.seq);
1896 } while (read_seqcount_retry(&tk_core.seq, seq));
1900 EXPORT_SYMBOL(current_kernel_time64);
1902 struct timespec64 get_monotonic_coarse64(void)
1904 struct timekeeper *tk = &tk_core.timekeeper;
1905 struct timespec64 now, mono;
1909 seq = read_seqcount_begin(&tk_core.seq);
1912 mono = tk->wall_to_monotonic;
1913 } while (read_seqcount_retry(&tk_core.seq, seq));
1915 set_normalized_timespec64(&now, now.tv_sec + mono.tv_sec,
1916 now.tv_nsec + mono.tv_nsec);
1922 * Must hold jiffies_lock
1924 void do_timer(unsigned long ticks)
1926 jiffies_64 += ticks;
1927 calc_global_load(ticks);
1931 * ktime_get_update_offsets_now - hrtimer helper
1932 * @cwsseq: pointer to check and store the clock was set sequence number
1933 * @offs_real: pointer to storage for monotonic -> realtime offset
1934 * @offs_boot: pointer to storage for monotonic -> boottime offset
1935 * @offs_tai: pointer to storage for monotonic -> clock tai offset
1937 * Returns current monotonic time and updates the offsets if the
1938 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
1941 * Called from hrtimer_interrupt() or retrigger_next_event()
1943 ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
1944 ktime_t *offs_boot, ktime_t *offs_tai)
1946 struct timekeeper *tk = &tk_core.timekeeper;
1952 seq = read_seqcount_begin(&tk_core.seq);
1954 base = tk->tkr_mono.base;
1955 nsecs = timekeeping_get_ns(&tk->tkr_mono);
1956 base = ktime_add_ns(base, nsecs);
1958 if (*cwsseq != tk->clock_was_set_seq) {
1959 *cwsseq = tk->clock_was_set_seq;
1960 *offs_real = tk->offs_real;
1961 *offs_boot = tk->offs_boot;
1962 *offs_tai = tk->offs_tai;
1965 /* Handle leapsecond insertion adjustments */
1966 if (unlikely(base.tv64 >= tk->next_leap_ktime.tv64))
1967 *offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
1969 } while (read_seqcount_retry(&tk_core.seq, seq));
1975 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
1977 int do_adjtimex(struct timex *txc)
1979 struct timekeeper *tk = &tk_core.timekeeper;
1980 unsigned long flags;
1981 struct timespec64 ts;
1985 /* Validate the data before disabling interrupts */
1986 ret = ntp_validate_timex(txc);
1990 if (txc->modes & ADJ_SETOFFSET) {
1991 struct timespec delta;
1992 delta.tv_sec = txc->time.tv_sec;
1993 delta.tv_nsec = txc->time.tv_usec;
1994 if (!(txc->modes & ADJ_NANO))
1995 delta.tv_nsec *= 1000;
1996 ret = timekeeping_inject_offset(&delta);
2001 getnstimeofday64(&ts);
2003 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2004 write_seqcount_begin(&tk_core.seq);
2006 orig_tai = tai = tk->tai_offset;
2007 ret = __do_adjtimex(txc, &ts, &tai);
2009 if (tai != orig_tai) {
2010 __timekeeping_set_tai_offset(tk, tai);
2011 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
2013 tk_update_leap_state(tk);
2015 write_seqcount_end(&tk_core.seq);
2016 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2018 if (tai != orig_tai)
2021 ntp_notify_cmos_timer();
2026 #ifdef CONFIG_NTP_PPS
2028 * hardpps() - Accessor function to NTP __hardpps function
2030 void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2032 unsigned long flags;
2034 raw_spin_lock_irqsave(&timekeeper_lock, flags);
2035 write_seqcount_begin(&tk_core.seq);
2037 __hardpps(phase_ts, raw_ts);
2039 write_seqcount_end(&tk_core.seq);
2040 raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2042 EXPORT_SYMBOL(hardpps);
2046 * xtime_update() - advances the timekeeping infrastructure
2047 * @ticks: number of ticks, that have elapsed since the last call.
2049 * Must be called with interrupts disabled.
2051 void xtime_update(unsigned long ticks)
2053 write_seqlock(&jiffies_lock);
2055 write_sequnlock(&jiffies_lock);