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
;
311 /* If arch requires, add in get_arch_timeoffset() */
312 return nsec
+ arch_gettimeoffset();
316 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
317 * @tkr: Timekeeping readout base from which we take the update
319 * We want to use this from any context including NMI and tracing /
320 * instrumenting the timekeeping code itself.
322 * So we handle this differently than the other timekeeping accessor
323 * functions which retry when the sequence count has changed. The
326 * smp_wmb(); <- Ensure that the last base[1] update is visible
328 * smp_wmb(); <- Ensure that the seqcount update is visible
329 * update(tkf->base[0], tkr);
330 * smp_wmb(); <- Ensure that the base[0] update is visible
332 * smp_wmb(); <- Ensure that the seqcount update is visible
333 * update(tkf->base[1], tkr);
335 * The reader side does:
341 * now = now(tkf->base[idx]);
343 * } while (seq != tkf->seq)
345 * As long as we update base[0] readers are forced off to
346 * base[1]. Once base[0] is updated readers are redirected to base[0]
347 * and the base[1] update takes place.
349 * So if a NMI hits the update of base[0] then it will use base[1]
350 * which is still consistent. In the worst case this can result is a
351 * slightly wrong timestamp (a few nanoseconds). See
352 * @ktime_get_mono_fast_ns.
354 static void update_fast_timekeeper(struct tk_read_base
*tkr
, struct tk_fast
*tkf
)
356 struct tk_read_base
*base
= tkf
->base
;
358 /* Force readers off to base[1] */
359 raw_write_seqcount_latch(&tkf
->seq
);
362 memcpy(base
, tkr
, sizeof(*base
));
364 /* Force readers back to base[0] */
365 raw_write_seqcount_latch(&tkf
->seq
);
368 memcpy(base
+ 1, base
, sizeof(*base
));
372 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
374 * This timestamp is not guaranteed to be monotonic across an update.
375 * The timestamp is calculated by:
377 * now = base_mono + clock_delta * slope
379 * So if the update lowers the slope, readers who are forced to the
380 * not yet updated second array are still using the old steeper slope.
389 * |12345678---> reader order
395 * So reader 6 will observe time going backwards versus reader 5.
397 * While other CPUs are likely to be able observe that, the only way
398 * for a CPU local observation is when an NMI hits in the middle of
399 * the update. Timestamps taken from that NMI context might be ahead
400 * of the following timestamps. Callers need to be aware of that and
403 static __always_inline u64
__ktime_get_fast_ns(struct tk_fast
*tkf
)
405 struct tk_read_base
*tkr
;
410 seq
= raw_read_seqcount(&tkf
->seq
);
411 tkr
= tkf
->base
+ (seq
& 0x01);
412 now
= ktime_to_ns(tkr
->base
) + timekeeping_get_ns(tkr
);
413 } while (read_seqcount_retry(&tkf
->seq
, seq
));
418 u64
ktime_get_mono_fast_ns(void)
420 return __ktime_get_fast_ns(&tk_fast_mono
);
422 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns
);
424 u64
ktime_get_raw_fast_ns(void)
426 return __ktime_get_fast_ns(&tk_fast_raw
);
428 EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns
);
430 /* Suspend-time cycles value for halted fast timekeeper. */
431 static cycle_t cycles_at_suspend
;
433 static cycle_t
dummy_clock_read(struct clocksource
*cs
)
435 return cycles_at_suspend
;
439 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
440 * @tk: Timekeeper to snapshot.
442 * It generally is unsafe to access the clocksource after timekeeping has been
443 * suspended, so take a snapshot of the readout base of @tk and use it as the
444 * fast timekeeper's readout base while suspended. It will return the same
445 * number of cycles every time until timekeeping is resumed at which time the
446 * proper readout base for the fast timekeeper will be restored automatically.
448 static void halt_fast_timekeeper(struct timekeeper
*tk
)
450 static struct tk_read_base tkr_dummy
;
451 struct tk_read_base
*tkr
= &tk
->tkr_mono
;
453 memcpy(&tkr_dummy
, tkr
, sizeof(tkr_dummy
));
454 cycles_at_suspend
= tkr
->read(tkr
->clock
);
455 tkr_dummy
.read
= dummy_clock_read
;
456 update_fast_timekeeper(&tkr_dummy
, &tk_fast_mono
);
459 memcpy(&tkr_dummy
, tkr
, sizeof(tkr_dummy
));
460 tkr_dummy
.read
= dummy_clock_read
;
461 update_fast_timekeeper(&tkr_dummy
, &tk_fast_raw
);
464 #ifdef CONFIG_GENERIC_TIME_VSYSCALL_OLD
466 static inline void update_vsyscall(struct timekeeper
*tk
)
468 struct timespec xt
, wm
;
470 xt
= timespec64_to_timespec(tk_xtime(tk
));
471 wm
= timespec64_to_timespec(tk
->wall_to_monotonic
);
472 update_vsyscall_old(&xt
, &wm
, tk
->tkr_mono
.clock
, tk
->tkr_mono
.mult
,
473 tk
->tkr_mono
.cycle_last
);
476 static inline void old_vsyscall_fixup(struct timekeeper
*tk
)
481 * Store only full nanoseconds into xtime_nsec after rounding
482 * it up and add the remainder to the error difference.
483 * XXX - This is necessary to avoid small 1ns inconsistnecies caused
484 * by truncating the remainder in vsyscalls. However, it causes
485 * additional work to be done in timekeeping_adjust(). Once
486 * the vsyscall implementations are converted to use xtime_nsec
487 * (shifted nanoseconds), and CONFIG_GENERIC_TIME_VSYSCALL_OLD
488 * users are removed, this can be killed.
490 remainder
= tk
->tkr_mono
.xtime_nsec
& ((1ULL << tk
->tkr_mono
.shift
) - 1);
491 tk
->tkr_mono
.xtime_nsec
-= remainder
;
492 tk
->tkr_mono
.xtime_nsec
+= 1ULL << tk
->tkr_mono
.shift
;
493 tk
->ntp_error
+= remainder
<< tk
->ntp_error_shift
;
494 tk
->ntp_error
-= (1ULL << tk
->tkr_mono
.shift
) << tk
->ntp_error_shift
;
497 #define old_vsyscall_fixup(tk)
500 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain
);
502 static void update_pvclock_gtod(struct timekeeper
*tk
, bool was_set
)
504 raw_notifier_call_chain(&pvclock_gtod_chain
, was_set
, tk
);
508 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
510 int pvclock_gtod_register_notifier(struct notifier_block
*nb
)
512 struct timekeeper
*tk
= &tk_core
.timekeeper
;
516 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
517 ret
= raw_notifier_chain_register(&pvclock_gtod_chain
, nb
);
518 update_pvclock_gtod(tk
, true);
519 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
523 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier
);
526 * pvclock_gtod_unregister_notifier - unregister a pvclock
527 * timedata update listener
529 int pvclock_gtod_unregister_notifier(struct notifier_block
*nb
)
534 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
535 ret
= raw_notifier_chain_unregister(&pvclock_gtod_chain
, nb
);
536 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
540 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier
);
543 * tk_update_leap_state - helper to update the next_leap_ktime
545 static inline void tk_update_leap_state(struct timekeeper
*tk
)
547 tk
->next_leap_ktime
= ntp_get_next_leap();
548 if (tk
->next_leap_ktime
.tv64
!= KTIME_MAX
)
549 /* Convert to monotonic time */
550 tk
->next_leap_ktime
= ktime_sub(tk
->next_leap_ktime
, tk
->offs_real
);
554 * Update the ktime_t based scalar nsec members of the timekeeper
556 static inline void tk_update_ktime_data(struct timekeeper
*tk
)
562 * The xtime based monotonic readout is:
563 * nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
564 * The ktime based monotonic readout is:
565 * nsec = base_mono + now();
566 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
568 seconds
= (u64
)(tk
->xtime_sec
+ tk
->wall_to_monotonic
.tv_sec
);
569 nsec
= (u32
) tk
->wall_to_monotonic
.tv_nsec
;
570 tk
->tkr_mono
.base
= ns_to_ktime(seconds
* NSEC_PER_SEC
+ nsec
);
572 /* Update the monotonic raw base */
573 tk
->tkr_raw
.base
= timespec64_to_ktime(tk
->raw_time
);
576 * The sum of the nanoseconds portions of xtime and
577 * wall_to_monotonic can be greater/equal one second. Take
578 * this into account before updating tk->ktime_sec.
580 nsec
+= (u32
)(tk
->tkr_mono
.xtime_nsec
>> tk
->tkr_mono
.shift
);
581 if (nsec
>= NSEC_PER_SEC
)
583 tk
->ktime_sec
= seconds
;
586 /* must hold timekeeper_lock */
587 static void timekeeping_update(struct timekeeper
*tk
, unsigned int action
)
589 if (action
& TK_CLEAR_NTP
) {
594 tk_update_leap_state(tk
);
595 tk_update_ktime_data(tk
);
598 update_pvclock_gtod(tk
, action
& TK_CLOCK_WAS_SET
);
600 update_fast_timekeeper(&tk
->tkr_mono
, &tk_fast_mono
);
601 update_fast_timekeeper(&tk
->tkr_raw
, &tk_fast_raw
);
603 if (action
& TK_CLOCK_WAS_SET
)
604 tk
->clock_was_set_seq
++;
606 * The mirroring of the data to the shadow-timekeeper needs
607 * to happen last here to ensure we don't over-write the
608 * timekeeper structure on the next update with stale data
610 if (action
& TK_MIRROR
)
611 memcpy(&shadow_timekeeper
, &tk_core
.timekeeper
,
612 sizeof(tk_core
.timekeeper
));
616 * timekeeping_forward_now - update clock to the current time
618 * Forward the current clock to update its state since the last call to
619 * update_wall_time(). This is useful before significant clock changes,
620 * as it avoids having to deal with this time offset explicitly.
622 static void timekeeping_forward_now(struct timekeeper
*tk
)
624 struct clocksource
*clock
= tk
->tkr_mono
.clock
;
625 cycle_t cycle_now
, delta
;
628 cycle_now
= tk
->tkr_mono
.read(clock
);
629 delta
= clocksource_delta(cycle_now
, tk
->tkr_mono
.cycle_last
, tk
->tkr_mono
.mask
);
630 tk
->tkr_mono
.cycle_last
= cycle_now
;
631 tk
->tkr_raw
.cycle_last
= cycle_now
;
633 tk
->tkr_mono
.xtime_nsec
+= delta
* tk
->tkr_mono
.mult
;
635 /* If arch requires, add in get_arch_timeoffset() */
636 tk
->tkr_mono
.xtime_nsec
+= (u64
)arch_gettimeoffset() << tk
->tkr_mono
.shift
;
638 tk_normalize_xtime(tk
);
640 nsec
= clocksource_cyc2ns(delta
, tk
->tkr_raw
.mult
, tk
->tkr_raw
.shift
);
641 timespec64_add_ns(&tk
->raw_time
, nsec
);
645 * __getnstimeofday64 - Returns the time of day in a timespec64.
646 * @ts: pointer to the timespec to be set
648 * Updates the time of day in the timespec.
649 * Returns 0 on success, or -ve when suspended (timespec will be undefined).
651 int __getnstimeofday64(struct timespec64
*ts
)
653 struct timekeeper
*tk
= &tk_core
.timekeeper
;
658 seq
= read_seqcount_begin(&tk_core
.seq
);
660 ts
->tv_sec
= tk
->xtime_sec
;
661 nsecs
= timekeeping_get_ns(&tk
->tkr_mono
);
663 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
666 timespec64_add_ns(ts
, nsecs
);
669 * Do not bail out early, in case there were callers still using
670 * the value, even in the face of the WARN_ON.
672 if (unlikely(timekeeping_suspended
))
676 EXPORT_SYMBOL(__getnstimeofday64
);
679 * getnstimeofday64 - Returns the time of day in a timespec64.
680 * @ts: pointer to the timespec64 to be set
682 * Returns the time of day in a timespec64 (WARN if suspended).
684 void getnstimeofday64(struct timespec64
*ts
)
686 WARN_ON(__getnstimeofday64(ts
));
688 EXPORT_SYMBOL(getnstimeofday64
);
690 ktime_t
ktime_get(void)
692 struct timekeeper
*tk
= &tk_core
.timekeeper
;
697 WARN_ON(timekeeping_suspended
);
700 seq
= read_seqcount_begin(&tk_core
.seq
);
701 base
= tk
->tkr_mono
.base
;
702 nsecs
= timekeeping_get_ns(&tk
->tkr_mono
);
704 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
706 return ktime_add_ns(base
, nsecs
);
708 EXPORT_SYMBOL_GPL(ktime_get
);
710 u32
ktime_get_resolution_ns(void)
712 struct timekeeper
*tk
= &tk_core
.timekeeper
;
716 WARN_ON(timekeeping_suspended
);
719 seq
= read_seqcount_begin(&tk_core
.seq
);
720 nsecs
= tk
->tkr_mono
.mult
>> tk
->tkr_mono
.shift
;
721 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
725 EXPORT_SYMBOL_GPL(ktime_get_resolution_ns
);
727 static ktime_t
*offsets
[TK_OFFS_MAX
] = {
728 [TK_OFFS_REAL
] = &tk_core
.timekeeper
.offs_real
,
729 [TK_OFFS_BOOT
] = &tk_core
.timekeeper
.offs_boot
,
730 [TK_OFFS_TAI
] = &tk_core
.timekeeper
.offs_tai
,
733 ktime_t
ktime_get_with_offset(enum tk_offsets offs
)
735 struct timekeeper
*tk
= &tk_core
.timekeeper
;
737 ktime_t base
, *offset
= offsets
[offs
];
740 WARN_ON(timekeeping_suspended
);
743 seq
= read_seqcount_begin(&tk_core
.seq
);
744 base
= ktime_add(tk
->tkr_mono
.base
, *offset
);
745 nsecs
= timekeeping_get_ns(&tk
->tkr_mono
);
747 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
749 return ktime_add_ns(base
, nsecs
);
752 EXPORT_SYMBOL_GPL(ktime_get_with_offset
);
755 * ktime_mono_to_any() - convert mononotic time to any other time
756 * @tmono: time to convert.
757 * @offs: which offset to use
759 ktime_t
ktime_mono_to_any(ktime_t tmono
, enum tk_offsets offs
)
761 ktime_t
*offset
= offsets
[offs
];
766 seq
= read_seqcount_begin(&tk_core
.seq
);
767 tconv
= ktime_add(tmono
, *offset
);
768 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
772 EXPORT_SYMBOL_GPL(ktime_mono_to_any
);
775 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
777 ktime_t
ktime_get_raw(void)
779 struct timekeeper
*tk
= &tk_core
.timekeeper
;
785 seq
= read_seqcount_begin(&tk_core
.seq
);
786 base
= tk
->tkr_raw
.base
;
787 nsecs
= timekeeping_get_ns(&tk
->tkr_raw
);
789 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
791 return ktime_add_ns(base
, nsecs
);
793 EXPORT_SYMBOL_GPL(ktime_get_raw
);
796 * ktime_get_ts64 - get the monotonic clock in timespec64 format
797 * @ts: pointer to timespec variable
799 * The function calculates the monotonic clock from the realtime
800 * clock and the wall_to_monotonic offset and stores the result
801 * in normalized timespec64 format in the variable pointed to by @ts.
803 void ktime_get_ts64(struct timespec64
*ts
)
805 struct timekeeper
*tk
= &tk_core
.timekeeper
;
806 struct timespec64 tomono
;
810 WARN_ON(timekeeping_suspended
);
813 seq
= read_seqcount_begin(&tk_core
.seq
);
814 ts
->tv_sec
= tk
->xtime_sec
;
815 nsec
= timekeeping_get_ns(&tk
->tkr_mono
);
816 tomono
= tk
->wall_to_monotonic
;
818 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
820 ts
->tv_sec
+= tomono
.tv_sec
;
822 timespec64_add_ns(ts
, nsec
+ tomono
.tv_nsec
);
824 EXPORT_SYMBOL_GPL(ktime_get_ts64
);
827 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
829 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
830 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
831 * works on both 32 and 64 bit systems. On 32 bit systems the readout
832 * covers ~136 years of uptime which should be enough to prevent
833 * premature wrap arounds.
835 time64_t
ktime_get_seconds(void)
837 struct timekeeper
*tk
= &tk_core
.timekeeper
;
839 WARN_ON(timekeeping_suspended
);
840 return tk
->ktime_sec
;
842 EXPORT_SYMBOL_GPL(ktime_get_seconds
);
845 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
847 * Returns the wall clock seconds since 1970. This replaces the
848 * get_seconds() interface which is not y2038 safe on 32bit systems.
850 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
851 * 32bit systems the access must be protected with the sequence
852 * counter to provide "atomic" access to the 64bit tk->xtime_sec
855 time64_t
ktime_get_real_seconds(void)
857 struct timekeeper
*tk
= &tk_core
.timekeeper
;
861 if (IS_ENABLED(CONFIG_64BIT
))
862 return tk
->xtime_sec
;
865 seq
= read_seqcount_begin(&tk_core
.seq
);
866 seconds
= tk
->xtime_sec
;
868 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
872 EXPORT_SYMBOL_GPL(ktime_get_real_seconds
);
874 #ifdef CONFIG_NTP_PPS
877 * getnstime_raw_and_real - get day and raw monotonic time in timespec format
878 * @ts_raw: pointer to the timespec to be set to raw monotonic time
879 * @ts_real: pointer to the timespec to be set to the time of day
881 * This function reads both the time of day and raw monotonic time at the
882 * same time atomically and stores the resulting timestamps in timespec
885 void getnstime_raw_and_real(struct timespec
*ts_raw
, struct timespec
*ts_real
)
887 struct timekeeper
*tk
= &tk_core
.timekeeper
;
889 s64 nsecs_raw
, nsecs_real
;
891 WARN_ON_ONCE(timekeeping_suspended
);
894 seq
= read_seqcount_begin(&tk_core
.seq
);
896 *ts_raw
= timespec64_to_timespec(tk
->raw_time
);
897 ts_real
->tv_sec
= tk
->xtime_sec
;
898 ts_real
->tv_nsec
= 0;
900 nsecs_raw
= timekeeping_get_ns(&tk
->tkr_raw
);
901 nsecs_real
= timekeeping_get_ns(&tk
->tkr_mono
);
903 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
905 timespec_add_ns(ts_raw
, nsecs_raw
);
906 timespec_add_ns(ts_real
, nsecs_real
);
908 EXPORT_SYMBOL(getnstime_raw_and_real
);
910 #endif /* CONFIG_NTP_PPS */
913 * do_gettimeofday - Returns the time of day in a timeval
914 * @tv: pointer to the timeval to be set
916 * NOTE: Users should be converted to using getnstimeofday()
918 void do_gettimeofday(struct timeval
*tv
)
920 struct timespec64 now
;
922 getnstimeofday64(&now
);
923 tv
->tv_sec
= now
.tv_sec
;
924 tv
->tv_usec
= now
.tv_nsec
/1000;
926 EXPORT_SYMBOL(do_gettimeofday
);
929 * do_settimeofday64 - Sets the time of day.
930 * @ts: pointer to the timespec64 variable containing the new time
932 * Sets the time of day to the new time and update NTP and notify hrtimers
934 int do_settimeofday64(const struct timespec64
*ts
)
936 struct timekeeper
*tk
= &tk_core
.timekeeper
;
937 struct timespec64 ts_delta
, xt
;
940 if (!timespec64_valid_strict(ts
))
943 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
944 write_seqcount_begin(&tk_core
.seq
);
946 timekeeping_forward_now(tk
);
949 ts_delta
.tv_sec
= ts
->tv_sec
- xt
.tv_sec
;
950 ts_delta
.tv_nsec
= ts
->tv_nsec
- xt
.tv_nsec
;
952 tk_set_wall_to_mono(tk
, timespec64_sub(tk
->wall_to_monotonic
, ts_delta
));
954 tk_set_xtime(tk
, ts
);
956 timekeeping_update(tk
, TK_CLEAR_NTP
| TK_MIRROR
| TK_CLOCK_WAS_SET
);
958 write_seqcount_end(&tk_core
.seq
);
959 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
961 /* signal hrtimers about time change */
966 EXPORT_SYMBOL(do_settimeofday64
);
969 * timekeeping_inject_offset - Adds or subtracts from the current time.
970 * @tv: pointer to the timespec variable containing the offset
972 * Adds or subtracts an offset value from the current time.
974 int timekeeping_inject_offset(struct timespec
*ts
)
976 struct timekeeper
*tk
= &tk_core
.timekeeper
;
978 struct timespec64 ts64
, tmp
;
981 if ((unsigned long)ts
->tv_nsec
>= NSEC_PER_SEC
)
984 ts64
= timespec_to_timespec64(*ts
);
986 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
987 write_seqcount_begin(&tk_core
.seq
);
989 timekeeping_forward_now(tk
);
991 /* Make sure the proposed value is valid */
992 tmp
= timespec64_add(tk_xtime(tk
), ts64
);
993 if (!timespec64_valid_strict(&tmp
)) {
998 tk_xtime_add(tk
, &ts64
);
999 tk_set_wall_to_mono(tk
, timespec64_sub(tk
->wall_to_monotonic
, ts64
));
1001 error
: /* even if we error out, we forwarded the time, so call update */
1002 timekeeping_update(tk
, TK_CLEAR_NTP
| TK_MIRROR
| TK_CLOCK_WAS_SET
);
1004 write_seqcount_end(&tk_core
.seq
);
1005 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1007 /* signal hrtimers about time change */
1012 EXPORT_SYMBOL(timekeeping_inject_offset
);
1016 * timekeeping_get_tai_offset - Returns current TAI offset from UTC
1019 s32
timekeeping_get_tai_offset(void)
1021 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1026 seq
= read_seqcount_begin(&tk_core
.seq
);
1027 ret
= tk
->tai_offset
;
1028 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
1034 * __timekeeping_set_tai_offset - Lock free worker function
1037 static void __timekeeping_set_tai_offset(struct timekeeper
*tk
, s32 tai_offset
)
1039 tk
->tai_offset
= tai_offset
;
1040 tk
->offs_tai
= ktime_add(tk
->offs_real
, ktime_set(tai_offset
, 0));
1044 * timekeeping_set_tai_offset - Sets the current TAI offset from UTC
1047 void timekeeping_set_tai_offset(s32 tai_offset
)
1049 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1050 unsigned long flags
;
1052 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
1053 write_seqcount_begin(&tk_core
.seq
);
1054 __timekeeping_set_tai_offset(tk
, tai_offset
);
1055 timekeeping_update(tk
, TK_MIRROR
| TK_CLOCK_WAS_SET
);
1056 write_seqcount_end(&tk_core
.seq
);
1057 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1062 * change_clocksource - Swaps clocksources if a new one is available
1064 * Accumulates current time interval and initializes new clocksource
1066 static int change_clocksource(void *data
)
1068 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1069 struct clocksource
*new, *old
;
1070 unsigned long flags
;
1072 new = (struct clocksource
*) data
;
1074 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
1075 write_seqcount_begin(&tk_core
.seq
);
1077 timekeeping_forward_now(tk
);
1079 * If the cs is in module, get a module reference. Succeeds
1080 * for built-in code (owner == NULL) as well.
1082 if (try_module_get(new->owner
)) {
1083 if (!new->enable
|| new->enable(new) == 0) {
1084 old
= tk
->tkr_mono
.clock
;
1085 tk_setup_internals(tk
, new);
1088 module_put(old
->owner
);
1090 module_put(new->owner
);
1093 timekeeping_update(tk
, TK_CLEAR_NTP
| TK_MIRROR
| TK_CLOCK_WAS_SET
);
1095 write_seqcount_end(&tk_core
.seq
);
1096 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1102 * timekeeping_notify - Install a new clock source
1103 * @clock: pointer to the clock source
1105 * This function is called from clocksource.c after a new, better clock
1106 * source has been registered. The caller holds the clocksource_mutex.
1108 int timekeeping_notify(struct clocksource
*clock
)
1110 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1112 if (tk
->tkr_mono
.clock
== clock
)
1114 stop_machine(change_clocksource
, clock
, NULL
);
1115 tick_clock_notify();
1116 return tk
->tkr_mono
.clock
== clock
? 0 : -1;
1120 * getrawmonotonic64 - Returns the raw monotonic time in a timespec
1121 * @ts: pointer to the timespec64 to be set
1123 * Returns the raw monotonic time (completely un-modified by ntp)
1125 void getrawmonotonic64(struct timespec64
*ts
)
1127 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1128 struct timespec64 ts64
;
1133 seq
= read_seqcount_begin(&tk_core
.seq
);
1134 nsecs
= timekeeping_get_ns(&tk
->tkr_raw
);
1135 ts64
= tk
->raw_time
;
1137 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
1139 timespec64_add_ns(&ts64
, nsecs
);
1142 EXPORT_SYMBOL(getrawmonotonic64
);
1146 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1148 int timekeeping_valid_for_hres(void)
1150 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1155 seq
= read_seqcount_begin(&tk_core
.seq
);
1157 ret
= tk
->tkr_mono
.clock
->flags
& CLOCK_SOURCE_VALID_FOR_HRES
;
1159 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
1165 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1167 u64
timekeeping_max_deferment(void)
1169 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1174 seq
= read_seqcount_begin(&tk_core
.seq
);
1176 ret
= tk
->tkr_mono
.clock
->max_idle_ns
;
1178 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
1184 * read_persistent_clock - Return time from the persistent clock.
1186 * Weak dummy function for arches that do not yet support it.
1187 * Reads the time from the battery backed persistent clock.
1188 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1190 * XXX - Do be sure to remove it once all arches implement it.
1192 void __weak
read_persistent_clock(struct timespec
*ts
)
1198 void __weak
read_persistent_clock64(struct timespec64
*ts64
)
1202 read_persistent_clock(&ts
);
1203 *ts64
= timespec_to_timespec64(ts
);
1207 * read_boot_clock64 - Return time of the system start.
1209 * Weak dummy function for arches that do not yet support it.
1210 * Function to read the exact time the system has been started.
1211 * Returns a timespec64 with tv_sec=0 and tv_nsec=0 if unsupported.
1213 * XXX - Do be sure to remove it once all arches implement it.
1215 void __weak
read_boot_clock64(struct timespec64
*ts
)
1221 /* Flag for if timekeeping_resume() has injected sleeptime */
1222 static bool sleeptime_injected
;
1224 /* Flag for if there is a persistent clock on this platform */
1225 static bool persistent_clock_exists
;
1228 * timekeeping_init - Initializes the clocksource and common timekeeping values
1230 void __init
timekeeping_init(void)
1232 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1233 struct clocksource
*clock
;
1234 unsigned long flags
;
1235 struct timespec64 now
, boot
, tmp
;
1237 read_persistent_clock64(&now
);
1238 if (!timespec64_valid_strict(&now
)) {
1239 pr_warn("WARNING: Persistent clock returned invalid value!\n"
1240 " Check your CMOS/BIOS settings.\n");
1243 } else if (now
.tv_sec
|| now
.tv_nsec
)
1244 persistent_clock_exists
= true;
1246 read_boot_clock64(&boot
);
1247 if (!timespec64_valid_strict(&boot
)) {
1248 pr_warn("WARNING: Boot clock returned invalid value!\n"
1249 " Check your CMOS/BIOS settings.\n");
1254 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
1255 write_seqcount_begin(&tk_core
.seq
);
1258 clock
= clocksource_default_clock();
1260 clock
->enable(clock
);
1261 tk_setup_internals(tk
, clock
);
1263 tk_set_xtime(tk
, &now
);
1264 tk
->raw_time
.tv_sec
= 0;
1265 tk
->raw_time
.tv_nsec
= 0;
1266 if (boot
.tv_sec
== 0 && boot
.tv_nsec
== 0)
1267 boot
= tk_xtime(tk
);
1269 set_normalized_timespec64(&tmp
, -boot
.tv_sec
, -boot
.tv_nsec
);
1270 tk_set_wall_to_mono(tk
, tmp
);
1272 timekeeping_update(tk
, TK_MIRROR
);
1274 write_seqcount_end(&tk_core
.seq
);
1275 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1278 /* time in seconds when suspend began for persistent clock */
1279 static struct timespec64 timekeeping_suspend_time
;
1282 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1283 * @delta: pointer to a timespec delta value
1285 * Takes a timespec offset measuring a suspend interval and properly
1286 * adds the sleep offset to the timekeeping variables.
1288 static void __timekeeping_inject_sleeptime(struct timekeeper
*tk
,
1289 struct timespec64
*delta
)
1291 if (!timespec64_valid_strict(delta
)) {
1292 printk_deferred(KERN_WARNING
1293 "__timekeeping_inject_sleeptime: Invalid "
1294 "sleep delta value!\n");
1297 tk_xtime_add(tk
, delta
);
1298 tk_set_wall_to_mono(tk
, timespec64_sub(tk
->wall_to_monotonic
, *delta
));
1299 tk_update_sleep_time(tk
, timespec64_to_ktime(*delta
));
1300 tk_debug_account_sleep_time(delta
);
1303 #if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1305 * We have three kinds of time sources to use for sleep time
1306 * injection, the preference order is:
1307 * 1) non-stop clocksource
1308 * 2) persistent clock (ie: RTC accessible when irqs are off)
1311 * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1312 * If system has neither 1) nor 2), 3) will be used finally.
1315 * If timekeeping has injected sleeptime via either 1) or 2),
1316 * 3) becomes needless, so in this case we don't need to call
1317 * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1320 bool timekeeping_rtc_skipresume(void)
1322 return sleeptime_injected
;
1326 * 1) can be determined whether to use or not only when doing
1327 * timekeeping_resume() which is invoked after rtc_suspend(),
1328 * so we can't skip rtc_suspend() surely if system has 1).
1330 * But if system has 2), 2) will definitely be used, so in this
1331 * case we don't need to call rtc_suspend(), and this is what
1332 * timekeeping_rtc_skipsuspend() means.
1334 bool timekeeping_rtc_skipsuspend(void)
1336 return persistent_clock_exists
;
1340 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1341 * @delta: pointer to a timespec64 delta value
1343 * This hook is for architectures that cannot support read_persistent_clock64
1344 * because their RTC/persistent clock is only accessible when irqs are enabled.
1345 * and also don't have an effective nonstop clocksource.
1347 * This function should only be called by rtc_resume(), and allows
1348 * a suspend offset to be injected into the timekeeping values.
1350 void timekeeping_inject_sleeptime64(struct timespec64
*delta
)
1352 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1353 unsigned long flags
;
1355 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
1356 write_seqcount_begin(&tk_core
.seq
);
1358 timekeeping_forward_now(tk
);
1360 __timekeeping_inject_sleeptime(tk
, delta
);
1362 timekeeping_update(tk
, TK_CLEAR_NTP
| TK_MIRROR
| TK_CLOCK_WAS_SET
);
1364 write_seqcount_end(&tk_core
.seq
);
1365 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1367 /* signal hrtimers about time change */
1373 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1375 void timekeeping_resume(void)
1377 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1378 struct clocksource
*clock
= tk
->tkr_mono
.clock
;
1379 unsigned long flags
;
1380 struct timespec64 ts_new
, ts_delta
;
1381 cycle_t cycle_now
, cycle_delta
;
1383 sleeptime_injected
= false;
1384 read_persistent_clock64(&ts_new
);
1386 clockevents_resume();
1387 clocksource_resume();
1389 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
1390 write_seqcount_begin(&tk_core
.seq
);
1393 * After system resumes, we need to calculate the suspended time and
1394 * compensate it for the OS time. There are 3 sources that could be
1395 * used: Nonstop clocksource during suspend, persistent clock and rtc
1398 * One specific platform may have 1 or 2 or all of them, and the
1399 * preference will be:
1400 * suspend-nonstop clocksource -> persistent clock -> rtc
1401 * The less preferred source will only be tried if there is no better
1402 * usable source. The rtc part is handled separately in rtc core code.
1404 cycle_now
= tk
->tkr_mono
.read(clock
);
1405 if ((clock
->flags
& CLOCK_SOURCE_SUSPEND_NONSTOP
) &&
1406 cycle_now
> tk
->tkr_mono
.cycle_last
) {
1407 u64 num
, max
= ULLONG_MAX
;
1408 u32 mult
= clock
->mult
;
1409 u32 shift
= clock
->shift
;
1412 cycle_delta
= clocksource_delta(cycle_now
, tk
->tkr_mono
.cycle_last
,
1416 * "cycle_delta * mutl" may cause 64 bits overflow, if the
1417 * suspended time is too long. In that case we need do the
1418 * 64 bits math carefully
1421 if (cycle_delta
> max
) {
1422 num
= div64_u64(cycle_delta
, max
);
1423 nsec
= (((u64
) max
* mult
) >> shift
) * num
;
1424 cycle_delta
-= num
* max
;
1426 nsec
+= ((u64
) cycle_delta
* mult
) >> shift
;
1428 ts_delta
= ns_to_timespec64(nsec
);
1429 sleeptime_injected
= true;
1430 } else if (timespec64_compare(&ts_new
, &timekeeping_suspend_time
) > 0) {
1431 ts_delta
= timespec64_sub(ts_new
, timekeeping_suspend_time
);
1432 sleeptime_injected
= true;
1435 if (sleeptime_injected
)
1436 __timekeeping_inject_sleeptime(tk
, &ts_delta
);
1438 /* Re-base the last cycle value */
1439 tk
->tkr_mono
.cycle_last
= cycle_now
;
1440 tk
->tkr_raw
.cycle_last
= cycle_now
;
1443 timekeeping_suspended
= 0;
1444 timekeeping_update(tk
, TK_MIRROR
| TK_CLOCK_WAS_SET
);
1445 write_seqcount_end(&tk_core
.seq
);
1446 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1448 touch_softlockup_watchdog();
1454 int timekeeping_suspend(void)
1456 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1457 unsigned long flags
;
1458 struct timespec64 delta
, delta_delta
;
1459 static struct timespec64 old_delta
;
1461 read_persistent_clock64(&timekeeping_suspend_time
);
1464 * On some systems the persistent_clock can not be detected at
1465 * timekeeping_init by its return value, so if we see a valid
1466 * value returned, update the persistent_clock_exists flag.
1468 if (timekeeping_suspend_time
.tv_sec
|| timekeeping_suspend_time
.tv_nsec
)
1469 persistent_clock_exists
= true;
1471 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
1472 write_seqcount_begin(&tk_core
.seq
);
1473 timekeeping_forward_now(tk
);
1474 timekeeping_suspended
= 1;
1476 if (persistent_clock_exists
) {
1478 * To avoid drift caused by repeated suspend/resumes,
1479 * which each can add ~1 second drift error,
1480 * try to compensate so the difference in system time
1481 * and persistent_clock time stays close to constant.
1483 delta
= timespec64_sub(tk_xtime(tk
), timekeeping_suspend_time
);
1484 delta_delta
= timespec64_sub(delta
, old_delta
);
1485 if (abs(delta_delta
.tv_sec
) >= 2) {
1487 * if delta_delta is too large, assume time correction
1488 * has occurred and set old_delta to the current delta.
1492 /* Otherwise try to adjust old_system to compensate */
1493 timekeeping_suspend_time
=
1494 timespec64_add(timekeeping_suspend_time
, delta_delta
);
1498 timekeeping_update(tk
, TK_MIRROR
);
1499 halt_fast_timekeeper(tk
);
1500 write_seqcount_end(&tk_core
.seq
);
1501 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1504 clocksource_suspend();
1505 clockevents_suspend();
1510 /* sysfs resume/suspend bits for timekeeping */
1511 static struct syscore_ops timekeeping_syscore_ops
= {
1512 .resume
= timekeeping_resume
,
1513 .suspend
= timekeeping_suspend
,
1516 static int __init
timekeeping_init_ops(void)
1518 register_syscore_ops(&timekeeping_syscore_ops
);
1521 device_initcall(timekeeping_init_ops
);
1524 * Apply a multiplier adjustment to the timekeeper
1526 static __always_inline
void timekeeping_apply_adjustment(struct timekeeper
*tk
,
1531 s64 interval
= tk
->cycle_interval
;
1535 mult_adj
= -mult_adj
;
1536 interval
= -interval
;
1539 mult_adj
<<= adj_scale
;
1540 interval
<<= adj_scale
;
1541 offset
<<= adj_scale
;
1544 * So the following can be confusing.
1546 * To keep things simple, lets assume mult_adj == 1 for now.
1548 * When mult_adj != 1, remember that the interval and offset values
1549 * have been appropriately scaled so the math is the same.
1551 * The basic idea here is that we're increasing the multiplier
1552 * by one, this causes the xtime_interval to be incremented by
1553 * one cycle_interval. This is because:
1554 * xtime_interval = cycle_interval * mult
1555 * So if mult is being incremented by one:
1556 * xtime_interval = cycle_interval * (mult + 1)
1558 * xtime_interval = (cycle_interval * mult) + cycle_interval
1559 * Which can be shortened to:
1560 * xtime_interval += cycle_interval
1562 * So offset stores the non-accumulated cycles. Thus the current
1563 * time (in shifted nanoseconds) is:
1564 * now = (offset * adj) + xtime_nsec
1565 * Now, even though we're adjusting the clock frequency, we have
1566 * to keep time consistent. In other words, we can't jump back
1567 * in time, and we also want to avoid jumping forward in time.
1569 * So given the same offset value, we need the time to be the same
1570 * both before and after the freq adjustment.
1571 * now = (offset * adj_1) + xtime_nsec_1
1572 * now = (offset * adj_2) + xtime_nsec_2
1574 * (offset * adj_1) + xtime_nsec_1 =
1575 * (offset * adj_2) + xtime_nsec_2
1579 * (offset * adj_1) + xtime_nsec_1 =
1580 * (offset * (adj_1+1)) + xtime_nsec_2
1581 * (offset * adj_1) + xtime_nsec_1 =
1582 * (offset * adj_1) + offset + xtime_nsec_2
1583 * Canceling the sides:
1584 * xtime_nsec_1 = offset + xtime_nsec_2
1586 * xtime_nsec_2 = xtime_nsec_1 - offset
1587 * Which simplfies to:
1588 * xtime_nsec -= offset
1590 * XXX - TODO: Doc ntp_error calculation.
1592 if ((mult_adj
> 0) && (tk
->tkr_mono
.mult
+ mult_adj
< mult_adj
)) {
1593 /* NTP adjustment caused clocksource mult overflow */
1598 tk
->tkr_mono
.mult
+= mult_adj
;
1599 tk
->xtime_interval
+= interval
;
1600 tk
->tkr_mono
.xtime_nsec
-= offset
;
1601 tk
->ntp_error
-= (interval
- offset
) << tk
->ntp_error_shift
;
1605 * Calculate the multiplier adjustment needed to match the frequency
1608 static __always_inline
void timekeeping_freqadjust(struct timekeeper
*tk
,
1611 s64 interval
= tk
->cycle_interval
;
1612 s64 xinterval
= tk
->xtime_interval
;
1617 /* Remove any current error adj from freq calculation */
1618 if (tk
->ntp_err_mult
)
1619 xinterval
-= tk
->cycle_interval
;
1621 tk
->ntp_tick
= ntp_tick_length();
1623 /* Calculate current error per tick */
1624 tick_error
= ntp_tick_length() >> tk
->ntp_error_shift
;
1625 tick_error
-= (xinterval
+ tk
->xtime_remainder
);
1627 /* Don't worry about correcting it if its small */
1628 if (likely((tick_error
>= 0) && (tick_error
<= interval
)))
1631 /* preserve the direction of correction */
1632 negative
= (tick_error
< 0);
1634 /* Sort out the magnitude of the correction */
1635 tick_error
= abs(tick_error
);
1636 for (adj
= 0; tick_error
> interval
; adj
++)
1639 /* scale the corrections */
1640 timekeeping_apply_adjustment(tk
, offset
, negative
, adj
);
1644 * Adjust the timekeeper's multiplier to the correct frequency
1645 * and also to reduce the accumulated error value.
1647 static void timekeeping_adjust(struct timekeeper
*tk
, s64 offset
)
1649 /* Correct for the current frequency error */
1650 timekeeping_freqadjust(tk
, offset
);
1652 /* Next make a small adjustment to fix any cumulative error */
1653 if (!tk
->ntp_err_mult
&& (tk
->ntp_error
> 0)) {
1654 tk
->ntp_err_mult
= 1;
1655 timekeeping_apply_adjustment(tk
, offset
, 0, 0);
1656 } else if (tk
->ntp_err_mult
&& (tk
->ntp_error
<= 0)) {
1657 /* Undo any existing error adjustment */
1658 timekeeping_apply_adjustment(tk
, offset
, 1, 0);
1659 tk
->ntp_err_mult
= 0;
1662 if (unlikely(tk
->tkr_mono
.clock
->maxadj
&&
1663 (abs(tk
->tkr_mono
.mult
- tk
->tkr_mono
.clock
->mult
)
1664 > tk
->tkr_mono
.clock
->maxadj
))) {
1665 printk_once(KERN_WARNING
1666 "Adjusting %s more than 11%% (%ld vs %ld)\n",
1667 tk
->tkr_mono
.clock
->name
, (long)tk
->tkr_mono
.mult
,
1668 (long)tk
->tkr_mono
.clock
->mult
+ tk
->tkr_mono
.clock
->maxadj
);
1672 * It may be possible that when we entered this function, xtime_nsec
1673 * was very small. Further, if we're slightly speeding the clocksource
1674 * in the code above, its possible the required corrective factor to
1675 * xtime_nsec could cause it to underflow.
1677 * Now, since we already accumulated the second, cannot simply roll
1678 * the accumulated second back, since the NTP subsystem has been
1679 * notified via second_overflow. So instead we push xtime_nsec forward
1680 * by the amount we underflowed, and add that amount into the error.
1682 * We'll correct this error next time through this function, when
1683 * xtime_nsec is not as small.
1685 if (unlikely((s64
)tk
->tkr_mono
.xtime_nsec
< 0)) {
1686 s64 neg
= -(s64
)tk
->tkr_mono
.xtime_nsec
;
1687 tk
->tkr_mono
.xtime_nsec
= 0;
1688 tk
->ntp_error
+= neg
<< tk
->ntp_error_shift
;
1693 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1695 * Helper function that accumulates a the nsecs greater then a second
1696 * from the xtime_nsec field to the xtime_secs field.
1697 * It also calls into the NTP code to handle leapsecond processing.
1700 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper
*tk
)
1702 u64 nsecps
= (u64
)NSEC_PER_SEC
<< tk
->tkr_mono
.shift
;
1703 unsigned int clock_set
= 0;
1705 while (tk
->tkr_mono
.xtime_nsec
>= nsecps
) {
1708 tk
->tkr_mono
.xtime_nsec
-= nsecps
;
1711 /* Figure out if its a leap sec and apply if needed */
1712 leap
= second_overflow(tk
->xtime_sec
);
1713 if (unlikely(leap
)) {
1714 struct timespec64 ts
;
1716 tk
->xtime_sec
+= leap
;
1720 tk_set_wall_to_mono(tk
,
1721 timespec64_sub(tk
->wall_to_monotonic
, ts
));
1723 __timekeeping_set_tai_offset(tk
, tk
->tai_offset
- leap
);
1725 clock_set
= TK_CLOCK_WAS_SET
;
1732 * logarithmic_accumulation - shifted accumulation of cycles
1734 * This functions accumulates a shifted interval of cycles into
1735 * into a shifted interval nanoseconds. Allows for O(log) accumulation
1738 * Returns the unconsumed cycles.
1740 static cycle_t
logarithmic_accumulation(struct timekeeper
*tk
, cycle_t offset
,
1742 unsigned int *clock_set
)
1744 cycle_t interval
= tk
->cycle_interval
<< shift
;
1747 /* If the offset is smaller then a shifted interval, do nothing */
1748 if (offset
< interval
)
1751 /* Accumulate one shifted interval */
1753 tk
->tkr_mono
.cycle_last
+= interval
;
1754 tk
->tkr_raw
.cycle_last
+= interval
;
1756 tk
->tkr_mono
.xtime_nsec
+= tk
->xtime_interval
<< shift
;
1757 *clock_set
|= accumulate_nsecs_to_secs(tk
);
1759 /* Accumulate raw time */
1760 raw_nsecs
= (u64
)tk
->raw_interval
<< shift
;
1761 raw_nsecs
+= tk
->raw_time
.tv_nsec
;
1762 if (raw_nsecs
>= NSEC_PER_SEC
) {
1763 u64 raw_secs
= raw_nsecs
;
1764 raw_nsecs
= do_div(raw_secs
, NSEC_PER_SEC
);
1765 tk
->raw_time
.tv_sec
+= raw_secs
;
1767 tk
->raw_time
.tv_nsec
= raw_nsecs
;
1769 /* Accumulate error between NTP and clock interval */
1770 tk
->ntp_error
+= tk
->ntp_tick
<< shift
;
1771 tk
->ntp_error
-= (tk
->xtime_interval
+ tk
->xtime_remainder
) <<
1772 (tk
->ntp_error_shift
+ shift
);
1778 * update_wall_time - Uses the current clocksource to increment the wall time
1781 void update_wall_time(void)
1783 struct timekeeper
*real_tk
= &tk_core
.timekeeper
;
1784 struct timekeeper
*tk
= &shadow_timekeeper
;
1786 int shift
= 0, maxshift
;
1787 unsigned int clock_set
= 0;
1788 unsigned long flags
;
1790 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
1792 /* Make sure we're fully resumed: */
1793 if (unlikely(timekeeping_suspended
))
1796 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
1797 offset
= real_tk
->cycle_interval
;
1799 offset
= clocksource_delta(tk
->tkr_mono
.read(tk
->tkr_mono
.clock
),
1800 tk
->tkr_mono
.cycle_last
, tk
->tkr_mono
.mask
);
1803 /* Check if there's really nothing to do */
1804 if (offset
< real_tk
->cycle_interval
)
1807 /* Do some additional sanity checking */
1808 timekeeping_check_update(real_tk
, offset
);
1811 * With NO_HZ we may have to accumulate many cycle_intervals
1812 * (think "ticks") worth of time at once. To do this efficiently,
1813 * we calculate the largest doubling multiple of cycle_intervals
1814 * that is smaller than the offset. We then accumulate that
1815 * chunk in one go, and then try to consume the next smaller
1818 shift
= ilog2(offset
) - ilog2(tk
->cycle_interval
);
1819 shift
= max(0, shift
);
1820 /* Bound shift to one less than what overflows tick_length */
1821 maxshift
= (64 - (ilog2(ntp_tick_length())+1)) - 1;
1822 shift
= min(shift
, maxshift
);
1823 while (offset
>= tk
->cycle_interval
) {
1824 offset
= logarithmic_accumulation(tk
, offset
, shift
,
1826 if (offset
< tk
->cycle_interval
<<shift
)
1830 /* correct the clock when NTP error is too big */
1831 timekeeping_adjust(tk
, offset
);
1834 * XXX This can be killed once everyone converts
1835 * to the new update_vsyscall.
1837 old_vsyscall_fixup(tk
);
1840 * Finally, make sure that after the rounding
1841 * xtime_nsec isn't larger than NSEC_PER_SEC
1843 clock_set
|= accumulate_nsecs_to_secs(tk
);
1845 write_seqcount_begin(&tk_core
.seq
);
1847 * Update the real timekeeper.
1849 * We could avoid this memcpy by switching pointers, but that
1850 * requires changes to all other timekeeper usage sites as
1851 * well, i.e. move the timekeeper pointer getter into the
1852 * spinlocked/seqcount protected sections. And we trade this
1853 * memcpy under the tk_core.seq against one before we start
1856 timekeeping_update(tk
, clock_set
);
1857 memcpy(real_tk
, tk
, sizeof(*tk
));
1858 /* The memcpy must come last. Do not put anything here! */
1859 write_seqcount_end(&tk_core
.seq
);
1861 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1863 /* Have to call _delayed version, since in irq context*/
1864 clock_was_set_delayed();
1868 * getboottime64 - Return the real time of system boot.
1869 * @ts: pointer to the timespec64 to be set
1871 * Returns the wall-time of boot in a timespec64.
1873 * This is based on the wall_to_monotonic offset and the total suspend
1874 * time. Calls to settimeofday will affect the value returned (which
1875 * basically means that however wrong your real time clock is at boot time,
1876 * you get the right time here).
1878 void getboottime64(struct timespec64
*ts
)
1880 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1881 ktime_t t
= ktime_sub(tk
->offs_real
, tk
->offs_boot
);
1883 *ts
= ktime_to_timespec64(t
);
1885 EXPORT_SYMBOL_GPL(getboottime64
);
1887 unsigned long get_seconds(void)
1889 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1891 return tk
->xtime_sec
;
1893 EXPORT_SYMBOL(get_seconds
);
1895 struct timespec
__current_kernel_time(void)
1897 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1899 return timespec64_to_timespec(tk_xtime(tk
));
1902 struct timespec
current_kernel_time(void)
1904 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1905 struct timespec64 now
;
1909 seq
= read_seqcount_begin(&tk_core
.seq
);
1912 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
1914 return timespec64_to_timespec(now
);
1916 EXPORT_SYMBOL(current_kernel_time
);
1918 struct timespec64
get_monotonic_coarse64(void)
1920 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1921 struct timespec64 now
, mono
;
1925 seq
= read_seqcount_begin(&tk_core
.seq
);
1928 mono
= tk
->wall_to_monotonic
;
1929 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
1931 set_normalized_timespec64(&now
, now
.tv_sec
+ mono
.tv_sec
,
1932 now
.tv_nsec
+ mono
.tv_nsec
);
1938 * Must hold jiffies_lock
1940 void do_timer(unsigned long ticks
)
1942 jiffies_64
+= ticks
;
1943 calc_global_load(ticks
);
1947 * ktime_get_update_offsets_now - hrtimer helper
1948 * @cwsseq: pointer to check and store the clock was set sequence number
1949 * @offs_real: pointer to storage for monotonic -> realtime offset
1950 * @offs_boot: pointer to storage for monotonic -> boottime offset
1951 * @offs_tai: pointer to storage for monotonic -> clock tai offset
1953 * Returns current monotonic time and updates the offsets if the
1954 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
1957 * Called from hrtimer_interrupt() or retrigger_next_event()
1959 ktime_t
ktime_get_update_offsets_now(unsigned int *cwsseq
, ktime_t
*offs_real
,
1960 ktime_t
*offs_boot
, ktime_t
*offs_tai
)
1962 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1968 seq
= read_seqcount_begin(&tk_core
.seq
);
1970 base
= tk
->tkr_mono
.base
;
1971 nsecs
= timekeeping_get_ns(&tk
->tkr_mono
);
1972 base
= ktime_add_ns(base
, nsecs
);
1974 if (*cwsseq
!= tk
->clock_was_set_seq
) {
1975 *cwsseq
= tk
->clock_was_set_seq
;
1976 *offs_real
= tk
->offs_real
;
1977 *offs_boot
= tk
->offs_boot
;
1978 *offs_tai
= tk
->offs_tai
;
1981 /* Handle leapsecond insertion adjustments */
1982 if (unlikely(base
.tv64
>= tk
->next_leap_ktime
.tv64
))
1983 *offs_real
= ktime_sub(tk
->offs_real
, ktime_set(1, 0));
1985 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
1991 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
1993 int do_adjtimex(struct timex
*txc
)
1995 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1996 unsigned long flags
;
1997 struct timespec64 ts
;
2001 /* Validate the data before disabling interrupts */
2002 ret
= ntp_validate_timex(txc
);
2006 if (txc
->modes
& ADJ_SETOFFSET
) {
2007 struct timespec delta
;
2008 delta
.tv_sec
= txc
->time
.tv_sec
;
2009 delta
.tv_nsec
= txc
->time
.tv_usec
;
2010 if (!(txc
->modes
& ADJ_NANO
))
2011 delta
.tv_nsec
*= 1000;
2012 ret
= timekeeping_inject_offset(&delta
);
2017 getnstimeofday64(&ts
);
2019 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
2020 write_seqcount_begin(&tk_core
.seq
);
2022 orig_tai
= tai
= tk
->tai_offset
;
2023 ret
= __do_adjtimex(txc
, &ts
, &tai
);
2025 if (tai
!= orig_tai
) {
2026 __timekeeping_set_tai_offset(tk
, tai
);
2027 timekeeping_update(tk
, TK_MIRROR
| TK_CLOCK_WAS_SET
);
2029 tk_update_leap_state(tk
);
2031 write_seqcount_end(&tk_core
.seq
);
2032 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
2034 if (tai
!= orig_tai
)
2037 ntp_notify_cmos_timer();
2042 #ifdef CONFIG_NTP_PPS
2044 * hardpps() - Accessor function to NTP __hardpps function
2046 void hardpps(const struct timespec
*phase_ts
, const struct timespec
*raw_ts
)
2048 unsigned long flags
;
2050 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
2051 write_seqcount_begin(&tk_core
.seq
);
2053 __hardpps(phase_ts
, raw_ts
);
2055 write_seqcount_end(&tk_core
.seq
);
2056 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
2058 EXPORT_SYMBOL(hardpps
);
2062 * xtime_update() - advances the timekeeping infrastructure
2063 * @ticks: number of ticks, that have elapsed since the last call.
2065 * Must be called with interrupts disabled.
2067 void xtime_update(unsigned long ticks
)
2069 write_seqlock(&jiffies_lock
);
2071 write_sequnlock(&jiffies_lock
);