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 * These simple flag variables are managed
123 * without locks, which is racy, but ok since
124 * we don't really care about being super
125 * precise about how many events were seen,
126 * just that a problem was observed.
128 static int timekeeping_underflow_seen
;
129 static int timekeeping_overflow_seen
;
131 /* last_warning is only modified under the timekeeping lock */
132 static long timekeeping_last_warning
;
134 static void timekeeping_check_update(struct timekeeper
*tk
, cycle_t offset
)
137 cycle_t max_cycles
= tk
->tkr_mono
.clock
->max_cycles
;
138 const char *name
= tk
->tkr_mono
.clock
->name
;
140 if (offset
> max_cycles
) {
141 printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n",
142 offset
, name
, max_cycles
);
143 printk_deferred(" timekeeping: Your kernel is sick, but tries to cope by capping time updates\n");
145 if (offset
> (max_cycles
>> 1)) {
146 printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the the '%s' clock's 50%% safety margin (%lld)\n",
147 offset
, name
, max_cycles
>> 1);
148 printk_deferred(" timekeeping: Your kernel is still fine, but is feeling a bit nervous\n");
152 if (timekeeping_underflow_seen
) {
153 if (jiffies
- timekeeping_last_warning
> WARNING_FREQ
) {
154 printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name
);
155 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
156 printk_deferred(" Your kernel is probably still fine.\n");
157 timekeeping_last_warning
= jiffies
;
159 timekeeping_underflow_seen
= 0;
162 if (timekeeping_overflow_seen
) {
163 if (jiffies
- timekeeping_last_warning
> WARNING_FREQ
) {
164 printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name
);
165 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
166 printk_deferred(" Your kernel is probably still fine.\n");
167 timekeeping_last_warning
= jiffies
;
169 timekeeping_overflow_seen
= 0;
173 static inline cycle_t
timekeeping_get_delta(struct tk_read_base
*tkr
)
175 cycle_t now
, last
, mask
, max
, delta
;
179 * Since we're called holding a seqlock, the data may shift
180 * under us while we're doing the calculation. This can cause
181 * false positives, since we'd note a problem but throw the
182 * results away. So nest another seqlock here to atomically
183 * grab the points we are checking with.
186 seq
= read_seqcount_begin(&tk_core
.seq
);
187 now
= tkr
->read(tkr
->clock
);
188 last
= tkr
->cycle_last
;
190 max
= tkr
->clock
->max_cycles
;
191 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
193 delta
= clocksource_delta(now
, last
, mask
);
196 * Try to catch underflows by checking if we are seeing small
197 * mask-relative negative values.
199 if (unlikely((~delta
& mask
) < (mask
>> 3))) {
200 timekeeping_underflow_seen
= 1;
204 /* Cap delta value to the max_cycles values to avoid mult overflows */
205 if (unlikely(delta
> max
)) {
206 timekeeping_overflow_seen
= 1;
207 delta
= tkr
->clock
->max_cycles
;
213 static inline void timekeeping_check_update(struct timekeeper
*tk
, cycle_t offset
)
216 static inline cycle_t
timekeeping_get_delta(struct tk_read_base
*tkr
)
218 cycle_t cycle_now
, delta
;
220 /* read clocksource */
221 cycle_now
= tkr
->read(tkr
->clock
);
223 /* calculate the delta since the last update_wall_time */
224 delta
= clocksource_delta(cycle_now
, tkr
->cycle_last
, tkr
->mask
);
231 * tk_setup_internals - Set up internals to use clocksource clock.
233 * @tk: The target timekeeper to setup.
234 * @clock: Pointer to clocksource.
236 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
237 * pair and interval request.
239 * Unless you're the timekeeping code, you should not be using this!
241 static void tk_setup_internals(struct timekeeper
*tk
, struct clocksource
*clock
)
244 u64 tmp
, ntpinterval
;
245 struct clocksource
*old_clock
;
247 old_clock
= tk
->tkr_mono
.clock
;
248 tk
->tkr_mono
.clock
= clock
;
249 tk
->tkr_mono
.read
= clock
->read
;
250 tk
->tkr_mono
.mask
= clock
->mask
;
251 tk
->tkr_mono
.cycle_last
= tk
->tkr_mono
.read(clock
);
253 tk
->tkr_raw
.clock
= clock
;
254 tk
->tkr_raw
.read
= clock
->read
;
255 tk
->tkr_raw
.mask
= clock
->mask
;
256 tk
->tkr_raw
.cycle_last
= tk
->tkr_mono
.cycle_last
;
258 /* Do the ns -> cycle conversion first, using original mult */
259 tmp
= NTP_INTERVAL_LENGTH
;
260 tmp
<<= clock
->shift
;
262 tmp
+= clock
->mult
/2;
263 do_div(tmp
, clock
->mult
);
267 interval
= (cycle_t
) tmp
;
268 tk
->cycle_interval
= interval
;
270 /* Go back from cycles -> shifted ns */
271 tk
->xtime_interval
= (u64
) interval
* clock
->mult
;
272 tk
->xtime_remainder
= ntpinterval
- tk
->xtime_interval
;
274 ((u64
) interval
* clock
->mult
) >> clock
->shift
;
276 /* if changing clocks, convert xtime_nsec shift units */
278 int shift_change
= clock
->shift
- old_clock
->shift
;
279 if (shift_change
< 0)
280 tk
->tkr_mono
.xtime_nsec
>>= -shift_change
;
282 tk
->tkr_mono
.xtime_nsec
<<= shift_change
;
284 tk
->tkr_raw
.xtime_nsec
= 0;
286 tk
->tkr_mono
.shift
= clock
->shift
;
287 tk
->tkr_raw
.shift
= clock
->shift
;
290 tk
->ntp_error_shift
= NTP_SCALE_SHIFT
- clock
->shift
;
291 tk
->ntp_tick
= ntpinterval
<< tk
->ntp_error_shift
;
294 * The timekeeper keeps its own mult values for the currently
295 * active clocksource. These value will be adjusted via NTP
296 * to counteract clock drifting.
298 tk
->tkr_mono
.mult
= clock
->mult
;
299 tk
->tkr_raw
.mult
= clock
->mult
;
300 tk
->ntp_err_mult
= 0;
303 /* Timekeeper helper functions. */
305 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
306 static u32
default_arch_gettimeoffset(void) { return 0; }
307 u32 (*arch_gettimeoffset
)(void) = default_arch_gettimeoffset
;
309 static inline u32
arch_gettimeoffset(void) { return 0; }
312 static inline s64
timekeeping_get_ns(struct tk_read_base
*tkr
)
317 delta
= timekeeping_get_delta(tkr
);
319 nsec
= delta
* tkr
->mult
+ tkr
->xtime_nsec
;
322 /* If arch requires, add in get_arch_timeoffset() */
323 return nsec
+ arch_gettimeoffset();
327 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
328 * @tkr: Timekeeping readout base from which we take the update
330 * We want to use this from any context including NMI and tracing /
331 * instrumenting the timekeeping code itself.
333 * So we handle this differently than the other timekeeping accessor
334 * functions which retry when the sequence count has changed. The
337 * smp_wmb(); <- Ensure that the last base[1] update is visible
339 * smp_wmb(); <- Ensure that the seqcount update is visible
340 * update(tkf->base[0], tkr);
341 * smp_wmb(); <- Ensure that the base[0] update is visible
343 * smp_wmb(); <- Ensure that the seqcount update is visible
344 * update(tkf->base[1], tkr);
346 * The reader side does:
352 * now = now(tkf->base[idx]);
354 * } while (seq != tkf->seq)
356 * As long as we update base[0] readers are forced off to
357 * base[1]. Once base[0] is updated readers are redirected to base[0]
358 * and the base[1] update takes place.
360 * So if a NMI hits the update of base[0] then it will use base[1]
361 * which is still consistent. In the worst case this can result is a
362 * slightly wrong timestamp (a few nanoseconds). See
363 * @ktime_get_mono_fast_ns.
365 static void update_fast_timekeeper(struct tk_read_base
*tkr
, struct tk_fast
*tkf
)
367 struct tk_read_base
*base
= tkf
->base
;
369 /* Force readers off to base[1] */
370 raw_write_seqcount_latch(&tkf
->seq
);
373 memcpy(base
, tkr
, sizeof(*base
));
375 /* Force readers back to base[0] */
376 raw_write_seqcount_latch(&tkf
->seq
);
379 memcpy(base
+ 1, base
, sizeof(*base
));
383 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
385 * This timestamp is not guaranteed to be monotonic across an update.
386 * The timestamp is calculated by:
388 * now = base_mono + clock_delta * slope
390 * So if the update lowers the slope, readers who are forced to the
391 * not yet updated second array are still using the old steeper slope.
400 * |12345678---> reader order
406 * So reader 6 will observe time going backwards versus reader 5.
408 * While other CPUs are likely to be able observe that, the only way
409 * for a CPU local observation is when an NMI hits in the middle of
410 * the update. Timestamps taken from that NMI context might be ahead
411 * of the following timestamps. Callers need to be aware of that and
414 static __always_inline u64
__ktime_get_fast_ns(struct tk_fast
*tkf
)
416 struct tk_read_base
*tkr
;
421 seq
= raw_read_seqcount(&tkf
->seq
);
422 tkr
= tkf
->base
+ (seq
& 0x01);
423 now
= ktime_to_ns(tkr
->base
) + timekeeping_get_ns(tkr
);
424 } while (read_seqcount_retry(&tkf
->seq
, seq
));
429 u64
ktime_get_mono_fast_ns(void)
431 return __ktime_get_fast_ns(&tk_fast_mono
);
433 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns
);
435 u64
ktime_get_raw_fast_ns(void)
437 return __ktime_get_fast_ns(&tk_fast_raw
);
439 EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns
);
441 /* Suspend-time cycles value for halted fast timekeeper. */
442 static cycle_t cycles_at_suspend
;
444 static cycle_t
dummy_clock_read(struct clocksource
*cs
)
446 return cycles_at_suspend
;
450 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
451 * @tk: Timekeeper to snapshot.
453 * It generally is unsafe to access the clocksource after timekeeping has been
454 * suspended, so take a snapshot of the readout base of @tk and use it as the
455 * fast timekeeper's readout base while suspended. It will return the same
456 * number of cycles every time until timekeeping is resumed at which time the
457 * proper readout base for the fast timekeeper will be restored automatically.
459 static void halt_fast_timekeeper(struct timekeeper
*tk
)
461 static struct tk_read_base tkr_dummy
;
462 struct tk_read_base
*tkr
= &tk
->tkr_mono
;
464 memcpy(&tkr_dummy
, tkr
, sizeof(tkr_dummy
));
465 cycles_at_suspend
= tkr
->read(tkr
->clock
);
466 tkr_dummy
.read
= dummy_clock_read
;
467 update_fast_timekeeper(&tkr_dummy
, &tk_fast_mono
);
470 memcpy(&tkr_dummy
, tkr
, sizeof(tkr_dummy
));
471 tkr_dummy
.read
= dummy_clock_read
;
472 update_fast_timekeeper(&tkr_dummy
, &tk_fast_raw
);
475 #ifdef CONFIG_GENERIC_TIME_VSYSCALL_OLD
477 static inline void update_vsyscall(struct timekeeper
*tk
)
479 struct timespec xt
, wm
;
481 xt
= timespec64_to_timespec(tk_xtime(tk
));
482 wm
= timespec64_to_timespec(tk
->wall_to_monotonic
);
483 update_vsyscall_old(&xt
, &wm
, tk
->tkr_mono
.clock
, tk
->tkr_mono
.mult
,
484 tk
->tkr_mono
.cycle_last
);
487 static inline void old_vsyscall_fixup(struct timekeeper
*tk
)
492 * Store only full nanoseconds into xtime_nsec after rounding
493 * it up and add the remainder to the error difference.
494 * XXX - This is necessary to avoid small 1ns inconsistnecies caused
495 * by truncating the remainder in vsyscalls. However, it causes
496 * additional work to be done in timekeeping_adjust(). Once
497 * the vsyscall implementations are converted to use xtime_nsec
498 * (shifted nanoseconds), and CONFIG_GENERIC_TIME_VSYSCALL_OLD
499 * users are removed, this can be killed.
501 remainder
= tk
->tkr_mono
.xtime_nsec
& ((1ULL << tk
->tkr_mono
.shift
) - 1);
502 tk
->tkr_mono
.xtime_nsec
-= remainder
;
503 tk
->tkr_mono
.xtime_nsec
+= 1ULL << tk
->tkr_mono
.shift
;
504 tk
->ntp_error
+= remainder
<< tk
->ntp_error_shift
;
505 tk
->ntp_error
-= (1ULL << tk
->tkr_mono
.shift
) << tk
->ntp_error_shift
;
508 #define old_vsyscall_fixup(tk)
511 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain
);
513 static void update_pvclock_gtod(struct timekeeper
*tk
, bool was_set
)
515 raw_notifier_call_chain(&pvclock_gtod_chain
, was_set
, tk
);
519 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
521 int pvclock_gtod_register_notifier(struct notifier_block
*nb
)
523 struct timekeeper
*tk
= &tk_core
.timekeeper
;
527 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
528 ret
= raw_notifier_chain_register(&pvclock_gtod_chain
, nb
);
529 update_pvclock_gtod(tk
, true);
530 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
534 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier
);
537 * pvclock_gtod_unregister_notifier - unregister a pvclock
538 * timedata update listener
540 int pvclock_gtod_unregister_notifier(struct notifier_block
*nb
)
545 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
546 ret
= raw_notifier_chain_unregister(&pvclock_gtod_chain
, nb
);
547 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
551 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier
);
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_ktime_data(tk
);
597 update_pvclock_gtod(tk
, action
& TK_CLOCK_WAS_SET
);
599 if (action
& TK_MIRROR
)
600 memcpy(&shadow_timekeeper
, &tk_core
.timekeeper
,
601 sizeof(tk_core
.timekeeper
));
603 update_fast_timekeeper(&tk
->tkr_mono
, &tk_fast_mono
);
604 update_fast_timekeeper(&tk
->tkr_raw
, &tk_fast_raw
);
608 * timekeeping_forward_now - update clock to the current time
610 * Forward the current clock to update its state since the last call to
611 * update_wall_time(). This is useful before significant clock changes,
612 * as it avoids having to deal with this time offset explicitly.
614 static void timekeeping_forward_now(struct timekeeper
*tk
)
616 struct clocksource
*clock
= tk
->tkr_mono
.clock
;
617 cycle_t cycle_now
, delta
;
620 cycle_now
= tk
->tkr_mono
.read(clock
);
621 delta
= clocksource_delta(cycle_now
, tk
->tkr_mono
.cycle_last
, tk
->tkr_mono
.mask
);
622 tk
->tkr_mono
.cycle_last
= cycle_now
;
623 tk
->tkr_raw
.cycle_last
= cycle_now
;
625 tk
->tkr_mono
.xtime_nsec
+= delta
* tk
->tkr_mono
.mult
;
627 /* If arch requires, add in get_arch_timeoffset() */
628 tk
->tkr_mono
.xtime_nsec
+= (u64
)arch_gettimeoffset() << tk
->tkr_mono
.shift
;
630 tk_normalize_xtime(tk
);
632 nsec
= clocksource_cyc2ns(delta
, tk
->tkr_raw
.mult
, tk
->tkr_raw
.shift
);
633 timespec64_add_ns(&tk
->raw_time
, nsec
);
637 * __getnstimeofday64 - Returns the time of day in a timespec64.
638 * @ts: pointer to the timespec to be set
640 * Updates the time of day in the timespec.
641 * Returns 0 on success, or -ve when suspended (timespec will be undefined).
643 int __getnstimeofday64(struct timespec64
*ts
)
645 struct timekeeper
*tk
= &tk_core
.timekeeper
;
650 seq
= read_seqcount_begin(&tk_core
.seq
);
652 ts
->tv_sec
= tk
->xtime_sec
;
653 nsecs
= timekeeping_get_ns(&tk
->tkr_mono
);
655 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
658 timespec64_add_ns(ts
, nsecs
);
661 * Do not bail out early, in case there were callers still using
662 * the value, even in the face of the WARN_ON.
664 if (unlikely(timekeeping_suspended
))
668 EXPORT_SYMBOL(__getnstimeofday64
);
671 * getnstimeofday64 - Returns the time of day in a timespec64.
672 * @ts: pointer to the timespec64 to be set
674 * Returns the time of day in a timespec64 (WARN if suspended).
676 void getnstimeofday64(struct timespec64
*ts
)
678 WARN_ON(__getnstimeofday64(ts
));
680 EXPORT_SYMBOL(getnstimeofday64
);
682 ktime_t
ktime_get(void)
684 struct timekeeper
*tk
= &tk_core
.timekeeper
;
689 WARN_ON(timekeeping_suspended
);
692 seq
= read_seqcount_begin(&tk_core
.seq
);
693 base
= tk
->tkr_mono
.base
;
694 nsecs
= timekeeping_get_ns(&tk
->tkr_mono
);
696 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
698 return ktime_add_ns(base
, nsecs
);
700 EXPORT_SYMBOL_GPL(ktime_get
);
702 static ktime_t
*offsets
[TK_OFFS_MAX
] = {
703 [TK_OFFS_REAL
] = &tk_core
.timekeeper
.offs_real
,
704 [TK_OFFS_BOOT
] = &tk_core
.timekeeper
.offs_boot
,
705 [TK_OFFS_TAI
] = &tk_core
.timekeeper
.offs_tai
,
708 ktime_t
ktime_get_with_offset(enum tk_offsets offs
)
710 struct timekeeper
*tk
= &tk_core
.timekeeper
;
712 ktime_t base
, *offset
= offsets
[offs
];
715 WARN_ON(timekeeping_suspended
);
718 seq
= read_seqcount_begin(&tk_core
.seq
);
719 base
= ktime_add(tk
->tkr_mono
.base
, *offset
);
720 nsecs
= timekeeping_get_ns(&tk
->tkr_mono
);
722 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
724 return ktime_add_ns(base
, nsecs
);
727 EXPORT_SYMBOL_GPL(ktime_get_with_offset
);
730 * ktime_mono_to_any() - convert mononotic time to any other time
731 * @tmono: time to convert.
732 * @offs: which offset to use
734 ktime_t
ktime_mono_to_any(ktime_t tmono
, enum tk_offsets offs
)
736 ktime_t
*offset
= offsets
[offs
];
741 seq
= read_seqcount_begin(&tk_core
.seq
);
742 tconv
= ktime_add(tmono
, *offset
);
743 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
747 EXPORT_SYMBOL_GPL(ktime_mono_to_any
);
750 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
752 ktime_t
ktime_get_raw(void)
754 struct timekeeper
*tk
= &tk_core
.timekeeper
;
760 seq
= read_seqcount_begin(&tk_core
.seq
);
761 base
= tk
->tkr_raw
.base
;
762 nsecs
= timekeeping_get_ns(&tk
->tkr_raw
);
764 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
766 return ktime_add_ns(base
, nsecs
);
768 EXPORT_SYMBOL_GPL(ktime_get_raw
);
771 * ktime_get_ts64 - get the monotonic clock in timespec64 format
772 * @ts: pointer to timespec variable
774 * The function calculates the monotonic clock from the realtime
775 * clock and the wall_to_monotonic offset and stores the result
776 * in normalized timespec64 format in the variable pointed to by @ts.
778 void ktime_get_ts64(struct timespec64
*ts
)
780 struct timekeeper
*tk
= &tk_core
.timekeeper
;
781 struct timespec64 tomono
;
785 WARN_ON(timekeeping_suspended
);
788 seq
= read_seqcount_begin(&tk_core
.seq
);
789 ts
->tv_sec
= tk
->xtime_sec
;
790 nsec
= timekeeping_get_ns(&tk
->tkr_mono
);
791 tomono
= tk
->wall_to_monotonic
;
793 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
795 ts
->tv_sec
+= tomono
.tv_sec
;
797 timespec64_add_ns(ts
, nsec
+ tomono
.tv_nsec
);
799 EXPORT_SYMBOL_GPL(ktime_get_ts64
);
802 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
804 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
805 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
806 * works on both 32 and 64 bit systems. On 32 bit systems the readout
807 * covers ~136 years of uptime which should be enough to prevent
808 * premature wrap arounds.
810 time64_t
ktime_get_seconds(void)
812 struct timekeeper
*tk
= &tk_core
.timekeeper
;
814 WARN_ON(timekeeping_suspended
);
815 return tk
->ktime_sec
;
817 EXPORT_SYMBOL_GPL(ktime_get_seconds
);
820 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
822 * Returns the wall clock seconds since 1970. This replaces the
823 * get_seconds() interface which is not y2038 safe on 32bit systems.
825 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
826 * 32bit systems the access must be protected with the sequence
827 * counter to provide "atomic" access to the 64bit tk->xtime_sec
830 time64_t
ktime_get_real_seconds(void)
832 struct timekeeper
*tk
= &tk_core
.timekeeper
;
836 if (IS_ENABLED(CONFIG_64BIT
))
837 return tk
->xtime_sec
;
840 seq
= read_seqcount_begin(&tk_core
.seq
);
841 seconds
= tk
->xtime_sec
;
843 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
847 EXPORT_SYMBOL_GPL(ktime_get_real_seconds
);
849 #ifdef CONFIG_NTP_PPS
852 * getnstime_raw_and_real - get day and raw monotonic time in timespec format
853 * @ts_raw: pointer to the timespec to be set to raw monotonic time
854 * @ts_real: pointer to the timespec to be set to the time of day
856 * This function reads both the time of day and raw monotonic time at the
857 * same time atomically and stores the resulting timestamps in timespec
860 void getnstime_raw_and_real(struct timespec
*ts_raw
, struct timespec
*ts_real
)
862 struct timekeeper
*tk
= &tk_core
.timekeeper
;
864 s64 nsecs_raw
, nsecs_real
;
866 WARN_ON_ONCE(timekeeping_suspended
);
869 seq
= read_seqcount_begin(&tk_core
.seq
);
871 *ts_raw
= timespec64_to_timespec(tk
->raw_time
);
872 ts_real
->tv_sec
= tk
->xtime_sec
;
873 ts_real
->tv_nsec
= 0;
875 nsecs_raw
= timekeeping_get_ns(&tk
->tkr_raw
);
876 nsecs_real
= timekeeping_get_ns(&tk
->tkr_mono
);
878 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
880 timespec_add_ns(ts_raw
, nsecs_raw
);
881 timespec_add_ns(ts_real
, nsecs_real
);
883 EXPORT_SYMBOL(getnstime_raw_and_real
);
885 #endif /* CONFIG_NTP_PPS */
888 * do_gettimeofday - Returns the time of day in a timeval
889 * @tv: pointer to the timeval to be set
891 * NOTE: Users should be converted to using getnstimeofday()
893 void do_gettimeofday(struct timeval
*tv
)
895 struct timespec64 now
;
897 getnstimeofday64(&now
);
898 tv
->tv_sec
= now
.tv_sec
;
899 tv
->tv_usec
= now
.tv_nsec
/1000;
901 EXPORT_SYMBOL(do_gettimeofday
);
904 * do_settimeofday64 - Sets the time of day.
905 * @ts: pointer to the timespec64 variable containing the new time
907 * Sets the time of day to the new time and update NTP and notify hrtimers
909 int do_settimeofday64(const struct timespec64
*ts
)
911 struct timekeeper
*tk
= &tk_core
.timekeeper
;
912 struct timespec64 ts_delta
, xt
;
915 if (!timespec64_valid_strict(ts
))
918 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
919 write_seqcount_begin(&tk_core
.seq
);
921 timekeeping_forward_now(tk
);
924 ts_delta
.tv_sec
= ts
->tv_sec
- xt
.tv_sec
;
925 ts_delta
.tv_nsec
= ts
->tv_nsec
- xt
.tv_nsec
;
927 tk_set_wall_to_mono(tk
, timespec64_sub(tk
->wall_to_monotonic
, ts_delta
));
929 tk_set_xtime(tk
, ts
);
931 timekeeping_update(tk
, TK_CLEAR_NTP
| TK_MIRROR
| TK_CLOCK_WAS_SET
);
933 write_seqcount_end(&tk_core
.seq
);
934 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
936 /* signal hrtimers about time change */
941 EXPORT_SYMBOL(do_settimeofday64
);
944 * timekeeping_inject_offset - Adds or subtracts from the current time.
945 * @tv: pointer to the timespec variable containing the offset
947 * Adds or subtracts an offset value from the current time.
949 int timekeeping_inject_offset(struct timespec
*ts
)
951 struct timekeeper
*tk
= &tk_core
.timekeeper
;
953 struct timespec64 ts64
, tmp
;
956 if ((unsigned long)ts
->tv_nsec
>= NSEC_PER_SEC
)
959 ts64
= timespec_to_timespec64(*ts
);
961 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
962 write_seqcount_begin(&tk_core
.seq
);
964 timekeeping_forward_now(tk
);
966 /* Make sure the proposed value is valid */
967 tmp
= timespec64_add(tk_xtime(tk
), ts64
);
968 if (!timespec64_valid_strict(&tmp
)) {
973 tk_xtime_add(tk
, &ts64
);
974 tk_set_wall_to_mono(tk
, timespec64_sub(tk
->wall_to_monotonic
, ts64
));
976 error
: /* even if we error out, we forwarded the time, so call update */
977 timekeeping_update(tk
, TK_CLEAR_NTP
| TK_MIRROR
| TK_CLOCK_WAS_SET
);
979 write_seqcount_end(&tk_core
.seq
);
980 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
982 /* signal hrtimers about time change */
987 EXPORT_SYMBOL(timekeeping_inject_offset
);
991 * timekeeping_get_tai_offset - Returns current TAI offset from UTC
994 s32
timekeeping_get_tai_offset(void)
996 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1001 seq
= read_seqcount_begin(&tk_core
.seq
);
1002 ret
= tk
->tai_offset
;
1003 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
1009 * __timekeeping_set_tai_offset - Lock free worker function
1012 static void __timekeeping_set_tai_offset(struct timekeeper
*tk
, s32 tai_offset
)
1014 tk
->tai_offset
= tai_offset
;
1015 tk
->offs_tai
= ktime_add(tk
->offs_real
, ktime_set(tai_offset
, 0));
1019 * timekeeping_set_tai_offset - Sets the current TAI offset from UTC
1022 void timekeeping_set_tai_offset(s32 tai_offset
)
1024 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1025 unsigned long flags
;
1027 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
1028 write_seqcount_begin(&tk_core
.seq
);
1029 __timekeeping_set_tai_offset(tk
, tai_offset
);
1030 timekeeping_update(tk
, TK_MIRROR
| TK_CLOCK_WAS_SET
);
1031 write_seqcount_end(&tk_core
.seq
);
1032 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1037 * change_clocksource - Swaps clocksources if a new one is available
1039 * Accumulates current time interval and initializes new clocksource
1041 static int change_clocksource(void *data
)
1043 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1044 struct clocksource
*new, *old
;
1045 unsigned long flags
;
1047 new = (struct clocksource
*) data
;
1049 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
1050 write_seqcount_begin(&tk_core
.seq
);
1052 timekeeping_forward_now(tk
);
1054 * If the cs is in module, get a module reference. Succeeds
1055 * for built-in code (owner == NULL) as well.
1057 if (try_module_get(new->owner
)) {
1058 if (!new->enable
|| new->enable(new) == 0) {
1059 old
= tk
->tkr_mono
.clock
;
1060 tk_setup_internals(tk
, new);
1063 module_put(old
->owner
);
1065 module_put(new->owner
);
1068 timekeeping_update(tk
, TK_CLEAR_NTP
| TK_MIRROR
| TK_CLOCK_WAS_SET
);
1070 write_seqcount_end(&tk_core
.seq
);
1071 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1077 * timekeeping_notify - Install a new clock source
1078 * @clock: pointer to the clock source
1080 * This function is called from clocksource.c after a new, better clock
1081 * source has been registered. The caller holds the clocksource_mutex.
1083 int timekeeping_notify(struct clocksource
*clock
)
1085 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1087 if (tk
->tkr_mono
.clock
== clock
)
1089 stop_machine(change_clocksource
, clock
, NULL
);
1090 tick_clock_notify();
1091 return tk
->tkr_mono
.clock
== clock
? 0 : -1;
1095 * getrawmonotonic64 - Returns the raw monotonic time in a timespec
1096 * @ts: pointer to the timespec64 to be set
1098 * Returns the raw monotonic time (completely un-modified by ntp)
1100 void getrawmonotonic64(struct timespec64
*ts
)
1102 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1103 struct timespec64 ts64
;
1108 seq
= read_seqcount_begin(&tk_core
.seq
);
1109 nsecs
= timekeeping_get_ns(&tk
->tkr_raw
);
1110 ts64
= tk
->raw_time
;
1112 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
1114 timespec64_add_ns(&ts64
, nsecs
);
1117 EXPORT_SYMBOL(getrawmonotonic64
);
1121 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1123 int timekeeping_valid_for_hres(void)
1125 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1130 seq
= read_seqcount_begin(&tk_core
.seq
);
1132 ret
= tk
->tkr_mono
.clock
->flags
& CLOCK_SOURCE_VALID_FOR_HRES
;
1134 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
1140 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1142 u64
timekeeping_max_deferment(void)
1144 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1149 seq
= read_seqcount_begin(&tk_core
.seq
);
1151 ret
= tk
->tkr_mono
.clock
->max_idle_ns
;
1153 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
1159 * read_persistent_clock - Return time from the persistent clock.
1161 * Weak dummy function for arches that do not yet support it.
1162 * Reads the time from the battery backed persistent clock.
1163 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1165 * XXX - Do be sure to remove it once all arches implement it.
1167 void __weak
read_persistent_clock(struct timespec
*ts
)
1173 void __weak
read_persistent_clock64(struct timespec64
*ts64
)
1177 read_persistent_clock(&ts
);
1178 *ts64
= timespec_to_timespec64(ts
);
1182 * read_boot_clock - Return time of the system start.
1184 * Weak dummy function for arches that do not yet support it.
1185 * Function to read the exact time the system has been started.
1186 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1188 * XXX - Do be sure to remove it once all arches implement it.
1190 void __weak
read_boot_clock(struct timespec
*ts
)
1196 void __weak
read_boot_clock64(struct timespec64
*ts64
)
1200 read_boot_clock(&ts
);
1201 *ts64
= timespec_to_timespec64(ts
);
1204 /* Flag for if timekeeping_resume() has injected sleeptime */
1205 static bool sleeptime_injected
;
1207 /* Flag for if there is a persistent clock on this platform */
1208 static bool persistent_clock_exists
;
1211 * timekeeping_init - Initializes the clocksource and common timekeeping values
1213 void __init
timekeeping_init(void)
1215 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1216 struct clocksource
*clock
;
1217 unsigned long flags
;
1218 struct timespec64 now
, boot
, tmp
;
1220 read_persistent_clock64(&now
);
1221 if (!timespec64_valid_strict(&now
)) {
1222 pr_warn("WARNING: Persistent clock returned invalid value!\n"
1223 " Check your CMOS/BIOS settings.\n");
1226 } else if (now
.tv_sec
|| now
.tv_nsec
)
1227 persistent_clock_exists
= true;
1229 read_boot_clock64(&boot
);
1230 if (!timespec64_valid_strict(&boot
)) {
1231 pr_warn("WARNING: Boot clock returned invalid value!\n"
1232 " Check your CMOS/BIOS settings.\n");
1237 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
1238 write_seqcount_begin(&tk_core
.seq
);
1241 clock
= clocksource_default_clock();
1243 clock
->enable(clock
);
1244 tk_setup_internals(tk
, clock
);
1246 tk_set_xtime(tk
, &now
);
1247 tk
->raw_time
.tv_sec
= 0;
1248 tk
->raw_time
.tv_nsec
= 0;
1249 if (boot
.tv_sec
== 0 && boot
.tv_nsec
== 0)
1250 boot
= tk_xtime(tk
);
1252 set_normalized_timespec64(&tmp
, -boot
.tv_sec
, -boot
.tv_nsec
);
1253 tk_set_wall_to_mono(tk
, tmp
);
1255 timekeeping_update(tk
, TK_MIRROR
);
1257 write_seqcount_end(&tk_core
.seq
);
1258 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1261 /* time in seconds when suspend began for persistent clock */
1262 static struct timespec64 timekeeping_suspend_time
;
1265 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1266 * @delta: pointer to a timespec delta value
1268 * Takes a timespec offset measuring a suspend interval and properly
1269 * adds the sleep offset to the timekeeping variables.
1271 static void __timekeeping_inject_sleeptime(struct timekeeper
*tk
,
1272 struct timespec64
*delta
)
1274 if (!timespec64_valid_strict(delta
)) {
1275 printk_deferred(KERN_WARNING
1276 "__timekeeping_inject_sleeptime: Invalid "
1277 "sleep delta value!\n");
1280 tk_xtime_add(tk
, delta
);
1281 tk_set_wall_to_mono(tk
, timespec64_sub(tk
->wall_to_monotonic
, *delta
));
1282 tk_update_sleep_time(tk
, timespec64_to_ktime(*delta
));
1283 tk_debug_account_sleep_time(delta
);
1286 #if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1288 * We have three kinds of time sources to use for sleep time
1289 * injection, the preference order is:
1290 * 1) non-stop clocksource
1291 * 2) persistent clock (ie: RTC accessible when irqs are off)
1294 * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1295 * If system has neither 1) nor 2), 3) will be used finally.
1298 * If timekeeping has injected sleeptime via either 1) or 2),
1299 * 3) becomes needless, so in this case we don't need to call
1300 * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1303 bool timekeeping_rtc_skipresume(void)
1305 return sleeptime_injected
;
1309 * 1) can be determined whether to use or not only when doing
1310 * timekeeping_resume() which is invoked after rtc_suspend(),
1311 * so we can't skip rtc_suspend() surely if system has 1).
1313 * But if system has 2), 2) will definitely be used, so in this
1314 * case we don't need to call rtc_suspend(), and this is what
1315 * timekeeping_rtc_skipsuspend() means.
1317 bool timekeeping_rtc_skipsuspend(void)
1319 return persistent_clock_exists
;
1323 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1324 * @delta: pointer to a timespec64 delta value
1326 * This hook is for architectures that cannot support read_persistent_clock64
1327 * because their RTC/persistent clock is only accessible when irqs are enabled.
1328 * and also don't have an effective nonstop clocksource.
1330 * This function should only be called by rtc_resume(), and allows
1331 * a suspend offset to be injected into the timekeeping values.
1333 void timekeeping_inject_sleeptime64(struct timespec64
*delta
)
1335 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1336 unsigned long flags
;
1338 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
1339 write_seqcount_begin(&tk_core
.seq
);
1341 timekeeping_forward_now(tk
);
1343 __timekeeping_inject_sleeptime(tk
, delta
);
1345 timekeeping_update(tk
, TK_CLEAR_NTP
| TK_MIRROR
| TK_CLOCK_WAS_SET
);
1347 write_seqcount_end(&tk_core
.seq
);
1348 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1350 /* signal hrtimers about time change */
1356 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1358 * This is for the generic clocksource timekeeping.
1359 * xtime/wall_to_monotonic/jiffies/etc are
1360 * still managed by arch specific suspend/resume code.
1362 void timekeeping_resume(void)
1364 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1365 struct clocksource
*clock
= tk
->tkr_mono
.clock
;
1366 unsigned long flags
;
1367 struct timespec64 ts_new
, ts_delta
;
1368 cycle_t cycle_now
, cycle_delta
;
1370 sleeptime_injected
= false;
1371 read_persistent_clock64(&ts_new
);
1373 clockevents_resume();
1374 clocksource_resume();
1376 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
1377 write_seqcount_begin(&tk_core
.seq
);
1380 * After system resumes, we need to calculate the suspended time and
1381 * compensate it for the OS time. There are 3 sources that could be
1382 * used: Nonstop clocksource during suspend, persistent clock and rtc
1385 * One specific platform may have 1 or 2 or all of them, and the
1386 * preference will be:
1387 * suspend-nonstop clocksource -> persistent clock -> rtc
1388 * The less preferred source will only be tried if there is no better
1389 * usable source. The rtc part is handled separately in rtc core code.
1391 cycle_now
= tk
->tkr_mono
.read(clock
);
1392 if ((clock
->flags
& CLOCK_SOURCE_SUSPEND_NONSTOP
) &&
1393 cycle_now
> tk
->tkr_mono
.cycle_last
) {
1394 u64 num
, max
= ULLONG_MAX
;
1395 u32 mult
= clock
->mult
;
1396 u32 shift
= clock
->shift
;
1399 cycle_delta
= clocksource_delta(cycle_now
, tk
->tkr_mono
.cycle_last
,
1403 * "cycle_delta * mutl" may cause 64 bits overflow, if the
1404 * suspended time is too long. In that case we need do the
1405 * 64 bits math carefully
1408 if (cycle_delta
> max
) {
1409 num
= div64_u64(cycle_delta
, max
);
1410 nsec
= (((u64
) max
* mult
) >> shift
) * num
;
1411 cycle_delta
-= num
* max
;
1413 nsec
+= ((u64
) cycle_delta
* mult
) >> shift
;
1415 ts_delta
= ns_to_timespec64(nsec
);
1416 sleeptime_injected
= true;
1417 } else if (timespec64_compare(&ts_new
, &timekeeping_suspend_time
) > 0) {
1418 ts_delta
= timespec64_sub(ts_new
, timekeeping_suspend_time
);
1419 sleeptime_injected
= true;
1422 if (sleeptime_injected
)
1423 __timekeeping_inject_sleeptime(tk
, &ts_delta
);
1425 /* Re-base the last cycle value */
1426 tk
->tkr_mono
.cycle_last
= cycle_now
;
1427 tk
->tkr_raw
.cycle_last
= cycle_now
;
1430 timekeeping_suspended
= 0;
1431 timekeeping_update(tk
, TK_MIRROR
| TK_CLOCK_WAS_SET
);
1432 write_seqcount_end(&tk_core
.seq
);
1433 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1435 touch_softlockup_watchdog();
1441 int timekeeping_suspend(void)
1443 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1444 unsigned long flags
;
1445 struct timespec64 delta
, delta_delta
;
1446 static struct timespec64 old_delta
;
1448 read_persistent_clock64(&timekeeping_suspend_time
);
1451 * On some systems the persistent_clock can not be detected at
1452 * timekeeping_init by its return value, so if we see a valid
1453 * value returned, update the persistent_clock_exists flag.
1455 if (timekeeping_suspend_time
.tv_sec
|| timekeeping_suspend_time
.tv_nsec
)
1456 persistent_clock_exists
= true;
1458 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
1459 write_seqcount_begin(&tk_core
.seq
);
1460 timekeeping_forward_now(tk
);
1461 timekeeping_suspended
= 1;
1463 if (persistent_clock_exists
) {
1465 * To avoid drift caused by repeated suspend/resumes,
1466 * which each can add ~1 second drift error,
1467 * try to compensate so the difference in system time
1468 * and persistent_clock time stays close to constant.
1470 delta
= timespec64_sub(tk_xtime(tk
), timekeeping_suspend_time
);
1471 delta_delta
= timespec64_sub(delta
, old_delta
);
1472 if (abs(delta_delta
.tv_sec
) >= 2) {
1474 * if delta_delta is too large, assume time correction
1475 * has occurred and set old_delta to the current delta.
1479 /* Otherwise try to adjust old_system to compensate */
1480 timekeeping_suspend_time
=
1481 timespec64_add(timekeeping_suspend_time
, delta_delta
);
1485 timekeeping_update(tk
, TK_MIRROR
);
1486 halt_fast_timekeeper(tk
);
1487 write_seqcount_end(&tk_core
.seq
);
1488 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1491 clocksource_suspend();
1492 clockevents_suspend();
1497 /* sysfs resume/suspend bits for timekeeping */
1498 static struct syscore_ops timekeeping_syscore_ops
= {
1499 .resume
= timekeeping_resume
,
1500 .suspend
= timekeeping_suspend
,
1503 static int __init
timekeeping_init_ops(void)
1505 register_syscore_ops(&timekeeping_syscore_ops
);
1508 device_initcall(timekeeping_init_ops
);
1511 * Apply a multiplier adjustment to the timekeeper
1513 static __always_inline
void timekeeping_apply_adjustment(struct timekeeper
*tk
,
1518 s64 interval
= tk
->cycle_interval
;
1522 mult_adj
= -mult_adj
;
1523 interval
= -interval
;
1526 mult_adj
<<= adj_scale
;
1527 interval
<<= adj_scale
;
1528 offset
<<= adj_scale
;
1531 * So the following can be confusing.
1533 * To keep things simple, lets assume mult_adj == 1 for now.
1535 * When mult_adj != 1, remember that the interval and offset values
1536 * have been appropriately scaled so the math is the same.
1538 * The basic idea here is that we're increasing the multiplier
1539 * by one, this causes the xtime_interval to be incremented by
1540 * one cycle_interval. This is because:
1541 * xtime_interval = cycle_interval * mult
1542 * So if mult is being incremented by one:
1543 * xtime_interval = cycle_interval * (mult + 1)
1545 * xtime_interval = (cycle_interval * mult) + cycle_interval
1546 * Which can be shortened to:
1547 * xtime_interval += cycle_interval
1549 * So offset stores the non-accumulated cycles. Thus the current
1550 * time (in shifted nanoseconds) is:
1551 * now = (offset * adj) + xtime_nsec
1552 * Now, even though we're adjusting the clock frequency, we have
1553 * to keep time consistent. In other words, we can't jump back
1554 * in time, and we also want to avoid jumping forward in time.
1556 * So given the same offset value, we need the time to be the same
1557 * both before and after the freq adjustment.
1558 * now = (offset * adj_1) + xtime_nsec_1
1559 * now = (offset * adj_2) + xtime_nsec_2
1561 * (offset * adj_1) + xtime_nsec_1 =
1562 * (offset * adj_2) + xtime_nsec_2
1566 * (offset * adj_1) + xtime_nsec_1 =
1567 * (offset * (adj_1+1)) + xtime_nsec_2
1568 * (offset * adj_1) + xtime_nsec_1 =
1569 * (offset * adj_1) + offset + xtime_nsec_2
1570 * Canceling the sides:
1571 * xtime_nsec_1 = offset + xtime_nsec_2
1573 * xtime_nsec_2 = xtime_nsec_1 - offset
1574 * Which simplfies to:
1575 * xtime_nsec -= offset
1577 * XXX - TODO: Doc ntp_error calculation.
1579 if ((mult_adj
> 0) && (tk
->tkr_mono
.mult
+ mult_adj
< mult_adj
)) {
1580 /* NTP adjustment caused clocksource mult overflow */
1585 tk
->tkr_mono
.mult
+= mult_adj
;
1586 tk
->xtime_interval
+= interval
;
1587 tk
->tkr_mono
.xtime_nsec
-= offset
;
1588 tk
->ntp_error
-= (interval
- offset
) << tk
->ntp_error_shift
;
1592 * Calculate the multiplier adjustment needed to match the frequency
1595 static __always_inline
void timekeeping_freqadjust(struct timekeeper
*tk
,
1598 s64 interval
= tk
->cycle_interval
;
1599 s64 xinterval
= tk
->xtime_interval
;
1604 /* Remove any current error adj from freq calculation */
1605 if (tk
->ntp_err_mult
)
1606 xinterval
-= tk
->cycle_interval
;
1608 tk
->ntp_tick
= ntp_tick_length();
1610 /* Calculate current error per tick */
1611 tick_error
= ntp_tick_length() >> tk
->ntp_error_shift
;
1612 tick_error
-= (xinterval
+ tk
->xtime_remainder
);
1614 /* Don't worry about correcting it if its small */
1615 if (likely((tick_error
>= 0) && (tick_error
<= interval
)))
1618 /* preserve the direction of correction */
1619 negative
= (tick_error
< 0);
1621 /* Sort out the magnitude of the correction */
1622 tick_error
= abs(tick_error
);
1623 for (adj
= 0; tick_error
> interval
; adj
++)
1626 /* scale the corrections */
1627 timekeeping_apply_adjustment(tk
, offset
, negative
, adj
);
1631 * Adjust the timekeeper's multiplier to the correct frequency
1632 * and also to reduce the accumulated error value.
1634 static void timekeeping_adjust(struct timekeeper
*tk
, s64 offset
)
1636 /* Correct for the current frequency error */
1637 timekeeping_freqadjust(tk
, offset
);
1639 /* Next make a small adjustment to fix any cumulative error */
1640 if (!tk
->ntp_err_mult
&& (tk
->ntp_error
> 0)) {
1641 tk
->ntp_err_mult
= 1;
1642 timekeeping_apply_adjustment(tk
, offset
, 0, 0);
1643 } else if (tk
->ntp_err_mult
&& (tk
->ntp_error
<= 0)) {
1644 /* Undo any existing error adjustment */
1645 timekeeping_apply_adjustment(tk
, offset
, 1, 0);
1646 tk
->ntp_err_mult
= 0;
1649 if (unlikely(tk
->tkr_mono
.clock
->maxadj
&&
1650 (abs(tk
->tkr_mono
.mult
- tk
->tkr_mono
.clock
->mult
)
1651 > tk
->tkr_mono
.clock
->maxadj
))) {
1652 printk_once(KERN_WARNING
1653 "Adjusting %s more than 11%% (%ld vs %ld)\n",
1654 tk
->tkr_mono
.clock
->name
, (long)tk
->tkr_mono
.mult
,
1655 (long)tk
->tkr_mono
.clock
->mult
+ tk
->tkr_mono
.clock
->maxadj
);
1659 * It may be possible that when we entered this function, xtime_nsec
1660 * was very small. Further, if we're slightly speeding the clocksource
1661 * in the code above, its possible the required corrective factor to
1662 * xtime_nsec could cause it to underflow.
1664 * Now, since we already accumulated the second, cannot simply roll
1665 * the accumulated second back, since the NTP subsystem has been
1666 * notified via second_overflow. So instead we push xtime_nsec forward
1667 * by the amount we underflowed, and add that amount into the error.
1669 * We'll correct this error next time through this function, when
1670 * xtime_nsec is not as small.
1672 if (unlikely((s64
)tk
->tkr_mono
.xtime_nsec
< 0)) {
1673 s64 neg
= -(s64
)tk
->tkr_mono
.xtime_nsec
;
1674 tk
->tkr_mono
.xtime_nsec
= 0;
1675 tk
->ntp_error
+= neg
<< tk
->ntp_error_shift
;
1680 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1682 * Helper function that accumulates a the nsecs greater then a second
1683 * from the xtime_nsec field to the xtime_secs field.
1684 * It also calls into the NTP code to handle leapsecond processing.
1687 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper
*tk
)
1689 u64 nsecps
= (u64
)NSEC_PER_SEC
<< tk
->tkr_mono
.shift
;
1690 unsigned int clock_set
= 0;
1692 while (tk
->tkr_mono
.xtime_nsec
>= nsecps
) {
1695 tk
->tkr_mono
.xtime_nsec
-= nsecps
;
1698 /* Figure out if its a leap sec and apply if needed */
1699 leap
= second_overflow(tk
->xtime_sec
);
1700 if (unlikely(leap
)) {
1701 struct timespec64 ts
;
1703 tk
->xtime_sec
+= leap
;
1707 tk_set_wall_to_mono(tk
,
1708 timespec64_sub(tk
->wall_to_monotonic
, ts
));
1710 __timekeeping_set_tai_offset(tk
, tk
->tai_offset
- leap
);
1712 clock_set
= TK_CLOCK_WAS_SET
;
1719 * logarithmic_accumulation - shifted accumulation of cycles
1721 * This functions accumulates a shifted interval of cycles into
1722 * into a shifted interval nanoseconds. Allows for O(log) accumulation
1725 * Returns the unconsumed cycles.
1727 static cycle_t
logarithmic_accumulation(struct timekeeper
*tk
, cycle_t offset
,
1729 unsigned int *clock_set
)
1731 cycle_t interval
= tk
->cycle_interval
<< shift
;
1734 /* If the offset is smaller then a shifted interval, do nothing */
1735 if (offset
< interval
)
1738 /* Accumulate one shifted interval */
1740 tk
->tkr_mono
.cycle_last
+= interval
;
1741 tk
->tkr_raw
.cycle_last
+= interval
;
1743 tk
->tkr_mono
.xtime_nsec
+= tk
->xtime_interval
<< shift
;
1744 *clock_set
|= accumulate_nsecs_to_secs(tk
);
1746 /* Accumulate raw time */
1747 raw_nsecs
= (u64
)tk
->raw_interval
<< shift
;
1748 raw_nsecs
+= tk
->raw_time
.tv_nsec
;
1749 if (raw_nsecs
>= NSEC_PER_SEC
) {
1750 u64 raw_secs
= raw_nsecs
;
1751 raw_nsecs
= do_div(raw_secs
, NSEC_PER_SEC
);
1752 tk
->raw_time
.tv_sec
+= raw_secs
;
1754 tk
->raw_time
.tv_nsec
= raw_nsecs
;
1756 /* Accumulate error between NTP and clock interval */
1757 tk
->ntp_error
+= tk
->ntp_tick
<< shift
;
1758 tk
->ntp_error
-= (tk
->xtime_interval
+ tk
->xtime_remainder
) <<
1759 (tk
->ntp_error_shift
+ shift
);
1765 * update_wall_time - Uses the current clocksource to increment the wall time
1768 void update_wall_time(void)
1770 struct timekeeper
*real_tk
= &tk_core
.timekeeper
;
1771 struct timekeeper
*tk
= &shadow_timekeeper
;
1773 int shift
= 0, maxshift
;
1774 unsigned int clock_set
= 0;
1775 unsigned long flags
;
1777 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
1779 /* Make sure we're fully resumed: */
1780 if (unlikely(timekeeping_suspended
))
1783 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
1784 offset
= real_tk
->cycle_interval
;
1786 offset
= clocksource_delta(tk
->tkr_mono
.read(tk
->tkr_mono
.clock
),
1787 tk
->tkr_mono
.cycle_last
, tk
->tkr_mono
.mask
);
1790 /* Check if there's really nothing to do */
1791 if (offset
< real_tk
->cycle_interval
)
1794 /* Do some additional sanity checking */
1795 timekeeping_check_update(real_tk
, offset
);
1798 * With NO_HZ we may have to accumulate many cycle_intervals
1799 * (think "ticks") worth of time at once. To do this efficiently,
1800 * we calculate the largest doubling multiple of cycle_intervals
1801 * that is smaller than the offset. We then accumulate that
1802 * chunk in one go, and then try to consume the next smaller
1805 shift
= ilog2(offset
) - ilog2(tk
->cycle_interval
);
1806 shift
= max(0, shift
);
1807 /* Bound shift to one less than what overflows tick_length */
1808 maxshift
= (64 - (ilog2(ntp_tick_length())+1)) - 1;
1809 shift
= min(shift
, maxshift
);
1810 while (offset
>= tk
->cycle_interval
) {
1811 offset
= logarithmic_accumulation(tk
, offset
, shift
,
1813 if (offset
< tk
->cycle_interval
<<shift
)
1817 /* correct the clock when NTP error is too big */
1818 timekeeping_adjust(tk
, offset
);
1821 * XXX This can be killed once everyone converts
1822 * to the new update_vsyscall.
1824 old_vsyscall_fixup(tk
);
1827 * Finally, make sure that after the rounding
1828 * xtime_nsec isn't larger than NSEC_PER_SEC
1830 clock_set
|= accumulate_nsecs_to_secs(tk
);
1832 write_seqcount_begin(&tk_core
.seq
);
1834 * Update the real timekeeper.
1836 * We could avoid this memcpy by switching pointers, but that
1837 * requires changes to all other timekeeper usage sites as
1838 * well, i.e. move the timekeeper pointer getter into the
1839 * spinlocked/seqcount protected sections. And we trade this
1840 * memcpy under the tk_core.seq against one before we start
1843 memcpy(real_tk
, tk
, sizeof(*tk
));
1844 timekeeping_update(real_tk
, clock_set
);
1845 write_seqcount_end(&tk_core
.seq
);
1847 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1849 /* Have to call _delayed version, since in irq context*/
1850 clock_was_set_delayed();
1854 * getboottime64 - Return the real time of system boot.
1855 * @ts: pointer to the timespec64 to be set
1857 * Returns the wall-time of boot in a timespec64.
1859 * This is based on the wall_to_monotonic offset and the total suspend
1860 * time. Calls to settimeofday will affect the value returned (which
1861 * basically means that however wrong your real time clock is at boot time,
1862 * you get the right time here).
1864 void getboottime64(struct timespec64
*ts
)
1866 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1867 ktime_t t
= ktime_sub(tk
->offs_real
, tk
->offs_boot
);
1869 *ts
= ktime_to_timespec64(t
);
1871 EXPORT_SYMBOL_GPL(getboottime64
);
1873 unsigned long get_seconds(void)
1875 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1877 return tk
->xtime_sec
;
1879 EXPORT_SYMBOL(get_seconds
);
1881 struct timespec
__current_kernel_time(void)
1883 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1885 return timespec64_to_timespec(tk_xtime(tk
));
1888 struct timespec
current_kernel_time(void)
1890 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1891 struct timespec64 now
;
1895 seq
= read_seqcount_begin(&tk_core
.seq
);
1898 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
1900 return timespec64_to_timespec(now
);
1902 EXPORT_SYMBOL(current_kernel_time
);
1904 struct timespec64
get_monotonic_coarse64(void)
1906 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1907 struct timespec64 now
, mono
;
1911 seq
= read_seqcount_begin(&tk_core
.seq
);
1914 mono
= tk
->wall_to_monotonic
;
1915 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
1917 set_normalized_timespec64(&now
, now
.tv_sec
+ mono
.tv_sec
,
1918 now
.tv_nsec
+ mono
.tv_nsec
);
1924 * Must hold jiffies_lock
1926 void do_timer(unsigned long ticks
)
1928 jiffies_64
+= ticks
;
1929 calc_global_load(ticks
);
1933 * ktime_get_update_offsets_tick - hrtimer helper
1934 * @offs_real: pointer to storage for monotonic -> realtime offset
1935 * @offs_boot: pointer to storage for monotonic -> boottime offset
1936 * @offs_tai: pointer to storage for monotonic -> clock tai offset
1938 * Returns monotonic time at last tick and various offsets
1940 ktime_t
ktime_get_update_offsets_tick(ktime_t
*offs_real
, ktime_t
*offs_boot
,
1943 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1949 seq
= read_seqcount_begin(&tk_core
.seq
);
1951 base
= tk
->tkr_mono
.base
;
1952 nsecs
= tk
->tkr_mono
.xtime_nsec
>> tk
->tkr_mono
.shift
;
1954 *offs_real
= tk
->offs_real
;
1955 *offs_boot
= tk
->offs_boot
;
1956 *offs_tai
= tk
->offs_tai
;
1957 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
1959 return ktime_add_ns(base
, nsecs
);
1962 #ifdef CONFIG_HIGH_RES_TIMERS
1964 * ktime_get_update_offsets_now - hrtimer helper
1965 * @offs_real: pointer to storage for monotonic -> realtime offset
1966 * @offs_boot: pointer to storage for monotonic -> boottime offset
1967 * @offs_tai: pointer to storage for monotonic -> clock tai offset
1969 * Returns current monotonic time and updates the offsets
1970 * Called from hrtimer_interrupt() or retrigger_next_event()
1972 ktime_t
ktime_get_update_offsets_now(ktime_t
*offs_real
, ktime_t
*offs_boot
,
1975 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1981 seq
= read_seqcount_begin(&tk_core
.seq
);
1983 base
= tk
->tkr_mono
.base
;
1984 nsecs
= timekeeping_get_ns(&tk
->tkr_mono
);
1986 *offs_real
= tk
->offs_real
;
1987 *offs_boot
= tk
->offs_boot
;
1988 *offs_tai
= tk
->offs_tai
;
1989 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
1991 return ktime_add_ns(base
, nsecs
);
1996 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
1998 int do_adjtimex(struct timex
*txc
)
2000 struct timekeeper
*tk
= &tk_core
.timekeeper
;
2001 unsigned long flags
;
2002 struct timespec64 ts
;
2006 /* Validate the data before disabling interrupts */
2007 ret
= ntp_validate_timex(txc
);
2011 if (txc
->modes
& ADJ_SETOFFSET
) {
2012 struct timespec delta
;
2013 delta
.tv_sec
= txc
->time
.tv_sec
;
2014 delta
.tv_nsec
= txc
->time
.tv_usec
;
2015 if (!(txc
->modes
& ADJ_NANO
))
2016 delta
.tv_nsec
*= 1000;
2017 ret
= timekeeping_inject_offset(&delta
);
2022 getnstimeofday64(&ts
);
2024 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
2025 write_seqcount_begin(&tk_core
.seq
);
2027 orig_tai
= tai
= tk
->tai_offset
;
2028 ret
= __do_adjtimex(txc
, &ts
, &tai
);
2030 if (tai
!= orig_tai
) {
2031 __timekeeping_set_tai_offset(tk
, tai
);
2032 timekeeping_update(tk
, TK_MIRROR
| TK_CLOCK_WAS_SET
);
2034 write_seqcount_end(&tk_core
.seq
);
2035 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
2037 if (tai
!= orig_tai
)
2040 ntp_notify_cmos_timer();
2045 #ifdef CONFIG_NTP_PPS
2047 * hardpps() - Accessor function to NTP __hardpps function
2049 void hardpps(const struct timespec
*phase_ts
, const struct timespec
*raw_ts
)
2051 unsigned long flags
;
2053 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
2054 write_seqcount_begin(&tk_core
.seq
);
2056 __hardpps(phase_ts
, raw_ts
);
2058 write_seqcount_end(&tk_core
.seq
);
2059 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
2061 EXPORT_SYMBOL(hardpps
);
2065 * xtime_update() - advances the timekeeping infrastructure
2066 * @ticks: number of ticks, that have elapsed since the last call.
2068 * Must be called with interrupts disabled.
2070 void xtime_update(unsigned long ticks
)
2072 write_seqlock(&jiffies_lock
);
2074 write_sequnlock(&jiffies_lock
);