1 // SPDX-License-Identifier: GPL-2.0
3 * Kernel timekeeping code and accessor functions. Based on code from
4 * timer.c, moved in commit 8524070b7982.
6 #include <linux/timekeeper_internal.h>
7 #include <linux/module.h>
8 #include <linux/interrupt.h>
9 #include <linux/percpu.h>
10 #include <linux/init.h>
12 #include <linux/nmi.h>
13 #include <linux/sched.h>
14 #include <linux/sched/loadavg.h>
15 #include <linux/sched/clock.h>
16 #include <linux/syscore_ops.h>
17 #include <linux/clocksource.h>
18 #include <linux/jiffies.h>
19 #include <linux/time.h>
20 #include <linux/tick.h>
21 #include <linux/stop_machine.h>
22 #include <linux/pvclock_gtod.h>
23 #include <linux/compiler.h>
24 #include <linux/audit.h>
26 #include "tick-internal.h"
27 #include "ntp_internal.h"
28 #include "timekeeping_internal.h"
30 #define TK_CLEAR_NTP (1 << 0)
31 #define TK_MIRROR (1 << 1)
32 #define TK_CLOCK_WAS_SET (1 << 2)
34 enum timekeeping_adv_mode
{
35 /* Update timekeeper when a tick has passed */
38 /* Update timekeeper on a direct frequency change */
42 DEFINE_RAW_SPINLOCK(timekeeper_lock
);
45 * The most important data for readout fits into a single 64 byte
49 seqcount_raw_spinlock_t seq
;
50 struct timekeeper timekeeper
;
51 } tk_core ____cacheline_aligned
= {
52 .seq
= SEQCNT_RAW_SPINLOCK_ZERO(tk_core
.seq
, &timekeeper_lock
),
55 static struct timekeeper shadow_timekeeper
;
57 /* flag for if timekeeping is suspended */
58 int __read_mostly timekeeping_suspended
;
61 * struct tk_fast - NMI safe timekeeper
62 * @seq: Sequence counter for protecting updates. The lowest bit
63 * is the index for the tk_read_base array
64 * @base: tk_read_base array. Access is indexed by the lowest bit of
67 * See @update_fast_timekeeper() below.
71 struct tk_read_base base
[2];
74 /* Suspend-time cycles value for halted fast timekeeper. */
75 static u64 cycles_at_suspend
;
77 static u64
dummy_clock_read(struct clocksource
*cs
)
79 if (timekeeping_suspended
)
80 return cycles_at_suspend
;
84 static struct clocksource dummy_clock
= {
85 .read
= dummy_clock_read
,
89 * Boot time initialization which allows local_clock() to be utilized
90 * during early boot when clocksources are not available. local_clock()
91 * returns nanoseconds already so no conversion is required, hence mult=1
92 * and shift=0. When the first proper clocksource is installed then
93 * the fast time keepers are updated with the correct values.
95 #define FAST_TK_INIT \
97 .clock = &dummy_clock, \
98 .mask = CLOCKSOURCE_MASK(64), \
103 static struct tk_fast tk_fast_mono ____cacheline_aligned
= {
104 .seq
= SEQCNT_LATCH_ZERO(tk_fast_mono
.seq
),
105 .base
[0] = FAST_TK_INIT
,
106 .base
[1] = FAST_TK_INIT
,
109 static struct tk_fast tk_fast_raw ____cacheline_aligned
= {
110 .seq
= SEQCNT_LATCH_ZERO(tk_fast_raw
.seq
),
111 .base
[0] = FAST_TK_INIT
,
112 .base
[1] = FAST_TK_INIT
,
115 static inline void tk_normalize_xtime(struct timekeeper
*tk
)
117 while (tk
->tkr_mono
.xtime_nsec
>= ((u64
)NSEC_PER_SEC
<< tk
->tkr_mono
.shift
)) {
118 tk
->tkr_mono
.xtime_nsec
-= (u64
)NSEC_PER_SEC
<< tk
->tkr_mono
.shift
;
121 while (tk
->tkr_raw
.xtime_nsec
>= ((u64
)NSEC_PER_SEC
<< tk
->tkr_raw
.shift
)) {
122 tk
->tkr_raw
.xtime_nsec
-= (u64
)NSEC_PER_SEC
<< tk
->tkr_raw
.shift
;
127 static inline struct timespec64
tk_xtime(const struct timekeeper
*tk
)
129 struct timespec64 ts
;
131 ts
.tv_sec
= tk
->xtime_sec
;
132 ts
.tv_nsec
= (long)(tk
->tkr_mono
.xtime_nsec
>> tk
->tkr_mono
.shift
);
136 static void tk_set_xtime(struct timekeeper
*tk
, const struct timespec64
*ts
)
138 tk
->xtime_sec
= ts
->tv_sec
;
139 tk
->tkr_mono
.xtime_nsec
= (u64
)ts
->tv_nsec
<< tk
->tkr_mono
.shift
;
142 static void tk_xtime_add(struct timekeeper
*tk
, const struct timespec64
*ts
)
144 tk
->xtime_sec
+= ts
->tv_sec
;
145 tk
->tkr_mono
.xtime_nsec
+= (u64
)ts
->tv_nsec
<< tk
->tkr_mono
.shift
;
146 tk_normalize_xtime(tk
);
149 static void tk_set_wall_to_mono(struct timekeeper
*tk
, struct timespec64 wtm
)
151 struct timespec64 tmp
;
154 * Verify consistency of: offset_real = -wall_to_monotonic
155 * before modifying anything
157 set_normalized_timespec64(&tmp
, -tk
->wall_to_monotonic
.tv_sec
,
158 -tk
->wall_to_monotonic
.tv_nsec
);
159 WARN_ON_ONCE(tk
->offs_real
!= timespec64_to_ktime(tmp
));
160 tk
->wall_to_monotonic
= wtm
;
161 set_normalized_timespec64(&tmp
, -wtm
.tv_sec
, -wtm
.tv_nsec
);
162 tk
->offs_real
= timespec64_to_ktime(tmp
);
163 tk
->offs_tai
= ktime_add(tk
->offs_real
, ktime_set(tk
->tai_offset
, 0));
166 static inline void tk_update_sleep_time(struct timekeeper
*tk
, ktime_t delta
)
168 tk
->offs_boot
= ktime_add(tk
->offs_boot
, delta
);
170 * Timespec representation for VDSO update to avoid 64bit division
173 tk
->monotonic_to_boot
= ktime_to_timespec64(tk
->offs_boot
);
177 * tk_clock_read - atomic clocksource read() helper
179 * This helper is necessary to use in the read paths because, while the
180 * seqcount ensures we don't return a bad value while structures are updated,
181 * it doesn't protect from potential crashes. There is the possibility that
182 * the tkr's clocksource may change between the read reference, and the
183 * clock reference passed to the read function. This can cause crashes if
184 * the wrong clocksource is passed to the wrong read function.
185 * This isn't necessary to use when holding the timekeeper_lock or doing
186 * a read of the fast-timekeeper tkrs (which is protected by its own locking
189 static inline u64
tk_clock_read(const struct tk_read_base
*tkr
)
191 struct clocksource
*clock
= READ_ONCE(tkr
->clock
);
193 return clock
->read(clock
);
196 #ifdef CONFIG_DEBUG_TIMEKEEPING
197 #define WARNING_FREQ (HZ*300) /* 5 minute rate-limiting */
199 static void timekeeping_check_update(struct timekeeper
*tk
, u64 offset
)
202 u64 max_cycles
= tk
->tkr_mono
.clock
->max_cycles
;
203 const char *name
= tk
->tkr_mono
.clock
->name
;
205 if (offset
> max_cycles
) {
206 printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n",
207 offset
, name
, max_cycles
);
208 printk_deferred(" timekeeping: Your kernel is sick, but tries to cope by capping time updates\n");
210 if (offset
> (max_cycles
>> 1)) {
211 printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the '%s' clock's 50%% safety margin (%lld)\n",
212 offset
, name
, max_cycles
>> 1);
213 printk_deferred(" timekeeping: Your kernel is still fine, but is feeling a bit nervous\n");
217 if (tk
->underflow_seen
) {
218 if (jiffies
- tk
->last_warning
> WARNING_FREQ
) {
219 printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name
);
220 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
221 printk_deferred(" Your kernel is probably still fine.\n");
222 tk
->last_warning
= jiffies
;
224 tk
->underflow_seen
= 0;
227 if (tk
->overflow_seen
) {
228 if (jiffies
- tk
->last_warning
> WARNING_FREQ
) {
229 printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name
);
230 printk_deferred(" Please report this, consider using a different clocksource, if possible.\n");
231 printk_deferred(" Your kernel is probably still fine.\n");
232 tk
->last_warning
= jiffies
;
234 tk
->overflow_seen
= 0;
238 static inline u64
timekeeping_get_delta(const struct tk_read_base
*tkr
)
240 struct timekeeper
*tk
= &tk_core
.timekeeper
;
241 u64 now
, last
, mask
, max
, delta
;
245 * Since we're called holding a seqcount, the data may shift
246 * under us while we're doing the calculation. This can cause
247 * false positives, since we'd note a problem but throw the
248 * results away. So nest another seqcount here to atomically
249 * grab the points we are checking with.
252 seq
= read_seqcount_begin(&tk_core
.seq
);
253 now
= tk_clock_read(tkr
);
254 last
= tkr
->cycle_last
;
256 max
= tkr
->clock
->max_cycles
;
257 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
259 delta
= clocksource_delta(now
, last
, mask
);
262 * Try to catch underflows by checking if we are seeing small
263 * mask-relative negative values.
265 if (unlikely((~delta
& mask
) < (mask
>> 3))) {
266 tk
->underflow_seen
= 1;
270 /* Cap delta value to the max_cycles values to avoid mult overflows */
271 if (unlikely(delta
> max
)) {
272 tk
->overflow_seen
= 1;
273 delta
= tkr
->clock
->max_cycles
;
279 static inline void timekeeping_check_update(struct timekeeper
*tk
, u64 offset
)
282 static inline u64
timekeeping_get_delta(const struct tk_read_base
*tkr
)
284 u64 cycle_now
, delta
;
286 /* read clocksource */
287 cycle_now
= tk_clock_read(tkr
);
289 /* calculate the delta since the last update_wall_time */
290 delta
= clocksource_delta(cycle_now
, tkr
->cycle_last
, tkr
->mask
);
297 * tk_setup_internals - Set up internals to use clocksource clock.
299 * @tk: The target timekeeper to setup.
300 * @clock: Pointer to clocksource.
302 * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
303 * pair and interval request.
305 * Unless you're the timekeeping code, you should not be using this!
307 static void tk_setup_internals(struct timekeeper
*tk
, struct clocksource
*clock
)
310 u64 tmp
, ntpinterval
;
311 struct clocksource
*old_clock
;
313 ++tk
->cs_was_changed_seq
;
314 old_clock
= tk
->tkr_mono
.clock
;
315 tk
->tkr_mono
.clock
= clock
;
316 tk
->tkr_mono
.mask
= clock
->mask
;
317 tk
->tkr_mono
.cycle_last
= tk_clock_read(&tk
->tkr_mono
);
319 tk
->tkr_raw
.clock
= clock
;
320 tk
->tkr_raw
.mask
= clock
->mask
;
321 tk
->tkr_raw
.cycle_last
= tk
->tkr_mono
.cycle_last
;
323 /* Do the ns -> cycle conversion first, using original mult */
324 tmp
= NTP_INTERVAL_LENGTH
;
325 tmp
<<= clock
->shift
;
327 tmp
+= clock
->mult
/2;
328 do_div(tmp
, clock
->mult
);
332 interval
= (u64
) tmp
;
333 tk
->cycle_interval
= interval
;
335 /* Go back from cycles -> shifted ns */
336 tk
->xtime_interval
= interval
* clock
->mult
;
337 tk
->xtime_remainder
= ntpinterval
- tk
->xtime_interval
;
338 tk
->raw_interval
= interval
* clock
->mult
;
340 /* if changing clocks, convert xtime_nsec shift units */
342 int shift_change
= clock
->shift
- old_clock
->shift
;
343 if (shift_change
< 0) {
344 tk
->tkr_mono
.xtime_nsec
>>= -shift_change
;
345 tk
->tkr_raw
.xtime_nsec
>>= -shift_change
;
347 tk
->tkr_mono
.xtime_nsec
<<= shift_change
;
348 tk
->tkr_raw
.xtime_nsec
<<= shift_change
;
352 tk
->tkr_mono
.shift
= clock
->shift
;
353 tk
->tkr_raw
.shift
= clock
->shift
;
356 tk
->ntp_error_shift
= NTP_SCALE_SHIFT
- clock
->shift
;
357 tk
->ntp_tick
= ntpinterval
<< tk
->ntp_error_shift
;
360 * The timekeeper keeps its own mult values for the currently
361 * active clocksource. These value will be adjusted via NTP
362 * to counteract clock drifting.
364 tk
->tkr_mono
.mult
= clock
->mult
;
365 tk
->tkr_raw
.mult
= clock
->mult
;
366 tk
->ntp_err_mult
= 0;
367 tk
->skip_second_overflow
= 0;
370 /* Timekeeper helper functions. */
372 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
373 static u32
default_arch_gettimeoffset(void) { return 0; }
374 u32 (*arch_gettimeoffset
)(void) = default_arch_gettimeoffset
;
376 static inline u32
arch_gettimeoffset(void) { return 0; }
379 static inline u64
timekeeping_delta_to_ns(const struct tk_read_base
*tkr
, u64 delta
)
383 nsec
= delta
* tkr
->mult
+ tkr
->xtime_nsec
;
386 /* If arch requires, add in get_arch_timeoffset() */
387 return nsec
+ arch_gettimeoffset();
390 static inline u64
timekeeping_get_ns(const struct tk_read_base
*tkr
)
394 delta
= timekeeping_get_delta(tkr
);
395 return timekeeping_delta_to_ns(tkr
, delta
);
398 static inline u64
timekeeping_cycles_to_ns(const struct tk_read_base
*tkr
, u64 cycles
)
402 /* calculate the delta since the last update_wall_time */
403 delta
= clocksource_delta(cycles
, tkr
->cycle_last
, tkr
->mask
);
404 return timekeeping_delta_to_ns(tkr
, delta
);
408 * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
409 * @tkr: Timekeeping readout base from which we take the update
410 * @tkf: Pointer to NMI safe timekeeper
412 * We want to use this from any context including NMI and tracing /
413 * instrumenting the timekeeping code itself.
415 * Employ the latch technique; see @raw_write_seqcount_latch.
417 * So if a NMI hits the update of base[0] then it will use base[1]
418 * which is still consistent. In the worst case this can result is a
419 * slightly wrong timestamp (a few nanoseconds). See
420 * @ktime_get_mono_fast_ns.
422 static void update_fast_timekeeper(const struct tk_read_base
*tkr
,
425 struct tk_read_base
*base
= tkf
->base
;
427 /* Force readers off to base[1] */
428 raw_write_seqcount_latch(&tkf
->seq
);
431 memcpy(base
, tkr
, sizeof(*base
));
433 /* Force readers back to base[0] */
434 raw_write_seqcount_latch(&tkf
->seq
);
437 memcpy(base
+ 1, base
, sizeof(*base
));
440 static __always_inline u64
__ktime_get_fast_ns(struct tk_fast
*tkf
)
442 struct tk_read_base
*tkr
;
447 seq
= raw_read_seqcount_latch(&tkf
->seq
);
448 tkr
= tkf
->base
+ (seq
& 0x01);
449 now
= ktime_to_ns(tkr
->base
);
451 now
+= timekeeping_delta_to_ns(tkr
,
456 } while (read_seqcount_latch_retry(&tkf
->seq
, seq
));
462 * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
464 * This timestamp is not guaranteed to be monotonic across an update.
465 * The timestamp is calculated by:
467 * now = base_mono + clock_delta * slope
469 * So if the update lowers the slope, readers who are forced to the
470 * not yet updated second array are still using the old steeper slope.
479 * |12345678---> reader order
485 * So reader 6 will observe time going backwards versus reader 5.
487 * While other CPUs are likely to be able to observe that, the only way
488 * for a CPU local observation is when an NMI hits in the middle of
489 * the update. Timestamps taken from that NMI context might be ahead
490 * of the following timestamps. Callers need to be aware of that and
493 u64
ktime_get_mono_fast_ns(void)
495 return __ktime_get_fast_ns(&tk_fast_mono
);
497 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns
);
500 * ktime_get_raw_fast_ns - Fast NMI safe access to clock monotonic raw
502 * Contrary to ktime_get_mono_fast_ns() this is always correct because the
503 * conversion factor is not affected by NTP/PTP correction.
505 u64
ktime_get_raw_fast_ns(void)
507 return __ktime_get_fast_ns(&tk_fast_raw
);
509 EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns
);
512 * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock.
514 * To keep it NMI safe since we're accessing from tracing, we're not using a
515 * separate timekeeper with updates to monotonic clock and boot offset
516 * protected with seqcounts. This has the following minor side effects:
518 * (1) Its possible that a timestamp be taken after the boot offset is updated
519 * but before the timekeeper is updated. If this happens, the new boot offset
520 * is added to the old timekeeping making the clock appear to update slightly
523 * timekeeping_inject_sleeptime64()
524 * __timekeeping_inject_sleeptime(tk, delta);
526 * timekeeping_update(tk, TK_CLEAR_NTP...);
528 * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be
529 * partially updated. Since the tk->offs_boot update is a rare event, this
530 * should be a rare occurrence which postprocessing should be able to handle.
532 * The caveats vs. timestamp ordering as documented for ktime_get_fast_ns()
535 u64 notrace
ktime_get_boot_fast_ns(void)
537 struct timekeeper
*tk
= &tk_core
.timekeeper
;
539 return (ktime_get_mono_fast_ns() + ktime_to_ns(tk
->offs_boot
));
541 EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns
);
543 static __always_inline u64
__ktime_get_real_fast(struct tk_fast
*tkf
, u64
*mono
)
545 struct tk_read_base
*tkr
;
546 u64 basem
, baser
, delta
;
550 seq
= raw_read_seqcount_latch(&tkf
->seq
);
551 tkr
= tkf
->base
+ (seq
& 0x01);
552 basem
= ktime_to_ns(tkr
->base
);
553 baser
= ktime_to_ns(tkr
->base_real
);
555 delta
= timekeeping_delta_to_ns(tkr
,
556 clocksource_delta(tk_clock_read(tkr
),
557 tkr
->cycle_last
, tkr
->mask
));
558 } while (read_seqcount_latch_retry(&tkf
->seq
, seq
));
561 *mono
= basem
+ delta
;
562 return baser
+ delta
;
566 * ktime_get_real_fast_ns: - NMI safe and fast access to clock realtime.
568 * See ktime_get_fast_ns() for documentation of the time stamp ordering.
570 u64
ktime_get_real_fast_ns(void)
572 return __ktime_get_real_fast(&tk_fast_mono
, NULL
);
574 EXPORT_SYMBOL_GPL(ktime_get_real_fast_ns
);
577 * ktime_get_fast_timestamps: - NMI safe timestamps
578 * @snapshot: Pointer to timestamp storage
580 * Stores clock monotonic, boottime and realtime timestamps.
582 * Boot time is a racy access on 32bit systems if the sleep time injection
583 * happens late during resume and not in timekeeping_resume(). That could
584 * be avoided by expanding struct tk_read_base with boot offset for 32bit
585 * and adding more overhead to the update. As this is a hard to observe
586 * once per resume event which can be filtered with reasonable effort using
587 * the accurate mono/real timestamps, it's probably not worth the trouble.
589 * Aside of that it might be possible on 32 and 64 bit to observe the
590 * following when the sleep time injection happens late:
593 * timekeeping_resume()
594 * ktime_get_fast_timestamps()
595 * mono, real = __ktime_get_real_fast()
596 * inject_sleep_time()
598 * boot = mono + bootoffset;
600 * That means that boot time already has the sleep time adjustment, but
601 * real time does not. On the next readout both are in sync again.
603 * Preventing this for 64bit is not really feasible without destroying the
604 * careful cache layout of the timekeeper because the sequence count and
605 * struct tk_read_base would then need two cache lines instead of one.
607 * Access to the time keeper clock source is disabled accross the innermost
608 * steps of suspend/resume. The accessors still work, but the timestamps
609 * are frozen until time keeping is resumed which happens very early.
611 * For regular suspend/resume there is no observable difference vs. sched
612 * clock, but it might affect some of the nasty low level debug printks.
614 * OTOH, access to sched clock is not guaranteed accross suspend/resume on
615 * all systems either so it depends on the hardware in use.
617 * If that turns out to be a real problem then this could be mitigated by
618 * using sched clock in a similar way as during early boot. But it's not as
619 * trivial as on early boot because it needs some careful protection
620 * against the clock monotonic timestamp jumping backwards on resume.
622 void ktime_get_fast_timestamps(struct ktime_timestamps
*snapshot
)
624 struct timekeeper
*tk
= &tk_core
.timekeeper
;
626 snapshot
->real
= __ktime_get_real_fast(&tk_fast_mono
, &snapshot
->mono
);
627 snapshot
->boot
= snapshot
->mono
+ ktime_to_ns(data_race(tk
->offs_boot
));
631 * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
632 * @tk: Timekeeper to snapshot.
634 * It generally is unsafe to access the clocksource after timekeeping has been
635 * suspended, so take a snapshot of the readout base of @tk and use it as the
636 * fast timekeeper's readout base while suspended. It will return the same
637 * number of cycles every time until timekeeping is resumed at which time the
638 * proper readout base for the fast timekeeper will be restored automatically.
640 static void halt_fast_timekeeper(const struct timekeeper
*tk
)
642 static struct tk_read_base tkr_dummy
;
643 const struct tk_read_base
*tkr
= &tk
->tkr_mono
;
645 memcpy(&tkr_dummy
, tkr
, sizeof(tkr_dummy
));
646 cycles_at_suspend
= tk_clock_read(tkr
);
647 tkr_dummy
.clock
= &dummy_clock
;
648 tkr_dummy
.base_real
= tkr
->base
+ tk
->offs_real
;
649 update_fast_timekeeper(&tkr_dummy
, &tk_fast_mono
);
652 memcpy(&tkr_dummy
, tkr
, sizeof(tkr_dummy
));
653 tkr_dummy
.clock
= &dummy_clock
;
654 update_fast_timekeeper(&tkr_dummy
, &tk_fast_raw
);
657 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain
);
659 static void update_pvclock_gtod(struct timekeeper
*tk
, bool was_set
)
661 raw_notifier_call_chain(&pvclock_gtod_chain
, was_set
, tk
);
665 * pvclock_gtod_register_notifier - register a pvclock timedata update listener
666 * @nb: Pointer to the notifier block to register
668 int pvclock_gtod_register_notifier(struct notifier_block
*nb
)
670 struct timekeeper
*tk
= &tk_core
.timekeeper
;
674 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
675 ret
= raw_notifier_chain_register(&pvclock_gtod_chain
, nb
);
676 update_pvclock_gtod(tk
, true);
677 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
681 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier
);
684 * pvclock_gtod_unregister_notifier - unregister a pvclock
685 * timedata update listener
686 * @nb: Pointer to the notifier block to unregister
688 int pvclock_gtod_unregister_notifier(struct notifier_block
*nb
)
693 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
694 ret
= raw_notifier_chain_unregister(&pvclock_gtod_chain
, nb
);
695 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
699 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier
);
702 * tk_update_leap_state - helper to update the next_leap_ktime
704 static inline void tk_update_leap_state(struct timekeeper
*tk
)
706 tk
->next_leap_ktime
= ntp_get_next_leap();
707 if (tk
->next_leap_ktime
!= KTIME_MAX
)
708 /* Convert to monotonic time */
709 tk
->next_leap_ktime
= ktime_sub(tk
->next_leap_ktime
, tk
->offs_real
);
713 * Update the ktime_t based scalar nsec members of the timekeeper
715 static inline void tk_update_ktime_data(struct timekeeper
*tk
)
721 * The xtime based monotonic readout is:
722 * nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
723 * The ktime based monotonic readout is:
724 * nsec = base_mono + now();
725 * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
727 seconds
= (u64
)(tk
->xtime_sec
+ tk
->wall_to_monotonic
.tv_sec
);
728 nsec
= (u32
) tk
->wall_to_monotonic
.tv_nsec
;
729 tk
->tkr_mono
.base
= ns_to_ktime(seconds
* NSEC_PER_SEC
+ nsec
);
732 * The sum of the nanoseconds portions of xtime and
733 * wall_to_monotonic can be greater/equal one second. Take
734 * this into account before updating tk->ktime_sec.
736 nsec
+= (u32
)(tk
->tkr_mono
.xtime_nsec
>> tk
->tkr_mono
.shift
);
737 if (nsec
>= NSEC_PER_SEC
)
739 tk
->ktime_sec
= seconds
;
741 /* Update the monotonic raw base */
742 tk
->tkr_raw
.base
= ns_to_ktime(tk
->raw_sec
* NSEC_PER_SEC
);
745 /* must hold timekeeper_lock */
746 static void timekeeping_update(struct timekeeper
*tk
, unsigned int action
)
748 if (action
& TK_CLEAR_NTP
) {
753 tk_update_leap_state(tk
);
754 tk_update_ktime_data(tk
);
757 update_pvclock_gtod(tk
, action
& TK_CLOCK_WAS_SET
);
759 tk
->tkr_mono
.base_real
= tk
->tkr_mono
.base
+ tk
->offs_real
;
760 update_fast_timekeeper(&tk
->tkr_mono
, &tk_fast_mono
);
761 update_fast_timekeeper(&tk
->tkr_raw
, &tk_fast_raw
);
763 if (action
& TK_CLOCK_WAS_SET
)
764 tk
->clock_was_set_seq
++;
766 * The mirroring of the data to the shadow-timekeeper needs
767 * to happen last here to ensure we don't over-write the
768 * timekeeper structure on the next update with stale data
770 if (action
& TK_MIRROR
)
771 memcpy(&shadow_timekeeper
, &tk_core
.timekeeper
,
772 sizeof(tk_core
.timekeeper
));
776 * timekeeping_forward_now - update clock to the current time
777 * @tk: Pointer to the timekeeper to update
779 * Forward the current clock to update its state since the last call to
780 * update_wall_time(). This is useful before significant clock changes,
781 * as it avoids having to deal with this time offset explicitly.
783 static void timekeeping_forward_now(struct timekeeper
*tk
)
785 u64 cycle_now
, delta
;
787 cycle_now
= tk_clock_read(&tk
->tkr_mono
);
788 delta
= clocksource_delta(cycle_now
, tk
->tkr_mono
.cycle_last
, tk
->tkr_mono
.mask
);
789 tk
->tkr_mono
.cycle_last
= cycle_now
;
790 tk
->tkr_raw
.cycle_last
= cycle_now
;
792 tk
->tkr_mono
.xtime_nsec
+= delta
* tk
->tkr_mono
.mult
;
794 /* If arch requires, add in get_arch_timeoffset() */
795 tk
->tkr_mono
.xtime_nsec
+= (u64
)arch_gettimeoffset() << tk
->tkr_mono
.shift
;
798 tk
->tkr_raw
.xtime_nsec
+= delta
* tk
->tkr_raw
.mult
;
800 /* If arch requires, add in get_arch_timeoffset() */
801 tk
->tkr_raw
.xtime_nsec
+= (u64
)arch_gettimeoffset() << tk
->tkr_raw
.shift
;
803 tk_normalize_xtime(tk
);
807 * ktime_get_real_ts64 - Returns the time of day in a timespec64.
808 * @ts: pointer to the timespec to be set
810 * Returns the time of day in a timespec64 (WARN if suspended).
812 void ktime_get_real_ts64(struct timespec64
*ts
)
814 struct timekeeper
*tk
= &tk_core
.timekeeper
;
818 WARN_ON(timekeeping_suspended
);
821 seq
= read_seqcount_begin(&tk_core
.seq
);
823 ts
->tv_sec
= tk
->xtime_sec
;
824 nsecs
= timekeeping_get_ns(&tk
->tkr_mono
);
826 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
829 timespec64_add_ns(ts
, nsecs
);
831 EXPORT_SYMBOL(ktime_get_real_ts64
);
833 ktime_t
ktime_get(void)
835 struct timekeeper
*tk
= &tk_core
.timekeeper
;
840 WARN_ON(timekeeping_suspended
);
843 seq
= read_seqcount_begin(&tk_core
.seq
);
844 base
= tk
->tkr_mono
.base
;
845 nsecs
= timekeeping_get_ns(&tk
->tkr_mono
);
847 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
849 return ktime_add_ns(base
, nsecs
);
851 EXPORT_SYMBOL_GPL(ktime_get
);
853 u32
ktime_get_resolution_ns(void)
855 struct timekeeper
*tk
= &tk_core
.timekeeper
;
859 WARN_ON(timekeeping_suspended
);
862 seq
= read_seqcount_begin(&tk_core
.seq
);
863 nsecs
= tk
->tkr_mono
.mult
>> tk
->tkr_mono
.shift
;
864 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
868 EXPORT_SYMBOL_GPL(ktime_get_resolution_ns
);
870 static ktime_t
*offsets
[TK_OFFS_MAX
] = {
871 [TK_OFFS_REAL
] = &tk_core
.timekeeper
.offs_real
,
872 [TK_OFFS_BOOT
] = &tk_core
.timekeeper
.offs_boot
,
873 [TK_OFFS_TAI
] = &tk_core
.timekeeper
.offs_tai
,
876 ktime_t
ktime_get_with_offset(enum tk_offsets offs
)
878 struct timekeeper
*tk
= &tk_core
.timekeeper
;
880 ktime_t base
, *offset
= offsets
[offs
];
883 WARN_ON(timekeeping_suspended
);
886 seq
= read_seqcount_begin(&tk_core
.seq
);
887 base
= ktime_add(tk
->tkr_mono
.base
, *offset
);
888 nsecs
= timekeeping_get_ns(&tk
->tkr_mono
);
890 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
892 return ktime_add_ns(base
, nsecs
);
895 EXPORT_SYMBOL_GPL(ktime_get_with_offset
);
897 ktime_t
ktime_get_coarse_with_offset(enum tk_offsets offs
)
899 struct timekeeper
*tk
= &tk_core
.timekeeper
;
901 ktime_t base
, *offset
= offsets
[offs
];
904 WARN_ON(timekeeping_suspended
);
907 seq
= read_seqcount_begin(&tk_core
.seq
);
908 base
= ktime_add(tk
->tkr_mono
.base
, *offset
);
909 nsecs
= tk
->tkr_mono
.xtime_nsec
>> tk
->tkr_mono
.shift
;
911 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
913 return ktime_add_ns(base
, nsecs
);
915 EXPORT_SYMBOL_GPL(ktime_get_coarse_with_offset
);
918 * ktime_mono_to_any() - convert mononotic time to any other time
919 * @tmono: time to convert.
920 * @offs: which offset to use
922 ktime_t
ktime_mono_to_any(ktime_t tmono
, enum tk_offsets offs
)
924 ktime_t
*offset
= offsets
[offs
];
929 seq
= read_seqcount_begin(&tk_core
.seq
);
930 tconv
= ktime_add(tmono
, *offset
);
931 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
935 EXPORT_SYMBOL_GPL(ktime_mono_to_any
);
938 * ktime_get_raw - Returns the raw monotonic time in ktime_t format
940 ktime_t
ktime_get_raw(void)
942 struct timekeeper
*tk
= &tk_core
.timekeeper
;
948 seq
= read_seqcount_begin(&tk_core
.seq
);
949 base
= tk
->tkr_raw
.base
;
950 nsecs
= timekeeping_get_ns(&tk
->tkr_raw
);
952 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
954 return ktime_add_ns(base
, nsecs
);
956 EXPORT_SYMBOL_GPL(ktime_get_raw
);
959 * ktime_get_ts64 - get the monotonic clock in timespec64 format
960 * @ts: pointer to timespec variable
962 * The function calculates the monotonic clock from the realtime
963 * clock and the wall_to_monotonic offset and stores the result
964 * in normalized timespec64 format in the variable pointed to by @ts.
966 void ktime_get_ts64(struct timespec64
*ts
)
968 struct timekeeper
*tk
= &tk_core
.timekeeper
;
969 struct timespec64 tomono
;
973 WARN_ON(timekeeping_suspended
);
976 seq
= read_seqcount_begin(&tk_core
.seq
);
977 ts
->tv_sec
= tk
->xtime_sec
;
978 nsec
= timekeeping_get_ns(&tk
->tkr_mono
);
979 tomono
= tk
->wall_to_monotonic
;
981 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
983 ts
->tv_sec
+= tomono
.tv_sec
;
985 timespec64_add_ns(ts
, nsec
+ tomono
.tv_nsec
);
987 EXPORT_SYMBOL_GPL(ktime_get_ts64
);
990 * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
992 * Returns the seconds portion of CLOCK_MONOTONIC with a single non
993 * serialized read. tk->ktime_sec is of type 'unsigned long' so this
994 * works on both 32 and 64 bit systems. On 32 bit systems the readout
995 * covers ~136 years of uptime which should be enough to prevent
996 * premature wrap arounds.
998 time64_t
ktime_get_seconds(void)
1000 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1002 WARN_ON(timekeeping_suspended
);
1003 return tk
->ktime_sec
;
1005 EXPORT_SYMBOL_GPL(ktime_get_seconds
);
1008 * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
1010 * Returns the wall clock seconds since 1970. This replaces the
1011 * get_seconds() interface which is not y2038 safe on 32bit systems.
1013 * For 64bit systems the fast access to tk->xtime_sec is preserved. On
1014 * 32bit systems the access must be protected with the sequence
1015 * counter to provide "atomic" access to the 64bit tk->xtime_sec
1018 time64_t
ktime_get_real_seconds(void)
1020 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1024 if (IS_ENABLED(CONFIG_64BIT
))
1025 return tk
->xtime_sec
;
1028 seq
= read_seqcount_begin(&tk_core
.seq
);
1029 seconds
= tk
->xtime_sec
;
1031 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
1035 EXPORT_SYMBOL_GPL(ktime_get_real_seconds
);
1038 * __ktime_get_real_seconds - The same as ktime_get_real_seconds
1039 * but without the sequence counter protect. This internal function
1040 * is called just when timekeeping lock is already held.
1042 noinstr time64_t
__ktime_get_real_seconds(void)
1044 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1046 return tk
->xtime_sec
;
1050 * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter
1051 * @systime_snapshot: pointer to struct receiving the system time snapshot
1053 void ktime_get_snapshot(struct system_time_snapshot
*systime_snapshot
)
1055 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1063 WARN_ON_ONCE(timekeeping_suspended
);
1066 seq
= read_seqcount_begin(&tk_core
.seq
);
1067 now
= tk_clock_read(&tk
->tkr_mono
);
1068 systime_snapshot
->cs_was_changed_seq
= tk
->cs_was_changed_seq
;
1069 systime_snapshot
->clock_was_set_seq
= tk
->clock_was_set_seq
;
1070 base_real
= ktime_add(tk
->tkr_mono
.base
,
1071 tk_core
.timekeeper
.offs_real
);
1072 base_raw
= tk
->tkr_raw
.base
;
1073 nsec_real
= timekeeping_cycles_to_ns(&tk
->tkr_mono
, now
);
1074 nsec_raw
= timekeeping_cycles_to_ns(&tk
->tkr_raw
, now
);
1075 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
1077 systime_snapshot
->cycles
= now
;
1078 systime_snapshot
->real
= ktime_add_ns(base_real
, nsec_real
);
1079 systime_snapshot
->raw
= ktime_add_ns(base_raw
, nsec_raw
);
1081 EXPORT_SYMBOL_GPL(ktime_get_snapshot
);
1083 /* Scale base by mult/div checking for overflow */
1084 static int scale64_check_overflow(u64 mult
, u64 div
, u64
*base
)
1088 tmp
= div64_u64_rem(*base
, div
, &rem
);
1090 if (((int)sizeof(u64
)*8 - fls64(mult
) < fls64(tmp
)) ||
1091 ((int)sizeof(u64
)*8 - fls64(mult
) < fls64(rem
)))
1095 rem
= div64_u64(rem
* mult
, div
);
1101 * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval
1102 * @history: Snapshot representing start of history
1103 * @partial_history_cycles: Cycle offset into history (fractional part)
1104 * @total_history_cycles: Total history length in cycles
1105 * @discontinuity: True indicates clock was set on history period
1106 * @ts: Cross timestamp that should be adjusted using
1107 * partial/total ratio
1109 * Helper function used by get_device_system_crosststamp() to correct the
1110 * crosstimestamp corresponding to the start of the current interval to the
1111 * system counter value (timestamp point) provided by the driver. The
1112 * total_history_* quantities are the total history starting at the provided
1113 * reference point and ending at the start of the current interval. The cycle
1114 * count between the driver timestamp point and the start of the current
1115 * interval is partial_history_cycles.
1117 static int adjust_historical_crosststamp(struct system_time_snapshot
*history
,
1118 u64 partial_history_cycles
,
1119 u64 total_history_cycles
,
1121 struct system_device_crosststamp
*ts
)
1123 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1124 u64 corr_raw
, corr_real
;
1125 bool interp_forward
;
1128 if (total_history_cycles
== 0 || partial_history_cycles
== 0)
1131 /* Interpolate shortest distance from beginning or end of history */
1132 interp_forward
= partial_history_cycles
> total_history_cycles
/ 2;
1133 partial_history_cycles
= interp_forward
?
1134 total_history_cycles
- partial_history_cycles
:
1135 partial_history_cycles
;
1138 * Scale the monotonic raw time delta by:
1139 * partial_history_cycles / total_history_cycles
1141 corr_raw
= (u64
)ktime_to_ns(
1142 ktime_sub(ts
->sys_monoraw
, history
->raw
));
1143 ret
= scale64_check_overflow(partial_history_cycles
,
1144 total_history_cycles
, &corr_raw
);
1149 * If there is a discontinuity in the history, scale monotonic raw
1151 * mult(real)/mult(raw) yielding the realtime correction
1152 * Otherwise, calculate the realtime correction similar to monotonic
1155 if (discontinuity
) {
1156 corr_real
= mul_u64_u32_div
1157 (corr_raw
, tk
->tkr_mono
.mult
, tk
->tkr_raw
.mult
);
1159 corr_real
= (u64
)ktime_to_ns(
1160 ktime_sub(ts
->sys_realtime
, history
->real
));
1161 ret
= scale64_check_overflow(partial_history_cycles
,
1162 total_history_cycles
, &corr_real
);
1167 /* Fixup monotonic raw and real time time values */
1168 if (interp_forward
) {
1169 ts
->sys_monoraw
= ktime_add_ns(history
->raw
, corr_raw
);
1170 ts
->sys_realtime
= ktime_add_ns(history
->real
, corr_real
);
1172 ts
->sys_monoraw
= ktime_sub_ns(ts
->sys_monoraw
, corr_raw
);
1173 ts
->sys_realtime
= ktime_sub_ns(ts
->sys_realtime
, corr_real
);
1180 * cycle_between - true if test occurs chronologically between before and after
1182 static bool cycle_between(u64 before
, u64 test
, u64 after
)
1184 if (test
> before
&& test
< after
)
1186 if (test
< before
&& before
> after
)
1192 * get_device_system_crosststamp - Synchronously capture system/device timestamp
1193 * @get_time_fn: Callback to get simultaneous device time and
1194 * system counter from the device driver
1195 * @ctx: Context passed to get_time_fn()
1196 * @history_begin: Historical reference point used to interpolate system
1197 * time when counter provided by the driver is before the current interval
1198 * @xtstamp: Receives simultaneously captured system and device time
1200 * Reads a timestamp from a device and correlates it to system time
1202 int get_device_system_crosststamp(int (*get_time_fn
)
1203 (ktime_t
*device_time
,
1204 struct system_counterval_t
*sys_counterval
,
1207 struct system_time_snapshot
*history_begin
,
1208 struct system_device_crosststamp
*xtstamp
)
1210 struct system_counterval_t system_counterval
;
1211 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1212 u64 cycles
, now
, interval_start
;
1213 unsigned int clock_was_set_seq
= 0;
1214 ktime_t base_real
, base_raw
;
1215 u64 nsec_real
, nsec_raw
;
1216 u8 cs_was_changed_seq
;
1222 seq
= read_seqcount_begin(&tk_core
.seq
);
1224 * Try to synchronously capture device time and a system
1225 * counter value calling back into the device driver
1227 ret
= get_time_fn(&xtstamp
->device
, &system_counterval
, ctx
);
1232 * Verify that the clocksource associated with the captured
1233 * system counter value is the same as the currently installed
1234 * timekeeper clocksource
1236 if (tk
->tkr_mono
.clock
!= system_counterval
.cs
)
1238 cycles
= system_counterval
.cycles
;
1241 * Check whether the system counter value provided by the
1242 * device driver is on the current timekeeping interval.
1244 now
= tk_clock_read(&tk
->tkr_mono
);
1245 interval_start
= tk
->tkr_mono
.cycle_last
;
1246 if (!cycle_between(interval_start
, cycles
, now
)) {
1247 clock_was_set_seq
= tk
->clock_was_set_seq
;
1248 cs_was_changed_seq
= tk
->cs_was_changed_seq
;
1249 cycles
= interval_start
;
1255 base_real
= ktime_add(tk
->tkr_mono
.base
,
1256 tk_core
.timekeeper
.offs_real
);
1257 base_raw
= tk
->tkr_raw
.base
;
1259 nsec_real
= timekeeping_cycles_to_ns(&tk
->tkr_mono
,
1260 system_counterval
.cycles
);
1261 nsec_raw
= timekeeping_cycles_to_ns(&tk
->tkr_raw
,
1262 system_counterval
.cycles
);
1263 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
1265 xtstamp
->sys_realtime
= ktime_add_ns(base_real
, nsec_real
);
1266 xtstamp
->sys_monoraw
= ktime_add_ns(base_raw
, nsec_raw
);
1269 * Interpolate if necessary, adjusting back from the start of the
1273 u64 partial_history_cycles
, total_history_cycles
;
1277 * Check that the counter value occurs after the provided
1278 * history reference and that the history doesn't cross a
1279 * clocksource change
1281 if (!history_begin
||
1282 !cycle_between(history_begin
->cycles
,
1283 system_counterval
.cycles
, cycles
) ||
1284 history_begin
->cs_was_changed_seq
!= cs_was_changed_seq
)
1286 partial_history_cycles
= cycles
- system_counterval
.cycles
;
1287 total_history_cycles
= cycles
- history_begin
->cycles
;
1289 history_begin
->clock_was_set_seq
!= clock_was_set_seq
;
1291 ret
= adjust_historical_crosststamp(history_begin
,
1292 partial_history_cycles
,
1293 total_history_cycles
,
1294 discontinuity
, xtstamp
);
1301 EXPORT_SYMBOL_GPL(get_device_system_crosststamp
);
1304 * do_settimeofday64 - Sets the time of day.
1305 * @ts: pointer to the timespec64 variable containing the new time
1307 * Sets the time of day to the new time and update NTP and notify hrtimers
1309 int do_settimeofday64(const struct timespec64
*ts
)
1311 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1312 struct timespec64 ts_delta
, xt
;
1313 unsigned long flags
;
1316 if (!timespec64_valid_settod(ts
))
1319 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
1320 write_seqcount_begin(&tk_core
.seq
);
1322 timekeeping_forward_now(tk
);
1325 ts_delta
.tv_sec
= ts
->tv_sec
- xt
.tv_sec
;
1326 ts_delta
.tv_nsec
= ts
->tv_nsec
- xt
.tv_nsec
;
1328 if (timespec64_compare(&tk
->wall_to_monotonic
, &ts_delta
) > 0) {
1333 tk_set_wall_to_mono(tk
, timespec64_sub(tk
->wall_to_monotonic
, ts_delta
));
1335 tk_set_xtime(tk
, ts
);
1337 timekeeping_update(tk
, TK_CLEAR_NTP
| TK_MIRROR
| TK_CLOCK_WAS_SET
);
1339 write_seqcount_end(&tk_core
.seq
);
1340 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1342 /* signal hrtimers about time change */
1346 audit_tk_injoffset(ts_delta
);
1350 EXPORT_SYMBOL(do_settimeofday64
);
1353 * timekeeping_inject_offset - Adds or subtracts from the current time.
1354 * @ts: Pointer to the timespec variable containing the offset
1356 * Adds or subtracts an offset value from the current time.
1358 static int timekeeping_inject_offset(const struct timespec64
*ts
)
1360 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1361 unsigned long flags
;
1362 struct timespec64 tmp
;
1365 if (ts
->tv_nsec
< 0 || ts
->tv_nsec
>= NSEC_PER_SEC
)
1368 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
1369 write_seqcount_begin(&tk_core
.seq
);
1371 timekeeping_forward_now(tk
);
1373 /* Make sure the proposed value is valid */
1374 tmp
= timespec64_add(tk_xtime(tk
), *ts
);
1375 if (timespec64_compare(&tk
->wall_to_monotonic
, ts
) > 0 ||
1376 !timespec64_valid_settod(&tmp
)) {
1381 tk_xtime_add(tk
, ts
);
1382 tk_set_wall_to_mono(tk
, timespec64_sub(tk
->wall_to_monotonic
, *ts
));
1384 error
: /* even if we error out, we forwarded the time, so call update */
1385 timekeeping_update(tk
, TK_CLEAR_NTP
| TK_MIRROR
| TK_CLOCK_WAS_SET
);
1387 write_seqcount_end(&tk_core
.seq
);
1388 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1390 /* signal hrtimers about time change */
1397 * Indicates if there is an offset between the system clock and the hardware
1398 * clock/persistent clock/rtc.
1400 int persistent_clock_is_local
;
1403 * Adjust the time obtained from the CMOS to be UTC time instead of
1406 * This is ugly, but preferable to the alternatives. Otherwise we
1407 * would either need to write a program to do it in /etc/rc (and risk
1408 * confusion if the program gets run more than once; it would also be
1409 * hard to make the program warp the clock precisely n hours) or
1410 * compile in the timezone information into the kernel. Bad, bad....
1414 * The best thing to do is to keep the CMOS clock in universal time (UTC)
1415 * as real UNIX machines always do it. This avoids all headaches about
1416 * daylight saving times and warping kernel clocks.
1418 void timekeeping_warp_clock(void)
1420 if (sys_tz
.tz_minuteswest
!= 0) {
1421 struct timespec64 adjust
;
1423 persistent_clock_is_local
= 1;
1424 adjust
.tv_sec
= sys_tz
.tz_minuteswest
* 60;
1426 timekeeping_inject_offset(&adjust
);
1431 * __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic
1433 static void __timekeeping_set_tai_offset(struct timekeeper
*tk
, s32 tai_offset
)
1435 tk
->tai_offset
= tai_offset
;
1436 tk
->offs_tai
= ktime_add(tk
->offs_real
, ktime_set(tai_offset
, 0));
1440 * change_clocksource - Swaps clocksources if a new one is available
1442 * Accumulates current time interval and initializes new clocksource
1444 static int change_clocksource(void *data
)
1446 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1447 struct clocksource
*new, *old
;
1448 unsigned long flags
;
1450 new = (struct clocksource
*) data
;
1452 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
1453 write_seqcount_begin(&tk_core
.seq
);
1455 timekeeping_forward_now(tk
);
1457 * If the cs is in module, get a module reference. Succeeds
1458 * for built-in code (owner == NULL) as well.
1460 if (try_module_get(new->owner
)) {
1461 if (!new->enable
|| new->enable(new) == 0) {
1462 old
= tk
->tkr_mono
.clock
;
1463 tk_setup_internals(tk
, new);
1466 module_put(old
->owner
);
1468 module_put(new->owner
);
1471 timekeeping_update(tk
, TK_CLEAR_NTP
| TK_MIRROR
| TK_CLOCK_WAS_SET
);
1473 write_seqcount_end(&tk_core
.seq
);
1474 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1480 * timekeeping_notify - Install a new clock source
1481 * @clock: pointer to the clock source
1483 * This function is called from clocksource.c after a new, better clock
1484 * source has been registered. The caller holds the clocksource_mutex.
1486 int timekeeping_notify(struct clocksource
*clock
)
1488 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1490 if (tk
->tkr_mono
.clock
== clock
)
1492 stop_machine(change_clocksource
, clock
, NULL
);
1493 tick_clock_notify();
1494 return tk
->tkr_mono
.clock
== clock
? 0 : -1;
1498 * ktime_get_raw_ts64 - Returns the raw monotonic time in a timespec
1499 * @ts: pointer to the timespec64 to be set
1501 * Returns the raw monotonic time (completely un-modified by ntp)
1503 void ktime_get_raw_ts64(struct timespec64
*ts
)
1505 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1510 seq
= read_seqcount_begin(&tk_core
.seq
);
1511 ts
->tv_sec
= tk
->raw_sec
;
1512 nsecs
= timekeeping_get_ns(&tk
->tkr_raw
);
1514 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
1517 timespec64_add_ns(ts
, nsecs
);
1519 EXPORT_SYMBOL(ktime_get_raw_ts64
);
1523 * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1525 int timekeeping_valid_for_hres(void)
1527 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1532 seq
= read_seqcount_begin(&tk_core
.seq
);
1534 ret
= tk
->tkr_mono
.clock
->flags
& CLOCK_SOURCE_VALID_FOR_HRES
;
1536 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
1542 * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1544 u64
timekeeping_max_deferment(void)
1546 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1551 seq
= read_seqcount_begin(&tk_core
.seq
);
1553 ret
= tk
->tkr_mono
.clock
->max_idle_ns
;
1555 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
1561 * read_persistent_clock64 - Return time from the persistent clock.
1562 * @ts: Pointer to the storage for the readout value
1564 * Weak dummy function for arches that do not yet support it.
1565 * Reads the time from the battery backed persistent clock.
1566 * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1568 * XXX - Do be sure to remove it once all arches implement it.
1570 void __weak
read_persistent_clock64(struct timespec64
*ts
)
1577 * read_persistent_wall_and_boot_offset - Read persistent clock, and also offset
1580 * Weak dummy function for arches that do not yet support it.
1581 * @wall_time: - current time as returned by persistent clock
1582 * @boot_offset: - offset that is defined as wall_time - boot_time
1584 * The default function calculates offset based on the current value of
1585 * local_clock(). This way architectures that support sched_clock() but don't
1586 * support dedicated boot time clock will provide the best estimate of the
1590 read_persistent_wall_and_boot_offset(struct timespec64
*wall_time
,
1591 struct timespec64
*boot_offset
)
1593 read_persistent_clock64(wall_time
);
1594 *boot_offset
= ns_to_timespec64(local_clock());
1598 * Flag reflecting whether timekeeping_resume() has injected sleeptime.
1600 * The flag starts of false and is only set when a suspend reaches
1601 * timekeeping_suspend(), timekeeping_resume() sets it to false when the
1602 * timekeeper clocksource is not stopping across suspend and has been
1603 * used to update sleep time. If the timekeeper clocksource has stopped
1604 * then the flag stays true and is used by the RTC resume code to decide
1605 * whether sleeptime must be injected and if so the flag gets false then.
1607 * If a suspend fails before reaching timekeeping_resume() then the flag
1608 * stays false and prevents erroneous sleeptime injection.
1610 static bool suspend_timing_needed
;
1612 /* Flag for if there is a persistent clock on this platform */
1613 static bool persistent_clock_exists
;
1616 * timekeeping_init - Initializes the clocksource and common timekeeping values
1618 void __init
timekeeping_init(void)
1620 struct timespec64 wall_time
, boot_offset
, wall_to_mono
;
1621 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1622 struct clocksource
*clock
;
1623 unsigned long flags
;
1625 read_persistent_wall_and_boot_offset(&wall_time
, &boot_offset
);
1626 if (timespec64_valid_settod(&wall_time
) &&
1627 timespec64_to_ns(&wall_time
) > 0) {
1628 persistent_clock_exists
= true;
1629 } else if (timespec64_to_ns(&wall_time
) != 0) {
1630 pr_warn("Persistent clock returned invalid value");
1631 wall_time
= (struct timespec64
){0};
1634 if (timespec64_compare(&wall_time
, &boot_offset
) < 0)
1635 boot_offset
= (struct timespec64
){0};
1638 * We want set wall_to_mono, so the following is true:
1639 * wall time + wall_to_mono = boot time
1641 wall_to_mono
= timespec64_sub(boot_offset
, wall_time
);
1643 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
1644 write_seqcount_begin(&tk_core
.seq
);
1647 clock
= clocksource_default_clock();
1649 clock
->enable(clock
);
1650 tk_setup_internals(tk
, clock
);
1652 tk_set_xtime(tk
, &wall_time
);
1655 tk_set_wall_to_mono(tk
, wall_to_mono
);
1657 timekeeping_update(tk
, TK_MIRROR
| TK_CLOCK_WAS_SET
);
1659 write_seqcount_end(&tk_core
.seq
);
1660 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1663 /* time in seconds when suspend began for persistent clock */
1664 static struct timespec64 timekeeping_suspend_time
;
1667 * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1668 * @tk: Pointer to the timekeeper to be updated
1669 * @delta: Pointer to the delta value in timespec64 format
1671 * Takes a timespec offset measuring a suspend interval and properly
1672 * adds the sleep offset to the timekeeping variables.
1674 static void __timekeeping_inject_sleeptime(struct timekeeper
*tk
,
1675 const struct timespec64
*delta
)
1677 if (!timespec64_valid_strict(delta
)) {
1678 printk_deferred(KERN_WARNING
1679 "__timekeeping_inject_sleeptime: Invalid "
1680 "sleep delta value!\n");
1683 tk_xtime_add(tk
, delta
);
1684 tk_set_wall_to_mono(tk
, timespec64_sub(tk
->wall_to_monotonic
, *delta
));
1685 tk_update_sleep_time(tk
, timespec64_to_ktime(*delta
));
1686 tk_debug_account_sleep_time(delta
);
1689 #if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1691 * We have three kinds of time sources to use for sleep time
1692 * injection, the preference order is:
1693 * 1) non-stop clocksource
1694 * 2) persistent clock (ie: RTC accessible when irqs are off)
1697 * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1698 * If system has neither 1) nor 2), 3) will be used finally.
1701 * If timekeeping has injected sleeptime via either 1) or 2),
1702 * 3) becomes needless, so in this case we don't need to call
1703 * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1706 bool timekeeping_rtc_skipresume(void)
1708 return !suspend_timing_needed
;
1712 * 1) can be determined whether to use or not only when doing
1713 * timekeeping_resume() which is invoked after rtc_suspend(),
1714 * so we can't skip rtc_suspend() surely if system has 1).
1716 * But if system has 2), 2) will definitely be used, so in this
1717 * case we don't need to call rtc_suspend(), and this is what
1718 * timekeeping_rtc_skipsuspend() means.
1720 bool timekeeping_rtc_skipsuspend(void)
1722 return persistent_clock_exists
;
1726 * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1727 * @delta: pointer to a timespec64 delta value
1729 * This hook is for architectures that cannot support read_persistent_clock64
1730 * because their RTC/persistent clock is only accessible when irqs are enabled.
1731 * and also don't have an effective nonstop clocksource.
1733 * This function should only be called by rtc_resume(), and allows
1734 * a suspend offset to be injected into the timekeeping values.
1736 void timekeeping_inject_sleeptime64(const struct timespec64
*delta
)
1738 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1739 unsigned long flags
;
1741 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
1742 write_seqcount_begin(&tk_core
.seq
);
1744 suspend_timing_needed
= false;
1746 timekeeping_forward_now(tk
);
1748 __timekeeping_inject_sleeptime(tk
, delta
);
1750 timekeeping_update(tk
, TK_CLEAR_NTP
| TK_MIRROR
| TK_CLOCK_WAS_SET
);
1752 write_seqcount_end(&tk_core
.seq
);
1753 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1755 /* signal hrtimers about time change */
1761 * timekeeping_resume - Resumes the generic timekeeping subsystem.
1763 void timekeeping_resume(void)
1765 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1766 struct clocksource
*clock
= tk
->tkr_mono
.clock
;
1767 unsigned long flags
;
1768 struct timespec64 ts_new
, ts_delta
;
1769 u64 cycle_now
, nsec
;
1770 bool inject_sleeptime
= false;
1772 read_persistent_clock64(&ts_new
);
1774 clockevents_resume();
1775 clocksource_resume();
1777 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
1778 write_seqcount_begin(&tk_core
.seq
);
1781 * After system resumes, we need to calculate the suspended time and
1782 * compensate it for the OS time. There are 3 sources that could be
1783 * used: Nonstop clocksource during suspend, persistent clock and rtc
1786 * One specific platform may have 1 or 2 or all of them, and the
1787 * preference will be:
1788 * suspend-nonstop clocksource -> persistent clock -> rtc
1789 * The less preferred source will only be tried if there is no better
1790 * usable source. The rtc part is handled separately in rtc core code.
1792 cycle_now
= tk_clock_read(&tk
->tkr_mono
);
1793 nsec
= clocksource_stop_suspend_timing(clock
, cycle_now
);
1795 ts_delta
= ns_to_timespec64(nsec
);
1796 inject_sleeptime
= true;
1797 } else if (timespec64_compare(&ts_new
, &timekeeping_suspend_time
) > 0) {
1798 ts_delta
= timespec64_sub(ts_new
, timekeeping_suspend_time
);
1799 inject_sleeptime
= true;
1802 if (inject_sleeptime
) {
1803 suspend_timing_needed
= false;
1804 __timekeeping_inject_sleeptime(tk
, &ts_delta
);
1807 /* Re-base the last cycle value */
1808 tk
->tkr_mono
.cycle_last
= cycle_now
;
1809 tk
->tkr_raw
.cycle_last
= cycle_now
;
1812 timekeeping_suspended
= 0;
1813 timekeeping_update(tk
, TK_MIRROR
| TK_CLOCK_WAS_SET
);
1814 write_seqcount_end(&tk_core
.seq
);
1815 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1817 touch_softlockup_watchdog();
1823 int timekeeping_suspend(void)
1825 struct timekeeper
*tk
= &tk_core
.timekeeper
;
1826 unsigned long flags
;
1827 struct timespec64 delta
, delta_delta
;
1828 static struct timespec64 old_delta
;
1829 struct clocksource
*curr_clock
;
1832 read_persistent_clock64(&timekeeping_suspend_time
);
1835 * On some systems the persistent_clock can not be detected at
1836 * timekeeping_init by its return value, so if we see a valid
1837 * value returned, update the persistent_clock_exists flag.
1839 if (timekeeping_suspend_time
.tv_sec
|| timekeeping_suspend_time
.tv_nsec
)
1840 persistent_clock_exists
= true;
1842 suspend_timing_needed
= true;
1844 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
1845 write_seqcount_begin(&tk_core
.seq
);
1846 timekeeping_forward_now(tk
);
1847 timekeeping_suspended
= 1;
1850 * Since we've called forward_now, cycle_last stores the value
1851 * just read from the current clocksource. Save this to potentially
1852 * use in suspend timing.
1854 curr_clock
= tk
->tkr_mono
.clock
;
1855 cycle_now
= tk
->tkr_mono
.cycle_last
;
1856 clocksource_start_suspend_timing(curr_clock
, cycle_now
);
1858 if (persistent_clock_exists
) {
1860 * To avoid drift caused by repeated suspend/resumes,
1861 * which each can add ~1 second drift error,
1862 * try to compensate so the difference in system time
1863 * and persistent_clock time stays close to constant.
1865 delta
= timespec64_sub(tk_xtime(tk
), timekeeping_suspend_time
);
1866 delta_delta
= timespec64_sub(delta
, old_delta
);
1867 if (abs(delta_delta
.tv_sec
) >= 2) {
1869 * if delta_delta is too large, assume time correction
1870 * has occurred and set old_delta to the current delta.
1874 /* Otherwise try to adjust old_system to compensate */
1875 timekeeping_suspend_time
=
1876 timespec64_add(timekeeping_suspend_time
, delta_delta
);
1880 timekeeping_update(tk
, TK_MIRROR
);
1881 halt_fast_timekeeper(tk
);
1882 write_seqcount_end(&tk_core
.seq
);
1883 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
1886 clocksource_suspend();
1887 clockevents_suspend();
1892 /* sysfs resume/suspend bits for timekeeping */
1893 static struct syscore_ops timekeeping_syscore_ops
= {
1894 .resume
= timekeeping_resume
,
1895 .suspend
= timekeeping_suspend
,
1898 static int __init
timekeeping_init_ops(void)
1900 register_syscore_ops(&timekeeping_syscore_ops
);
1903 device_initcall(timekeeping_init_ops
);
1906 * Apply a multiplier adjustment to the timekeeper
1908 static __always_inline
void timekeeping_apply_adjustment(struct timekeeper
*tk
,
1912 s64 interval
= tk
->cycle_interval
;
1914 if (mult_adj
== 0) {
1916 } else if (mult_adj
== -1) {
1917 interval
= -interval
;
1919 } else if (mult_adj
!= 1) {
1920 interval
*= mult_adj
;
1925 * So the following can be confusing.
1927 * To keep things simple, lets assume mult_adj == 1 for now.
1929 * When mult_adj != 1, remember that the interval and offset values
1930 * have been appropriately scaled so the math is the same.
1932 * The basic idea here is that we're increasing the multiplier
1933 * by one, this causes the xtime_interval to be incremented by
1934 * one cycle_interval. This is because:
1935 * xtime_interval = cycle_interval * mult
1936 * So if mult is being incremented by one:
1937 * xtime_interval = cycle_interval * (mult + 1)
1939 * xtime_interval = (cycle_interval * mult) + cycle_interval
1940 * Which can be shortened to:
1941 * xtime_interval += cycle_interval
1943 * So offset stores the non-accumulated cycles. Thus the current
1944 * time (in shifted nanoseconds) is:
1945 * now = (offset * adj) + xtime_nsec
1946 * Now, even though we're adjusting the clock frequency, we have
1947 * to keep time consistent. In other words, we can't jump back
1948 * in time, and we also want to avoid jumping forward in time.
1950 * So given the same offset value, we need the time to be the same
1951 * both before and after the freq adjustment.
1952 * now = (offset * adj_1) + xtime_nsec_1
1953 * now = (offset * adj_2) + xtime_nsec_2
1955 * (offset * adj_1) + xtime_nsec_1 =
1956 * (offset * adj_2) + xtime_nsec_2
1960 * (offset * adj_1) + xtime_nsec_1 =
1961 * (offset * (adj_1+1)) + xtime_nsec_2
1962 * (offset * adj_1) + xtime_nsec_1 =
1963 * (offset * adj_1) + offset + xtime_nsec_2
1964 * Canceling the sides:
1965 * xtime_nsec_1 = offset + xtime_nsec_2
1967 * xtime_nsec_2 = xtime_nsec_1 - offset
1968 * Which simplfies to:
1969 * xtime_nsec -= offset
1971 if ((mult_adj
> 0) && (tk
->tkr_mono
.mult
+ mult_adj
< mult_adj
)) {
1972 /* NTP adjustment caused clocksource mult overflow */
1977 tk
->tkr_mono
.mult
+= mult_adj
;
1978 tk
->xtime_interval
+= interval
;
1979 tk
->tkr_mono
.xtime_nsec
-= offset
;
1983 * Adjust the timekeeper's multiplier to the correct frequency
1984 * and also to reduce the accumulated error value.
1986 static void timekeeping_adjust(struct timekeeper
*tk
, s64 offset
)
1991 * Determine the multiplier from the current NTP tick length.
1992 * Avoid expensive division when the tick length doesn't change.
1994 if (likely(tk
->ntp_tick
== ntp_tick_length())) {
1995 mult
= tk
->tkr_mono
.mult
- tk
->ntp_err_mult
;
1997 tk
->ntp_tick
= ntp_tick_length();
1998 mult
= div64_u64((tk
->ntp_tick
>> tk
->ntp_error_shift
) -
1999 tk
->xtime_remainder
, tk
->cycle_interval
);
2003 * If the clock is behind the NTP time, increase the multiplier by 1
2004 * to catch up with it. If it's ahead and there was a remainder in the
2005 * tick division, the clock will slow down. Otherwise it will stay
2006 * ahead until the tick length changes to a non-divisible value.
2008 tk
->ntp_err_mult
= tk
->ntp_error
> 0 ? 1 : 0;
2009 mult
+= tk
->ntp_err_mult
;
2011 timekeeping_apply_adjustment(tk
, offset
, mult
- tk
->tkr_mono
.mult
);
2013 if (unlikely(tk
->tkr_mono
.clock
->maxadj
&&
2014 (abs(tk
->tkr_mono
.mult
- tk
->tkr_mono
.clock
->mult
)
2015 > tk
->tkr_mono
.clock
->maxadj
))) {
2016 printk_once(KERN_WARNING
2017 "Adjusting %s more than 11%% (%ld vs %ld)\n",
2018 tk
->tkr_mono
.clock
->name
, (long)tk
->tkr_mono
.mult
,
2019 (long)tk
->tkr_mono
.clock
->mult
+ tk
->tkr_mono
.clock
->maxadj
);
2023 * It may be possible that when we entered this function, xtime_nsec
2024 * was very small. Further, if we're slightly speeding the clocksource
2025 * in the code above, its possible the required corrective factor to
2026 * xtime_nsec could cause it to underflow.
2028 * Now, since we have already accumulated the second and the NTP
2029 * subsystem has been notified via second_overflow(), we need to skip
2032 if (unlikely((s64
)tk
->tkr_mono
.xtime_nsec
< 0)) {
2033 tk
->tkr_mono
.xtime_nsec
+= (u64
)NSEC_PER_SEC
<<
2036 tk
->skip_second_overflow
= 1;
2041 * accumulate_nsecs_to_secs - Accumulates nsecs into secs
2043 * Helper function that accumulates the nsecs greater than a second
2044 * from the xtime_nsec field to the xtime_secs field.
2045 * It also calls into the NTP code to handle leapsecond processing.
2047 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper
*tk
)
2049 u64 nsecps
= (u64
)NSEC_PER_SEC
<< tk
->tkr_mono
.shift
;
2050 unsigned int clock_set
= 0;
2052 while (tk
->tkr_mono
.xtime_nsec
>= nsecps
) {
2055 tk
->tkr_mono
.xtime_nsec
-= nsecps
;
2059 * Skip NTP update if this second was accumulated before,
2060 * i.e. xtime_nsec underflowed in timekeeping_adjust()
2062 if (unlikely(tk
->skip_second_overflow
)) {
2063 tk
->skip_second_overflow
= 0;
2067 /* Figure out if its a leap sec and apply if needed */
2068 leap
= second_overflow(tk
->xtime_sec
);
2069 if (unlikely(leap
)) {
2070 struct timespec64 ts
;
2072 tk
->xtime_sec
+= leap
;
2076 tk_set_wall_to_mono(tk
,
2077 timespec64_sub(tk
->wall_to_monotonic
, ts
));
2079 __timekeeping_set_tai_offset(tk
, tk
->tai_offset
- leap
);
2081 clock_set
= TK_CLOCK_WAS_SET
;
2088 * logarithmic_accumulation - shifted accumulation of cycles
2090 * This functions accumulates a shifted interval of cycles into
2091 * a shifted interval nanoseconds. Allows for O(log) accumulation
2094 * Returns the unconsumed cycles.
2096 static u64
logarithmic_accumulation(struct timekeeper
*tk
, u64 offset
,
2097 u32 shift
, unsigned int *clock_set
)
2099 u64 interval
= tk
->cycle_interval
<< shift
;
2102 /* If the offset is smaller than a shifted interval, do nothing */
2103 if (offset
< interval
)
2106 /* Accumulate one shifted interval */
2108 tk
->tkr_mono
.cycle_last
+= interval
;
2109 tk
->tkr_raw
.cycle_last
+= interval
;
2111 tk
->tkr_mono
.xtime_nsec
+= tk
->xtime_interval
<< shift
;
2112 *clock_set
|= accumulate_nsecs_to_secs(tk
);
2114 /* Accumulate raw time */
2115 tk
->tkr_raw
.xtime_nsec
+= tk
->raw_interval
<< shift
;
2116 snsec_per_sec
= (u64
)NSEC_PER_SEC
<< tk
->tkr_raw
.shift
;
2117 while (tk
->tkr_raw
.xtime_nsec
>= snsec_per_sec
) {
2118 tk
->tkr_raw
.xtime_nsec
-= snsec_per_sec
;
2122 /* Accumulate error between NTP and clock interval */
2123 tk
->ntp_error
+= tk
->ntp_tick
<< shift
;
2124 tk
->ntp_error
-= (tk
->xtime_interval
+ tk
->xtime_remainder
) <<
2125 (tk
->ntp_error_shift
+ shift
);
2131 * timekeeping_advance - Updates the timekeeper to the current time and
2132 * current NTP tick length
2134 static void timekeeping_advance(enum timekeeping_adv_mode mode
)
2136 struct timekeeper
*real_tk
= &tk_core
.timekeeper
;
2137 struct timekeeper
*tk
= &shadow_timekeeper
;
2139 int shift
= 0, maxshift
;
2140 unsigned int clock_set
= 0;
2141 unsigned long flags
;
2143 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
2145 /* Make sure we're fully resumed: */
2146 if (unlikely(timekeeping_suspended
))
2149 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
2150 offset
= real_tk
->cycle_interval
;
2152 if (mode
!= TK_ADV_TICK
)
2155 offset
= clocksource_delta(tk_clock_read(&tk
->tkr_mono
),
2156 tk
->tkr_mono
.cycle_last
, tk
->tkr_mono
.mask
);
2158 /* Check if there's really nothing to do */
2159 if (offset
< real_tk
->cycle_interval
&& mode
== TK_ADV_TICK
)
2163 /* Do some additional sanity checking */
2164 timekeeping_check_update(tk
, offset
);
2167 * With NO_HZ we may have to accumulate many cycle_intervals
2168 * (think "ticks") worth of time at once. To do this efficiently,
2169 * we calculate the largest doubling multiple of cycle_intervals
2170 * that is smaller than the offset. We then accumulate that
2171 * chunk in one go, and then try to consume the next smaller
2174 shift
= ilog2(offset
) - ilog2(tk
->cycle_interval
);
2175 shift
= max(0, shift
);
2176 /* Bound shift to one less than what overflows tick_length */
2177 maxshift
= (64 - (ilog2(ntp_tick_length())+1)) - 1;
2178 shift
= min(shift
, maxshift
);
2179 while (offset
>= tk
->cycle_interval
) {
2180 offset
= logarithmic_accumulation(tk
, offset
, shift
,
2182 if (offset
< tk
->cycle_interval
<<shift
)
2186 /* Adjust the multiplier to correct NTP error */
2187 timekeeping_adjust(tk
, offset
);
2190 * Finally, make sure that after the rounding
2191 * xtime_nsec isn't larger than NSEC_PER_SEC
2193 clock_set
|= accumulate_nsecs_to_secs(tk
);
2195 write_seqcount_begin(&tk_core
.seq
);
2197 * Update the real timekeeper.
2199 * We could avoid this memcpy by switching pointers, but that
2200 * requires changes to all other timekeeper usage sites as
2201 * well, i.e. move the timekeeper pointer getter into the
2202 * spinlocked/seqcount protected sections. And we trade this
2203 * memcpy under the tk_core.seq against one before we start
2206 timekeeping_update(tk
, clock_set
);
2207 memcpy(real_tk
, tk
, sizeof(*tk
));
2208 /* The memcpy must come last. Do not put anything here! */
2209 write_seqcount_end(&tk_core
.seq
);
2211 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
2213 /* Have to call _delayed version, since in irq context*/
2214 clock_was_set_delayed();
2218 * update_wall_time - Uses the current clocksource to increment the wall time
2221 void update_wall_time(void)
2223 timekeeping_advance(TK_ADV_TICK
);
2227 * getboottime64 - Return the real time of system boot.
2228 * @ts: pointer to the timespec64 to be set
2230 * Returns the wall-time of boot in a timespec64.
2232 * This is based on the wall_to_monotonic offset and the total suspend
2233 * time. Calls to settimeofday will affect the value returned (which
2234 * basically means that however wrong your real time clock is at boot time,
2235 * you get the right time here).
2237 void getboottime64(struct timespec64
*ts
)
2239 struct timekeeper
*tk
= &tk_core
.timekeeper
;
2240 ktime_t t
= ktime_sub(tk
->offs_real
, tk
->offs_boot
);
2242 *ts
= ktime_to_timespec64(t
);
2244 EXPORT_SYMBOL_GPL(getboottime64
);
2246 void ktime_get_coarse_real_ts64(struct timespec64
*ts
)
2248 struct timekeeper
*tk
= &tk_core
.timekeeper
;
2252 seq
= read_seqcount_begin(&tk_core
.seq
);
2255 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
2257 EXPORT_SYMBOL(ktime_get_coarse_real_ts64
);
2259 void ktime_get_coarse_ts64(struct timespec64
*ts
)
2261 struct timekeeper
*tk
= &tk_core
.timekeeper
;
2262 struct timespec64 now
, mono
;
2266 seq
= read_seqcount_begin(&tk_core
.seq
);
2269 mono
= tk
->wall_to_monotonic
;
2270 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
2272 set_normalized_timespec64(ts
, now
.tv_sec
+ mono
.tv_sec
,
2273 now
.tv_nsec
+ mono
.tv_nsec
);
2275 EXPORT_SYMBOL(ktime_get_coarse_ts64
);
2278 * Must hold jiffies_lock
2280 void do_timer(unsigned long ticks
)
2282 jiffies_64
+= ticks
;
2287 * ktime_get_update_offsets_now - hrtimer helper
2288 * @cwsseq: pointer to check and store the clock was set sequence number
2289 * @offs_real: pointer to storage for monotonic -> realtime offset
2290 * @offs_boot: pointer to storage for monotonic -> boottime offset
2291 * @offs_tai: pointer to storage for monotonic -> clock tai offset
2293 * Returns current monotonic time and updates the offsets if the
2294 * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
2297 * Called from hrtimer_interrupt() or retrigger_next_event()
2299 ktime_t
ktime_get_update_offsets_now(unsigned int *cwsseq
, ktime_t
*offs_real
,
2300 ktime_t
*offs_boot
, ktime_t
*offs_tai
)
2302 struct timekeeper
*tk
= &tk_core
.timekeeper
;
2308 seq
= read_seqcount_begin(&tk_core
.seq
);
2310 base
= tk
->tkr_mono
.base
;
2311 nsecs
= timekeeping_get_ns(&tk
->tkr_mono
);
2312 base
= ktime_add_ns(base
, nsecs
);
2314 if (*cwsseq
!= tk
->clock_was_set_seq
) {
2315 *cwsseq
= tk
->clock_was_set_seq
;
2316 *offs_real
= tk
->offs_real
;
2317 *offs_boot
= tk
->offs_boot
;
2318 *offs_tai
= tk
->offs_tai
;
2321 /* Handle leapsecond insertion adjustments */
2322 if (unlikely(base
>= tk
->next_leap_ktime
))
2323 *offs_real
= ktime_sub(tk
->offs_real
, ktime_set(1, 0));
2325 } while (read_seqcount_retry(&tk_core
.seq
, seq
));
2331 * timekeeping_validate_timex - Ensures the timex is ok for use in do_adjtimex
2333 static int timekeeping_validate_timex(const struct __kernel_timex
*txc
)
2335 if (txc
->modes
& ADJ_ADJTIME
) {
2336 /* singleshot must not be used with any other mode bits */
2337 if (!(txc
->modes
& ADJ_OFFSET_SINGLESHOT
))
2339 if (!(txc
->modes
& ADJ_OFFSET_READONLY
) &&
2340 !capable(CAP_SYS_TIME
))
2343 /* In order to modify anything, you gotta be super-user! */
2344 if (txc
->modes
&& !capable(CAP_SYS_TIME
))
2347 * if the quartz is off by more than 10% then
2348 * something is VERY wrong!
2350 if (txc
->modes
& ADJ_TICK
&&
2351 (txc
->tick
< 900000/USER_HZ
||
2352 txc
->tick
> 1100000/USER_HZ
))
2356 if (txc
->modes
& ADJ_SETOFFSET
) {
2357 /* In order to inject time, you gotta be super-user! */
2358 if (!capable(CAP_SYS_TIME
))
2362 * Validate if a timespec/timeval used to inject a time
2363 * offset is valid. Offsets can be postive or negative, so
2364 * we don't check tv_sec. The value of the timeval/timespec
2365 * is the sum of its fields,but *NOTE*:
2366 * The field tv_usec/tv_nsec must always be non-negative and
2367 * we can't have more nanoseconds/microseconds than a second.
2369 if (txc
->time
.tv_usec
< 0)
2372 if (txc
->modes
& ADJ_NANO
) {
2373 if (txc
->time
.tv_usec
>= NSEC_PER_SEC
)
2376 if (txc
->time
.tv_usec
>= USEC_PER_SEC
)
2382 * Check for potential multiplication overflows that can
2383 * only happen on 64-bit systems:
2385 if ((txc
->modes
& ADJ_FREQUENCY
) && (BITS_PER_LONG
== 64)) {
2386 if (LLONG_MIN
/ PPM_SCALE
> txc
->freq
)
2388 if (LLONG_MAX
/ PPM_SCALE
< txc
->freq
)
2397 * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2399 int do_adjtimex(struct __kernel_timex
*txc
)
2401 struct timekeeper
*tk
= &tk_core
.timekeeper
;
2402 struct audit_ntp_data ad
;
2403 unsigned long flags
;
2404 struct timespec64 ts
;
2408 /* Validate the data before disabling interrupts */
2409 ret
= timekeeping_validate_timex(txc
);
2413 if (txc
->modes
& ADJ_SETOFFSET
) {
2414 struct timespec64 delta
;
2415 delta
.tv_sec
= txc
->time
.tv_sec
;
2416 delta
.tv_nsec
= txc
->time
.tv_usec
;
2417 if (!(txc
->modes
& ADJ_NANO
))
2418 delta
.tv_nsec
*= 1000;
2419 ret
= timekeeping_inject_offset(&delta
);
2423 audit_tk_injoffset(delta
);
2426 audit_ntp_init(&ad
);
2428 ktime_get_real_ts64(&ts
);
2430 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
2431 write_seqcount_begin(&tk_core
.seq
);
2433 orig_tai
= tai
= tk
->tai_offset
;
2434 ret
= __do_adjtimex(txc
, &ts
, &tai
, &ad
);
2436 if (tai
!= orig_tai
) {
2437 __timekeeping_set_tai_offset(tk
, tai
);
2438 timekeeping_update(tk
, TK_MIRROR
| TK_CLOCK_WAS_SET
);
2440 tk_update_leap_state(tk
);
2442 write_seqcount_end(&tk_core
.seq
);
2443 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
2447 /* Update the multiplier immediately if frequency was set directly */
2448 if (txc
->modes
& (ADJ_FREQUENCY
| ADJ_TICK
))
2449 timekeeping_advance(TK_ADV_FREQ
);
2451 if (tai
!= orig_tai
)
2454 ntp_notify_cmos_timer();
2459 #ifdef CONFIG_NTP_PPS
2461 * hardpps() - Accessor function to NTP __hardpps function
2463 void hardpps(const struct timespec64
*phase_ts
, const struct timespec64
*raw_ts
)
2465 unsigned long flags
;
2467 raw_spin_lock_irqsave(&timekeeper_lock
, flags
);
2468 write_seqcount_begin(&tk_core
.seq
);
2470 __hardpps(phase_ts
, raw_ts
);
2472 write_seqcount_end(&tk_core
.seq
);
2473 raw_spin_unlock_irqrestore(&timekeeper_lock
, flags
);
2475 EXPORT_SYMBOL(hardpps
);
2476 #endif /* CONFIG_NTP_PPS */
2479 * xtime_update() - advances the timekeeping infrastructure
2480 * @ticks: number of ticks, that have elapsed since the last call.
2482 * Must be called with interrupts disabled.
2484 void xtime_update(unsigned long ticks
)
2486 raw_spin_lock(&jiffies_lock
);
2487 write_seqcount_begin(&jiffies_seq
);
2489 write_seqcount_end(&jiffies_seq
);
2490 raw_spin_unlock(&jiffies_lock
);