ARM: dts: Move the PMIC interrupt pinctrl line to rk3288-evb common
[cascardo/linux.git] / kernel / time / timekeeping.c
1 /*
2  *  linux/kernel/time/timekeeping.c
3  *
4  *  Kernel timekeeping code and accessor functions
5  *
6  *  This code was moved from linux/kernel/timer.c.
7  *  Please see that file for copyright and history logs.
8  *
9  */
10
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>
16 #include <linux/mm.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>
26
27 #include "tick-internal.h"
28 #include "ntp_internal.h"
29 #include "timekeeping_internal.h"
30
31 #define TK_CLEAR_NTP            (1 << 0)
32 #define TK_MIRROR               (1 << 1)
33 #define TK_CLOCK_WAS_SET        (1 << 2)
34
35 /*
36  * The most important data for readout fits into a single 64 byte
37  * cache line.
38  */
39 static struct {
40         seqcount_t              seq;
41         struct timekeeper       timekeeper;
42 } tk_core ____cacheline_aligned;
43
44 static DEFINE_RAW_SPINLOCK(timekeeper_lock);
45 static struct timekeeper shadow_timekeeper;
46
47 /**
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
52  *              @seq.
53  *
54  * See @update_fast_timekeeper() below.
55  */
56 struct tk_fast {
57         seqcount_t              seq;
58         struct tk_read_base     base[2];
59 };
60
61 static struct tk_fast tk_fast_mono ____cacheline_aligned;
62
63 /* flag for if timekeeping is suspended */
64 int __read_mostly timekeeping_suspended;
65
66 /* Flag for if there is a persistent clock on this platform */
67 bool __read_mostly persistent_clock_exist = false;
68
69 static inline void tk_normalize_xtime(struct timekeeper *tk)
70 {
71         while (tk->tkr.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr.shift)) {
72                 tk->tkr.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr.shift;
73                 tk->xtime_sec++;
74         }
75 }
76
77 static inline struct timespec64 tk_xtime(struct timekeeper *tk)
78 {
79         struct timespec64 ts;
80
81         ts.tv_sec = tk->xtime_sec;
82         ts.tv_nsec = (long)(tk->tkr.xtime_nsec >> tk->tkr.shift);
83         return ts;
84 }
85
86 static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
87 {
88         tk->xtime_sec = ts->tv_sec;
89         tk->tkr.xtime_nsec = (u64)ts->tv_nsec << tk->tkr.shift;
90 }
91
92 static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
93 {
94         tk->xtime_sec += ts->tv_sec;
95         tk->tkr.xtime_nsec += (u64)ts->tv_nsec << tk->tkr.shift;
96         tk_normalize_xtime(tk);
97 }
98
99 static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
100 {
101         struct timespec64 tmp;
102
103         /*
104          * Verify consistency of: offset_real = -wall_to_monotonic
105          * before modifying anything
106          */
107         set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
108                                         -tk->wall_to_monotonic.tv_nsec);
109         WARN_ON_ONCE(tk->offs_real.tv64 != timespec64_to_ktime(tmp).tv64);
110         tk->wall_to_monotonic = wtm;
111         set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
112         tk->offs_real = timespec64_to_ktime(tmp);
113         tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0));
114 }
115
116 static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
117 {
118         tk->offs_boot = ktime_add(tk->offs_boot, delta);
119 }
120
121 /**
122  * tk_setup_internals - Set up internals to use clocksource clock.
123  *
124  * @tk:         The target timekeeper to setup.
125  * @clock:              Pointer to clocksource.
126  *
127  * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
128  * pair and interval request.
129  *
130  * Unless you're the timekeeping code, you should not be using this!
131  */
132 static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
133 {
134         cycle_t interval;
135         u64 tmp, ntpinterval;
136         struct clocksource *old_clock;
137
138         old_clock = tk->tkr.clock;
139         tk->tkr.clock = clock;
140         tk->tkr.read = clock->read;
141         tk->tkr.mask = clock->mask;
142         tk->tkr.cycle_last = tk->tkr.read(clock);
143
144         /* Do the ns -> cycle conversion first, using original mult */
145         tmp = NTP_INTERVAL_LENGTH;
146         tmp <<= clock->shift;
147         ntpinterval = tmp;
148         tmp += clock->mult/2;
149         do_div(tmp, clock->mult);
150         if (tmp == 0)
151                 tmp = 1;
152
153         interval = (cycle_t) tmp;
154         tk->cycle_interval = interval;
155
156         /* Go back from cycles -> shifted ns */
157         tk->xtime_interval = (u64) interval * clock->mult;
158         tk->xtime_remainder = ntpinterval - tk->xtime_interval;
159         tk->raw_interval =
160                 ((u64) interval * clock->mult) >> clock->shift;
161
162          /* if changing clocks, convert xtime_nsec shift units */
163         if (old_clock) {
164                 int shift_change = clock->shift - old_clock->shift;
165                 if (shift_change < 0)
166                         tk->tkr.xtime_nsec >>= -shift_change;
167                 else
168                         tk->tkr.xtime_nsec <<= shift_change;
169         }
170         tk->tkr.shift = clock->shift;
171
172         tk->ntp_error = 0;
173         tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
174         tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
175
176         /*
177          * The timekeeper keeps its own mult values for the currently
178          * active clocksource. These value will be adjusted via NTP
179          * to counteract clock drifting.
180          */
181         tk->tkr.mult = clock->mult;
182         tk->ntp_err_mult = 0;
183 }
184
185 /* Timekeeper helper functions. */
186
187 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
188 static u32 default_arch_gettimeoffset(void) { return 0; }
189 u32 (*arch_gettimeoffset)(void) = default_arch_gettimeoffset;
190 #else
191 static inline u32 arch_gettimeoffset(void) { return 0; }
192 #endif
193
194 static inline s64 timekeeping_get_ns(struct tk_read_base *tkr)
195 {
196         cycle_t cycle_now, delta;
197         s64 nsec;
198
199         /* read clocksource: */
200         cycle_now = tkr->read(tkr->clock);
201
202         /* calculate the delta since the last update_wall_time: */
203         delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
204
205         nsec = delta * tkr->mult + tkr->xtime_nsec;
206         nsec >>= tkr->shift;
207
208         /* If arch requires, add in get_arch_timeoffset() */
209         return nsec + arch_gettimeoffset();
210 }
211
212 static inline s64 timekeeping_get_ns_raw(struct timekeeper *tk)
213 {
214         struct clocksource *clock = tk->tkr.clock;
215         cycle_t cycle_now, delta;
216         s64 nsec;
217
218         /* read clocksource: */
219         cycle_now = tk->tkr.read(clock);
220
221         /* calculate the delta since the last update_wall_time: */
222         delta = clocksource_delta(cycle_now, tk->tkr.cycle_last, tk->tkr.mask);
223
224         /* convert delta to nanoseconds. */
225         nsec = clocksource_cyc2ns(delta, clock->mult, clock->shift);
226
227         /* If arch requires, add in get_arch_timeoffset() */
228         return nsec + arch_gettimeoffset();
229 }
230
231 /**
232  * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
233  * @tk:         The timekeeper from which we take the update
234  * @tkf:        The fast timekeeper to update
235  * @tbase:      The time base for the fast timekeeper (mono/raw)
236  *
237  * We want to use this from any context including NMI and tracing /
238  * instrumenting the timekeeping code itself.
239  *
240  * So we handle this differently than the other timekeeping accessor
241  * functions which retry when the sequence count has changed. The
242  * update side does:
243  *
244  * smp_wmb();   <- Ensure that the last base[1] update is visible
245  * tkf->seq++;
246  * smp_wmb();   <- Ensure that the seqcount update is visible
247  * update(tkf->base[0], tk);
248  * smp_wmb();   <- Ensure that the base[0] update is visible
249  * tkf->seq++;
250  * smp_wmb();   <- Ensure that the seqcount update is visible
251  * update(tkf->base[1], tk);
252  *
253  * The reader side does:
254  *
255  * do {
256  *      seq = tkf->seq;
257  *      smp_rmb();
258  *      idx = seq & 0x01;
259  *      now = now(tkf->base[idx]);
260  *      smp_rmb();
261  * } while (seq != tkf->seq)
262  *
263  * As long as we update base[0] readers are forced off to
264  * base[1]. Once base[0] is updated readers are redirected to base[0]
265  * and the base[1] update takes place.
266  *
267  * So if a NMI hits the update of base[0] then it will use base[1]
268  * which is still consistent. In the worst case this can result is a
269  * slightly wrong timestamp (a few nanoseconds). See
270  * @ktime_get_mono_fast_ns.
271  */
272 static void update_fast_timekeeper(struct timekeeper *tk)
273 {
274         struct tk_read_base *base = tk_fast_mono.base;
275
276         /* Force readers off to base[1] */
277         raw_write_seqcount_latch(&tk_fast_mono.seq);
278
279         /* Update base[0] */
280         memcpy(base, &tk->tkr, sizeof(*base));
281
282         /* Force readers back to base[0] */
283         raw_write_seqcount_latch(&tk_fast_mono.seq);
284
285         /* Update base[1] */
286         memcpy(base + 1, base, sizeof(*base));
287 }
288
289 /**
290  * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
291  *
292  * This timestamp is not guaranteed to be monotonic across an update.
293  * The timestamp is calculated by:
294  *
295  *      now = base_mono + clock_delta * slope
296  *
297  * So if the update lowers the slope, readers who are forced to the
298  * not yet updated second array are still using the old steeper slope.
299  *
300  * tmono
301  * ^
302  * |    o  n
303  * |   o n
304  * |  u
305  * | o
306  * |o
307  * |12345678---> reader order
308  *
309  * o = old slope
310  * u = update
311  * n = new slope
312  *
313  * So reader 6 will observe time going backwards versus reader 5.
314  *
315  * While other CPUs are likely to be able observe that, the only way
316  * for a CPU local observation is when an NMI hits in the middle of
317  * the update. Timestamps taken from that NMI context might be ahead
318  * of the following timestamps. Callers need to be aware of that and
319  * deal with it.
320  */
321 u64 notrace ktime_get_mono_fast_ns(void)
322 {
323         struct tk_read_base *tkr;
324         unsigned int seq;
325         u64 now;
326
327         do {
328                 seq = raw_read_seqcount(&tk_fast_mono.seq);
329                 tkr = tk_fast_mono.base + (seq & 0x01);
330                 now = ktime_to_ns(tkr->base_mono) + timekeeping_get_ns(tkr);
331
332         } while (read_seqcount_retry(&tk_fast_mono.seq, seq));
333         return now;
334 }
335 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
336
337 #ifdef CONFIG_GENERIC_TIME_VSYSCALL_OLD
338
339 static inline void update_vsyscall(struct timekeeper *tk)
340 {
341         struct timespec xt, wm;
342
343         xt = timespec64_to_timespec(tk_xtime(tk));
344         wm = timespec64_to_timespec(tk->wall_to_monotonic);
345         update_vsyscall_old(&xt, &wm, tk->tkr.clock, tk->tkr.mult,
346                             tk->tkr.cycle_last);
347 }
348
349 static inline void old_vsyscall_fixup(struct timekeeper *tk)
350 {
351         s64 remainder;
352
353         /*
354         * Store only full nanoseconds into xtime_nsec after rounding
355         * it up and add the remainder to the error difference.
356         * XXX - This is necessary to avoid small 1ns inconsistnecies caused
357         * by truncating the remainder in vsyscalls. However, it causes
358         * additional work to be done in timekeeping_adjust(). Once
359         * the vsyscall implementations are converted to use xtime_nsec
360         * (shifted nanoseconds), and CONFIG_GENERIC_TIME_VSYSCALL_OLD
361         * users are removed, this can be killed.
362         */
363         remainder = tk->tkr.xtime_nsec & ((1ULL << tk->tkr.shift) - 1);
364         tk->tkr.xtime_nsec -= remainder;
365         tk->tkr.xtime_nsec += 1ULL << tk->tkr.shift;
366         tk->ntp_error += remainder << tk->ntp_error_shift;
367         tk->ntp_error -= (1ULL << tk->tkr.shift) << tk->ntp_error_shift;
368 }
369 #else
370 #define old_vsyscall_fixup(tk)
371 #endif
372
373 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
374
375 static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
376 {
377         raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
378 }
379
380 /**
381  * pvclock_gtod_register_notifier - register a pvclock timedata update listener
382  */
383 int pvclock_gtod_register_notifier(struct notifier_block *nb)
384 {
385         struct timekeeper *tk = &tk_core.timekeeper;
386         unsigned long flags;
387         int ret;
388
389         raw_spin_lock_irqsave(&timekeeper_lock, flags);
390         ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
391         update_pvclock_gtod(tk, true);
392         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
393
394         return ret;
395 }
396 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
397
398 /**
399  * pvclock_gtod_unregister_notifier - unregister a pvclock
400  * timedata update listener
401  */
402 int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
403 {
404         unsigned long flags;
405         int ret;
406
407         raw_spin_lock_irqsave(&timekeeper_lock, flags);
408         ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
409         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
410
411         return ret;
412 }
413 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
414
415 /*
416  * Update the ktime_t based scalar nsec members of the timekeeper
417  */
418 static inline void tk_update_ktime_data(struct timekeeper *tk)
419 {
420         s64 nsec;
421
422         /*
423          * The xtime based monotonic readout is:
424          *      nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
425          * The ktime based monotonic readout is:
426          *      nsec = base_mono + now();
427          * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
428          */
429         nsec = (s64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
430         nsec *= NSEC_PER_SEC;
431         nsec += tk->wall_to_monotonic.tv_nsec;
432         tk->tkr.base_mono = ns_to_ktime(nsec);
433
434         /* Update the monotonic raw base */
435         tk->base_raw = timespec64_to_ktime(tk->raw_time);
436 }
437
438 /* must hold timekeeper_lock */
439 static void timekeeping_update(struct timekeeper *tk, unsigned int action)
440 {
441         if (action & TK_CLEAR_NTP) {
442                 tk->ntp_error = 0;
443                 ntp_clear();
444         }
445         update_vsyscall(tk);
446         update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
447
448         tk_update_ktime_data(tk);
449
450         if (action & TK_MIRROR)
451                 memcpy(&shadow_timekeeper, &tk_core.timekeeper,
452                        sizeof(tk_core.timekeeper));
453
454         update_fast_timekeeper(tk);
455 }
456
457 /**
458  * timekeeping_forward_now - update clock to the current time
459  *
460  * Forward the current clock to update its state since the last call to
461  * update_wall_time(). This is useful before significant clock changes,
462  * as it avoids having to deal with this time offset explicitly.
463  */
464 static void timekeeping_forward_now(struct timekeeper *tk)
465 {
466         struct clocksource *clock = tk->tkr.clock;
467         cycle_t cycle_now, delta;
468         s64 nsec;
469
470         cycle_now = tk->tkr.read(clock);
471         delta = clocksource_delta(cycle_now, tk->tkr.cycle_last, tk->tkr.mask);
472         tk->tkr.cycle_last = cycle_now;
473
474         tk->tkr.xtime_nsec += delta * tk->tkr.mult;
475
476         /* If arch requires, add in get_arch_timeoffset() */
477         tk->tkr.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr.shift;
478
479         tk_normalize_xtime(tk);
480
481         nsec = clocksource_cyc2ns(delta, clock->mult, clock->shift);
482         timespec64_add_ns(&tk->raw_time, nsec);
483 }
484
485 /**
486  * __getnstimeofday64 - Returns the time of day in a timespec64.
487  * @ts:         pointer to the timespec to be set
488  *
489  * Updates the time of day in the timespec.
490  * Returns 0 on success, or -ve when suspended (timespec will be undefined).
491  */
492 int __getnstimeofday64(struct timespec64 *ts)
493 {
494         struct timekeeper *tk = &tk_core.timekeeper;
495         unsigned long seq;
496         s64 nsecs = 0;
497
498         do {
499                 seq = read_seqcount_begin(&tk_core.seq);
500
501                 ts->tv_sec = tk->xtime_sec;
502                 nsecs = timekeeping_get_ns(&tk->tkr);
503
504         } while (read_seqcount_retry(&tk_core.seq, seq));
505
506         ts->tv_nsec = 0;
507         timespec64_add_ns(ts, nsecs);
508
509         /*
510          * Do not bail out early, in case there were callers still using
511          * the value, even in the face of the WARN_ON.
512          */
513         if (unlikely(timekeeping_suspended))
514                 return -EAGAIN;
515         return 0;
516 }
517 EXPORT_SYMBOL(__getnstimeofday64);
518
519 /**
520  * getnstimeofday64 - Returns the time of day in a timespec64.
521  * @ts:         pointer to the timespec to be set
522  *
523  * Returns the time of day in a timespec (WARN if suspended).
524  */
525 void getnstimeofday64(struct timespec64 *ts)
526 {
527         WARN_ON(__getnstimeofday64(ts));
528 }
529 EXPORT_SYMBOL(getnstimeofday64);
530
531 ktime_t ktime_get(void)
532 {
533         struct timekeeper *tk = &tk_core.timekeeper;
534         unsigned int seq;
535         ktime_t base;
536         s64 nsecs;
537
538         WARN_ON(timekeeping_suspended);
539
540         do {
541                 seq = read_seqcount_begin(&tk_core.seq);
542                 base = tk->tkr.base_mono;
543                 nsecs = timekeeping_get_ns(&tk->tkr);
544
545         } while (read_seqcount_retry(&tk_core.seq, seq));
546
547         return ktime_add_ns(base, nsecs);
548 }
549 EXPORT_SYMBOL_GPL(ktime_get);
550
551 static ktime_t *offsets[TK_OFFS_MAX] = {
552         [TK_OFFS_REAL]  = &tk_core.timekeeper.offs_real,
553         [TK_OFFS_BOOT]  = &tk_core.timekeeper.offs_boot,
554         [TK_OFFS_TAI]   = &tk_core.timekeeper.offs_tai,
555 };
556
557 ktime_t ktime_get_with_offset(enum tk_offsets offs)
558 {
559         struct timekeeper *tk = &tk_core.timekeeper;
560         unsigned int seq;
561         ktime_t base, *offset = offsets[offs];
562         s64 nsecs;
563
564         WARN_ON(timekeeping_suspended);
565
566         do {
567                 seq = read_seqcount_begin(&tk_core.seq);
568                 base = ktime_add(tk->tkr.base_mono, *offset);
569                 nsecs = timekeeping_get_ns(&tk->tkr);
570
571         } while (read_seqcount_retry(&tk_core.seq, seq));
572
573         return ktime_add_ns(base, nsecs);
574
575 }
576 EXPORT_SYMBOL_GPL(ktime_get_with_offset);
577
578 /**
579  * ktime_mono_to_any() - convert mononotic time to any other time
580  * @tmono:      time to convert.
581  * @offs:       which offset to use
582  */
583 ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
584 {
585         ktime_t *offset = offsets[offs];
586         unsigned long seq;
587         ktime_t tconv;
588
589         do {
590                 seq = read_seqcount_begin(&tk_core.seq);
591                 tconv = ktime_add(tmono, *offset);
592         } while (read_seqcount_retry(&tk_core.seq, seq));
593
594         return tconv;
595 }
596 EXPORT_SYMBOL_GPL(ktime_mono_to_any);
597
598 /**
599  * ktime_get_raw - Returns the raw monotonic time in ktime_t format
600  */
601 ktime_t ktime_get_raw(void)
602 {
603         struct timekeeper *tk = &tk_core.timekeeper;
604         unsigned int seq;
605         ktime_t base;
606         s64 nsecs;
607
608         do {
609                 seq = read_seqcount_begin(&tk_core.seq);
610                 base = tk->base_raw;
611                 nsecs = timekeeping_get_ns_raw(tk);
612
613         } while (read_seqcount_retry(&tk_core.seq, seq));
614
615         return ktime_add_ns(base, nsecs);
616 }
617 EXPORT_SYMBOL_GPL(ktime_get_raw);
618
619 /**
620  * ktime_get_ts64 - get the monotonic clock in timespec64 format
621  * @ts:         pointer to timespec variable
622  *
623  * The function calculates the monotonic clock from the realtime
624  * clock and the wall_to_monotonic offset and stores the result
625  * in normalized timespec format in the variable pointed to by @ts.
626  */
627 void ktime_get_ts64(struct timespec64 *ts)
628 {
629         struct timekeeper *tk = &tk_core.timekeeper;
630         struct timespec64 tomono;
631         s64 nsec;
632         unsigned int seq;
633
634         WARN_ON(timekeeping_suspended);
635
636         do {
637                 seq = read_seqcount_begin(&tk_core.seq);
638                 ts->tv_sec = tk->xtime_sec;
639                 nsec = timekeeping_get_ns(&tk->tkr);
640                 tomono = tk->wall_to_monotonic;
641
642         } while (read_seqcount_retry(&tk_core.seq, seq));
643
644         ts->tv_sec += tomono.tv_sec;
645         ts->tv_nsec = 0;
646         timespec64_add_ns(ts, nsec + tomono.tv_nsec);
647 }
648 EXPORT_SYMBOL_GPL(ktime_get_ts64);
649
650 #ifdef CONFIG_NTP_PPS
651
652 /**
653  * getnstime_raw_and_real - get day and raw monotonic time in timespec format
654  * @ts_raw:     pointer to the timespec to be set to raw monotonic time
655  * @ts_real:    pointer to the timespec to be set to the time of day
656  *
657  * This function reads both the time of day and raw monotonic time at the
658  * same time atomically and stores the resulting timestamps in timespec
659  * format.
660  */
661 void getnstime_raw_and_real(struct timespec *ts_raw, struct timespec *ts_real)
662 {
663         struct timekeeper *tk = &tk_core.timekeeper;
664         unsigned long seq;
665         s64 nsecs_raw, nsecs_real;
666
667         WARN_ON_ONCE(timekeeping_suspended);
668
669         do {
670                 seq = read_seqcount_begin(&tk_core.seq);
671
672                 *ts_raw = timespec64_to_timespec(tk->raw_time);
673                 ts_real->tv_sec = tk->xtime_sec;
674                 ts_real->tv_nsec = 0;
675
676                 nsecs_raw = timekeeping_get_ns_raw(tk);
677                 nsecs_real = timekeeping_get_ns(&tk->tkr);
678
679         } while (read_seqcount_retry(&tk_core.seq, seq));
680
681         timespec_add_ns(ts_raw, nsecs_raw);
682         timespec_add_ns(ts_real, nsecs_real);
683 }
684 EXPORT_SYMBOL(getnstime_raw_and_real);
685
686 #endif /* CONFIG_NTP_PPS */
687
688 /**
689  * do_gettimeofday - Returns the time of day in a timeval
690  * @tv:         pointer to the timeval to be set
691  *
692  * NOTE: Users should be converted to using getnstimeofday()
693  */
694 void do_gettimeofday(struct timeval *tv)
695 {
696         struct timespec64 now;
697
698         getnstimeofday64(&now);
699         tv->tv_sec = now.tv_sec;
700         tv->tv_usec = now.tv_nsec/1000;
701 }
702 EXPORT_SYMBOL(do_gettimeofday);
703
704 /**
705  * do_settimeofday - Sets the time of day
706  * @tv:         pointer to the timespec variable containing the new time
707  *
708  * Sets the time of day to the new time and update NTP and notify hrtimers
709  */
710 int do_settimeofday(const struct timespec *tv)
711 {
712         struct timekeeper *tk = &tk_core.timekeeper;
713         struct timespec64 ts_delta, xt, tmp;
714         unsigned long flags;
715
716         if (!timespec_valid_strict(tv))
717                 return -EINVAL;
718
719         raw_spin_lock_irqsave(&timekeeper_lock, flags);
720         write_seqcount_begin(&tk_core.seq);
721
722         timekeeping_forward_now(tk);
723
724         xt = tk_xtime(tk);
725         ts_delta.tv_sec = tv->tv_sec - xt.tv_sec;
726         ts_delta.tv_nsec = tv->tv_nsec - xt.tv_nsec;
727
728         tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
729
730         tmp = timespec_to_timespec64(*tv);
731         tk_set_xtime(tk, &tmp);
732
733         timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
734
735         write_seqcount_end(&tk_core.seq);
736         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
737
738         /* signal hrtimers about time change */
739         clock_was_set();
740
741         return 0;
742 }
743 EXPORT_SYMBOL(do_settimeofday);
744
745 /**
746  * timekeeping_inject_offset - Adds or subtracts from the current time.
747  * @tv:         pointer to the timespec variable containing the offset
748  *
749  * Adds or subtracts an offset value from the current time.
750  */
751 int timekeeping_inject_offset(struct timespec *ts)
752 {
753         struct timekeeper *tk = &tk_core.timekeeper;
754         unsigned long flags;
755         struct timespec64 ts64, tmp;
756         int ret = 0;
757
758         if ((unsigned long)ts->tv_nsec >= NSEC_PER_SEC)
759                 return -EINVAL;
760
761         ts64 = timespec_to_timespec64(*ts);
762
763         raw_spin_lock_irqsave(&timekeeper_lock, flags);
764         write_seqcount_begin(&tk_core.seq);
765
766         timekeeping_forward_now(tk);
767
768         /* Make sure the proposed value is valid */
769         tmp = timespec64_add(tk_xtime(tk),  ts64);
770         if (!timespec64_valid_strict(&tmp)) {
771                 ret = -EINVAL;
772                 goto error;
773         }
774
775         tk_xtime_add(tk, &ts64);
776         tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts64));
777
778 error: /* even if we error out, we forwarded the time, so call update */
779         timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
780
781         write_seqcount_end(&tk_core.seq);
782         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
783
784         /* signal hrtimers about time change */
785         clock_was_set();
786
787         return ret;
788 }
789 EXPORT_SYMBOL(timekeeping_inject_offset);
790
791
792 /**
793  * timekeeping_get_tai_offset - Returns current TAI offset from UTC
794  *
795  */
796 s32 timekeeping_get_tai_offset(void)
797 {
798         struct timekeeper *tk = &tk_core.timekeeper;
799         unsigned int seq;
800         s32 ret;
801
802         do {
803                 seq = read_seqcount_begin(&tk_core.seq);
804                 ret = tk->tai_offset;
805         } while (read_seqcount_retry(&tk_core.seq, seq));
806
807         return ret;
808 }
809
810 /**
811  * __timekeeping_set_tai_offset - Lock free worker function
812  *
813  */
814 static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
815 {
816         tk->tai_offset = tai_offset;
817         tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
818 }
819
820 /**
821  * timekeeping_set_tai_offset - Sets the current TAI offset from UTC
822  *
823  */
824 void timekeeping_set_tai_offset(s32 tai_offset)
825 {
826         struct timekeeper *tk = &tk_core.timekeeper;
827         unsigned long flags;
828
829         raw_spin_lock_irqsave(&timekeeper_lock, flags);
830         write_seqcount_begin(&tk_core.seq);
831         __timekeeping_set_tai_offset(tk, tai_offset);
832         timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
833         write_seqcount_end(&tk_core.seq);
834         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
835         clock_was_set();
836 }
837
838 /**
839  * change_clocksource - Swaps clocksources if a new one is available
840  *
841  * Accumulates current time interval and initializes new clocksource
842  */
843 static int change_clocksource(void *data)
844 {
845         struct timekeeper *tk = &tk_core.timekeeper;
846         struct clocksource *new, *old;
847         unsigned long flags;
848
849         new = (struct clocksource *) data;
850
851         raw_spin_lock_irqsave(&timekeeper_lock, flags);
852         write_seqcount_begin(&tk_core.seq);
853
854         timekeeping_forward_now(tk);
855         /*
856          * If the cs is in module, get a module reference. Succeeds
857          * for built-in code (owner == NULL) as well.
858          */
859         if (try_module_get(new->owner)) {
860                 if (!new->enable || new->enable(new) == 0) {
861                         old = tk->tkr.clock;
862                         tk_setup_internals(tk, new);
863                         if (old->disable)
864                                 old->disable(old);
865                         module_put(old->owner);
866                 } else {
867                         module_put(new->owner);
868                 }
869         }
870         timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
871
872         write_seqcount_end(&tk_core.seq);
873         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
874
875         return 0;
876 }
877
878 /**
879  * timekeeping_notify - Install a new clock source
880  * @clock:              pointer to the clock source
881  *
882  * This function is called from clocksource.c after a new, better clock
883  * source has been registered. The caller holds the clocksource_mutex.
884  */
885 int timekeeping_notify(struct clocksource *clock)
886 {
887         struct timekeeper *tk = &tk_core.timekeeper;
888
889         if (tk->tkr.clock == clock)
890                 return 0;
891         stop_machine(change_clocksource, clock, NULL);
892         tick_clock_notify();
893         return tk->tkr.clock == clock ? 0 : -1;
894 }
895
896 /**
897  * getrawmonotonic - Returns the raw monotonic time in a timespec
898  * @ts:         pointer to the timespec to be set
899  *
900  * Returns the raw monotonic time (completely un-modified by ntp)
901  */
902 void getrawmonotonic(struct timespec *ts)
903 {
904         struct timekeeper *tk = &tk_core.timekeeper;
905         struct timespec64 ts64;
906         unsigned long seq;
907         s64 nsecs;
908
909         do {
910                 seq = read_seqcount_begin(&tk_core.seq);
911                 nsecs = timekeeping_get_ns_raw(tk);
912                 ts64 = tk->raw_time;
913
914         } while (read_seqcount_retry(&tk_core.seq, seq));
915
916         timespec64_add_ns(&ts64, nsecs);
917         *ts = timespec64_to_timespec(ts64);
918 }
919 EXPORT_SYMBOL(getrawmonotonic);
920
921 /**
922  * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
923  */
924 int timekeeping_valid_for_hres(void)
925 {
926         struct timekeeper *tk = &tk_core.timekeeper;
927         unsigned long seq;
928         int ret;
929
930         do {
931                 seq = read_seqcount_begin(&tk_core.seq);
932
933                 ret = tk->tkr.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
934
935         } while (read_seqcount_retry(&tk_core.seq, seq));
936
937         return ret;
938 }
939
940 /**
941  * timekeeping_max_deferment - Returns max time the clocksource can be deferred
942  */
943 u64 timekeeping_max_deferment(void)
944 {
945         struct timekeeper *tk = &tk_core.timekeeper;
946         unsigned long seq;
947         u64 ret;
948
949         do {
950                 seq = read_seqcount_begin(&tk_core.seq);
951
952                 ret = tk->tkr.clock->max_idle_ns;
953
954         } while (read_seqcount_retry(&tk_core.seq, seq));
955
956         return ret;
957 }
958
959 /**
960  * read_persistent_clock -  Return time from the persistent clock.
961  *
962  * Weak dummy function for arches that do not yet support it.
963  * Reads the time from the battery backed persistent clock.
964  * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
965  *
966  *  XXX - Do be sure to remove it once all arches implement it.
967  */
968 void __weak read_persistent_clock(struct timespec *ts)
969 {
970         ts->tv_sec = 0;
971         ts->tv_nsec = 0;
972 }
973
974 /**
975  * read_boot_clock -  Return time of the system start.
976  *
977  * Weak dummy function for arches that do not yet support it.
978  * Function to read the exact time the system has been started.
979  * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
980  *
981  *  XXX - Do be sure to remove it once all arches implement it.
982  */
983 void __weak read_boot_clock(struct timespec *ts)
984 {
985         ts->tv_sec = 0;
986         ts->tv_nsec = 0;
987 }
988
989 /*
990  * timekeeping_init - Initializes the clocksource and common timekeeping values
991  */
992 void __init timekeeping_init(void)
993 {
994         struct timekeeper *tk = &tk_core.timekeeper;
995         struct clocksource *clock;
996         unsigned long flags;
997         struct timespec64 now, boot, tmp;
998         struct timespec ts;
999
1000         read_persistent_clock(&ts);
1001         now = timespec_to_timespec64(ts);
1002         if (!timespec64_valid_strict(&now)) {
1003                 pr_warn("WARNING: Persistent clock returned invalid value!\n"
1004                         "         Check your CMOS/BIOS settings.\n");
1005                 now.tv_sec = 0;
1006                 now.tv_nsec = 0;
1007         } else if (now.tv_sec || now.tv_nsec)
1008                 persistent_clock_exist = true;
1009
1010         read_boot_clock(&ts);
1011         boot = timespec_to_timespec64(ts);
1012         if (!timespec64_valid_strict(&boot)) {
1013                 pr_warn("WARNING: Boot clock returned invalid value!\n"
1014                         "         Check your CMOS/BIOS settings.\n");
1015                 boot.tv_sec = 0;
1016                 boot.tv_nsec = 0;
1017         }
1018
1019         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1020         write_seqcount_begin(&tk_core.seq);
1021         ntp_init();
1022
1023         clock = clocksource_default_clock();
1024         if (clock->enable)
1025                 clock->enable(clock);
1026         tk_setup_internals(tk, clock);
1027
1028         tk_set_xtime(tk, &now);
1029         tk->raw_time.tv_sec = 0;
1030         tk->raw_time.tv_nsec = 0;
1031         tk->base_raw.tv64 = 0;
1032         if (boot.tv_sec == 0 && boot.tv_nsec == 0)
1033                 boot = tk_xtime(tk);
1034
1035         set_normalized_timespec64(&tmp, -boot.tv_sec, -boot.tv_nsec);
1036         tk_set_wall_to_mono(tk, tmp);
1037
1038         timekeeping_update(tk, TK_MIRROR);
1039
1040         write_seqcount_end(&tk_core.seq);
1041         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1042 }
1043
1044 /* time in seconds when suspend began */
1045 static struct timespec64 timekeeping_suspend_time;
1046
1047 /**
1048  * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1049  * @delta: pointer to a timespec delta value
1050  *
1051  * Takes a timespec offset measuring a suspend interval and properly
1052  * adds the sleep offset to the timekeeping variables.
1053  */
1054 static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1055                                            struct timespec64 *delta)
1056 {
1057         if (!timespec64_valid_strict(delta)) {
1058                 printk_deferred(KERN_WARNING
1059                                 "__timekeeping_inject_sleeptime: Invalid "
1060                                 "sleep delta value!\n");
1061                 return;
1062         }
1063         tk_xtime_add(tk, delta);
1064         tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1065         tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1066         tk_debug_account_sleep_time(delta);
1067 }
1068
1069 /**
1070  * timekeeping_inject_sleeptime - Adds suspend interval to timeekeeping values
1071  * @delta: pointer to a timespec delta value
1072  *
1073  * This hook is for architectures that cannot support read_persistent_clock
1074  * because their RTC/persistent clock is only accessible when irqs are enabled.
1075  *
1076  * This function should only be called by rtc_resume(), and allows
1077  * a suspend offset to be injected into the timekeeping values.
1078  */
1079 void timekeeping_inject_sleeptime(struct timespec *delta)
1080 {
1081         struct timekeeper *tk = &tk_core.timekeeper;
1082         struct timespec64 tmp;
1083         unsigned long flags;
1084
1085         /*
1086          * Make sure we don't set the clock twice, as timekeeping_resume()
1087          * already did it
1088          */
1089         if (has_persistent_clock())
1090                 return;
1091
1092         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1093         write_seqcount_begin(&tk_core.seq);
1094
1095         timekeeping_forward_now(tk);
1096
1097         tmp = timespec_to_timespec64(*delta);
1098         __timekeeping_inject_sleeptime(tk, &tmp);
1099
1100         timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1101
1102         write_seqcount_end(&tk_core.seq);
1103         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1104
1105         /* signal hrtimers about time change */
1106         clock_was_set();
1107 }
1108
1109 /**
1110  * timekeeping_resume - Resumes the generic timekeeping subsystem.
1111  *
1112  * This is for the generic clocksource timekeeping.
1113  * xtime/wall_to_monotonic/jiffies/etc are
1114  * still managed by arch specific suspend/resume code.
1115  */
1116 static void timekeeping_resume(void)
1117 {
1118         struct timekeeper *tk = &tk_core.timekeeper;
1119         struct clocksource *clock = tk->tkr.clock;
1120         unsigned long flags;
1121         struct timespec64 ts_new, ts_delta;
1122         struct timespec tmp;
1123         cycle_t cycle_now, cycle_delta;
1124         bool suspendtime_found = false;
1125
1126         read_persistent_clock(&tmp);
1127         ts_new = timespec_to_timespec64(tmp);
1128
1129         clockevents_resume();
1130         clocksource_resume();
1131
1132         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1133         write_seqcount_begin(&tk_core.seq);
1134
1135         /*
1136          * After system resumes, we need to calculate the suspended time and
1137          * compensate it for the OS time. There are 3 sources that could be
1138          * used: Nonstop clocksource during suspend, persistent clock and rtc
1139          * device.
1140          *
1141          * One specific platform may have 1 or 2 or all of them, and the
1142          * preference will be:
1143          *      suspend-nonstop clocksource -> persistent clock -> rtc
1144          * The less preferred source will only be tried if there is no better
1145          * usable source. The rtc part is handled separately in rtc core code.
1146          */
1147         cycle_now = tk->tkr.read(clock);
1148         if ((clock->flags & CLOCK_SOURCE_SUSPEND_NONSTOP) &&
1149                 cycle_now > tk->tkr.cycle_last) {
1150                 u64 num, max = ULLONG_MAX;
1151                 u32 mult = clock->mult;
1152                 u32 shift = clock->shift;
1153                 s64 nsec = 0;
1154
1155                 cycle_delta = clocksource_delta(cycle_now, tk->tkr.cycle_last,
1156                                                 tk->tkr.mask);
1157
1158                 /*
1159                  * "cycle_delta * mutl" may cause 64 bits overflow, if the
1160                  * suspended time is too long. In that case we need do the
1161                  * 64 bits math carefully
1162                  */
1163                 do_div(max, mult);
1164                 if (cycle_delta > max) {
1165                         num = div64_u64(cycle_delta, max);
1166                         nsec = (((u64) max * mult) >> shift) * num;
1167                         cycle_delta -= num * max;
1168                 }
1169                 nsec += ((u64) cycle_delta * mult) >> shift;
1170
1171                 ts_delta = ns_to_timespec64(nsec);
1172                 suspendtime_found = true;
1173         } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1174                 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1175                 suspendtime_found = true;
1176         }
1177
1178         if (suspendtime_found)
1179                 __timekeeping_inject_sleeptime(tk, &ts_delta);
1180
1181         /* Re-base the last cycle value */
1182         tk->tkr.cycle_last = cycle_now;
1183         tk->ntp_error = 0;
1184         timekeeping_suspended = 0;
1185         timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1186         write_seqcount_end(&tk_core.seq);
1187         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1188
1189         touch_softlockup_watchdog();
1190
1191         clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
1192
1193         /* Resume hrtimers */
1194         hrtimers_resume();
1195 }
1196
1197 static int timekeeping_suspend(void)
1198 {
1199         struct timekeeper *tk = &tk_core.timekeeper;
1200         unsigned long flags;
1201         struct timespec64               delta, delta_delta;
1202         static struct timespec64        old_delta;
1203         struct timespec tmp;
1204
1205         read_persistent_clock(&tmp);
1206         timekeeping_suspend_time = timespec_to_timespec64(tmp);
1207
1208         /*
1209          * On some systems the persistent_clock can not be detected at
1210          * timekeeping_init by its return value, so if we see a valid
1211          * value returned, update the persistent_clock_exists flag.
1212          */
1213         if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1214                 persistent_clock_exist = true;
1215
1216         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1217         write_seqcount_begin(&tk_core.seq);
1218         timekeeping_forward_now(tk);
1219         timekeeping_suspended = 1;
1220
1221         /*
1222          * To avoid drift caused by repeated suspend/resumes,
1223          * which each can add ~1 second drift error,
1224          * try to compensate so the difference in system time
1225          * and persistent_clock time stays close to constant.
1226          */
1227         delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1228         delta_delta = timespec64_sub(delta, old_delta);
1229         if (abs(delta_delta.tv_sec)  >= 2) {
1230                 /*
1231                  * if delta_delta is too large, assume time correction
1232                  * has occured and set old_delta to the current delta.
1233                  */
1234                 old_delta = delta;
1235         } else {
1236                 /* Otherwise try to adjust old_system to compensate */
1237                 timekeeping_suspend_time =
1238                         timespec64_add(timekeeping_suspend_time, delta_delta);
1239         }
1240
1241         timekeeping_update(tk, TK_MIRROR);
1242         write_seqcount_end(&tk_core.seq);
1243         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1244
1245         clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
1246         clocksource_suspend();
1247         clockevents_suspend();
1248
1249         return 0;
1250 }
1251
1252 /* sysfs resume/suspend bits for timekeeping */
1253 static struct syscore_ops timekeeping_syscore_ops = {
1254         .resume         = timekeeping_resume,
1255         .suspend        = timekeeping_suspend,
1256 };
1257
1258 static int __init timekeeping_init_ops(void)
1259 {
1260         register_syscore_ops(&timekeeping_syscore_ops);
1261         return 0;
1262 }
1263 device_initcall(timekeeping_init_ops);
1264
1265 /*
1266  * Apply a multiplier adjustment to the timekeeper
1267  */
1268 static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1269                                                          s64 offset,
1270                                                          bool negative,
1271                                                          int adj_scale)
1272 {
1273         s64 interval = tk->cycle_interval;
1274         s32 mult_adj = 1;
1275
1276         if (negative) {
1277                 mult_adj = -mult_adj;
1278                 interval = -interval;
1279                 offset  = -offset;
1280         }
1281         mult_adj <<= adj_scale;
1282         interval <<= adj_scale;
1283         offset <<= adj_scale;
1284
1285         /*
1286          * So the following can be confusing.
1287          *
1288          * To keep things simple, lets assume mult_adj == 1 for now.
1289          *
1290          * When mult_adj != 1, remember that the interval and offset values
1291          * have been appropriately scaled so the math is the same.
1292          *
1293          * The basic idea here is that we're increasing the multiplier
1294          * by one, this causes the xtime_interval to be incremented by
1295          * one cycle_interval. This is because:
1296          *      xtime_interval = cycle_interval * mult
1297          * So if mult is being incremented by one:
1298          *      xtime_interval = cycle_interval * (mult + 1)
1299          * Its the same as:
1300          *      xtime_interval = (cycle_interval * mult) + cycle_interval
1301          * Which can be shortened to:
1302          *      xtime_interval += cycle_interval
1303          *
1304          * So offset stores the non-accumulated cycles. Thus the current
1305          * time (in shifted nanoseconds) is:
1306          *      now = (offset * adj) + xtime_nsec
1307          * Now, even though we're adjusting the clock frequency, we have
1308          * to keep time consistent. In other words, we can't jump back
1309          * in time, and we also want to avoid jumping forward in time.
1310          *
1311          * So given the same offset value, we need the time to be the same
1312          * both before and after the freq adjustment.
1313          *      now = (offset * adj_1) + xtime_nsec_1
1314          *      now = (offset * adj_2) + xtime_nsec_2
1315          * So:
1316          *      (offset * adj_1) + xtime_nsec_1 =
1317          *              (offset * adj_2) + xtime_nsec_2
1318          * And we know:
1319          *      adj_2 = adj_1 + 1
1320          * So:
1321          *      (offset * adj_1) + xtime_nsec_1 =
1322          *              (offset * (adj_1+1)) + xtime_nsec_2
1323          *      (offset * adj_1) + xtime_nsec_1 =
1324          *              (offset * adj_1) + offset + xtime_nsec_2
1325          * Canceling the sides:
1326          *      xtime_nsec_1 = offset + xtime_nsec_2
1327          * Which gives us:
1328          *      xtime_nsec_2 = xtime_nsec_1 - offset
1329          * Which simplfies to:
1330          *      xtime_nsec -= offset
1331          *
1332          * XXX - TODO: Doc ntp_error calculation.
1333          */
1334         tk->tkr.mult += mult_adj;
1335         tk->xtime_interval += interval;
1336         tk->tkr.xtime_nsec -= offset;
1337         tk->ntp_error -= (interval - offset) << tk->ntp_error_shift;
1338 }
1339
1340 /*
1341  * Calculate the multiplier adjustment needed to match the frequency
1342  * specified by NTP
1343  */
1344 static __always_inline void timekeeping_freqadjust(struct timekeeper *tk,
1345                                                         s64 offset)
1346 {
1347         s64 interval = tk->cycle_interval;
1348         s64 xinterval = tk->xtime_interval;
1349         s64 tick_error;
1350         bool negative;
1351         u32 adj;
1352
1353         /* Remove any current error adj from freq calculation */
1354         if (tk->ntp_err_mult)
1355                 xinterval -= tk->cycle_interval;
1356
1357         tk->ntp_tick = ntp_tick_length();
1358
1359         /* Calculate current error per tick */
1360         tick_error = ntp_tick_length() >> tk->ntp_error_shift;
1361         tick_error -= (xinterval + tk->xtime_remainder);
1362
1363         /* Don't worry about correcting it if its small */
1364         if (likely((tick_error >= 0) && (tick_error <= interval)))
1365                 return;
1366
1367         /* preserve the direction of correction */
1368         negative = (tick_error < 0);
1369
1370         /* Sort out the magnitude of the correction */
1371         tick_error = abs(tick_error);
1372         for (adj = 0; tick_error > interval; adj++)
1373                 tick_error >>= 1;
1374
1375         /* scale the corrections */
1376         timekeeping_apply_adjustment(tk, offset, negative, adj);
1377 }
1378
1379 /*
1380  * Adjust the timekeeper's multiplier to the correct frequency
1381  * and also to reduce the accumulated error value.
1382  */
1383 static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
1384 {
1385         /* Correct for the current frequency error */
1386         timekeeping_freqadjust(tk, offset);
1387
1388         /* Next make a small adjustment to fix any cumulative error */
1389         if (!tk->ntp_err_mult && (tk->ntp_error > 0)) {
1390                 tk->ntp_err_mult = 1;
1391                 timekeeping_apply_adjustment(tk, offset, 0, 0);
1392         } else if (tk->ntp_err_mult && (tk->ntp_error <= 0)) {
1393                 /* Undo any existing error adjustment */
1394                 timekeeping_apply_adjustment(tk, offset, 1, 0);
1395                 tk->ntp_err_mult = 0;
1396         }
1397
1398         if (unlikely(tk->tkr.clock->maxadj &&
1399                 (tk->tkr.mult > tk->tkr.clock->mult + tk->tkr.clock->maxadj))) {
1400                 printk_once(KERN_WARNING
1401                         "Adjusting %s more than 11%% (%ld vs %ld)\n",
1402                         tk->tkr.clock->name, (long)tk->tkr.mult,
1403                         (long)tk->tkr.clock->mult + tk->tkr.clock->maxadj);
1404         }
1405
1406         /*
1407          * It may be possible that when we entered this function, xtime_nsec
1408          * was very small.  Further, if we're slightly speeding the clocksource
1409          * in the code above, its possible the required corrective factor to
1410          * xtime_nsec could cause it to underflow.
1411          *
1412          * Now, since we already accumulated the second, cannot simply roll
1413          * the accumulated second back, since the NTP subsystem has been
1414          * notified via second_overflow. So instead we push xtime_nsec forward
1415          * by the amount we underflowed, and add that amount into the error.
1416          *
1417          * We'll correct this error next time through this function, when
1418          * xtime_nsec is not as small.
1419          */
1420         if (unlikely((s64)tk->tkr.xtime_nsec < 0)) {
1421                 s64 neg = -(s64)tk->tkr.xtime_nsec;
1422                 tk->tkr.xtime_nsec = 0;
1423                 tk->ntp_error += neg << tk->ntp_error_shift;
1424         }
1425 }
1426
1427 /**
1428  * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1429  *
1430  * Helper function that accumulates a the nsecs greater then a second
1431  * from the xtime_nsec field to the xtime_secs field.
1432  * It also calls into the NTP code to handle leapsecond processing.
1433  *
1434  */
1435 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
1436 {
1437         u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr.shift;
1438         unsigned int clock_set = 0;
1439
1440         while (tk->tkr.xtime_nsec >= nsecps) {
1441                 int leap;
1442
1443                 tk->tkr.xtime_nsec -= nsecps;
1444                 tk->xtime_sec++;
1445
1446                 /* Figure out if its a leap sec and apply if needed */
1447                 leap = second_overflow(tk->xtime_sec);
1448                 if (unlikely(leap)) {
1449                         struct timespec64 ts;
1450
1451                         tk->xtime_sec += leap;
1452
1453                         ts.tv_sec = leap;
1454                         ts.tv_nsec = 0;
1455                         tk_set_wall_to_mono(tk,
1456                                 timespec64_sub(tk->wall_to_monotonic, ts));
1457
1458                         __timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
1459
1460                         clock_set = TK_CLOCK_WAS_SET;
1461                 }
1462         }
1463         return clock_set;
1464 }
1465
1466 /**
1467  * logarithmic_accumulation - shifted accumulation of cycles
1468  *
1469  * This functions accumulates a shifted interval of cycles into
1470  * into a shifted interval nanoseconds. Allows for O(log) accumulation
1471  * loop.
1472  *
1473  * Returns the unconsumed cycles.
1474  */
1475 static cycle_t logarithmic_accumulation(struct timekeeper *tk, cycle_t offset,
1476                                                 u32 shift,
1477                                                 unsigned int *clock_set)
1478 {
1479         cycle_t interval = tk->cycle_interval << shift;
1480         u64 raw_nsecs;
1481
1482         /* If the offset is smaller then a shifted interval, do nothing */
1483         if (offset < interval)
1484                 return offset;
1485
1486         /* Accumulate one shifted interval */
1487         offset -= interval;
1488         tk->tkr.cycle_last += interval;
1489
1490         tk->tkr.xtime_nsec += tk->xtime_interval << shift;
1491         *clock_set |= accumulate_nsecs_to_secs(tk);
1492
1493         /* Accumulate raw time */
1494         raw_nsecs = (u64)tk->raw_interval << shift;
1495         raw_nsecs += tk->raw_time.tv_nsec;
1496         if (raw_nsecs >= NSEC_PER_SEC) {
1497                 u64 raw_secs = raw_nsecs;
1498                 raw_nsecs = do_div(raw_secs, NSEC_PER_SEC);
1499                 tk->raw_time.tv_sec += raw_secs;
1500         }
1501         tk->raw_time.tv_nsec = raw_nsecs;
1502
1503         /* Accumulate error between NTP and clock interval */
1504         tk->ntp_error += tk->ntp_tick << shift;
1505         tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
1506                                                 (tk->ntp_error_shift + shift);
1507
1508         return offset;
1509 }
1510
1511 /**
1512  * update_wall_time - Uses the current clocksource to increment the wall time
1513  *
1514  */
1515 void update_wall_time(void)
1516 {
1517         struct timekeeper *real_tk = &tk_core.timekeeper;
1518         struct timekeeper *tk = &shadow_timekeeper;
1519         cycle_t offset;
1520         int shift = 0, maxshift;
1521         unsigned int clock_set = 0;
1522         unsigned long flags;
1523
1524         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1525
1526         /* Make sure we're fully resumed: */
1527         if (unlikely(timekeeping_suspended))
1528                 goto out;
1529
1530 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
1531         offset = real_tk->cycle_interval;
1532 #else
1533         offset = clocksource_delta(tk->tkr.read(tk->tkr.clock),
1534                                    tk->tkr.cycle_last, tk->tkr.mask);
1535 #endif
1536
1537         /* Check if there's really nothing to do */
1538         if (offset < real_tk->cycle_interval)
1539                 goto out;
1540
1541         /*
1542          * With NO_HZ we may have to accumulate many cycle_intervals
1543          * (think "ticks") worth of time at once. To do this efficiently,
1544          * we calculate the largest doubling multiple of cycle_intervals
1545          * that is smaller than the offset.  We then accumulate that
1546          * chunk in one go, and then try to consume the next smaller
1547          * doubled multiple.
1548          */
1549         shift = ilog2(offset) - ilog2(tk->cycle_interval);
1550         shift = max(0, shift);
1551         /* Bound shift to one less than what overflows tick_length */
1552         maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
1553         shift = min(shift, maxshift);
1554         while (offset >= tk->cycle_interval) {
1555                 offset = logarithmic_accumulation(tk, offset, shift,
1556                                                         &clock_set);
1557                 if (offset < tk->cycle_interval<<shift)
1558                         shift--;
1559         }
1560
1561         /* correct the clock when NTP error is too big */
1562         timekeeping_adjust(tk, offset);
1563
1564         /*
1565          * XXX This can be killed once everyone converts
1566          * to the new update_vsyscall.
1567          */
1568         old_vsyscall_fixup(tk);
1569
1570         /*
1571          * Finally, make sure that after the rounding
1572          * xtime_nsec isn't larger than NSEC_PER_SEC
1573          */
1574         clock_set |= accumulate_nsecs_to_secs(tk);
1575
1576         write_seqcount_begin(&tk_core.seq);
1577         /*
1578          * Update the real timekeeper.
1579          *
1580          * We could avoid this memcpy by switching pointers, but that
1581          * requires changes to all other timekeeper usage sites as
1582          * well, i.e. move the timekeeper pointer getter into the
1583          * spinlocked/seqcount protected sections. And we trade this
1584          * memcpy under the tk_core.seq against one before we start
1585          * updating.
1586          */
1587         memcpy(real_tk, tk, sizeof(*tk));
1588         timekeeping_update(real_tk, clock_set);
1589         write_seqcount_end(&tk_core.seq);
1590 out:
1591         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1592         if (clock_set)
1593                 /* Have to call _delayed version, since in irq context*/
1594                 clock_was_set_delayed();
1595 }
1596
1597 /**
1598  * getboottime - Return the real time of system boot.
1599  * @ts:         pointer to the timespec to be set
1600  *
1601  * Returns the wall-time of boot in a timespec.
1602  *
1603  * This is based on the wall_to_monotonic offset and the total suspend
1604  * time. Calls to settimeofday will affect the value returned (which
1605  * basically means that however wrong your real time clock is at boot time,
1606  * you get the right time here).
1607  */
1608 void getboottime(struct timespec *ts)
1609 {
1610         struct timekeeper *tk = &tk_core.timekeeper;
1611         ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
1612
1613         *ts = ktime_to_timespec(t);
1614 }
1615 EXPORT_SYMBOL_GPL(getboottime);
1616
1617 unsigned long get_seconds(void)
1618 {
1619         struct timekeeper *tk = &tk_core.timekeeper;
1620
1621         return tk->xtime_sec;
1622 }
1623 EXPORT_SYMBOL(get_seconds);
1624
1625 struct timespec __current_kernel_time(void)
1626 {
1627         struct timekeeper *tk = &tk_core.timekeeper;
1628
1629         return timespec64_to_timespec(tk_xtime(tk));
1630 }
1631
1632 struct timespec current_kernel_time(void)
1633 {
1634         struct timekeeper *tk = &tk_core.timekeeper;
1635         struct timespec64 now;
1636         unsigned long seq;
1637
1638         do {
1639                 seq = read_seqcount_begin(&tk_core.seq);
1640
1641                 now = tk_xtime(tk);
1642         } while (read_seqcount_retry(&tk_core.seq, seq));
1643
1644         return timespec64_to_timespec(now);
1645 }
1646 EXPORT_SYMBOL(current_kernel_time);
1647
1648 struct timespec get_monotonic_coarse(void)
1649 {
1650         struct timekeeper *tk = &tk_core.timekeeper;
1651         struct timespec64 now, mono;
1652         unsigned long seq;
1653
1654         do {
1655                 seq = read_seqcount_begin(&tk_core.seq);
1656
1657                 now = tk_xtime(tk);
1658                 mono = tk->wall_to_monotonic;
1659         } while (read_seqcount_retry(&tk_core.seq, seq));
1660
1661         set_normalized_timespec64(&now, now.tv_sec + mono.tv_sec,
1662                                 now.tv_nsec + mono.tv_nsec);
1663
1664         return timespec64_to_timespec(now);
1665 }
1666
1667 /*
1668  * Must hold jiffies_lock
1669  */
1670 void do_timer(unsigned long ticks)
1671 {
1672         jiffies_64 += ticks;
1673         calc_global_load(ticks);
1674 }
1675
1676 /**
1677  * ktime_get_update_offsets_tick - hrtimer helper
1678  * @offs_real:  pointer to storage for monotonic -> realtime offset
1679  * @offs_boot:  pointer to storage for monotonic -> boottime offset
1680  * @offs_tai:   pointer to storage for monotonic -> clock tai offset
1681  *
1682  * Returns monotonic time at last tick and various offsets
1683  */
1684 ktime_t ktime_get_update_offsets_tick(ktime_t *offs_real, ktime_t *offs_boot,
1685                                                         ktime_t *offs_tai)
1686 {
1687         struct timekeeper *tk = &tk_core.timekeeper;
1688         unsigned int seq;
1689         ktime_t base;
1690         u64 nsecs;
1691
1692         do {
1693                 seq = read_seqcount_begin(&tk_core.seq);
1694
1695                 base = tk->tkr.base_mono;
1696                 nsecs = tk->tkr.xtime_nsec >> tk->tkr.shift;
1697
1698                 *offs_real = tk->offs_real;
1699                 *offs_boot = tk->offs_boot;
1700                 *offs_tai = tk->offs_tai;
1701         } while (read_seqcount_retry(&tk_core.seq, seq));
1702
1703         return ktime_add_ns(base, nsecs);
1704 }
1705
1706 #ifdef CONFIG_HIGH_RES_TIMERS
1707 /**
1708  * ktime_get_update_offsets_now - hrtimer helper
1709  * @offs_real:  pointer to storage for monotonic -> realtime offset
1710  * @offs_boot:  pointer to storage for monotonic -> boottime offset
1711  * @offs_tai:   pointer to storage for monotonic -> clock tai offset
1712  *
1713  * Returns current monotonic time and updates the offsets
1714  * Called from hrtimer_interrupt() or retrigger_next_event()
1715  */
1716 ktime_t ktime_get_update_offsets_now(ktime_t *offs_real, ktime_t *offs_boot,
1717                                                         ktime_t *offs_tai)
1718 {
1719         struct timekeeper *tk = &tk_core.timekeeper;
1720         unsigned int seq;
1721         ktime_t base;
1722         u64 nsecs;
1723
1724         do {
1725                 seq = read_seqcount_begin(&tk_core.seq);
1726
1727                 base = tk->tkr.base_mono;
1728                 nsecs = timekeeping_get_ns(&tk->tkr);
1729
1730                 *offs_real = tk->offs_real;
1731                 *offs_boot = tk->offs_boot;
1732                 *offs_tai = tk->offs_tai;
1733         } while (read_seqcount_retry(&tk_core.seq, seq));
1734
1735         return ktime_add_ns(base, nsecs);
1736 }
1737 #endif
1738
1739 /**
1740  * do_adjtimex() - Accessor function to NTP __do_adjtimex function
1741  */
1742 int do_adjtimex(struct timex *txc)
1743 {
1744         struct timekeeper *tk = &tk_core.timekeeper;
1745         unsigned long flags;
1746         struct timespec64 ts;
1747         s32 orig_tai, tai;
1748         int ret;
1749
1750         /* Validate the data before disabling interrupts */
1751         ret = ntp_validate_timex(txc);
1752         if (ret)
1753                 return ret;
1754
1755         if (txc->modes & ADJ_SETOFFSET) {
1756                 struct timespec delta;
1757                 delta.tv_sec  = txc->time.tv_sec;
1758                 delta.tv_nsec = txc->time.tv_usec;
1759                 if (!(txc->modes & ADJ_NANO))
1760                         delta.tv_nsec *= 1000;
1761                 ret = timekeeping_inject_offset(&delta);
1762                 if (ret)
1763                         return ret;
1764         }
1765
1766         getnstimeofday64(&ts);
1767
1768         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1769         write_seqcount_begin(&tk_core.seq);
1770
1771         orig_tai = tai = tk->tai_offset;
1772         ret = __do_adjtimex(txc, &ts, &tai);
1773
1774         if (tai != orig_tai) {
1775                 __timekeeping_set_tai_offset(tk, tai);
1776                 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1777         }
1778         write_seqcount_end(&tk_core.seq);
1779         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1780
1781         if (tai != orig_tai)
1782                 clock_was_set();
1783
1784         ntp_notify_cmos_timer();
1785
1786         return ret;
1787 }
1788
1789 #ifdef CONFIG_NTP_PPS
1790 /**
1791  * hardpps() - Accessor function to NTP __hardpps function
1792  */
1793 void hardpps(const struct timespec *phase_ts, const struct timespec *raw_ts)
1794 {
1795         unsigned long flags;
1796
1797         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1798         write_seqcount_begin(&tk_core.seq);
1799
1800         __hardpps(phase_ts, raw_ts);
1801
1802         write_seqcount_end(&tk_core.seq);
1803         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1804 }
1805 EXPORT_SYMBOL(hardpps);
1806 #endif
1807
1808 /**
1809  * xtime_update() - advances the timekeeping infrastructure
1810  * @ticks:      number of ticks, that have elapsed since the last call.
1811  *
1812  * Must be called with interrupts disabled.
1813  */
1814 void xtime_update(unsigned long ticks)
1815 {
1816         write_seqlock(&jiffies_lock);
1817         do_timer(ticks);
1818         write_sequnlock(&jiffies_lock);
1819         update_wall_time();
1820 }