8ea4fb31571928913781b75b70b75210a46b3d8d
[sfrench/cifs-2.6.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/nmi.h>
18 #include <linux/sched.h>
19 #include <linux/sched/loadavg.h>
20 #include <linux/syscore_ops.h>
21 #include <linux/clocksource.h>
22 #include <linux/jiffies.h>
23 #include <linux/time.h>
24 #include <linux/tick.h>
25 #include <linux/stop_machine.h>
26 #include <linux/pvclock_gtod.h>
27 #include <linux/compiler.h>
28
29 #include "tick-internal.h"
30 #include "ntp_internal.h"
31 #include "timekeeping_internal.h"
32
33 #define TK_CLEAR_NTP            (1 << 0)
34 #define TK_MIRROR               (1 << 1)
35 #define TK_CLOCK_WAS_SET        (1 << 2)
36
37 /*
38  * The most important data for readout fits into a single 64 byte
39  * cache line.
40  */
41 static struct {
42         seqcount_t              seq;
43         struct timekeeper       timekeeper;
44 } tk_core ____cacheline_aligned;
45
46 static DEFINE_RAW_SPINLOCK(timekeeper_lock);
47 static struct timekeeper shadow_timekeeper;
48
49 /**
50  * struct tk_fast - NMI safe timekeeper
51  * @seq:        Sequence counter for protecting updates. The lowest bit
52  *              is the index for the tk_read_base array
53  * @base:       tk_read_base array. Access is indexed by the lowest bit of
54  *              @seq.
55  *
56  * See @update_fast_timekeeper() below.
57  */
58 struct tk_fast {
59         seqcount_t              seq;
60         struct tk_read_base     base[2];
61 };
62
63 static struct tk_fast tk_fast_mono ____cacheline_aligned;
64 static struct tk_fast tk_fast_raw  ____cacheline_aligned;
65
66 /* flag for if timekeeping is suspended */
67 int __read_mostly timekeeping_suspended;
68
69 static inline void tk_normalize_xtime(struct timekeeper *tk)
70 {
71         while (tk->tkr_mono.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_mono.shift)) {
72                 tk->tkr_mono.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
73                 tk->xtime_sec++;
74         }
75         while (tk->tkr_raw.xtime_nsec >= ((u64)NSEC_PER_SEC << tk->tkr_raw.shift)) {
76                 tk->tkr_raw.xtime_nsec -= (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
77                 tk->raw_sec++;
78         }
79 }
80
81 static inline struct timespec64 tk_xtime(struct timekeeper *tk)
82 {
83         struct timespec64 ts;
84
85         ts.tv_sec = tk->xtime_sec;
86         ts.tv_nsec = (long)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
87         return ts;
88 }
89
90 static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
91 {
92         tk->xtime_sec = ts->tv_sec;
93         tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift;
94 }
95
96 static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
97 {
98         tk->xtime_sec += ts->tv_sec;
99         tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift;
100         tk_normalize_xtime(tk);
101 }
102
103 static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
104 {
105         struct timespec64 tmp;
106
107         /*
108          * Verify consistency of: offset_real = -wall_to_monotonic
109          * before modifying anything
110          */
111         set_normalized_timespec64(&tmp, -tk->wall_to_monotonic.tv_sec,
112                                         -tk->wall_to_monotonic.tv_nsec);
113         WARN_ON_ONCE(tk->offs_real != timespec64_to_ktime(tmp));
114         tk->wall_to_monotonic = wtm;
115         set_normalized_timespec64(&tmp, -wtm.tv_sec, -wtm.tv_nsec);
116         tk->offs_real = timespec64_to_ktime(tmp);
117         tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tk->tai_offset, 0));
118 }
119
120 static inline void tk_update_sleep_time(struct timekeeper *tk, ktime_t delta)
121 {
122         tk->offs_boot = ktime_add(tk->offs_boot, delta);
123 }
124
125 /*
126  * tk_clock_read - atomic clocksource read() helper
127  *
128  * This helper is necessary to use in the read paths because, while the
129  * seqlock ensures we don't return a bad value while structures are updated,
130  * it doesn't protect from potential crashes. There is the possibility that
131  * the tkr's clocksource may change between the read reference, and the
132  * clock reference passed to the read function.  This can cause crashes if
133  * the wrong clocksource is passed to the wrong read function.
134  * This isn't necessary to use when holding the timekeeper_lock or doing
135  * a read of the fast-timekeeper tkrs (which is protected by its own locking
136  * and update logic).
137  */
138 static inline u64 tk_clock_read(struct tk_read_base *tkr)
139 {
140         struct clocksource *clock = READ_ONCE(tkr->clock);
141
142         return clock->read(clock);
143 }
144
145 #ifdef CONFIG_DEBUG_TIMEKEEPING
146 #define WARNING_FREQ (HZ*300) /* 5 minute rate-limiting */
147
148 static void timekeeping_check_update(struct timekeeper *tk, u64 offset)
149 {
150
151         u64 max_cycles = tk->tkr_mono.clock->max_cycles;
152         const char *name = tk->tkr_mono.clock->name;
153
154         if (offset > max_cycles) {
155                 printk_deferred("WARNING: timekeeping: Cycle offset (%lld) is larger than allowed by the '%s' clock's max_cycles value (%lld): time overflow danger\n",
156                                 offset, name, max_cycles);
157                 printk_deferred("         timekeeping: Your kernel is sick, but tries to cope by capping time updates\n");
158         } else {
159                 if (offset > (max_cycles >> 1)) {
160                         printk_deferred("INFO: timekeeping: Cycle offset (%lld) is larger than the '%s' clock's 50%% safety margin (%lld)\n",
161                                         offset, name, max_cycles >> 1);
162                         printk_deferred("      timekeeping: Your kernel is still fine, but is feeling a bit nervous\n");
163                 }
164         }
165
166         if (tk->underflow_seen) {
167                 if (jiffies - tk->last_warning > WARNING_FREQ) {
168                         printk_deferred("WARNING: Underflow in clocksource '%s' observed, time update ignored.\n", name);
169                         printk_deferred("         Please report this, consider using a different clocksource, if possible.\n");
170                         printk_deferred("         Your kernel is probably still fine.\n");
171                         tk->last_warning = jiffies;
172                 }
173                 tk->underflow_seen = 0;
174         }
175
176         if (tk->overflow_seen) {
177                 if (jiffies - tk->last_warning > WARNING_FREQ) {
178                         printk_deferred("WARNING: Overflow in clocksource '%s' observed, time update capped.\n", name);
179                         printk_deferred("         Please report this, consider using a different clocksource, if possible.\n");
180                         printk_deferred("         Your kernel is probably still fine.\n");
181                         tk->last_warning = jiffies;
182                 }
183                 tk->overflow_seen = 0;
184         }
185 }
186
187 static inline u64 timekeeping_get_delta(struct tk_read_base *tkr)
188 {
189         struct timekeeper *tk = &tk_core.timekeeper;
190         u64 now, last, mask, max, delta;
191         unsigned int seq;
192
193         /*
194          * Since we're called holding a seqlock, the data may shift
195          * under us while we're doing the calculation. This can cause
196          * false positives, since we'd note a problem but throw the
197          * results away. So nest another seqlock here to atomically
198          * grab the points we are checking with.
199          */
200         do {
201                 seq = read_seqcount_begin(&tk_core.seq);
202                 now = tk_clock_read(tkr);
203                 last = tkr->cycle_last;
204                 mask = tkr->mask;
205                 max = tkr->clock->max_cycles;
206         } while (read_seqcount_retry(&tk_core.seq, seq));
207
208         delta = clocksource_delta(now, last, mask);
209
210         /*
211          * Try to catch underflows by checking if we are seeing small
212          * mask-relative negative values.
213          */
214         if (unlikely((~delta & mask) < (mask >> 3))) {
215                 tk->underflow_seen = 1;
216                 delta = 0;
217         }
218
219         /* Cap delta value to the max_cycles values to avoid mult overflows */
220         if (unlikely(delta > max)) {
221                 tk->overflow_seen = 1;
222                 delta = tkr->clock->max_cycles;
223         }
224
225         return delta;
226 }
227 #else
228 static inline void timekeeping_check_update(struct timekeeper *tk, u64 offset)
229 {
230 }
231 static inline u64 timekeeping_get_delta(struct tk_read_base *tkr)
232 {
233         u64 cycle_now, delta;
234
235         /* read clocksource */
236         cycle_now = tk_clock_read(tkr);
237
238         /* calculate the delta since the last update_wall_time */
239         delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
240
241         return delta;
242 }
243 #endif
244
245 /**
246  * tk_setup_internals - Set up internals to use clocksource clock.
247  *
248  * @tk:         The target timekeeper to setup.
249  * @clock:              Pointer to clocksource.
250  *
251  * Calculates a fixed cycle/nsec interval for a given clocksource/adjustment
252  * pair and interval request.
253  *
254  * Unless you're the timekeeping code, you should not be using this!
255  */
256 static void tk_setup_internals(struct timekeeper *tk, struct clocksource *clock)
257 {
258         u64 interval;
259         u64 tmp, ntpinterval;
260         struct clocksource *old_clock;
261
262         ++tk->cs_was_changed_seq;
263         old_clock = tk->tkr_mono.clock;
264         tk->tkr_mono.clock = clock;
265         tk->tkr_mono.mask = clock->mask;
266         tk->tkr_mono.cycle_last = tk_clock_read(&tk->tkr_mono);
267
268         tk->tkr_raw.clock = clock;
269         tk->tkr_raw.mask = clock->mask;
270         tk->tkr_raw.cycle_last = tk->tkr_mono.cycle_last;
271
272         /* Do the ns -> cycle conversion first, using original mult */
273         tmp = NTP_INTERVAL_LENGTH;
274         tmp <<= clock->shift;
275         ntpinterval = tmp;
276         tmp += clock->mult/2;
277         do_div(tmp, clock->mult);
278         if (tmp == 0)
279                 tmp = 1;
280
281         interval = (u64) tmp;
282         tk->cycle_interval = interval;
283
284         /* Go back from cycles -> shifted ns */
285         tk->xtime_interval = interval * clock->mult;
286         tk->xtime_remainder = ntpinterval - tk->xtime_interval;
287         tk->raw_interval = interval * clock->mult;
288
289          /* if changing clocks, convert xtime_nsec shift units */
290         if (old_clock) {
291                 int shift_change = clock->shift - old_clock->shift;
292                 if (shift_change < 0) {
293                         tk->tkr_mono.xtime_nsec >>= -shift_change;
294                         tk->tkr_raw.xtime_nsec >>= -shift_change;
295                 } else {
296                         tk->tkr_mono.xtime_nsec <<= shift_change;
297                         tk->tkr_raw.xtime_nsec <<= shift_change;
298                 }
299         }
300
301         tk->tkr_mono.shift = clock->shift;
302         tk->tkr_raw.shift = clock->shift;
303
304         tk->ntp_error = 0;
305         tk->ntp_error_shift = NTP_SCALE_SHIFT - clock->shift;
306         tk->ntp_tick = ntpinterval << tk->ntp_error_shift;
307
308         /*
309          * The timekeeper keeps its own mult values for the currently
310          * active clocksource. These value will be adjusted via NTP
311          * to counteract clock drifting.
312          */
313         tk->tkr_mono.mult = clock->mult;
314         tk->tkr_raw.mult = clock->mult;
315         tk->ntp_err_mult = 0;
316 }
317
318 /* Timekeeper helper functions. */
319
320 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
321 static u32 default_arch_gettimeoffset(void) { return 0; }
322 u32 (*arch_gettimeoffset)(void) = default_arch_gettimeoffset;
323 #else
324 static inline u32 arch_gettimeoffset(void) { return 0; }
325 #endif
326
327 static inline u64 timekeeping_delta_to_ns(struct tk_read_base *tkr, u64 delta)
328 {
329         u64 nsec;
330
331         nsec = delta * tkr->mult + tkr->xtime_nsec;
332         nsec >>= tkr->shift;
333
334         /* If arch requires, add in get_arch_timeoffset() */
335         return nsec + arch_gettimeoffset();
336 }
337
338 static inline u64 timekeeping_get_ns(struct tk_read_base *tkr)
339 {
340         u64 delta;
341
342         delta = timekeeping_get_delta(tkr);
343         return timekeeping_delta_to_ns(tkr, delta);
344 }
345
346 static inline u64 timekeeping_cycles_to_ns(struct tk_read_base *tkr, u64 cycles)
347 {
348         u64 delta;
349
350         /* calculate the delta since the last update_wall_time */
351         delta = clocksource_delta(cycles, tkr->cycle_last, tkr->mask);
352         return timekeeping_delta_to_ns(tkr, delta);
353 }
354
355 /**
356  * update_fast_timekeeper - Update the fast and NMI safe monotonic timekeeper.
357  * @tkr: Timekeeping readout base from which we take the update
358  *
359  * We want to use this from any context including NMI and tracing /
360  * instrumenting the timekeeping code itself.
361  *
362  * Employ the latch technique; see @raw_write_seqcount_latch.
363  *
364  * So if a NMI hits the update of base[0] then it will use base[1]
365  * which is still consistent. In the worst case this can result is a
366  * slightly wrong timestamp (a few nanoseconds). See
367  * @ktime_get_mono_fast_ns.
368  */
369 static void update_fast_timekeeper(struct tk_read_base *tkr, struct tk_fast *tkf)
370 {
371         struct tk_read_base *base = tkf->base;
372
373         /* Force readers off to base[1] */
374         raw_write_seqcount_latch(&tkf->seq);
375
376         /* Update base[0] */
377         memcpy(base, tkr, sizeof(*base));
378
379         /* Force readers back to base[0] */
380         raw_write_seqcount_latch(&tkf->seq);
381
382         /* Update base[1] */
383         memcpy(base + 1, base, sizeof(*base));
384 }
385
386 /**
387  * ktime_get_mono_fast_ns - Fast NMI safe access to clock monotonic
388  *
389  * This timestamp is not guaranteed to be monotonic across an update.
390  * The timestamp is calculated by:
391  *
392  *      now = base_mono + clock_delta * slope
393  *
394  * So if the update lowers the slope, readers who are forced to the
395  * not yet updated second array are still using the old steeper slope.
396  *
397  * tmono
398  * ^
399  * |    o  n
400  * |   o n
401  * |  u
402  * | o
403  * |o
404  * |12345678---> reader order
405  *
406  * o = old slope
407  * u = update
408  * n = new slope
409  *
410  * So reader 6 will observe time going backwards versus reader 5.
411  *
412  * While other CPUs are likely to be able observe that, the only way
413  * for a CPU local observation is when an NMI hits in the middle of
414  * the update. Timestamps taken from that NMI context might be ahead
415  * of the following timestamps. Callers need to be aware of that and
416  * deal with it.
417  */
418 static __always_inline u64 __ktime_get_fast_ns(struct tk_fast *tkf)
419 {
420         struct tk_read_base *tkr;
421         unsigned int seq;
422         u64 now;
423
424         do {
425                 seq = raw_read_seqcount_latch(&tkf->seq);
426                 tkr = tkf->base + (seq & 0x01);
427                 now = ktime_to_ns(tkr->base);
428
429                 now += timekeeping_delta_to_ns(tkr,
430                                 clocksource_delta(
431                                         tk_clock_read(tkr),
432                                         tkr->cycle_last,
433                                         tkr->mask));
434         } while (read_seqcount_retry(&tkf->seq, seq));
435
436         return now;
437 }
438
439 u64 ktime_get_mono_fast_ns(void)
440 {
441         return __ktime_get_fast_ns(&tk_fast_mono);
442 }
443 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
444
445 u64 ktime_get_raw_fast_ns(void)
446 {
447         return __ktime_get_fast_ns(&tk_fast_raw);
448 }
449 EXPORT_SYMBOL_GPL(ktime_get_raw_fast_ns);
450
451 /**
452  * ktime_get_boot_fast_ns - NMI safe and fast access to boot clock.
453  *
454  * To keep it NMI safe since we're accessing from tracing, we're not using a
455  * separate timekeeper with updates to monotonic clock and boot offset
456  * protected with seqlocks. This has the following minor side effects:
457  *
458  * (1) Its possible that a timestamp be taken after the boot offset is updated
459  * but before the timekeeper is updated. If this happens, the new boot offset
460  * is added to the old timekeeping making the clock appear to update slightly
461  * earlier:
462  *    CPU 0                                        CPU 1
463  *    timekeeping_inject_sleeptime64()
464  *    __timekeeping_inject_sleeptime(tk, delta);
465  *                                                 timestamp();
466  *    timekeeping_update(tk, TK_CLEAR_NTP...);
467  *
468  * (2) On 32-bit systems, the 64-bit boot offset (tk->offs_boot) may be
469  * partially updated.  Since the tk->offs_boot update is a rare event, this
470  * should be a rare occurrence which postprocessing should be able to handle.
471  */
472 u64 notrace ktime_get_boot_fast_ns(void)
473 {
474         struct timekeeper *tk = &tk_core.timekeeper;
475
476         return (ktime_get_mono_fast_ns() + ktime_to_ns(tk->offs_boot));
477 }
478 EXPORT_SYMBOL_GPL(ktime_get_boot_fast_ns);
479
480 /* Suspend-time cycles value for halted fast timekeeper. */
481 static u64 cycles_at_suspend;
482
483 static u64 dummy_clock_read(struct clocksource *cs)
484 {
485         return cycles_at_suspend;
486 }
487
488 static struct clocksource dummy_clock = {
489         .read = dummy_clock_read,
490 };
491
492 /**
493  * halt_fast_timekeeper - Prevent fast timekeeper from accessing clocksource.
494  * @tk: Timekeeper to snapshot.
495  *
496  * It generally is unsafe to access the clocksource after timekeeping has been
497  * suspended, so take a snapshot of the readout base of @tk and use it as the
498  * fast timekeeper's readout base while suspended.  It will return the same
499  * number of cycles every time until timekeeping is resumed at which time the
500  * proper readout base for the fast timekeeper will be restored automatically.
501  */
502 static void halt_fast_timekeeper(struct timekeeper *tk)
503 {
504         static struct tk_read_base tkr_dummy;
505         struct tk_read_base *tkr = &tk->tkr_mono;
506
507         memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
508         cycles_at_suspend = tk_clock_read(tkr);
509         tkr_dummy.clock = &dummy_clock;
510         update_fast_timekeeper(&tkr_dummy, &tk_fast_mono);
511
512         tkr = &tk->tkr_raw;
513         memcpy(&tkr_dummy, tkr, sizeof(tkr_dummy));
514         tkr_dummy.clock = &dummy_clock;
515         update_fast_timekeeper(&tkr_dummy, &tk_fast_raw);
516 }
517
518 #ifdef CONFIG_GENERIC_TIME_VSYSCALL_OLD
519 #warning Please contact your maintainers, as GENERIC_TIME_VSYSCALL_OLD compatibity will disappear soon.
520
521 static inline void update_vsyscall(struct timekeeper *tk)
522 {
523         struct timespec xt, wm;
524
525         xt = timespec64_to_timespec(tk_xtime(tk));
526         wm = timespec64_to_timespec(tk->wall_to_monotonic);
527         update_vsyscall_old(&xt, &wm, tk->tkr_mono.clock, tk->tkr_mono.mult,
528                             tk->tkr_mono.cycle_last);
529 }
530
531 static inline void old_vsyscall_fixup(struct timekeeper *tk)
532 {
533         s64 remainder;
534
535         /*
536         * Store only full nanoseconds into xtime_nsec after rounding
537         * it up and add the remainder to the error difference.
538         * XXX - This is necessary to avoid small 1ns inconsistnecies caused
539         * by truncating the remainder in vsyscalls. However, it causes
540         * additional work to be done in timekeeping_adjust(). Once
541         * the vsyscall implementations are converted to use xtime_nsec
542         * (shifted nanoseconds), and CONFIG_GENERIC_TIME_VSYSCALL_OLD
543         * users are removed, this can be killed.
544         */
545         remainder = tk->tkr_mono.xtime_nsec & ((1ULL << tk->tkr_mono.shift) - 1);
546         if (remainder != 0) {
547                 tk->tkr_mono.xtime_nsec -= remainder;
548                 tk->tkr_mono.xtime_nsec += 1ULL << tk->tkr_mono.shift;
549                 tk->ntp_error += remainder << tk->ntp_error_shift;
550                 tk->ntp_error -= (1ULL << tk->tkr_mono.shift) << tk->ntp_error_shift;
551         }
552 }
553 #else
554 #define old_vsyscall_fixup(tk)
555 #endif
556
557 static RAW_NOTIFIER_HEAD(pvclock_gtod_chain);
558
559 static void update_pvclock_gtod(struct timekeeper *tk, bool was_set)
560 {
561         raw_notifier_call_chain(&pvclock_gtod_chain, was_set, tk);
562 }
563
564 /**
565  * pvclock_gtod_register_notifier - register a pvclock timedata update listener
566  */
567 int pvclock_gtod_register_notifier(struct notifier_block *nb)
568 {
569         struct timekeeper *tk = &tk_core.timekeeper;
570         unsigned long flags;
571         int ret;
572
573         raw_spin_lock_irqsave(&timekeeper_lock, flags);
574         ret = raw_notifier_chain_register(&pvclock_gtod_chain, nb);
575         update_pvclock_gtod(tk, true);
576         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
577
578         return ret;
579 }
580 EXPORT_SYMBOL_GPL(pvclock_gtod_register_notifier);
581
582 /**
583  * pvclock_gtod_unregister_notifier - unregister a pvclock
584  * timedata update listener
585  */
586 int pvclock_gtod_unregister_notifier(struct notifier_block *nb)
587 {
588         unsigned long flags;
589         int ret;
590
591         raw_spin_lock_irqsave(&timekeeper_lock, flags);
592         ret = raw_notifier_chain_unregister(&pvclock_gtod_chain, nb);
593         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
594
595         return ret;
596 }
597 EXPORT_SYMBOL_GPL(pvclock_gtod_unregister_notifier);
598
599 /*
600  * tk_update_leap_state - helper to update the next_leap_ktime
601  */
602 static inline void tk_update_leap_state(struct timekeeper *tk)
603 {
604         tk->next_leap_ktime = ntp_get_next_leap();
605         if (tk->next_leap_ktime != KTIME_MAX)
606                 /* Convert to monotonic time */
607                 tk->next_leap_ktime = ktime_sub(tk->next_leap_ktime, tk->offs_real);
608 }
609
610 /*
611  * Update the ktime_t based scalar nsec members of the timekeeper
612  */
613 static inline void tk_update_ktime_data(struct timekeeper *tk)
614 {
615         u64 seconds;
616         u32 nsec;
617
618         /*
619          * The xtime based monotonic readout is:
620          *      nsec = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec + now();
621          * The ktime based monotonic readout is:
622          *      nsec = base_mono + now();
623          * ==> base_mono = (xtime_sec + wtm_sec) * 1e9 + wtm_nsec
624          */
625         seconds = (u64)(tk->xtime_sec + tk->wall_to_monotonic.tv_sec);
626         nsec = (u32) tk->wall_to_monotonic.tv_nsec;
627         tk->tkr_mono.base = ns_to_ktime(seconds * NSEC_PER_SEC + nsec);
628
629         /*
630          * The sum of the nanoseconds portions of xtime and
631          * wall_to_monotonic can be greater/equal one second. Take
632          * this into account before updating tk->ktime_sec.
633          */
634         nsec += (u32)(tk->tkr_mono.xtime_nsec >> tk->tkr_mono.shift);
635         if (nsec >= NSEC_PER_SEC)
636                 seconds++;
637         tk->ktime_sec = seconds;
638
639         /* Update the monotonic raw base */
640         tk->tkr_raw.base = ns_to_ktime(tk->raw_sec * NSEC_PER_SEC);
641 }
642
643 /* must hold timekeeper_lock */
644 static void timekeeping_update(struct timekeeper *tk, unsigned int action)
645 {
646         if (action & TK_CLEAR_NTP) {
647                 tk->ntp_error = 0;
648                 ntp_clear();
649         }
650
651         tk_update_leap_state(tk);
652         tk_update_ktime_data(tk);
653
654         update_vsyscall(tk);
655         update_pvclock_gtod(tk, action & TK_CLOCK_WAS_SET);
656
657         update_fast_timekeeper(&tk->tkr_mono, &tk_fast_mono);
658         update_fast_timekeeper(&tk->tkr_raw,  &tk_fast_raw);
659
660         if (action & TK_CLOCK_WAS_SET)
661                 tk->clock_was_set_seq++;
662         /*
663          * The mirroring of the data to the shadow-timekeeper needs
664          * to happen last here to ensure we don't over-write the
665          * timekeeper structure on the next update with stale data
666          */
667         if (action & TK_MIRROR)
668                 memcpy(&shadow_timekeeper, &tk_core.timekeeper,
669                        sizeof(tk_core.timekeeper));
670 }
671
672 /**
673  * timekeeping_forward_now - update clock to the current time
674  *
675  * Forward the current clock to update its state since the last call to
676  * update_wall_time(). This is useful before significant clock changes,
677  * as it avoids having to deal with this time offset explicitly.
678  */
679 static void timekeeping_forward_now(struct timekeeper *tk)
680 {
681         u64 cycle_now, delta;
682
683         cycle_now = tk_clock_read(&tk->tkr_mono);
684         delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
685         tk->tkr_mono.cycle_last = cycle_now;
686         tk->tkr_raw.cycle_last  = cycle_now;
687
688         tk->tkr_mono.xtime_nsec += delta * tk->tkr_mono.mult;
689
690         /* If arch requires, add in get_arch_timeoffset() */
691         tk->tkr_mono.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_mono.shift;
692
693
694         tk->tkr_raw.xtime_nsec += delta * tk->tkr_raw.mult;
695
696         /* If arch requires, add in get_arch_timeoffset() */
697         tk->tkr_raw.xtime_nsec += (u64)arch_gettimeoffset() << tk->tkr_raw.shift;
698
699         tk_normalize_xtime(tk);
700 }
701
702 /**
703  * __getnstimeofday64 - Returns the time of day in a timespec64.
704  * @ts:         pointer to the timespec to be set
705  *
706  * Updates the time of day in the timespec.
707  * Returns 0 on success, or -ve when suspended (timespec will be undefined).
708  */
709 int __getnstimeofday64(struct timespec64 *ts)
710 {
711         struct timekeeper *tk = &tk_core.timekeeper;
712         unsigned long seq;
713         u64 nsecs;
714
715         do {
716                 seq = read_seqcount_begin(&tk_core.seq);
717
718                 ts->tv_sec = tk->xtime_sec;
719                 nsecs = timekeeping_get_ns(&tk->tkr_mono);
720
721         } while (read_seqcount_retry(&tk_core.seq, seq));
722
723         ts->tv_nsec = 0;
724         timespec64_add_ns(ts, nsecs);
725
726         /*
727          * Do not bail out early, in case there were callers still using
728          * the value, even in the face of the WARN_ON.
729          */
730         if (unlikely(timekeeping_suspended))
731                 return -EAGAIN;
732         return 0;
733 }
734 EXPORT_SYMBOL(__getnstimeofday64);
735
736 /**
737  * getnstimeofday64 - Returns the time of day in a timespec64.
738  * @ts:         pointer to the timespec64 to be set
739  *
740  * Returns the time of day in a timespec64 (WARN if suspended).
741  */
742 void getnstimeofday64(struct timespec64 *ts)
743 {
744         WARN_ON(__getnstimeofday64(ts));
745 }
746 EXPORT_SYMBOL(getnstimeofday64);
747
748 ktime_t ktime_get(void)
749 {
750         struct timekeeper *tk = &tk_core.timekeeper;
751         unsigned int seq;
752         ktime_t base;
753         u64 nsecs;
754
755         WARN_ON(timekeeping_suspended);
756
757         do {
758                 seq = read_seqcount_begin(&tk_core.seq);
759                 base = tk->tkr_mono.base;
760                 nsecs = timekeeping_get_ns(&tk->tkr_mono);
761
762         } while (read_seqcount_retry(&tk_core.seq, seq));
763
764         return ktime_add_ns(base, nsecs);
765 }
766 EXPORT_SYMBOL_GPL(ktime_get);
767
768 u32 ktime_get_resolution_ns(void)
769 {
770         struct timekeeper *tk = &tk_core.timekeeper;
771         unsigned int seq;
772         u32 nsecs;
773
774         WARN_ON(timekeeping_suspended);
775
776         do {
777                 seq = read_seqcount_begin(&tk_core.seq);
778                 nsecs = tk->tkr_mono.mult >> tk->tkr_mono.shift;
779         } while (read_seqcount_retry(&tk_core.seq, seq));
780
781         return nsecs;
782 }
783 EXPORT_SYMBOL_GPL(ktime_get_resolution_ns);
784
785 static ktime_t *offsets[TK_OFFS_MAX] = {
786         [TK_OFFS_REAL]  = &tk_core.timekeeper.offs_real,
787         [TK_OFFS_BOOT]  = &tk_core.timekeeper.offs_boot,
788         [TK_OFFS_TAI]   = &tk_core.timekeeper.offs_tai,
789 };
790
791 ktime_t ktime_get_with_offset(enum tk_offsets offs)
792 {
793         struct timekeeper *tk = &tk_core.timekeeper;
794         unsigned int seq;
795         ktime_t base, *offset = offsets[offs];
796         u64 nsecs;
797
798         WARN_ON(timekeeping_suspended);
799
800         do {
801                 seq = read_seqcount_begin(&tk_core.seq);
802                 base = ktime_add(tk->tkr_mono.base, *offset);
803                 nsecs = timekeeping_get_ns(&tk->tkr_mono);
804
805         } while (read_seqcount_retry(&tk_core.seq, seq));
806
807         return ktime_add_ns(base, nsecs);
808
809 }
810 EXPORT_SYMBOL_GPL(ktime_get_with_offset);
811
812 /**
813  * ktime_mono_to_any() - convert mononotic time to any other time
814  * @tmono:      time to convert.
815  * @offs:       which offset to use
816  */
817 ktime_t ktime_mono_to_any(ktime_t tmono, enum tk_offsets offs)
818 {
819         ktime_t *offset = offsets[offs];
820         unsigned long seq;
821         ktime_t tconv;
822
823         do {
824                 seq = read_seqcount_begin(&tk_core.seq);
825                 tconv = ktime_add(tmono, *offset);
826         } while (read_seqcount_retry(&tk_core.seq, seq));
827
828         return tconv;
829 }
830 EXPORT_SYMBOL_GPL(ktime_mono_to_any);
831
832 /**
833  * ktime_get_raw - Returns the raw monotonic time in ktime_t format
834  */
835 ktime_t ktime_get_raw(void)
836 {
837         struct timekeeper *tk = &tk_core.timekeeper;
838         unsigned int seq;
839         ktime_t base;
840         u64 nsecs;
841
842         do {
843                 seq = read_seqcount_begin(&tk_core.seq);
844                 base = tk->tkr_raw.base;
845                 nsecs = timekeeping_get_ns(&tk->tkr_raw);
846
847         } while (read_seqcount_retry(&tk_core.seq, seq));
848
849         return ktime_add_ns(base, nsecs);
850 }
851 EXPORT_SYMBOL_GPL(ktime_get_raw);
852
853 /**
854  * ktime_get_ts64 - get the monotonic clock in timespec64 format
855  * @ts:         pointer to timespec variable
856  *
857  * The function calculates the monotonic clock from the realtime
858  * clock and the wall_to_monotonic offset and stores the result
859  * in normalized timespec64 format in the variable pointed to by @ts.
860  */
861 void ktime_get_ts64(struct timespec64 *ts)
862 {
863         struct timekeeper *tk = &tk_core.timekeeper;
864         struct timespec64 tomono;
865         unsigned int seq;
866         u64 nsec;
867
868         WARN_ON(timekeeping_suspended);
869
870         do {
871                 seq = read_seqcount_begin(&tk_core.seq);
872                 ts->tv_sec = tk->xtime_sec;
873                 nsec = timekeeping_get_ns(&tk->tkr_mono);
874                 tomono = tk->wall_to_monotonic;
875
876         } while (read_seqcount_retry(&tk_core.seq, seq));
877
878         ts->tv_sec += tomono.tv_sec;
879         ts->tv_nsec = 0;
880         timespec64_add_ns(ts, nsec + tomono.tv_nsec);
881 }
882 EXPORT_SYMBOL_GPL(ktime_get_ts64);
883
884 /**
885  * ktime_get_seconds - Get the seconds portion of CLOCK_MONOTONIC
886  *
887  * Returns the seconds portion of CLOCK_MONOTONIC with a single non
888  * serialized read. tk->ktime_sec is of type 'unsigned long' so this
889  * works on both 32 and 64 bit systems. On 32 bit systems the readout
890  * covers ~136 years of uptime which should be enough to prevent
891  * premature wrap arounds.
892  */
893 time64_t ktime_get_seconds(void)
894 {
895         struct timekeeper *tk = &tk_core.timekeeper;
896
897         WARN_ON(timekeeping_suspended);
898         return tk->ktime_sec;
899 }
900 EXPORT_SYMBOL_GPL(ktime_get_seconds);
901
902 /**
903  * ktime_get_real_seconds - Get the seconds portion of CLOCK_REALTIME
904  *
905  * Returns the wall clock seconds since 1970. This replaces the
906  * get_seconds() interface which is not y2038 safe on 32bit systems.
907  *
908  * For 64bit systems the fast access to tk->xtime_sec is preserved. On
909  * 32bit systems the access must be protected with the sequence
910  * counter to provide "atomic" access to the 64bit tk->xtime_sec
911  * value.
912  */
913 time64_t ktime_get_real_seconds(void)
914 {
915         struct timekeeper *tk = &tk_core.timekeeper;
916         time64_t seconds;
917         unsigned int seq;
918
919         if (IS_ENABLED(CONFIG_64BIT))
920                 return tk->xtime_sec;
921
922         do {
923                 seq = read_seqcount_begin(&tk_core.seq);
924                 seconds = tk->xtime_sec;
925
926         } while (read_seqcount_retry(&tk_core.seq, seq));
927
928         return seconds;
929 }
930 EXPORT_SYMBOL_GPL(ktime_get_real_seconds);
931
932 /**
933  * __ktime_get_real_seconds - The same as ktime_get_real_seconds
934  * but without the sequence counter protect. This internal function
935  * is called just when timekeeping lock is already held.
936  */
937 time64_t __ktime_get_real_seconds(void)
938 {
939         struct timekeeper *tk = &tk_core.timekeeper;
940
941         return tk->xtime_sec;
942 }
943
944 /**
945  * ktime_get_snapshot - snapshots the realtime/monotonic raw clocks with counter
946  * @systime_snapshot:   pointer to struct receiving the system time snapshot
947  */
948 void ktime_get_snapshot(struct system_time_snapshot *systime_snapshot)
949 {
950         struct timekeeper *tk = &tk_core.timekeeper;
951         unsigned long seq;
952         ktime_t base_raw;
953         ktime_t base_real;
954         u64 nsec_raw;
955         u64 nsec_real;
956         u64 now;
957
958         WARN_ON_ONCE(timekeeping_suspended);
959
960         do {
961                 seq = read_seqcount_begin(&tk_core.seq);
962                 now = tk_clock_read(&tk->tkr_mono);
963                 systime_snapshot->cs_was_changed_seq = tk->cs_was_changed_seq;
964                 systime_snapshot->clock_was_set_seq = tk->clock_was_set_seq;
965                 base_real = ktime_add(tk->tkr_mono.base,
966                                       tk_core.timekeeper.offs_real);
967                 base_raw = tk->tkr_raw.base;
968                 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono, now);
969                 nsec_raw  = timekeeping_cycles_to_ns(&tk->tkr_raw, now);
970         } while (read_seqcount_retry(&tk_core.seq, seq));
971
972         systime_snapshot->cycles = now;
973         systime_snapshot->real = ktime_add_ns(base_real, nsec_real);
974         systime_snapshot->raw = ktime_add_ns(base_raw, nsec_raw);
975 }
976 EXPORT_SYMBOL_GPL(ktime_get_snapshot);
977
978 /* Scale base by mult/div checking for overflow */
979 static int scale64_check_overflow(u64 mult, u64 div, u64 *base)
980 {
981         u64 tmp, rem;
982
983         tmp = div64_u64_rem(*base, div, &rem);
984
985         if (((int)sizeof(u64)*8 - fls64(mult) < fls64(tmp)) ||
986             ((int)sizeof(u64)*8 - fls64(mult) < fls64(rem)))
987                 return -EOVERFLOW;
988         tmp *= mult;
989         rem *= mult;
990
991         do_div(rem, div);
992         *base = tmp + rem;
993         return 0;
994 }
995
996 /**
997  * adjust_historical_crosststamp - adjust crosstimestamp previous to current interval
998  * @history:                    Snapshot representing start of history
999  * @partial_history_cycles:     Cycle offset into history (fractional part)
1000  * @total_history_cycles:       Total history length in cycles
1001  * @discontinuity:              True indicates clock was set on history period
1002  * @ts:                         Cross timestamp that should be adjusted using
1003  *      partial/total ratio
1004  *
1005  * Helper function used by get_device_system_crosststamp() to correct the
1006  * crosstimestamp corresponding to the start of the current interval to the
1007  * system counter value (timestamp point) provided by the driver. The
1008  * total_history_* quantities are the total history starting at the provided
1009  * reference point and ending at the start of the current interval. The cycle
1010  * count between the driver timestamp point and the start of the current
1011  * interval is partial_history_cycles.
1012  */
1013 static int adjust_historical_crosststamp(struct system_time_snapshot *history,
1014                                          u64 partial_history_cycles,
1015                                          u64 total_history_cycles,
1016                                          bool discontinuity,
1017                                          struct system_device_crosststamp *ts)
1018 {
1019         struct timekeeper *tk = &tk_core.timekeeper;
1020         u64 corr_raw, corr_real;
1021         bool interp_forward;
1022         int ret;
1023
1024         if (total_history_cycles == 0 || partial_history_cycles == 0)
1025                 return 0;
1026
1027         /* Interpolate shortest distance from beginning or end of history */
1028         interp_forward = partial_history_cycles > total_history_cycles / 2;
1029         partial_history_cycles = interp_forward ?
1030                 total_history_cycles - partial_history_cycles :
1031                 partial_history_cycles;
1032
1033         /*
1034          * Scale the monotonic raw time delta by:
1035          *      partial_history_cycles / total_history_cycles
1036          */
1037         corr_raw = (u64)ktime_to_ns(
1038                 ktime_sub(ts->sys_monoraw, history->raw));
1039         ret = scale64_check_overflow(partial_history_cycles,
1040                                      total_history_cycles, &corr_raw);
1041         if (ret)
1042                 return ret;
1043
1044         /*
1045          * If there is a discontinuity in the history, scale monotonic raw
1046          *      correction by:
1047          *      mult(real)/mult(raw) yielding the realtime correction
1048          * Otherwise, calculate the realtime correction similar to monotonic
1049          *      raw calculation
1050          */
1051         if (discontinuity) {
1052                 corr_real = mul_u64_u32_div
1053                         (corr_raw, tk->tkr_mono.mult, tk->tkr_raw.mult);
1054         } else {
1055                 corr_real = (u64)ktime_to_ns(
1056                         ktime_sub(ts->sys_realtime, history->real));
1057                 ret = scale64_check_overflow(partial_history_cycles,
1058                                              total_history_cycles, &corr_real);
1059                 if (ret)
1060                         return ret;
1061         }
1062
1063         /* Fixup monotonic raw and real time time values */
1064         if (interp_forward) {
1065                 ts->sys_monoraw = ktime_add_ns(history->raw, corr_raw);
1066                 ts->sys_realtime = ktime_add_ns(history->real, corr_real);
1067         } else {
1068                 ts->sys_monoraw = ktime_sub_ns(ts->sys_monoraw, corr_raw);
1069                 ts->sys_realtime = ktime_sub_ns(ts->sys_realtime, corr_real);
1070         }
1071
1072         return 0;
1073 }
1074
1075 /*
1076  * cycle_between - true if test occurs chronologically between before and after
1077  */
1078 static bool cycle_between(u64 before, u64 test, u64 after)
1079 {
1080         if (test > before && test < after)
1081                 return true;
1082         if (test < before && before > after)
1083                 return true;
1084         return false;
1085 }
1086
1087 /**
1088  * get_device_system_crosststamp - Synchronously capture system/device timestamp
1089  * @get_time_fn:        Callback to get simultaneous device time and
1090  *      system counter from the device driver
1091  * @ctx:                Context passed to get_time_fn()
1092  * @history_begin:      Historical reference point used to interpolate system
1093  *      time when counter provided by the driver is before the current interval
1094  * @xtstamp:            Receives simultaneously captured system and device time
1095  *
1096  * Reads a timestamp from a device and correlates it to system time
1097  */
1098 int get_device_system_crosststamp(int (*get_time_fn)
1099                                   (ktime_t *device_time,
1100                                    struct system_counterval_t *sys_counterval,
1101                                    void *ctx),
1102                                   void *ctx,
1103                                   struct system_time_snapshot *history_begin,
1104                                   struct system_device_crosststamp *xtstamp)
1105 {
1106         struct system_counterval_t system_counterval;
1107         struct timekeeper *tk = &tk_core.timekeeper;
1108         u64 cycles, now, interval_start;
1109         unsigned int clock_was_set_seq = 0;
1110         ktime_t base_real, base_raw;
1111         u64 nsec_real, nsec_raw;
1112         u8 cs_was_changed_seq;
1113         unsigned long seq;
1114         bool do_interp;
1115         int ret;
1116
1117         do {
1118                 seq = read_seqcount_begin(&tk_core.seq);
1119                 /*
1120                  * Try to synchronously capture device time and a system
1121                  * counter value calling back into the device driver
1122                  */
1123                 ret = get_time_fn(&xtstamp->device, &system_counterval, ctx);
1124                 if (ret)
1125                         return ret;
1126
1127                 /*
1128                  * Verify that the clocksource associated with the captured
1129                  * system counter value is the same as the currently installed
1130                  * timekeeper clocksource
1131                  */
1132                 if (tk->tkr_mono.clock != system_counterval.cs)
1133                         return -ENODEV;
1134                 cycles = system_counterval.cycles;
1135
1136                 /*
1137                  * Check whether the system counter value provided by the
1138                  * device driver is on the current timekeeping interval.
1139                  */
1140                 now = tk_clock_read(&tk->tkr_mono);
1141                 interval_start = tk->tkr_mono.cycle_last;
1142                 if (!cycle_between(interval_start, cycles, now)) {
1143                         clock_was_set_seq = tk->clock_was_set_seq;
1144                         cs_was_changed_seq = tk->cs_was_changed_seq;
1145                         cycles = interval_start;
1146                         do_interp = true;
1147                 } else {
1148                         do_interp = false;
1149                 }
1150
1151                 base_real = ktime_add(tk->tkr_mono.base,
1152                                       tk_core.timekeeper.offs_real);
1153                 base_raw = tk->tkr_raw.base;
1154
1155                 nsec_real = timekeeping_cycles_to_ns(&tk->tkr_mono,
1156                                                      system_counterval.cycles);
1157                 nsec_raw = timekeeping_cycles_to_ns(&tk->tkr_raw,
1158                                                     system_counterval.cycles);
1159         } while (read_seqcount_retry(&tk_core.seq, seq));
1160
1161         xtstamp->sys_realtime = ktime_add_ns(base_real, nsec_real);
1162         xtstamp->sys_monoraw = ktime_add_ns(base_raw, nsec_raw);
1163
1164         /*
1165          * Interpolate if necessary, adjusting back from the start of the
1166          * current interval
1167          */
1168         if (do_interp) {
1169                 u64 partial_history_cycles, total_history_cycles;
1170                 bool discontinuity;
1171
1172                 /*
1173                  * Check that the counter value occurs after the provided
1174                  * history reference and that the history doesn't cross a
1175                  * clocksource change
1176                  */
1177                 if (!history_begin ||
1178                     !cycle_between(history_begin->cycles,
1179                                    system_counterval.cycles, cycles) ||
1180                     history_begin->cs_was_changed_seq != cs_was_changed_seq)
1181                         return -EINVAL;
1182                 partial_history_cycles = cycles - system_counterval.cycles;
1183                 total_history_cycles = cycles - history_begin->cycles;
1184                 discontinuity =
1185                         history_begin->clock_was_set_seq != clock_was_set_seq;
1186
1187                 ret = adjust_historical_crosststamp(history_begin,
1188                                                     partial_history_cycles,
1189                                                     total_history_cycles,
1190                                                     discontinuity, xtstamp);
1191                 if (ret)
1192                         return ret;
1193         }
1194
1195         return 0;
1196 }
1197 EXPORT_SYMBOL_GPL(get_device_system_crosststamp);
1198
1199 /**
1200  * do_gettimeofday - Returns the time of day in a timeval
1201  * @tv:         pointer to the timeval to be set
1202  *
1203  * NOTE: Users should be converted to using getnstimeofday()
1204  */
1205 void do_gettimeofday(struct timeval *tv)
1206 {
1207         struct timespec64 now;
1208
1209         getnstimeofday64(&now);
1210         tv->tv_sec = now.tv_sec;
1211         tv->tv_usec = now.tv_nsec/1000;
1212 }
1213 EXPORT_SYMBOL(do_gettimeofday);
1214
1215 /**
1216  * do_settimeofday64 - Sets the time of day.
1217  * @ts:     pointer to the timespec64 variable containing the new time
1218  *
1219  * Sets the time of day to the new time and update NTP and notify hrtimers
1220  */
1221 int do_settimeofday64(const struct timespec64 *ts)
1222 {
1223         struct timekeeper *tk = &tk_core.timekeeper;
1224         struct timespec64 ts_delta, xt;
1225         unsigned long flags;
1226         int ret = 0;
1227
1228         if (!timespec64_valid_strict(ts))
1229                 return -EINVAL;
1230
1231         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1232         write_seqcount_begin(&tk_core.seq);
1233
1234         timekeeping_forward_now(tk);
1235
1236         xt = tk_xtime(tk);
1237         ts_delta.tv_sec = ts->tv_sec - xt.tv_sec;
1238         ts_delta.tv_nsec = ts->tv_nsec - xt.tv_nsec;
1239
1240         if (timespec64_compare(&tk->wall_to_monotonic, &ts_delta) > 0) {
1241                 ret = -EINVAL;
1242                 goto out;
1243         }
1244
1245         tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts_delta));
1246
1247         tk_set_xtime(tk, ts);
1248 out:
1249         timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1250
1251         write_seqcount_end(&tk_core.seq);
1252         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1253
1254         /* signal hrtimers about time change */
1255         clock_was_set();
1256
1257         return ret;
1258 }
1259 EXPORT_SYMBOL(do_settimeofday64);
1260
1261 /**
1262  * timekeeping_inject_offset - Adds or subtracts from the current time.
1263  * @tv:         pointer to the timespec variable containing the offset
1264  *
1265  * Adds or subtracts an offset value from the current time.
1266  */
1267 int timekeeping_inject_offset(struct timespec *ts)
1268 {
1269         struct timekeeper *tk = &tk_core.timekeeper;
1270         unsigned long flags;
1271         struct timespec64 ts64, tmp;
1272         int ret = 0;
1273
1274         if (!timespec_inject_offset_valid(ts))
1275                 return -EINVAL;
1276
1277         ts64 = timespec_to_timespec64(*ts);
1278
1279         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1280         write_seqcount_begin(&tk_core.seq);
1281
1282         timekeeping_forward_now(tk);
1283
1284         /* Make sure the proposed value is valid */
1285         tmp = timespec64_add(tk_xtime(tk),  ts64);
1286         if (timespec64_compare(&tk->wall_to_monotonic, &ts64) > 0 ||
1287             !timespec64_valid_strict(&tmp)) {
1288                 ret = -EINVAL;
1289                 goto error;
1290         }
1291
1292         tk_xtime_add(tk, &ts64);
1293         tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, ts64));
1294
1295 error: /* even if we error out, we forwarded the time, so call update */
1296         timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1297
1298         write_seqcount_end(&tk_core.seq);
1299         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1300
1301         /* signal hrtimers about time change */
1302         clock_was_set();
1303
1304         return ret;
1305 }
1306 EXPORT_SYMBOL(timekeeping_inject_offset);
1307
1308 /**
1309  * __timekeeping_set_tai_offset - Sets the TAI offset from UTC and monotonic
1310  *
1311  */
1312 static void __timekeeping_set_tai_offset(struct timekeeper *tk, s32 tai_offset)
1313 {
1314         tk->tai_offset = tai_offset;
1315         tk->offs_tai = ktime_add(tk->offs_real, ktime_set(tai_offset, 0));
1316 }
1317
1318 /**
1319  * change_clocksource - Swaps clocksources if a new one is available
1320  *
1321  * Accumulates current time interval and initializes new clocksource
1322  */
1323 static int change_clocksource(void *data)
1324 {
1325         struct timekeeper *tk = &tk_core.timekeeper;
1326         struct clocksource *new, *old;
1327         unsigned long flags;
1328
1329         new = (struct clocksource *) data;
1330
1331         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1332         write_seqcount_begin(&tk_core.seq);
1333
1334         timekeeping_forward_now(tk);
1335         /*
1336          * If the cs is in module, get a module reference. Succeeds
1337          * for built-in code (owner == NULL) as well.
1338          */
1339         if (try_module_get(new->owner)) {
1340                 if (!new->enable || new->enable(new) == 0) {
1341                         old = tk->tkr_mono.clock;
1342                         tk_setup_internals(tk, new);
1343                         if (old->disable)
1344                                 old->disable(old);
1345                         module_put(old->owner);
1346                 } else {
1347                         module_put(new->owner);
1348                 }
1349         }
1350         timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1351
1352         write_seqcount_end(&tk_core.seq);
1353         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1354
1355         return 0;
1356 }
1357
1358 /**
1359  * timekeeping_notify - Install a new clock source
1360  * @clock:              pointer to the clock source
1361  *
1362  * This function is called from clocksource.c after a new, better clock
1363  * source has been registered. The caller holds the clocksource_mutex.
1364  */
1365 int timekeeping_notify(struct clocksource *clock)
1366 {
1367         struct timekeeper *tk = &tk_core.timekeeper;
1368
1369         if (tk->tkr_mono.clock == clock)
1370                 return 0;
1371         stop_machine(change_clocksource, clock, NULL);
1372         tick_clock_notify();
1373         return tk->tkr_mono.clock == clock ? 0 : -1;
1374 }
1375
1376 /**
1377  * getrawmonotonic64 - Returns the raw monotonic time in a timespec
1378  * @ts:         pointer to the timespec64 to be set
1379  *
1380  * Returns the raw monotonic time (completely un-modified by ntp)
1381  */
1382 void getrawmonotonic64(struct timespec64 *ts)
1383 {
1384         struct timekeeper *tk = &tk_core.timekeeper;
1385         unsigned long seq;
1386         u64 nsecs;
1387
1388         do {
1389                 seq = read_seqcount_begin(&tk_core.seq);
1390                 ts->tv_sec = tk->raw_sec;
1391                 nsecs = timekeeping_get_ns(&tk->tkr_raw);
1392
1393         } while (read_seqcount_retry(&tk_core.seq, seq));
1394
1395         ts->tv_nsec = 0;
1396         timespec64_add_ns(ts, nsecs);
1397 }
1398 EXPORT_SYMBOL(getrawmonotonic64);
1399
1400
1401 /**
1402  * timekeeping_valid_for_hres - Check if timekeeping is suitable for hres
1403  */
1404 int timekeeping_valid_for_hres(void)
1405 {
1406         struct timekeeper *tk = &tk_core.timekeeper;
1407         unsigned long seq;
1408         int ret;
1409
1410         do {
1411                 seq = read_seqcount_begin(&tk_core.seq);
1412
1413                 ret = tk->tkr_mono.clock->flags & CLOCK_SOURCE_VALID_FOR_HRES;
1414
1415         } while (read_seqcount_retry(&tk_core.seq, seq));
1416
1417         return ret;
1418 }
1419
1420 /**
1421  * timekeeping_max_deferment - Returns max time the clocksource can be deferred
1422  */
1423 u64 timekeeping_max_deferment(void)
1424 {
1425         struct timekeeper *tk = &tk_core.timekeeper;
1426         unsigned long seq;
1427         u64 ret;
1428
1429         do {
1430                 seq = read_seqcount_begin(&tk_core.seq);
1431
1432                 ret = tk->tkr_mono.clock->max_idle_ns;
1433
1434         } while (read_seqcount_retry(&tk_core.seq, seq));
1435
1436         return ret;
1437 }
1438
1439 /**
1440  * read_persistent_clock -  Return time from the persistent clock.
1441  *
1442  * Weak dummy function for arches that do not yet support it.
1443  * Reads the time from the battery backed persistent clock.
1444  * Returns a timespec with tv_sec=0 and tv_nsec=0 if unsupported.
1445  *
1446  *  XXX - Do be sure to remove it once all arches implement it.
1447  */
1448 void __weak read_persistent_clock(struct timespec *ts)
1449 {
1450         ts->tv_sec = 0;
1451         ts->tv_nsec = 0;
1452 }
1453
1454 void __weak read_persistent_clock64(struct timespec64 *ts64)
1455 {
1456         struct timespec ts;
1457
1458         read_persistent_clock(&ts);
1459         *ts64 = timespec_to_timespec64(ts);
1460 }
1461
1462 /**
1463  * read_boot_clock64 -  Return time of the system start.
1464  *
1465  * Weak dummy function for arches that do not yet support it.
1466  * Function to read the exact time the system has been started.
1467  * Returns a timespec64 with tv_sec=0 and tv_nsec=0 if unsupported.
1468  *
1469  *  XXX - Do be sure to remove it once all arches implement it.
1470  */
1471 void __weak read_boot_clock64(struct timespec64 *ts)
1472 {
1473         ts->tv_sec = 0;
1474         ts->tv_nsec = 0;
1475 }
1476
1477 /* Flag for if timekeeping_resume() has injected sleeptime */
1478 static bool sleeptime_injected;
1479
1480 /* Flag for if there is a persistent clock on this platform */
1481 static bool persistent_clock_exists;
1482
1483 /*
1484  * timekeeping_init - Initializes the clocksource and common timekeeping values
1485  */
1486 void __init timekeeping_init(void)
1487 {
1488         struct timekeeper *tk = &tk_core.timekeeper;
1489         struct clocksource *clock;
1490         unsigned long flags;
1491         struct timespec64 now, boot, tmp;
1492
1493         read_persistent_clock64(&now);
1494         if (!timespec64_valid_strict(&now)) {
1495                 pr_warn("WARNING: Persistent clock returned invalid value!\n"
1496                         "         Check your CMOS/BIOS settings.\n");
1497                 now.tv_sec = 0;
1498                 now.tv_nsec = 0;
1499         } else if (now.tv_sec || now.tv_nsec)
1500                 persistent_clock_exists = true;
1501
1502         read_boot_clock64(&boot);
1503         if (!timespec64_valid_strict(&boot)) {
1504                 pr_warn("WARNING: Boot clock returned invalid value!\n"
1505                         "         Check your CMOS/BIOS settings.\n");
1506                 boot.tv_sec = 0;
1507                 boot.tv_nsec = 0;
1508         }
1509
1510         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1511         write_seqcount_begin(&tk_core.seq);
1512         ntp_init();
1513
1514         clock = clocksource_default_clock();
1515         if (clock->enable)
1516                 clock->enable(clock);
1517         tk_setup_internals(tk, clock);
1518
1519         tk_set_xtime(tk, &now);
1520         tk->raw_sec = 0;
1521         if (boot.tv_sec == 0 && boot.tv_nsec == 0)
1522                 boot = tk_xtime(tk);
1523
1524         set_normalized_timespec64(&tmp, -boot.tv_sec, -boot.tv_nsec);
1525         tk_set_wall_to_mono(tk, tmp);
1526
1527         timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1528
1529         write_seqcount_end(&tk_core.seq);
1530         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1531 }
1532
1533 /* time in seconds when suspend began for persistent clock */
1534 static struct timespec64 timekeeping_suspend_time;
1535
1536 /**
1537  * __timekeeping_inject_sleeptime - Internal function to add sleep interval
1538  * @delta: pointer to a timespec delta value
1539  *
1540  * Takes a timespec offset measuring a suspend interval and properly
1541  * adds the sleep offset to the timekeeping variables.
1542  */
1543 static void __timekeeping_inject_sleeptime(struct timekeeper *tk,
1544                                            struct timespec64 *delta)
1545 {
1546         if (!timespec64_valid_strict(delta)) {
1547                 printk_deferred(KERN_WARNING
1548                                 "__timekeeping_inject_sleeptime: Invalid "
1549                                 "sleep delta value!\n");
1550                 return;
1551         }
1552         tk_xtime_add(tk, delta);
1553         tk_set_wall_to_mono(tk, timespec64_sub(tk->wall_to_monotonic, *delta));
1554         tk_update_sleep_time(tk, timespec64_to_ktime(*delta));
1555         tk_debug_account_sleep_time(delta);
1556 }
1557
1558 #if defined(CONFIG_PM_SLEEP) && defined(CONFIG_RTC_HCTOSYS_DEVICE)
1559 /**
1560  * We have three kinds of time sources to use for sleep time
1561  * injection, the preference order is:
1562  * 1) non-stop clocksource
1563  * 2) persistent clock (ie: RTC accessible when irqs are off)
1564  * 3) RTC
1565  *
1566  * 1) and 2) are used by timekeeping, 3) by RTC subsystem.
1567  * If system has neither 1) nor 2), 3) will be used finally.
1568  *
1569  *
1570  * If timekeeping has injected sleeptime via either 1) or 2),
1571  * 3) becomes needless, so in this case we don't need to call
1572  * rtc_resume(), and this is what timekeeping_rtc_skipresume()
1573  * means.
1574  */
1575 bool timekeeping_rtc_skipresume(void)
1576 {
1577         return sleeptime_injected;
1578 }
1579
1580 /**
1581  * 1) can be determined whether to use or not only when doing
1582  * timekeeping_resume() which is invoked after rtc_suspend(),
1583  * so we can't skip rtc_suspend() surely if system has 1).
1584  *
1585  * But if system has 2), 2) will definitely be used, so in this
1586  * case we don't need to call rtc_suspend(), and this is what
1587  * timekeeping_rtc_skipsuspend() means.
1588  */
1589 bool timekeeping_rtc_skipsuspend(void)
1590 {
1591         return persistent_clock_exists;
1592 }
1593
1594 /**
1595  * timekeeping_inject_sleeptime64 - Adds suspend interval to timeekeeping values
1596  * @delta: pointer to a timespec64 delta value
1597  *
1598  * This hook is for architectures that cannot support read_persistent_clock64
1599  * because their RTC/persistent clock is only accessible when irqs are enabled.
1600  * and also don't have an effective nonstop clocksource.
1601  *
1602  * This function should only be called by rtc_resume(), and allows
1603  * a suspend offset to be injected into the timekeeping values.
1604  */
1605 void timekeeping_inject_sleeptime64(struct timespec64 *delta)
1606 {
1607         struct timekeeper *tk = &tk_core.timekeeper;
1608         unsigned long flags;
1609
1610         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1611         write_seqcount_begin(&tk_core.seq);
1612
1613         timekeeping_forward_now(tk);
1614
1615         __timekeeping_inject_sleeptime(tk, delta);
1616
1617         timekeeping_update(tk, TK_CLEAR_NTP | TK_MIRROR | TK_CLOCK_WAS_SET);
1618
1619         write_seqcount_end(&tk_core.seq);
1620         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1621
1622         /* signal hrtimers about time change */
1623         clock_was_set();
1624 }
1625 #endif
1626
1627 /**
1628  * timekeeping_resume - Resumes the generic timekeeping subsystem.
1629  */
1630 void timekeeping_resume(void)
1631 {
1632         struct timekeeper *tk = &tk_core.timekeeper;
1633         struct clocksource *clock = tk->tkr_mono.clock;
1634         unsigned long flags;
1635         struct timespec64 ts_new, ts_delta;
1636         u64 cycle_now;
1637
1638         sleeptime_injected = false;
1639         read_persistent_clock64(&ts_new);
1640
1641         clockevents_resume();
1642         clocksource_resume();
1643
1644         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1645         write_seqcount_begin(&tk_core.seq);
1646
1647         /*
1648          * After system resumes, we need to calculate the suspended time and
1649          * compensate it for the OS time. There are 3 sources that could be
1650          * used: Nonstop clocksource during suspend, persistent clock and rtc
1651          * device.
1652          *
1653          * One specific platform may have 1 or 2 or all of them, and the
1654          * preference will be:
1655          *      suspend-nonstop clocksource -> persistent clock -> rtc
1656          * The less preferred source will only be tried if there is no better
1657          * usable source. The rtc part is handled separately in rtc core code.
1658          */
1659         cycle_now = tk_clock_read(&tk->tkr_mono);
1660         if ((clock->flags & CLOCK_SOURCE_SUSPEND_NONSTOP) &&
1661                 cycle_now > tk->tkr_mono.cycle_last) {
1662                 u64 nsec, cyc_delta;
1663
1664                 cyc_delta = clocksource_delta(cycle_now, tk->tkr_mono.cycle_last,
1665                                               tk->tkr_mono.mask);
1666                 nsec = mul_u64_u32_shr(cyc_delta, clock->mult, clock->shift);
1667                 ts_delta = ns_to_timespec64(nsec);
1668                 sleeptime_injected = true;
1669         } else if (timespec64_compare(&ts_new, &timekeeping_suspend_time) > 0) {
1670                 ts_delta = timespec64_sub(ts_new, timekeeping_suspend_time);
1671                 sleeptime_injected = true;
1672         }
1673
1674         if (sleeptime_injected)
1675                 __timekeeping_inject_sleeptime(tk, &ts_delta);
1676
1677         /* Re-base the last cycle value */
1678         tk->tkr_mono.cycle_last = cycle_now;
1679         tk->tkr_raw.cycle_last  = cycle_now;
1680
1681         tk->ntp_error = 0;
1682         timekeeping_suspended = 0;
1683         timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
1684         write_seqcount_end(&tk_core.seq);
1685         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1686
1687         touch_softlockup_watchdog();
1688
1689         tick_resume();
1690         hrtimers_resume();
1691 }
1692
1693 int timekeeping_suspend(void)
1694 {
1695         struct timekeeper *tk = &tk_core.timekeeper;
1696         unsigned long flags;
1697         struct timespec64               delta, delta_delta;
1698         static struct timespec64        old_delta;
1699
1700         read_persistent_clock64(&timekeeping_suspend_time);
1701
1702         /*
1703          * On some systems the persistent_clock can not be detected at
1704          * timekeeping_init by its return value, so if we see a valid
1705          * value returned, update the persistent_clock_exists flag.
1706          */
1707         if (timekeeping_suspend_time.tv_sec || timekeeping_suspend_time.tv_nsec)
1708                 persistent_clock_exists = true;
1709
1710         raw_spin_lock_irqsave(&timekeeper_lock, flags);
1711         write_seqcount_begin(&tk_core.seq);
1712         timekeeping_forward_now(tk);
1713         timekeeping_suspended = 1;
1714
1715         if (persistent_clock_exists) {
1716                 /*
1717                  * To avoid drift caused by repeated suspend/resumes,
1718                  * which each can add ~1 second drift error,
1719                  * try to compensate so the difference in system time
1720                  * and persistent_clock time stays close to constant.
1721                  */
1722                 delta = timespec64_sub(tk_xtime(tk), timekeeping_suspend_time);
1723                 delta_delta = timespec64_sub(delta, old_delta);
1724                 if (abs(delta_delta.tv_sec) >= 2) {
1725                         /*
1726                          * if delta_delta is too large, assume time correction
1727                          * has occurred and set old_delta to the current delta.
1728                          */
1729                         old_delta = delta;
1730                 } else {
1731                         /* Otherwise try to adjust old_system to compensate */
1732                         timekeeping_suspend_time =
1733                                 timespec64_add(timekeeping_suspend_time, delta_delta);
1734                 }
1735         }
1736
1737         timekeeping_update(tk, TK_MIRROR);
1738         halt_fast_timekeeper(tk);
1739         write_seqcount_end(&tk_core.seq);
1740         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
1741
1742         tick_suspend();
1743         clocksource_suspend();
1744         clockevents_suspend();
1745
1746         return 0;
1747 }
1748
1749 /* sysfs resume/suspend bits for timekeeping */
1750 static struct syscore_ops timekeeping_syscore_ops = {
1751         .resume         = timekeeping_resume,
1752         .suspend        = timekeeping_suspend,
1753 };
1754
1755 static int __init timekeeping_init_ops(void)
1756 {
1757         register_syscore_ops(&timekeeping_syscore_ops);
1758         return 0;
1759 }
1760 device_initcall(timekeeping_init_ops);
1761
1762 /*
1763  * Apply a multiplier adjustment to the timekeeper
1764  */
1765 static __always_inline void timekeeping_apply_adjustment(struct timekeeper *tk,
1766                                                          s64 offset,
1767                                                          bool negative,
1768                                                          int adj_scale)
1769 {
1770         s64 interval = tk->cycle_interval;
1771         s32 mult_adj = 1;
1772
1773         if (negative) {
1774                 mult_adj = -mult_adj;
1775                 interval = -interval;
1776                 offset  = -offset;
1777         }
1778         mult_adj <<= adj_scale;
1779         interval <<= adj_scale;
1780         offset <<= adj_scale;
1781
1782         /*
1783          * So the following can be confusing.
1784          *
1785          * To keep things simple, lets assume mult_adj == 1 for now.
1786          *
1787          * When mult_adj != 1, remember that the interval and offset values
1788          * have been appropriately scaled so the math is the same.
1789          *
1790          * The basic idea here is that we're increasing the multiplier
1791          * by one, this causes the xtime_interval to be incremented by
1792          * one cycle_interval. This is because:
1793          *      xtime_interval = cycle_interval * mult
1794          * So if mult is being incremented by one:
1795          *      xtime_interval = cycle_interval * (mult + 1)
1796          * Its the same as:
1797          *      xtime_interval = (cycle_interval * mult) + cycle_interval
1798          * Which can be shortened to:
1799          *      xtime_interval += cycle_interval
1800          *
1801          * So offset stores the non-accumulated cycles. Thus the current
1802          * time (in shifted nanoseconds) is:
1803          *      now = (offset * adj) + xtime_nsec
1804          * Now, even though we're adjusting the clock frequency, we have
1805          * to keep time consistent. In other words, we can't jump back
1806          * in time, and we also want to avoid jumping forward in time.
1807          *
1808          * So given the same offset value, we need the time to be the same
1809          * both before and after the freq adjustment.
1810          *      now = (offset * adj_1) + xtime_nsec_1
1811          *      now = (offset * adj_2) + xtime_nsec_2
1812          * So:
1813          *      (offset * adj_1) + xtime_nsec_1 =
1814          *              (offset * adj_2) + xtime_nsec_2
1815          * And we know:
1816          *      adj_2 = adj_1 + 1
1817          * So:
1818          *      (offset * adj_1) + xtime_nsec_1 =
1819          *              (offset * (adj_1+1)) + xtime_nsec_2
1820          *      (offset * adj_1) + xtime_nsec_1 =
1821          *              (offset * adj_1) + offset + xtime_nsec_2
1822          * Canceling the sides:
1823          *      xtime_nsec_1 = offset + xtime_nsec_2
1824          * Which gives us:
1825          *      xtime_nsec_2 = xtime_nsec_1 - offset
1826          * Which simplfies to:
1827          *      xtime_nsec -= offset
1828          *
1829          * XXX - TODO: Doc ntp_error calculation.
1830          */
1831         if ((mult_adj > 0) && (tk->tkr_mono.mult + mult_adj < mult_adj)) {
1832                 /* NTP adjustment caused clocksource mult overflow */
1833                 WARN_ON_ONCE(1);
1834                 return;
1835         }
1836
1837         tk->tkr_mono.mult += mult_adj;
1838         tk->xtime_interval += interval;
1839         tk->tkr_mono.xtime_nsec -= offset;
1840         tk->ntp_error -= (interval - offset) << tk->ntp_error_shift;
1841 }
1842
1843 /*
1844  * Calculate the multiplier adjustment needed to match the frequency
1845  * specified by NTP
1846  */
1847 static __always_inline void timekeeping_freqadjust(struct timekeeper *tk,
1848                                                         s64 offset)
1849 {
1850         s64 interval = tk->cycle_interval;
1851         s64 xinterval = tk->xtime_interval;
1852         u32 base = tk->tkr_mono.clock->mult;
1853         u32 max = tk->tkr_mono.clock->maxadj;
1854         u32 cur_adj = tk->tkr_mono.mult;
1855         s64 tick_error;
1856         bool negative;
1857         u32 adj_scale;
1858
1859         /* Remove any current error adj from freq calculation */
1860         if (tk->ntp_err_mult)
1861                 xinterval -= tk->cycle_interval;
1862
1863         tk->ntp_tick = ntp_tick_length();
1864
1865         /* Calculate current error per tick */
1866         tick_error = ntp_tick_length() >> tk->ntp_error_shift;
1867         tick_error -= (xinterval + tk->xtime_remainder);
1868
1869         /* Don't worry about correcting it if its small */
1870         if (likely((tick_error >= 0) && (tick_error <= interval)))
1871                 return;
1872
1873         /* preserve the direction of correction */
1874         negative = (tick_error < 0);
1875
1876         /* If any adjustment would pass the max, just return */
1877         if (negative && (cur_adj - 1) <= (base - max))
1878                 return;
1879         if (!negative && (cur_adj + 1) >= (base + max))
1880                 return;
1881         /*
1882          * Sort out the magnitude of the correction, but
1883          * avoid making so large a correction that we go
1884          * over the max adjustment.
1885          */
1886         adj_scale = 0;
1887         tick_error = abs(tick_error);
1888         while (tick_error > interval) {
1889                 u32 adj = 1 << (adj_scale + 1);
1890
1891                 /* Check if adjustment gets us within 1 unit from the max */
1892                 if (negative && (cur_adj - adj) <= (base - max))
1893                         break;
1894                 if (!negative && (cur_adj + adj) >= (base + max))
1895                         break;
1896
1897                 adj_scale++;
1898                 tick_error >>= 1;
1899         }
1900
1901         /* scale the corrections */
1902         timekeeping_apply_adjustment(tk, offset, negative, adj_scale);
1903 }
1904
1905 /*
1906  * Adjust the timekeeper's multiplier to the correct frequency
1907  * and also to reduce the accumulated error value.
1908  */
1909 static void timekeeping_adjust(struct timekeeper *tk, s64 offset)
1910 {
1911         /* Correct for the current frequency error */
1912         timekeeping_freqadjust(tk, offset);
1913
1914         /* Next make a small adjustment to fix any cumulative error */
1915         if (!tk->ntp_err_mult && (tk->ntp_error > 0)) {
1916                 tk->ntp_err_mult = 1;
1917                 timekeeping_apply_adjustment(tk, offset, 0, 0);
1918         } else if (tk->ntp_err_mult && (tk->ntp_error <= 0)) {
1919                 /* Undo any existing error adjustment */
1920                 timekeeping_apply_adjustment(tk, offset, 1, 0);
1921                 tk->ntp_err_mult = 0;
1922         }
1923
1924         if (unlikely(tk->tkr_mono.clock->maxadj &&
1925                 (abs(tk->tkr_mono.mult - tk->tkr_mono.clock->mult)
1926                         > tk->tkr_mono.clock->maxadj))) {
1927                 printk_once(KERN_WARNING
1928                         "Adjusting %s more than 11%% (%ld vs %ld)\n",
1929                         tk->tkr_mono.clock->name, (long)tk->tkr_mono.mult,
1930                         (long)tk->tkr_mono.clock->mult + tk->tkr_mono.clock->maxadj);
1931         }
1932
1933         /*
1934          * It may be possible that when we entered this function, xtime_nsec
1935          * was very small.  Further, if we're slightly speeding the clocksource
1936          * in the code above, its possible the required corrective factor to
1937          * xtime_nsec could cause it to underflow.
1938          *
1939          * Now, since we already accumulated the second, cannot simply roll
1940          * the accumulated second back, since the NTP subsystem has been
1941          * notified via second_overflow. So instead we push xtime_nsec forward
1942          * by the amount we underflowed, and add that amount into the error.
1943          *
1944          * We'll correct this error next time through this function, when
1945          * xtime_nsec is not as small.
1946          */
1947         if (unlikely((s64)tk->tkr_mono.xtime_nsec < 0)) {
1948                 s64 neg = -(s64)tk->tkr_mono.xtime_nsec;
1949                 tk->tkr_mono.xtime_nsec = 0;
1950                 tk->ntp_error += neg << tk->ntp_error_shift;
1951         }
1952 }
1953
1954 /**
1955  * accumulate_nsecs_to_secs - Accumulates nsecs into secs
1956  *
1957  * Helper function that accumulates the nsecs greater than a second
1958  * from the xtime_nsec field to the xtime_secs field.
1959  * It also calls into the NTP code to handle leapsecond processing.
1960  *
1961  */
1962 static inline unsigned int accumulate_nsecs_to_secs(struct timekeeper *tk)
1963 {
1964         u64 nsecps = (u64)NSEC_PER_SEC << tk->tkr_mono.shift;
1965         unsigned int clock_set = 0;
1966
1967         while (tk->tkr_mono.xtime_nsec >= nsecps) {
1968                 int leap;
1969
1970                 tk->tkr_mono.xtime_nsec -= nsecps;
1971                 tk->xtime_sec++;
1972
1973                 /* Figure out if its a leap sec and apply if needed */
1974                 leap = second_overflow(tk->xtime_sec);
1975                 if (unlikely(leap)) {
1976                         struct timespec64 ts;
1977
1978                         tk->xtime_sec += leap;
1979
1980                         ts.tv_sec = leap;
1981                         ts.tv_nsec = 0;
1982                         tk_set_wall_to_mono(tk,
1983                                 timespec64_sub(tk->wall_to_monotonic, ts));
1984
1985                         __timekeeping_set_tai_offset(tk, tk->tai_offset - leap);
1986
1987                         clock_set = TK_CLOCK_WAS_SET;
1988                 }
1989         }
1990         return clock_set;
1991 }
1992
1993 /**
1994  * logarithmic_accumulation - shifted accumulation of cycles
1995  *
1996  * This functions accumulates a shifted interval of cycles into
1997  * into a shifted interval nanoseconds. Allows for O(log) accumulation
1998  * loop.
1999  *
2000  * Returns the unconsumed cycles.
2001  */
2002 static u64 logarithmic_accumulation(struct timekeeper *tk, u64 offset,
2003                                     u32 shift, unsigned int *clock_set)
2004 {
2005         u64 interval = tk->cycle_interval << shift;
2006         u64 snsec_per_sec;
2007
2008         /* If the offset is smaller than a shifted interval, do nothing */
2009         if (offset < interval)
2010                 return offset;
2011
2012         /* Accumulate one shifted interval */
2013         offset -= interval;
2014         tk->tkr_mono.cycle_last += interval;
2015         tk->tkr_raw.cycle_last  += interval;
2016
2017         tk->tkr_mono.xtime_nsec += tk->xtime_interval << shift;
2018         *clock_set |= accumulate_nsecs_to_secs(tk);
2019
2020         /* Accumulate raw time */
2021         tk->tkr_raw.xtime_nsec += tk->raw_interval << shift;
2022         snsec_per_sec = (u64)NSEC_PER_SEC << tk->tkr_raw.shift;
2023         while (tk->tkr_raw.xtime_nsec >= snsec_per_sec) {
2024                 tk->tkr_raw.xtime_nsec -= snsec_per_sec;
2025                 tk->raw_sec++;
2026         }
2027
2028         /* Accumulate error between NTP and clock interval */
2029         tk->ntp_error += tk->ntp_tick << shift;
2030         tk->ntp_error -= (tk->xtime_interval + tk->xtime_remainder) <<
2031                                                 (tk->ntp_error_shift + shift);
2032
2033         return offset;
2034 }
2035
2036 /**
2037  * update_wall_time - Uses the current clocksource to increment the wall time
2038  *
2039  */
2040 void update_wall_time(void)
2041 {
2042         struct timekeeper *real_tk = &tk_core.timekeeper;
2043         struct timekeeper *tk = &shadow_timekeeper;
2044         u64 offset;
2045         int shift = 0, maxshift;
2046         unsigned int clock_set = 0;
2047         unsigned long flags;
2048
2049         raw_spin_lock_irqsave(&timekeeper_lock, flags);
2050
2051         /* Make sure we're fully resumed: */
2052         if (unlikely(timekeeping_suspended))
2053                 goto out;
2054
2055 #ifdef CONFIG_ARCH_USES_GETTIMEOFFSET
2056         offset = real_tk->cycle_interval;
2057 #else
2058         offset = clocksource_delta(tk_clock_read(&tk->tkr_mono),
2059                                    tk->tkr_mono.cycle_last, tk->tkr_mono.mask);
2060 #endif
2061
2062         /* Check if there's really nothing to do */
2063         if (offset < real_tk->cycle_interval)
2064                 goto out;
2065
2066         /* Do some additional sanity checking */
2067         timekeeping_check_update(tk, offset);
2068
2069         /*
2070          * With NO_HZ we may have to accumulate many cycle_intervals
2071          * (think "ticks") worth of time at once. To do this efficiently,
2072          * we calculate the largest doubling multiple of cycle_intervals
2073          * that is smaller than the offset.  We then accumulate that
2074          * chunk in one go, and then try to consume the next smaller
2075          * doubled multiple.
2076          */
2077         shift = ilog2(offset) - ilog2(tk->cycle_interval);
2078         shift = max(0, shift);
2079         /* Bound shift to one less than what overflows tick_length */
2080         maxshift = (64 - (ilog2(ntp_tick_length())+1)) - 1;
2081         shift = min(shift, maxshift);
2082         while (offset >= tk->cycle_interval) {
2083                 offset = logarithmic_accumulation(tk, offset, shift,
2084                                                         &clock_set);
2085                 if (offset < tk->cycle_interval<<shift)
2086                         shift--;
2087         }
2088
2089         /* correct the clock when NTP error is too big */
2090         timekeeping_adjust(tk, offset);
2091
2092         /*
2093          * XXX This can be killed once everyone converts
2094          * to the new update_vsyscall.
2095          */
2096         old_vsyscall_fixup(tk);
2097
2098         /*
2099          * Finally, make sure that after the rounding
2100          * xtime_nsec isn't larger than NSEC_PER_SEC
2101          */
2102         clock_set |= accumulate_nsecs_to_secs(tk);
2103
2104         write_seqcount_begin(&tk_core.seq);
2105         /*
2106          * Update the real timekeeper.
2107          *
2108          * We could avoid this memcpy by switching pointers, but that
2109          * requires changes to all other timekeeper usage sites as
2110          * well, i.e. move the timekeeper pointer getter into the
2111          * spinlocked/seqcount protected sections. And we trade this
2112          * memcpy under the tk_core.seq against one before we start
2113          * updating.
2114          */
2115         timekeeping_update(tk, clock_set);
2116         memcpy(real_tk, tk, sizeof(*tk));
2117         /* The memcpy must come last. Do not put anything here! */
2118         write_seqcount_end(&tk_core.seq);
2119 out:
2120         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2121         if (clock_set)
2122                 /* Have to call _delayed version, since in irq context*/
2123                 clock_was_set_delayed();
2124 }
2125
2126 /**
2127  * getboottime64 - Return the real time of system boot.
2128  * @ts:         pointer to the timespec64 to be set
2129  *
2130  * Returns the wall-time of boot in a timespec64.
2131  *
2132  * This is based on the wall_to_monotonic offset and the total suspend
2133  * time. Calls to settimeofday will affect the value returned (which
2134  * basically means that however wrong your real time clock is at boot time,
2135  * you get the right time here).
2136  */
2137 void getboottime64(struct timespec64 *ts)
2138 {
2139         struct timekeeper *tk = &tk_core.timekeeper;
2140         ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
2141
2142         *ts = ktime_to_timespec64(t);
2143 }
2144 EXPORT_SYMBOL_GPL(getboottime64);
2145
2146 unsigned long get_seconds(void)
2147 {
2148         struct timekeeper *tk = &tk_core.timekeeper;
2149
2150         return tk->xtime_sec;
2151 }
2152 EXPORT_SYMBOL(get_seconds);
2153
2154 struct timespec __current_kernel_time(void)
2155 {
2156         struct timekeeper *tk = &tk_core.timekeeper;
2157
2158         return timespec64_to_timespec(tk_xtime(tk));
2159 }
2160
2161 struct timespec64 current_kernel_time64(void)
2162 {
2163         struct timekeeper *tk = &tk_core.timekeeper;
2164         struct timespec64 now;
2165         unsigned long seq;
2166
2167         do {
2168                 seq = read_seqcount_begin(&tk_core.seq);
2169
2170                 now = tk_xtime(tk);
2171         } while (read_seqcount_retry(&tk_core.seq, seq));
2172
2173         return now;
2174 }
2175 EXPORT_SYMBOL(current_kernel_time64);
2176
2177 struct timespec64 get_monotonic_coarse64(void)
2178 {
2179         struct timekeeper *tk = &tk_core.timekeeper;
2180         struct timespec64 now, mono;
2181         unsigned long seq;
2182
2183         do {
2184                 seq = read_seqcount_begin(&tk_core.seq);
2185
2186                 now = tk_xtime(tk);
2187                 mono = tk->wall_to_monotonic;
2188         } while (read_seqcount_retry(&tk_core.seq, seq));
2189
2190         set_normalized_timespec64(&now, now.tv_sec + mono.tv_sec,
2191                                 now.tv_nsec + mono.tv_nsec);
2192
2193         return now;
2194 }
2195 EXPORT_SYMBOL(get_monotonic_coarse64);
2196
2197 /*
2198  * Must hold jiffies_lock
2199  */
2200 void do_timer(unsigned long ticks)
2201 {
2202         jiffies_64 += ticks;
2203         calc_global_load(ticks);
2204 }
2205
2206 /**
2207  * ktime_get_update_offsets_now - hrtimer helper
2208  * @cwsseq:     pointer to check and store the clock was set sequence number
2209  * @offs_real:  pointer to storage for monotonic -> realtime offset
2210  * @offs_boot:  pointer to storage for monotonic -> boottime offset
2211  * @offs_tai:   pointer to storage for monotonic -> clock tai offset
2212  *
2213  * Returns current monotonic time and updates the offsets if the
2214  * sequence number in @cwsseq and timekeeper.clock_was_set_seq are
2215  * different.
2216  *
2217  * Called from hrtimer_interrupt() or retrigger_next_event()
2218  */
2219 ktime_t ktime_get_update_offsets_now(unsigned int *cwsseq, ktime_t *offs_real,
2220                                      ktime_t *offs_boot, ktime_t *offs_tai)
2221 {
2222         struct timekeeper *tk = &tk_core.timekeeper;
2223         unsigned int seq;
2224         ktime_t base;
2225         u64 nsecs;
2226
2227         do {
2228                 seq = read_seqcount_begin(&tk_core.seq);
2229
2230                 base = tk->tkr_mono.base;
2231                 nsecs = timekeeping_get_ns(&tk->tkr_mono);
2232                 base = ktime_add_ns(base, nsecs);
2233
2234                 if (*cwsseq != tk->clock_was_set_seq) {
2235                         *cwsseq = tk->clock_was_set_seq;
2236                         *offs_real = tk->offs_real;
2237                         *offs_boot = tk->offs_boot;
2238                         *offs_tai = tk->offs_tai;
2239                 }
2240
2241                 /* Handle leapsecond insertion adjustments */
2242                 if (unlikely(base >= tk->next_leap_ktime))
2243                         *offs_real = ktime_sub(tk->offs_real, ktime_set(1, 0));
2244
2245         } while (read_seqcount_retry(&tk_core.seq, seq));
2246
2247         return base;
2248 }
2249
2250 /**
2251  * do_adjtimex() - Accessor function to NTP __do_adjtimex function
2252  */
2253 int do_adjtimex(struct timex *txc)
2254 {
2255         struct timekeeper *tk = &tk_core.timekeeper;
2256         unsigned long flags;
2257         struct timespec64 ts;
2258         s32 orig_tai, tai;
2259         int ret;
2260
2261         /* Validate the data before disabling interrupts */
2262         ret = ntp_validate_timex(txc);
2263         if (ret)
2264                 return ret;
2265
2266         if (txc->modes & ADJ_SETOFFSET) {
2267                 struct timespec delta;
2268                 delta.tv_sec  = txc->time.tv_sec;
2269                 delta.tv_nsec = txc->time.tv_usec;
2270                 if (!(txc->modes & ADJ_NANO))
2271                         delta.tv_nsec *= 1000;
2272                 ret = timekeeping_inject_offset(&delta);
2273                 if (ret)
2274                         return ret;
2275         }
2276
2277         getnstimeofday64(&ts);
2278
2279         raw_spin_lock_irqsave(&timekeeper_lock, flags);
2280         write_seqcount_begin(&tk_core.seq);
2281
2282         orig_tai = tai = tk->tai_offset;
2283         ret = __do_adjtimex(txc, &ts, &tai);
2284
2285         if (tai != orig_tai) {
2286                 __timekeeping_set_tai_offset(tk, tai);
2287                 timekeeping_update(tk, TK_MIRROR | TK_CLOCK_WAS_SET);
2288         }
2289         tk_update_leap_state(tk);
2290
2291         write_seqcount_end(&tk_core.seq);
2292         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2293
2294         if (tai != orig_tai)
2295                 clock_was_set();
2296
2297         ntp_notify_cmos_timer();
2298
2299         return ret;
2300 }
2301
2302 #ifdef CONFIG_NTP_PPS
2303 /**
2304  * hardpps() - Accessor function to NTP __hardpps function
2305  */
2306 void hardpps(const struct timespec64 *phase_ts, const struct timespec64 *raw_ts)
2307 {
2308         unsigned long flags;
2309
2310         raw_spin_lock_irqsave(&timekeeper_lock, flags);
2311         write_seqcount_begin(&tk_core.seq);
2312
2313         __hardpps(phase_ts, raw_ts);
2314
2315         write_seqcount_end(&tk_core.seq);
2316         raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
2317 }
2318 EXPORT_SYMBOL(hardpps);
2319 #endif
2320
2321 /**
2322  * xtime_update() - advances the timekeeping infrastructure
2323  * @ticks:      number of ticks, that have elapsed since the last call.
2324  *
2325  * Must be called with interrupts disabled.
2326  */
2327 void xtime_update(unsigned long ticks)
2328 {
2329         write_seqlock(&jiffies_lock);
2330         do_timer(ticks);
2331         write_sequnlock(&jiffies_lock);
2332         update_wall_time();
2333 }