Loading...
1/*
2 * linux/kernel/time/clocksource.c
3 *
4 * This file contains the functions which manage clocksource drivers.
5 *
6 * Copyright (C) 2004, 2005 IBM, John Stultz (johnstul@us.ibm.com)
7 *
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation; either version 2 of the License, or
11 * (at your option) any later version.
12 *
13 * This program is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
17 *
18 * You should have received a copy of the GNU General Public License
19 * along with this program; if not, write to the Free Software
20 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21 *
22 * TODO WishList:
23 * o Allow clocksource drivers to be unregistered
24 */
25
26#include <linux/device.h>
27#include <linux/clocksource.h>
28#include <linux/init.h>
29#include <linux/module.h>
30#include <linux/sched.h> /* for spin_unlock_irq() using preempt_count() m68k */
31#include <linux/tick.h>
32#include <linux/kthread.h>
33
34#include "tick-internal.h"
35
36void timecounter_init(struct timecounter *tc,
37 const struct cyclecounter *cc,
38 u64 start_tstamp)
39{
40 tc->cc = cc;
41 tc->cycle_last = cc->read(cc);
42 tc->nsec = start_tstamp;
43}
44EXPORT_SYMBOL_GPL(timecounter_init);
45
46/**
47 * timecounter_read_delta - get nanoseconds since last call of this function
48 * @tc: Pointer to time counter
49 *
50 * When the underlying cycle counter runs over, this will be handled
51 * correctly as long as it does not run over more than once between
52 * calls.
53 *
54 * The first call to this function for a new time counter initializes
55 * the time tracking and returns an undefined result.
56 */
57static u64 timecounter_read_delta(struct timecounter *tc)
58{
59 cycle_t cycle_now, cycle_delta;
60 u64 ns_offset;
61
62 /* read cycle counter: */
63 cycle_now = tc->cc->read(tc->cc);
64
65 /* calculate the delta since the last timecounter_read_delta(): */
66 cycle_delta = (cycle_now - tc->cycle_last) & tc->cc->mask;
67
68 /* convert to nanoseconds: */
69 ns_offset = cyclecounter_cyc2ns(tc->cc, cycle_delta);
70
71 /* update time stamp of timecounter_read_delta() call: */
72 tc->cycle_last = cycle_now;
73
74 return ns_offset;
75}
76
77u64 timecounter_read(struct timecounter *tc)
78{
79 u64 nsec;
80
81 /* increment time by nanoseconds since last call */
82 nsec = timecounter_read_delta(tc);
83 nsec += tc->nsec;
84 tc->nsec = nsec;
85
86 return nsec;
87}
88EXPORT_SYMBOL_GPL(timecounter_read);
89
90u64 timecounter_cyc2time(struct timecounter *tc,
91 cycle_t cycle_tstamp)
92{
93 u64 cycle_delta = (cycle_tstamp - tc->cycle_last) & tc->cc->mask;
94 u64 nsec;
95
96 /*
97 * Instead of always treating cycle_tstamp as more recent
98 * than tc->cycle_last, detect when it is too far in the
99 * future and treat it as old time stamp instead.
100 */
101 if (cycle_delta > tc->cc->mask / 2) {
102 cycle_delta = (tc->cycle_last - cycle_tstamp) & tc->cc->mask;
103 nsec = tc->nsec - cyclecounter_cyc2ns(tc->cc, cycle_delta);
104 } else {
105 nsec = cyclecounter_cyc2ns(tc->cc, cycle_delta) + tc->nsec;
106 }
107
108 return nsec;
109}
110EXPORT_SYMBOL_GPL(timecounter_cyc2time);
111
112/**
113 * clocks_calc_mult_shift - calculate mult/shift factors for scaled math of clocks
114 * @mult: pointer to mult variable
115 * @shift: pointer to shift variable
116 * @from: frequency to convert from
117 * @to: frequency to convert to
118 * @maxsec: guaranteed runtime conversion range in seconds
119 *
120 * The function evaluates the shift/mult pair for the scaled math
121 * operations of clocksources and clockevents.
122 *
123 * @to and @from are frequency values in HZ. For clock sources @to is
124 * NSEC_PER_SEC == 1GHz and @from is the counter frequency. For clock
125 * event @to is the counter frequency and @from is NSEC_PER_SEC.
126 *
127 * The @maxsec conversion range argument controls the time frame in
128 * seconds which must be covered by the runtime conversion with the
129 * calculated mult and shift factors. This guarantees that no 64bit
130 * overflow happens when the input value of the conversion is
131 * multiplied with the calculated mult factor. Larger ranges may
132 * reduce the conversion accuracy by chosing smaller mult and shift
133 * factors.
134 */
135void
136clocks_calc_mult_shift(u32 *mult, u32 *shift, u32 from, u32 to, u32 maxsec)
137{
138 u64 tmp;
139 u32 sft, sftacc= 32;
140
141 /*
142 * Calculate the shift factor which is limiting the conversion
143 * range:
144 */
145 tmp = ((u64)maxsec * from) >> 32;
146 while (tmp) {
147 tmp >>=1;
148 sftacc--;
149 }
150
151 /*
152 * Find the conversion shift/mult pair which has the best
153 * accuracy and fits the maxsec conversion range:
154 */
155 for (sft = 32; sft > 0; sft--) {
156 tmp = (u64) to << sft;
157 tmp += from / 2;
158 do_div(tmp, from);
159 if ((tmp >> sftacc) == 0)
160 break;
161 }
162 *mult = tmp;
163 *shift = sft;
164}
165
166/*[Clocksource internal variables]---------
167 * curr_clocksource:
168 * currently selected clocksource.
169 * clocksource_list:
170 * linked list with the registered clocksources
171 * clocksource_mutex:
172 * protects manipulations to curr_clocksource and the clocksource_list
173 * override_name:
174 * Name of the user-specified clocksource.
175 */
176static struct clocksource *curr_clocksource;
177static LIST_HEAD(clocksource_list);
178static DEFINE_MUTEX(clocksource_mutex);
179static char override_name[CS_NAME_LEN];
180static int finished_booting;
181
182#ifdef CONFIG_CLOCKSOURCE_WATCHDOG
183static void clocksource_watchdog_work(struct work_struct *work);
184static void clocksource_select(void);
185
186static LIST_HEAD(watchdog_list);
187static struct clocksource *watchdog;
188static struct timer_list watchdog_timer;
189static DECLARE_WORK(watchdog_work, clocksource_watchdog_work);
190static DEFINE_SPINLOCK(watchdog_lock);
191static int watchdog_running;
192static atomic_t watchdog_reset_pending;
193
194static int clocksource_watchdog_kthread(void *data);
195static void __clocksource_change_rating(struct clocksource *cs, int rating);
196
197/*
198 * Interval: 0.5sec Threshold: 0.0625s
199 */
200#define WATCHDOG_INTERVAL (HZ >> 1)
201#define WATCHDOG_THRESHOLD (NSEC_PER_SEC >> 4)
202
203static void clocksource_watchdog_work(struct work_struct *work)
204{
205 /*
206 * If kthread_run fails the next watchdog scan over the
207 * watchdog_list will find the unstable clock again.
208 */
209 kthread_run(clocksource_watchdog_kthread, NULL, "kwatchdog");
210}
211
212static void __clocksource_unstable(struct clocksource *cs)
213{
214 cs->flags &= ~(CLOCK_SOURCE_VALID_FOR_HRES | CLOCK_SOURCE_WATCHDOG);
215 cs->flags |= CLOCK_SOURCE_UNSTABLE;
216 if (finished_booting)
217 schedule_work(&watchdog_work);
218}
219
220static void clocksource_unstable(struct clocksource *cs, int64_t delta)
221{
222 printk(KERN_WARNING "Clocksource %s unstable (delta = %Ld ns)\n",
223 cs->name, delta);
224 __clocksource_unstable(cs);
225}
226
227/**
228 * clocksource_mark_unstable - mark clocksource unstable via watchdog
229 * @cs: clocksource to be marked unstable
230 *
231 * This function is called instead of clocksource_change_rating from
232 * cpu hotplug code to avoid a deadlock between the clocksource mutex
233 * and the cpu hotplug mutex. It defers the update of the clocksource
234 * to the watchdog thread.
235 */
236void clocksource_mark_unstable(struct clocksource *cs)
237{
238 unsigned long flags;
239
240 spin_lock_irqsave(&watchdog_lock, flags);
241 if (!(cs->flags & CLOCK_SOURCE_UNSTABLE)) {
242 if (list_empty(&cs->wd_list))
243 list_add(&cs->wd_list, &watchdog_list);
244 __clocksource_unstable(cs);
245 }
246 spin_unlock_irqrestore(&watchdog_lock, flags);
247}
248
249static void clocksource_watchdog(unsigned long data)
250{
251 struct clocksource *cs;
252 cycle_t csnow, wdnow;
253 int64_t wd_nsec, cs_nsec;
254 int next_cpu, reset_pending;
255
256 spin_lock(&watchdog_lock);
257 if (!watchdog_running)
258 goto out;
259
260 reset_pending = atomic_read(&watchdog_reset_pending);
261
262 list_for_each_entry(cs, &watchdog_list, wd_list) {
263
264 /* Clocksource already marked unstable? */
265 if (cs->flags & CLOCK_SOURCE_UNSTABLE) {
266 if (finished_booting)
267 schedule_work(&watchdog_work);
268 continue;
269 }
270
271 local_irq_disable();
272 csnow = cs->read(cs);
273 wdnow = watchdog->read(watchdog);
274 local_irq_enable();
275
276 /* Clocksource initialized ? */
277 if (!(cs->flags & CLOCK_SOURCE_WATCHDOG) ||
278 atomic_read(&watchdog_reset_pending)) {
279 cs->flags |= CLOCK_SOURCE_WATCHDOG;
280 cs->wd_last = wdnow;
281 cs->cs_last = csnow;
282 continue;
283 }
284
285 wd_nsec = clocksource_cyc2ns((wdnow - cs->wd_last) & watchdog->mask,
286 watchdog->mult, watchdog->shift);
287
288 cs_nsec = clocksource_cyc2ns((csnow - cs->cs_last) &
289 cs->mask, cs->mult, cs->shift);
290 cs->cs_last = csnow;
291 cs->wd_last = wdnow;
292
293 if (atomic_read(&watchdog_reset_pending))
294 continue;
295
296 /* Check the deviation from the watchdog clocksource. */
297 if ((abs(cs_nsec - wd_nsec) > WATCHDOG_THRESHOLD)) {
298 clocksource_unstable(cs, cs_nsec - wd_nsec);
299 continue;
300 }
301
302 if (!(cs->flags & CLOCK_SOURCE_VALID_FOR_HRES) &&
303 (cs->flags & CLOCK_SOURCE_IS_CONTINUOUS) &&
304 (watchdog->flags & CLOCK_SOURCE_IS_CONTINUOUS)) {
305 /* Mark it valid for high-res. */
306 cs->flags |= CLOCK_SOURCE_VALID_FOR_HRES;
307
308 /*
309 * clocksource_done_booting() will sort it if
310 * finished_booting is not set yet.
311 */
312 if (!finished_booting)
313 continue;
314
315 /*
316 * If this is not the current clocksource let
317 * the watchdog thread reselect it. Due to the
318 * change to high res this clocksource might
319 * be preferred now. If it is the current
320 * clocksource let the tick code know about
321 * that change.
322 */
323 if (cs != curr_clocksource) {
324 cs->flags |= CLOCK_SOURCE_RESELECT;
325 schedule_work(&watchdog_work);
326 } else {
327 tick_clock_notify();
328 }
329 }
330 }
331
332 /*
333 * We only clear the watchdog_reset_pending, when we did a
334 * full cycle through all clocksources.
335 */
336 if (reset_pending)
337 atomic_dec(&watchdog_reset_pending);
338
339 /*
340 * Cycle through CPUs to check if the CPUs stay synchronized
341 * to each other.
342 */
343 next_cpu = cpumask_next(raw_smp_processor_id(), cpu_online_mask);
344 if (next_cpu >= nr_cpu_ids)
345 next_cpu = cpumask_first(cpu_online_mask);
346 watchdog_timer.expires += WATCHDOG_INTERVAL;
347 add_timer_on(&watchdog_timer, next_cpu);
348out:
349 spin_unlock(&watchdog_lock);
350}
351
352static inline void clocksource_start_watchdog(void)
353{
354 if (watchdog_running || !watchdog || list_empty(&watchdog_list))
355 return;
356 init_timer(&watchdog_timer);
357 watchdog_timer.function = clocksource_watchdog;
358 watchdog_timer.expires = jiffies + WATCHDOG_INTERVAL;
359 add_timer_on(&watchdog_timer, cpumask_first(cpu_online_mask));
360 watchdog_running = 1;
361}
362
363static inline void clocksource_stop_watchdog(void)
364{
365 if (!watchdog_running || (watchdog && !list_empty(&watchdog_list)))
366 return;
367 del_timer(&watchdog_timer);
368 watchdog_running = 0;
369}
370
371static inline void clocksource_reset_watchdog(void)
372{
373 struct clocksource *cs;
374
375 list_for_each_entry(cs, &watchdog_list, wd_list)
376 cs->flags &= ~CLOCK_SOURCE_WATCHDOG;
377}
378
379static void clocksource_resume_watchdog(void)
380{
381 atomic_inc(&watchdog_reset_pending);
382}
383
384static void clocksource_enqueue_watchdog(struct clocksource *cs)
385{
386 unsigned long flags;
387
388 spin_lock_irqsave(&watchdog_lock, flags);
389 if (cs->flags & CLOCK_SOURCE_MUST_VERIFY) {
390 /* cs is a clocksource to be watched. */
391 list_add(&cs->wd_list, &watchdog_list);
392 cs->flags &= ~CLOCK_SOURCE_WATCHDOG;
393 } else {
394 /* cs is a watchdog. */
395 if (cs->flags & CLOCK_SOURCE_IS_CONTINUOUS)
396 cs->flags |= CLOCK_SOURCE_VALID_FOR_HRES;
397 /* Pick the best watchdog. */
398 if (!watchdog || cs->rating > watchdog->rating) {
399 watchdog = cs;
400 /* Reset watchdog cycles */
401 clocksource_reset_watchdog();
402 }
403 }
404 /* Check if the watchdog timer needs to be started. */
405 clocksource_start_watchdog();
406 spin_unlock_irqrestore(&watchdog_lock, flags);
407}
408
409static void clocksource_dequeue_watchdog(struct clocksource *cs)
410{
411 unsigned long flags;
412
413 spin_lock_irqsave(&watchdog_lock, flags);
414 if (cs != watchdog) {
415 if (cs->flags & CLOCK_SOURCE_MUST_VERIFY) {
416 /* cs is a watched clocksource. */
417 list_del_init(&cs->wd_list);
418 /* Check if the watchdog timer needs to be stopped. */
419 clocksource_stop_watchdog();
420 }
421 }
422 spin_unlock_irqrestore(&watchdog_lock, flags);
423}
424
425static int __clocksource_watchdog_kthread(void)
426{
427 struct clocksource *cs, *tmp;
428 unsigned long flags;
429 LIST_HEAD(unstable);
430 int select = 0;
431
432 spin_lock_irqsave(&watchdog_lock, flags);
433 list_for_each_entry_safe(cs, tmp, &watchdog_list, wd_list) {
434 if (cs->flags & CLOCK_SOURCE_UNSTABLE) {
435 list_del_init(&cs->wd_list);
436 list_add(&cs->wd_list, &unstable);
437 select = 1;
438 }
439 if (cs->flags & CLOCK_SOURCE_RESELECT) {
440 cs->flags &= ~CLOCK_SOURCE_RESELECT;
441 select = 1;
442 }
443 }
444 /* Check if the watchdog timer needs to be stopped. */
445 clocksource_stop_watchdog();
446 spin_unlock_irqrestore(&watchdog_lock, flags);
447
448 /* Needs to be done outside of watchdog lock */
449 list_for_each_entry_safe(cs, tmp, &unstable, wd_list) {
450 list_del_init(&cs->wd_list);
451 __clocksource_change_rating(cs, 0);
452 }
453 return select;
454}
455
456static int clocksource_watchdog_kthread(void *data)
457{
458 mutex_lock(&clocksource_mutex);
459 if (__clocksource_watchdog_kthread())
460 clocksource_select();
461 mutex_unlock(&clocksource_mutex);
462 return 0;
463}
464
465static bool clocksource_is_watchdog(struct clocksource *cs)
466{
467 return cs == watchdog;
468}
469
470#else /* CONFIG_CLOCKSOURCE_WATCHDOG */
471
472static void clocksource_enqueue_watchdog(struct clocksource *cs)
473{
474 if (cs->flags & CLOCK_SOURCE_IS_CONTINUOUS)
475 cs->flags |= CLOCK_SOURCE_VALID_FOR_HRES;
476}
477
478static inline void clocksource_dequeue_watchdog(struct clocksource *cs) { }
479static inline void clocksource_resume_watchdog(void) { }
480static inline int __clocksource_watchdog_kthread(void) { return 0; }
481static bool clocksource_is_watchdog(struct clocksource *cs) { return false; }
482void clocksource_mark_unstable(struct clocksource *cs) { }
483
484#endif /* CONFIG_CLOCKSOURCE_WATCHDOG */
485
486/**
487 * clocksource_suspend - suspend the clocksource(s)
488 */
489void clocksource_suspend(void)
490{
491 struct clocksource *cs;
492
493 list_for_each_entry_reverse(cs, &clocksource_list, list)
494 if (cs->suspend)
495 cs->suspend(cs);
496}
497
498/**
499 * clocksource_resume - resume the clocksource(s)
500 */
501void clocksource_resume(void)
502{
503 struct clocksource *cs;
504
505 list_for_each_entry(cs, &clocksource_list, list)
506 if (cs->resume)
507 cs->resume(cs);
508
509 clocksource_resume_watchdog();
510}
511
512/**
513 * clocksource_touch_watchdog - Update watchdog
514 *
515 * Update the watchdog after exception contexts such as kgdb so as not
516 * to incorrectly trip the watchdog. This might fail when the kernel
517 * was stopped in code which holds watchdog_lock.
518 */
519void clocksource_touch_watchdog(void)
520{
521 clocksource_resume_watchdog();
522}
523
524/**
525 * clocksource_max_adjustment- Returns max adjustment amount
526 * @cs: Pointer to clocksource
527 *
528 */
529static u32 clocksource_max_adjustment(struct clocksource *cs)
530{
531 u64 ret;
532 /*
533 * We won't try to correct for more than 11% adjustments (110,000 ppm),
534 */
535 ret = (u64)cs->mult * 11;
536 do_div(ret,100);
537 return (u32)ret;
538}
539
540/**
541 * clocks_calc_max_nsecs - Returns maximum nanoseconds that can be converted
542 * @mult: cycle to nanosecond multiplier
543 * @shift: cycle to nanosecond divisor (power of two)
544 * @maxadj: maximum adjustment value to mult (~11%)
545 * @mask: bitmask for two's complement subtraction of non 64 bit counters
546 */
547u64 clocks_calc_max_nsecs(u32 mult, u32 shift, u32 maxadj, u64 mask)
548{
549 u64 max_nsecs, max_cycles;
550
551 /*
552 * Calculate the maximum number of cycles that we can pass to the
553 * cyc2ns function without overflowing a 64-bit signed result. The
554 * maximum number of cycles is equal to ULLONG_MAX/(mult+maxadj)
555 * which is equivalent to the below.
556 * max_cycles < (2^63)/(mult + maxadj)
557 * max_cycles < 2^(log2((2^63)/(mult + maxadj)))
558 * max_cycles < 2^(log2(2^63) - log2(mult + maxadj))
559 * max_cycles < 2^(63 - log2(mult + maxadj))
560 * max_cycles < 1 << (63 - log2(mult + maxadj))
561 * Please note that we add 1 to the result of the log2 to account for
562 * any rounding errors, ensure the above inequality is satisfied and
563 * no overflow will occur.
564 */
565 max_cycles = 1ULL << (63 - (ilog2(mult + maxadj) + 1));
566
567 /*
568 * The actual maximum number of cycles we can defer the clocksource is
569 * determined by the minimum of max_cycles and mask.
570 * Note: Here we subtract the maxadj to make sure we don't sleep for
571 * too long if there's a large negative adjustment.
572 */
573 max_cycles = min(max_cycles, mask);
574 max_nsecs = clocksource_cyc2ns(max_cycles, mult - maxadj, shift);
575
576 return max_nsecs;
577}
578
579/**
580 * clocksource_max_deferment - Returns max time the clocksource can be deferred
581 * @cs: Pointer to clocksource
582 *
583 */
584static u64 clocksource_max_deferment(struct clocksource *cs)
585{
586 u64 max_nsecs;
587
588 max_nsecs = clocks_calc_max_nsecs(cs->mult, cs->shift, cs->maxadj,
589 cs->mask);
590 /*
591 * To ensure that the clocksource does not wrap whilst we are idle,
592 * limit the time the clocksource can be deferred by 12.5%. Please
593 * note a margin of 12.5% is used because this can be computed with
594 * a shift, versus say 10% which would require division.
595 */
596 return max_nsecs - (max_nsecs >> 3);
597}
598
599#ifndef CONFIG_ARCH_USES_GETTIMEOFFSET
600
601static struct clocksource *clocksource_find_best(bool oneshot, bool skipcur)
602{
603 struct clocksource *cs;
604
605 if (!finished_booting || list_empty(&clocksource_list))
606 return NULL;
607
608 /*
609 * We pick the clocksource with the highest rating. If oneshot
610 * mode is active, we pick the highres valid clocksource with
611 * the best rating.
612 */
613 list_for_each_entry(cs, &clocksource_list, list) {
614 if (skipcur && cs == curr_clocksource)
615 continue;
616 if (oneshot && !(cs->flags & CLOCK_SOURCE_VALID_FOR_HRES))
617 continue;
618 return cs;
619 }
620 return NULL;
621}
622
623static void __clocksource_select(bool skipcur)
624{
625 bool oneshot = tick_oneshot_mode_active();
626 struct clocksource *best, *cs;
627
628 /* Find the best suitable clocksource */
629 best = clocksource_find_best(oneshot, skipcur);
630 if (!best)
631 return;
632
633 /* Check for the override clocksource. */
634 list_for_each_entry(cs, &clocksource_list, list) {
635 if (skipcur && cs == curr_clocksource)
636 continue;
637 if (strcmp(cs->name, override_name) != 0)
638 continue;
639 /*
640 * Check to make sure we don't switch to a non-highres
641 * capable clocksource if the tick code is in oneshot
642 * mode (highres or nohz)
643 */
644 if (!(cs->flags & CLOCK_SOURCE_VALID_FOR_HRES) && oneshot) {
645 /* Override clocksource cannot be used. */
646 printk(KERN_WARNING "Override clocksource %s is not "
647 "HRT compatible. Cannot switch while in "
648 "HRT/NOHZ mode\n", cs->name);
649 override_name[0] = 0;
650 } else
651 /* Override clocksource can be used. */
652 best = cs;
653 break;
654 }
655
656 if (curr_clocksource != best && !timekeeping_notify(best)) {
657 pr_info("Switched to clocksource %s\n", best->name);
658 curr_clocksource = best;
659 }
660}
661
662/**
663 * clocksource_select - Select the best clocksource available
664 *
665 * Private function. Must hold clocksource_mutex when called.
666 *
667 * Select the clocksource with the best rating, or the clocksource,
668 * which is selected by userspace override.
669 */
670static void clocksource_select(void)
671{
672 return __clocksource_select(false);
673}
674
675static void clocksource_select_fallback(void)
676{
677 return __clocksource_select(true);
678}
679
680#else /* !CONFIG_ARCH_USES_GETTIMEOFFSET */
681
682static inline void clocksource_select(void) { }
683static inline void clocksource_select_fallback(void) { }
684
685#endif
686
687/*
688 * clocksource_done_booting - Called near the end of core bootup
689 *
690 * Hack to avoid lots of clocksource churn at boot time.
691 * We use fs_initcall because we want this to start before
692 * device_initcall but after subsys_initcall.
693 */
694static int __init clocksource_done_booting(void)
695{
696 mutex_lock(&clocksource_mutex);
697 curr_clocksource = clocksource_default_clock();
698 finished_booting = 1;
699 /*
700 * Run the watchdog first to eliminate unstable clock sources
701 */
702 __clocksource_watchdog_kthread();
703 clocksource_select();
704 mutex_unlock(&clocksource_mutex);
705 return 0;
706}
707fs_initcall(clocksource_done_booting);
708
709/*
710 * Enqueue the clocksource sorted by rating
711 */
712static void clocksource_enqueue(struct clocksource *cs)
713{
714 struct list_head *entry = &clocksource_list;
715 struct clocksource *tmp;
716
717 list_for_each_entry(tmp, &clocksource_list, list)
718 /* Keep track of the place, where to insert */
719 if (tmp->rating >= cs->rating)
720 entry = &tmp->list;
721 list_add(&cs->list, entry);
722}
723
724/**
725 * __clocksource_updatefreq_scale - Used update clocksource with new freq
726 * @cs: clocksource to be registered
727 * @scale: Scale factor multiplied against freq to get clocksource hz
728 * @freq: clocksource frequency (cycles per second) divided by scale
729 *
730 * This should only be called from the clocksource->enable() method.
731 *
732 * This *SHOULD NOT* be called directly! Please use the
733 * clocksource_updatefreq_hz() or clocksource_updatefreq_khz helper functions.
734 */
735void __clocksource_updatefreq_scale(struct clocksource *cs, u32 scale, u32 freq)
736{
737 u64 sec;
738 /*
739 * Calc the maximum number of seconds which we can run before
740 * wrapping around. For clocksources which have a mask > 32bit
741 * we need to limit the max sleep time to have a good
742 * conversion precision. 10 minutes is still a reasonable
743 * amount. That results in a shift value of 24 for a
744 * clocksource with mask >= 40bit and f >= 4GHz. That maps to
745 * ~ 0.06ppm granularity for NTP. We apply the same 12.5%
746 * margin as we do in clocksource_max_deferment()
747 */
748 sec = (cs->mask - (cs->mask >> 3));
749 do_div(sec, freq);
750 do_div(sec, scale);
751 if (!sec)
752 sec = 1;
753 else if (sec > 600 && cs->mask > UINT_MAX)
754 sec = 600;
755
756 clocks_calc_mult_shift(&cs->mult, &cs->shift, freq,
757 NSEC_PER_SEC / scale, sec * scale);
758
759 /*
760 * for clocksources that have large mults, to avoid overflow.
761 * Since mult may be adjusted by ntp, add an safety extra margin
762 *
763 */
764 cs->maxadj = clocksource_max_adjustment(cs);
765 while ((cs->mult + cs->maxadj < cs->mult)
766 || (cs->mult - cs->maxadj > cs->mult)) {
767 cs->mult >>= 1;
768 cs->shift--;
769 cs->maxadj = clocksource_max_adjustment(cs);
770 }
771
772 cs->max_idle_ns = clocksource_max_deferment(cs);
773}
774EXPORT_SYMBOL_GPL(__clocksource_updatefreq_scale);
775
776/**
777 * __clocksource_register_scale - Used to install new clocksources
778 * @cs: clocksource to be registered
779 * @scale: Scale factor multiplied against freq to get clocksource hz
780 * @freq: clocksource frequency (cycles per second) divided by scale
781 *
782 * Returns -EBUSY if registration fails, zero otherwise.
783 *
784 * This *SHOULD NOT* be called directly! Please use the
785 * clocksource_register_hz() or clocksource_register_khz helper functions.
786 */
787int __clocksource_register_scale(struct clocksource *cs, u32 scale, u32 freq)
788{
789
790 /* Initialize mult/shift and max_idle_ns */
791 __clocksource_updatefreq_scale(cs, scale, freq);
792
793 /* Add clocksource to the clcoksource list */
794 mutex_lock(&clocksource_mutex);
795 clocksource_enqueue(cs);
796 clocksource_enqueue_watchdog(cs);
797 clocksource_select();
798 mutex_unlock(&clocksource_mutex);
799 return 0;
800}
801EXPORT_SYMBOL_GPL(__clocksource_register_scale);
802
803
804/**
805 * clocksource_register - Used to install new clocksources
806 * @cs: clocksource to be registered
807 *
808 * Returns -EBUSY if registration fails, zero otherwise.
809 */
810int clocksource_register(struct clocksource *cs)
811{
812 /* calculate max adjustment for given mult/shift */
813 cs->maxadj = clocksource_max_adjustment(cs);
814 WARN_ONCE(cs->mult + cs->maxadj < cs->mult,
815 "Clocksource %s might overflow on 11%% adjustment\n",
816 cs->name);
817
818 /* calculate max idle time permitted for this clocksource */
819 cs->max_idle_ns = clocksource_max_deferment(cs);
820
821 mutex_lock(&clocksource_mutex);
822 clocksource_enqueue(cs);
823 clocksource_enqueue_watchdog(cs);
824 clocksource_select();
825 mutex_unlock(&clocksource_mutex);
826 return 0;
827}
828EXPORT_SYMBOL(clocksource_register);
829
830static void __clocksource_change_rating(struct clocksource *cs, int rating)
831{
832 list_del(&cs->list);
833 cs->rating = rating;
834 clocksource_enqueue(cs);
835}
836
837/**
838 * clocksource_change_rating - Change the rating of a registered clocksource
839 * @cs: clocksource to be changed
840 * @rating: new rating
841 */
842void clocksource_change_rating(struct clocksource *cs, int rating)
843{
844 mutex_lock(&clocksource_mutex);
845 __clocksource_change_rating(cs, rating);
846 clocksource_select();
847 mutex_unlock(&clocksource_mutex);
848}
849EXPORT_SYMBOL(clocksource_change_rating);
850
851/*
852 * Unbind clocksource @cs. Called with clocksource_mutex held
853 */
854static int clocksource_unbind(struct clocksource *cs)
855{
856 /*
857 * I really can't convince myself to support this on hardware
858 * designed by lobotomized monkeys.
859 */
860 if (clocksource_is_watchdog(cs))
861 return -EBUSY;
862
863 if (cs == curr_clocksource) {
864 /* Select and try to install a replacement clock source */
865 clocksource_select_fallback();
866 if (curr_clocksource == cs)
867 return -EBUSY;
868 }
869 clocksource_dequeue_watchdog(cs);
870 list_del_init(&cs->list);
871 return 0;
872}
873
874/**
875 * clocksource_unregister - remove a registered clocksource
876 * @cs: clocksource to be unregistered
877 */
878int clocksource_unregister(struct clocksource *cs)
879{
880 int ret = 0;
881
882 mutex_lock(&clocksource_mutex);
883 if (!list_empty(&cs->list))
884 ret = clocksource_unbind(cs);
885 mutex_unlock(&clocksource_mutex);
886 return ret;
887}
888EXPORT_SYMBOL(clocksource_unregister);
889
890#ifdef CONFIG_SYSFS
891/**
892 * sysfs_show_current_clocksources - sysfs interface for current clocksource
893 * @dev: unused
894 * @attr: unused
895 * @buf: char buffer to be filled with clocksource list
896 *
897 * Provides sysfs interface for listing current clocksource.
898 */
899static ssize_t
900sysfs_show_current_clocksources(struct device *dev,
901 struct device_attribute *attr, char *buf)
902{
903 ssize_t count = 0;
904
905 mutex_lock(&clocksource_mutex);
906 count = snprintf(buf, PAGE_SIZE, "%s\n", curr_clocksource->name);
907 mutex_unlock(&clocksource_mutex);
908
909 return count;
910}
911
912ssize_t sysfs_get_uname(const char *buf, char *dst, size_t cnt)
913{
914 size_t ret = cnt;
915
916 /* strings from sysfs write are not 0 terminated! */
917 if (!cnt || cnt >= CS_NAME_LEN)
918 return -EINVAL;
919
920 /* strip of \n: */
921 if (buf[cnt-1] == '\n')
922 cnt--;
923 if (cnt > 0)
924 memcpy(dst, buf, cnt);
925 dst[cnt] = 0;
926 return ret;
927}
928
929/**
930 * sysfs_override_clocksource - interface for manually overriding clocksource
931 * @dev: unused
932 * @attr: unused
933 * @buf: name of override clocksource
934 * @count: length of buffer
935 *
936 * Takes input from sysfs interface for manually overriding the default
937 * clocksource selection.
938 */
939static ssize_t sysfs_override_clocksource(struct device *dev,
940 struct device_attribute *attr,
941 const char *buf, size_t count)
942{
943 ssize_t ret;
944
945 mutex_lock(&clocksource_mutex);
946
947 ret = sysfs_get_uname(buf, override_name, count);
948 if (ret >= 0)
949 clocksource_select();
950
951 mutex_unlock(&clocksource_mutex);
952
953 return ret;
954}
955
956/**
957 * sysfs_unbind_current_clocksource - interface for manually unbinding clocksource
958 * @dev: unused
959 * @attr: unused
960 * @buf: unused
961 * @count: length of buffer
962 *
963 * Takes input from sysfs interface for manually unbinding a clocksource.
964 */
965static ssize_t sysfs_unbind_clocksource(struct device *dev,
966 struct device_attribute *attr,
967 const char *buf, size_t count)
968{
969 struct clocksource *cs;
970 char name[CS_NAME_LEN];
971 ssize_t ret;
972
973 ret = sysfs_get_uname(buf, name, count);
974 if (ret < 0)
975 return ret;
976
977 ret = -ENODEV;
978 mutex_lock(&clocksource_mutex);
979 list_for_each_entry(cs, &clocksource_list, list) {
980 if (strcmp(cs->name, name))
981 continue;
982 ret = clocksource_unbind(cs);
983 break;
984 }
985 mutex_unlock(&clocksource_mutex);
986
987 return ret ? ret : count;
988}
989
990/**
991 * sysfs_show_available_clocksources - sysfs interface for listing clocksource
992 * @dev: unused
993 * @attr: unused
994 * @buf: char buffer to be filled with clocksource list
995 *
996 * Provides sysfs interface for listing registered clocksources
997 */
998static ssize_t
999sysfs_show_available_clocksources(struct device *dev,
1000 struct device_attribute *attr,
1001 char *buf)
1002{
1003 struct clocksource *src;
1004 ssize_t count = 0;
1005
1006 mutex_lock(&clocksource_mutex);
1007 list_for_each_entry(src, &clocksource_list, list) {
1008 /*
1009 * Don't show non-HRES clocksource if the tick code is
1010 * in one shot mode (highres=on or nohz=on)
1011 */
1012 if (!tick_oneshot_mode_active() ||
1013 (src->flags & CLOCK_SOURCE_VALID_FOR_HRES))
1014 count += snprintf(buf + count,
1015 max((ssize_t)PAGE_SIZE - count, (ssize_t)0),
1016 "%s ", src->name);
1017 }
1018 mutex_unlock(&clocksource_mutex);
1019
1020 count += snprintf(buf + count,
1021 max((ssize_t)PAGE_SIZE - count, (ssize_t)0), "\n");
1022
1023 return count;
1024}
1025
1026/*
1027 * Sysfs setup bits:
1028 */
1029static DEVICE_ATTR(current_clocksource, 0644, sysfs_show_current_clocksources,
1030 sysfs_override_clocksource);
1031
1032static DEVICE_ATTR(unbind_clocksource, 0200, NULL, sysfs_unbind_clocksource);
1033
1034static DEVICE_ATTR(available_clocksource, 0444,
1035 sysfs_show_available_clocksources, NULL);
1036
1037static struct bus_type clocksource_subsys = {
1038 .name = "clocksource",
1039 .dev_name = "clocksource",
1040};
1041
1042static struct device device_clocksource = {
1043 .id = 0,
1044 .bus = &clocksource_subsys,
1045};
1046
1047static int __init init_clocksource_sysfs(void)
1048{
1049 int error = subsys_system_register(&clocksource_subsys, NULL);
1050
1051 if (!error)
1052 error = device_register(&device_clocksource);
1053 if (!error)
1054 error = device_create_file(
1055 &device_clocksource,
1056 &dev_attr_current_clocksource);
1057 if (!error)
1058 error = device_create_file(&device_clocksource,
1059 &dev_attr_unbind_clocksource);
1060 if (!error)
1061 error = device_create_file(
1062 &device_clocksource,
1063 &dev_attr_available_clocksource);
1064 return error;
1065}
1066
1067device_initcall(init_clocksource_sysfs);
1068#endif /* CONFIG_SYSFS */
1069
1070/**
1071 * boot_override_clocksource - boot clock override
1072 * @str: override name
1073 *
1074 * Takes a clocksource= boot argument and uses it
1075 * as the clocksource override name.
1076 */
1077static int __init boot_override_clocksource(char* str)
1078{
1079 mutex_lock(&clocksource_mutex);
1080 if (str)
1081 strlcpy(override_name, str, sizeof(override_name));
1082 mutex_unlock(&clocksource_mutex);
1083 return 1;
1084}
1085
1086__setup("clocksource=", boot_override_clocksource);
1087
1088/**
1089 * boot_override_clock - Compatibility layer for deprecated boot option
1090 * @str: override name
1091 *
1092 * DEPRECATED! Takes a clock= boot argument and uses it
1093 * as the clocksource override name
1094 */
1095static int __init boot_override_clock(char* str)
1096{
1097 if (!strcmp(str, "pmtmr")) {
1098 printk("Warning: clock=pmtmr is deprecated. "
1099 "Use clocksource=acpi_pm.\n");
1100 return boot_override_clocksource("acpi_pm");
1101 }
1102 printk("Warning! clock= boot option is deprecated. "
1103 "Use clocksource=xyz\n");
1104 return boot_override_clocksource(str);
1105}
1106
1107__setup("clock=", boot_override_clock);
1// SPDX-License-Identifier: GPL-2.0+
2/*
3 * This file contains the functions which manage clocksource drivers.
4 *
5 * Copyright (C) 2004, 2005 IBM, John Stultz (johnstul@us.ibm.com)
6 */
7
8#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9
10#include <linux/device.h>
11#include <linux/clocksource.h>
12#include <linux/init.h>
13#include <linux/module.h>
14#include <linux/sched.h> /* for spin_unlock_irq() using preempt_count() m68k */
15#include <linux/tick.h>
16#include <linux/kthread.h>
17#include <linux/prandom.h>
18#include <linux/cpu.h>
19
20#include "tick-internal.h"
21#include "timekeeping_internal.h"
22
23static void clocksource_enqueue(struct clocksource *cs);
24
25static noinline u64 cycles_to_nsec_safe(struct clocksource *cs, u64 start, u64 end)
26{
27 u64 delta = clocksource_delta(end, start, cs->mask, cs->max_raw_delta);
28
29 if (likely(delta < cs->max_cycles))
30 return clocksource_cyc2ns(delta, cs->mult, cs->shift);
31
32 return mul_u64_u32_shr(delta, cs->mult, cs->shift);
33}
34
35/**
36 * clocks_calc_mult_shift - calculate mult/shift factors for scaled math of clocks
37 * @mult: pointer to mult variable
38 * @shift: pointer to shift variable
39 * @from: frequency to convert from
40 * @to: frequency to convert to
41 * @maxsec: guaranteed runtime conversion range in seconds
42 *
43 * The function evaluates the shift/mult pair for the scaled math
44 * operations of clocksources and clockevents.
45 *
46 * @to and @from are frequency values in HZ. For clock sources @to is
47 * NSEC_PER_SEC == 1GHz and @from is the counter frequency. For clock
48 * event @to is the counter frequency and @from is NSEC_PER_SEC.
49 *
50 * The @maxsec conversion range argument controls the time frame in
51 * seconds which must be covered by the runtime conversion with the
52 * calculated mult and shift factors. This guarantees that no 64bit
53 * overflow happens when the input value of the conversion is
54 * multiplied with the calculated mult factor. Larger ranges may
55 * reduce the conversion accuracy by choosing smaller mult and shift
56 * factors.
57 */
58void
59clocks_calc_mult_shift(u32 *mult, u32 *shift, u32 from, u32 to, u32 maxsec)
60{
61 u64 tmp;
62 u32 sft, sftacc= 32;
63
64 /*
65 * Calculate the shift factor which is limiting the conversion
66 * range:
67 */
68 tmp = ((u64)maxsec * from) >> 32;
69 while (tmp) {
70 tmp >>=1;
71 sftacc--;
72 }
73
74 /*
75 * Find the conversion shift/mult pair which has the best
76 * accuracy and fits the maxsec conversion range:
77 */
78 for (sft = 32; sft > 0; sft--) {
79 tmp = (u64) to << sft;
80 tmp += from / 2;
81 do_div(tmp, from);
82 if ((tmp >> sftacc) == 0)
83 break;
84 }
85 *mult = tmp;
86 *shift = sft;
87}
88EXPORT_SYMBOL_GPL(clocks_calc_mult_shift);
89
90/*[Clocksource internal variables]---------
91 * curr_clocksource:
92 * currently selected clocksource.
93 * suspend_clocksource:
94 * used to calculate the suspend time.
95 * clocksource_list:
96 * linked list with the registered clocksources
97 * clocksource_mutex:
98 * protects manipulations to curr_clocksource and the clocksource_list
99 * override_name:
100 * Name of the user-specified clocksource.
101 */
102static struct clocksource *curr_clocksource;
103static struct clocksource *suspend_clocksource;
104static LIST_HEAD(clocksource_list);
105static DEFINE_MUTEX(clocksource_mutex);
106static char override_name[CS_NAME_LEN];
107static int finished_booting;
108static u64 suspend_start;
109
110/*
111 * Interval: 0.5sec.
112 */
113#define WATCHDOG_INTERVAL (HZ >> 1)
114#define WATCHDOG_INTERVAL_MAX_NS ((2 * WATCHDOG_INTERVAL) * (NSEC_PER_SEC / HZ))
115
116/*
117 * Threshold: 0.0312s, when doubled: 0.0625s.
118 */
119#define WATCHDOG_THRESHOLD (NSEC_PER_SEC >> 5)
120
121/*
122 * Maximum permissible delay between two readouts of the watchdog
123 * clocksource surrounding a read of the clocksource being validated.
124 * This delay could be due to SMIs, NMIs, or to VCPU preemptions. Used as
125 * a lower bound for cs->uncertainty_margin values when registering clocks.
126 *
127 * The default of 500 parts per million is based on NTP's limits.
128 * If a clocksource is good enough for NTP, it is good enough for us!
129 *
130 * In other words, by default, even if a clocksource is extremely
131 * precise (for example, with a sub-nanosecond period), the maximum
132 * permissible skew between the clocksource watchdog and the clocksource
133 * under test is not permitted to go below the 500ppm minimum defined
134 * by MAX_SKEW_USEC. This 500ppm minimum may be overridden using the
135 * CLOCKSOURCE_WATCHDOG_MAX_SKEW_US Kconfig option.
136 */
137#ifdef CONFIG_CLOCKSOURCE_WATCHDOG_MAX_SKEW_US
138#define MAX_SKEW_USEC CONFIG_CLOCKSOURCE_WATCHDOG_MAX_SKEW_US
139#else
140#define MAX_SKEW_USEC (125 * WATCHDOG_INTERVAL / HZ)
141#endif
142
143/*
144 * Default for maximum permissible skew when cs->uncertainty_margin is
145 * not specified, and the lower bound even when cs->uncertainty_margin
146 * is specified. This is also the default that is used when registering
147 * clocks with unspecifed cs->uncertainty_margin, so this macro is used
148 * even in CONFIG_CLOCKSOURCE_WATCHDOG=n kernels.
149 */
150#define WATCHDOG_MAX_SKEW (MAX_SKEW_USEC * NSEC_PER_USEC)
151
152#ifdef CONFIG_CLOCKSOURCE_WATCHDOG
153static void clocksource_watchdog_work(struct work_struct *work);
154static void clocksource_select(void);
155
156static LIST_HEAD(watchdog_list);
157static struct clocksource *watchdog;
158static struct timer_list watchdog_timer;
159static DECLARE_WORK(watchdog_work, clocksource_watchdog_work);
160static DEFINE_SPINLOCK(watchdog_lock);
161static int watchdog_running;
162static atomic_t watchdog_reset_pending;
163static int64_t watchdog_max_interval;
164
165static inline void clocksource_watchdog_lock(unsigned long *flags)
166{
167 spin_lock_irqsave(&watchdog_lock, *flags);
168}
169
170static inline void clocksource_watchdog_unlock(unsigned long *flags)
171{
172 spin_unlock_irqrestore(&watchdog_lock, *flags);
173}
174
175static int clocksource_watchdog_kthread(void *data);
176
177static void clocksource_watchdog_work(struct work_struct *work)
178{
179 /*
180 * We cannot directly run clocksource_watchdog_kthread() here, because
181 * clocksource_select() calls timekeeping_notify() which uses
182 * stop_machine(). One cannot use stop_machine() from a workqueue() due
183 * lock inversions wrt CPU hotplug.
184 *
185 * Also, we only ever run this work once or twice during the lifetime
186 * of the kernel, so there is no point in creating a more permanent
187 * kthread for this.
188 *
189 * If kthread_run fails the next watchdog scan over the
190 * watchdog_list will find the unstable clock again.
191 */
192 kthread_run(clocksource_watchdog_kthread, NULL, "kwatchdog");
193}
194
195static void clocksource_change_rating(struct clocksource *cs, int rating)
196{
197 list_del(&cs->list);
198 cs->rating = rating;
199 clocksource_enqueue(cs);
200}
201
202static void __clocksource_unstable(struct clocksource *cs)
203{
204 cs->flags &= ~(CLOCK_SOURCE_VALID_FOR_HRES | CLOCK_SOURCE_WATCHDOG);
205 cs->flags |= CLOCK_SOURCE_UNSTABLE;
206
207 /*
208 * If the clocksource is registered clocksource_watchdog_kthread() will
209 * re-rate and re-select.
210 */
211 if (list_empty(&cs->list)) {
212 cs->rating = 0;
213 return;
214 }
215
216 if (cs->mark_unstable)
217 cs->mark_unstable(cs);
218
219 /* kick clocksource_watchdog_kthread() */
220 if (finished_booting)
221 schedule_work(&watchdog_work);
222}
223
224/**
225 * clocksource_mark_unstable - mark clocksource unstable via watchdog
226 * @cs: clocksource to be marked unstable
227 *
228 * This function is called by the x86 TSC code to mark clocksources as unstable;
229 * it defers demotion and re-selection to a kthread.
230 */
231void clocksource_mark_unstable(struct clocksource *cs)
232{
233 unsigned long flags;
234
235 spin_lock_irqsave(&watchdog_lock, flags);
236 if (!(cs->flags & CLOCK_SOURCE_UNSTABLE)) {
237 if (!list_empty(&cs->list) && list_empty(&cs->wd_list))
238 list_add(&cs->wd_list, &watchdog_list);
239 __clocksource_unstable(cs);
240 }
241 spin_unlock_irqrestore(&watchdog_lock, flags);
242}
243
244static int verify_n_cpus = 8;
245module_param(verify_n_cpus, int, 0644);
246
247enum wd_read_status {
248 WD_READ_SUCCESS,
249 WD_READ_UNSTABLE,
250 WD_READ_SKIP
251};
252
253static enum wd_read_status cs_watchdog_read(struct clocksource *cs, u64 *csnow, u64 *wdnow)
254{
255 int64_t md = 2 * watchdog->uncertainty_margin;
256 unsigned int nretries, max_retries;
257 int64_t wd_delay, wd_seq_delay;
258 u64 wd_end, wd_end2;
259
260 max_retries = clocksource_get_max_watchdog_retry();
261 for (nretries = 0; nretries <= max_retries; nretries++) {
262 local_irq_disable();
263 *wdnow = watchdog->read(watchdog);
264 *csnow = cs->read(cs);
265 wd_end = watchdog->read(watchdog);
266 wd_end2 = watchdog->read(watchdog);
267 local_irq_enable();
268
269 wd_delay = cycles_to_nsec_safe(watchdog, *wdnow, wd_end);
270 if (wd_delay <= md + cs->uncertainty_margin) {
271 if (nretries > 1 && nretries >= max_retries) {
272 pr_warn("timekeeping watchdog on CPU%d: %s retried %d times before success\n",
273 smp_processor_id(), watchdog->name, nretries);
274 }
275 return WD_READ_SUCCESS;
276 }
277
278 /*
279 * Now compute delay in consecutive watchdog read to see if
280 * there is too much external interferences that cause
281 * significant delay in reading both clocksource and watchdog.
282 *
283 * If consecutive WD read-back delay > md, report
284 * system busy, reinit the watchdog and skip the current
285 * watchdog test.
286 */
287 wd_seq_delay = cycles_to_nsec_safe(watchdog, wd_end, wd_end2);
288 if (wd_seq_delay > md)
289 goto skip_test;
290 }
291
292 pr_warn("timekeeping watchdog on CPU%d: wd-%s-wd excessive read-back delay of %lldns vs. limit of %ldns, wd-wd read-back delay only %lldns, attempt %d, marking %s unstable\n",
293 smp_processor_id(), cs->name, wd_delay, WATCHDOG_MAX_SKEW, wd_seq_delay, nretries, cs->name);
294 return WD_READ_UNSTABLE;
295
296skip_test:
297 pr_info("timekeeping watchdog on CPU%d: %s wd-wd read-back delay of %lldns\n",
298 smp_processor_id(), watchdog->name, wd_seq_delay);
299 pr_info("wd-%s-wd read-back delay of %lldns, clock-skew test skipped!\n",
300 cs->name, wd_delay);
301 return WD_READ_SKIP;
302}
303
304static u64 csnow_mid;
305static cpumask_t cpus_ahead;
306static cpumask_t cpus_behind;
307static cpumask_t cpus_chosen;
308
309static void clocksource_verify_choose_cpus(void)
310{
311 int cpu, i, n = verify_n_cpus;
312
313 if (n < 0) {
314 /* Check all of the CPUs. */
315 cpumask_copy(&cpus_chosen, cpu_online_mask);
316 cpumask_clear_cpu(smp_processor_id(), &cpus_chosen);
317 return;
318 }
319
320 /* If no checking desired, or no other CPU to check, leave. */
321 cpumask_clear(&cpus_chosen);
322 if (n == 0 || num_online_cpus() <= 1)
323 return;
324
325 /* Make sure to select at least one CPU other than the current CPU. */
326 cpu = cpumask_first(cpu_online_mask);
327 if (cpu == smp_processor_id())
328 cpu = cpumask_next(cpu, cpu_online_mask);
329 if (WARN_ON_ONCE(cpu >= nr_cpu_ids))
330 return;
331 cpumask_set_cpu(cpu, &cpus_chosen);
332
333 /* Force a sane value for the boot parameter. */
334 if (n > nr_cpu_ids)
335 n = nr_cpu_ids;
336
337 /*
338 * Randomly select the specified number of CPUs. If the same
339 * CPU is selected multiple times, that CPU is checked only once,
340 * and no replacement CPU is selected. This gracefully handles
341 * situations where verify_n_cpus is greater than the number of
342 * CPUs that are currently online.
343 */
344 for (i = 1; i < n; i++) {
345 cpu = get_random_u32_below(nr_cpu_ids);
346 cpu = cpumask_next(cpu - 1, cpu_online_mask);
347 if (cpu >= nr_cpu_ids)
348 cpu = cpumask_first(cpu_online_mask);
349 if (!WARN_ON_ONCE(cpu >= nr_cpu_ids))
350 cpumask_set_cpu(cpu, &cpus_chosen);
351 }
352
353 /* Don't verify ourselves. */
354 cpumask_clear_cpu(smp_processor_id(), &cpus_chosen);
355}
356
357static void clocksource_verify_one_cpu(void *csin)
358{
359 struct clocksource *cs = (struct clocksource *)csin;
360
361 csnow_mid = cs->read(cs);
362}
363
364void clocksource_verify_percpu(struct clocksource *cs)
365{
366 int64_t cs_nsec, cs_nsec_max = 0, cs_nsec_min = LLONG_MAX;
367 u64 csnow_begin, csnow_end;
368 int cpu, testcpu;
369 s64 delta;
370
371 if (verify_n_cpus == 0)
372 return;
373 cpumask_clear(&cpus_ahead);
374 cpumask_clear(&cpus_behind);
375 cpus_read_lock();
376 migrate_disable();
377 clocksource_verify_choose_cpus();
378 if (cpumask_empty(&cpus_chosen)) {
379 migrate_enable();
380 cpus_read_unlock();
381 pr_warn("Not enough CPUs to check clocksource '%s'.\n", cs->name);
382 return;
383 }
384 testcpu = smp_processor_id();
385 pr_info("Checking clocksource %s synchronization from CPU %d to CPUs %*pbl.\n",
386 cs->name, testcpu, cpumask_pr_args(&cpus_chosen));
387 preempt_disable();
388 for_each_cpu(cpu, &cpus_chosen) {
389 if (cpu == testcpu)
390 continue;
391 csnow_begin = cs->read(cs);
392 smp_call_function_single(cpu, clocksource_verify_one_cpu, cs, 1);
393 csnow_end = cs->read(cs);
394 delta = (s64)((csnow_mid - csnow_begin) & cs->mask);
395 if (delta < 0)
396 cpumask_set_cpu(cpu, &cpus_behind);
397 delta = (csnow_end - csnow_mid) & cs->mask;
398 if (delta < 0)
399 cpumask_set_cpu(cpu, &cpus_ahead);
400 cs_nsec = cycles_to_nsec_safe(cs, csnow_begin, csnow_end);
401 if (cs_nsec > cs_nsec_max)
402 cs_nsec_max = cs_nsec;
403 if (cs_nsec < cs_nsec_min)
404 cs_nsec_min = cs_nsec;
405 }
406 preempt_enable();
407 migrate_enable();
408 cpus_read_unlock();
409 if (!cpumask_empty(&cpus_ahead))
410 pr_warn(" CPUs %*pbl ahead of CPU %d for clocksource %s.\n",
411 cpumask_pr_args(&cpus_ahead), testcpu, cs->name);
412 if (!cpumask_empty(&cpus_behind))
413 pr_warn(" CPUs %*pbl behind CPU %d for clocksource %s.\n",
414 cpumask_pr_args(&cpus_behind), testcpu, cs->name);
415 if (!cpumask_empty(&cpus_ahead) || !cpumask_empty(&cpus_behind))
416 pr_warn(" CPU %d check durations %lldns - %lldns for clocksource %s.\n",
417 testcpu, cs_nsec_min, cs_nsec_max, cs->name);
418}
419EXPORT_SYMBOL_GPL(clocksource_verify_percpu);
420
421static inline void clocksource_reset_watchdog(void)
422{
423 struct clocksource *cs;
424
425 list_for_each_entry(cs, &watchdog_list, wd_list)
426 cs->flags &= ~CLOCK_SOURCE_WATCHDOG;
427}
428
429
430static void clocksource_watchdog(struct timer_list *unused)
431{
432 int64_t wd_nsec, cs_nsec, interval;
433 u64 csnow, wdnow, cslast, wdlast;
434 int next_cpu, reset_pending;
435 struct clocksource *cs;
436 enum wd_read_status read_ret;
437 unsigned long extra_wait = 0;
438 u32 md;
439
440 spin_lock(&watchdog_lock);
441 if (!watchdog_running)
442 goto out;
443
444 reset_pending = atomic_read(&watchdog_reset_pending);
445
446 list_for_each_entry(cs, &watchdog_list, wd_list) {
447
448 /* Clocksource already marked unstable? */
449 if (cs->flags & CLOCK_SOURCE_UNSTABLE) {
450 if (finished_booting)
451 schedule_work(&watchdog_work);
452 continue;
453 }
454
455 read_ret = cs_watchdog_read(cs, &csnow, &wdnow);
456
457 if (read_ret == WD_READ_UNSTABLE) {
458 /* Clock readout unreliable, so give it up. */
459 __clocksource_unstable(cs);
460 continue;
461 }
462
463 /*
464 * When WD_READ_SKIP is returned, it means the system is likely
465 * under very heavy load, where the latency of reading
466 * watchdog/clocksource is very big, and affect the accuracy of
467 * watchdog check. So give system some space and suspend the
468 * watchdog check for 5 minutes.
469 */
470 if (read_ret == WD_READ_SKIP) {
471 /*
472 * As the watchdog timer will be suspended, and
473 * cs->last could keep unchanged for 5 minutes, reset
474 * the counters.
475 */
476 clocksource_reset_watchdog();
477 extra_wait = HZ * 300;
478 break;
479 }
480
481 /* Clocksource initialized ? */
482 if (!(cs->flags & CLOCK_SOURCE_WATCHDOG) ||
483 atomic_read(&watchdog_reset_pending)) {
484 cs->flags |= CLOCK_SOURCE_WATCHDOG;
485 cs->wd_last = wdnow;
486 cs->cs_last = csnow;
487 continue;
488 }
489
490 wd_nsec = cycles_to_nsec_safe(watchdog, cs->wd_last, wdnow);
491 cs_nsec = cycles_to_nsec_safe(cs, cs->cs_last, csnow);
492 wdlast = cs->wd_last; /* save these in case we print them */
493 cslast = cs->cs_last;
494 cs->cs_last = csnow;
495 cs->wd_last = wdnow;
496
497 if (atomic_read(&watchdog_reset_pending))
498 continue;
499
500 /*
501 * The processing of timer softirqs can get delayed (usually
502 * on account of ksoftirqd not getting to run in a timely
503 * manner), which causes the watchdog interval to stretch.
504 * Skew detection may fail for longer watchdog intervals
505 * on account of fixed margins being used.
506 * Some clocksources, e.g. acpi_pm, cannot tolerate
507 * watchdog intervals longer than a few seconds.
508 */
509 interval = max(cs_nsec, wd_nsec);
510 if (unlikely(interval > WATCHDOG_INTERVAL_MAX_NS)) {
511 if (system_state > SYSTEM_SCHEDULING &&
512 interval > 2 * watchdog_max_interval) {
513 watchdog_max_interval = interval;
514 pr_warn("Long readout interval, skipping watchdog check: cs_nsec: %lld wd_nsec: %lld\n",
515 cs_nsec, wd_nsec);
516 }
517 watchdog_timer.expires = jiffies;
518 continue;
519 }
520
521 /* Check the deviation from the watchdog clocksource. */
522 md = cs->uncertainty_margin + watchdog->uncertainty_margin;
523 if (abs(cs_nsec - wd_nsec) > md) {
524 s64 cs_wd_msec;
525 s64 wd_msec;
526 u32 wd_rem;
527
528 pr_warn("timekeeping watchdog on CPU%d: Marking clocksource '%s' as unstable because the skew is too large:\n",
529 smp_processor_id(), cs->name);
530 pr_warn(" '%s' wd_nsec: %lld wd_now: %llx wd_last: %llx mask: %llx\n",
531 watchdog->name, wd_nsec, wdnow, wdlast, watchdog->mask);
532 pr_warn(" '%s' cs_nsec: %lld cs_now: %llx cs_last: %llx mask: %llx\n",
533 cs->name, cs_nsec, csnow, cslast, cs->mask);
534 cs_wd_msec = div_s64_rem(cs_nsec - wd_nsec, 1000 * 1000, &wd_rem);
535 wd_msec = div_s64_rem(wd_nsec, 1000 * 1000, &wd_rem);
536 pr_warn(" Clocksource '%s' skewed %lld ns (%lld ms) over watchdog '%s' interval of %lld ns (%lld ms)\n",
537 cs->name, cs_nsec - wd_nsec, cs_wd_msec, watchdog->name, wd_nsec, wd_msec);
538 if (curr_clocksource == cs)
539 pr_warn(" '%s' is current clocksource.\n", cs->name);
540 else if (curr_clocksource)
541 pr_warn(" '%s' (not '%s') is current clocksource.\n", curr_clocksource->name, cs->name);
542 else
543 pr_warn(" No current clocksource.\n");
544 __clocksource_unstable(cs);
545 continue;
546 }
547
548 if (cs == curr_clocksource && cs->tick_stable)
549 cs->tick_stable(cs);
550
551 if (!(cs->flags & CLOCK_SOURCE_VALID_FOR_HRES) &&
552 (cs->flags & CLOCK_SOURCE_IS_CONTINUOUS) &&
553 (watchdog->flags & CLOCK_SOURCE_IS_CONTINUOUS)) {
554 /* Mark it valid for high-res. */
555 cs->flags |= CLOCK_SOURCE_VALID_FOR_HRES;
556
557 /*
558 * clocksource_done_booting() will sort it if
559 * finished_booting is not set yet.
560 */
561 if (!finished_booting)
562 continue;
563
564 /*
565 * If this is not the current clocksource let
566 * the watchdog thread reselect it. Due to the
567 * change to high res this clocksource might
568 * be preferred now. If it is the current
569 * clocksource let the tick code know about
570 * that change.
571 */
572 if (cs != curr_clocksource) {
573 cs->flags |= CLOCK_SOURCE_RESELECT;
574 schedule_work(&watchdog_work);
575 } else {
576 tick_clock_notify();
577 }
578 }
579 }
580
581 /*
582 * We only clear the watchdog_reset_pending, when we did a
583 * full cycle through all clocksources.
584 */
585 if (reset_pending)
586 atomic_dec(&watchdog_reset_pending);
587
588 /*
589 * Cycle through CPUs to check if the CPUs stay synchronized
590 * to each other.
591 */
592 next_cpu = cpumask_next(raw_smp_processor_id(), cpu_online_mask);
593 if (next_cpu >= nr_cpu_ids)
594 next_cpu = cpumask_first(cpu_online_mask);
595
596 /*
597 * Arm timer if not already pending: could race with concurrent
598 * pair clocksource_stop_watchdog() clocksource_start_watchdog().
599 */
600 if (!timer_pending(&watchdog_timer)) {
601 watchdog_timer.expires += WATCHDOG_INTERVAL + extra_wait;
602 add_timer_on(&watchdog_timer, next_cpu);
603 }
604out:
605 spin_unlock(&watchdog_lock);
606}
607
608static inline void clocksource_start_watchdog(void)
609{
610 if (watchdog_running || !watchdog || list_empty(&watchdog_list))
611 return;
612 timer_setup(&watchdog_timer, clocksource_watchdog, 0);
613 watchdog_timer.expires = jiffies + WATCHDOG_INTERVAL;
614 add_timer_on(&watchdog_timer, cpumask_first(cpu_online_mask));
615 watchdog_running = 1;
616}
617
618static inline void clocksource_stop_watchdog(void)
619{
620 if (!watchdog_running || (watchdog && !list_empty(&watchdog_list)))
621 return;
622 del_timer(&watchdog_timer);
623 watchdog_running = 0;
624}
625
626static void clocksource_resume_watchdog(void)
627{
628 atomic_inc(&watchdog_reset_pending);
629}
630
631static void clocksource_enqueue_watchdog(struct clocksource *cs)
632{
633 INIT_LIST_HEAD(&cs->wd_list);
634
635 if (cs->flags & CLOCK_SOURCE_MUST_VERIFY) {
636 /* cs is a clocksource to be watched. */
637 list_add(&cs->wd_list, &watchdog_list);
638 cs->flags &= ~CLOCK_SOURCE_WATCHDOG;
639 } else {
640 /* cs is a watchdog. */
641 if (cs->flags & CLOCK_SOURCE_IS_CONTINUOUS)
642 cs->flags |= CLOCK_SOURCE_VALID_FOR_HRES;
643 }
644}
645
646static void clocksource_select_watchdog(bool fallback)
647{
648 struct clocksource *cs, *old_wd;
649 unsigned long flags;
650
651 spin_lock_irqsave(&watchdog_lock, flags);
652 /* save current watchdog */
653 old_wd = watchdog;
654 if (fallback)
655 watchdog = NULL;
656
657 list_for_each_entry(cs, &clocksource_list, list) {
658 /* cs is a clocksource to be watched. */
659 if (cs->flags & CLOCK_SOURCE_MUST_VERIFY)
660 continue;
661
662 /* Skip current if we were requested for a fallback. */
663 if (fallback && cs == old_wd)
664 continue;
665
666 /* Pick the best watchdog. */
667 if (!watchdog || cs->rating > watchdog->rating)
668 watchdog = cs;
669 }
670 /* If we failed to find a fallback restore the old one. */
671 if (!watchdog)
672 watchdog = old_wd;
673
674 /* If we changed the watchdog we need to reset cycles. */
675 if (watchdog != old_wd)
676 clocksource_reset_watchdog();
677
678 /* Check if the watchdog timer needs to be started. */
679 clocksource_start_watchdog();
680 spin_unlock_irqrestore(&watchdog_lock, flags);
681}
682
683static void clocksource_dequeue_watchdog(struct clocksource *cs)
684{
685 if (cs != watchdog) {
686 if (cs->flags & CLOCK_SOURCE_MUST_VERIFY) {
687 /* cs is a watched clocksource. */
688 list_del_init(&cs->wd_list);
689 /* Check if the watchdog timer needs to be stopped. */
690 clocksource_stop_watchdog();
691 }
692 }
693}
694
695static int __clocksource_watchdog_kthread(void)
696{
697 struct clocksource *cs, *tmp;
698 unsigned long flags;
699 int select = 0;
700
701 /* Do any required per-CPU skew verification. */
702 if (curr_clocksource &&
703 curr_clocksource->flags & CLOCK_SOURCE_UNSTABLE &&
704 curr_clocksource->flags & CLOCK_SOURCE_VERIFY_PERCPU)
705 clocksource_verify_percpu(curr_clocksource);
706
707 spin_lock_irqsave(&watchdog_lock, flags);
708 list_for_each_entry_safe(cs, tmp, &watchdog_list, wd_list) {
709 if (cs->flags & CLOCK_SOURCE_UNSTABLE) {
710 list_del_init(&cs->wd_list);
711 clocksource_change_rating(cs, 0);
712 select = 1;
713 }
714 if (cs->flags & CLOCK_SOURCE_RESELECT) {
715 cs->flags &= ~CLOCK_SOURCE_RESELECT;
716 select = 1;
717 }
718 }
719 /* Check if the watchdog timer needs to be stopped. */
720 clocksource_stop_watchdog();
721 spin_unlock_irqrestore(&watchdog_lock, flags);
722
723 return select;
724}
725
726static int clocksource_watchdog_kthread(void *data)
727{
728 mutex_lock(&clocksource_mutex);
729 if (__clocksource_watchdog_kthread())
730 clocksource_select();
731 mutex_unlock(&clocksource_mutex);
732 return 0;
733}
734
735static bool clocksource_is_watchdog(struct clocksource *cs)
736{
737 return cs == watchdog;
738}
739
740#else /* CONFIG_CLOCKSOURCE_WATCHDOG */
741
742static void clocksource_enqueue_watchdog(struct clocksource *cs)
743{
744 if (cs->flags & CLOCK_SOURCE_IS_CONTINUOUS)
745 cs->flags |= CLOCK_SOURCE_VALID_FOR_HRES;
746}
747
748static void clocksource_select_watchdog(bool fallback) { }
749static inline void clocksource_dequeue_watchdog(struct clocksource *cs) { }
750static inline void clocksource_resume_watchdog(void) { }
751static inline int __clocksource_watchdog_kthread(void) { return 0; }
752static bool clocksource_is_watchdog(struct clocksource *cs) { return false; }
753void clocksource_mark_unstable(struct clocksource *cs) { }
754
755static inline void clocksource_watchdog_lock(unsigned long *flags) { }
756static inline void clocksource_watchdog_unlock(unsigned long *flags) { }
757
758#endif /* CONFIG_CLOCKSOURCE_WATCHDOG */
759
760static bool clocksource_is_suspend(struct clocksource *cs)
761{
762 return cs == suspend_clocksource;
763}
764
765static void __clocksource_suspend_select(struct clocksource *cs)
766{
767 /*
768 * Skip the clocksource which will be stopped in suspend state.
769 */
770 if (!(cs->flags & CLOCK_SOURCE_SUSPEND_NONSTOP))
771 return;
772
773 /*
774 * The nonstop clocksource can be selected as the suspend clocksource to
775 * calculate the suspend time, so it should not supply suspend/resume
776 * interfaces to suspend the nonstop clocksource when system suspends.
777 */
778 if (cs->suspend || cs->resume) {
779 pr_warn("Nonstop clocksource %s should not supply suspend/resume interfaces\n",
780 cs->name);
781 }
782
783 /* Pick the best rating. */
784 if (!suspend_clocksource || cs->rating > suspend_clocksource->rating)
785 suspend_clocksource = cs;
786}
787
788/**
789 * clocksource_suspend_select - Select the best clocksource for suspend timing
790 * @fallback: if select a fallback clocksource
791 */
792static void clocksource_suspend_select(bool fallback)
793{
794 struct clocksource *cs, *old_suspend;
795
796 old_suspend = suspend_clocksource;
797 if (fallback)
798 suspend_clocksource = NULL;
799
800 list_for_each_entry(cs, &clocksource_list, list) {
801 /* Skip current if we were requested for a fallback. */
802 if (fallback && cs == old_suspend)
803 continue;
804
805 __clocksource_suspend_select(cs);
806 }
807}
808
809/**
810 * clocksource_start_suspend_timing - Start measuring the suspend timing
811 * @cs: current clocksource from timekeeping
812 * @start_cycles: current cycles from timekeeping
813 *
814 * This function will save the start cycle values of suspend timer to calculate
815 * the suspend time when resuming system.
816 *
817 * This function is called late in the suspend process from timekeeping_suspend(),
818 * that means processes are frozen, non-boot cpus and interrupts are disabled
819 * now. It is therefore possible to start the suspend timer without taking the
820 * clocksource mutex.
821 */
822void clocksource_start_suspend_timing(struct clocksource *cs, u64 start_cycles)
823{
824 if (!suspend_clocksource)
825 return;
826
827 /*
828 * If current clocksource is the suspend timer, we should use the
829 * tkr_mono.cycle_last value as suspend_start to avoid same reading
830 * from suspend timer.
831 */
832 if (clocksource_is_suspend(cs)) {
833 suspend_start = start_cycles;
834 return;
835 }
836
837 if (suspend_clocksource->enable &&
838 suspend_clocksource->enable(suspend_clocksource)) {
839 pr_warn_once("Failed to enable the non-suspend-able clocksource.\n");
840 return;
841 }
842
843 suspend_start = suspend_clocksource->read(suspend_clocksource);
844}
845
846/**
847 * clocksource_stop_suspend_timing - Stop measuring the suspend timing
848 * @cs: current clocksource from timekeeping
849 * @cycle_now: current cycles from timekeeping
850 *
851 * This function will calculate the suspend time from suspend timer.
852 *
853 * Returns nanoseconds since suspend started, 0 if no usable suspend clocksource.
854 *
855 * This function is called early in the resume process from timekeeping_resume(),
856 * that means there is only one cpu, no processes are running and the interrupts
857 * are disabled. It is therefore possible to stop the suspend timer without
858 * taking the clocksource mutex.
859 */
860u64 clocksource_stop_suspend_timing(struct clocksource *cs, u64 cycle_now)
861{
862 u64 now, nsec = 0;
863
864 if (!suspend_clocksource)
865 return 0;
866
867 /*
868 * If current clocksource is the suspend timer, we should use the
869 * tkr_mono.cycle_last value from timekeeping as current cycle to
870 * avoid same reading from suspend timer.
871 */
872 if (clocksource_is_suspend(cs))
873 now = cycle_now;
874 else
875 now = suspend_clocksource->read(suspend_clocksource);
876
877 if (now > suspend_start)
878 nsec = cycles_to_nsec_safe(suspend_clocksource, suspend_start, now);
879
880 /*
881 * Disable the suspend timer to save power if current clocksource is
882 * not the suspend timer.
883 */
884 if (!clocksource_is_suspend(cs) && suspend_clocksource->disable)
885 suspend_clocksource->disable(suspend_clocksource);
886
887 return nsec;
888}
889
890/**
891 * clocksource_suspend - suspend the clocksource(s)
892 */
893void clocksource_suspend(void)
894{
895 struct clocksource *cs;
896
897 list_for_each_entry_reverse(cs, &clocksource_list, list)
898 if (cs->suspend)
899 cs->suspend(cs);
900}
901
902/**
903 * clocksource_resume - resume the clocksource(s)
904 */
905void clocksource_resume(void)
906{
907 struct clocksource *cs;
908
909 list_for_each_entry(cs, &clocksource_list, list)
910 if (cs->resume)
911 cs->resume(cs);
912
913 clocksource_resume_watchdog();
914}
915
916/**
917 * clocksource_touch_watchdog - Update watchdog
918 *
919 * Update the watchdog after exception contexts such as kgdb so as not
920 * to incorrectly trip the watchdog. This might fail when the kernel
921 * was stopped in code which holds watchdog_lock.
922 */
923void clocksource_touch_watchdog(void)
924{
925 clocksource_resume_watchdog();
926}
927
928/**
929 * clocksource_max_adjustment- Returns max adjustment amount
930 * @cs: Pointer to clocksource
931 *
932 */
933static u32 clocksource_max_adjustment(struct clocksource *cs)
934{
935 u64 ret;
936 /*
937 * We won't try to correct for more than 11% adjustments (110,000 ppm),
938 */
939 ret = (u64)cs->mult * 11;
940 do_div(ret,100);
941 return (u32)ret;
942}
943
944/**
945 * clocks_calc_max_nsecs - Returns maximum nanoseconds that can be converted
946 * @mult: cycle to nanosecond multiplier
947 * @shift: cycle to nanosecond divisor (power of two)
948 * @maxadj: maximum adjustment value to mult (~11%)
949 * @mask: bitmask for two's complement subtraction of non 64 bit counters
950 * @max_cyc: maximum cycle value before potential overflow (does not include
951 * any safety margin)
952 *
953 * NOTE: This function includes a safety margin of 50%, in other words, we
954 * return half the number of nanoseconds the hardware counter can technically
955 * cover. This is done so that we can potentially detect problems caused by
956 * delayed timers or bad hardware, which might result in time intervals that
957 * are larger than what the math used can handle without overflows.
958 */
959u64 clocks_calc_max_nsecs(u32 mult, u32 shift, u32 maxadj, u64 mask, u64 *max_cyc)
960{
961 u64 max_nsecs, max_cycles;
962
963 /*
964 * Calculate the maximum number of cycles that we can pass to the
965 * cyc2ns() function without overflowing a 64-bit result.
966 */
967 max_cycles = ULLONG_MAX;
968 do_div(max_cycles, mult+maxadj);
969
970 /*
971 * The actual maximum number of cycles we can defer the clocksource is
972 * determined by the minimum of max_cycles and mask.
973 * Note: Here we subtract the maxadj to make sure we don't sleep for
974 * too long if there's a large negative adjustment.
975 */
976 max_cycles = min(max_cycles, mask);
977 max_nsecs = clocksource_cyc2ns(max_cycles, mult - maxadj, shift);
978
979 /* return the max_cycles value as well if requested */
980 if (max_cyc)
981 *max_cyc = max_cycles;
982
983 /* Return 50% of the actual maximum, so we can detect bad values */
984 max_nsecs >>= 1;
985
986 return max_nsecs;
987}
988
989/**
990 * clocksource_update_max_deferment - Updates the clocksource max_idle_ns & max_cycles
991 * @cs: Pointer to clocksource to be updated
992 *
993 */
994static inline void clocksource_update_max_deferment(struct clocksource *cs)
995{
996 cs->max_idle_ns = clocks_calc_max_nsecs(cs->mult, cs->shift,
997 cs->maxadj, cs->mask,
998 &cs->max_cycles);
999
1000 /*
1001 * Threshold for detecting negative motion in clocksource_delta().
1002 *
1003 * Allow for 0.875 of the counter width so that overly long idle
1004 * sleeps, which go slightly over mask/2, do not trigger the
1005 * negative motion detection.
1006 */
1007 cs->max_raw_delta = (cs->mask >> 1) + (cs->mask >> 2) + (cs->mask >> 3);
1008}
1009
1010static struct clocksource *clocksource_find_best(bool oneshot, bool skipcur)
1011{
1012 struct clocksource *cs;
1013
1014 if (!finished_booting || list_empty(&clocksource_list))
1015 return NULL;
1016
1017 /*
1018 * We pick the clocksource with the highest rating. If oneshot
1019 * mode is active, we pick the highres valid clocksource with
1020 * the best rating.
1021 */
1022 list_for_each_entry(cs, &clocksource_list, list) {
1023 if (skipcur && cs == curr_clocksource)
1024 continue;
1025 if (oneshot && !(cs->flags & CLOCK_SOURCE_VALID_FOR_HRES))
1026 continue;
1027 return cs;
1028 }
1029 return NULL;
1030}
1031
1032static void __clocksource_select(bool skipcur)
1033{
1034 bool oneshot = tick_oneshot_mode_active();
1035 struct clocksource *best, *cs;
1036
1037 /* Find the best suitable clocksource */
1038 best = clocksource_find_best(oneshot, skipcur);
1039 if (!best)
1040 return;
1041
1042 if (!strlen(override_name))
1043 goto found;
1044
1045 /* Check for the override clocksource. */
1046 list_for_each_entry(cs, &clocksource_list, list) {
1047 if (skipcur && cs == curr_clocksource)
1048 continue;
1049 if (strcmp(cs->name, override_name) != 0)
1050 continue;
1051 /*
1052 * Check to make sure we don't switch to a non-highres
1053 * capable clocksource if the tick code is in oneshot
1054 * mode (highres or nohz)
1055 */
1056 if (!(cs->flags & CLOCK_SOURCE_VALID_FOR_HRES) && oneshot) {
1057 /* Override clocksource cannot be used. */
1058 if (cs->flags & CLOCK_SOURCE_UNSTABLE) {
1059 pr_warn("Override clocksource %s is unstable and not HRT compatible - cannot switch while in HRT/NOHZ mode\n",
1060 cs->name);
1061 override_name[0] = 0;
1062 } else {
1063 /*
1064 * The override cannot be currently verified.
1065 * Deferring to let the watchdog check.
1066 */
1067 pr_info("Override clocksource %s is not currently HRT compatible - deferring\n",
1068 cs->name);
1069 }
1070 } else
1071 /* Override clocksource can be used. */
1072 best = cs;
1073 break;
1074 }
1075
1076found:
1077 if (curr_clocksource != best && !timekeeping_notify(best)) {
1078 pr_info("Switched to clocksource %s\n", best->name);
1079 curr_clocksource = best;
1080 }
1081}
1082
1083/**
1084 * clocksource_select - Select the best clocksource available
1085 *
1086 * Private function. Must hold clocksource_mutex when called.
1087 *
1088 * Select the clocksource with the best rating, or the clocksource,
1089 * which is selected by userspace override.
1090 */
1091static void clocksource_select(void)
1092{
1093 __clocksource_select(false);
1094}
1095
1096static void clocksource_select_fallback(void)
1097{
1098 __clocksource_select(true);
1099}
1100
1101/*
1102 * clocksource_done_booting - Called near the end of core bootup
1103 *
1104 * Hack to avoid lots of clocksource churn at boot time.
1105 * We use fs_initcall because we want this to start before
1106 * device_initcall but after subsys_initcall.
1107 */
1108static int __init clocksource_done_booting(void)
1109{
1110 mutex_lock(&clocksource_mutex);
1111 curr_clocksource = clocksource_default_clock();
1112 finished_booting = 1;
1113 /*
1114 * Run the watchdog first to eliminate unstable clock sources
1115 */
1116 __clocksource_watchdog_kthread();
1117 clocksource_select();
1118 mutex_unlock(&clocksource_mutex);
1119 return 0;
1120}
1121fs_initcall(clocksource_done_booting);
1122
1123/*
1124 * Enqueue the clocksource sorted by rating
1125 */
1126static void clocksource_enqueue(struct clocksource *cs)
1127{
1128 struct list_head *entry = &clocksource_list;
1129 struct clocksource *tmp;
1130
1131 list_for_each_entry(tmp, &clocksource_list, list) {
1132 /* Keep track of the place, where to insert */
1133 if (tmp->rating < cs->rating)
1134 break;
1135 entry = &tmp->list;
1136 }
1137 list_add(&cs->list, entry);
1138}
1139
1140/**
1141 * __clocksource_update_freq_scale - Used update clocksource with new freq
1142 * @cs: clocksource to be registered
1143 * @scale: Scale factor multiplied against freq to get clocksource hz
1144 * @freq: clocksource frequency (cycles per second) divided by scale
1145 *
1146 * This should only be called from the clocksource->enable() method.
1147 *
1148 * This *SHOULD NOT* be called directly! Please use the
1149 * __clocksource_update_freq_hz() or __clocksource_update_freq_khz() helper
1150 * functions.
1151 */
1152void __clocksource_update_freq_scale(struct clocksource *cs, u32 scale, u32 freq)
1153{
1154 u64 sec;
1155
1156 /*
1157 * Default clocksources are *special* and self-define their mult/shift.
1158 * But, you're not special, so you should specify a freq value.
1159 */
1160 if (freq) {
1161 /*
1162 * Calc the maximum number of seconds which we can run before
1163 * wrapping around. For clocksources which have a mask > 32-bit
1164 * we need to limit the max sleep time to have a good
1165 * conversion precision. 10 minutes is still a reasonable
1166 * amount. That results in a shift value of 24 for a
1167 * clocksource with mask >= 40-bit and f >= 4GHz. That maps to
1168 * ~ 0.06ppm granularity for NTP.
1169 */
1170 sec = cs->mask;
1171 do_div(sec, freq);
1172 do_div(sec, scale);
1173 if (!sec)
1174 sec = 1;
1175 else if (sec > 600 && cs->mask > UINT_MAX)
1176 sec = 600;
1177
1178 clocks_calc_mult_shift(&cs->mult, &cs->shift, freq,
1179 NSEC_PER_SEC / scale, sec * scale);
1180 }
1181
1182 /*
1183 * If the uncertainty margin is not specified, calculate it. If
1184 * both scale and freq are non-zero, calculate the clock period, but
1185 * bound below at 2*WATCHDOG_MAX_SKEW, that is, 500ppm by default.
1186 * However, if either of scale or freq is zero, be very conservative
1187 * and take the tens-of-milliseconds WATCHDOG_THRESHOLD value
1188 * for the uncertainty margin. Allow stupidly small uncertainty
1189 * margins to be specified by the caller for testing purposes,
1190 * but warn to discourage production use of this capability.
1191 *
1192 * Bottom line: The sum of the uncertainty margins of the
1193 * watchdog clocksource and the clocksource under test will be at
1194 * least 500ppm by default. For more information, please see the
1195 * comment preceding CONFIG_CLOCKSOURCE_WATCHDOG_MAX_SKEW_US above.
1196 */
1197 if (scale && freq && !cs->uncertainty_margin) {
1198 cs->uncertainty_margin = NSEC_PER_SEC / (scale * freq);
1199 if (cs->uncertainty_margin < 2 * WATCHDOG_MAX_SKEW)
1200 cs->uncertainty_margin = 2 * WATCHDOG_MAX_SKEW;
1201 } else if (!cs->uncertainty_margin) {
1202 cs->uncertainty_margin = WATCHDOG_THRESHOLD;
1203 }
1204 WARN_ON_ONCE(cs->uncertainty_margin < 2 * WATCHDOG_MAX_SKEW);
1205
1206 /*
1207 * Ensure clocksources that have large 'mult' values don't overflow
1208 * when adjusted.
1209 */
1210 cs->maxadj = clocksource_max_adjustment(cs);
1211 while (freq && ((cs->mult + cs->maxadj < cs->mult)
1212 || (cs->mult - cs->maxadj > cs->mult))) {
1213 cs->mult >>= 1;
1214 cs->shift--;
1215 cs->maxadj = clocksource_max_adjustment(cs);
1216 }
1217
1218 /*
1219 * Only warn for *special* clocksources that self-define
1220 * their mult/shift values and don't specify a freq.
1221 */
1222 WARN_ONCE(cs->mult + cs->maxadj < cs->mult,
1223 "timekeeping: Clocksource %s might overflow on 11%% adjustment\n",
1224 cs->name);
1225
1226 clocksource_update_max_deferment(cs);
1227
1228 pr_info("%s: mask: 0x%llx max_cycles: 0x%llx, max_idle_ns: %lld ns\n",
1229 cs->name, cs->mask, cs->max_cycles, cs->max_idle_ns);
1230}
1231EXPORT_SYMBOL_GPL(__clocksource_update_freq_scale);
1232
1233/**
1234 * __clocksource_register_scale - Used to install new clocksources
1235 * @cs: clocksource to be registered
1236 * @scale: Scale factor multiplied against freq to get clocksource hz
1237 * @freq: clocksource frequency (cycles per second) divided by scale
1238 *
1239 * Returns -EBUSY if registration fails, zero otherwise.
1240 *
1241 * This *SHOULD NOT* be called directly! Please use the
1242 * clocksource_register_hz() or clocksource_register_khz helper functions.
1243 */
1244int __clocksource_register_scale(struct clocksource *cs, u32 scale, u32 freq)
1245{
1246 unsigned long flags;
1247
1248 clocksource_arch_init(cs);
1249
1250 if (WARN_ON_ONCE((unsigned int)cs->id >= CSID_MAX))
1251 cs->id = CSID_GENERIC;
1252 if (cs->vdso_clock_mode < 0 ||
1253 cs->vdso_clock_mode >= VDSO_CLOCKMODE_MAX) {
1254 pr_warn("clocksource %s registered with invalid VDSO mode %d. Disabling VDSO support.\n",
1255 cs->name, cs->vdso_clock_mode);
1256 cs->vdso_clock_mode = VDSO_CLOCKMODE_NONE;
1257 }
1258
1259 /* Initialize mult/shift and max_idle_ns */
1260 __clocksource_update_freq_scale(cs, scale, freq);
1261
1262 /* Add clocksource to the clocksource list */
1263 mutex_lock(&clocksource_mutex);
1264
1265 clocksource_watchdog_lock(&flags);
1266 clocksource_enqueue(cs);
1267 clocksource_enqueue_watchdog(cs);
1268 clocksource_watchdog_unlock(&flags);
1269
1270 clocksource_select();
1271 clocksource_select_watchdog(false);
1272 __clocksource_suspend_select(cs);
1273 mutex_unlock(&clocksource_mutex);
1274 return 0;
1275}
1276EXPORT_SYMBOL_GPL(__clocksource_register_scale);
1277
1278/*
1279 * Unbind clocksource @cs. Called with clocksource_mutex held
1280 */
1281static int clocksource_unbind(struct clocksource *cs)
1282{
1283 unsigned long flags;
1284
1285 if (clocksource_is_watchdog(cs)) {
1286 /* Select and try to install a replacement watchdog. */
1287 clocksource_select_watchdog(true);
1288 if (clocksource_is_watchdog(cs))
1289 return -EBUSY;
1290 }
1291
1292 if (cs == curr_clocksource) {
1293 /* Select and try to install a replacement clock source */
1294 clocksource_select_fallback();
1295 if (curr_clocksource == cs)
1296 return -EBUSY;
1297 }
1298
1299 if (clocksource_is_suspend(cs)) {
1300 /*
1301 * Select and try to install a replacement suspend clocksource.
1302 * If no replacement suspend clocksource, we will just let the
1303 * clocksource go and have no suspend clocksource.
1304 */
1305 clocksource_suspend_select(true);
1306 }
1307
1308 clocksource_watchdog_lock(&flags);
1309 clocksource_dequeue_watchdog(cs);
1310 list_del_init(&cs->list);
1311 clocksource_watchdog_unlock(&flags);
1312
1313 return 0;
1314}
1315
1316/**
1317 * clocksource_unregister - remove a registered clocksource
1318 * @cs: clocksource to be unregistered
1319 */
1320int clocksource_unregister(struct clocksource *cs)
1321{
1322 int ret = 0;
1323
1324 mutex_lock(&clocksource_mutex);
1325 if (!list_empty(&cs->list))
1326 ret = clocksource_unbind(cs);
1327 mutex_unlock(&clocksource_mutex);
1328 return ret;
1329}
1330EXPORT_SYMBOL(clocksource_unregister);
1331
1332#ifdef CONFIG_SYSFS
1333/**
1334 * current_clocksource_show - sysfs interface for current clocksource
1335 * @dev: unused
1336 * @attr: unused
1337 * @buf: char buffer to be filled with clocksource list
1338 *
1339 * Provides sysfs interface for listing current clocksource.
1340 */
1341static ssize_t current_clocksource_show(struct device *dev,
1342 struct device_attribute *attr,
1343 char *buf)
1344{
1345 ssize_t count = 0;
1346
1347 mutex_lock(&clocksource_mutex);
1348 count = sysfs_emit(buf, "%s\n", curr_clocksource->name);
1349 mutex_unlock(&clocksource_mutex);
1350
1351 return count;
1352}
1353
1354ssize_t sysfs_get_uname(const char *buf, char *dst, size_t cnt)
1355{
1356 size_t ret = cnt;
1357
1358 /* strings from sysfs write are not 0 terminated! */
1359 if (!cnt || cnt >= CS_NAME_LEN)
1360 return -EINVAL;
1361
1362 /* strip of \n: */
1363 if (buf[cnt-1] == '\n')
1364 cnt--;
1365 if (cnt > 0)
1366 memcpy(dst, buf, cnt);
1367 dst[cnt] = 0;
1368 return ret;
1369}
1370
1371/**
1372 * current_clocksource_store - interface for manually overriding clocksource
1373 * @dev: unused
1374 * @attr: unused
1375 * @buf: name of override clocksource
1376 * @count: length of buffer
1377 *
1378 * Takes input from sysfs interface for manually overriding the default
1379 * clocksource selection.
1380 */
1381static ssize_t current_clocksource_store(struct device *dev,
1382 struct device_attribute *attr,
1383 const char *buf, size_t count)
1384{
1385 ssize_t ret;
1386
1387 mutex_lock(&clocksource_mutex);
1388
1389 ret = sysfs_get_uname(buf, override_name, count);
1390 if (ret >= 0)
1391 clocksource_select();
1392
1393 mutex_unlock(&clocksource_mutex);
1394
1395 return ret;
1396}
1397static DEVICE_ATTR_RW(current_clocksource);
1398
1399/**
1400 * unbind_clocksource_store - interface for manually unbinding clocksource
1401 * @dev: unused
1402 * @attr: unused
1403 * @buf: unused
1404 * @count: length of buffer
1405 *
1406 * Takes input from sysfs interface for manually unbinding a clocksource.
1407 */
1408static ssize_t unbind_clocksource_store(struct device *dev,
1409 struct device_attribute *attr,
1410 const char *buf, size_t count)
1411{
1412 struct clocksource *cs;
1413 char name[CS_NAME_LEN];
1414 ssize_t ret;
1415
1416 ret = sysfs_get_uname(buf, name, count);
1417 if (ret < 0)
1418 return ret;
1419
1420 ret = -ENODEV;
1421 mutex_lock(&clocksource_mutex);
1422 list_for_each_entry(cs, &clocksource_list, list) {
1423 if (strcmp(cs->name, name))
1424 continue;
1425 ret = clocksource_unbind(cs);
1426 break;
1427 }
1428 mutex_unlock(&clocksource_mutex);
1429
1430 return ret ? ret : count;
1431}
1432static DEVICE_ATTR_WO(unbind_clocksource);
1433
1434/**
1435 * available_clocksource_show - sysfs interface for listing clocksource
1436 * @dev: unused
1437 * @attr: unused
1438 * @buf: char buffer to be filled with clocksource list
1439 *
1440 * Provides sysfs interface for listing registered clocksources
1441 */
1442static ssize_t available_clocksource_show(struct device *dev,
1443 struct device_attribute *attr,
1444 char *buf)
1445{
1446 struct clocksource *src;
1447 ssize_t count = 0;
1448
1449 mutex_lock(&clocksource_mutex);
1450 list_for_each_entry(src, &clocksource_list, list) {
1451 /*
1452 * Don't show non-HRES clocksource if the tick code is
1453 * in one shot mode (highres=on or nohz=on)
1454 */
1455 if (!tick_oneshot_mode_active() ||
1456 (src->flags & CLOCK_SOURCE_VALID_FOR_HRES))
1457 count += snprintf(buf + count,
1458 max((ssize_t)PAGE_SIZE - count, (ssize_t)0),
1459 "%s ", src->name);
1460 }
1461 mutex_unlock(&clocksource_mutex);
1462
1463 count += snprintf(buf + count,
1464 max((ssize_t)PAGE_SIZE - count, (ssize_t)0), "\n");
1465
1466 return count;
1467}
1468static DEVICE_ATTR_RO(available_clocksource);
1469
1470static struct attribute *clocksource_attrs[] = {
1471 &dev_attr_current_clocksource.attr,
1472 &dev_attr_unbind_clocksource.attr,
1473 &dev_attr_available_clocksource.attr,
1474 NULL
1475};
1476ATTRIBUTE_GROUPS(clocksource);
1477
1478static const struct bus_type clocksource_subsys = {
1479 .name = "clocksource",
1480 .dev_name = "clocksource",
1481};
1482
1483static struct device device_clocksource = {
1484 .id = 0,
1485 .bus = &clocksource_subsys,
1486 .groups = clocksource_groups,
1487};
1488
1489static int __init init_clocksource_sysfs(void)
1490{
1491 int error = subsys_system_register(&clocksource_subsys, NULL);
1492
1493 if (!error)
1494 error = device_register(&device_clocksource);
1495
1496 return error;
1497}
1498
1499device_initcall(init_clocksource_sysfs);
1500#endif /* CONFIG_SYSFS */
1501
1502/**
1503 * boot_override_clocksource - boot clock override
1504 * @str: override name
1505 *
1506 * Takes a clocksource= boot argument and uses it
1507 * as the clocksource override name.
1508 */
1509static int __init boot_override_clocksource(char* str)
1510{
1511 mutex_lock(&clocksource_mutex);
1512 if (str)
1513 strscpy(override_name, str, sizeof(override_name));
1514 mutex_unlock(&clocksource_mutex);
1515 return 1;
1516}
1517
1518__setup("clocksource=", boot_override_clocksource);
1519
1520/**
1521 * boot_override_clock - Compatibility layer for deprecated boot option
1522 * @str: override name
1523 *
1524 * DEPRECATED! Takes a clock= boot argument and uses it
1525 * as the clocksource override name
1526 */
1527static int __init boot_override_clock(char* str)
1528{
1529 if (!strcmp(str, "pmtmr")) {
1530 pr_warn("clock=pmtmr is deprecated - use clocksource=acpi_pm\n");
1531 return boot_override_clocksource("acpi_pm");
1532 }
1533 pr_warn("clock= boot option is deprecated - use clocksource=xyz\n");
1534 return boot_override_clocksource(str);
1535}
1536
1537__setup("clock=", boot_override_clock);