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#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
27
28#include <linux/device.h>
29#include <linux/clocksource.h>
30#include <linux/init.h>
31#include <linux/module.h>
32#include <linux/sched.h> /* for spin_unlock_irq() using preempt_count() m68k */
33#include <linux/tick.h>
34#include <linux/kthread.h>
35
36#include "tick-internal.h"
37#include "timekeeping_internal.h"
38
39/**
40 * clocks_calc_mult_shift - calculate mult/shift factors for scaled math of clocks
41 * @mult: pointer to mult variable
42 * @shift: pointer to shift variable
43 * @from: frequency to convert from
44 * @to: frequency to convert to
45 * @maxsec: guaranteed runtime conversion range in seconds
46 *
47 * The function evaluates the shift/mult pair for the scaled math
48 * operations of clocksources and clockevents.
49 *
50 * @to and @from are frequency values in HZ. For clock sources @to is
51 * NSEC_PER_SEC == 1GHz and @from is the counter frequency. For clock
52 * event @to is the counter frequency and @from is NSEC_PER_SEC.
53 *
54 * The @maxsec conversion range argument controls the time frame in
55 * seconds which must be covered by the runtime conversion with the
56 * calculated mult and shift factors. This guarantees that no 64bit
57 * overflow happens when the input value of the conversion is
58 * multiplied with the calculated mult factor. Larger ranges may
59 * reduce the conversion accuracy by chosing smaller mult and shift
60 * factors.
61 */
62void
63clocks_calc_mult_shift(u32 *mult, u32 *shift, u32 from, u32 to, u32 maxsec)
64{
65 u64 tmp;
66 u32 sft, sftacc= 32;
67
68 /*
69 * Calculate the shift factor which is limiting the conversion
70 * range:
71 */
72 tmp = ((u64)maxsec * from) >> 32;
73 while (tmp) {
74 tmp >>=1;
75 sftacc--;
76 }
77
78 /*
79 * Find the conversion shift/mult pair which has the best
80 * accuracy and fits the maxsec conversion range:
81 */
82 for (sft = 32; sft > 0; sft--) {
83 tmp = (u64) to << sft;
84 tmp += from / 2;
85 do_div(tmp, from);
86 if ((tmp >> sftacc) == 0)
87 break;
88 }
89 *mult = tmp;
90 *shift = sft;
91}
92
93/*[Clocksource internal variables]---------
94 * curr_clocksource:
95 * currently selected clocksource.
96 * clocksource_list:
97 * linked list with the registered clocksources
98 * clocksource_mutex:
99 * protects manipulations to curr_clocksource and the clocksource_list
100 * override_name:
101 * Name of the user-specified clocksource.
102 */
103static struct clocksource *curr_clocksource;
104static LIST_HEAD(clocksource_list);
105static DEFINE_MUTEX(clocksource_mutex);
106static char override_name[CS_NAME_LEN];
107static int finished_booting;
108
109#ifdef CONFIG_CLOCKSOURCE_WATCHDOG
110static void clocksource_watchdog_work(struct work_struct *work);
111static void clocksource_select(void);
112
113static LIST_HEAD(watchdog_list);
114static struct clocksource *watchdog;
115static struct timer_list watchdog_timer;
116static DECLARE_WORK(watchdog_work, clocksource_watchdog_work);
117static DEFINE_SPINLOCK(watchdog_lock);
118static int watchdog_running;
119static atomic_t watchdog_reset_pending;
120
121static int clocksource_watchdog_kthread(void *data);
122static void __clocksource_change_rating(struct clocksource *cs, int rating);
123
124/*
125 * Interval: 0.5sec Threshold: 0.0625s
126 */
127#define WATCHDOG_INTERVAL (HZ >> 1)
128#define WATCHDOG_THRESHOLD (NSEC_PER_SEC >> 4)
129
130static void clocksource_watchdog_work(struct work_struct *work)
131{
132 /*
133 * If kthread_run fails the next watchdog scan over the
134 * watchdog_list will find the unstable clock again.
135 */
136 kthread_run(clocksource_watchdog_kthread, NULL, "kwatchdog");
137}
138
139static void __clocksource_unstable(struct clocksource *cs)
140{
141 cs->flags &= ~(CLOCK_SOURCE_VALID_FOR_HRES | CLOCK_SOURCE_WATCHDOG);
142 cs->flags |= CLOCK_SOURCE_UNSTABLE;
143 if (finished_booting)
144 schedule_work(&watchdog_work);
145}
146
147/**
148 * clocksource_mark_unstable - mark clocksource unstable via watchdog
149 * @cs: clocksource to be marked unstable
150 *
151 * This function is called instead of clocksource_change_rating from
152 * cpu hotplug code to avoid a deadlock between the clocksource mutex
153 * and the cpu hotplug mutex. It defers the update of the clocksource
154 * to the watchdog thread.
155 */
156void clocksource_mark_unstable(struct clocksource *cs)
157{
158 unsigned long flags;
159
160 spin_lock_irqsave(&watchdog_lock, flags);
161 if (!(cs->flags & CLOCK_SOURCE_UNSTABLE)) {
162 if (list_empty(&cs->wd_list))
163 list_add(&cs->wd_list, &watchdog_list);
164 __clocksource_unstable(cs);
165 }
166 spin_unlock_irqrestore(&watchdog_lock, flags);
167}
168
169static void clocksource_watchdog(unsigned long data)
170{
171 struct clocksource *cs;
172 cycle_t csnow, wdnow, cslast, wdlast, delta;
173 int64_t wd_nsec, cs_nsec;
174 int next_cpu, reset_pending;
175
176 spin_lock(&watchdog_lock);
177 if (!watchdog_running)
178 goto out;
179
180 reset_pending = atomic_read(&watchdog_reset_pending);
181
182 list_for_each_entry(cs, &watchdog_list, wd_list) {
183
184 /* Clocksource already marked unstable? */
185 if (cs->flags & CLOCK_SOURCE_UNSTABLE) {
186 if (finished_booting)
187 schedule_work(&watchdog_work);
188 continue;
189 }
190
191 local_irq_disable();
192 csnow = cs->read(cs);
193 wdnow = watchdog->read(watchdog);
194 local_irq_enable();
195
196 /* Clocksource initialized ? */
197 if (!(cs->flags & CLOCK_SOURCE_WATCHDOG) ||
198 atomic_read(&watchdog_reset_pending)) {
199 cs->flags |= CLOCK_SOURCE_WATCHDOG;
200 cs->wd_last = wdnow;
201 cs->cs_last = csnow;
202 continue;
203 }
204
205 delta = clocksource_delta(wdnow, cs->wd_last, watchdog->mask);
206 wd_nsec = clocksource_cyc2ns(delta, watchdog->mult,
207 watchdog->shift);
208
209 delta = clocksource_delta(csnow, cs->cs_last, cs->mask);
210 cs_nsec = clocksource_cyc2ns(delta, cs->mult, cs->shift);
211 wdlast = cs->wd_last; /* save these in case we print them */
212 cslast = cs->cs_last;
213 cs->cs_last = csnow;
214 cs->wd_last = wdnow;
215
216 if (atomic_read(&watchdog_reset_pending))
217 continue;
218
219 /* Check the deviation from the watchdog clocksource. */
220 if (abs(cs_nsec - wd_nsec) > WATCHDOG_THRESHOLD) {
221 pr_warn("timekeeping watchdog on CPU%d: Marking clocksource '%s' as unstable because the skew is too large:\n",
222 smp_processor_id(), cs->name);
223 pr_warn(" '%s' wd_now: %llx wd_last: %llx mask: %llx\n",
224 watchdog->name, wdnow, wdlast, watchdog->mask);
225 pr_warn(" '%s' cs_now: %llx cs_last: %llx mask: %llx\n",
226 cs->name, csnow, cslast, cs->mask);
227 __clocksource_unstable(cs);
228 continue;
229 }
230
231 if (!(cs->flags & CLOCK_SOURCE_VALID_FOR_HRES) &&
232 (cs->flags & CLOCK_SOURCE_IS_CONTINUOUS) &&
233 (watchdog->flags & CLOCK_SOURCE_IS_CONTINUOUS)) {
234 /* Mark it valid for high-res. */
235 cs->flags |= CLOCK_SOURCE_VALID_FOR_HRES;
236
237 /*
238 * clocksource_done_booting() will sort it if
239 * finished_booting is not set yet.
240 */
241 if (!finished_booting)
242 continue;
243
244 /*
245 * If this is not the current clocksource let
246 * the watchdog thread reselect it. Due to the
247 * change to high res this clocksource might
248 * be preferred now. If it is the current
249 * clocksource let the tick code know about
250 * that change.
251 */
252 if (cs != curr_clocksource) {
253 cs->flags |= CLOCK_SOURCE_RESELECT;
254 schedule_work(&watchdog_work);
255 } else {
256 tick_clock_notify();
257 }
258 }
259 }
260
261 /*
262 * We only clear the watchdog_reset_pending, when we did a
263 * full cycle through all clocksources.
264 */
265 if (reset_pending)
266 atomic_dec(&watchdog_reset_pending);
267
268 /*
269 * Cycle through CPUs to check if the CPUs stay synchronized
270 * to each other.
271 */
272 next_cpu = cpumask_next(raw_smp_processor_id(), cpu_online_mask);
273 if (next_cpu >= nr_cpu_ids)
274 next_cpu = cpumask_first(cpu_online_mask);
275 watchdog_timer.expires += WATCHDOG_INTERVAL;
276 add_timer_on(&watchdog_timer, next_cpu);
277out:
278 spin_unlock(&watchdog_lock);
279}
280
281static inline void clocksource_start_watchdog(void)
282{
283 if (watchdog_running || !watchdog || list_empty(&watchdog_list))
284 return;
285 init_timer(&watchdog_timer);
286 watchdog_timer.function = clocksource_watchdog;
287 watchdog_timer.expires = jiffies + WATCHDOG_INTERVAL;
288 add_timer_on(&watchdog_timer, cpumask_first(cpu_online_mask));
289 watchdog_running = 1;
290}
291
292static inline void clocksource_stop_watchdog(void)
293{
294 if (!watchdog_running || (watchdog && !list_empty(&watchdog_list)))
295 return;
296 del_timer(&watchdog_timer);
297 watchdog_running = 0;
298}
299
300static inline void clocksource_reset_watchdog(void)
301{
302 struct clocksource *cs;
303
304 list_for_each_entry(cs, &watchdog_list, wd_list)
305 cs->flags &= ~CLOCK_SOURCE_WATCHDOG;
306}
307
308static void clocksource_resume_watchdog(void)
309{
310 atomic_inc(&watchdog_reset_pending);
311}
312
313static void clocksource_enqueue_watchdog(struct clocksource *cs)
314{
315 unsigned long flags;
316
317 spin_lock_irqsave(&watchdog_lock, flags);
318 if (cs->flags & CLOCK_SOURCE_MUST_VERIFY) {
319 /* cs is a clocksource to be watched. */
320 list_add(&cs->wd_list, &watchdog_list);
321 cs->flags &= ~CLOCK_SOURCE_WATCHDOG;
322 } else {
323 /* cs is a watchdog. */
324 if (cs->flags & CLOCK_SOURCE_IS_CONTINUOUS)
325 cs->flags |= CLOCK_SOURCE_VALID_FOR_HRES;
326 }
327 spin_unlock_irqrestore(&watchdog_lock, flags);
328}
329
330static void clocksource_select_watchdog(bool fallback)
331{
332 struct clocksource *cs, *old_wd;
333 unsigned long flags;
334
335 spin_lock_irqsave(&watchdog_lock, flags);
336 /* save current watchdog */
337 old_wd = watchdog;
338 if (fallback)
339 watchdog = NULL;
340
341 list_for_each_entry(cs, &clocksource_list, list) {
342 /* cs is a clocksource to be watched. */
343 if (cs->flags & CLOCK_SOURCE_MUST_VERIFY)
344 continue;
345
346 /* Skip current if we were requested for a fallback. */
347 if (fallback && cs == old_wd)
348 continue;
349
350 /* Pick the best watchdog. */
351 if (!watchdog || cs->rating > watchdog->rating)
352 watchdog = cs;
353 }
354 /* If we failed to find a fallback restore the old one. */
355 if (!watchdog)
356 watchdog = old_wd;
357
358 /* If we changed the watchdog we need to reset cycles. */
359 if (watchdog != old_wd)
360 clocksource_reset_watchdog();
361
362 /* Check if the watchdog timer needs to be started. */
363 clocksource_start_watchdog();
364 spin_unlock_irqrestore(&watchdog_lock, flags);
365}
366
367static void clocksource_dequeue_watchdog(struct clocksource *cs)
368{
369 unsigned long flags;
370
371 spin_lock_irqsave(&watchdog_lock, flags);
372 if (cs != watchdog) {
373 if (cs->flags & CLOCK_SOURCE_MUST_VERIFY) {
374 /* cs is a watched clocksource. */
375 list_del_init(&cs->wd_list);
376 /* Check if the watchdog timer needs to be stopped. */
377 clocksource_stop_watchdog();
378 }
379 }
380 spin_unlock_irqrestore(&watchdog_lock, flags);
381}
382
383static int __clocksource_watchdog_kthread(void)
384{
385 struct clocksource *cs, *tmp;
386 unsigned long flags;
387 LIST_HEAD(unstable);
388 int select = 0;
389
390 spin_lock_irqsave(&watchdog_lock, flags);
391 list_for_each_entry_safe(cs, tmp, &watchdog_list, wd_list) {
392 if (cs->flags & CLOCK_SOURCE_UNSTABLE) {
393 list_del_init(&cs->wd_list);
394 list_add(&cs->wd_list, &unstable);
395 select = 1;
396 }
397 if (cs->flags & CLOCK_SOURCE_RESELECT) {
398 cs->flags &= ~CLOCK_SOURCE_RESELECT;
399 select = 1;
400 }
401 }
402 /* Check if the watchdog timer needs to be stopped. */
403 clocksource_stop_watchdog();
404 spin_unlock_irqrestore(&watchdog_lock, flags);
405
406 /* Needs to be done outside of watchdog lock */
407 list_for_each_entry_safe(cs, tmp, &unstable, wd_list) {
408 list_del_init(&cs->wd_list);
409 __clocksource_change_rating(cs, 0);
410 }
411 return select;
412}
413
414static int clocksource_watchdog_kthread(void *data)
415{
416 mutex_lock(&clocksource_mutex);
417 if (__clocksource_watchdog_kthread())
418 clocksource_select();
419 mutex_unlock(&clocksource_mutex);
420 return 0;
421}
422
423static bool clocksource_is_watchdog(struct clocksource *cs)
424{
425 return cs == watchdog;
426}
427
428#else /* CONFIG_CLOCKSOURCE_WATCHDOG */
429
430static void clocksource_enqueue_watchdog(struct clocksource *cs)
431{
432 if (cs->flags & CLOCK_SOURCE_IS_CONTINUOUS)
433 cs->flags |= CLOCK_SOURCE_VALID_FOR_HRES;
434}
435
436static void clocksource_select_watchdog(bool fallback) { }
437static inline void clocksource_dequeue_watchdog(struct clocksource *cs) { }
438static inline void clocksource_resume_watchdog(void) { }
439static inline int __clocksource_watchdog_kthread(void) { return 0; }
440static bool clocksource_is_watchdog(struct clocksource *cs) { return false; }
441void clocksource_mark_unstable(struct clocksource *cs) { }
442
443#endif /* CONFIG_CLOCKSOURCE_WATCHDOG */
444
445/**
446 * clocksource_suspend - suspend the clocksource(s)
447 */
448void clocksource_suspend(void)
449{
450 struct clocksource *cs;
451
452 list_for_each_entry_reverse(cs, &clocksource_list, list)
453 if (cs->suspend)
454 cs->suspend(cs);
455}
456
457/**
458 * clocksource_resume - resume the clocksource(s)
459 */
460void clocksource_resume(void)
461{
462 struct clocksource *cs;
463
464 list_for_each_entry(cs, &clocksource_list, list)
465 if (cs->resume)
466 cs->resume(cs);
467
468 clocksource_resume_watchdog();
469}
470
471/**
472 * clocksource_touch_watchdog - Update watchdog
473 *
474 * Update the watchdog after exception contexts such as kgdb so as not
475 * to incorrectly trip the watchdog. This might fail when the kernel
476 * was stopped in code which holds watchdog_lock.
477 */
478void clocksource_touch_watchdog(void)
479{
480 clocksource_resume_watchdog();
481}
482
483/**
484 * clocksource_max_adjustment- Returns max adjustment amount
485 * @cs: Pointer to clocksource
486 *
487 */
488static u32 clocksource_max_adjustment(struct clocksource *cs)
489{
490 u64 ret;
491 /*
492 * We won't try to correct for more than 11% adjustments (110,000 ppm),
493 */
494 ret = (u64)cs->mult * 11;
495 do_div(ret,100);
496 return (u32)ret;
497}
498
499/**
500 * clocks_calc_max_nsecs - Returns maximum nanoseconds that can be converted
501 * @mult: cycle to nanosecond multiplier
502 * @shift: cycle to nanosecond divisor (power of two)
503 * @maxadj: maximum adjustment value to mult (~11%)
504 * @mask: bitmask for two's complement subtraction of non 64 bit counters
505 * @max_cyc: maximum cycle value before potential overflow (does not include
506 * any safety margin)
507 *
508 * NOTE: This function includes a safety margin of 50%, in other words, we
509 * return half the number of nanoseconds the hardware counter can technically
510 * cover. This is done so that we can potentially detect problems caused by
511 * delayed timers or bad hardware, which might result in time intervals that
512 * are larger than what the math used can handle without overflows.
513 */
514u64 clocks_calc_max_nsecs(u32 mult, u32 shift, u32 maxadj, u64 mask, u64 *max_cyc)
515{
516 u64 max_nsecs, max_cycles;
517
518 /*
519 * Calculate the maximum number of cycles that we can pass to the
520 * cyc2ns() function without overflowing a 64-bit result.
521 */
522 max_cycles = ULLONG_MAX;
523 do_div(max_cycles, mult+maxadj);
524
525 /*
526 * The actual maximum number of cycles we can defer the clocksource is
527 * determined by the minimum of max_cycles and mask.
528 * Note: Here we subtract the maxadj to make sure we don't sleep for
529 * too long if there's a large negative adjustment.
530 */
531 max_cycles = min(max_cycles, mask);
532 max_nsecs = clocksource_cyc2ns(max_cycles, mult - maxadj, shift);
533
534 /* return the max_cycles value as well if requested */
535 if (max_cyc)
536 *max_cyc = max_cycles;
537
538 /* Return 50% of the actual maximum, so we can detect bad values */
539 max_nsecs >>= 1;
540
541 return max_nsecs;
542}
543
544/**
545 * clocksource_update_max_deferment - Updates the clocksource max_idle_ns & max_cycles
546 * @cs: Pointer to clocksource to be updated
547 *
548 */
549static inline void clocksource_update_max_deferment(struct clocksource *cs)
550{
551 cs->max_idle_ns = clocks_calc_max_nsecs(cs->mult, cs->shift,
552 cs->maxadj, cs->mask,
553 &cs->max_cycles);
554}
555
556#ifndef CONFIG_ARCH_USES_GETTIMEOFFSET
557
558static struct clocksource *clocksource_find_best(bool oneshot, bool skipcur)
559{
560 struct clocksource *cs;
561
562 if (!finished_booting || list_empty(&clocksource_list))
563 return NULL;
564
565 /*
566 * We pick the clocksource with the highest rating. If oneshot
567 * mode is active, we pick the highres valid clocksource with
568 * the best rating.
569 */
570 list_for_each_entry(cs, &clocksource_list, list) {
571 if (skipcur && cs == curr_clocksource)
572 continue;
573 if (oneshot && !(cs->flags & CLOCK_SOURCE_VALID_FOR_HRES))
574 continue;
575 return cs;
576 }
577 return NULL;
578}
579
580static void __clocksource_select(bool skipcur)
581{
582 bool oneshot = tick_oneshot_mode_active();
583 struct clocksource *best, *cs;
584
585 /* Find the best suitable clocksource */
586 best = clocksource_find_best(oneshot, skipcur);
587 if (!best)
588 return;
589
590 /* Check for the override clocksource. */
591 list_for_each_entry(cs, &clocksource_list, list) {
592 if (skipcur && cs == curr_clocksource)
593 continue;
594 if (strcmp(cs->name, override_name) != 0)
595 continue;
596 /*
597 * Check to make sure we don't switch to a non-highres
598 * capable clocksource if the tick code is in oneshot
599 * mode (highres or nohz)
600 */
601 if (!(cs->flags & CLOCK_SOURCE_VALID_FOR_HRES) && oneshot) {
602 /* Override clocksource cannot be used. */
603 pr_warn("Override clocksource %s is not HRT compatible - cannot switch while in HRT/NOHZ mode\n",
604 cs->name);
605 override_name[0] = 0;
606 } else
607 /* Override clocksource can be used. */
608 best = cs;
609 break;
610 }
611
612 if (curr_clocksource != best && !timekeeping_notify(best)) {
613 pr_info("Switched to clocksource %s\n", best->name);
614 curr_clocksource = best;
615 }
616}
617
618/**
619 * clocksource_select - Select the best clocksource available
620 *
621 * Private function. Must hold clocksource_mutex when called.
622 *
623 * Select the clocksource with the best rating, or the clocksource,
624 * which is selected by userspace override.
625 */
626static void clocksource_select(void)
627{
628 __clocksource_select(false);
629}
630
631static void clocksource_select_fallback(void)
632{
633 __clocksource_select(true);
634}
635
636#else /* !CONFIG_ARCH_USES_GETTIMEOFFSET */
637static inline void clocksource_select(void) { }
638static inline void clocksource_select_fallback(void) { }
639
640#endif
641
642/*
643 * clocksource_done_booting - Called near the end of core bootup
644 *
645 * Hack to avoid lots of clocksource churn at boot time.
646 * We use fs_initcall because we want this to start before
647 * device_initcall but after subsys_initcall.
648 */
649static int __init clocksource_done_booting(void)
650{
651 mutex_lock(&clocksource_mutex);
652 curr_clocksource = clocksource_default_clock();
653 finished_booting = 1;
654 /*
655 * Run the watchdog first to eliminate unstable clock sources
656 */
657 __clocksource_watchdog_kthread();
658 clocksource_select();
659 mutex_unlock(&clocksource_mutex);
660 return 0;
661}
662fs_initcall(clocksource_done_booting);
663
664/*
665 * Enqueue the clocksource sorted by rating
666 */
667static void clocksource_enqueue(struct clocksource *cs)
668{
669 struct list_head *entry = &clocksource_list;
670 struct clocksource *tmp;
671
672 list_for_each_entry(tmp, &clocksource_list, list)
673 /* Keep track of the place, where to insert */
674 if (tmp->rating >= cs->rating)
675 entry = &tmp->list;
676 list_add(&cs->list, entry);
677}
678
679/**
680 * __clocksource_update_freq_scale - Used update clocksource with new freq
681 * @cs: clocksource to be registered
682 * @scale: Scale factor multiplied against freq to get clocksource hz
683 * @freq: clocksource frequency (cycles per second) divided by scale
684 *
685 * This should only be called from the clocksource->enable() method.
686 *
687 * This *SHOULD NOT* be called directly! Please use the
688 * __clocksource_update_freq_hz() or __clocksource_update_freq_khz() helper
689 * functions.
690 */
691void __clocksource_update_freq_scale(struct clocksource *cs, u32 scale, u32 freq)
692{
693 u64 sec;
694
695 /*
696 * Default clocksources are *special* and self-define their mult/shift.
697 * But, you're not special, so you should specify a freq value.
698 */
699 if (freq) {
700 /*
701 * Calc the maximum number of seconds which we can run before
702 * wrapping around. For clocksources which have a mask > 32-bit
703 * we need to limit the max sleep time to have a good
704 * conversion precision. 10 minutes is still a reasonable
705 * amount. That results in a shift value of 24 for a
706 * clocksource with mask >= 40-bit and f >= 4GHz. That maps to
707 * ~ 0.06ppm granularity for NTP.
708 */
709 sec = cs->mask;
710 do_div(sec, freq);
711 do_div(sec, scale);
712 if (!sec)
713 sec = 1;
714 else if (sec > 600 && cs->mask > UINT_MAX)
715 sec = 600;
716
717 clocks_calc_mult_shift(&cs->mult, &cs->shift, freq,
718 NSEC_PER_SEC / scale, sec * scale);
719 }
720 /*
721 * Ensure clocksources that have large 'mult' values don't overflow
722 * when adjusted.
723 */
724 cs->maxadj = clocksource_max_adjustment(cs);
725 while (freq && ((cs->mult + cs->maxadj < cs->mult)
726 || (cs->mult - cs->maxadj > cs->mult))) {
727 cs->mult >>= 1;
728 cs->shift--;
729 cs->maxadj = clocksource_max_adjustment(cs);
730 }
731
732 /*
733 * Only warn for *special* clocksources that self-define
734 * their mult/shift values and don't specify a freq.
735 */
736 WARN_ONCE(cs->mult + cs->maxadj < cs->mult,
737 "timekeeping: Clocksource %s might overflow on 11%% adjustment\n",
738 cs->name);
739
740 clocksource_update_max_deferment(cs);
741
742 pr_info("%s: mask: 0x%llx max_cycles: 0x%llx, max_idle_ns: %lld ns\n",
743 cs->name, cs->mask, cs->max_cycles, cs->max_idle_ns);
744}
745EXPORT_SYMBOL_GPL(__clocksource_update_freq_scale);
746
747/**
748 * __clocksource_register_scale - Used to install new clocksources
749 * @cs: clocksource to be registered
750 * @scale: Scale factor multiplied against freq to get clocksource hz
751 * @freq: clocksource frequency (cycles per second) divided by scale
752 *
753 * Returns -EBUSY if registration fails, zero otherwise.
754 *
755 * This *SHOULD NOT* be called directly! Please use the
756 * clocksource_register_hz() or clocksource_register_khz helper functions.
757 */
758int __clocksource_register_scale(struct clocksource *cs, u32 scale, u32 freq)
759{
760
761 /* Initialize mult/shift and max_idle_ns */
762 __clocksource_update_freq_scale(cs, scale, freq);
763
764 /* Add clocksource to the clocksource list */
765 mutex_lock(&clocksource_mutex);
766 clocksource_enqueue(cs);
767 clocksource_enqueue_watchdog(cs);
768 clocksource_select();
769 clocksource_select_watchdog(false);
770 mutex_unlock(&clocksource_mutex);
771 return 0;
772}
773EXPORT_SYMBOL_GPL(__clocksource_register_scale);
774
775static void __clocksource_change_rating(struct clocksource *cs, int rating)
776{
777 list_del(&cs->list);
778 cs->rating = rating;
779 clocksource_enqueue(cs);
780}
781
782/**
783 * clocksource_change_rating - Change the rating of a registered clocksource
784 * @cs: clocksource to be changed
785 * @rating: new rating
786 */
787void clocksource_change_rating(struct clocksource *cs, int rating)
788{
789 mutex_lock(&clocksource_mutex);
790 __clocksource_change_rating(cs, rating);
791 clocksource_select();
792 clocksource_select_watchdog(false);
793 mutex_unlock(&clocksource_mutex);
794}
795EXPORT_SYMBOL(clocksource_change_rating);
796
797/*
798 * Unbind clocksource @cs. Called with clocksource_mutex held
799 */
800static int clocksource_unbind(struct clocksource *cs)
801{
802 if (clocksource_is_watchdog(cs)) {
803 /* Select and try to install a replacement watchdog. */
804 clocksource_select_watchdog(true);
805 if (clocksource_is_watchdog(cs))
806 return -EBUSY;
807 }
808
809 if (cs == curr_clocksource) {
810 /* Select and try to install a replacement clock source */
811 clocksource_select_fallback();
812 if (curr_clocksource == cs)
813 return -EBUSY;
814 }
815 clocksource_dequeue_watchdog(cs);
816 list_del_init(&cs->list);
817 return 0;
818}
819
820/**
821 * clocksource_unregister - remove a registered clocksource
822 * @cs: clocksource to be unregistered
823 */
824int clocksource_unregister(struct clocksource *cs)
825{
826 int ret = 0;
827
828 mutex_lock(&clocksource_mutex);
829 if (!list_empty(&cs->list))
830 ret = clocksource_unbind(cs);
831 mutex_unlock(&clocksource_mutex);
832 return ret;
833}
834EXPORT_SYMBOL(clocksource_unregister);
835
836#ifdef CONFIG_SYSFS
837/**
838 * sysfs_show_current_clocksources - sysfs interface for current clocksource
839 * @dev: unused
840 * @attr: unused
841 * @buf: char buffer to be filled with clocksource list
842 *
843 * Provides sysfs interface for listing current clocksource.
844 */
845static ssize_t
846sysfs_show_current_clocksources(struct device *dev,
847 struct device_attribute *attr, char *buf)
848{
849 ssize_t count = 0;
850
851 mutex_lock(&clocksource_mutex);
852 count = snprintf(buf, PAGE_SIZE, "%s\n", curr_clocksource->name);
853 mutex_unlock(&clocksource_mutex);
854
855 return count;
856}
857
858ssize_t sysfs_get_uname(const char *buf, char *dst, size_t cnt)
859{
860 size_t ret = cnt;
861
862 /* strings from sysfs write are not 0 terminated! */
863 if (!cnt || cnt >= CS_NAME_LEN)
864 return -EINVAL;
865
866 /* strip of \n: */
867 if (buf[cnt-1] == '\n')
868 cnt--;
869 if (cnt > 0)
870 memcpy(dst, buf, cnt);
871 dst[cnt] = 0;
872 return ret;
873}
874
875/**
876 * sysfs_override_clocksource - interface for manually overriding clocksource
877 * @dev: unused
878 * @attr: unused
879 * @buf: name of override clocksource
880 * @count: length of buffer
881 *
882 * Takes input from sysfs interface for manually overriding the default
883 * clocksource selection.
884 */
885static ssize_t sysfs_override_clocksource(struct device *dev,
886 struct device_attribute *attr,
887 const char *buf, size_t count)
888{
889 ssize_t ret;
890
891 mutex_lock(&clocksource_mutex);
892
893 ret = sysfs_get_uname(buf, override_name, count);
894 if (ret >= 0)
895 clocksource_select();
896
897 mutex_unlock(&clocksource_mutex);
898
899 return ret;
900}
901
902/**
903 * sysfs_unbind_current_clocksource - interface for manually unbinding clocksource
904 * @dev: unused
905 * @attr: unused
906 * @buf: unused
907 * @count: length of buffer
908 *
909 * Takes input from sysfs interface for manually unbinding a clocksource.
910 */
911static ssize_t sysfs_unbind_clocksource(struct device *dev,
912 struct device_attribute *attr,
913 const char *buf, size_t count)
914{
915 struct clocksource *cs;
916 char name[CS_NAME_LEN];
917 ssize_t ret;
918
919 ret = sysfs_get_uname(buf, name, count);
920 if (ret < 0)
921 return ret;
922
923 ret = -ENODEV;
924 mutex_lock(&clocksource_mutex);
925 list_for_each_entry(cs, &clocksource_list, list) {
926 if (strcmp(cs->name, name))
927 continue;
928 ret = clocksource_unbind(cs);
929 break;
930 }
931 mutex_unlock(&clocksource_mutex);
932
933 return ret ? ret : count;
934}
935
936/**
937 * sysfs_show_available_clocksources - sysfs interface for listing clocksource
938 * @dev: unused
939 * @attr: unused
940 * @buf: char buffer to be filled with clocksource list
941 *
942 * Provides sysfs interface for listing registered clocksources
943 */
944static ssize_t
945sysfs_show_available_clocksources(struct device *dev,
946 struct device_attribute *attr,
947 char *buf)
948{
949 struct clocksource *src;
950 ssize_t count = 0;
951
952 mutex_lock(&clocksource_mutex);
953 list_for_each_entry(src, &clocksource_list, list) {
954 /*
955 * Don't show non-HRES clocksource if the tick code is
956 * in one shot mode (highres=on or nohz=on)
957 */
958 if (!tick_oneshot_mode_active() ||
959 (src->flags & CLOCK_SOURCE_VALID_FOR_HRES))
960 count += snprintf(buf + count,
961 max((ssize_t)PAGE_SIZE - count, (ssize_t)0),
962 "%s ", src->name);
963 }
964 mutex_unlock(&clocksource_mutex);
965
966 count += snprintf(buf + count,
967 max((ssize_t)PAGE_SIZE - count, (ssize_t)0), "\n");
968
969 return count;
970}
971
972/*
973 * Sysfs setup bits:
974 */
975static DEVICE_ATTR(current_clocksource, 0644, sysfs_show_current_clocksources,
976 sysfs_override_clocksource);
977
978static DEVICE_ATTR(unbind_clocksource, 0200, NULL, sysfs_unbind_clocksource);
979
980static DEVICE_ATTR(available_clocksource, 0444,
981 sysfs_show_available_clocksources, NULL);
982
983static struct bus_type clocksource_subsys = {
984 .name = "clocksource",
985 .dev_name = "clocksource",
986};
987
988static struct device device_clocksource = {
989 .id = 0,
990 .bus = &clocksource_subsys,
991};
992
993static int __init init_clocksource_sysfs(void)
994{
995 int error = subsys_system_register(&clocksource_subsys, NULL);
996
997 if (!error)
998 error = device_register(&device_clocksource);
999 if (!error)
1000 error = device_create_file(
1001 &device_clocksource,
1002 &dev_attr_current_clocksource);
1003 if (!error)
1004 error = device_create_file(&device_clocksource,
1005 &dev_attr_unbind_clocksource);
1006 if (!error)
1007 error = device_create_file(
1008 &device_clocksource,
1009 &dev_attr_available_clocksource);
1010 return error;
1011}
1012
1013device_initcall(init_clocksource_sysfs);
1014#endif /* CONFIG_SYSFS */
1015
1016/**
1017 * boot_override_clocksource - boot clock override
1018 * @str: override name
1019 *
1020 * Takes a clocksource= boot argument and uses it
1021 * as the clocksource override name.
1022 */
1023static int __init boot_override_clocksource(char* str)
1024{
1025 mutex_lock(&clocksource_mutex);
1026 if (str)
1027 strlcpy(override_name, str, sizeof(override_name));
1028 mutex_unlock(&clocksource_mutex);
1029 return 1;
1030}
1031
1032__setup("clocksource=", boot_override_clocksource);
1033
1034/**
1035 * boot_override_clock - Compatibility layer for deprecated boot option
1036 * @str: override name
1037 *
1038 * DEPRECATED! Takes a clock= boot argument and uses it
1039 * as the clocksource override name
1040 */
1041static int __init boot_override_clock(char* str)
1042{
1043 if (!strcmp(str, "pmtmr")) {
1044 pr_warn("clock=pmtmr is deprecated - use clocksource=acpi_pm\n");
1045 return boot_override_clocksource("acpi_pm");
1046 }
1047 pr_warn("clock= boot option is deprecated - use clocksource=xyz\n");
1048 return boot_override_clocksource(str);
1049}
1050
1051__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);