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/*
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}
92EXPORT_SYMBOL_GPL(clocks_calc_mult_shift);
93
94/*[Clocksource internal variables]---------
95 * curr_clocksource:
96 * currently selected clocksource.
97 * clocksource_list:
98 * linked list with the registered clocksources
99 * clocksource_mutex:
100 * protects manipulations to curr_clocksource and the clocksource_list
101 * override_name:
102 * Name of the user-specified clocksource.
103 */
104static struct clocksource *curr_clocksource;
105static LIST_HEAD(clocksource_list);
106static DEFINE_MUTEX(clocksource_mutex);
107static char override_name[CS_NAME_LEN];
108static int finished_booting;
109
110#ifdef CONFIG_CLOCKSOURCE_WATCHDOG
111static void clocksource_watchdog_work(struct work_struct *work);
112static void clocksource_select(void);
113
114static LIST_HEAD(watchdog_list);
115static struct clocksource *watchdog;
116static struct timer_list watchdog_timer;
117static DECLARE_WORK(watchdog_work, clocksource_watchdog_work);
118static DEFINE_SPINLOCK(watchdog_lock);
119static int watchdog_running;
120static atomic_t watchdog_reset_pending;
121
122static void inline clocksource_watchdog_lock(unsigned long *flags)
123{
124 spin_lock_irqsave(&watchdog_lock, *flags);
125}
126
127static void inline clocksource_watchdog_unlock(unsigned long *flags)
128{
129 spin_unlock_irqrestore(&watchdog_lock, *flags);
130}
131
132static int clocksource_watchdog_kthread(void *data);
133static void __clocksource_change_rating(struct clocksource *cs, int rating);
134
135/*
136 * Interval: 0.5sec Threshold: 0.0625s
137 */
138#define WATCHDOG_INTERVAL (HZ >> 1)
139#define WATCHDOG_THRESHOLD (NSEC_PER_SEC >> 4)
140
141static void clocksource_watchdog_work(struct work_struct *work)
142{
143 /*
144 * If kthread_run fails the next watchdog scan over the
145 * watchdog_list will find the unstable clock again.
146 */
147 kthread_run(clocksource_watchdog_kthread, NULL, "kwatchdog");
148}
149
150static void __clocksource_unstable(struct clocksource *cs)
151{
152 cs->flags &= ~(CLOCK_SOURCE_VALID_FOR_HRES | CLOCK_SOURCE_WATCHDOG);
153 cs->flags |= CLOCK_SOURCE_UNSTABLE;
154
155 /*
156 * If the clocksource is registered clocksource_watchdog_kthread() will
157 * re-rate and re-select.
158 */
159 if (list_empty(&cs->list)) {
160 cs->rating = 0;
161 return;
162 }
163
164 if (cs->mark_unstable)
165 cs->mark_unstable(cs);
166
167 /* kick clocksource_watchdog_kthread() */
168 if (finished_booting)
169 schedule_work(&watchdog_work);
170}
171
172/**
173 * clocksource_mark_unstable - mark clocksource unstable via watchdog
174 * @cs: clocksource to be marked unstable
175 *
176 * This function is called by the x86 TSC code to mark clocksources as unstable;
177 * it defers demotion and re-selection to a kthread.
178 */
179void clocksource_mark_unstable(struct clocksource *cs)
180{
181 unsigned long flags;
182
183 spin_lock_irqsave(&watchdog_lock, flags);
184 if (!(cs->flags & CLOCK_SOURCE_UNSTABLE)) {
185 if (!list_empty(&cs->list) && list_empty(&cs->wd_list))
186 list_add(&cs->wd_list, &watchdog_list);
187 __clocksource_unstable(cs);
188 }
189 spin_unlock_irqrestore(&watchdog_lock, flags);
190}
191
192static void clocksource_watchdog(struct timer_list *unused)
193{
194 struct clocksource *cs;
195 u64 csnow, wdnow, cslast, wdlast, delta;
196 int64_t wd_nsec, cs_nsec;
197 int next_cpu, reset_pending;
198
199 spin_lock(&watchdog_lock);
200 if (!watchdog_running)
201 goto out;
202
203 reset_pending = atomic_read(&watchdog_reset_pending);
204
205 list_for_each_entry(cs, &watchdog_list, wd_list) {
206
207 /* Clocksource already marked unstable? */
208 if (cs->flags & CLOCK_SOURCE_UNSTABLE) {
209 if (finished_booting)
210 schedule_work(&watchdog_work);
211 continue;
212 }
213
214 local_irq_disable();
215 csnow = cs->read(cs);
216 wdnow = watchdog->read(watchdog);
217 local_irq_enable();
218
219 /* Clocksource initialized ? */
220 if (!(cs->flags & CLOCK_SOURCE_WATCHDOG) ||
221 atomic_read(&watchdog_reset_pending)) {
222 cs->flags |= CLOCK_SOURCE_WATCHDOG;
223 cs->wd_last = wdnow;
224 cs->cs_last = csnow;
225 continue;
226 }
227
228 delta = clocksource_delta(wdnow, cs->wd_last, watchdog->mask);
229 wd_nsec = clocksource_cyc2ns(delta, watchdog->mult,
230 watchdog->shift);
231
232 delta = clocksource_delta(csnow, cs->cs_last, cs->mask);
233 cs_nsec = clocksource_cyc2ns(delta, cs->mult, cs->shift);
234 wdlast = cs->wd_last; /* save these in case we print them */
235 cslast = cs->cs_last;
236 cs->cs_last = csnow;
237 cs->wd_last = wdnow;
238
239 if (atomic_read(&watchdog_reset_pending))
240 continue;
241
242 /* Check the deviation from the watchdog clocksource. */
243 if (abs(cs_nsec - wd_nsec) > WATCHDOG_THRESHOLD) {
244 pr_warn("timekeeping watchdog on CPU%d: Marking clocksource '%s' as unstable because the skew is too large:\n",
245 smp_processor_id(), cs->name);
246 pr_warn(" '%s' wd_now: %llx wd_last: %llx mask: %llx\n",
247 watchdog->name, wdnow, wdlast, watchdog->mask);
248 pr_warn(" '%s' cs_now: %llx cs_last: %llx mask: %llx\n",
249 cs->name, csnow, cslast, cs->mask);
250 __clocksource_unstable(cs);
251 continue;
252 }
253
254 if (cs == curr_clocksource && cs->tick_stable)
255 cs->tick_stable(cs);
256
257 if (!(cs->flags & CLOCK_SOURCE_VALID_FOR_HRES) &&
258 (cs->flags & CLOCK_SOURCE_IS_CONTINUOUS) &&
259 (watchdog->flags & CLOCK_SOURCE_IS_CONTINUOUS)) {
260 /* Mark it valid for high-res. */
261 cs->flags |= CLOCK_SOURCE_VALID_FOR_HRES;
262
263 /*
264 * clocksource_done_booting() will sort it if
265 * finished_booting is not set yet.
266 */
267 if (!finished_booting)
268 continue;
269
270 /*
271 * If this is not the current clocksource let
272 * the watchdog thread reselect it. Due to the
273 * change to high res this clocksource might
274 * be preferred now. If it is the current
275 * clocksource let the tick code know about
276 * that change.
277 */
278 if (cs != curr_clocksource) {
279 cs->flags |= CLOCK_SOURCE_RESELECT;
280 schedule_work(&watchdog_work);
281 } else {
282 tick_clock_notify();
283 }
284 }
285 }
286
287 /*
288 * We only clear the watchdog_reset_pending, when we did a
289 * full cycle through all clocksources.
290 */
291 if (reset_pending)
292 atomic_dec(&watchdog_reset_pending);
293
294 /*
295 * Cycle through CPUs to check if the CPUs stay synchronized
296 * to each other.
297 */
298 next_cpu = cpumask_next(raw_smp_processor_id(), cpu_online_mask);
299 if (next_cpu >= nr_cpu_ids)
300 next_cpu = cpumask_first(cpu_online_mask);
301 watchdog_timer.expires += WATCHDOG_INTERVAL;
302 add_timer_on(&watchdog_timer, next_cpu);
303out:
304 spin_unlock(&watchdog_lock);
305}
306
307static inline void clocksource_start_watchdog(void)
308{
309 if (watchdog_running || !watchdog || list_empty(&watchdog_list))
310 return;
311 timer_setup(&watchdog_timer, clocksource_watchdog, 0);
312 watchdog_timer.expires = jiffies + WATCHDOG_INTERVAL;
313 add_timer_on(&watchdog_timer, cpumask_first(cpu_online_mask));
314 watchdog_running = 1;
315}
316
317static inline void clocksource_stop_watchdog(void)
318{
319 if (!watchdog_running || (watchdog && !list_empty(&watchdog_list)))
320 return;
321 del_timer(&watchdog_timer);
322 watchdog_running = 0;
323}
324
325static inline void clocksource_reset_watchdog(void)
326{
327 struct clocksource *cs;
328
329 list_for_each_entry(cs, &watchdog_list, wd_list)
330 cs->flags &= ~CLOCK_SOURCE_WATCHDOG;
331}
332
333static void clocksource_resume_watchdog(void)
334{
335 atomic_inc(&watchdog_reset_pending);
336}
337
338static void clocksource_enqueue_watchdog(struct clocksource *cs)
339{
340 INIT_LIST_HEAD(&cs->wd_list);
341
342 if (cs->flags & CLOCK_SOURCE_MUST_VERIFY) {
343 /* cs is a clocksource to be watched. */
344 list_add(&cs->wd_list, &watchdog_list);
345 cs->flags &= ~CLOCK_SOURCE_WATCHDOG;
346 } else {
347 /* cs is a watchdog. */
348 if (cs->flags & CLOCK_SOURCE_IS_CONTINUOUS)
349 cs->flags |= CLOCK_SOURCE_VALID_FOR_HRES;
350 }
351}
352
353static void clocksource_select_watchdog(bool fallback)
354{
355 struct clocksource *cs, *old_wd;
356 unsigned long flags;
357
358 spin_lock_irqsave(&watchdog_lock, flags);
359 /* save current watchdog */
360 old_wd = watchdog;
361 if (fallback)
362 watchdog = NULL;
363
364 list_for_each_entry(cs, &clocksource_list, list) {
365 /* cs is a clocksource to be watched. */
366 if (cs->flags & CLOCK_SOURCE_MUST_VERIFY)
367 continue;
368
369 /* Skip current if we were requested for a fallback. */
370 if (fallback && cs == old_wd)
371 continue;
372
373 /* Pick the best watchdog. */
374 if (!watchdog || cs->rating > watchdog->rating)
375 watchdog = cs;
376 }
377 /* If we failed to find a fallback restore the old one. */
378 if (!watchdog)
379 watchdog = old_wd;
380
381 /* If we changed the watchdog we need to reset cycles. */
382 if (watchdog != old_wd)
383 clocksource_reset_watchdog();
384
385 /* Check if the watchdog timer needs to be started. */
386 clocksource_start_watchdog();
387 spin_unlock_irqrestore(&watchdog_lock, flags);
388}
389
390static void clocksource_dequeue_watchdog(struct clocksource *cs)
391{
392 if (cs != watchdog) {
393 if (cs->flags & CLOCK_SOURCE_MUST_VERIFY) {
394 /* cs is a watched clocksource. */
395 list_del_init(&cs->wd_list);
396 /* Check if the watchdog timer needs to be stopped. */
397 clocksource_stop_watchdog();
398 }
399 }
400}
401
402static int __clocksource_watchdog_kthread(void)
403{
404 struct clocksource *cs, *tmp;
405 unsigned long flags;
406 int select = 0;
407
408 spin_lock_irqsave(&watchdog_lock, flags);
409 list_for_each_entry_safe(cs, tmp, &watchdog_list, wd_list) {
410 if (cs->flags & CLOCK_SOURCE_UNSTABLE) {
411 list_del_init(&cs->wd_list);
412 __clocksource_change_rating(cs, 0);
413 select = 1;
414 }
415 if (cs->flags & CLOCK_SOURCE_RESELECT) {
416 cs->flags &= ~CLOCK_SOURCE_RESELECT;
417 select = 1;
418 }
419 }
420 /* Check if the watchdog timer needs to be stopped. */
421 clocksource_stop_watchdog();
422 spin_unlock_irqrestore(&watchdog_lock, flags);
423
424 return select;
425}
426
427static int clocksource_watchdog_kthread(void *data)
428{
429 mutex_lock(&clocksource_mutex);
430 if (__clocksource_watchdog_kthread())
431 clocksource_select();
432 mutex_unlock(&clocksource_mutex);
433 return 0;
434}
435
436static bool clocksource_is_watchdog(struct clocksource *cs)
437{
438 return cs == watchdog;
439}
440
441#else /* CONFIG_CLOCKSOURCE_WATCHDOG */
442
443static void clocksource_enqueue_watchdog(struct clocksource *cs)
444{
445 if (cs->flags & CLOCK_SOURCE_IS_CONTINUOUS)
446 cs->flags |= CLOCK_SOURCE_VALID_FOR_HRES;
447}
448
449static void clocksource_select_watchdog(bool fallback) { }
450static inline void clocksource_dequeue_watchdog(struct clocksource *cs) { }
451static inline void clocksource_resume_watchdog(void) { }
452static inline int __clocksource_watchdog_kthread(void) { return 0; }
453static bool clocksource_is_watchdog(struct clocksource *cs) { return false; }
454void clocksource_mark_unstable(struct clocksource *cs) { }
455
456static void inline clocksource_watchdog_lock(unsigned long *flags) { }
457static void inline clocksource_watchdog_unlock(unsigned long *flags) { }
458
459#endif /* CONFIG_CLOCKSOURCE_WATCHDOG */
460
461/**
462 * clocksource_suspend - suspend the clocksource(s)
463 */
464void clocksource_suspend(void)
465{
466 struct clocksource *cs;
467
468 list_for_each_entry_reverse(cs, &clocksource_list, list)
469 if (cs->suspend)
470 cs->suspend(cs);
471}
472
473/**
474 * clocksource_resume - resume the clocksource(s)
475 */
476void clocksource_resume(void)
477{
478 struct clocksource *cs;
479
480 list_for_each_entry(cs, &clocksource_list, list)
481 if (cs->resume)
482 cs->resume(cs);
483
484 clocksource_resume_watchdog();
485}
486
487/**
488 * clocksource_touch_watchdog - Update watchdog
489 *
490 * Update the watchdog after exception contexts such as kgdb so as not
491 * to incorrectly trip the watchdog. This might fail when the kernel
492 * was stopped in code which holds watchdog_lock.
493 */
494void clocksource_touch_watchdog(void)
495{
496 clocksource_resume_watchdog();
497}
498
499/**
500 * clocksource_max_adjustment- Returns max adjustment amount
501 * @cs: Pointer to clocksource
502 *
503 */
504static u32 clocksource_max_adjustment(struct clocksource *cs)
505{
506 u64 ret;
507 /*
508 * We won't try to correct for more than 11% adjustments (110,000 ppm),
509 */
510 ret = (u64)cs->mult * 11;
511 do_div(ret,100);
512 return (u32)ret;
513}
514
515/**
516 * clocks_calc_max_nsecs - Returns maximum nanoseconds that can be converted
517 * @mult: cycle to nanosecond multiplier
518 * @shift: cycle to nanosecond divisor (power of two)
519 * @maxadj: maximum adjustment value to mult (~11%)
520 * @mask: bitmask for two's complement subtraction of non 64 bit counters
521 * @max_cyc: maximum cycle value before potential overflow (does not include
522 * any safety margin)
523 *
524 * NOTE: This function includes a safety margin of 50%, in other words, we
525 * return half the number of nanoseconds the hardware counter can technically
526 * cover. This is done so that we can potentially detect problems caused by
527 * delayed timers or bad hardware, which might result in time intervals that
528 * are larger than what the math used can handle without overflows.
529 */
530u64 clocks_calc_max_nsecs(u32 mult, u32 shift, u32 maxadj, u64 mask, u64 *max_cyc)
531{
532 u64 max_nsecs, max_cycles;
533
534 /*
535 * Calculate the maximum number of cycles that we can pass to the
536 * cyc2ns() function without overflowing a 64-bit result.
537 */
538 max_cycles = ULLONG_MAX;
539 do_div(max_cycles, mult+maxadj);
540
541 /*
542 * The actual maximum number of cycles we can defer the clocksource is
543 * determined by the minimum of max_cycles and mask.
544 * Note: Here we subtract the maxadj to make sure we don't sleep for
545 * too long if there's a large negative adjustment.
546 */
547 max_cycles = min(max_cycles, mask);
548 max_nsecs = clocksource_cyc2ns(max_cycles, mult - maxadj, shift);
549
550 /* return the max_cycles value as well if requested */
551 if (max_cyc)
552 *max_cyc = max_cycles;
553
554 /* Return 50% of the actual maximum, so we can detect bad values */
555 max_nsecs >>= 1;
556
557 return max_nsecs;
558}
559
560/**
561 * clocksource_update_max_deferment - Updates the clocksource max_idle_ns & max_cycles
562 * @cs: Pointer to clocksource to be updated
563 *
564 */
565static inline void clocksource_update_max_deferment(struct clocksource *cs)
566{
567 cs->max_idle_ns = clocks_calc_max_nsecs(cs->mult, cs->shift,
568 cs->maxadj, cs->mask,
569 &cs->max_cycles);
570}
571
572#ifndef CONFIG_ARCH_USES_GETTIMEOFFSET
573
574static struct clocksource *clocksource_find_best(bool oneshot, bool skipcur)
575{
576 struct clocksource *cs;
577
578 if (!finished_booting || list_empty(&clocksource_list))
579 return NULL;
580
581 /*
582 * We pick the clocksource with the highest rating. If oneshot
583 * mode is active, we pick the highres valid clocksource with
584 * the best rating.
585 */
586 list_for_each_entry(cs, &clocksource_list, list) {
587 if (skipcur && cs == curr_clocksource)
588 continue;
589 if (oneshot && !(cs->flags & CLOCK_SOURCE_VALID_FOR_HRES))
590 continue;
591 return cs;
592 }
593 return NULL;
594}
595
596static void __clocksource_select(bool skipcur)
597{
598 bool oneshot = tick_oneshot_mode_active();
599 struct clocksource *best, *cs;
600
601 /* Find the best suitable clocksource */
602 best = clocksource_find_best(oneshot, skipcur);
603 if (!best)
604 return;
605
606 if (!strlen(override_name))
607 goto found;
608
609 /* Check for the override clocksource. */
610 list_for_each_entry(cs, &clocksource_list, list) {
611 if (skipcur && cs == curr_clocksource)
612 continue;
613 if (strcmp(cs->name, override_name) != 0)
614 continue;
615 /*
616 * Check to make sure we don't switch to a non-highres
617 * capable clocksource if the tick code is in oneshot
618 * mode (highres or nohz)
619 */
620 if (!(cs->flags & CLOCK_SOURCE_VALID_FOR_HRES) && oneshot) {
621 /* Override clocksource cannot be used. */
622 if (cs->flags & CLOCK_SOURCE_UNSTABLE) {
623 pr_warn("Override clocksource %s is unstable and not HRT compatible - cannot switch while in HRT/NOHZ mode\n",
624 cs->name);
625 override_name[0] = 0;
626 } else {
627 /*
628 * The override cannot be currently verified.
629 * Deferring to let the watchdog check.
630 */
631 pr_info("Override clocksource %s is not currently HRT compatible - deferring\n",
632 cs->name);
633 }
634 } else
635 /* Override clocksource can be used. */
636 best = cs;
637 break;
638 }
639
640found:
641 if (curr_clocksource != best && !timekeeping_notify(best)) {
642 pr_info("Switched to clocksource %s\n", best->name);
643 curr_clocksource = best;
644 }
645}
646
647/**
648 * clocksource_select - Select the best clocksource available
649 *
650 * Private function. Must hold clocksource_mutex when called.
651 *
652 * Select the clocksource with the best rating, or the clocksource,
653 * which is selected by userspace override.
654 */
655static void clocksource_select(void)
656{
657 __clocksource_select(false);
658}
659
660static void clocksource_select_fallback(void)
661{
662 __clocksource_select(true);
663}
664
665#else /* !CONFIG_ARCH_USES_GETTIMEOFFSET */
666static inline void clocksource_select(void) { }
667static inline void clocksource_select_fallback(void) { }
668
669#endif
670
671/*
672 * clocksource_done_booting - Called near the end of core bootup
673 *
674 * Hack to avoid lots of clocksource churn at boot time.
675 * We use fs_initcall because we want this to start before
676 * device_initcall but after subsys_initcall.
677 */
678static int __init clocksource_done_booting(void)
679{
680 mutex_lock(&clocksource_mutex);
681 curr_clocksource = clocksource_default_clock();
682 finished_booting = 1;
683 /*
684 * Run the watchdog first to eliminate unstable clock sources
685 */
686 __clocksource_watchdog_kthread();
687 clocksource_select();
688 mutex_unlock(&clocksource_mutex);
689 return 0;
690}
691fs_initcall(clocksource_done_booting);
692
693/*
694 * Enqueue the clocksource sorted by rating
695 */
696static void clocksource_enqueue(struct clocksource *cs)
697{
698 struct list_head *entry = &clocksource_list;
699 struct clocksource *tmp;
700
701 list_for_each_entry(tmp, &clocksource_list, list) {
702 /* Keep track of the place, where to insert */
703 if (tmp->rating < cs->rating)
704 break;
705 entry = &tmp->list;
706 }
707 list_add(&cs->list, entry);
708}
709
710/**
711 * __clocksource_update_freq_scale - Used update clocksource with new freq
712 * @cs: clocksource to be registered
713 * @scale: Scale factor multiplied against freq to get clocksource hz
714 * @freq: clocksource frequency (cycles per second) divided by scale
715 *
716 * This should only be called from the clocksource->enable() method.
717 *
718 * This *SHOULD NOT* be called directly! Please use the
719 * __clocksource_update_freq_hz() or __clocksource_update_freq_khz() helper
720 * functions.
721 */
722void __clocksource_update_freq_scale(struct clocksource *cs, u32 scale, u32 freq)
723{
724 u64 sec;
725
726 /*
727 * Default clocksources are *special* and self-define their mult/shift.
728 * But, you're not special, so you should specify a freq value.
729 */
730 if (freq) {
731 /*
732 * Calc the maximum number of seconds which we can run before
733 * wrapping around. For clocksources which have a mask > 32-bit
734 * we need to limit the max sleep time to have a good
735 * conversion precision. 10 minutes is still a reasonable
736 * amount. That results in a shift value of 24 for a
737 * clocksource with mask >= 40-bit and f >= 4GHz. That maps to
738 * ~ 0.06ppm granularity for NTP.
739 */
740 sec = cs->mask;
741 do_div(sec, freq);
742 do_div(sec, scale);
743 if (!sec)
744 sec = 1;
745 else if (sec > 600 && cs->mask > UINT_MAX)
746 sec = 600;
747
748 clocks_calc_mult_shift(&cs->mult, &cs->shift, freq,
749 NSEC_PER_SEC / scale, sec * scale);
750 }
751 /*
752 * Ensure clocksources that have large 'mult' values don't overflow
753 * when adjusted.
754 */
755 cs->maxadj = clocksource_max_adjustment(cs);
756 while (freq && ((cs->mult + cs->maxadj < cs->mult)
757 || (cs->mult - cs->maxadj > cs->mult))) {
758 cs->mult >>= 1;
759 cs->shift--;
760 cs->maxadj = clocksource_max_adjustment(cs);
761 }
762
763 /*
764 * Only warn for *special* clocksources that self-define
765 * their mult/shift values and don't specify a freq.
766 */
767 WARN_ONCE(cs->mult + cs->maxadj < cs->mult,
768 "timekeeping: Clocksource %s might overflow on 11%% adjustment\n",
769 cs->name);
770
771 clocksource_update_max_deferment(cs);
772
773 pr_info("%s: mask: 0x%llx max_cycles: 0x%llx, max_idle_ns: %lld ns\n",
774 cs->name, cs->mask, cs->max_cycles, cs->max_idle_ns);
775}
776EXPORT_SYMBOL_GPL(__clocksource_update_freq_scale);
777
778/**
779 * __clocksource_register_scale - Used to install new clocksources
780 * @cs: clocksource to be registered
781 * @scale: Scale factor multiplied against freq to get clocksource hz
782 * @freq: clocksource frequency (cycles per second) divided by scale
783 *
784 * Returns -EBUSY if registration fails, zero otherwise.
785 *
786 * This *SHOULD NOT* be called directly! Please use the
787 * clocksource_register_hz() or clocksource_register_khz helper functions.
788 */
789int __clocksource_register_scale(struct clocksource *cs, u32 scale, u32 freq)
790{
791 unsigned long flags;
792
793 /* Initialize mult/shift and max_idle_ns */
794 __clocksource_update_freq_scale(cs, scale, freq);
795
796 /* Add clocksource to the clocksource list */
797 mutex_lock(&clocksource_mutex);
798
799 clocksource_watchdog_lock(&flags);
800 clocksource_enqueue(cs);
801 clocksource_enqueue_watchdog(cs);
802 clocksource_watchdog_unlock(&flags);
803
804 clocksource_select();
805 clocksource_select_watchdog(false);
806 mutex_unlock(&clocksource_mutex);
807 return 0;
808}
809EXPORT_SYMBOL_GPL(__clocksource_register_scale);
810
811static void __clocksource_change_rating(struct clocksource *cs, int rating)
812{
813 list_del(&cs->list);
814 cs->rating = rating;
815 clocksource_enqueue(cs);
816}
817
818/**
819 * clocksource_change_rating - Change the rating of a registered clocksource
820 * @cs: clocksource to be changed
821 * @rating: new rating
822 */
823void clocksource_change_rating(struct clocksource *cs, int rating)
824{
825 unsigned long flags;
826
827 mutex_lock(&clocksource_mutex);
828 clocksource_watchdog_lock(&flags);
829 __clocksource_change_rating(cs, rating);
830 clocksource_watchdog_unlock(&flags);
831
832 clocksource_select();
833 clocksource_select_watchdog(false);
834 mutex_unlock(&clocksource_mutex);
835}
836EXPORT_SYMBOL(clocksource_change_rating);
837
838/*
839 * Unbind clocksource @cs. Called with clocksource_mutex held
840 */
841static int clocksource_unbind(struct clocksource *cs)
842{
843 unsigned long flags;
844
845 if (clocksource_is_watchdog(cs)) {
846 /* Select and try to install a replacement watchdog. */
847 clocksource_select_watchdog(true);
848 if (clocksource_is_watchdog(cs))
849 return -EBUSY;
850 }
851
852 if (cs == curr_clocksource) {
853 /* Select and try to install a replacement clock source */
854 clocksource_select_fallback();
855 if (curr_clocksource == cs)
856 return -EBUSY;
857 }
858
859 clocksource_watchdog_lock(&flags);
860 clocksource_dequeue_watchdog(cs);
861 list_del_init(&cs->list);
862 clocksource_watchdog_unlock(&flags);
863
864 return 0;
865}
866
867/**
868 * clocksource_unregister - remove a registered clocksource
869 * @cs: clocksource to be unregistered
870 */
871int clocksource_unregister(struct clocksource *cs)
872{
873 int ret = 0;
874
875 mutex_lock(&clocksource_mutex);
876 if (!list_empty(&cs->list))
877 ret = clocksource_unbind(cs);
878 mutex_unlock(&clocksource_mutex);
879 return ret;
880}
881EXPORT_SYMBOL(clocksource_unregister);
882
883#ifdef CONFIG_SYSFS
884/**
885 * current_clocksource_show - sysfs interface for current clocksource
886 * @dev: unused
887 * @attr: unused
888 * @buf: char buffer to be filled with clocksource list
889 *
890 * Provides sysfs interface for listing current clocksource.
891 */
892static ssize_t current_clocksource_show(struct device *dev,
893 struct device_attribute *attr,
894 char *buf)
895{
896 ssize_t count = 0;
897
898 mutex_lock(&clocksource_mutex);
899 count = snprintf(buf, PAGE_SIZE, "%s\n", curr_clocksource->name);
900 mutex_unlock(&clocksource_mutex);
901
902 return count;
903}
904
905ssize_t sysfs_get_uname(const char *buf, char *dst, size_t cnt)
906{
907 size_t ret = cnt;
908
909 /* strings from sysfs write are not 0 terminated! */
910 if (!cnt || cnt >= CS_NAME_LEN)
911 return -EINVAL;
912
913 /* strip of \n: */
914 if (buf[cnt-1] == '\n')
915 cnt--;
916 if (cnt > 0)
917 memcpy(dst, buf, cnt);
918 dst[cnt] = 0;
919 return ret;
920}
921
922/**
923 * current_clocksource_store - interface for manually overriding clocksource
924 * @dev: unused
925 * @attr: unused
926 * @buf: name of override clocksource
927 * @count: length of buffer
928 *
929 * Takes input from sysfs interface for manually overriding the default
930 * clocksource selection.
931 */
932static ssize_t current_clocksource_store(struct device *dev,
933 struct device_attribute *attr,
934 const char *buf, size_t count)
935{
936 ssize_t ret;
937
938 mutex_lock(&clocksource_mutex);
939
940 ret = sysfs_get_uname(buf, override_name, count);
941 if (ret >= 0)
942 clocksource_select();
943
944 mutex_unlock(&clocksource_mutex);
945
946 return ret;
947}
948static DEVICE_ATTR_RW(current_clocksource);
949
950/**
951 * unbind_clocksource_store - interface for manually unbinding clocksource
952 * @dev: unused
953 * @attr: unused
954 * @buf: unused
955 * @count: length of buffer
956 *
957 * Takes input from sysfs interface for manually unbinding a clocksource.
958 */
959static ssize_t unbind_clocksource_store(struct device *dev,
960 struct device_attribute *attr,
961 const char *buf, size_t count)
962{
963 struct clocksource *cs;
964 char name[CS_NAME_LEN];
965 ssize_t ret;
966
967 ret = sysfs_get_uname(buf, name, count);
968 if (ret < 0)
969 return ret;
970
971 ret = -ENODEV;
972 mutex_lock(&clocksource_mutex);
973 list_for_each_entry(cs, &clocksource_list, list) {
974 if (strcmp(cs->name, name))
975 continue;
976 ret = clocksource_unbind(cs);
977 break;
978 }
979 mutex_unlock(&clocksource_mutex);
980
981 return ret ? ret : count;
982}
983static DEVICE_ATTR_WO(unbind_clocksource);
984
985/**
986 * available_clocksource_show - sysfs interface for listing clocksource
987 * @dev: unused
988 * @attr: unused
989 * @buf: char buffer to be filled with clocksource list
990 *
991 * Provides sysfs interface for listing registered clocksources
992 */
993static ssize_t available_clocksource_show(struct device *dev,
994 struct device_attribute *attr,
995 char *buf)
996{
997 struct clocksource *src;
998 ssize_t count = 0;
999
1000 mutex_lock(&clocksource_mutex);
1001 list_for_each_entry(src, &clocksource_list, list) {
1002 /*
1003 * Don't show non-HRES clocksource if the tick code is
1004 * in one shot mode (highres=on or nohz=on)
1005 */
1006 if (!tick_oneshot_mode_active() ||
1007 (src->flags & CLOCK_SOURCE_VALID_FOR_HRES))
1008 count += snprintf(buf + count,
1009 max((ssize_t)PAGE_SIZE - count, (ssize_t)0),
1010 "%s ", src->name);
1011 }
1012 mutex_unlock(&clocksource_mutex);
1013
1014 count += snprintf(buf + count,
1015 max((ssize_t)PAGE_SIZE - count, (ssize_t)0), "\n");
1016
1017 return count;
1018}
1019static DEVICE_ATTR_RO(available_clocksource);
1020
1021static struct attribute *clocksource_attrs[] = {
1022 &dev_attr_current_clocksource.attr,
1023 &dev_attr_unbind_clocksource.attr,
1024 &dev_attr_available_clocksource.attr,
1025 NULL
1026};
1027ATTRIBUTE_GROUPS(clocksource);
1028
1029static struct bus_type clocksource_subsys = {
1030 .name = "clocksource",
1031 .dev_name = "clocksource",
1032};
1033
1034static struct device device_clocksource = {
1035 .id = 0,
1036 .bus = &clocksource_subsys,
1037 .groups = clocksource_groups,
1038};
1039
1040static int __init init_clocksource_sysfs(void)
1041{
1042 int error = subsys_system_register(&clocksource_subsys, NULL);
1043
1044 if (!error)
1045 error = device_register(&device_clocksource);
1046
1047 return error;
1048}
1049
1050device_initcall(init_clocksource_sysfs);
1051#endif /* CONFIG_SYSFS */
1052
1053/**
1054 * boot_override_clocksource - boot clock override
1055 * @str: override name
1056 *
1057 * Takes a clocksource= boot argument and uses it
1058 * as the clocksource override name.
1059 */
1060static int __init boot_override_clocksource(char* str)
1061{
1062 mutex_lock(&clocksource_mutex);
1063 if (str)
1064 strlcpy(override_name, str, sizeof(override_name));
1065 mutex_unlock(&clocksource_mutex);
1066 return 1;
1067}
1068
1069__setup("clocksource=", boot_override_clocksource);
1070
1071/**
1072 * boot_override_clock - Compatibility layer for deprecated boot option
1073 * @str: override name
1074 *
1075 * DEPRECATED! Takes a clock= boot argument and uses it
1076 * as the clocksource override name
1077 */
1078static int __init boot_override_clock(char* str)
1079{
1080 if (!strcmp(str, "pmtmr")) {
1081 pr_warn("clock=pmtmr is deprecated - use clocksource=acpi_pm\n");
1082 return boot_override_clocksource("acpi_pm");
1083 }
1084 pr_warn("clock= boot option is deprecated - use clocksource=xyz\n");
1085 return boot_override_clocksource(str);
1086}
1087
1088__setup("clock=", boot_override_clock);