2 * ----------------------------------------------------------------------------
3 * "THE BEER-WARE LICENSE" (Revision 42):
4 * <phk@FreeBSD.ORG> wrote this file. As long as you retain this notice you
5 * can do whatever you want with this stuff. If we meet some day, and you think
6 * this stuff is worth it, you can buy me a beer in return. Poul-Henning Kamp
7 * ----------------------------------------------------------------------------
9 * Copyright (c) 2011, 2015, 2016 The FreeBSD Foundation
10 * All rights reserved.
12 * Portions of this software were developed by Julien Ridoux at the University
13 * of Melbourne under sponsorship from the FreeBSD Foundation.
15 * Portions of this software were developed by Konstantin Belousov
16 * under sponsorship from the FreeBSD Foundation.
19 #include <sys/cdefs.h>
20 __FBSDID("$FreeBSD$");
22 #include "opt_compat.h"
24 #include "opt_ffclock.h"
26 #include <sys/param.h>
27 #include <sys/kernel.h>
28 #include <sys/limits.h>
30 #include <sys/mutex.h>
33 #include <sys/sleepqueue.h>
34 #include <sys/sysctl.h>
35 #include <sys/syslog.h>
36 #include <sys/systm.h>
37 #include <sys/timeffc.h>
38 #include <sys/timepps.h>
39 #include <sys/timetc.h>
40 #include <sys/timex.h>
44 * A large step happens on boot. This constant detects such steps.
45 * It is relatively small so that ntp_update_second gets called enough
46 * in the typical 'missed a couple of seconds' case, but doesn't loop
47 * forever when the time step is large.
49 #define LARGE_STEP 200
52 * Implement a dummy timecounter which we can use until we get a real one
53 * in the air. This allows the console and other early stuff to use
58 dummy_get_timecount(struct timecounter *tc)
65 static struct timecounter dummy_timecounter = {
66 dummy_get_timecount, 0, ~0u, 1000000, "dummy", -1000000
70 /* These fields must be initialized by the driver. */
71 struct timecounter *th_counter;
72 int64_t th_adjustment;
74 u_int th_offset_count;
75 struct bintime th_offset;
76 struct bintime th_bintime;
77 struct timeval th_microtime;
78 struct timespec th_nanotime;
79 struct bintime th_boottime;
80 /* Fields not to be copied in tc_windup start with th_generation. */
82 struct timehands *th_next;
85 static struct timehands th0;
86 static struct timehands th1 = {
89 static struct timehands th0 = {
90 .th_counter = &dummy_timecounter,
91 .th_scale = (uint64_t)-1 / 1000000,
92 .th_offset = { .sec = 1 },
97 static struct timehands *volatile timehands = &th0;
98 struct timecounter *timecounter = &dummy_timecounter;
99 static struct timecounter *timecounters = &dummy_timecounter;
101 int tc_min_ticktock_freq = 1;
103 volatile time_t time_second = 1;
104 volatile time_t time_uptime = 1;
106 static int sysctl_kern_boottime(SYSCTL_HANDLER_ARGS);
107 SYSCTL_PROC(_kern, KERN_BOOTTIME, boottime, CTLTYPE_STRUCT|CTLFLAG_RD,
108 NULL, 0, sysctl_kern_boottime, "S,timeval", "System boottime");
110 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW, 0, "");
111 static SYSCTL_NODE(_kern_timecounter, OID_AUTO, tc, CTLFLAG_RW, 0, "");
113 static int timestepwarnings;
114 SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RW,
115 ×tepwarnings, 0, "Log time steps");
117 struct bintime bt_timethreshold;
118 struct bintime bt_tickthreshold;
119 sbintime_t sbt_timethreshold;
120 sbintime_t sbt_tickthreshold;
121 struct bintime tc_tick_bt;
122 sbintime_t tc_tick_sbt;
124 int tc_timepercentage = TC_DEFAULTPERC;
125 static int sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS);
126 SYSCTL_PROC(_kern_timecounter, OID_AUTO, alloweddeviation,
127 CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, 0, 0,
128 sysctl_kern_timecounter_adjprecision, "I",
129 "Allowed time interval deviation in percents");
131 volatile int rtc_generation = 1;
133 static int tc_chosen; /* Non-zero if a specific tc was chosen via sysctl. */
135 static void tc_windup(struct bintime *new_boottimebin);
136 static void cpu_tick_calibrate(int);
138 void dtrace_getnanotime(struct timespec *tsp);
141 sysctl_kern_boottime(SYSCTL_HANDLER_ARGS)
143 struct timeval boottime;
145 getboottime(&boottime);
151 if (req->flags & SCTL_MASK32) {
152 tv[0] = boottime.tv_sec;
153 tv[1] = boottime.tv_usec;
154 return (SYSCTL_OUT(req, tv, sizeof(tv)));
158 return (SYSCTL_OUT(req, &boottime, sizeof(boottime)));
162 sysctl_kern_timecounter_get(SYSCTL_HANDLER_ARGS)
165 struct timecounter *tc = arg1;
167 ncount = tc->tc_get_timecount(tc);
168 return (sysctl_handle_int(oidp, &ncount, 0, req));
172 sysctl_kern_timecounter_freq(SYSCTL_HANDLER_ARGS)
175 struct timecounter *tc = arg1;
177 freq = tc->tc_frequency;
178 return (sysctl_handle_64(oidp, &freq, 0, req));
182 * Return the difference between the timehands' counter value now and what
183 * was when we copied it to the timehands' offset_count.
185 static __inline u_int
186 tc_delta(struct timehands *th)
188 struct timecounter *tc;
191 return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
192 tc->tc_counter_mask);
196 * Functions for reading the time. We have to loop until we are sure that
197 * the timehands that we operated on was not updated under our feet. See
198 * the comment in <sys/time.h> for a description of these 12 functions.
203 fbclock_binuptime(struct bintime *bt)
205 struct timehands *th;
210 gen = atomic_load_acq_int(&th->th_generation);
212 bintime_addx(bt, th->th_scale * tc_delta(th));
213 atomic_thread_fence_acq();
214 } while (gen == 0 || gen != th->th_generation);
218 fbclock_nanouptime(struct timespec *tsp)
222 fbclock_binuptime(&bt);
223 bintime2timespec(&bt, tsp);
227 fbclock_microuptime(struct timeval *tvp)
231 fbclock_binuptime(&bt);
232 bintime2timeval(&bt, tvp);
236 fbclock_bintime(struct bintime *bt)
238 struct timehands *th;
243 gen = atomic_load_acq_int(&th->th_generation);
244 *bt = th->th_bintime;
245 bintime_addx(bt, th->th_scale * tc_delta(th));
246 atomic_thread_fence_acq();
247 } while (gen == 0 || gen != th->th_generation);
251 fbclock_nanotime(struct timespec *tsp)
255 fbclock_bintime(&bt);
256 bintime2timespec(&bt, tsp);
260 fbclock_microtime(struct timeval *tvp)
264 fbclock_bintime(&bt);
265 bintime2timeval(&bt, tvp);
269 fbclock_getbinuptime(struct bintime *bt)
271 struct timehands *th;
276 gen = atomic_load_acq_int(&th->th_generation);
278 atomic_thread_fence_acq();
279 } while (gen == 0 || gen != th->th_generation);
283 fbclock_getnanouptime(struct timespec *tsp)
285 struct timehands *th;
290 gen = atomic_load_acq_int(&th->th_generation);
291 bintime2timespec(&th->th_offset, tsp);
292 atomic_thread_fence_acq();
293 } while (gen == 0 || gen != th->th_generation);
297 fbclock_getmicrouptime(struct timeval *tvp)
299 struct timehands *th;
304 gen = atomic_load_acq_int(&th->th_generation);
305 bintime2timeval(&th->th_offset, tvp);
306 atomic_thread_fence_acq();
307 } while (gen == 0 || gen != th->th_generation);
311 fbclock_getbintime(struct bintime *bt)
313 struct timehands *th;
318 gen = atomic_load_acq_int(&th->th_generation);
319 *bt = th->th_bintime;
320 atomic_thread_fence_acq();
321 } while (gen == 0 || gen != th->th_generation);
325 fbclock_getnanotime(struct timespec *tsp)
327 struct timehands *th;
332 gen = atomic_load_acq_int(&th->th_generation);
333 *tsp = th->th_nanotime;
334 atomic_thread_fence_acq();
335 } while (gen == 0 || gen != th->th_generation);
339 fbclock_getmicrotime(struct timeval *tvp)
341 struct timehands *th;
346 gen = atomic_load_acq_int(&th->th_generation);
347 *tvp = th->th_microtime;
348 atomic_thread_fence_acq();
349 } while (gen == 0 || gen != th->th_generation);
353 binuptime(struct bintime *bt)
355 struct timehands *th;
360 gen = atomic_load_acq_int(&th->th_generation);
362 bintime_addx(bt, th->th_scale * tc_delta(th));
363 atomic_thread_fence_acq();
364 } while (gen == 0 || gen != th->th_generation);
368 nanouptime(struct timespec *tsp)
373 bintime2timespec(&bt, tsp);
377 microuptime(struct timeval *tvp)
382 bintime2timeval(&bt, tvp);
386 bintime(struct bintime *bt)
388 struct timehands *th;
393 gen = atomic_load_acq_int(&th->th_generation);
394 *bt = th->th_bintime;
395 bintime_addx(bt, th->th_scale * tc_delta(th));
396 atomic_thread_fence_acq();
397 } while (gen == 0 || gen != th->th_generation);
401 nanotime(struct timespec *tsp)
406 bintime2timespec(&bt, tsp);
410 microtime(struct timeval *tvp)
415 bintime2timeval(&bt, tvp);
419 getbinuptime(struct bintime *bt)
421 struct timehands *th;
426 gen = atomic_load_acq_int(&th->th_generation);
428 atomic_thread_fence_acq();
429 } while (gen == 0 || gen != th->th_generation);
433 getnanouptime(struct timespec *tsp)
435 struct timehands *th;
440 gen = atomic_load_acq_int(&th->th_generation);
441 bintime2timespec(&th->th_offset, tsp);
442 atomic_thread_fence_acq();
443 } while (gen == 0 || gen != th->th_generation);
447 getmicrouptime(struct timeval *tvp)
449 struct timehands *th;
454 gen = atomic_load_acq_int(&th->th_generation);
455 bintime2timeval(&th->th_offset, tvp);
456 atomic_thread_fence_acq();
457 } while (gen == 0 || gen != th->th_generation);
461 getbintime(struct bintime *bt)
463 struct timehands *th;
468 gen = atomic_load_acq_int(&th->th_generation);
469 *bt = th->th_bintime;
470 atomic_thread_fence_acq();
471 } while (gen == 0 || gen != th->th_generation);
475 getnanotime(struct timespec *tsp)
477 struct timehands *th;
482 gen = atomic_load_acq_int(&th->th_generation);
483 *tsp = th->th_nanotime;
484 atomic_thread_fence_acq();
485 } while (gen == 0 || gen != th->th_generation);
489 getmicrotime(struct timeval *tvp)
491 struct timehands *th;
496 gen = atomic_load_acq_int(&th->th_generation);
497 *tvp = th->th_microtime;
498 atomic_thread_fence_acq();
499 } while (gen == 0 || gen != th->th_generation);
504 getboottime(struct timeval *boottime)
506 struct bintime boottimebin;
508 getboottimebin(&boottimebin);
509 bintime2timeval(&boottimebin, boottime);
513 getboottimebin(struct bintime *boottimebin)
515 struct timehands *th;
520 gen = atomic_load_acq_int(&th->th_generation);
521 *boottimebin = th->th_boottime;
522 atomic_thread_fence_acq();
523 } while (gen == 0 || gen != th->th_generation);
528 * Support for feed-forward synchronization algorithms. This is heavily inspired
529 * by the timehands mechanism but kept independent from it. *_windup() functions
530 * have some connection to avoid accessing the timecounter hardware more than
534 /* Feed-forward clock estimates kept updated by the synchronization daemon. */
535 struct ffclock_estimate ffclock_estimate;
536 struct bintime ffclock_boottime; /* Feed-forward boot time estimate. */
537 uint32_t ffclock_status; /* Feed-forward clock status. */
538 int8_t ffclock_updated; /* New estimates are available. */
539 struct mtx ffclock_mtx; /* Mutex on ffclock_estimate. */
542 struct ffclock_estimate cest;
543 struct bintime tick_time;
544 struct bintime tick_time_lerp;
545 ffcounter tick_ffcount;
546 uint64_t period_lerp;
547 volatile uint8_t gen;
548 struct fftimehands *next;
551 #define NUM_ELEMENTS(x) (sizeof(x) / sizeof(*x))
553 static struct fftimehands ffth[10];
554 static struct fftimehands *volatile fftimehands = ffth;
559 struct fftimehands *cur;
560 struct fftimehands *last;
562 memset(ffth, 0, sizeof(ffth));
564 last = ffth + NUM_ELEMENTS(ffth) - 1;
565 for (cur = ffth; cur < last; cur++)
570 ffclock_status = FFCLOCK_STA_UNSYNC;
571 mtx_init(&ffclock_mtx, "ffclock lock", NULL, MTX_DEF);
575 * Reset the feed-forward clock estimates. Called from inittodr() to get things
576 * kick started and uses the timecounter nominal frequency as a first period
577 * estimate. Note: this function may be called several time just after boot.
578 * Note: this is the only function that sets the value of boot time for the
579 * monotonic (i.e. uptime) version of the feed-forward clock.
582 ffclock_reset_clock(struct timespec *ts)
584 struct timecounter *tc;
585 struct ffclock_estimate cest;
587 tc = timehands->th_counter;
588 memset(&cest, 0, sizeof(struct ffclock_estimate));
590 timespec2bintime(ts, &ffclock_boottime);
591 timespec2bintime(ts, &(cest.update_time));
592 ffclock_read_counter(&cest.update_ffcount);
593 cest.leapsec_next = 0;
594 cest.period = ((1ULL << 63) / tc->tc_frequency) << 1;
597 cest.status = FFCLOCK_STA_UNSYNC;
598 cest.leapsec_total = 0;
601 mtx_lock(&ffclock_mtx);
602 bcopy(&cest, &ffclock_estimate, sizeof(struct ffclock_estimate));
603 ffclock_updated = INT8_MAX;
604 mtx_unlock(&ffclock_mtx);
606 printf("ffclock reset: %s (%llu Hz), time = %ld.%09lu\n", tc->tc_name,
607 (unsigned long long)tc->tc_frequency, (long)ts->tv_sec,
608 (unsigned long)ts->tv_nsec);
612 * Sub-routine to convert a time interval measured in RAW counter units to time
613 * in seconds stored in bintime format.
614 * NOTE: bintime_mul requires u_int, but the value of the ffcounter may be
615 * larger than the max value of u_int (on 32 bit architecture). Loop to consume
619 ffclock_convert_delta(ffcounter ffdelta, uint64_t period, struct bintime *bt)
622 ffcounter delta, delta_max;
624 delta_max = (1ULL << (8 * sizeof(unsigned int))) - 1;
627 if (ffdelta > delta_max)
633 bintime_mul(&bt2, (unsigned int)delta);
634 bintime_add(bt, &bt2);
636 } while (ffdelta > 0);
640 * Update the fftimehands.
641 * Push the tick ffcount and time(s) forward based on current clock estimate.
642 * The conversion from ffcounter to bintime relies on the difference clock
643 * principle, whose accuracy relies on computing small time intervals. If a new
644 * clock estimate has been passed by the synchronisation daemon, make it
645 * current, and compute the linear interpolation for monotonic time if needed.
648 ffclock_windup(unsigned int delta)
650 struct ffclock_estimate *cest;
651 struct fftimehands *ffth;
652 struct bintime bt, gap_lerp;
655 unsigned int polling;
656 uint8_t forward_jump, ogen;
659 * Pick the next timehand, copy current ffclock estimates and move tick
660 * times and counter forward.
663 ffth = fftimehands->next;
667 bcopy(&fftimehands->cest, cest, sizeof(struct ffclock_estimate));
668 ffdelta = (ffcounter)delta;
669 ffth->period_lerp = fftimehands->period_lerp;
671 ffth->tick_time = fftimehands->tick_time;
672 ffclock_convert_delta(ffdelta, cest->period, &bt);
673 bintime_add(&ffth->tick_time, &bt);
675 ffth->tick_time_lerp = fftimehands->tick_time_lerp;
676 ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt);
677 bintime_add(&ffth->tick_time_lerp, &bt);
679 ffth->tick_ffcount = fftimehands->tick_ffcount + ffdelta;
682 * Assess the status of the clock, if the last update is too old, it is
683 * likely the synchronisation daemon is dead and the clock is free
686 if (ffclock_updated == 0) {
687 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
688 ffclock_convert_delta(ffdelta, cest->period, &bt);
689 if (bt.sec > 2 * FFCLOCK_SKM_SCALE)
690 ffclock_status |= FFCLOCK_STA_UNSYNC;
694 * If available, grab updated clock estimates and make them current.
695 * Recompute time at this tick using the updated estimates. The clock
696 * estimates passed the feed-forward synchronisation daemon may result
697 * in time conversion that is not monotonically increasing (just after
698 * the update). time_lerp is a particular linear interpolation over the
699 * synchronisation algo polling period that ensures monotonicity for the
700 * clock ids requesting it.
702 if (ffclock_updated > 0) {
703 bcopy(&ffclock_estimate, cest, sizeof(struct ffclock_estimate));
704 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
705 ffth->tick_time = cest->update_time;
706 ffclock_convert_delta(ffdelta, cest->period, &bt);
707 bintime_add(&ffth->tick_time, &bt);
709 /* ffclock_reset sets ffclock_updated to INT8_MAX */
710 if (ffclock_updated == INT8_MAX)
711 ffth->tick_time_lerp = ffth->tick_time;
713 if (bintime_cmp(&ffth->tick_time, &ffth->tick_time_lerp, >))
718 bintime_clear(&gap_lerp);
720 gap_lerp = ffth->tick_time;
721 bintime_sub(&gap_lerp, &ffth->tick_time_lerp);
723 gap_lerp = ffth->tick_time_lerp;
724 bintime_sub(&gap_lerp, &ffth->tick_time);
728 * The reset from the RTC clock may be far from accurate, and
729 * reducing the gap between real time and interpolated time
730 * could take a very long time if the interpolated clock insists
731 * on strict monotonicity. The clock is reset under very strict
732 * conditions (kernel time is known to be wrong and
733 * synchronization daemon has been restarted recently.
734 * ffclock_boottime absorbs the jump to ensure boot time is
735 * correct and uptime functions stay consistent.
737 if (((ffclock_status & FFCLOCK_STA_UNSYNC) == FFCLOCK_STA_UNSYNC) &&
738 ((cest->status & FFCLOCK_STA_UNSYNC) == 0) &&
739 ((cest->status & FFCLOCK_STA_WARMUP) == FFCLOCK_STA_WARMUP)) {
741 bintime_add(&ffclock_boottime, &gap_lerp);
743 bintime_sub(&ffclock_boottime, &gap_lerp);
744 ffth->tick_time_lerp = ffth->tick_time;
745 bintime_clear(&gap_lerp);
748 ffclock_status = cest->status;
749 ffth->period_lerp = cest->period;
752 * Compute corrected period used for the linear interpolation of
753 * time. The rate of linear interpolation is capped to 5000PPM
756 if (bintime_isset(&gap_lerp)) {
757 ffdelta = cest->update_ffcount;
758 ffdelta -= fftimehands->cest.update_ffcount;
759 ffclock_convert_delta(ffdelta, cest->period, &bt);
762 bt.frac = 5000000 * (uint64_t)18446744073LL;
763 bintime_mul(&bt, polling);
764 if (bintime_cmp(&gap_lerp, &bt, >))
767 /* Approximate 1 sec by 1-(1/2^64) to ease arithmetic */
769 if (gap_lerp.sec > 0) {
771 frac /= ffdelta / gap_lerp.sec;
773 frac += gap_lerp.frac / ffdelta;
776 ffth->period_lerp += frac;
778 ffth->period_lerp -= frac;
790 * Adjust the fftimehands when the timecounter is changed. Stating the obvious,
791 * the old and new hardware counter cannot be read simultaneously. tc_windup()
792 * does read the two counters 'back to back', but a few cycles are effectively
793 * lost, and not accumulated in tick_ffcount. This is a fairly radical
794 * operation for a feed-forward synchronization daemon, and it is its job to not
795 * pushing irrelevant data to the kernel. Because there is no locking here,
796 * simply force to ignore pending or next update to give daemon a chance to
797 * realize the counter has changed.
800 ffclock_change_tc(struct timehands *th)
802 struct fftimehands *ffth;
803 struct ffclock_estimate *cest;
804 struct timecounter *tc;
808 ffth = fftimehands->next;
813 bcopy(&(fftimehands->cest), cest, sizeof(struct ffclock_estimate));
814 cest->period = ((1ULL << 63) / tc->tc_frequency ) << 1;
817 cest->status |= FFCLOCK_STA_UNSYNC;
819 ffth->tick_ffcount = fftimehands->tick_ffcount;
820 ffth->tick_time_lerp = fftimehands->tick_time_lerp;
821 ffth->tick_time = fftimehands->tick_time;
822 ffth->period_lerp = cest->period;
824 /* Do not lock but ignore next update from synchronization daemon. */
834 * Retrieve feed-forward counter and time of last kernel tick.
837 ffclock_last_tick(ffcounter *ffcount, struct bintime *bt, uint32_t flags)
839 struct fftimehands *ffth;
843 * No locking but check generation has not changed. Also need to make
844 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
849 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP)
850 *bt = ffth->tick_time_lerp;
852 *bt = ffth->tick_time;
853 *ffcount = ffth->tick_ffcount;
854 } while (gen == 0 || gen != ffth->gen);
858 * Absolute clock conversion. Low level function to convert ffcounter to
859 * bintime. The ffcounter is converted using the current ffclock period estimate
860 * or the "interpolated period" to ensure monotonicity.
861 * NOTE: this conversion may have been deferred, and the clock updated since the
862 * hardware counter has been read.
865 ffclock_convert_abs(ffcounter ffcount, struct bintime *bt, uint32_t flags)
867 struct fftimehands *ffth;
873 * No locking but check generation has not changed. Also need to make
874 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
879 if (ffcount > ffth->tick_ffcount)
880 ffdelta = ffcount - ffth->tick_ffcount;
882 ffdelta = ffth->tick_ffcount - ffcount;
884 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP) {
885 *bt = ffth->tick_time_lerp;
886 ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt2);
888 *bt = ffth->tick_time;
889 ffclock_convert_delta(ffdelta, ffth->cest.period, &bt2);
892 if (ffcount > ffth->tick_ffcount)
893 bintime_add(bt, &bt2);
895 bintime_sub(bt, &bt2);
896 } while (gen == 0 || gen != ffth->gen);
900 * Difference clock conversion.
901 * Low level function to Convert a time interval measured in RAW counter units
902 * into bintime. The difference clock allows measuring small intervals much more
903 * reliably than the absolute clock.
906 ffclock_convert_diff(ffcounter ffdelta, struct bintime *bt)
908 struct fftimehands *ffth;
911 /* No locking but check generation has not changed. */
915 ffclock_convert_delta(ffdelta, ffth->cest.period, bt);
916 } while (gen == 0 || gen != ffth->gen);
920 * Access to current ffcounter value.
923 ffclock_read_counter(ffcounter *ffcount)
925 struct timehands *th;
926 struct fftimehands *ffth;
927 unsigned int gen, delta;
930 * ffclock_windup() called from tc_windup(), safe to rely on
931 * th->th_generation only, for correct delta and ffcounter.
935 gen = atomic_load_acq_int(&th->th_generation);
937 delta = tc_delta(th);
938 *ffcount = ffth->tick_ffcount;
939 atomic_thread_fence_acq();
940 } while (gen == 0 || gen != th->th_generation);
946 binuptime(struct bintime *bt)
949 binuptime_fromclock(bt, sysclock_active);
953 nanouptime(struct timespec *tsp)
956 nanouptime_fromclock(tsp, sysclock_active);
960 microuptime(struct timeval *tvp)
963 microuptime_fromclock(tvp, sysclock_active);
967 bintime(struct bintime *bt)
970 bintime_fromclock(bt, sysclock_active);
974 nanotime(struct timespec *tsp)
977 nanotime_fromclock(tsp, sysclock_active);
981 microtime(struct timeval *tvp)
984 microtime_fromclock(tvp, sysclock_active);
988 getbinuptime(struct bintime *bt)
991 getbinuptime_fromclock(bt, sysclock_active);
995 getnanouptime(struct timespec *tsp)
998 getnanouptime_fromclock(tsp, sysclock_active);
1002 getmicrouptime(struct timeval *tvp)
1005 getmicrouptime_fromclock(tvp, sysclock_active);
1009 getbintime(struct bintime *bt)
1012 getbintime_fromclock(bt, sysclock_active);
1016 getnanotime(struct timespec *tsp)
1019 getnanotime_fromclock(tsp, sysclock_active);
1023 getmicrotime(struct timeval *tvp)
1026 getmicrouptime_fromclock(tvp, sysclock_active);
1029 #endif /* FFCLOCK */
1032 * This is a clone of getnanotime and used for walltimestamps.
1033 * The dtrace_ prefix prevents fbt from creating probes for
1034 * it so walltimestamp can be safely used in all fbt probes.
1037 dtrace_getnanotime(struct timespec *tsp)
1039 struct timehands *th;
1044 gen = atomic_load_acq_int(&th->th_generation);
1045 *tsp = th->th_nanotime;
1046 atomic_thread_fence_acq();
1047 } while (gen == 0 || gen != th->th_generation);
1051 * System clock currently providing time to the system. Modifiable via sysctl
1052 * when the FFCLOCK option is defined.
1054 int sysclock_active = SYSCLOCK_FBCK;
1056 /* Internal NTP status and error estimates. */
1057 extern int time_status;
1058 extern long time_esterror;
1061 * Take a snapshot of sysclock data which can be used to compare system clocks
1062 * and generate timestamps after the fact.
1065 sysclock_getsnapshot(struct sysclock_snap *clock_snap, int fast)
1067 struct fbclock_info *fbi;
1068 struct timehands *th;
1070 unsigned int delta, gen;
1073 struct fftimehands *ffth;
1074 struct ffclock_info *ffi;
1075 struct ffclock_estimate cest;
1077 ffi = &clock_snap->ff_info;
1080 fbi = &clock_snap->fb_info;
1085 gen = atomic_load_acq_int(&th->th_generation);
1086 fbi->th_scale = th->th_scale;
1087 fbi->tick_time = th->th_offset;
1090 ffi->tick_time = ffth->tick_time_lerp;
1091 ffi->tick_time_lerp = ffth->tick_time_lerp;
1092 ffi->period = ffth->cest.period;
1093 ffi->period_lerp = ffth->period_lerp;
1094 clock_snap->ffcount = ffth->tick_ffcount;
1098 delta = tc_delta(th);
1099 atomic_thread_fence_acq();
1100 } while (gen == 0 || gen != th->th_generation);
1102 clock_snap->delta = delta;
1103 clock_snap->sysclock_active = sysclock_active;
1105 /* Record feedback clock status and error. */
1106 clock_snap->fb_info.status = time_status;
1107 /* XXX: Very crude estimate of feedback clock error. */
1108 bt.sec = time_esterror / 1000000;
1109 bt.frac = ((time_esterror - bt.sec) * 1000000) *
1110 (uint64_t)18446744073709ULL;
1111 clock_snap->fb_info.error = bt;
1115 clock_snap->ffcount += delta;
1117 /* Record feed-forward clock leap second adjustment. */
1118 ffi->leapsec_adjustment = cest.leapsec_total;
1119 if (clock_snap->ffcount > cest.leapsec_next)
1120 ffi->leapsec_adjustment -= cest.leapsec;
1122 /* Record feed-forward clock status and error. */
1123 clock_snap->ff_info.status = cest.status;
1124 ffcount = clock_snap->ffcount - cest.update_ffcount;
1125 ffclock_convert_delta(ffcount, cest.period, &bt);
1126 /* 18446744073709 = int(2^64/1e12), err_bound_rate in [ps/s]. */
1127 bintime_mul(&bt, cest.errb_rate * (uint64_t)18446744073709ULL);
1128 /* 18446744073 = int(2^64 / 1e9), since err_abs in [ns]. */
1129 bintime_addx(&bt, cest.errb_abs * (uint64_t)18446744073ULL);
1130 clock_snap->ff_info.error = bt;
1135 * Convert a sysclock snapshot into a struct bintime based on the specified
1136 * clock source and flags.
1139 sysclock_snap2bintime(struct sysclock_snap *cs, struct bintime *bt,
1140 int whichclock, uint32_t flags)
1142 struct bintime boottimebin;
1148 switch (whichclock) {
1150 *bt = cs->fb_info.tick_time;
1152 /* If snapshot was created with !fast, delta will be >0. */
1154 bintime_addx(bt, cs->fb_info.th_scale * cs->delta);
1156 if ((flags & FBCLOCK_UPTIME) == 0) {
1157 getboottimebin(&boottimebin);
1158 bintime_add(bt, &boottimebin);
1163 if (flags & FFCLOCK_LERP) {
1164 *bt = cs->ff_info.tick_time_lerp;
1165 period = cs->ff_info.period_lerp;
1167 *bt = cs->ff_info.tick_time;
1168 period = cs->ff_info.period;
1171 /* If snapshot was created with !fast, delta will be >0. */
1172 if (cs->delta > 0) {
1173 ffclock_convert_delta(cs->delta, period, &bt2);
1174 bintime_add(bt, &bt2);
1177 /* Leap second adjustment. */
1178 if (flags & FFCLOCK_LEAPSEC)
1179 bt->sec -= cs->ff_info.leapsec_adjustment;
1181 /* Boot time adjustment, for uptime/monotonic clocks. */
1182 if (flags & FFCLOCK_UPTIME)
1183 bintime_sub(bt, &ffclock_boottime);
1195 * Initialize a new timecounter and possibly use it.
1198 tc_init(struct timecounter *tc)
1201 struct sysctl_oid *tc_root;
1203 u = tc->tc_frequency / tc->tc_counter_mask;
1204 /* XXX: We need some margin here, 10% is a guess */
1207 if (u > hz && tc->tc_quality >= 0) {
1208 tc->tc_quality = -2000;
1210 printf("Timecounter \"%s\" frequency %ju Hz",
1211 tc->tc_name, (uintmax_t)tc->tc_frequency);
1212 printf(" -- Insufficient hz, needs at least %u\n", u);
1214 } else if (tc->tc_quality >= 0 || bootverbose) {
1215 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
1216 tc->tc_name, (uintmax_t)tc->tc_frequency,
1220 tc->tc_next = timecounters;
1223 * Set up sysctl tree for this counter.
1225 tc_root = SYSCTL_ADD_NODE_WITH_LABEL(NULL,
1226 SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name,
1227 CTLFLAG_RW, 0, "timecounter description", "timecounter");
1228 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1229 "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0,
1230 "mask for implemented bits");
1231 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1232 "counter", CTLTYPE_UINT | CTLFLAG_RD, tc, sizeof(*tc),
1233 sysctl_kern_timecounter_get, "IU", "current timecounter value");
1234 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1235 "frequency", CTLTYPE_U64 | CTLFLAG_RD, tc, sizeof(*tc),
1236 sysctl_kern_timecounter_freq, "QU", "timecounter frequency");
1237 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1238 "quality", CTLFLAG_RD, &(tc->tc_quality), 0,
1239 "goodness of time counter");
1241 * Do not automatically switch if the current tc was specifically
1242 * chosen. Never automatically use a timecounter with negative quality.
1243 * Even though we run on the dummy counter, switching here may be
1244 * worse since this timecounter may not be monotonic.
1248 if (tc->tc_quality < 0)
1250 if (tc->tc_quality < timecounter->tc_quality)
1252 if (tc->tc_quality == timecounter->tc_quality &&
1253 tc->tc_frequency < timecounter->tc_frequency)
1255 (void)tc->tc_get_timecount(tc);
1256 (void)tc->tc_get_timecount(tc);
1260 /* Report the frequency of the current timecounter. */
1262 tc_getfrequency(void)
1265 return (timehands->th_counter->tc_frequency);
1269 sleeping_on_old_rtc(struct thread *td)
1273 * td_rtcgen is modified by curthread when it is running,
1274 * and by other threads in this function. By finding the thread
1275 * on a sleepqueue and holding the lock on the sleepqueue
1276 * chain, we guarantee that the thread is not running and that
1277 * modifying td_rtcgen is safe. Setting td_rtcgen to zero informs
1278 * the thread that it was woken due to a real-time clock adjustment.
1279 * (The declaration of td_rtcgen refers to this comment.)
1281 if (td->td_rtcgen != 0 && td->td_rtcgen != rtc_generation) {
1288 static struct mtx tc_setclock_mtx;
1289 MTX_SYSINIT(tc_setclock_init, &tc_setclock_mtx, "tcsetc", MTX_SPIN);
1292 * Step our concept of UTC. This is done by modifying our estimate of
1296 tc_setclock(struct timespec *ts)
1298 struct timespec tbef, taft;
1299 struct bintime bt, bt2;
1301 timespec2bintime(ts, &bt);
1303 mtx_lock_spin(&tc_setclock_mtx);
1304 cpu_tick_calibrate(1);
1306 bintime_sub(&bt, &bt2);
1308 /* XXX fiddle all the little crinkly bits around the fiords... */
1310 mtx_unlock_spin(&tc_setclock_mtx);
1312 /* Avoid rtc_generation == 0, since td_rtcgen == 0 is special. */
1313 atomic_add_rel_int(&rtc_generation, 2);
1314 sleepq_chains_remove_matching(sleeping_on_old_rtc);
1315 if (timestepwarnings) {
1318 "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
1319 (intmax_t)tbef.tv_sec, tbef.tv_nsec,
1320 (intmax_t)taft.tv_sec, taft.tv_nsec,
1321 (intmax_t)ts->tv_sec, ts->tv_nsec);
1326 * Initialize the next struct timehands in the ring and make
1327 * it the active timehands. Along the way we might switch to a different
1328 * timecounter and/or do seconds processing in NTP. Slightly magic.
1331 tc_windup(struct bintime *new_boottimebin)
1334 struct timehands *th, *tho;
1336 u_int delta, ncount, ogen;
1341 * Make the next timehands a copy of the current one, but do
1342 * not overwrite the generation or next pointer. While we
1343 * update the contents, the generation must be zero. We need
1344 * to ensure that the zero generation is visible before the
1345 * data updates become visible, which requires release fence.
1346 * For similar reasons, re-reading of the generation after the
1347 * data is read should use acquire fence.
1351 ogen = th->th_generation;
1352 th->th_generation = 0;
1353 atomic_thread_fence_rel();
1354 bcopy(tho, th, offsetof(struct timehands, th_generation));
1355 if (new_boottimebin != NULL)
1356 th->th_boottime = *new_boottimebin;
1359 * Capture a timecounter delta on the current timecounter and if
1360 * changing timecounters, a counter value from the new timecounter.
1361 * Update the offset fields accordingly.
1363 delta = tc_delta(th);
1364 if (th->th_counter != timecounter)
1365 ncount = timecounter->tc_get_timecount(timecounter);
1369 ffclock_windup(delta);
1371 th->th_offset_count += delta;
1372 th->th_offset_count &= th->th_counter->tc_counter_mask;
1373 while (delta > th->th_counter->tc_frequency) {
1374 /* Eat complete unadjusted seconds. */
1375 delta -= th->th_counter->tc_frequency;
1376 th->th_offset.sec++;
1378 if ((delta > th->th_counter->tc_frequency / 2) &&
1379 (th->th_scale * delta < ((uint64_t)1 << 63))) {
1380 /* The product th_scale * delta just barely overflows. */
1381 th->th_offset.sec++;
1383 bintime_addx(&th->th_offset, th->th_scale * delta);
1386 * Hardware latching timecounters may not generate interrupts on
1387 * PPS events, so instead we poll them. There is a finite risk that
1388 * the hardware might capture a count which is later than the one we
1389 * got above, and therefore possibly in the next NTP second which might
1390 * have a different rate than the current NTP second. It doesn't
1391 * matter in practice.
1393 if (tho->th_counter->tc_poll_pps)
1394 tho->th_counter->tc_poll_pps(tho->th_counter);
1397 * Deal with NTP second processing. The for loop normally
1398 * iterates at most once, but in extreme situations it might
1399 * keep NTP sane if timeouts are not run for several seconds.
1400 * At boot, the time step can be large when the TOD hardware
1401 * has been read, so on really large steps, we call
1402 * ntp_update_second only twice. We need to call it twice in
1403 * case we missed a leap second.
1406 bintime_add(&bt, &th->th_boottime);
1407 i = bt.sec - tho->th_microtime.tv_sec;
1410 for (; i > 0; i--) {
1412 ntp_update_second(&th->th_adjustment, &bt.sec);
1414 th->th_boottime.sec += bt.sec - t;
1416 th->th_bintime = th->th_offset;
1417 bintime_add(&th->th_bintime, &th->th_boottime);
1418 /* Update the UTC timestamps used by the get*() functions. */
1419 /* XXX shouldn't do this here. Should force non-`get' versions. */
1420 bintime2timeval(&bt, &th->th_microtime);
1421 bintime2timespec(&bt, &th->th_nanotime);
1423 /* Now is a good time to change timecounters. */
1424 if (th->th_counter != timecounter) {
1426 if ((timecounter->tc_flags & TC_FLAGS_C2STOP) != 0)
1427 cpu_disable_c2_sleep++;
1428 if ((th->th_counter->tc_flags & TC_FLAGS_C2STOP) != 0)
1429 cpu_disable_c2_sleep--;
1431 th->th_counter = timecounter;
1432 th->th_offset_count = ncount;
1433 tc_min_ticktock_freq = max(1, timecounter->tc_frequency /
1434 (((uint64_t)timecounter->tc_counter_mask + 1) / 3));
1436 ffclock_change_tc(th);
1441 * Recalculate the scaling factor. We want the number of 1/2^64
1442 * fractions of a second per period of the hardware counter, taking
1443 * into account the th_adjustment factor which the NTP PLL/adjtime(2)
1444 * processing provides us with.
1446 * The th_adjustment is nanoseconds per second with 32 bit binary
1447 * fraction and we want 64 bit binary fraction of second:
1449 * x = a * 2^32 / 10^9 = a * 4.294967296
1451 * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
1452 * we can only multiply by about 850 without overflowing, that
1453 * leaves no suitably precise fractions for multiply before divide.
1455 * Divide before multiply with a fraction of 2199/512 results in a
1456 * systematic undercompensation of 10PPM of th_adjustment. On a
1457 * 5000PPM adjustment this is a 0.05PPM error. This is acceptable.
1459 * We happily sacrifice the lowest of the 64 bits of our result
1460 * to the goddess of code clarity.
1463 scale = (uint64_t)1 << 63;
1464 scale += (th->th_adjustment / 1024) * 2199;
1465 scale /= th->th_counter->tc_frequency;
1466 th->th_scale = scale * 2;
1469 * Now that the struct timehands is again consistent, set the new
1470 * generation number, making sure to not make it zero.
1474 atomic_store_rel_int(&th->th_generation, ogen);
1476 /* Go live with the new struct timehands. */
1478 switch (sysclock_active) {
1481 time_second = th->th_microtime.tv_sec;
1482 time_uptime = th->th_offset.sec;
1486 time_second = fftimehands->tick_time_lerp.sec;
1487 time_uptime = fftimehands->tick_time_lerp.sec - ffclock_boottime.sec;
1493 timekeep_push_vdso();
1496 /* Report or change the active timecounter hardware. */
1498 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
1501 struct timecounter *newtc, *tc;
1505 strlcpy(newname, tc->tc_name, sizeof(newname));
1507 error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
1508 if (error != 0 || req->newptr == NULL)
1510 /* Record that the tc in use now was specifically chosen. */
1512 if (strcmp(newname, tc->tc_name) == 0)
1514 for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
1515 if (strcmp(newname, newtc->tc_name) != 0)
1518 /* Warm up new timecounter. */
1519 (void)newtc->tc_get_timecount(newtc);
1520 (void)newtc->tc_get_timecount(newtc);
1522 timecounter = newtc;
1525 * The vdso timehands update is deferred until the next
1528 * This is prudent given that 'timekeep_push_vdso()' does not
1529 * use any locking and that it can be called in hard interrupt
1530 * context via 'tc_windup()'.
1537 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW,
1538 0, 0, sysctl_kern_timecounter_hardware, "A",
1539 "Timecounter hardware selected");
1542 /* Report the available timecounter hardware. */
1544 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
1547 struct timecounter *tc;
1550 sbuf_new_for_sysctl(&sb, NULL, 0, req);
1551 for (tc = timecounters; tc != NULL; tc = tc->tc_next) {
1552 if (tc != timecounters)
1553 sbuf_putc(&sb, ' ');
1554 sbuf_printf(&sb, "%s(%d)", tc->tc_name, tc->tc_quality);
1556 error = sbuf_finish(&sb);
1561 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice, CTLTYPE_STRING | CTLFLAG_RD,
1562 0, 0, sysctl_kern_timecounter_choice, "A", "Timecounter hardware detected");
1565 * RFC 2783 PPS-API implementation.
1569 * Return true if the driver is aware of the abi version extensions in the
1570 * pps_state structure, and it supports at least the given abi version number.
1573 abi_aware(struct pps_state *pps, int vers)
1576 return ((pps->kcmode & KCMODE_ABIFLAG) && pps->driver_abi >= vers);
1580 pps_fetch(struct pps_fetch_args *fapi, struct pps_state *pps)
1583 pps_seq_t aseq, cseq;
1586 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1590 * If no timeout is requested, immediately return whatever values were
1591 * most recently captured. If timeout seconds is -1, that's a request
1592 * to block without a timeout. WITNESS won't let us sleep forever
1593 * without a lock (we really don't need a lock), so just repeatedly
1594 * sleep a long time.
1596 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) {
1597 if (fapi->timeout.tv_sec == -1)
1600 tv.tv_sec = fapi->timeout.tv_sec;
1601 tv.tv_usec = fapi->timeout.tv_nsec / 1000;
1604 aseq = pps->ppsinfo.assert_sequence;
1605 cseq = pps->ppsinfo.clear_sequence;
1606 while (aseq == pps->ppsinfo.assert_sequence &&
1607 cseq == pps->ppsinfo.clear_sequence) {
1608 if (abi_aware(pps, 1) && pps->driver_mtx != NULL) {
1609 if (pps->flags & PPSFLAG_MTX_SPIN) {
1610 err = msleep_spin(pps, pps->driver_mtx,
1613 err = msleep(pps, pps->driver_mtx, PCATCH,
1617 err = tsleep(pps, PCATCH, "ppsfch", timo);
1619 if (err == EWOULDBLOCK) {
1620 if (fapi->timeout.tv_sec == -1) {
1625 } else if (err != 0) {
1631 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1632 fapi->pps_info_buf = pps->ppsinfo;
1638 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1641 struct pps_fetch_args *fapi;
1643 struct pps_fetch_ffc_args *fapi_ffc;
1646 struct pps_kcbind_args *kapi;
1649 KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
1651 case PPS_IOC_CREATE:
1653 case PPS_IOC_DESTROY:
1655 case PPS_IOC_SETPARAMS:
1656 app = (pps_params_t *)data;
1657 if (app->mode & ~pps->ppscap)
1660 /* Ensure only a single clock is selected for ffc timestamp. */
1661 if ((app->mode & PPS_TSCLK_MASK) == PPS_TSCLK_MASK)
1664 pps->ppsparam = *app;
1666 case PPS_IOC_GETPARAMS:
1667 app = (pps_params_t *)data;
1668 *app = pps->ppsparam;
1669 app->api_version = PPS_API_VERS_1;
1671 case PPS_IOC_GETCAP:
1672 *(int*)data = pps->ppscap;
1675 fapi = (struct pps_fetch_args *)data;
1676 return (pps_fetch(fapi, pps));
1678 case PPS_IOC_FETCH_FFCOUNTER:
1679 fapi_ffc = (struct pps_fetch_ffc_args *)data;
1680 if (fapi_ffc->tsformat && fapi_ffc->tsformat !=
1683 if (fapi_ffc->timeout.tv_sec || fapi_ffc->timeout.tv_nsec)
1684 return (EOPNOTSUPP);
1685 pps->ppsinfo_ffc.current_mode = pps->ppsparam.mode;
1686 fapi_ffc->pps_info_buf_ffc = pps->ppsinfo_ffc;
1687 /* Overwrite timestamps if feedback clock selected. */
1688 switch (pps->ppsparam.mode & PPS_TSCLK_MASK) {
1689 case PPS_TSCLK_FBCK:
1690 fapi_ffc->pps_info_buf_ffc.assert_timestamp =
1691 pps->ppsinfo.assert_timestamp;
1692 fapi_ffc->pps_info_buf_ffc.clear_timestamp =
1693 pps->ppsinfo.clear_timestamp;
1695 case PPS_TSCLK_FFWD:
1701 #endif /* FFCLOCK */
1702 case PPS_IOC_KCBIND:
1704 kapi = (struct pps_kcbind_args *)data;
1705 /* XXX Only root should be able to do this */
1706 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1708 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1710 if (kapi->edge & ~pps->ppscap)
1712 pps->kcmode = (kapi->edge & KCMODE_EDGEMASK) |
1713 (pps->kcmode & KCMODE_ABIFLAG);
1716 return (EOPNOTSUPP);
1724 pps_init(struct pps_state *pps)
1726 pps->ppscap |= PPS_TSFMT_TSPEC | PPS_CANWAIT;
1727 if (pps->ppscap & PPS_CAPTUREASSERT)
1728 pps->ppscap |= PPS_OFFSETASSERT;
1729 if (pps->ppscap & PPS_CAPTURECLEAR)
1730 pps->ppscap |= PPS_OFFSETCLEAR;
1732 pps->ppscap |= PPS_TSCLK_MASK;
1734 pps->kcmode &= ~KCMODE_ABIFLAG;
1738 pps_init_abi(struct pps_state *pps)
1742 if (pps->driver_abi > 0) {
1743 pps->kcmode |= KCMODE_ABIFLAG;
1744 pps->kernel_abi = PPS_ABI_VERSION;
1749 pps_capture(struct pps_state *pps)
1751 struct timehands *th;
1753 KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
1755 pps->capgen = atomic_load_acq_int(&th->th_generation);
1758 pps->capffth = fftimehands;
1760 pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
1761 atomic_thread_fence_acq();
1762 if (pps->capgen != th->th_generation)
1767 pps_event(struct pps_state *pps, int event)
1770 struct timespec ts, *tsp, *osp;
1771 u_int tcount, *pcount;
1775 struct timespec *tsp_ffc;
1776 pps_seq_t *pseq_ffc;
1783 KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
1784 /* Nothing to do if not currently set to capture this event type. */
1785 if ((event & pps->ppsparam.mode) == 0)
1787 /* If the timecounter was wound up underneath us, bail out. */
1788 if (pps->capgen == 0 || pps->capgen !=
1789 atomic_load_acq_int(&pps->capth->th_generation))
1792 /* Things would be easier with arrays. */
1793 if (event == PPS_CAPTUREASSERT) {
1794 tsp = &pps->ppsinfo.assert_timestamp;
1795 osp = &pps->ppsparam.assert_offset;
1796 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1798 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1800 pcount = &pps->ppscount[0];
1801 pseq = &pps->ppsinfo.assert_sequence;
1803 ffcount = &pps->ppsinfo_ffc.assert_ffcount;
1804 tsp_ffc = &pps->ppsinfo_ffc.assert_timestamp;
1805 pseq_ffc = &pps->ppsinfo_ffc.assert_sequence;
1808 tsp = &pps->ppsinfo.clear_timestamp;
1809 osp = &pps->ppsparam.clear_offset;
1810 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1812 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1814 pcount = &pps->ppscount[1];
1815 pseq = &pps->ppsinfo.clear_sequence;
1817 ffcount = &pps->ppsinfo_ffc.clear_ffcount;
1818 tsp_ffc = &pps->ppsinfo_ffc.clear_timestamp;
1819 pseq_ffc = &pps->ppsinfo_ffc.clear_sequence;
1824 * If the timecounter changed, we cannot compare the count values, so
1825 * we have to drop the rest of the PPS-stuff until the next event.
1827 if (pps->ppstc != pps->capth->th_counter) {
1828 pps->ppstc = pps->capth->th_counter;
1829 *pcount = pps->capcount;
1830 pps->ppscount[2] = pps->capcount;
1834 /* Convert the count to a timespec. */
1835 tcount = pps->capcount - pps->capth->th_offset_count;
1836 tcount &= pps->capth->th_counter->tc_counter_mask;
1837 bt = pps->capth->th_bintime;
1838 bintime_addx(&bt, pps->capth->th_scale * tcount);
1839 bintime2timespec(&bt, &ts);
1841 /* If the timecounter was wound up underneath us, bail out. */
1842 atomic_thread_fence_acq();
1843 if (pps->capgen != pps->capth->th_generation)
1846 *pcount = pps->capcount;
1851 timespecadd(tsp, osp);
1852 if (tsp->tv_nsec < 0) {
1853 tsp->tv_nsec += 1000000000;
1859 *ffcount = pps->capffth->tick_ffcount + tcount;
1860 bt = pps->capffth->tick_time;
1861 ffclock_convert_delta(tcount, pps->capffth->cest.period, &bt);
1862 bintime_add(&bt, &pps->capffth->tick_time);
1863 bintime2timespec(&bt, &ts);
1873 * Feed the NTP PLL/FLL.
1874 * The FLL wants to know how many (hardware) nanoseconds
1875 * elapsed since the previous event.
1877 tcount = pps->capcount - pps->ppscount[2];
1878 pps->ppscount[2] = pps->capcount;
1879 tcount &= pps->capth->th_counter->tc_counter_mask;
1880 scale = (uint64_t)1 << 63;
1881 scale /= pps->capth->th_counter->tc_frequency;
1885 bintime_addx(&bt, scale * tcount);
1886 bintime2timespec(&bt, &ts);
1887 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
1891 /* Wakeup anyone sleeping in pps_fetch(). */
1896 * Timecounters need to be updated every so often to prevent the hardware
1897 * counter from overflowing. Updating also recalculates the cached values
1898 * used by the get*() family of functions, so their precision depends on
1899 * the update frequency.
1903 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0,
1904 "Approximate number of hardclock ticks in a millisecond");
1907 tc_ticktock(int cnt)
1911 if (mtx_trylock_spin(&tc_setclock_mtx)) {
1913 if (count >= tc_tick) {
1917 mtx_unlock_spin(&tc_setclock_mtx);
1921 static void __inline
1922 tc_adjprecision(void)
1926 if (tc_timepercentage > 0) {
1927 t = (99 + tc_timepercentage) / tc_timepercentage;
1928 tc_precexp = fls(t + (t >> 1)) - 1;
1929 FREQ2BT(hz / tc_tick, &bt_timethreshold);
1930 FREQ2BT(hz, &bt_tickthreshold);
1931 bintime_shift(&bt_timethreshold, tc_precexp);
1932 bintime_shift(&bt_tickthreshold, tc_precexp);
1935 bt_timethreshold.sec = INT_MAX;
1936 bt_timethreshold.frac = ~(uint64_t)0;
1937 bt_tickthreshold = bt_timethreshold;
1939 sbt_timethreshold = bttosbt(bt_timethreshold);
1940 sbt_tickthreshold = bttosbt(bt_tickthreshold);
1944 sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS)
1948 val = tc_timepercentage;
1949 error = sysctl_handle_int(oidp, &val, 0, req);
1950 if (error != 0 || req->newptr == NULL)
1952 tc_timepercentage = val;
1961 inittimecounter(void *dummy)
1967 * Set the initial timeout to
1968 * max(1, <approx. number of hardclock ticks in a millisecond>).
1969 * People should probably not use the sysctl to set the timeout
1970 * to smaller than its initial value, since that value is the
1971 * smallest reasonable one. If they want better timestamps they
1972 * should use the non-"get"* functions.
1975 tc_tick = (hz + 500) / 1000;
1979 FREQ2BT(hz, &tick_bt);
1980 tick_sbt = bttosbt(tick_bt);
1981 tick_rate = hz / tc_tick;
1982 FREQ2BT(tick_rate, &tc_tick_bt);
1983 tc_tick_sbt = bttosbt(tc_tick_bt);
1984 p = (tc_tick * 1000000) / hz;
1985 printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
1990 /* warm up new timecounter (again) and get rolling. */
1991 (void)timecounter->tc_get_timecount(timecounter);
1992 (void)timecounter->tc_get_timecount(timecounter);
1993 mtx_lock_spin(&tc_setclock_mtx);
1995 mtx_unlock_spin(&tc_setclock_mtx);
1998 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL);
2000 /* Cpu tick handling -------------------------------------------------*/
2002 static int cpu_tick_variable;
2003 static uint64_t cpu_tick_frequency;
2005 static DPCPU_DEFINE(uint64_t, tc_cpu_ticks_base);
2006 static DPCPU_DEFINE(unsigned, tc_cpu_ticks_last);
2011 struct timecounter *tc;
2012 uint64_t res, *base;
2016 base = DPCPU_PTR(tc_cpu_ticks_base);
2017 last = DPCPU_PTR(tc_cpu_ticks_last);
2018 tc = timehands->th_counter;
2019 u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
2021 *base += (uint64_t)tc->tc_counter_mask + 1;
2029 cpu_tick_calibration(void)
2031 static time_t last_calib;
2033 if (time_uptime != last_calib && !(time_uptime & 0xf)) {
2034 cpu_tick_calibrate(0);
2035 last_calib = time_uptime;
2040 * This function gets called every 16 seconds on only one designated
2041 * CPU in the system from hardclock() via cpu_tick_calibration()().
2043 * Whenever the real time clock is stepped we get called with reset=1
2044 * to make sure we handle suspend/resume and similar events correctly.
2048 cpu_tick_calibrate(int reset)
2050 static uint64_t c_last;
2051 uint64_t c_this, c_delta;
2052 static struct bintime t_last;
2053 struct bintime t_this, t_delta;
2057 /* The clock was stepped, abort & reset */
2062 /* we don't calibrate fixed rate cputicks */
2063 if (!cpu_tick_variable)
2066 getbinuptime(&t_this);
2067 c_this = cpu_ticks();
2068 if (t_last.sec != 0) {
2069 c_delta = c_this - c_last;
2071 bintime_sub(&t_delta, &t_last);
2074 * 2^(64-20) / 16[s] =
2076 * 17.592.186.044.416 / 16 =
2077 * 1.099.511.627.776 [Hz]
2079 divi = t_delta.sec << 20;
2080 divi |= t_delta.frac >> (64 - 20);
2083 if (c_delta > cpu_tick_frequency) {
2084 if (0 && bootverbose)
2085 printf("cpu_tick increased to %ju Hz\n",
2087 cpu_tick_frequency = c_delta;
2095 set_cputicker(cpu_tick_f *func, uint64_t freq, unsigned var)
2099 cpu_ticks = tc_cpu_ticks;
2101 cpu_tick_frequency = freq;
2102 cpu_tick_variable = var;
2111 if (cpu_ticks == tc_cpu_ticks)
2112 return (tc_getfrequency());
2113 return (cpu_tick_frequency);
2117 * We need to be slightly careful converting cputicks to microseconds.
2118 * There is plenty of margin in 64 bits of microseconds (half a million
2119 * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
2120 * before divide conversion (to retain precision) we find that the
2121 * margin shrinks to 1.5 hours (one millionth of 146y).
2122 * With a three prong approach we never lose significant bits, no
2123 * matter what the cputick rate and length of timeinterval is.
2127 cputick2usec(uint64_t tick)
2130 if (tick > 18446744073709551LL) /* floor(2^64 / 1000) */
2131 return (tick / (cpu_tickrate() / 1000000LL));
2132 else if (tick > 18446744073709LL) /* floor(2^64 / 1000000) */
2133 return ((tick * 1000LL) / (cpu_tickrate() / 1000LL));
2135 return ((tick * 1000000LL) / cpu_tickrate());
2138 cpu_tick_f *cpu_ticks = tc_cpu_ticks;
2140 static int vdso_th_enable = 1;
2142 sysctl_fast_gettime(SYSCTL_HANDLER_ARGS)
2144 int old_vdso_th_enable, error;
2146 old_vdso_th_enable = vdso_th_enable;
2147 error = sysctl_handle_int(oidp, &old_vdso_th_enable, 0, req);
2150 vdso_th_enable = old_vdso_th_enable;
2153 SYSCTL_PROC(_kern_timecounter, OID_AUTO, fast_gettime,
2154 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
2155 NULL, 0, sysctl_fast_gettime, "I", "Enable fast time of day");
2158 tc_fill_vdso_timehands(struct vdso_timehands *vdso_th)
2160 struct timehands *th;
2164 vdso_th->th_scale = th->th_scale;
2165 vdso_th->th_offset_count = th->th_offset_count;
2166 vdso_th->th_counter_mask = th->th_counter->tc_counter_mask;
2167 vdso_th->th_offset = th->th_offset;
2168 vdso_th->th_boottime = th->th_boottime;
2169 if (th->th_counter->tc_fill_vdso_timehands != NULL) {
2170 enabled = th->th_counter->tc_fill_vdso_timehands(vdso_th,
2174 if (!vdso_th_enable)
2179 #ifdef COMPAT_FREEBSD32
2181 tc_fill_vdso_timehands32(struct vdso_timehands32 *vdso_th32)
2183 struct timehands *th;
2187 *(uint64_t *)&vdso_th32->th_scale[0] = th->th_scale;
2188 vdso_th32->th_offset_count = th->th_offset_count;
2189 vdso_th32->th_counter_mask = th->th_counter->tc_counter_mask;
2190 vdso_th32->th_offset.sec = th->th_offset.sec;
2191 *(uint64_t *)&vdso_th32->th_offset.frac[0] = th->th_offset.frac;
2192 vdso_th32->th_boottime.sec = th->th_boottime.sec;
2193 *(uint64_t *)&vdso_th32->th_boottime.frac[0] = th->th_boottime.frac;
2194 if (th->th_counter->tc_fill_vdso_timehands32 != NULL) {
2195 enabled = th->th_counter->tc_fill_vdso_timehands32(vdso_th32,
2199 if (!vdso_th_enable)