2 * SPDX-License-Identifier: Beerware
4 * ----------------------------------------------------------------------------
5 * "THE BEER-WARE LICENSE" (Revision 42):
6 * <phk@FreeBSD.ORG> wrote this file. As long as you retain this notice you
7 * can do whatever you want with this stuff. If we meet some day, and you think
8 * this stuff is worth it, you can buy me a beer in return. Poul-Henning Kamp
9 * ----------------------------------------------------------------------------
11 * Copyright (c) 2011, 2015, 2016 The FreeBSD Foundation
12 * All rights reserved.
14 * Portions of this software were developed by Julien Ridoux at the University
15 * of Melbourne under sponsorship from the FreeBSD Foundation.
17 * Portions of this software were developed by Konstantin Belousov
18 * under sponsorship from the FreeBSD Foundation.
21 #include <sys/cdefs.h>
22 __FBSDID("$FreeBSD$");
25 #include "opt_ffclock.h"
27 #include <sys/param.h>
28 #include <sys/kernel.h>
29 #include <sys/limits.h>
31 #include <sys/mutex.h>
34 #include <sys/sleepqueue.h>
35 #include <sys/sysctl.h>
36 #include <sys/syslog.h>
37 #include <sys/systm.h>
38 #include <sys/timeffc.h>
39 #include <sys/timepps.h>
40 #include <sys/timetc.h>
41 #include <sys/timex.h>
45 * A large step happens on boot. This constant detects such steps.
46 * It is relatively small so that ntp_update_second gets called enough
47 * in the typical 'missed a couple of seconds' case, but doesn't loop
48 * forever when the time step is large.
50 #define LARGE_STEP 200
53 * Implement a dummy timecounter which we can use until we get a real one
54 * in the air. This allows the console and other early stuff to use
59 dummy_get_timecount(struct timecounter *tc)
66 static struct timecounter dummy_timecounter = {
67 dummy_get_timecount, 0, ~0u, 1000000, "dummy", -1000000
71 /* These fields must be initialized by the driver. */
72 struct timecounter *th_counter;
73 int64_t th_adjustment;
75 u_int th_offset_count;
76 struct bintime th_offset;
77 struct bintime th_bintime;
78 struct timeval th_microtime;
79 struct timespec th_nanotime;
80 struct bintime th_boottime;
81 /* Fields not to be copied in tc_windup start with th_generation. */
83 struct timehands *th_next;
86 static struct timehands th0;
87 static struct timehands th1 = {
90 static struct timehands th0 = {
91 .th_counter = &dummy_timecounter,
92 .th_scale = (uint64_t)-1 / 1000000,
93 .th_offset = { .sec = 1 },
98 static struct timehands *volatile timehands = &th0;
99 struct timecounter *timecounter = &dummy_timecounter;
100 static struct timecounter *timecounters = &dummy_timecounter;
102 int tc_min_ticktock_freq = 1;
104 volatile time_t time_second = 1;
105 volatile time_t time_uptime = 1;
107 static int sysctl_kern_boottime(SYSCTL_HANDLER_ARGS);
108 SYSCTL_PROC(_kern, KERN_BOOTTIME, boottime, CTLTYPE_STRUCT|CTLFLAG_RD,
109 NULL, 0, sysctl_kern_boottime, "S,timeval", "System boottime");
111 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW, 0, "");
112 static SYSCTL_NODE(_kern_timecounter, OID_AUTO, tc, CTLFLAG_RW, 0, "");
114 static int timestepwarnings;
115 SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RW,
116 ×tepwarnings, 0, "Log time steps");
118 struct bintime bt_timethreshold;
119 struct bintime bt_tickthreshold;
120 sbintime_t sbt_timethreshold;
121 sbintime_t sbt_tickthreshold;
122 struct bintime tc_tick_bt;
123 sbintime_t tc_tick_sbt;
125 int tc_timepercentage = TC_DEFAULTPERC;
126 static int sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS);
127 SYSCTL_PROC(_kern_timecounter, OID_AUTO, alloweddeviation,
128 CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, 0, 0,
129 sysctl_kern_timecounter_adjprecision, "I",
130 "Allowed time interval deviation in percents");
132 volatile int rtc_generation = 1;
134 static int tc_chosen; /* Non-zero if a specific tc was chosen via sysctl. */
136 static void tc_windup(struct bintime *new_boottimebin);
137 static void cpu_tick_calibrate(int);
139 void dtrace_getnanotime(struct timespec *tsp);
142 sysctl_kern_boottime(SYSCTL_HANDLER_ARGS)
144 struct timeval boottime;
146 getboottime(&boottime);
152 if (req->flags & SCTL_MASK32) {
153 tv[0] = boottime.tv_sec;
154 tv[1] = boottime.tv_usec;
155 return (SYSCTL_OUT(req, tv, sizeof(tv)));
159 return (SYSCTL_OUT(req, &boottime, sizeof(boottime)));
163 sysctl_kern_timecounter_get(SYSCTL_HANDLER_ARGS)
166 struct timecounter *tc = arg1;
168 ncount = tc->tc_get_timecount(tc);
169 return (sysctl_handle_int(oidp, &ncount, 0, req));
173 sysctl_kern_timecounter_freq(SYSCTL_HANDLER_ARGS)
176 struct timecounter *tc = arg1;
178 freq = tc->tc_frequency;
179 return (sysctl_handle_64(oidp, &freq, 0, req));
183 * Return the difference between the timehands' counter value now and what
184 * was when we copied it to the timehands' offset_count.
186 static __inline u_int
187 tc_delta(struct timehands *th)
189 struct timecounter *tc;
192 return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
193 tc->tc_counter_mask);
197 * Functions for reading the time. We have to loop until we are sure that
198 * the timehands that we operated on was not updated under our feet. See
199 * the comment in <sys/time.h> for a description of these 12 functions.
204 fbclock_binuptime(struct bintime *bt)
206 struct timehands *th;
211 gen = atomic_load_acq_int(&th->th_generation);
213 bintime_addx(bt, th->th_scale * tc_delta(th));
214 atomic_thread_fence_acq();
215 } while (gen == 0 || gen != th->th_generation);
219 fbclock_nanouptime(struct timespec *tsp)
223 fbclock_binuptime(&bt);
224 bintime2timespec(&bt, tsp);
228 fbclock_microuptime(struct timeval *tvp)
232 fbclock_binuptime(&bt);
233 bintime2timeval(&bt, tvp);
237 fbclock_bintime(struct bintime *bt)
239 struct timehands *th;
244 gen = atomic_load_acq_int(&th->th_generation);
245 *bt = th->th_bintime;
246 bintime_addx(bt, th->th_scale * tc_delta(th));
247 atomic_thread_fence_acq();
248 } while (gen == 0 || gen != th->th_generation);
252 fbclock_nanotime(struct timespec *tsp)
256 fbclock_bintime(&bt);
257 bintime2timespec(&bt, tsp);
261 fbclock_microtime(struct timeval *tvp)
265 fbclock_bintime(&bt);
266 bintime2timeval(&bt, tvp);
270 fbclock_getbinuptime(struct bintime *bt)
272 struct timehands *th;
277 gen = atomic_load_acq_int(&th->th_generation);
279 atomic_thread_fence_acq();
280 } while (gen == 0 || gen != th->th_generation);
284 fbclock_getnanouptime(struct timespec *tsp)
286 struct timehands *th;
291 gen = atomic_load_acq_int(&th->th_generation);
292 bintime2timespec(&th->th_offset, tsp);
293 atomic_thread_fence_acq();
294 } while (gen == 0 || gen != th->th_generation);
298 fbclock_getmicrouptime(struct timeval *tvp)
300 struct timehands *th;
305 gen = atomic_load_acq_int(&th->th_generation);
306 bintime2timeval(&th->th_offset, tvp);
307 atomic_thread_fence_acq();
308 } while (gen == 0 || gen != th->th_generation);
312 fbclock_getbintime(struct bintime *bt)
314 struct timehands *th;
319 gen = atomic_load_acq_int(&th->th_generation);
320 *bt = th->th_bintime;
321 atomic_thread_fence_acq();
322 } while (gen == 0 || gen != th->th_generation);
326 fbclock_getnanotime(struct timespec *tsp)
328 struct timehands *th;
333 gen = atomic_load_acq_int(&th->th_generation);
334 *tsp = th->th_nanotime;
335 atomic_thread_fence_acq();
336 } while (gen == 0 || gen != th->th_generation);
340 fbclock_getmicrotime(struct timeval *tvp)
342 struct timehands *th;
347 gen = atomic_load_acq_int(&th->th_generation);
348 *tvp = th->th_microtime;
349 atomic_thread_fence_acq();
350 } while (gen == 0 || gen != th->th_generation);
354 binuptime(struct bintime *bt)
356 struct timehands *th;
361 gen = atomic_load_acq_int(&th->th_generation);
363 bintime_addx(bt, th->th_scale * tc_delta(th));
364 atomic_thread_fence_acq();
365 } while (gen == 0 || gen != th->th_generation);
369 nanouptime(struct timespec *tsp)
374 bintime2timespec(&bt, tsp);
378 microuptime(struct timeval *tvp)
383 bintime2timeval(&bt, tvp);
387 bintime(struct bintime *bt)
389 struct timehands *th;
394 gen = atomic_load_acq_int(&th->th_generation);
395 *bt = th->th_bintime;
396 bintime_addx(bt, th->th_scale * tc_delta(th));
397 atomic_thread_fence_acq();
398 } while (gen == 0 || gen != th->th_generation);
402 nanotime(struct timespec *tsp)
407 bintime2timespec(&bt, tsp);
411 microtime(struct timeval *tvp)
416 bintime2timeval(&bt, tvp);
420 getbinuptime(struct bintime *bt)
422 struct timehands *th;
427 gen = atomic_load_acq_int(&th->th_generation);
429 atomic_thread_fence_acq();
430 } while (gen == 0 || gen != th->th_generation);
434 getnanouptime(struct timespec *tsp)
436 struct timehands *th;
441 gen = atomic_load_acq_int(&th->th_generation);
442 bintime2timespec(&th->th_offset, tsp);
443 atomic_thread_fence_acq();
444 } while (gen == 0 || gen != th->th_generation);
448 getmicrouptime(struct timeval *tvp)
450 struct timehands *th;
455 gen = atomic_load_acq_int(&th->th_generation);
456 bintime2timeval(&th->th_offset, tvp);
457 atomic_thread_fence_acq();
458 } while (gen == 0 || gen != th->th_generation);
462 getbintime(struct bintime *bt)
464 struct timehands *th;
469 gen = atomic_load_acq_int(&th->th_generation);
470 *bt = th->th_bintime;
471 atomic_thread_fence_acq();
472 } while (gen == 0 || gen != th->th_generation);
476 getnanotime(struct timespec *tsp)
478 struct timehands *th;
483 gen = atomic_load_acq_int(&th->th_generation);
484 *tsp = th->th_nanotime;
485 atomic_thread_fence_acq();
486 } while (gen == 0 || gen != th->th_generation);
490 getmicrotime(struct timeval *tvp)
492 struct timehands *th;
497 gen = atomic_load_acq_int(&th->th_generation);
498 *tvp = th->th_microtime;
499 atomic_thread_fence_acq();
500 } while (gen == 0 || gen != th->th_generation);
505 getboottime(struct timeval *boottime)
507 struct bintime boottimebin;
509 getboottimebin(&boottimebin);
510 bintime2timeval(&boottimebin, boottime);
514 getboottimebin(struct bintime *boottimebin)
516 struct timehands *th;
521 gen = atomic_load_acq_int(&th->th_generation);
522 *boottimebin = th->th_boottime;
523 atomic_thread_fence_acq();
524 } while (gen == 0 || gen != th->th_generation);
529 * Support for feed-forward synchronization algorithms. This is heavily inspired
530 * by the timehands mechanism but kept independent from it. *_windup() functions
531 * have some connection to avoid accessing the timecounter hardware more than
535 /* Feed-forward clock estimates kept updated by the synchronization daemon. */
536 struct ffclock_estimate ffclock_estimate;
537 struct bintime ffclock_boottime; /* Feed-forward boot time estimate. */
538 uint32_t ffclock_status; /* Feed-forward clock status. */
539 int8_t ffclock_updated; /* New estimates are available. */
540 struct mtx ffclock_mtx; /* Mutex on ffclock_estimate. */
543 struct ffclock_estimate cest;
544 struct bintime tick_time;
545 struct bintime tick_time_lerp;
546 ffcounter tick_ffcount;
547 uint64_t period_lerp;
548 volatile uint8_t gen;
549 struct fftimehands *next;
552 #define NUM_ELEMENTS(x) (sizeof(x) / sizeof(*x))
554 static struct fftimehands ffth[10];
555 static struct fftimehands *volatile fftimehands = ffth;
560 struct fftimehands *cur;
561 struct fftimehands *last;
563 memset(ffth, 0, sizeof(ffth));
565 last = ffth + NUM_ELEMENTS(ffth) - 1;
566 for (cur = ffth; cur < last; cur++)
571 ffclock_status = FFCLOCK_STA_UNSYNC;
572 mtx_init(&ffclock_mtx, "ffclock lock", NULL, MTX_DEF);
576 * Reset the feed-forward clock estimates. Called from inittodr() to get things
577 * kick started and uses the timecounter nominal frequency as a first period
578 * estimate. Note: this function may be called several time just after boot.
579 * Note: this is the only function that sets the value of boot time for the
580 * monotonic (i.e. uptime) version of the feed-forward clock.
583 ffclock_reset_clock(struct timespec *ts)
585 struct timecounter *tc;
586 struct ffclock_estimate cest;
588 tc = timehands->th_counter;
589 memset(&cest, 0, sizeof(struct ffclock_estimate));
591 timespec2bintime(ts, &ffclock_boottime);
592 timespec2bintime(ts, &(cest.update_time));
593 ffclock_read_counter(&cest.update_ffcount);
594 cest.leapsec_next = 0;
595 cest.period = ((1ULL << 63) / tc->tc_frequency) << 1;
598 cest.status = FFCLOCK_STA_UNSYNC;
599 cest.leapsec_total = 0;
602 mtx_lock(&ffclock_mtx);
603 bcopy(&cest, &ffclock_estimate, sizeof(struct ffclock_estimate));
604 ffclock_updated = INT8_MAX;
605 mtx_unlock(&ffclock_mtx);
607 printf("ffclock reset: %s (%llu Hz), time = %ld.%09lu\n", tc->tc_name,
608 (unsigned long long)tc->tc_frequency, (long)ts->tv_sec,
609 (unsigned long)ts->tv_nsec);
613 * Sub-routine to convert a time interval measured in RAW counter units to time
614 * in seconds stored in bintime format.
615 * NOTE: bintime_mul requires u_int, but the value of the ffcounter may be
616 * larger than the max value of u_int (on 32 bit architecture). Loop to consume
620 ffclock_convert_delta(ffcounter ffdelta, uint64_t period, struct bintime *bt)
623 ffcounter delta, delta_max;
625 delta_max = (1ULL << (8 * sizeof(unsigned int))) - 1;
628 if (ffdelta > delta_max)
634 bintime_mul(&bt2, (unsigned int)delta);
635 bintime_add(bt, &bt2);
637 } while (ffdelta > 0);
641 * Update the fftimehands.
642 * Push the tick ffcount and time(s) forward based on current clock estimate.
643 * The conversion from ffcounter to bintime relies on the difference clock
644 * principle, whose accuracy relies on computing small time intervals. If a new
645 * clock estimate has been passed by the synchronisation daemon, make it
646 * current, and compute the linear interpolation for monotonic time if needed.
649 ffclock_windup(unsigned int delta)
651 struct ffclock_estimate *cest;
652 struct fftimehands *ffth;
653 struct bintime bt, gap_lerp;
656 unsigned int polling;
657 uint8_t forward_jump, ogen;
660 * Pick the next timehand, copy current ffclock estimates and move tick
661 * times and counter forward.
664 ffth = fftimehands->next;
668 bcopy(&fftimehands->cest, cest, sizeof(struct ffclock_estimate));
669 ffdelta = (ffcounter)delta;
670 ffth->period_lerp = fftimehands->period_lerp;
672 ffth->tick_time = fftimehands->tick_time;
673 ffclock_convert_delta(ffdelta, cest->period, &bt);
674 bintime_add(&ffth->tick_time, &bt);
676 ffth->tick_time_lerp = fftimehands->tick_time_lerp;
677 ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt);
678 bintime_add(&ffth->tick_time_lerp, &bt);
680 ffth->tick_ffcount = fftimehands->tick_ffcount + ffdelta;
683 * Assess the status of the clock, if the last update is too old, it is
684 * likely the synchronisation daemon is dead and the clock is free
687 if (ffclock_updated == 0) {
688 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
689 ffclock_convert_delta(ffdelta, cest->period, &bt);
690 if (bt.sec > 2 * FFCLOCK_SKM_SCALE)
691 ffclock_status |= FFCLOCK_STA_UNSYNC;
695 * If available, grab updated clock estimates and make them current.
696 * Recompute time at this tick using the updated estimates. The clock
697 * estimates passed the feed-forward synchronisation daemon may result
698 * in time conversion that is not monotonically increasing (just after
699 * the update). time_lerp is a particular linear interpolation over the
700 * synchronisation algo polling period that ensures monotonicity for the
701 * clock ids requesting it.
703 if (ffclock_updated > 0) {
704 bcopy(&ffclock_estimate, cest, sizeof(struct ffclock_estimate));
705 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
706 ffth->tick_time = cest->update_time;
707 ffclock_convert_delta(ffdelta, cest->period, &bt);
708 bintime_add(&ffth->tick_time, &bt);
710 /* ffclock_reset sets ffclock_updated to INT8_MAX */
711 if (ffclock_updated == INT8_MAX)
712 ffth->tick_time_lerp = ffth->tick_time;
714 if (bintime_cmp(&ffth->tick_time, &ffth->tick_time_lerp, >))
719 bintime_clear(&gap_lerp);
721 gap_lerp = ffth->tick_time;
722 bintime_sub(&gap_lerp, &ffth->tick_time_lerp);
724 gap_lerp = ffth->tick_time_lerp;
725 bintime_sub(&gap_lerp, &ffth->tick_time);
729 * The reset from the RTC clock may be far from accurate, and
730 * reducing the gap between real time and interpolated time
731 * could take a very long time if the interpolated clock insists
732 * on strict monotonicity. The clock is reset under very strict
733 * conditions (kernel time is known to be wrong and
734 * synchronization daemon has been restarted recently.
735 * ffclock_boottime absorbs the jump to ensure boot time is
736 * correct and uptime functions stay consistent.
738 if (((ffclock_status & FFCLOCK_STA_UNSYNC) == FFCLOCK_STA_UNSYNC) &&
739 ((cest->status & FFCLOCK_STA_UNSYNC) == 0) &&
740 ((cest->status & FFCLOCK_STA_WARMUP) == FFCLOCK_STA_WARMUP)) {
742 bintime_add(&ffclock_boottime, &gap_lerp);
744 bintime_sub(&ffclock_boottime, &gap_lerp);
745 ffth->tick_time_lerp = ffth->tick_time;
746 bintime_clear(&gap_lerp);
749 ffclock_status = cest->status;
750 ffth->period_lerp = cest->period;
753 * Compute corrected period used for the linear interpolation of
754 * time. The rate of linear interpolation is capped to 5000PPM
757 if (bintime_isset(&gap_lerp)) {
758 ffdelta = cest->update_ffcount;
759 ffdelta -= fftimehands->cest.update_ffcount;
760 ffclock_convert_delta(ffdelta, cest->period, &bt);
763 bt.frac = 5000000 * (uint64_t)18446744073LL;
764 bintime_mul(&bt, polling);
765 if (bintime_cmp(&gap_lerp, &bt, >))
768 /* Approximate 1 sec by 1-(1/2^64) to ease arithmetic */
770 if (gap_lerp.sec > 0) {
772 frac /= ffdelta / gap_lerp.sec;
774 frac += gap_lerp.frac / ffdelta;
777 ffth->period_lerp += frac;
779 ffth->period_lerp -= frac;
791 * Adjust the fftimehands when the timecounter is changed. Stating the obvious,
792 * the old and new hardware counter cannot be read simultaneously. tc_windup()
793 * does read the two counters 'back to back', but a few cycles are effectively
794 * lost, and not accumulated in tick_ffcount. This is a fairly radical
795 * operation for a feed-forward synchronization daemon, and it is its job to not
796 * pushing irrelevant data to the kernel. Because there is no locking here,
797 * simply force to ignore pending or next update to give daemon a chance to
798 * realize the counter has changed.
801 ffclock_change_tc(struct timehands *th)
803 struct fftimehands *ffth;
804 struct ffclock_estimate *cest;
805 struct timecounter *tc;
809 ffth = fftimehands->next;
814 bcopy(&(fftimehands->cest), cest, sizeof(struct ffclock_estimate));
815 cest->period = ((1ULL << 63) / tc->tc_frequency ) << 1;
818 cest->status |= FFCLOCK_STA_UNSYNC;
820 ffth->tick_ffcount = fftimehands->tick_ffcount;
821 ffth->tick_time_lerp = fftimehands->tick_time_lerp;
822 ffth->tick_time = fftimehands->tick_time;
823 ffth->period_lerp = cest->period;
825 /* Do not lock but ignore next update from synchronization daemon. */
835 * Retrieve feed-forward counter and time of last kernel tick.
838 ffclock_last_tick(ffcounter *ffcount, struct bintime *bt, uint32_t flags)
840 struct fftimehands *ffth;
844 * No locking but check generation has not changed. Also need to make
845 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
850 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP)
851 *bt = ffth->tick_time_lerp;
853 *bt = ffth->tick_time;
854 *ffcount = ffth->tick_ffcount;
855 } while (gen == 0 || gen != ffth->gen);
859 * Absolute clock conversion. Low level function to convert ffcounter to
860 * bintime. The ffcounter is converted using the current ffclock period estimate
861 * or the "interpolated period" to ensure monotonicity.
862 * NOTE: this conversion may have been deferred, and the clock updated since the
863 * hardware counter has been read.
866 ffclock_convert_abs(ffcounter ffcount, struct bintime *bt, uint32_t flags)
868 struct fftimehands *ffth;
874 * No locking but check generation has not changed. Also need to make
875 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
880 if (ffcount > ffth->tick_ffcount)
881 ffdelta = ffcount - ffth->tick_ffcount;
883 ffdelta = ffth->tick_ffcount - ffcount;
885 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP) {
886 *bt = ffth->tick_time_lerp;
887 ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt2);
889 *bt = ffth->tick_time;
890 ffclock_convert_delta(ffdelta, ffth->cest.period, &bt2);
893 if (ffcount > ffth->tick_ffcount)
894 bintime_add(bt, &bt2);
896 bintime_sub(bt, &bt2);
897 } while (gen == 0 || gen != ffth->gen);
901 * Difference clock conversion.
902 * Low level function to Convert a time interval measured in RAW counter units
903 * into bintime. The difference clock allows measuring small intervals much more
904 * reliably than the absolute clock.
907 ffclock_convert_diff(ffcounter ffdelta, struct bintime *bt)
909 struct fftimehands *ffth;
912 /* No locking but check generation has not changed. */
916 ffclock_convert_delta(ffdelta, ffth->cest.period, bt);
917 } while (gen == 0 || gen != ffth->gen);
921 * Access to current ffcounter value.
924 ffclock_read_counter(ffcounter *ffcount)
926 struct timehands *th;
927 struct fftimehands *ffth;
928 unsigned int gen, delta;
931 * ffclock_windup() called from tc_windup(), safe to rely on
932 * th->th_generation only, for correct delta and ffcounter.
936 gen = atomic_load_acq_int(&th->th_generation);
938 delta = tc_delta(th);
939 *ffcount = ffth->tick_ffcount;
940 atomic_thread_fence_acq();
941 } while (gen == 0 || gen != th->th_generation);
947 binuptime(struct bintime *bt)
950 binuptime_fromclock(bt, sysclock_active);
954 nanouptime(struct timespec *tsp)
957 nanouptime_fromclock(tsp, sysclock_active);
961 microuptime(struct timeval *tvp)
964 microuptime_fromclock(tvp, sysclock_active);
968 bintime(struct bintime *bt)
971 bintime_fromclock(bt, sysclock_active);
975 nanotime(struct timespec *tsp)
978 nanotime_fromclock(tsp, sysclock_active);
982 microtime(struct timeval *tvp)
985 microtime_fromclock(tvp, sysclock_active);
989 getbinuptime(struct bintime *bt)
992 getbinuptime_fromclock(bt, sysclock_active);
996 getnanouptime(struct timespec *tsp)
999 getnanouptime_fromclock(tsp, sysclock_active);
1003 getmicrouptime(struct timeval *tvp)
1006 getmicrouptime_fromclock(tvp, sysclock_active);
1010 getbintime(struct bintime *bt)
1013 getbintime_fromclock(bt, sysclock_active);
1017 getnanotime(struct timespec *tsp)
1020 getnanotime_fromclock(tsp, sysclock_active);
1024 getmicrotime(struct timeval *tvp)
1027 getmicrouptime_fromclock(tvp, sysclock_active);
1030 #endif /* FFCLOCK */
1033 * This is a clone of getnanotime and used for walltimestamps.
1034 * The dtrace_ prefix prevents fbt from creating probes for
1035 * it so walltimestamp can be safely used in all fbt probes.
1038 dtrace_getnanotime(struct timespec *tsp)
1040 struct timehands *th;
1045 gen = atomic_load_acq_int(&th->th_generation);
1046 *tsp = th->th_nanotime;
1047 atomic_thread_fence_acq();
1048 } while (gen == 0 || gen != th->th_generation);
1052 * System clock currently providing time to the system. Modifiable via sysctl
1053 * when the FFCLOCK option is defined.
1055 int sysclock_active = SYSCLOCK_FBCK;
1057 /* Internal NTP status and error estimates. */
1058 extern int time_status;
1059 extern long time_esterror;
1062 * Take a snapshot of sysclock data which can be used to compare system clocks
1063 * and generate timestamps after the fact.
1066 sysclock_getsnapshot(struct sysclock_snap *clock_snap, int fast)
1068 struct fbclock_info *fbi;
1069 struct timehands *th;
1071 unsigned int delta, gen;
1074 struct fftimehands *ffth;
1075 struct ffclock_info *ffi;
1076 struct ffclock_estimate cest;
1078 ffi = &clock_snap->ff_info;
1081 fbi = &clock_snap->fb_info;
1086 gen = atomic_load_acq_int(&th->th_generation);
1087 fbi->th_scale = th->th_scale;
1088 fbi->tick_time = th->th_offset;
1091 ffi->tick_time = ffth->tick_time_lerp;
1092 ffi->tick_time_lerp = ffth->tick_time_lerp;
1093 ffi->period = ffth->cest.period;
1094 ffi->period_lerp = ffth->period_lerp;
1095 clock_snap->ffcount = ffth->tick_ffcount;
1099 delta = tc_delta(th);
1100 atomic_thread_fence_acq();
1101 } while (gen == 0 || gen != th->th_generation);
1103 clock_snap->delta = delta;
1104 clock_snap->sysclock_active = sysclock_active;
1106 /* Record feedback clock status and error. */
1107 clock_snap->fb_info.status = time_status;
1108 /* XXX: Very crude estimate of feedback clock error. */
1109 bt.sec = time_esterror / 1000000;
1110 bt.frac = ((time_esterror - bt.sec) * 1000000) *
1111 (uint64_t)18446744073709ULL;
1112 clock_snap->fb_info.error = bt;
1116 clock_snap->ffcount += delta;
1118 /* Record feed-forward clock leap second adjustment. */
1119 ffi->leapsec_adjustment = cest.leapsec_total;
1120 if (clock_snap->ffcount > cest.leapsec_next)
1121 ffi->leapsec_adjustment -= cest.leapsec;
1123 /* Record feed-forward clock status and error. */
1124 clock_snap->ff_info.status = cest.status;
1125 ffcount = clock_snap->ffcount - cest.update_ffcount;
1126 ffclock_convert_delta(ffcount, cest.period, &bt);
1127 /* 18446744073709 = int(2^64/1e12), err_bound_rate in [ps/s]. */
1128 bintime_mul(&bt, cest.errb_rate * (uint64_t)18446744073709ULL);
1129 /* 18446744073 = int(2^64 / 1e9), since err_abs in [ns]. */
1130 bintime_addx(&bt, cest.errb_abs * (uint64_t)18446744073ULL);
1131 clock_snap->ff_info.error = bt;
1136 * Convert a sysclock snapshot into a struct bintime based on the specified
1137 * clock source and flags.
1140 sysclock_snap2bintime(struct sysclock_snap *cs, struct bintime *bt,
1141 int whichclock, uint32_t flags)
1143 struct bintime boottimebin;
1149 switch (whichclock) {
1151 *bt = cs->fb_info.tick_time;
1153 /* If snapshot was created with !fast, delta will be >0. */
1155 bintime_addx(bt, cs->fb_info.th_scale * cs->delta);
1157 if ((flags & FBCLOCK_UPTIME) == 0) {
1158 getboottimebin(&boottimebin);
1159 bintime_add(bt, &boottimebin);
1164 if (flags & FFCLOCK_LERP) {
1165 *bt = cs->ff_info.tick_time_lerp;
1166 period = cs->ff_info.period_lerp;
1168 *bt = cs->ff_info.tick_time;
1169 period = cs->ff_info.period;
1172 /* If snapshot was created with !fast, delta will be >0. */
1173 if (cs->delta > 0) {
1174 ffclock_convert_delta(cs->delta, period, &bt2);
1175 bintime_add(bt, &bt2);
1178 /* Leap second adjustment. */
1179 if (flags & FFCLOCK_LEAPSEC)
1180 bt->sec -= cs->ff_info.leapsec_adjustment;
1182 /* Boot time adjustment, for uptime/monotonic clocks. */
1183 if (flags & FFCLOCK_UPTIME)
1184 bintime_sub(bt, &ffclock_boottime);
1196 * Initialize a new timecounter and possibly use it.
1199 tc_init(struct timecounter *tc)
1202 struct sysctl_oid *tc_root;
1204 u = tc->tc_frequency / tc->tc_counter_mask;
1205 /* XXX: We need some margin here, 10% is a guess */
1208 if (u > hz && tc->tc_quality >= 0) {
1209 tc->tc_quality = -2000;
1211 printf("Timecounter \"%s\" frequency %ju Hz",
1212 tc->tc_name, (uintmax_t)tc->tc_frequency);
1213 printf(" -- Insufficient hz, needs at least %u\n", u);
1215 } else if (tc->tc_quality >= 0 || bootverbose) {
1216 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
1217 tc->tc_name, (uintmax_t)tc->tc_frequency,
1221 tc->tc_next = timecounters;
1224 * Set up sysctl tree for this counter.
1226 tc_root = SYSCTL_ADD_NODE_WITH_LABEL(NULL,
1227 SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name,
1228 CTLFLAG_RW, 0, "timecounter description", "timecounter");
1229 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1230 "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0,
1231 "mask for implemented bits");
1232 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1233 "counter", CTLTYPE_UINT | CTLFLAG_RD, tc, sizeof(*tc),
1234 sysctl_kern_timecounter_get, "IU", "current timecounter value");
1235 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1236 "frequency", CTLTYPE_U64 | CTLFLAG_RD, tc, sizeof(*tc),
1237 sysctl_kern_timecounter_freq, "QU", "timecounter frequency");
1238 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1239 "quality", CTLFLAG_RD, &(tc->tc_quality), 0,
1240 "goodness of time counter");
1242 * Do not automatically switch if the current tc was specifically
1243 * chosen. Never automatically use a timecounter with negative quality.
1244 * Even though we run on the dummy counter, switching here may be
1245 * worse since this timecounter may not be monotonic.
1249 if (tc->tc_quality < 0)
1251 if (tc->tc_quality < timecounter->tc_quality)
1253 if (tc->tc_quality == timecounter->tc_quality &&
1254 tc->tc_frequency < timecounter->tc_frequency)
1256 (void)tc->tc_get_timecount(tc);
1257 (void)tc->tc_get_timecount(tc);
1261 /* Report the frequency of the current timecounter. */
1263 tc_getfrequency(void)
1266 return (timehands->th_counter->tc_frequency);
1270 sleeping_on_old_rtc(struct thread *td)
1274 * td_rtcgen is modified by curthread when it is running,
1275 * and by other threads in this function. By finding the thread
1276 * on a sleepqueue and holding the lock on the sleepqueue
1277 * chain, we guarantee that the thread is not running and that
1278 * modifying td_rtcgen is safe. Setting td_rtcgen to zero informs
1279 * the thread that it was woken due to a real-time clock adjustment.
1280 * (The declaration of td_rtcgen refers to this comment.)
1282 if (td->td_rtcgen != 0 && td->td_rtcgen != rtc_generation) {
1289 static struct mtx tc_setclock_mtx;
1290 MTX_SYSINIT(tc_setclock_init, &tc_setclock_mtx, "tcsetc", MTX_SPIN);
1293 * Step our concept of UTC. This is done by modifying our estimate of
1297 tc_setclock(struct timespec *ts)
1299 struct timespec tbef, taft;
1300 struct bintime bt, bt2;
1302 timespec2bintime(ts, &bt);
1304 mtx_lock_spin(&tc_setclock_mtx);
1305 cpu_tick_calibrate(1);
1307 bintime_sub(&bt, &bt2);
1309 /* XXX fiddle all the little crinkly bits around the fiords... */
1311 mtx_unlock_spin(&tc_setclock_mtx);
1313 /* Avoid rtc_generation == 0, since td_rtcgen == 0 is special. */
1314 atomic_add_rel_int(&rtc_generation, 2);
1315 sleepq_chains_remove_matching(sleeping_on_old_rtc);
1316 if (timestepwarnings) {
1319 "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
1320 (intmax_t)tbef.tv_sec, tbef.tv_nsec,
1321 (intmax_t)taft.tv_sec, taft.tv_nsec,
1322 (intmax_t)ts->tv_sec, ts->tv_nsec);
1327 * Initialize the next struct timehands in the ring and make
1328 * it the active timehands. Along the way we might switch to a different
1329 * timecounter and/or do seconds processing in NTP. Slightly magic.
1332 tc_windup(struct bintime *new_boottimebin)
1335 struct timehands *th, *tho;
1337 u_int delta, ncount, ogen;
1342 * Make the next timehands a copy of the current one, but do
1343 * not overwrite the generation or next pointer. While we
1344 * update the contents, the generation must be zero. We need
1345 * to ensure that the zero generation is visible before the
1346 * data updates become visible, which requires release fence.
1347 * For similar reasons, re-reading of the generation after the
1348 * data is read should use acquire fence.
1352 ogen = th->th_generation;
1353 th->th_generation = 0;
1354 atomic_thread_fence_rel();
1355 memcpy(th, tho, offsetof(struct timehands, th_generation));
1356 if (new_boottimebin != NULL)
1357 th->th_boottime = *new_boottimebin;
1360 * Capture a timecounter delta on the current timecounter and if
1361 * changing timecounters, a counter value from the new timecounter.
1362 * Update the offset fields accordingly.
1364 delta = tc_delta(th);
1365 if (th->th_counter != timecounter)
1366 ncount = timecounter->tc_get_timecount(timecounter);
1370 ffclock_windup(delta);
1372 th->th_offset_count += delta;
1373 th->th_offset_count &= th->th_counter->tc_counter_mask;
1374 while (delta > th->th_counter->tc_frequency) {
1375 /* Eat complete unadjusted seconds. */
1376 delta -= th->th_counter->tc_frequency;
1377 th->th_offset.sec++;
1379 if ((delta > th->th_counter->tc_frequency / 2) &&
1380 (th->th_scale * delta < ((uint64_t)1 << 63))) {
1381 /* The product th_scale * delta just barely overflows. */
1382 th->th_offset.sec++;
1384 bintime_addx(&th->th_offset, th->th_scale * delta);
1387 * Hardware latching timecounters may not generate interrupts on
1388 * PPS events, so instead we poll them. There is a finite risk that
1389 * the hardware might capture a count which is later than the one we
1390 * got above, and therefore possibly in the next NTP second which might
1391 * have a different rate than the current NTP second. It doesn't
1392 * matter in practice.
1394 if (tho->th_counter->tc_poll_pps)
1395 tho->th_counter->tc_poll_pps(tho->th_counter);
1398 * Deal with NTP second processing. The for loop normally
1399 * iterates at most once, but in extreme situations it might
1400 * keep NTP sane if timeouts are not run for several seconds.
1401 * At boot, the time step can be large when the TOD hardware
1402 * has been read, so on really large steps, we call
1403 * ntp_update_second only twice. We need to call it twice in
1404 * case we missed a leap second.
1407 bintime_add(&bt, &th->th_boottime);
1408 i = bt.sec - tho->th_microtime.tv_sec;
1411 for (; i > 0; i--) {
1413 ntp_update_second(&th->th_adjustment, &bt.sec);
1415 th->th_boottime.sec += bt.sec - t;
1417 /* Update the UTC timestamps used by the get*() functions. */
1418 th->th_bintime = bt;
1419 bintime2timeval(&bt, &th->th_microtime);
1420 bintime2timespec(&bt, &th->th_nanotime);
1422 /* Now is a good time to change timecounters. */
1423 if (th->th_counter != timecounter) {
1425 if ((timecounter->tc_flags & TC_FLAGS_C2STOP) != 0)
1426 cpu_disable_c2_sleep++;
1427 if ((th->th_counter->tc_flags & TC_FLAGS_C2STOP) != 0)
1428 cpu_disable_c2_sleep--;
1430 th->th_counter = timecounter;
1431 th->th_offset_count = ncount;
1432 tc_min_ticktock_freq = max(1, timecounter->tc_frequency /
1433 (((uint64_t)timecounter->tc_counter_mask + 1) / 3));
1435 ffclock_change_tc(th);
1440 * Recalculate the scaling factor. We want the number of 1/2^64
1441 * fractions of a second per period of the hardware counter, taking
1442 * into account the th_adjustment factor which the NTP PLL/adjtime(2)
1443 * processing provides us with.
1445 * The th_adjustment is nanoseconds per second with 32 bit binary
1446 * fraction and we want 64 bit binary fraction of second:
1448 * x = a * 2^32 / 10^9 = a * 4.294967296
1450 * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
1451 * we can only multiply by about 850 without overflowing, that
1452 * leaves no suitably precise fractions for multiply before divide.
1454 * Divide before multiply with a fraction of 2199/512 results in a
1455 * systematic undercompensation of 10PPM of th_adjustment. On a
1456 * 5000PPM adjustment this is a 0.05PPM error. This is acceptable.
1458 * We happily sacrifice the lowest of the 64 bits of our result
1459 * to the goddess of code clarity.
1462 scale = (uint64_t)1 << 63;
1463 scale += (th->th_adjustment / 1024) * 2199;
1464 scale /= th->th_counter->tc_frequency;
1465 th->th_scale = scale * 2;
1468 * Now that the struct timehands is again consistent, set the new
1469 * generation number, making sure to not make it zero.
1473 atomic_store_rel_int(&th->th_generation, ogen);
1475 /* Go live with the new struct timehands. */
1477 switch (sysclock_active) {
1480 time_second = th->th_microtime.tv_sec;
1481 time_uptime = th->th_offset.sec;
1485 time_second = fftimehands->tick_time_lerp.sec;
1486 time_uptime = fftimehands->tick_time_lerp.sec - ffclock_boottime.sec;
1492 timekeep_push_vdso();
1495 /* Report or change the active timecounter hardware. */
1497 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
1500 struct timecounter *newtc, *tc;
1504 strlcpy(newname, tc->tc_name, sizeof(newname));
1506 error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
1507 if (error != 0 || req->newptr == NULL)
1509 /* Record that the tc in use now was specifically chosen. */
1511 if (strcmp(newname, tc->tc_name) == 0)
1513 for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
1514 if (strcmp(newname, newtc->tc_name) != 0)
1517 /* Warm up new timecounter. */
1518 (void)newtc->tc_get_timecount(newtc);
1519 (void)newtc->tc_get_timecount(newtc);
1521 timecounter = newtc;
1524 * The vdso timehands update is deferred until the next
1527 * This is prudent given that 'timekeep_push_vdso()' does not
1528 * use any locking and that it can be called in hard interrupt
1529 * context via 'tc_windup()'.
1536 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW,
1537 0, 0, sysctl_kern_timecounter_hardware, "A",
1538 "Timecounter hardware selected");
1541 /* Report the available timecounter hardware. */
1543 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
1546 struct timecounter *tc;
1549 sbuf_new_for_sysctl(&sb, NULL, 0, req);
1550 for (tc = timecounters; tc != NULL; tc = tc->tc_next) {
1551 if (tc != timecounters)
1552 sbuf_putc(&sb, ' ');
1553 sbuf_printf(&sb, "%s(%d)", tc->tc_name, tc->tc_quality);
1555 error = sbuf_finish(&sb);
1560 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice, CTLTYPE_STRING | CTLFLAG_RD,
1561 0, 0, sysctl_kern_timecounter_choice, "A", "Timecounter hardware detected");
1564 * RFC 2783 PPS-API implementation.
1568 * Return true if the driver is aware of the abi version extensions in the
1569 * pps_state structure, and it supports at least the given abi version number.
1572 abi_aware(struct pps_state *pps, int vers)
1575 return ((pps->kcmode & KCMODE_ABIFLAG) && pps->driver_abi >= vers);
1579 pps_fetch(struct pps_fetch_args *fapi, struct pps_state *pps)
1582 pps_seq_t aseq, cseq;
1585 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1589 * If no timeout is requested, immediately return whatever values were
1590 * most recently captured. If timeout seconds is -1, that's a request
1591 * to block without a timeout. WITNESS won't let us sleep forever
1592 * without a lock (we really don't need a lock), so just repeatedly
1593 * sleep a long time.
1595 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) {
1596 if (fapi->timeout.tv_sec == -1)
1599 tv.tv_sec = fapi->timeout.tv_sec;
1600 tv.tv_usec = fapi->timeout.tv_nsec / 1000;
1603 aseq = atomic_load_int(&pps->ppsinfo.assert_sequence);
1604 cseq = atomic_load_int(&pps->ppsinfo.clear_sequence);
1605 while (aseq == atomic_load_int(&pps->ppsinfo.assert_sequence) &&
1606 cseq == atomic_load_int(&pps->ppsinfo.clear_sequence)) {
1607 if (abi_aware(pps, 1) && pps->driver_mtx != NULL) {
1608 if (pps->flags & PPSFLAG_MTX_SPIN) {
1609 err = msleep_spin(pps, pps->driver_mtx,
1612 err = msleep(pps, pps->driver_mtx, PCATCH,
1616 err = tsleep(pps, PCATCH, "ppsfch", timo);
1618 if (err == EWOULDBLOCK) {
1619 if (fapi->timeout.tv_sec == -1) {
1624 } else if (err != 0) {
1630 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1631 fapi->pps_info_buf = pps->ppsinfo;
1637 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1640 struct pps_fetch_args *fapi;
1642 struct pps_fetch_ffc_args *fapi_ffc;
1645 struct pps_kcbind_args *kapi;
1648 KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
1650 case PPS_IOC_CREATE:
1652 case PPS_IOC_DESTROY:
1654 case PPS_IOC_SETPARAMS:
1655 app = (pps_params_t *)data;
1656 if (app->mode & ~pps->ppscap)
1659 /* Ensure only a single clock is selected for ffc timestamp. */
1660 if ((app->mode & PPS_TSCLK_MASK) == PPS_TSCLK_MASK)
1663 pps->ppsparam = *app;
1665 case PPS_IOC_GETPARAMS:
1666 app = (pps_params_t *)data;
1667 *app = pps->ppsparam;
1668 app->api_version = PPS_API_VERS_1;
1670 case PPS_IOC_GETCAP:
1671 *(int*)data = pps->ppscap;
1674 fapi = (struct pps_fetch_args *)data;
1675 return (pps_fetch(fapi, pps));
1677 case PPS_IOC_FETCH_FFCOUNTER:
1678 fapi_ffc = (struct pps_fetch_ffc_args *)data;
1679 if (fapi_ffc->tsformat && fapi_ffc->tsformat !=
1682 if (fapi_ffc->timeout.tv_sec || fapi_ffc->timeout.tv_nsec)
1683 return (EOPNOTSUPP);
1684 pps->ppsinfo_ffc.current_mode = pps->ppsparam.mode;
1685 fapi_ffc->pps_info_buf_ffc = pps->ppsinfo_ffc;
1686 /* Overwrite timestamps if feedback clock selected. */
1687 switch (pps->ppsparam.mode & PPS_TSCLK_MASK) {
1688 case PPS_TSCLK_FBCK:
1689 fapi_ffc->pps_info_buf_ffc.assert_timestamp =
1690 pps->ppsinfo.assert_timestamp;
1691 fapi_ffc->pps_info_buf_ffc.clear_timestamp =
1692 pps->ppsinfo.clear_timestamp;
1694 case PPS_TSCLK_FFWD:
1700 #endif /* FFCLOCK */
1701 case PPS_IOC_KCBIND:
1703 kapi = (struct pps_kcbind_args *)data;
1704 /* XXX Only root should be able to do this */
1705 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1707 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1709 if (kapi->edge & ~pps->ppscap)
1711 pps->kcmode = (kapi->edge & KCMODE_EDGEMASK) |
1712 (pps->kcmode & KCMODE_ABIFLAG);
1715 return (EOPNOTSUPP);
1723 pps_init(struct pps_state *pps)
1725 pps->ppscap |= PPS_TSFMT_TSPEC | PPS_CANWAIT;
1726 if (pps->ppscap & PPS_CAPTUREASSERT)
1727 pps->ppscap |= PPS_OFFSETASSERT;
1728 if (pps->ppscap & PPS_CAPTURECLEAR)
1729 pps->ppscap |= PPS_OFFSETCLEAR;
1731 pps->ppscap |= PPS_TSCLK_MASK;
1733 pps->kcmode &= ~KCMODE_ABIFLAG;
1737 pps_init_abi(struct pps_state *pps)
1741 if (pps->driver_abi > 0) {
1742 pps->kcmode |= KCMODE_ABIFLAG;
1743 pps->kernel_abi = PPS_ABI_VERSION;
1748 pps_capture(struct pps_state *pps)
1750 struct timehands *th;
1752 KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
1754 pps->capgen = atomic_load_acq_int(&th->th_generation);
1757 pps->capffth = fftimehands;
1759 pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
1760 atomic_thread_fence_acq();
1761 if (pps->capgen != th->th_generation)
1766 pps_event(struct pps_state *pps, int event)
1769 struct timespec ts, *tsp, *osp;
1770 u_int tcount, *pcount;
1774 struct timespec *tsp_ffc;
1775 pps_seq_t *pseq_ffc;
1782 KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
1783 /* Nothing to do if not currently set to capture this event type. */
1784 if ((event & pps->ppsparam.mode) == 0)
1786 /* If the timecounter was wound up underneath us, bail out. */
1787 if (pps->capgen == 0 || pps->capgen !=
1788 atomic_load_acq_int(&pps->capth->th_generation))
1791 /* Things would be easier with arrays. */
1792 if (event == PPS_CAPTUREASSERT) {
1793 tsp = &pps->ppsinfo.assert_timestamp;
1794 osp = &pps->ppsparam.assert_offset;
1795 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1797 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1799 pcount = &pps->ppscount[0];
1800 pseq = &pps->ppsinfo.assert_sequence;
1802 ffcount = &pps->ppsinfo_ffc.assert_ffcount;
1803 tsp_ffc = &pps->ppsinfo_ffc.assert_timestamp;
1804 pseq_ffc = &pps->ppsinfo_ffc.assert_sequence;
1807 tsp = &pps->ppsinfo.clear_timestamp;
1808 osp = &pps->ppsparam.clear_offset;
1809 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1811 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1813 pcount = &pps->ppscount[1];
1814 pseq = &pps->ppsinfo.clear_sequence;
1816 ffcount = &pps->ppsinfo_ffc.clear_ffcount;
1817 tsp_ffc = &pps->ppsinfo_ffc.clear_timestamp;
1818 pseq_ffc = &pps->ppsinfo_ffc.clear_sequence;
1823 * If the timecounter changed, we cannot compare the count values, so
1824 * we have to drop the rest of the PPS-stuff until the next event.
1826 if (pps->ppstc != pps->capth->th_counter) {
1827 pps->ppstc = pps->capth->th_counter;
1828 *pcount = pps->capcount;
1829 pps->ppscount[2] = pps->capcount;
1833 /* Convert the count to a timespec. */
1834 tcount = pps->capcount - pps->capth->th_offset_count;
1835 tcount &= pps->capth->th_counter->tc_counter_mask;
1836 bt = pps->capth->th_bintime;
1837 bintime_addx(&bt, pps->capth->th_scale * tcount);
1838 bintime2timespec(&bt, &ts);
1840 /* If the timecounter was wound up underneath us, bail out. */
1841 atomic_thread_fence_acq();
1842 if (pps->capgen != pps->capth->th_generation)
1845 *pcount = pps->capcount;
1850 timespecadd(tsp, osp, tsp);
1851 if (tsp->tv_nsec < 0) {
1852 tsp->tv_nsec += 1000000000;
1858 *ffcount = pps->capffth->tick_ffcount + tcount;
1859 bt = pps->capffth->tick_time;
1860 ffclock_convert_delta(tcount, pps->capffth->cest.period, &bt);
1861 bintime_add(&bt, &pps->capffth->tick_time);
1862 bintime2timespec(&bt, &ts);
1872 * Feed the NTP PLL/FLL.
1873 * The FLL wants to know how many (hardware) nanoseconds
1874 * elapsed since the previous event.
1876 tcount = pps->capcount - pps->ppscount[2];
1877 pps->ppscount[2] = pps->capcount;
1878 tcount &= pps->capth->th_counter->tc_counter_mask;
1879 scale = (uint64_t)1 << 63;
1880 scale /= pps->capth->th_counter->tc_frequency;
1884 bintime_addx(&bt, scale * tcount);
1885 bintime2timespec(&bt, &ts);
1886 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
1890 /* Wakeup anyone sleeping in pps_fetch(). */
1895 * Timecounters need to be updated every so often to prevent the hardware
1896 * counter from overflowing. Updating also recalculates the cached values
1897 * used by the get*() family of functions, so their precision depends on
1898 * the update frequency.
1902 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0,
1903 "Approximate number of hardclock ticks in a millisecond");
1906 tc_ticktock(int cnt)
1910 if (mtx_trylock_spin(&tc_setclock_mtx)) {
1912 if (count >= tc_tick) {
1916 mtx_unlock_spin(&tc_setclock_mtx);
1920 static void __inline
1921 tc_adjprecision(void)
1925 if (tc_timepercentage > 0) {
1926 t = (99 + tc_timepercentage) / tc_timepercentage;
1927 tc_precexp = fls(t + (t >> 1)) - 1;
1928 FREQ2BT(hz / tc_tick, &bt_timethreshold);
1929 FREQ2BT(hz, &bt_tickthreshold);
1930 bintime_shift(&bt_timethreshold, tc_precexp);
1931 bintime_shift(&bt_tickthreshold, tc_precexp);
1934 bt_timethreshold.sec = INT_MAX;
1935 bt_timethreshold.frac = ~(uint64_t)0;
1936 bt_tickthreshold = bt_timethreshold;
1938 sbt_timethreshold = bttosbt(bt_timethreshold);
1939 sbt_tickthreshold = bttosbt(bt_tickthreshold);
1943 sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS)
1947 val = tc_timepercentage;
1948 error = sysctl_handle_int(oidp, &val, 0, req);
1949 if (error != 0 || req->newptr == NULL)
1951 tc_timepercentage = val;
1960 inittimecounter(void *dummy)
1966 * Set the initial timeout to
1967 * max(1, <approx. number of hardclock ticks in a millisecond>).
1968 * People should probably not use the sysctl to set the timeout
1969 * to smaller than its initial value, since that value is the
1970 * smallest reasonable one. If they want better timestamps they
1971 * should use the non-"get"* functions.
1974 tc_tick = (hz + 500) / 1000;
1978 FREQ2BT(hz, &tick_bt);
1979 tick_sbt = bttosbt(tick_bt);
1980 tick_rate = hz / tc_tick;
1981 FREQ2BT(tick_rate, &tc_tick_bt);
1982 tc_tick_sbt = bttosbt(tc_tick_bt);
1983 p = (tc_tick * 1000000) / hz;
1984 printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
1989 /* warm up new timecounter (again) and get rolling. */
1990 (void)timecounter->tc_get_timecount(timecounter);
1991 (void)timecounter->tc_get_timecount(timecounter);
1992 mtx_lock_spin(&tc_setclock_mtx);
1994 mtx_unlock_spin(&tc_setclock_mtx);
1997 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL);
1999 /* Cpu tick handling -------------------------------------------------*/
2001 static int cpu_tick_variable;
2002 static uint64_t cpu_tick_frequency;
2004 DPCPU_DEFINE_STATIC(uint64_t, tc_cpu_ticks_base);
2005 DPCPU_DEFINE_STATIC(unsigned, tc_cpu_ticks_last);
2010 struct timecounter *tc;
2011 uint64_t res, *base;
2015 base = DPCPU_PTR(tc_cpu_ticks_base);
2016 last = DPCPU_PTR(tc_cpu_ticks_last);
2017 tc = timehands->th_counter;
2018 u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
2020 *base += (uint64_t)tc->tc_counter_mask + 1;
2028 cpu_tick_calibration(void)
2030 static time_t last_calib;
2032 if (time_uptime != last_calib && !(time_uptime & 0xf)) {
2033 cpu_tick_calibrate(0);
2034 last_calib = time_uptime;
2039 * This function gets called every 16 seconds on only one designated
2040 * CPU in the system from hardclock() via cpu_tick_calibration()().
2042 * Whenever the real time clock is stepped we get called with reset=1
2043 * to make sure we handle suspend/resume and similar events correctly.
2047 cpu_tick_calibrate(int reset)
2049 static uint64_t c_last;
2050 uint64_t c_this, c_delta;
2051 static struct bintime t_last;
2052 struct bintime t_this, t_delta;
2056 /* The clock was stepped, abort & reset */
2061 /* we don't calibrate fixed rate cputicks */
2062 if (!cpu_tick_variable)
2065 getbinuptime(&t_this);
2066 c_this = cpu_ticks();
2067 if (t_last.sec != 0) {
2068 c_delta = c_this - c_last;
2070 bintime_sub(&t_delta, &t_last);
2073 * 2^(64-20) / 16[s] =
2075 * 17.592.186.044.416 / 16 =
2076 * 1.099.511.627.776 [Hz]
2078 divi = t_delta.sec << 20;
2079 divi |= t_delta.frac >> (64 - 20);
2082 if (c_delta > cpu_tick_frequency) {
2083 if (0 && bootverbose)
2084 printf("cpu_tick increased to %ju Hz\n",
2086 cpu_tick_frequency = c_delta;
2094 set_cputicker(cpu_tick_f *func, uint64_t freq, unsigned var)
2098 cpu_ticks = tc_cpu_ticks;
2100 cpu_tick_frequency = freq;
2101 cpu_tick_variable = var;
2110 if (cpu_ticks == tc_cpu_ticks)
2111 return (tc_getfrequency());
2112 return (cpu_tick_frequency);
2116 * We need to be slightly careful converting cputicks to microseconds.
2117 * There is plenty of margin in 64 bits of microseconds (half a million
2118 * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
2119 * before divide conversion (to retain precision) we find that the
2120 * margin shrinks to 1.5 hours (one millionth of 146y).
2121 * With a three prong approach we never lose significant bits, no
2122 * matter what the cputick rate and length of timeinterval is.
2126 cputick2usec(uint64_t tick)
2129 if (tick > 18446744073709551LL) /* floor(2^64 / 1000) */
2130 return (tick / (cpu_tickrate() / 1000000LL));
2131 else if (tick > 18446744073709LL) /* floor(2^64 / 1000000) */
2132 return ((tick * 1000LL) / (cpu_tickrate() / 1000LL));
2134 return ((tick * 1000000LL) / cpu_tickrate());
2137 cpu_tick_f *cpu_ticks = tc_cpu_ticks;
2139 static int vdso_th_enable = 1;
2141 sysctl_fast_gettime(SYSCTL_HANDLER_ARGS)
2143 int old_vdso_th_enable, error;
2145 old_vdso_th_enable = vdso_th_enable;
2146 error = sysctl_handle_int(oidp, &old_vdso_th_enable, 0, req);
2149 vdso_th_enable = old_vdso_th_enable;
2152 SYSCTL_PROC(_kern_timecounter, OID_AUTO, fast_gettime,
2153 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
2154 NULL, 0, sysctl_fast_gettime, "I", "Enable fast time of day");
2157 tc_fill_vdso_timehands(struct vdso_timehands *vdso_th)
2159 struct timehands *th;
2163 vdso_th->th_scale = th->th_scale;
2164 vdso_th->th_offset_count = th->th_offset_count;
2165 vdso_th->th_counter_mask = th->th_counter->tc_counter_mask;
2166 vdso_th->th_offset = th->th_offset;
2167 vdso_th->th_boottime = th->th_boottime;
2168 if (th->th_counter->tc_fill_vdso_timehands != NULL) {
2169 enabled = th->th_counter->tc_fill_vdso_timehands(vdso_th,
2173 if (!vdso_th_enable)
2178 #ifdef COMPAT_FREEBSD32
2180 tc_fill_vdso_timehands32(struct vdso_timehands32 *vdso_th32)
2182 struct timehands *th;
2186 *(uint64_t *)&vdso_th32->th_scale[0] = th->th_scale;
2187 vdso_th32->th_offset_count = th->th_offset_count;
2188 vdso_th32->th_counter_mask = th->th_counter->tc_counter_mask;
2189 vdso_th32->th_offset.sec = th->th_offset.sec;
2190 *(uint64_t *)&vdso_th32->th_offset.frac[0] = th->th_offset.frac;
2191 vdso_th32->th_boottime.sec = th->th_boottime.sec;
2192 *(uint64_t *)&vdso_th32->th_boottime.frac[0] = th->th_boottime.frac;
2193 if (th->th_counter->tc_fill_vdso_timehands32 != NULL) {
2194 enabled = th->th_counter->tc_fill_vdso_timehands32(vdso_th32,
2198 if (!vdso_th_enable)