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;
76 u_int th_offset_count;
77 struct bintime th_offset;
78 struct bintime th_bintime;
79 struct timeval th_microtime;
80 struct timespec th_nanotime;
81 struct bintime th_boottime;
82 /* Fields not to be copied in tc_windup start with th_generation. */
84 struct timehands *th_next;
87 static struct timehands ths[16] = {
89 .th_counter = &dummy_timecounter,
90 .th_scale = (uint64_t)-1 / 1000000,
91 .th_large_delta = 1000000,
92 .th_offset = { .sec = 1 },
97 static struct timehands *volatile timehands = &ths[0];
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 static int timehands_count = 2;
118 SYSCTL_INT(_kern_timecounter, OID_AUTO, timehands_count,
119 CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
120 &timehands_count, 0, "Count of timehands in rotation");
122 struct bintime bt_timethreshold;
123 struct bintime bt_tickthreshold;
124 sbintime_t sbt_timethreshold;
125 sbintime_t sbt_tickthreshold;
126 struct bintime tc_tick_bt;
127 sbintime_t tc_tick_sbt;
129 int tc_timepercentage = TC_DEFAULTPERC;
130 static int sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS);
131 SYSCTL_PROC(_kern_timecounter, OID_AUTO, alloweddeviation,
132 CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, 0, 0,
133 sysctl_kern_timecounter_adjprecision, "I",
134 "Allowed time interval deviation in percents");
136 volatile int rtc_generation = 1;
138 static int tc_chosen; /* Non-zero if a specific tc was chosen via sysctl. */
140 static void tc_windup(struct bintime *new_boottimebin);
141 static void cpu_tick_calibrate(int);
143 void dtrace_getnanotime(struct timespec *tsp);
146 sysctl_kern_boottime(SYSCTL_HANDLER_ARGS)
148 struct timeval boottime;
150 getboottime(&boottime);
152 /* i386 is the only arch which uses a 32bits time_t */
157 if (req->flags & SCTL_MASK32) {
158 tv[0] = boottime.tv_sec;
159 tv[1] = boottime.tv_usec;
160 return (SYSCTL_OUT(req, tv, sizeof(tv)));
164 return (SYSCTL_OUT(req, &boottime, sizeof(boottime)));
168 sysctl_kern_timecounter_get(SYSCTL_HANDLER_ARGS)
171 struct timecounter *tc = arg1;
173 ncount = tc->tc_get_timecount(tc);
174 return (sysctl_handle_int(oidp, &ncount, 0, req));
178 sysctl_kern_timecounter_freq(SYSCTL_HANDLER_ARGS)
181 struct timecounter *tc = arg1;
183 freq = tc->tc_frequency;
184 return (sysctl_handle_64(oidp, &freq, 0, req));
188 * Return the difference between the timehands' counter value now and what
189 * was when we copied it to the timehands' offset_count.
191 static __inline u_int
192 tc_delta(struct timehands *th)
194 struct timecounter *tc;
197 return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
198 tc->tc_counter_mask);
202 * Functions for reading the time. We have to loop until we are sure that
203 * the timehands that we operated on was not updated under our feet. See
204 * the comment in <sys/time.h> for a description of these 12 functions.
208 bintime_off(struct bintime *bt, u_int off)
210 struct timehands *th;
213 u_int delta, gen, large_delta;
217 gen = atomic_load_acq_int(&th->th_generation);
218 btp = (struct bintime *)((vm_offset_t)th + off);
220 scale = th->th_scale;
221 delta = tc_delta(th);
222 large_delta = th->th_large_delta;
223 atomic_thread_fence_acq();
224 } while (gen == 0 || gen != th->th_generation);
226 if (__predict_false(delta >= large_delta)) {
227 /* Avoid overflow for scale * delta. */
228 x = (scale >> 32) * delta;
230 bintime_addx(bt, x << 32);
231 bintime_addx(bt, (scale & 0xffffffff) * delta);
233 bintime_addx(bt, scale * delta);
236 #define GETTHBINTIME(dst, member) \
238 _Static_assert(_Generic(((struct timehands *)NULL)->member, \
239 struct bintime: 1, default: 0) == 1, \
240 "struct timehands member is not of struct bintime type"); \
241 bintime_off(dst, __offsetof(struct timehands, member)); \
245 getthmember(void *out, size_t out_size, u_int off)
247 struct timehands *th;
252 gen = atomic_load_acq_int(&th->th_generation);
253 memcpy(out, (char *)th + off, out_size);
254 atomic_thread_fence_acq();
255 } while (gen == 0 || gen != th->th_generation);
257 #define GETTHMEMBER(dst, member) \
259 _Static_assert(_Generic(*dst, \
260 __typeof(((struct timehands *)NULL)->member): 1, \
262 "*dst and struct timehands member have different types"); \
263 getthmember(dst, sizeof(*dst), __offsetof(struct timehands, \
269 fbclock_binuptime(struct bintime *bt)
272 GETTHBINTIME(bt, th_offset);
276 fbclock_nanouptime(struct timespec *tsp)
280 fbclock_binuptime(&bt);
281 bintime2timespec(&bt, tsp);
285 fbclock_microuptime(struct timeval *tvp)
289 fbclock_binuptime(&bt);
290 bintime2timeval(&bt, tvp);
294 fbclock_bintime(struct bintime *bt)
297 GETTHBINTIME(bt, th_bintime);
301 fbclock_nanotime(struct timespec *tsp)
305 fbclock_bintime(&bt);
306 bintime2timespec(&bt, tsp);
310 fbclock_microtime(struct timeval *tvp)
314 fbclock_bintime(&bt);
315 bintime2timeval(&bt, tvp);
319 fbclock_getbinuptime(struct bintime *bt)
322 GETTHMEMBER(bt, th_offset);
326 fbclock_getnanouptime(struct timespec *tsp)
330 GETTHMEMBER(&bt, th_offset);
331 bintime2timespec(&bt, tsp);
335 fbclock_getmicrouptime(struct timeval *tvp)
339 GETTHMEMBER(&bt, th_offset);
340 bintime2timeval(&bt, tvp);
344 fbclock_getbintime(struct bintime *bt)
347 GETTHMEMBER(bt, th_bintime);
351 fbclock_getnanotime(struct timespec *tsp)
354 GETTHMEMBER(tsp, th_nanotime);
358 fbclock_getmicrotime(struct timeval *tvp)
361 GETTHMEMBER(tvp, th_microtime);
366 binuptime(struct bintime *bt)
369 GETTHBINTIME(bt, th_offset);
373 nanouptime(struct timespec *tsp)
378 bintime2timespec(&bt, tsp);
382 microuptime(struct timeval *tvp)
387 bintime2timeval(&bt, tvp);
391 bintime(struct bintime *bt)
394 GETTHBINTIME(bt, th_bintime);
398 nanotime(struct timespec *tsp)
403 bintime2timespec(&bt, tsp);
407 microtime(struct timeval *tvp)
412 bintime2timeval(&bt, tvp);
416 getbinuptime(struct bintime *bt)
419 GETTHMEMBER(bt, th_offset);
423 getnanouptime(struct timespec *tsp)
427 GETTHMEMBER(&bt, th_offset);
428 bintime2timespec(&bt, tsp);
432 getmicrouptime(struct timeval *tvp)
436 GETTHMEMBER(&bt, th_offset);
437 bintime2timeval(&bt, tvp);
441 getbintime(struct bintime *bt)
444 GETTHMEMBER(bt, th_bintime);
448 getnanotime(struct timespec *tsp)
451 GETTHMEMBER(tsp, th_nanotime);
455 getmicrotime(struct timeval *tvp)
458 GETTHMEMBER(tvp, th_microtime);
463 getboottime(struct timeval *boottime)
465 struct bintime boottimebin;
467 getboottimebin(&boottimebin);
468 bintime2timeval(&boottimebin, boottime);
472 getboottimebin(struct bintime *boottimebin)
475 GETTHMEMBER(boottimebin, th_boottime);
480 * Support for feed-forward synchronization algorithms. This is heavily inspired
481 * by the timehands mechanism but kept independent from it. *_windup() functions
482 * have some connection to avoid accessing the timecounter hardware more than
486 /* Feed-forward clock estimates kept updated by the synchronization daemon. */
487 struct ffclock_estimate ffclock_estimate;
488 struct bintime ffclock_boottime; /* Feed-forward boot time estimate. */
489 uint32_t ffclock_status; /* Feed-forward clock status. */
490 int8_t ffclock_updated; /* New estimates are available. */
491 struct mtx ffclock_mtx; /* Mutex on ffclock_estimate. */
494 struct ffclock_estimate cest;
495 struct bintime tick_time;
496 struct bintime tick_time_lerp;
497 ffcounter tick_ffcount;
498 uint64_t period_lerp;
499 volatile uint8_t gen;
500 struct fftimehands *next;
503 #define NUM_ELEMENTS(x) (sizeof(x) / sizeof(*x))
505 static struct fftimehands ffth[10];
506 static struct fftimehands *volatile fftimehands = ffth;
511 struct fftimehands *cur;
512 struct fftimehands *last;
514 memset(ffth, 0, sizeof(ffth));
516 last = ffth + NUM_ELEMENTS(ffth) - 1;
517 for (cur = ffth; cur < last; cur++)
522 ffclock_status = FFCLOCK_STA_UNSYNC;
523 mtx_init(&ffclock_mtx, "ffclock lock", NULL, MTX_DEF);
527 * Reset the feed-forward clock estimates. Called from inittodr() to get things
528 * kick started and uses the timecounter nominal frequency as a first period
529 * estimate. Note: this function may be called several time just after boot.
530 * Note: this is the only function that sets the value of boot time for the
531 * monotonic (i.e. uptime) version of the feed-forward clock.
534 ffclock_reset_clock(struct timespec *ts)
536 struct timecounter *tc;
537 struct ffclock_estimate cest;
539 tc = timehands->th_counter;
540 memset(&cest, 0, sizeof(struct ffclock_estimate));
542 timespec2bintime(ts, &ffclock_boottime);
543 timespec2bintime(ts, &(cest.update_time));
544 ffclock_read_counter(&cest.update_ffcount);
545 cest.leapsec_next = 0;
546 cest.period = ((1ULL << 63) / tc->tc_frequency) << 1;
549 cest.status = FFCLOCK_STA_UNSYNC;
550 cest.leapsec_total = 0;
553 mtx_lock(&ffclock_mtx);
554 bcopy(&cest, &ffclock_estimate, sizeof(struct ffclock_estimate));
555 ffclock_updated = INT8_MAX;
556 mtx_unlock(&ffclock_mtx);
558 printf("ffclock reset: %s (%llu Hz), time = %ld.%09lu\n", tc->tc_name,
559 (unsigned long long)tc->tc_frequency, (long)ts->tv_sec,
560 (unsigned long)ts->tv_nsec);
564 * Sub-routine to convert a time interval measured in RAW counter units to time
565 * in seconds stored in bintime format.
566 * NOTE: bintime_mul requires u_int, but the value of the ffcounter may be
567 * larger than the max value of u_int (on 32 bit architecture). Loop to consume
571 ffclock_convert_delta(ffcounter ffdelta, uint64_t period, struct bintime *bt)
574 ffcounter delta, delta_max;
576 delta_max = (1ULL << (8 * sizeof(unsigned int))) - 1;
579 if (ffdelta > delta_max)
585 bintime_mul(&bt2, (unsigned int)delta);
586 bintime_add(bt, &bt2);
588 } while (ffdelta > 0);
592 * Update the fftimehands.
593 * Push the tick ffcount and time(s) forward based on current clock estimate.
594 * The conversion from ffcounter to bintime relies on the difference clock
595 * principle, whose accuracy relies on computing small time intervals. If a new
596 * clock estimate has been passed by the synchronisation daemon, make it
597 * current, and compute the linear interpolation for monotonic time if needed.
600 ffclock_windup(unsigned int delta)
602 struct ffclock_estimate *cest;
603 struct fftimehands *ffth;
604 struct bintime bt, gap_lerp;
607 unsigned int polling;
608 uint8_t forward_jump, ogen;
611 * Pick the next timehand, copy current ffclock estimates and move tick
612 * times and counter forward.
615 ffth = fftimehands->next;
619 bcopy(&fftimehands->cest, cest, sizeof(struct ffclock_estimate));
620 ffdelta = (ffcounter)delta;
621 ffth->period_lerp = fftimehands->period_lerp;
623 ffth->tick_time = fftimehands->tick_time;
624 ffclock_convert_delta(ffdelta, cest->period, &bt);
625 bintime_add(&ffth->tick_time, &bt);
627 ffth->tick_time_lerp = fftimehands->tick_time_lerp;
628 ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt);
629 bintime_add(&ffth->tick_time_lerp, &bt);
631 ffth->tick_ffcount = fftimehands->tick_ffcount + ffdelta;
634 * Assess the status of the clock, if the last update is too old, it is
635 * likely the synchronisation daemon is dead and the clock is free
638 if (ffclock_updated == 0) {
639 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
640 ffclock_convert_delta(ffdelta, cest->period, &bt);
641 if (bt.sec > 2 * FFCLOCK_SKM_SCALE)
642 ffclock_status |= FFCLOCK_STA_UNSYNC;
646 * If available, grab updated clock estimates and make them current.
647 * Recompute time at this tick using the updated estimates. The clock
648 * estimates passed the feed-forward synchronisation daemon may result
649 * in time conversion that is not monotonically increasing (just after
650 * the update). time_lerp is a particular linear interpolation over the
651 * synchronisation algo polling period that ensures monotonicity for the
652 * clock ids requesting it.
654 if (ffclock_updated > 0) {
655 bcopy(&ffclock_estimate, cest, sizeof(struct ffclock_estimate));
656 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
657 ffth->tick_time = cest->update_time;
658 ffclock_convert_delta(ffdelta, cest->period, &bt);
659 bintime_add(&ffth->tick_time, &bt);
661 /* ffclock_reset sets ffclock_updated to INT8_MAX */
662 if (ffclock_updated == INT8_MAX)
663 ffth->tick_time_lerp = ffth->tick_time;
665 if (bintime_cmp(&ffth->tick_time, &ffth->tick_time_lerp, >))
670 bintime_clear(&gap_lerp);
672 gap_lerp = ffth->tick_time;
673 bintime_sub(&gap_lerp, &ffth->tick_time_lerp);
675 gap_lerp = ffth->tick_time_lerp;
676 bintime_sub(&gap_lerp, &ffth->tick_time);
680 * The reset from the RTC clock may be far from accurate, and
681 * reducing the gap between real time and interpolated time
682 * could take a very long time if the interpolated clock insists
683 * on strict monotonicity. The clock is reset under very strict
684 * conditions (kernel time is known to be wrong and
685 * synchronization daemon has been restarted recently.
686 * ffclock_boottime absorbs the jump to ensure boot time is
687 * correct and uptime functions stay consistent.
689 if (((ffclock_status & FFCLOCK_STA_UNSYNC) == FFCLOCK_STA_UNSYNC) &&
690 ((cest->status & FFCLOCK_STA_UNSYNC) == 0) &&
691 ((cest->status & FFCLOCK_STA_WARMUP) == FFCLOCK_STA_WARMUP)) {
693 bintime_add(&ffclock_boottime, &gap_lerp);
695 bintime_sub(&ffclock_boottime, &gap_lerp);
696 ffth->tick_time_lerp = ffth->tick_time;
697 bintime_clear(&gap_lerp);
700 ffclock_status = cest->status;
701 ffth->period_lerp = cest->period;
704 * Compute corrected period used for the linear interpolation of
705 * time. The rate of linear interpolation is capped to 5000PPM
708 if (bintime_isset(&gap_lerp)) {
709 ffdelta = cest->update_ffcount;
710 ffdelta -= fftimehands->cest.update_ffcount;
711 ffclock_convert_delta(ffdelta, cest->period, &bt);
714 bt.frac = 5000000 * (uint64_t)18446744073LL;
715 bintime_mul(&bt, polling);
716 if (bintime_cmp(&gap_lerp, &bt, >))
719 /* Approximate 1 sec by 1-(1/2^64) to ease arithmetic */
721 if (gap_lerp.sec > 0) {
723 frac /= ffdelta / gap_lerp.sec;
725 frac += gap_lerp.frac / ffdelta;
728 ffth->period_lerp += frac;
730 ffth->period_lerp -= frac;
742 * Adjust the fftimehands when the timecounter is changed. Stating the obvious,
743 * the old and new hardware counter cannot be read simultaneously. tc_windup()
744 * does read the two counters 'back to back', but a few cycles are effectively
745 * lost, and not accumulated in tick_ffcount. This is a fairly radical
746 * operation for a feed-forward synchronization daemon, and it is its job to not
747 * pushing irrelevant data to the kernel. Because there is no locking here,
748 * simply force to ignore pending or next update to give daemon a chance to
749 * realize the counter has changed.
752 ffclock_change_tc(struct timehands *th)
754 struct fftimehands *ffth;
755 struct ffclock_estimate *cest;
756 struct timecounter *tc;
760 ffth = fftimehands->next;
765 bcopy(&(fftimehands->cest), cest, sizeof(struct ffclock_estimate));
766 cest->period = ((1ULL << 63) / tc->tc_frequency ) << 1;
769 cest->status |= FFCLOCK_STA_UNSYNC;
771 ffth->tick_ffcount = fftimehands->tick_ffcount;
772 ffth->tick_time_lerp = fftimehands->tick_time_lerp;
773 ffth->tick_time = fftimehands->tick_time;
774 ffth->period_lerp = cest->period;
776 /* Do not lock but ignore next update from synchronization daemon. */
786 * Retrieve feed-forward counter and time of last kernel tick.
789 ffclock_last_tick(ffcounter *ffcount, struct bintime *bt, uint32_t flags)
791 struct fftimehands *ffth;
795 * No locking but check generation has not changed. Also need to make
796 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
801 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP)
802 *bt = ffth->tick_time_lerp;
804 *bt = ffth->tick_time;
805 *ffcount = ffth->tick_ffcount;
806 } while (gen == 0 || gen != ffth->gen);
810 * Absolute clock conversion. Low level function to convert ffcounter to
811 * bintime. The ffcounter is converted using the current ffclock period estimate
812 * or the "interpolated period" to ensure monotonicity.
813 * NOTE: this conversion may have been deferred, and the clock updated since the
814 * hardware counter has been read.
817 ffclock_convert_abs(ffcounter ffcount, struct bintime *bt, uint32_t flags)
819 struct fftimehands *ffth;
825 * No locking but check generation has not changed. Also need to make
826 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
831 if (ffcount > ffth->tick_ffcount)
832 ffdelta = ffcount - ffth->tick_ffcount;
834 ffdelta = ffth->tick_ffcount - ffcount;
836 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP) {
837 *bt = ffth->tick_time_lerp;
838 ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt2);
840 *bt = ffth->tick_time;
841 ffclock_convert_delta(ffdelta, ffth->cest.period, &bt2);
844 if (ffcount > ffth->tick_ffcount)
845 bintime_add(bt, &bt2);
847 bintime_sub(bt, &bt2);
848 } while (gen == 0 || gen != ffth->gen);
852 * Difference clock conversion.
853 * Low level function to Convert a time interval measured in RAW counter units
854 * into bintime. The difference clock allows measuring small intervals much more
855 * reliably than the absolute clock.
858 ffclock_convert_diff(ffcounter ffdelta, struct bintime *bt)
860 struct fftimehands *ffth;
863 /* No locking but check generation has not changed. */
867 ffclock_convert_delta(ffdelta, ffth->cest.period, bt);
868 } while (gen == 0 || gen != ffth->gen);
872 * Access to current ffcounter value.
875 ffclock_read_counter(ffcounter *ffcount)
877 struct timehands *th;
878 struct fftimehands *ffth;
879 unsigned int gen, delta;
882 * ffclock_windup() called from tc_windup(), safe to rely on
883 * th->th_generation only, for correct delta and ffcounter.
887 gen = atomic_load_acq_int(&th->th_generation);
889 delta = tc_delta(th);
890 *ffcount = ffth->tick_ffcount;
891 atomic_thread_fence_acq();
892 } while (gen == 0 || gen != th->th_generation);
898 binuptime(struct bintime *bt)
901 binuptime_fromclock(bt, sysclock_active);
905 nanouptime(struct timespec *tsp)
908 nanouptime_fromclock(tsp, sysclock_active);
912 microuptime(struct timeval *tvp)
915 microuptime_fromclock(tvp, sysclock_active);
919 bintime(struct bintime *bt)
922 bintime_fromclock(bt, sysclock_active);
926 nanotime(struct timespec *tsp)
929 nanotime_fromclock(tsp, sysclock_active);
933 microtime(struct timeval *tvp)
936 microtime_fromclock(tvp, sysclock_active);
940 getbinuptime(struct bintime *bt)
943 getbinuptime_fromclock(bt, sysclock_active);
947 getnanouptime(struct timespec *tsp)
950 getnanouptime_fromclock(tsp, sysclock_active);
954 getmicrouptime(struct timeval *tvp)
957 getmicrouptime_fromclock(tvp, sysclock_active);
961 getbintime(struct bintime *bt)
964 getbintime_fromclock(bt, sysclock_active);
968 getnanotime(struct timespec *tsp)
971 getnanotime_fromclock(tsp, sysclock_active);
975 getmicrotime(struct timeval *tvp)
978 getmicrouptime_fromclock(tvp, sysclock_active);
984 * This is a clone of getnanotime and used for walltimestamps.
985 * The dtrace_ prefix prevents fbt from creating probes for
986 * it so walltimestamp can be safely used in all fbt probes.
989 dtrace_getnanotime(struct timespec *tsp)
992 GETTHMEMBER(tsp, th_nanotime);
996 * System clock currently providing time to the system. Modifiable via sysctl
997 * when the FFCLOCK option is defined.
999 int sysclock_active = SYSCLOCK_FBCK;
1001 /* Internal NTP status and error estimates. */
1002 extern int time_status;
1003 extern long time_esterror;
1006 * Take a snapshot of sysclock data which can be used to compare system clocks
1007 * and generate timestamps after the fact.
1010 sysclock_getsnapshot(struct sysclock_snap *clock_snap, int fast)
1012 struct fbclock_info *fbi;
1013 struct timehands *th;
1015 unsigned int delta, gen;
1018 struct fftimehands *ffth;
1019 struct ffclock_info *ffi;
1020 struct ffclock_estimate cest;
1022 ffi = &clock_snap->ff_info;
1025 fbi = &clock_snap->fb_info;
1030 gen = atomic_load_acq_int(&th->th_generation);
1031 fbi->th_scale = th->th_scale;
1032 fbi->tick_time = th->th_offset;
1035 ffi->tick_time = ffth->tick_time_lerp;
1036 ffi->tick_time_lerp = ffth->tick_time_lerp;
1037 ffi->period = ffth->cest.period;
1038 ffi->period_lerp = ffth->period_lerp;
1039 clock_snap->ffcount = ffth->tick_ffcount;
1043 delta = tc_delta(th);
1044 atomic_thread_fence_acq();
1045 } while (gen == 0 || gen != th->th_generation);
1047 clock_snap->delta = delta;
1048 clock_snap->sysclock_active = sysclock_active;
1050 /* Record feedback clock status and error. */
1051 clock_snap->fb_info.status = time_status;
1052 /* XXX: Very crude estimate of feedback clock error. */
1053 bt.sec = time_esterror / 1000000;
1054 bt.frac = ((time_esterror - bt.sec) * 1000000) *
1055 (uint64_t)18446744073709ULL;
1056 clock_snap->fb_info.error = bt;
1060 clock_snap->ffcount += delta;
1062 /* Record feed-forward clock leap second adjustment. */
1063 ffi->leapsec_adjustment = cest.leapsec_total;
1064 if (clock_snap->ffcount > cest.leapsec_next)
1065 ffi->leapsec_adjustment -= cest.leapsec;
1067 /* Record feed-forward clock status and error. */
1068 clock_snap->ff_info.status = cest.status;
1069 ffcount = clock_snap->ffcount - cest.update_ffcount;
1070 ffclock_convert_delta(ffcount, cest.period, &bt);
1071 /* 18446744073709 = int(2^64/1e12), err_bound_rate in [ps/s]. */
1072 bintime_mul(&bt, cest.errb_rate * (uint64_t)18446744073709ULL);
1073 /* 18446744073 = int(2^64 / 1e9), since err_abs in [ns]. */
1074 bintime_addx(&bt, cest.errb_abs * (uint64_t)18446744073ULL);
1075 clock_snap->ff_info.error = bt;
1080 * Convert a sysclock snapshot into a struct bintime based on the specified
1081 * clock source and flags.
1084 sysclock_snap2bintime(struct sysclock_snap *cs, struct bintime *bt,
1085 int whichclock, uint32_t flags)
1087 struct bintime boottimebin;
1093 switch (whichclock) {
1095 *bt = cs->fb_info.tick_time;
1097 /* If snapshot was created with !fast, delta will be >0. */
1099 bintime_addx(bt, cs->fb_info.th_scale * cs->delta);
1101 if ((flags & FBCLOCK_UPTIME) == 0) {
1102 getboottimebin(&boottimebin);
1103 bintime_add(bt, &boottimebin);
1108 if (flags & FFCLOCK_LERP) {
1109 *bt = cs->ff_info.tick_time_lerp;
1110 period = cs->ff_info.period_lerp;
1112 *bt = cs->ff_info.tick_time;
1113 period = cs->ff_info.period;
1116 /* If snapshot was created with !fast, delta will be >0. */
1117 if (cs->delta > 0) {
1118 ffclock_convert_delta(cs->delta, period, &bt2);
1119 bintime_add(bt, &bt2);
1122 /* Leap second adjustment. */
1123 if (flags & FFCLOCK_LEAPSEC)
1124 bt->sec -= cs->ff_info.leapsec_adjustment;
1126 /* Boot time adjustment, for uptime/monotonic clocks. */
1127 if (flags & FFCLOCK_UPTIME)
1128 bintime_sub(bt, &ffclock_boottime);
1140 * Initialize a new timecounter and possibly use it.
1143 tc_init(struct timecounter *tc)
1146 struct sysctl_oid *tc_root;
1148 u = tc->tc_frequency / tc->tc_counter_mask;
1149 /* XXX: We need some margin here, 10% is a guess */
1152 if (u > hz && tc->tc_quality >= 0) {
1153 tc->tc_quality = -2000;
1155 printf("Timecounter \"%s\" frequency %ju Hz",
1156 tc->tc_name, (uintmax_t)tc->tc_frequency);
1157 printf(" -- Insufficient hz, needs at least %u\n", u);
1159 } else if (tc->tc_quality >= 0 || bootverbose) {
1160 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
1161 tc->tc_name, (uintmax_t)tc->tc_frequency,
1165 tc->tc_next = timecounters;
1168 * Set up sysctl tree for this counter.
1170 tc_root = SYSCTL_ADD_NODE_WITH_LABEL(NULL,
1171 SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name,
1172 CTLFLAG_RW, 0, "timecounter description", "timecounter");
1173 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1174 "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0,
1175 "mask for implemented bits");
1176 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1177 "counter", CTLTYPE_UINT | CTLFLAG_RD, tc, sizeof(*tc),
1178 sysctl_kern_timecounter_get, "IU", "current timecounter value");
1179 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1180 "frequency", CTLTYPE_U64 | CTLFLAG_RD, tc, sizeof(*tc),
1181 sysctl_kern_timecounter_freq, "QU", "timecounter frequency");
1182 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1183 "quality", CTLFLAG_RD, &(tc->tc_quality), 0,
1184 "goodness of time counter");
1186 * Do not automatically switch if the current tc was specifically
1187 * chosen. Never automatically use a timecounter with negative quality.
1188 * Even though we run on the dummy counter, switching here may be
1189 * worse since this timecounter may not be monotonic.
1193 if (tc->tc_quality < 0)
1195 if (tc->tc_quality < timecounter->tc_quality)
1197 if (tc->tc_quality == timecounter->tc_quality &&
1198 tc->tc_frequency < timecounter->tc_frequency)
1200 (void)tc->tc_get_timecount(tc);
1201 (void)tc->tc_get_timecount(tc);
1205 /* Report the frequency of the current timecounter. */
1207 tc_getfrequency(void)
1210 return (timehands->th_counter->tc_frequency);
1214 sleeping_on_old_rtc(struct thread *td)
1218 * td_rtcgen is modified by curthread when it is running,
1219 * and by other threads in this function. By finding the thread
1220 * on a sleepqueue and holding the lock on the sleepqueue
1221 * chain, we guarantee that the thread is not running and that
1222 * modifying td_rtcgen is safe. Setting td_rtcgen to zero informs
1223 * the thread that it was woken due to a real-time clock adjustment.
1224 * (The declaration of td_rtcgen refers to this comment.)
1226 if (td->td_rtcgen != 0 && td->td_rtcgen != rtc_generation) {
1233 static struct mtx tc_setclock_mtx;
1234 MTX_SYSINIT(tc_setclock_init, &tc_setclock_mtx, "tcsetc", MTX_SPIN);
1237 * Step our concept of UTC. This is done by modifying our estimate of
1241 tc_setclock(struct timespec *ts)
1243 struct timespec tbef, taft;
1244 struct bintime bt, bt2;
1246 timespec2bintime(ts, &bt);
1248 mtx_lock_spin(&tc_setclock_mtx);
1249 cpu_tick_calibrate(1);
1251 bintime_sub(&bt, &bt2);
1253 /* XXX fiddle all the little crinkly bits around the fiords... */
1255 mtx_unlock_spin(&tc_setclock_mtx);
1257 /* Avoid rtc_generation == 0, since td_rtcgen == 0 is special. */
1258 atomic_add_rel_int(&rtc_generation, 2);
1259 sleepq_chains_remove_matching(sleeping_on_old_rtc);
1260 if (timestepwarnings) {
1263 "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
1264 (intmax_t)tbef.tv_sec, tbef.tv_nsec,
1265 (intmax_t)taft.tv_sec, taft.tv_nsec,
1266 (intmax_t)ts->tv_sec, ts->tv_nsec);
1271 * Initialize the next struct timehands in the ring and make
1272 * it the active timehands. Along the way we might switch to a different
1273 * timecounter and/or do seconds processing in NTP. Slightly magic.
1276 tc_windup(struct bintime *new_boottimebin)
1279 struct timehands *th, *tho;
1281 u_int delta, ncount, ogen;
1286 * Make the next timehands a copy of the current one, but do
1287 * not overwrite the generation or next pointer. While we
1288 * update the contents, the generation must be zero. We need
1289 * to ensure that the zero generation is visible before the
1290 * data updates become visible, which requires release fence.
1291 * For similar reasons, re-reading of the generation after the
1292 * data is read should use acquire fence.
1296 ogen = th->th_generation;
1297 th->th_generation = 0;
1298 atomic_thread_fence_rel();
1299 memcpy(th, tho, offsetof(struct timehands, th_generation));
1300 if (new_boottimebin != NULL)
1301 th->th_boottime = *new_boottimebin;
1304 * Capture a timecounter delta on the current timecounter and if
1305 * changing timecounters, a counter value from the new timecounter.
1306 * Update the offset fields accordingly.
1308 delta = tc_delta(th);
1309 if (th->th_counter != timecounter)
1310 ncount = timecounter->tc_get_timecount(timecounter);
1314 ffclock_windup(delta);
1316 th->th_offset_count += delta;
1317 th->th_offset_count &= th->th_counter->tc_counter_mask;
1318 while (delta > th->th_counter->tc_frequency) {
1319 /* Eat complete unadjusted seconds. */
1320 delta -= th->th_counter->tc_frequency;
1321 th->th_offset.sec++;
1323 if ((delta > th->th_counter->tc_frequency / 2) &&
1324 (th->th_scale * delta < ((uint64_t)1 << 63))) {
1325 /* The product th_scale * delta just barely overflows. */
1326 th->th_offset.sec++;
1328 bintime_addx(&th->th_offset, th->th_scale * delta);
1331 * Hardware latching timecounters may not generate interrupts on
1332 * PPS events, so instead we poll them. There is a finite risk that
1333 * the hardware might capture a count which is later than the one we
1334 * got above, and therefore possibly in the next NTP second which might
1335 * have a different rate than the current NTP second. It doesn't
1336 * matter in practice.
1338 if (tho->th_counter->tc_poll_pps)
1339 tho->th_counter->tc_poll_pps(tho->th_counter);
1342 * Deal with NTP second processing. The for loop normally
1343 * iterates at most once, but in extreme situations it might
1344 * keep NTP sane if timeouts are not run for several seconds.
1345 * At boot, the time step can be large when the TOD hardware
1346 * has been read, so on really large steps, we call
1347 * ntp_update_second only twice. We need to call it twice in
1348 * case we missed a leap second.
1351 bintime_add(&bt, &th->th_boottime);
1352 i = bt.sec - tho->th_microtime.tv_sec;
1355 for (; i > 0; i--) {
1357 ntp_update_second(&th->th_adjustment, &bt.sec);
1359 th->th_boottime.sec += bt.sec - t;
1361 /* Update the UTC timestamps used by the get*() functions. */
1362 th->th_bintime = bt;
1363 bintime2timeval(&bt, &th->th_microtime);
1364 bintime2timespec(&bt, &th->th_nanotime);
1366 /* Now is a good time to change timecounters. */
1367 if (th->th_counter != timecounter) {
1369 if ((timecounter->tc_flags & TC_FLAGS_C2STOP) != 0)
1370 cpu_disable_c2_sleep++;
1371 if ((th->th_counter->tc_flags & TC_FLAGS_C2STOP) != 0)
1372 cpu_disable_c2_sleep--;
1374 th->th_counter = timecounter;
1375 th->th_offset_count = ncount;
1376 tc_min_ticktock_freq = max(1, timecounter->tc_frequency /
1377 (((uint64_t)timecounter->tc_counter_mask + 1) / 3));
1379 ffclock_change_tc(th);
1384 * Recalculate the scaling factor. We want the number of 1/2^64
1385 * fractions of a second per period of the hardware counter, taking
1386 * into account the th_adjustment factor which the NTP PLL/adjtime(2)
1387 * processing provides us with.
1389 * The th_adjustment is nanoseconds per second with 32 bit binary
1390 * fraction and we want 64 bit binary fraction of second:
1392 * x = a * 2^32 / 10^9 = a * 4.294967296
1394 * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
1395 * we can only multiply by about 850 without overflowing, that
1396 * leaves no suitably precise fractions for multiply before divide.
1398 * Divide before multiply with a fraction of 2199/512 results in a
1399 * systematic undercompensation of 10PPM of th_adjustment. On a
1400 * 5000PPM adjustment this is a 0.05PPM error. This is acceptable.
1402 * We happily sacrifice the lowest of the 64 bits of our result
1403 * to the goddess of code clarity.
1406 scale = (uint64_t)1 << 63;
1407 scale += (th->th_adjustment / 1024) * 2199;
1408 scale /= th->th_counter->tc_frequency;
1409 th->th_scale = scale * 2;
1410 th->th_large_delta = MIN(((uint64_t)1 << 63) / scale, UINT_MAX);
1413 * Now that the struct timehands is again consistent, set the new
1414 * generation number, making sure to not make it zero.
1418 atomic_store_rel_int(&th->th_generation, ogen);
1420 /* Go live with the new struct timehands. */
1422 switch (sysclock_active) {
1425 time_second = th->th_microtime.tv_sec;
1426 time_uptime = th->th_offset.sec;
1430 time_second = fftimehands->tick_time_lerp.sec;
1431 time_uptime = fftimehands->tick_time_lerp.sec - ffclock_boottime.sec;
1437 timekeep_push_vdso();
1440 /* Report or change the active timecounter hardware. */
1442 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
1445 struct timecounter *newtc, *tc;
1449 strlcpy(newname, tc->tc_name, sizeof(newname));
1451 error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
1452 if (error != 0 || req->newptr == NULL)
1454 /* Record that the tc in use now was specifically chosen. */
1456 if (strcmp(newname, tc->tc_name) == 0)
1458 for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
1459 if (strcmp(newname, newtc->tc_name) != 0)
1462 /* Warm up new timecounter. */
1463 (void)newtc->tc_get_timecount(newtc);
1464 (void)newtc->tc_get_timecount(newtc);
1466 timecounter = newtc;
1469 * The vdso timehands update is deferred until the next
1472 * This is prudent given that 'timekeep_push_vdso()' does not
1473 * use any locking and that it can be called in hard interrupt
1474 * context via 'tc_windup()'.
1481 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW,
1482 0, 0, sysctl_kern_timecounter_hardware, "A",
1483 "Timecounter hardware selected");
1485 /* Report the available timecounter hardware. */
1487 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
1490 struct timecounter *tc;
1493 sbuf_new_for_sysctl(&sb, NULL, 0, req);
1494 for (tc = timecounters; tc != NULL; tc = tc->tc_next) {
1495 if (tc != timecounters)
1496 sbuf_putc(&sb, ' ');
1497 sbuf_printf(&sb, "%s(%d)", tc->tc_name, tc->tc_quality);
1499 error = sbuf_finish(&sb);
1504 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice, CTLTYPE_STRING | CTLFLAG_RD,
1505 0, 0, sysctl_kern_timecounter_choice, "A", "Timecounter hardware detected");
1508 * RFC 2783 PPS-API implementation.
1512 * Return true if the driver is aware of the abi version extensions in the
1513 * pps_state structure, and it supports at least the given abi version number.
1516 abi_aware(struct pps_state *pps, int vers)
1519 return ((pps->kcmode & KCMODE_ABIFLAG) && pps->driver_abi >= vers);
1523 pps_fetch(struct pps_fetch_args *fapi, struct pps_state *pps)
1526 pps_seq_t aseq, cseq;
1529 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1533 * If no timeout is requested, immediately return whatever values were
1534 * most recently captured. If timeout seconds is -1, that's a request
1535 * to block without a timeout. WITNESS won't let us sleep forever
1536 * without a lock (we really don't need a lock), so just repeatedly
1537 * sleep a long time.
1539 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) {
1540 if (fapi->timeout.tv_sec == -1)
1543 tv.tv_sec = fapi->timeout.tv_sec;
1544 tv.tv_usec = fapi->timeout.tv_nsec / 1000;
1547 aseq = atomic_load_int(&pps->ppsinfo.assert_sequence);
1548 cseq = atomic_load_int(&pps->ppsinfo.clear_sequence);
1549 while (aseq == atomic_load_int(&pps->ppsinfo.assert_sequence) &&
1550 cseq == atomic_load_int(&pps->ppsinfo.clear_sequence)) {
1551 if (abi_aware(pps, 1) && pps->driver_mtx != NULL) {
1552 if (pps->flags & PPSFLAG_MTX_SPIN) {
1553 err = msleep_spin(pps, pps->driver_mtx,
1556 err = msleep(pps, pps->driver_mtx, PCATCH,
1560 err = tsleep(pps, PCATCH, "ppsfch", timo);
1562 if (err == EWOULDBLOCK) {
1563 if (fapi->timeout.tv_sec == -1) {
1568 } else if (err != 0) {
1574 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1575 fapi->pps_info_buf = pps->ppsinfo;
1581 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1584 struct pps_fetch_args *fapi;
1586 struct pps_fetch_ffc_args *fapi_ffc;
1589 struct pps_kcbind_args *kapi;
1592 KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
1594 case PPS_IOC_CREATE:
1596 case PPS_IOC_DESTROY:
1598 case PPS_IOC_SETPARAMS:
1599 app = (pps_params_t *)data;
1600 if (app->mode & ~pps->ppscap)
1603 /* Ensure only a single clock is selected for ffc timestamp. */
1604 if ((app->mode & PPS_TSCLK_MASK) == PPS_TSCLK_MASK)
1607 pps->ppsparam = *app;
1609 case PPS_IOC_GETPARAMS:
1610 app = (pps_params_t *)data;
1611 *app = pps->ppsparam;
1612 app->api_version = PPS_API_VERS_1;
1614 case PPS_IOC_GETCAP:
1615 *(int*)data = pps->ppscap;
1618 fapi = (struct pps_fetch_args *)data;
1619 return (pps_fetch(fapi, pps));
1621 case PPS_IOC_FETCH_FFCOUNTER:
1622 fapi_ffc = (struct pps_fetch_ffc_args *)data;
1623 if (fapi_ffc->tsformat && fapi_ffc->tsformat !=
1626 if (fapi_ffc->timeout.tv_sec || fapi_ffc->timeout.tv_nsec)
1627 return (EOPNOTSUPP);
1628 pps->ppsinfo_ffc.current_mode = pps->ppsparam.mode;
1629 fapi_ffc->pps_info_buf_ffc = pps->ppsinfo_ffc;
1630 /* Overwrite timestamps if feedback clock selected. */
1631 switch (pps->ppsparam.mode & PPS_TSCLK_MASK) {
1632 case PPS_TSCLK_FBCK:
1633 fapi_ffc->pps_info_buf_ffc.assert_timestamp =
1634 pps->ppsinfo.assert_timestamp;
1635 fapi_ffc->pps_info_buf_ffc.clear_timestamp =
1636 pps->ppsinfo.clear_timestamp;
1638 case PPS_TSCLK_FFWD:
1644 #endif /* FFCLOCK */
1645 case PPS_IOC_KCBIND:
1647 kapi = (struct pps_kcbind_args *)data;
1648 /* XXX Only root should be able to do this */
1649 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1651 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1653 if (kapi->edge & ~pps->ppscap)
1655 pps->kcmode = (kapi->edge & KCMODE_EDGEMASK) |
1656 (pps->kcmode & KCMODE_ABIFLAG);
1659 return (EOPNOTSUPP);
1667 pps_init(struct pps_state *pps)
1669 pps->ppscap |= PPS_TSFMT_TSPEC | PPS_CANWAIT;
1670 if (pps->ppscap & PPS_CAPTUREASSERT)
1671 pps->ppscap |= PPS_OFFSETASSERT;
1672 if (pps->ppscap & PPS_CAPTURECLEAR)
1673 pps->ppscap |= PPS_OFFSETCLEAR;
1675 pps->ppscap |= PPS_TSCLK_MASK;
1677 pps->kcmode &= ~KCMODE_ABIFLAG;
1681 pps_init_abi(struct pps_state *pps)
1685 if (pps->driver_abi > 0) {
1686 pps->kcmode |= KCMODE_ABIFLAG;
1687 pps->kernel_abi = PPS_ABI_VERSION;
1692 pps_capture(struct pps_state *pps)
1694 struct timehands *th;
1696 KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
1698 pps->capgen = atomic_load_acq_int(&th->th_generation);
1701 pps->capffth = fftimehands;
1703 pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
1704 atomic_thread_fence_acq();
1705 if (pps->capgen != th->th_generation)
1710 pps_event(struct pps_state *pps, int event)
1713 struct timespec ts, *tsp, *osp;
1714 u_int tcount, *pcount;
1718 struct timespec *tsp_ffc;
1719 pps_seq_t *pseq_ffc;
1726 KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
1727 /* Nothing to do if not currently set to capture this event type. */
1728 if ((event & pps->ppsparam.mode) == 0)
1730 /* If the timecounter was wound up underneath us, bail out. */
1731 if (pps->capgen == 0 || pps->capgen !=
1732 atomic_load_acq_int(&pps->capth->th_generation))
1735 /* Things would be easier with arrays. */
1736 if (event == PPS_CAPTUREASSERT) {
1737 tsp = &pps->ppsinfo.assert_timestamp;
1738 osp = &pps->ppsparam.assert_offset;
1739 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1741 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1743 pcount = &pps->ppscount[0];
1744 pseq = &pps->ppsinfo.assert_sequence;
1746 ffcount = &pps->ppsinfo_ffc.assert_ffcount;
1747 tsp_ffc = &pps->ppsinfo_ffc.assert_timestamp;
1748 pseq_ffc = &pps->ppsinfo_ffc.assert_sequence;
1751 tsp = &pps->ppsinfo.clear_timestamp;
1752 osp = &pps->ppsparam.clear_offset;
1753 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1755 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1757 pcount = &pps->ppscount[1];
1758 pseq = &pps->ppsinfo.clear_sequence;
1760 ffcount = &pps->ppsinfo_ffc.clear_ffcount;
1761 tsp_ffc = &pps->ppsinfo_ffc.clear_timestamp;
1762 pseq_ffc = &pps->ppsinfo_ffc.clear_sequence;
1767 * If the timecounter changed, we cannot compare the count values, so
1768 * we have to drop the rest of the PPS-stuff until the next event.
1770 if (pps->ppstc != pps->capth->th_counter) {
1771 pps->ppstc = pps->capth->th_counter;
1772 *pcount = pps->capcount;
1773 pps->ppscount[2] = pps->capcount;
1777 /* Convert the count to a timespec. */
1778 tcount = pps->capcount - pps->capth->th_offset_count;
1779 tcount &= pps->capth->th_counter->tc_counter_mask;
1780 bt = pps->capth->th_bintime;
1781 bintime_addx(&bt, pps->capth->th_scale * tcount);
1782 bintime2timespec(&bt, &ts);
1784 /* If the timecounter was wound up underneath us, bail out. */
1785 atomic_thread_fence_acq();
1786 if (pps->capgen != pps->capth->th_generation)
1789 *pcount = pps->capcount;
1794 timespecadd(tsp, osp, tsp);
1795 if (tsp->tv_nsec < 0) {
1796 tsp->tv_nsec += 1000000000;
1802 *ffcount = pps->capffth->tick_ffcount + tcount;
1803 bt = pps->capffth->tick_time;
1804 ffclock_convert_delta(tcount, pps->capffth->cest.period, &bt);
1805 bintime_add(&bt, &pps->capffth->tick_time);
1806 bintime2timespec(&bt, &ts);
1816 * Feed the NTP PLL/FLL.
1817 * The FLL wants to know how many (hardware) nanoseconds
1818 * elapsed since the previous event.
1820 tcount = pps->capcount - pps->ppscount[2];
1821 pps->ppscount[2] = pps->capcount;
1822 tcount &= pps->capth->th_counter->tc_counter_mask;
1823 scale = (uint64_t)1 << 63;
1824 scale /= pps->capth->th_counter->tc_frequency;
1828 bintime_addx(&bt, scale * tcount);
1829 bintime2timespec(&bt, &ts);
1830 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
1834 /* Wakeup anyone sleeping in pps_fetch(). */
1839 * Timecounters need to be updated every so often to prevent the hardware
1840 * counter from overflowing. Updating also recalculates the cached values
1841 * used by the get*() family of functions, so their precision depends on
1842 * the update frequency.
1846 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0,
1847 "Approximate number of hardclock ticks in a millisecond");
1850 tc_ticktock(int cnt)
1854 if (mtx_trylock_spin(&tc_setclock_mtx)) {
1856 if (count >= tc_tick) {
1860 mtx_unlock_spin(&tc_setclock_mtx);
1864 static void __inline
1865 tc_adjprecision(void)
1869 if (tc_timepercentage > 0) {
1870 t = (99 + tc_timepercentage) / tc_timepercentage;
1871 tc_precexp = fls(t + (t >> 1)) - 1;
1872 FREQ2BT(hz / tc_tick, &bt_timethreshold);
1873 FREQ2BT(hz, &bt_tickthreshold);
1874 bintime_shift(&bt_timethreshold, tc_precexp);
1875 bintime_shift(&bt_tickthreshold, tc_precexp);
1878 bt_timethreshold.sec = INT_MAX;
1879 bt_timethreshold.frac = ~(uint64_t)0;
1880 bt_tickthreshold = bt_timethreshold;
1882 sbt_timethreshold = bttosbt(bt_timethreshold);
1883 sbt_tickthreshold = bttosbt(bt_tickthreshold);
1887 sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS)
1891 val = tc_timepercentage;
1892 error = sysctl_handle_int(oidp, &val, 0, req);
1893 if (error != 0 || req->newptr == NULL)
1895 tc_timepercentage = val;
1903 /* Set up the requested number of timehands. */
1905 inittimehands(void *dummy)
1907 struct timehands *thp;
1910 TUNABLE_INT_FETCH("kern.timecounter.timehands_count",
1912 if (timehands_count < 1)
1913 timehands_count = 1;
1914 if (timehands_count > nitems(ths))
1915 timehands_count = nitems(ths);
1916 for (i = 1, thp = &ths[0]; i < timehands_count; thp = &ths[i++])
1917 thp->th_next = &ths[i];
1918 thp->th_next = &ths[0];
1920 SYSINIT(timehands, SI_SUB_TUNABLES, SI_ORDER_ANY, inittimehands, NULL);
1923 inittimecounter(void *dummy)
1929 * Set the initial timeout to
1930 * max(1, <approx. number of hardclock ticks in a millisecond>).
1931 * People should probably not use the sysctl to set the timeout
1932 * to smaller than its initial value, since that value is the
1933 * smallest reasonable one. If they want better timestamps they
1934 * should use the non-"get"* functions.
1937 tc_tick = (hz + 500) / 1000;
1941 FREQ2BT(hz, &tick_bt);
1942 tick_sbt = bttosbt(tick_bt);
1943 tick_rate = hz / tc_tick;
1944 FREQ2BT(tick_rate, &tc_tick_bt);
1945 tc_tick_sbt = bttosbt(tc_tick_bt);
1946 p = (tc_tick * 1000000) / hz;
1947 printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
1953 /* warm up new timecounter (again) and get rolling. */
1954 (void)timecounter->tc_get_timecount(timecounter);
1955 (void)timecounter->tc_get_timecount(timecounter);
1956 mtx_lock_spin(&tc_setclock_mtx);
1958 mtx_unlock_spin(&tc_setclock_mtx);
1961 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL);
1963 /* Cpu tick handling -------------------------------------------------*/
1965 static int cpu_tick_variable;
1966 static uint64_t cpu_tick_frequency;
1968 DPCPU_DEFINE_STATIC(uint64_t, tc_cpu_ticks_base);
1969 DPCPU_DEFINE_STATIC(unsigned, tc_cpu_ticks_last);
1974 struct timecounter *tc;
1975 uint64_t res, *base;
1979 base = DPCPU_PTR(tc_cpu_ticks_base);
1980 last = DPCPU_PTR(tc_cpu_ticks_last);
1981 tc = timehands->th_counter;
1982 u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
1984 *base += (uint64_t)tc->tc_counter_mask + 1;
1992 cpu_tick_calibration(void)
1994 static time_t last_calib;
1996 if (time_uptime != last_calib && !(time_uptime & 0xf)) {
1997 cpu_tick_calibrate(0);
1998 last_calib = time_uptime;
2003 * This function gets called every 16 seconds on only one designated
2004 * CPU in the system from hardclock() via cpu_tick_calibration()().
2006 * Whenever the real time clock is stepped we get called with reset=1
2007 * to make sure we handle suspend/resume and similar events correctly.
2011 cpu_tick_calibrate(int reset)
2013 static uint64_t c_last;
2014 uint64_t c_this, c_delta;
2015 static struct bintime t_last;
2016 struct bintime t_this, t_delta;
2020 /* The clock was stepped, abort & reset */
2025 /* we don't calibrate fixed rate cputicks */
2026 if (!cpu_tick_variable)
2029 getbinuptime(&t_this);
2030 c_this = cpu_ticks();
2031 if (t_last.sec != 0) {
2032 c_delta = c_this - c_last;
2034 bintime_sub(&t_delta, &t_last);
2037 * 2^(64-20) / 16[s] =
2039 * 17.592.186.044.416 / 16 =
2040 * 1.099.511.627.776 [Hz]
2042 divi = t_delta.sec << 20;
2043 divi |= t_delta.frac >> (64 - 20);
2046 if (c_delta > cpu_tick_frequency) {
2047 if (0 && bootverbose)
2048 printf("cpu_tick increased to %ju Hz\n",
2050 cpu_tick_frequency = c_delta;
2058 set_cputicker(cpu_tick_f *func, uint64_t freq, unsigned var)
2062 cpu_ticks = tc_cpu_ticks;
2064 cpu_tick_frequency = freq;
2065 cpu_tick_variable = var;
2074 if (cpu_ticks == tc_cpu_ticks)
2075 return (tc_getfrequency());
2076 return (cpu_tick_frequency);
2080 * We need to be slightly careful converting cputicks to microseconds.
2081 * There is plenty of margin in 64 bits of microseconds (half a million
2082 * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
2083 * before divide conversion (to retain precision) we find that the
2084 * margin shrinks to 1.5 hours (one millionth of 146y).
2085 * With a three prong approach we never lose significant bits, no
2086 * matter what the cputick rate and length of timeinterval is.
2090 cputick2usec(uint64_t tick)
2093 if (tick > 18446744073709551LL) /* floor(2^64 / 1000) */
2094 return (tick / (cpu_tickrate() / 1000000LL));
2095 else if (tick > 18446744073709LL) /* floor(2^64 / 1000000) */
2096 return ((tick * 1000LL) / (cpu_tickrate() / 1000LL));
2098 return ((tick * 1000000LL) / cpu_tickrate());
2101 cpu_tick_f *cpu_ticks = tc_cpu_ticks;
2103 static int vdso_th_enable = 1;
2105 sysctl_fast_gettime(SYSCTL_HANDLER_ARGS)
2107 int old_vdso_th_enable, error;
2109 old_vdso_th_enable = vdso_th_enable;
2110 error = sysctl_handle_int(oidp, &old_vdso_th_enable, 0, req);
2113 vdso_th_enable = old_vdso_th_enable;
2116 SYSCTL_PROC(_kern_timecounter, OID_AUTO, fast_gettime,
2117 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
2118 NULL, 0, sysctl_fast_gettime, "I", "Enable fast time of day");
2121 tc_fill_vdso_timehands(struct vdso_timehands *vdso_th)
2123 struct timehands *th;
2127 vdso_th->th_scale = th->th_scale;
2128 vdso_th->th_offset_count = th->th_offset_count;
2129 vdso_th->th_counter_mask = th->th_counter->tc_counter_mask;
2130 vdso_th->th_offset = th->th_offset;
2131 vdso_th->th_boottime = th->th_boottime;
2132 if (th->th_counter->tc_fill_vdso_timehands != NULL) {
2133 enabled = th->th_counter->tc_fill_vdso_timehands(vdso_th,
2137 if (!vdso_th_enable)
2142 #ifdef COMPAT_FREEBSD32
2144 tc_fill_vdso_timehands32(struct vdso_timehands32 *vdso_th32)
2146 struct timehands *th;
2150 *(uint64_t *)&vdso_th32->th_scale[0] = th->th_scale;
2151 vdso_th32->th_offset_count = th->th_offset_count;
2152 vdso_th32->th_counter_mask = th->th_counter->tc_counter_mask;
2153 vdso_th32->th_offset.sec = th->th_offset.sec;
2154 *(uint64_t *)&vdso_th32->th_offset.frac[0] = th->th_offset.frac;
2155 vdso_th32->th_boottime.sec = th->th_boottime.sec;
2156 *(uint64_t *)&vdso_th32->th_boottime.frac[0] = th->th_boottime.frac;
2157 if (th->th_counter->tc_fill_vdso_timehands32 != NULL) {
2158 enabled = th->th_counter->tc_fill_vdso_timehands32(vdso_th32,
2162 if (!vdso_th_enable)