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. */
139 static char tc_from_tunable[16];
141 static void tc_windup(struct bintime *new_boottimebin);
142 static void cpu_tick_calibrate(int);
144 void dtrace_getnanotime(struct timespec *tsp);
147 sysctl_kern_boottime(SYSCTL_HANDLER_ARGS)
149 struct timeval boottime;
151 getboottime(&boottime);
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) \
239 _Static_assert(_Generic(((struct timehands *)NULL)->member, \
240 struct bintime: 1, default: 0) == 1, \
241 "struct timehands member is not of struct bintime type"); \
243 bintime_off(dst, __offsetof(struct timehands, member)); \
247 getthmember(void *out, size_t out_size, u_int off)
249 struct timehands *th;
254 gen = atomic_load_acq_int(&th->th_generation);
255 memcpy(out, (char *)th + off, out_size);
256 atomic_thread_fence_acq();
257 } while (gen == 0 || gen != th->th_generation);
259 #define GETTHMEMBER(dst, member) \
262 _Static_assert(_Generic(*dst, \
263 __typeof(((struct timehands *)NULL)->member): 1, \
265 "*dst and struct timehands member have different types"); \
267 getthmember(dst, sizeof(*dst), __offsetof(struct timehands, \
273 fbclock_binuptime(struct bintime *bt)
276 GETTHBINTIME(bt, th_offset);
280 fbclock_nanouptime(struct timespec *tsp)
284 fbclock_binuptime(&bt);
285 bintime2timespec(&bt, tsp);
289 fbclock_microuptime(struct timeval *tvp)
293 fbclock_binuptime(&bt);
294 bintime2timeval(&bt, tvp);
298 fbclock_bintime(struct bintime *bt)
301 GETTHBINTIME(bt, th_bintime);
305 fbclock_nanotime(struct timespec *tsp)
309 fbclock_bintime(&bt);
310 bintime2timespec(&bt, tsp);
314 fbclock_microtime(struct timeval *tvp)
318 fbclock_bintime(&bt);
319 bintime2timeval(&bt, tvp);
323 fbclock_getbinuptime(struct bintime *bt)
326 GETTHMEMBER(bt, th_offset);
330 fbclock_getnanouptime(struct timespec *tsp)
334 GETTHMEMBER(&bt, th_offset);
335 bintime2timespec(&bt, tsp);
339 fbclock_getmicrouptime(struct timeval *tvp)
343 GETTHMEMBER(&bt, th_offset);
344 bintime2timeval(&bt, tvp);
348 fbclock_getbintime(struct bintime *bt)
351 GETTHMEMBER(bt, th_bintime);
355 fbclock_getnanotime(struct timespec *tsp)
358 GETTHMEMBER(tsp, th_nanotime);
362 fbclock_getmicrotime(struct timeval *tvp)
365 GETTHMEMBER(tvp, th_microtime);
370 binuptime(struct bintime *bt)
373 GETTHBINTIME(bt, th_offset);
377 nanouptime(struct timespec *tsp)
382 bintime2timespec(&bt, tsp);
386 microuptime(struct timeval *tvp)
391 bintime2timeval(&bt, tvp);
395 bintime(struct bintime *bt)
398 GETTHBINTIME(bt, th_bintime);
402 nanotime(struct timespec *tsp)
407 bintime2timespec(&bt, tsp);
411 microtime(struct timeval *tvp)
416 bintime2timeval(&bt, tvp);
420 getbinuptime(struct bintime *bt)
423 GETTHMEMBER(bt, th_offset);
427 getnanouptime(struct timespec *tsp)
431 GETTHMEMBER(&bt, th_offset);
432 bintime2timespec(&bt, tsp);
436 getmicrouptime(struct timeval *tvp)
440 GETTHMEMBER(&bt, th_offset);
441 bintime2timeval(&bt, tvp);
445 getbintime(struct bintime *bt)
448 GETTHMEMBER(bt, th_bintime);
452 getnanotime(struct timespec *tsp)
455 GETTHMEMBER(tsp, th_nanotime);
459 getmicrotime(struct timeval *tvp)
462 GETTHMEMBER(tvp, th_microtime);
467 getboottime(struct timeval *boottime)
469 struct bintime boottimebin;
471 getboottimebin(&boottimebin);
472 bintime2timeval(&boottimebin, boottime);
476 getboottimebin(struct bintime *boottimebin)
479 GETTHMEMBER(boottimebin, th_boottime);
484 * Support for feed-forward synchronization algorithms. This is heavily inspired
485 * by the timehands mechanism but kept independent from it. *_windup() functions
486 * have some connection to avoid accessing the timecounter hardware more than
490 /* Feed-forward clock estimates kept updated by the synchronization daemon. */
491 struct ffclock_estimate ffclock_estimate;
492 struct bintime ffclock_boottime; /* Feed-forward boot time estimate. */
493 uint32_t ffclock_status; /* Feed-forward clock status. */
494 int8_t ffclock_updated; /* New estimates are available. */
495 struct mtx ffclock_mtx; /* Mutex on ffclock_estimate. */
498 struct ffclock_estimate cest;
499 struct bintime tick_time;
500 struct bintime tick_time_lerp;
501 ffcounter tick_ffcount;
502 uint64_t period_lerp;
503 volatile uint8_t gen;
504 struct fftimehands *next;
507 #define NUM_ELEMENTS(x) (sizeof(x) / sizeof(*x))
509 static struct fftimehands ffth[10];
510 static struct fftimehands *volatile fftimehands = ffth;
515 struct fftimehands *cur;
516 struct fftimehands *last;
518 memset(ffth, 0, sizeof(ffth));
520 last = ffth + NUM_ELEMENTS(ffth) - 1;
521 for (cur = ffth; cur < last; cur++)
526 ffclock_status = FFCLOCK_STA_UNSYNC;
527 mtx_init(&ffclock_mtx, "ffclock lock", NULL, MTX_DEF);
531 * Reset the feed-forward clock estimates. Called from inittodr() to get things
532 * kick started and uses the timecounter nominal frequency as a first period
533 * estimate. Note: this function may be called several time just after boot.
534 * Note: this is the only function that sets the value of boot time for the
535 * monotonic (i.e. uptime) version of the feed-forward clock.
538 ffclock_reset_clock(struct timespec *ts)
540 struct timecounter *tc;
541 struct ffclock_estimate cest;
543 tc = timehands->th_counter;
544 memset(&cest, 0, sizeof(struct ffclock_estimate));
546 timespec2bintime(ts, &ffclock_boottime);
547 timespec2bintime(ts, &(cest.update_time));
548 ffclock_read_counter(&cest.update_ffcount);
549 cest.leapsec_next = 0;
550 cest.period = ((1ULL << 63) / tc->tc_frequency) << 1;
553 cest.status = FFCLOCK_STA_UNSYNC;
554 cest.leapsec_total = 0;
557 mtx_lock(&ffclock_mtx);
558 bcopy(&cest, &ffclock_estimate, sizeof(struct ffclock_estimate));
559 ffclock_updated = INT8_MAX;
560 mtx_unlock(&ffclock_mtx);
562 printf("ffclock reset: %s (%llu Hz), time = %ld.%09lu\n", tc->tc_name,
563 (unsigned long long)tc->tc_frequency, (long)ts->tv_sec,
564 (unsigned long)ts->tv_nsec);
568 * Sub-routine to convert a time interval measured in RAW counter units to time
569 * in seconds stored in bintime format.
570 * NOTE: bintime_mul requires u_int, but the value of the ffcounter may be
571 * larger than the max value of u_int (on 32 bit architecture). Loop to consume
575 ffclock_convert_delta(ffcounter ffdelta, uint64_t period, struct bintime *bt)
578 ffcounter delta, delta_max;
580 delta_max = (1ULL << (8 * sizeof(unsigned int))) - 1;
583 if (ffdelta > delta_max)
589 bintime_mul(&bt2, (unsigned int)delta);
590 bintime_add(bt, &bt2);
592 } while (ffdelta > 0);
596 * Update the fftimehands.
597 * Push the tick ffcount and time(s) forward based on current clock estimate.
598 * The conversion from ffcounter to bintime relies on the difference clock
599 * principle, whose accuracy relies on computing small time intervals. If a new
600 * clock estimate has been passed by the synchronisation daemon, make it
601 * current, and compute the linear interpolation for monotonic time if needed.
604 ffclock_windup(unsigned int delta)
606 struct ffclock_estimate *cest;
607 struct fftimehands *ffth;
608 struct bintime bt, gap_lerp;
611 unsigned int polling;
612 uint8_t forward_jump, ogen;
615 * Pick the next timehand, copy current ffclock estimates and move tick
616 * times and counter forward.
619 ffth = fftimehands->next;
623 bcopy(&fftimehands->cest, cest, sizeof(struct ffclock_estimate));
624 ffdelta = (ffcounter)delta;
625 ffth->period_lerp = fftimehands->period_lerp;
627 ffth->tick_time = fftimehands->tick_time;
628 ffclock_convert_delta(ffdelta, cest->period, &bt);
629 bintime_add(&ffth->tick_time, &bt);
631 ffth->tick_time_lerp = fftimehands->tick_time_lerp;
632 ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt);
633 bintime_add(&ffth->tick_time_lerp, &bt);
635 ffth->tick_ffcount = fftimehands->tick_ffcount + ffdelta;
638 * Assess the status of the clock, if the last update is too old, it is
639 * likely the synchronisation daemon is dead and the clock is free
642 if (ffclock_updated == 0) {
643 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
644 ffclock_convert_delta(ffdelta, cest->period, &bt);
645 if (bt.sec > 2 * FFCLOCK_SKM_SCALE)
646 ffclock_status |= FFCLOCK_STA_UNSYNC;
650 * If available, grab updated clock estimates and make them current.
651 * Recompute time at this tick using the updated estimates. The clock
652 * estimates passed the feed-forward synchronisation daemon may result
653 * in time conversion that is not monotonically increasing (just after
654 * the update). time_lerp is a particular linear interpolation over the
655 * synchronisation algo polling period that ensures monotonicity for the
656 * clock ids requesting it.
658 if (ffclock_updated > 0) {
659 bcopy(&ffclock_estimate, cest, sizeof(struct ffclock_estimate));
660 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
661 ffth->tick_time = cest->update_time;
662 ffclock_convert_delta(ffdelta, cest->period, &bt);
663 bintime_add(&ffth->tick_time, &bt);
665 /* ffclock_reset sets ffclock_updated to INT8_MAX */
666 if (ffclock_updated == INT8_MAX)
667 ffth->tick_time_lerp = ffth->tick_time;
669 if (bintime_cmp(&ffth->tick_time, &ffth->tick_time_lerp, >))
674 bintime_clear(&gap_lerp);
676 gap_lerp = ffth->tick_time;
677 bintime_sub(&gap_lerp, &ffth->tick_time_lerp);
679 gap_lerp = ffth->tick_time_lerp;
680 bintime_sub(&gap_lerp, &ffth->tick_time);
684 * The reset from the RTC clock may be far from accurate, and
685 * reducing the gap between real time and interpolated time
686 * could take a very long time if the interpolated clock insists
687 * on strict monotonicity. The clock is reset under very strict
688 * conditions (kernel time is known to be wrong and
689 * synchronization daemon has been restarted recently.
690 * ffclock_boottime absorbs the jump to ensure boot time is
691 * correct and uptime functions stay consistent.
693 if (((ffclock_status & FFCLOCK_STA_UNSYNC) == FFCLOCK_STA_UNSYNC) &&
694 ((cest->status & FFCLOCK_STA_UNSYNC) == 0) &&
695 ((cest->status & FFCLOCK_STA_WARMUP) == FFCLOCK_STA_WARMUP)) {
697 bintime_add(&ffclock_boottime, &gap_lerp);
699 bintime_sub(&ffclock_boottime, &gap_lerp);
700 ffth->tick_time_lerp = ffth->tick_time;
701 bintime_clear(&gap_lerp);
704 ffclock_status = cest->status;
705 ffth->period_lerp = cest->period;
708 * Compute corrected period used for the linear interpolation of
709 * time. The rate of linear interpolation is capped to 5000PPM
712 if (bintime_isset(&gap_lerp)) {
713 ffdelta = cest->update_ffcount;
714 ffdelta -= fftimehands->cest.update_ffcount;
715 ffclock_convert_delta(ffdelta, cest->period, &bt);
718 bt.frac = 5000000 * (uint64_t)18446744073LL;
719 bintime_mul(&bt, polling);
720 if (bintime_cmp(&gap_lerp, &bt, >))
723 /* Approximate 1 sec by 1-(1/2^64) to ease arithmetic */
725 if (gap_lerp.sec > 0) {
727 frac /= ffdelta / gap_lerp.sec;
729 frac += gap_lerp.frac / ffdelta;
732 ffth->period_lerp += frac;
734 ffth->period_lerp -= frac;
746 * Adjust the fftimehands when the timecounter is changed. Stating the obvious,
747 * the old and new hardware counter cannot be read simultaneously. tc_windup()
748 * does read the two counters 'back to back', but a few cycles are effectively
749 * lost, and not accumulated in tick_ffcount. This is a fairly radical
750 * operation for a feed-forward synchronization daemon, and it is its job to not
751 * pushing irrelevant data to the kernel. Because there is no locking here,
752 * simply force to ignore pending or next update to give daemon a chance to
753 * realize the counter has changed.
756 ffclock_change_tc(struct timehands *th)
758 struct fftimehands *ffth;
759 struct ffclock_estimate *cest;
760 struct timecounter *tc;
764 ffth = fftimehands->next;
769 bcopy(&(fftimehands->cest), cest, sizeof(struct ffclock_estimate));
770 cest->period = ((1ULL << 63) / tc->tc_frequency ) << 1;
773 cest->status |= FFCLOCK_STA_UNSYNC;
775 ffth->tick_ffcount = fftimehands->tick_ffcount;
776 ffth->tick_time_lerp = fftimehands->tick_time_lerp;
777 ffth->tick_time = fftimehands->tick_time;
778 ffth->period_lerp = cest->period;
780 /* Do not lock but ignore next update from synchronization daemon. */
790 * Retrieve feed-forward counter and time of last kernel tick.
793 ffclock_last_tick(ffcounter *ffcount, struct bintime *bt, uint32_t flags)
795 struct fftimehands *ffth;
799 * No locking but check generation has not changed. Also need to make
800 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
805 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP)
806 *bt = ffth->tick_time_lerp;
808 *bt = ffth->tick_time;
809 *ffcount = ffth->tick_ffcount;
810 } while (gen == 0 || gen != ffth->gen);
814 * Absolute clock conversion. Low level function to convert ffcounter to
815 * bintime. The ffcounter is converted using the current ffclock period estimate
816 * or the "interpolated period" to ensure monotonicity.
817 * NOTE: this conversion may have been deferred, and the clock updated since the
818 * hardware counter has been read.
821 ffclock_convert_abs(ffcounter ffcount, struct bintime *bt, uint32_t flags)
823 struct fftimehands *ffth;
829 * No locking but check generation has not changed. Also need to make
830 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
835 if (ffcount > ffth->tick_ffcount)
836 ffdelta = ffcount - ffth->tick_ffcount;
838 ffdelta = ffth->tick_ffcount - ffcount;
840 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP) {
841 *bt = ffth->tick_time_lerp;
842 ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt2);
844 *bt = ffth->tick_time;
845 ffclock_convert_delta(ffdelta, ffth->cest.period, &bt2);
848 if (ffcount > ffth->tick_ffcount)
849 bintime_add(bt, &bt2);
851 bintime_sub(bt, &bt2);
852 } while (gen == 0 || gen != ffth->gen);
856 * Difference clock conversion.
857 * Low level function to Convert a time interval measured in RAW counter units
858 * into bintime. The difference clock allows measuring small intervals much more
859 * reliably than the absolute clock.
862 ffclock_convert_diff(ffcounter ffdelta, struct bintime *bt)
864 struct fftimehands *ffth;
867 /* No locking but check generation has not changed. */
871 ffclock_convert_delta(ffdelta, ffth->cest.period, bt);
872 } while (gen == 0 || gen != ffth->gen);
876 * Access to current ffcounter value.
879 ffclock_read_counter(ffcounter *ffcount)
881 struct timehands *th;
882 struct fftimehands *ffth;
883 unsigned int gen, delta;
886 * ffclock_windup() called from tc_windup(), safe to rely on
887 * th->th_generation only, for correct delta and ffcounter.
891 gen = atomic_load_acq_int(&th->th_generation);
893 delta = tc_delta(th);
894 *ffcount = ffth->tick_ffcount;
895 atomic_thread_fence_acq();
896 } while (gen == 0 || gen != th->th_generation);
902 binuptime(struct bintime *bt)
905 binuptime_fromclock(bt, sysclock_active);
909 nanouptime(struct timespec *tsp)
912 nanouptime_fromclock(tsp, sysclock_active);
916 microuptime(struct timeval *tvp)
919 microuptime_fromclock(tvp, sysclock_active);
923 bintime(struct bintime *bt)
926 bintime_fromclock(bt, sysclock_active);
930 nanotime(struct timespec *tsp)
933 nanotime_fromclock(tsp, sysclock_active);
937 microtime(struct timeval *tvp)
940 microtime_fromclock(tvp, sysclock_active);
944 getbinuptime(struct bintime *bt)
947 getbinuptime_fromclock(bt, sysclock_active);
951 getnanouptime(struct timespec *tsp)
954 getnanouptime_fromclock(tsp, sysclock_active);
958 getmicrouptime(struct timeval *tvp)
961 getmicrouptime_fromclock(tvp, sysclock_active);
965 getbintime(struct bintime *bt)
968 getbintime_fromclock(bt, sysclock_active);
972 getnanotime(struct timespec *tsp)
975 getnanotime_fromclock(tsp, sysclock_active);
979 getmicrotime(struct timeval *tvp)
982 getmicrouptime_fromclock(tvp, sysclock_active);
988 * This is a clone of getnanotime and used for walltimestamps.
989 * The dtrace_ prefix prevents fbt from creating probes for
990 * it so walltimestamp can be safely used in all fbt probes.
993 dtrace_getnanotime(struct timespec *tsp)
996 GETTHMEMBER(tsp, th_nanotime);
1000 * System clock currently providing time to the system. Modifiable via sysctl
1001 * when the FFCLOCK option is defined.
1003 int sysclock_active = SYSCLOCK_FBCK;
1005 /* Internal NTP status and error estimates. */
1006 extern int time_status;
1007 extern long time_esterror;
1010 * Take a snapshot of sysclock data which can be used to compare system clocks
1011 * and generate timestamps after the fact.
1014 sysclock_getsnapshot(struct sysclock_snap *clock_snap, int fast)
1016 struct fbclock_info *fbi;
1017 struct timehands *th;
1019 unsigned int delta, gen;
1022 struct fftimehands *ffth;
1023 struct ffclock_info *ffi;
1024 struct ffclock_estimate cest;
1026 ffi = &clock_snap->ff_info;
1029 fbi = &clock_snap->fb_info;
1034 gen = atomic_load_acq_int(&th->th_generation);
1035 fbi->th_scale = th->th_scale;
1036 fbi->tick_time = th->th_offset;
1039 ffi->tick_time = ffth->tick_time_lerp;
1040 ffi->tick_time_lerp = ffth->tick_time_lerp;
1041 ffi->period = ffth->cest.period;
1042 ffi->period_lerp = ffth->period_lerp;
1043 clock_snap->ffcount = ffth->tick_ffcount;
1047 delta = tc_delta(th);
1048 atomic_thread_fence_acq();
1049 } while (gen == 0 || gen != th->th_generation);
1051 clock_snap->delta = delta;
1052 clock_snap->sysclock_active = sysclock_active;
1054 /* Record feedback clock status and error. */
1055 clock_snap->fb_info.status = time_status;
1056 /* XXX: Very crude estimate of feedback clock error. */
1057 bt.sec = time_esterror / 1000000;
1058 bt.frac = ((time_esterror - bt.sec) * 1000000) *
1059 (uint64_t)18446744073709ULL;
1060 clock_snap->fb_info.error = bt;
1064 clock_snap->ffcount += delta;
1066 /* Record feed-forward clock leap second adjustment. */
1067 ffi->leapsec_adjustment = cest.leapsec_total;
1068 if (clock_snap->ffcount > cest.leapsec_next)
1069 ffi->leapsec_adjustment -= cest.leapsec;
1071 /* Record feed-forward clock status and error. */
1072 clock_snap->ff_info.status = cest.status;
1073 ffcount = clock_snap->ffcount - cest.update_ffcount;
1074 ffclock_convert_delta(ffcount, cest.period, &bt);
1075 /* 18446744073709 = int(2^64/1e12), err_bound_rate in [ps/s]. */
1076 bintime_mul(&bt, cest.errb_rate * (uint64_t)18446744073709ULL);
1077 /* 18446744073 = int(2^64 / 1e9), since err_abs in [ns]. */
1078 bintime_addx(&bt, cest.errb_abs * (uint64_t)18446744073ULL);
1079 clock_snap->ff_info.error = bt;
1084 * Convert a sysclock snapshot into a struct bintime based on the specified
1085 * clock source and flags.
1088 sysclock_snap2bintime(struct sysclock_snap *cs, struct bintime *bt,
1089 int whichclock, uint32_t flags)
1091 struct bintime boottimebin;
1097 switch (whichclock) {
1099 *bt = cs->fb_info.tick_time;
1101 /* If snapshot was created with !fast, delta will be >0. */
1103 bintime_addx(bt, cs->fb_info.th_scale * cs->delta);
1105 if ((flags & FBCLOCK_UPTIME) == 0) {
1106 getboottimebin(&boottimebin);
1107 bintime_add(bt, &boottimebin);
1112 if (flags & FFCLOCK_LERP) {
1113 *bt = cs->ff_info.tick_time_lerp;
1114 period = cs->ff_info.period_lerp;
1116 *bt = cs->ff_info.tick_time;
1117 period = cs->ff_info.period;
1120 /* If snapshot was created with !fast, delta will be >0. */
1121 if (cs->delta > 0) {
1122 ffclock_convert_delta(cs->delta, period, &bt2);
1123 bintime_add(bt, &bt2);
1126 /* Leap second adjustment. */
1127 if (flags & FFCLOCK_LEAPSEC)
1128 bt->sec -= cs->ff_info.leapsec_adjustment;
1130 /* Boot time adjustment, for uptime/monotonic clocks. */
1131 if (flags & FFCLOCK_UPTIME)
1132 bintime_sub(bt, &ffclock_boottime);
1144 * Initialize a new timecounter and possibly use it.
1147 tc_init(struct timecounter *tc)
1150 struct sysctl_oid *tc_root;
1152 u = tc->tc_frequency / tc->tc_counter_mask;
1153 /* XXX: We need some margin here, 10% is a guess */
1156 if (u > hz && tc->tc_quality >= 0) {
1157 tc->tc_quality = -2000;
1159 printf("Timecounter \"%s\" frequency %ju Hz",
1160 tc->tc_name, (uintmax_t)tc->tc_frequency);
1161 printf(" -- Insufficient hz, needs at least %u\n", u);
1163 } else if (tc->tc_quality >= 0 || bootverbose) {
1164 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
1165 tc->tc_name, (uintmax_t)tc->tc_frequency,
1169 tc->tc_next = timecounters;
1172 * Set up sysctl tree for this counter.
1174 tc_root = SYSCTL_ADD_NODE_WITH_LABEL(NULL,
1175 SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name,
1176 CTLFLAG_RW, 0, "timecounter description", "timecounter");
1177 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1178 "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0,
1179 "mask for implemented bits");
1180 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1181 "counter", CTLTYPE_UINT | CTLFLAG_RD, tc, sizeof(*tc),
1182 sysctl_kern_timecounter_get, "IU", "current timecounter value");
1183 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1184 "frequency", CTLTYPE_U64 | CTLFLAG_RD, tc, sizeof(*tc),
1185 sysctl_kern_timecounter_freq, "QU", "timecounter frequency");
1186 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1187 "quality", CTLFLAG_RD, &(tc->tc_quality), 0,
1188 "goodness of time counter");
1190 * Do not automatically switch if the current tc was specifically
1191 * chosen. Never automatically use a timecounter with negative quality.
1192 * Even though we run on the dummy counter, switching here may be
1193 * worse since this timecounter may not be monotonic.
1197 if (tc->tc_quality < 0)
1199 if (tc_from_tunable[0] != '\0' &&
1200 strcmp(tc->tc_name, tc_from_tunable) == 0) {
1202 tc_from_tunable[0] = '\0';
1204 if (tc->tc_quality < timecounter->tc_quality)
1206 if (tc->tc_quality == timecounter->tc_quality &&
1207 tc->tc_frequency < timecounter->tc_frequency)
1210 (void)tc->tc_get_timecount(tc);
1214 /* Report the frequency of the current timecounter. */
1216 tc_getfrequency(void)
1219 return (timehands->th_counter->tc_frequency);
1223 sleeping_on_old_rtc(struct thread *td)
1227 * td_rtcgen is modified by curthread when it is running,
1228 * and by other threads in this function. By finding the thread
1229 * on a sleepqueue and holding the lock on the sleepqueue
1230 * chain, we guarantee that the thread is not running and that
1231 * modifying td_rtcgen is safe. Setting td_rtcgen to zero informs
1232 * the thread that it was woken due to a real-time clock adjustment.
1233 * (The declaration of td_rtcgen refers to this comment.)
1235 if (td->td_rtcgen != 0 && td->td_rtcgen != rtc_generation) {
1242 static struct mtx tc_setclock_mtx;
1243 MTX_SYSINIT(tc_setclock_init, &tc_setclock_mtx, "tcsetc", MTX_SPIN);
1246 * Step our concept of UTC. This is done by modifying our estimate of
1250 tc_setclock(struct timespec *ts)
1252 struct timespec tbef, taft;
1253 struct bintime bt, bt2;
1255 timespec2bintime(ts, &bt);
1257 mtx_lock_spin(&tc_setclock_mtx);
1258 cpu_tick_calibrate(1);
1260 bintime_sub(&bt, &bt2);
1262 /* XXX fiddle all the little crinkly bits around the fiords... */
1264 mtx_unlock_spin(&tc_setclock_mtx);
1266 /* Avoid rtc_generation == 0, since td_rtcgen == 0 is special. */
1267 atomic_add_rel_int(&rtc_generation, 2);
1268 sleepq_chains_remove_matching(sleeping_on_old_rtc);
1269 if (timestepwarnings) {
1272 "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
1273 (intmax_t)tbef.tv_sec, tbef.tv_nsec,
1274 (intmax_t)taft.tv_sec, taft.tv_nsec,
1275 (intmax_t)ts->tv_sec, ts->tv_nsec);
1280 * Initialize the next struct timehands in the ring and make
1281 * it the active timehands. Along the way we might switch to a different
1282 * timecounter and/or do seconds processing in NTP. Slightly magic.
1285 tc_windup(struct bintime *new_boottimebin)
1288 struct timehands *th, *tho;
1290 u_int delta, ncount, ogen;
1295 * Make the next timehands a copy of the current one, but do
1296 * not overwrite the generation or next pointer. While we
1297 * update the contents, the generation must be zero. We need
1298 * to ensure that the zero generation is visible before the
1299 * data updates become visible, which requires release fence.
1300 * For similar reasons, re-reading of the generation after the
1301 * data is read should use acquire fence.
1305 ogen = th->th_generation;
1306 th->th_generation = 0;
1307 atomic_thread_fence_rel();
1308 memcpy(th, tho, offsetof(struct timehands, th_generation));
1309 if (new_boottimebin != NULL)
1310 th->th_boottime = *new_boottimebin;
1313 * Capture a timecounter delta on the current timecounter and if
1314 * changing timecounters, a counter value from the new timecounter.
1315 * Update the offset fields accordingly.
1317 delta = tc_delta(th);
1318 if (th->th_counter != timecounter)
1319 ncount = timecounter->tc_get_timecount(timecounter);
1323 ffclock_windup(delta);
1325 th->th_offset_count += delta;
1326 th->th_offset_count &= th->th_counter->tc_counter_mask;
1327 while (delta > th->th_counter->tc_frequency) {
1328 /* Eat complete unadjusted seconds. */
1329 delta -= th->th_counter->tc_frequency;
1330 th->th_offset.sec++;
1332 if ((delta > th->th_counter->tc_frequency / 2) &&
1333 (th->th_scale * delta < ((uint64_t)1 << 63))) {
1334 /* The product th_scale * delta just barely overflows. */
1335 th->th_offset.sec++;
1337 bintime_addx(&th->th_offset, th->th_scale * delta);
1340 * Hardware latching timecounters may not generate interrupts on
1341 * PPS events, so instead we poll them. There is a finite risk that
1342 * the hardware might capture a count which is later than the one we
1343 * got above, and therefore possibly in the next NTP second which might
1344 * have a different rate than the current NTP second. It doesn't
1345 * matter in practice.
1347 if (tho->th_counter->tc_poll_pps)
1348 tho->th_counter->tc_poll_pps(tho->th_counter);
1351 * Deal with NTP second processing. The for loop normally
1352 * iterates at most once, but in extreme situations it might
1353 * keep NTP sane if timeouts are not run for several seconds.
1354 * At boot, the time step can be large when the TOD hardware
1355 * has been read, so on really large steps, we call
1356 * ntp_update_second only twice. We need to call it twice in
1357 * case we missed a leap second.
1360 bintime_add(&bt, &th->th_boottime);
1361 i = bt.sec - tho->th_microtime.tv_sec;
1364 for (; i > 0; i--) {
1366 ntp_update_second(&th->th_adjustment, &bt.sec);
1368 th->th_boottime.sec += bt.sec - t;
1370 /* Update the UTC timestamps used by the get*() functions. */
1371 th->th_bintime = bt;
1372 bintime2timeval(&bt, &th->th_microtime);
1373 bintime2timespec(&bt, &th->th_nanotime);
1375 /* Now is a good time to change timecounters. */
1376 if (th->th_counter != timecounter) {
1378 if ((timecounter->tc_flags & TC_FLAGS_C2STOP) != 0)
1379 cpu_disable_c2_sleep++;
1380 if ((th->th_counter->tc_flags & TC_FLAGS_C2STOP) != 0)
1381 cpu_disable_c2_sleep--;
1383 th->th_counter = timecounter;
1384 th->th_offset_count = ncount;
1385 tc_min_ticktock_freq = max(1, timecounter->tc_frequency /
1386 (((uint64_t)timecounter->tc_counter_mask + 1) / 3));
1388 ffclock_change_tc(th);
1393 * Recalculate the scaling factor. We want the number of 1/2^64
1394 * fractions of a second per period of the hardware counter, taking
1395 * into account the th_adjustment factor which the NTP PLL/adjtime(2)
1396 * processing provides us with.
1398 * The th_adjustment is nanoseconds per second with 32 bit binary
1399 * fraction and we want 64 bit binary fraction of second:
1401 * x = a * 2^32 / 10^9 = a * 4.294967296
1403 * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
1404 * we can only multiply by about 850 without overflowing, that
1405 * leaves no suitably precise fractions for multiply before divide.
1407 * Divide before multiply with a fraction of 2199/512 results in a
1408 * systematic undercompensation of 10PPM of th_adjustment. On a
1409 * 5000PPM adjustment this is a 0.05PPM error. This is acceptable.
1411 * We happily sacrifice the lowest of the 64 bits of our result
1412 * to the goddess of code clarity.
1415 scale = (uint64_t)1 << 63;
1416 scale += (th->th_adjustment / 1024) * 2199;
1417 scale /= th->th_counter->tc_frequency;
1418 th->th_scale = scale * 2;
1419 th->th_large_delta = MIN(((uint64_t)1 << 63) / scale, UINT_MAX);
1422 * Now that the struct timehands is again consistent, set the new
1423 * generation number, making sure to not make it zero.
1427 atomic_store_rel_int(&th->th_generation, ogen);
1429 /* Go live with the new struct timehands. */
1431 switch (sysclock_active) {
1434 time_second = th->th_microtime.tv_sec;
1435 time_uptime = th->th_offset.sec;
1439 time_second = fftimehands->tick_time_lerp.sec;
1440 time_uptime = fftimehands->tick_time_lerp.sec - ffclock_boottime.sec;
1446 timekeep_push_vdso();
1449 /* Report or change the active timecounter hardware. */
1451 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
1454 struct timecounter *newtc, *tc;
1458 strlcpy(newname, tc->tc_name, sizeof(newname));
1460 error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
1461 if (error != 0 || req->newptr == NULL)
1463 /* Record that the tc in use now was specifically chosen. */
1465 if (strcmp(newname, tc->tc_name) == 0)
1467 for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
1468 if (strcmp(newname, newtc->tc_name) != 0)
1471 /* Warm up new timecounter. */
1472 (void)newtc->tc_get_timecount(newtc);
1474 timecounter = newtc;
1477 * The vdso timehands update is deferred until the next
1480 * This is prudent given that 'timekeep_push_vdso()' does not
1481 * use any locking and that it can be called in hard interrupt
1482 * context via 'tc_windup()'.
1489 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware,
1490 CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, 0, 0,
1491 sysctl_kern_timecounter_hardware, "A",
1492 "Timecounter hardware selected");
1495 /* Report the available timecounter hardware. */
1497 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
1500 struct timecounter *tc;
1503 sbuf_new_for_sysctl(&sb, NULL, 0, req);
1504 for (tc = timecounters; tc != NULL; tc = tc->tc_next) {
1505 if (tc != timecounters)
1506 sbuf_putc(&sb, ' ');
1507 sbuf_printf(&sb, "%s(%d)", tc->tc_name, tc->tc_quality);
1509 error = sbuf_finish(&sb);
1514 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice, CTLTYPE_STRING | CTLFLAG_RD,
1515 0, 0, sysctl_kern_timecounter_choice, "A", "Timecounter hardware detected");
1518 * RFC 2783 PPS-API implementation.
1522 * Return true if the driver is aware of the abi version extensions in the
1523 * pps_state structure, and it supports at least the given abi version number.
1526 abi_aware(struct pps_state *pps, int vers)
1529 return ((pps->kcmode & KCMODE_ABIFLAG) && pps->driver_abi >= vers);
1533 pps_fetch(struct pps_fetch_args *fapi, struct pps_state *pps)
1536 pps_seq_t aseq, cseq;
1539 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1543 * If no timeout is requested, immediately return whatever values were
1544 * most recently captured. If timeout seconds is -1, that's a request
1545 * to block without a timeout. WITNESS won't let us sleep forever
1546 * without a lock (we really don't need a lock), so just repeatedly
1547 * sleep a long time.
1549 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) {
1550 if (fapi->timeout.tv_sec == -1)
1553 tv.tv_sec = fapi->timeout.tv_sec;
1554 tv.tv_usec = fapi->timeout.tv_nsec / 1000;
1557 aseq = atomic_load_int(&pps->ppsinfo.assert_sequence);
1558 cseq = atomic_load_int(&pps->ppsinfo.clear_sequence);
1559 while (aseq == atomic_load_int(&pps->ppsinfo.assert_sequence) &&
1560 cseq == atomic_load_int(&pps->ppsinfo.clear_sequence)) {
1561 if (abi_aware(pps, 1) && pps->driver_mtx != NULL) {
1562 if (pps->flags & PPSFLAG_MTX_SPIN) {
1563 err = msleep_spin(pps, pps->driver_mtx,
1566 err = msleep(pps, pps->driver_mtx, PCATCH,
1570 err = tsleep(pps, PCATCH, "ppsfch", timo);
1572 if (err == EWOULDBLOCK) {
1573 if (fapi->timeout.tv_sec == -1) {
1578 } else if (err != 0) {
1584 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1585 fapi->pps_info_buf = pps->ppsinfo;
1591 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1594 struct pps_fetch_args *fapi;
1596 struct pps_fetch_ffc_args *fapi_ffc;
1599 struct pps_kcbind_args *kapi;
1602 KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
1604 case PPS_IOC_CREATE:
1606 case PPS_IOC_DESTROY:
1608 case PPS_IOC_SETPARAMS:
1609 app = (pps_params_t *)data;
1610 if (app->mode & ~pps->ppscap)
1613 /* Ensure only a single clock is selected for ffc timestamp. */
1614 if ((app->mode & PPS_TSCLK_MASK) == PPS_TSCLK_MASK)
1617 pps->ppsparam = *app;
1619 case PPS_IOC_GETPARAMS:
1620 app = (pps_params_t *)data;
1621 *app = pps->ppsparam;
1622 app->api_version = PPS_API_VERS_1;
1624 case PPS_IOC_GETCAP:
1625 *(int*)data = pps->ppscap;
1628 fapi = (struct pps_fetch_args *)data;
1629 return (pps_fetch(fapi, pps));
1631 case PPS_IOC_FETCH_FFCOUNTER:
1632 fapi_ffc = (struct pps_fetch_ffc_args *)data;
1633 if (fapi_ffc->tsformat && fapi_ffc->tsformat !=
1636 if (fapi_ffc->timeout.tv_sec || fapi_ffc->timeout.tv_nsec)
1637 return (EOPNOTSUPP);
1638 pps->ppsinfo_ffc.current_mode = pps->ppsparam.mode;
1639 fapi_ffc->pps_info_buf_ffc = pps->ppsinfo_ffc;
1640 /* Overwrite timestamps if feedback clock selected. */
1641 switch (pps->ppsparam.mode & PPS_TSCLK_MASK) {
1642 case PPS_TSCLK_FBCK:
1643 fapi_ffc->pps_info_buf_ffc.assert_timestamp =
1644 pps->ppsinfo.assert_timestamp;
1645 fapi_ffc->pps_info_buf_ffc.clear_timestamp =
1646 pps->ppsinfo.clear_timestamp;
1648 case PPS_TSCLK_FFWD:
1654 #endif /* FFCLOCK */
1655 case PPS_IOC_KCBIND:
1657 kapi = (struct pps_kcbind_args *)data;
1658 /* XXX Only root should be able to do this */
1659 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1661 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1663 if (kapi->edge & ~pps->ppscap)
1665 pps->kcmode = (kapi->edge & KCMODE_EDGEMASK) |
1666 (pps->kcmode & KCMODE_ABIFLAG);
1669 return (EOPNOTSUPP);
1677 pps_init(struct pps_state *pps)
1679 pps->ppscap |= PPS_TSFMT_TSPEC | PPS_CANWAIT;
1680 if (pps->ppscap & PPS_CAPTUREASSERT)
1681 pps->ppscap |= PPS_OFFSETASSERT;
1682 if (pps->ppscap & PPS_CAPTURECLEAR)
1683 pps->ppscap |= PPS_OFFSETCLEAR;
1685 pps->ppscap |= PPS_TSCLK_MASK;
1687 pps->kcmode &= ~KCMODE_ABIFLAG;
1691 pps_init_abi(struct pps_state *pps)
1695 if (pps->driver_abi > 0) {
1696 pps->kcmode |= KCMODE_ABIFLAG;
1697 pps->kernel_abi = PPS_ABI_VERSION;
1702 pps_capture(struct pps_state *pps)
1704 struct timehands *th;
1706 KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
1708 pps->capgen = atomic_load_acq_int(&th->th_generation);
1711 pps->capffth = fftimehands;
1713 pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
1714 atomic_thread_fence_acq();
1715 if (pps->capgen != th->th_generation)
1720 pps_event(struct pps_state *pps, int event)
1723 struct timespec ts, *tsp, *osp;
1724 u_int tcount, *pcount;
1728 struct timespec *tsp_ffc;
1729 pps_seq_t *pseq_ffc;
1736 KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
1737 /* Nothing to do if not currently set to capture this event type. */
1738 if ((event & pps->ppsparam.mode) == 0)
1740 /* If the timecounter was wound up underneath us, bail out. */
1741 if (pps->capgen == 0 || pps->capgen !=
1742 atomic_load_acq_int(&pps->capth->th_generation))
1745 /* Things would be easier with arrays. */
1746 if (event == PPS_CAPTUREASSERT) {
1747 tsp = &pps->ppsinfo.assert_timestamp;
1748 osp = &pps->ppsparam.assert_offset;
1749 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1751 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1753 pcount = &pps->ppscount[0];
1754 pseq = &pps->ppsinfo.assert_sequence;
1756 ffcount = &pps->ppsinfo_ffc.assert_ffcount;
1757 tsp_ffc = &pps->ppsinfo_ffc.assert_timestamp;
1758 pseq_ffc = &pps->ppsinfo_ffc.assert_sequence;
1761 tsp = &pps->ppsinfo.clear_timestamp;
1762 osp = &pps->ppsparam.clear_offset;
1763 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1765 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1767 pcount = &pps->ppscount[1];
1768 pseq = &pps->ppsinfo.clear_sequence;
1770 ffcount = &pps->ppsinfo_ffc.clear_ffcount;
1771 tsp_ffc = &pps->ppsinfo_ffc.clear_timestamp;
1772 pseq_ffc = &pps->ppsinfo_ffc.clear_sequence;
1777 * If the timecounter changed, we cannot compare the count values, so
1778 * we have to drop the rest of the PPS-stuff until the next event.
1780 if (pps->ppstc != pps->capth->th_counter) {
1781 pps->ppstc = pps->capth->th_counter;
1782 *pcount = pps->capcount;
1783 pps->ppscount[2] = pps->capcount;
1787 /* Convert the count to a timespec. */
1788 tcount = pps->capcount - pps->capth->th_offset_count;
1789 tcount &= pps->capth->th_counter->tc_counter_mask;
1790 bt = pps->capth->th_bintime;
1791 bintime_addx(&bt, pps->capth->th_scale * tcount);
1792 bintime2timespec(&bt, &ts);
1794 /* If the timecounter was wound up underneath us, bail out. */
1795 atomic_thread_fence_acq();
1796 if (pps->capgen != pps->capth->th_generation)
1799 *pcount = pps->capcount;
1804 timespecadd(tsp, osp, tsp);
1805 if (tsp->tv_nsec < 0) {
1806 tsp->tv_nsec += 1000000000;
1812 *ffcount = pps->capffth->tick_ffcount + tcount;
1813 bt = pps->capffth->tick_time;
1814 ffclock_convert_delta(tcount, pps->capffth->cest.period, &bt);
1815 bintime_add(&bt, &pps->capffth->tick_time);
1816 bintime2timespec(&bt, &ts);
1826 * Feed the NTP PLL/FLL.
1827 * The FLL wants to know how many (hardware) nanoseconds
1828 * elapsed since the previous event.
1830 tcount = pps->capcount - pps->ppscount[2];
1831 pps->ppscount[2] = pps->capcount;
1832 tcount &= pps->capth->th_counter->tc_counter_mask;
1833 scale = (uint64_t)1 << 63;
1834 scale /= pps->capth->th_counter->tc_frequency;
1838 bintime_addx(&bt, scale * tcount);
1839 bintime2timespec(&bt, &ts);
1840 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
1844 /* Wakeup anyone sleeping in pps_fetch(). */
1849 * Timecounters need to be updated every so often to prevent the hardware
1850 * counter from overflowing. Updating also recalculates the cached values
1851 * used by the get*() family of functions, so their precision depends on
1852 * the update frequency.
1856 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0,
1857 "Approximate number of hardclock ticks in a millisecond");
1860 tc_ticktock(int cnt)
1864 if (mtx_trylock_spin(&tc_setclock_mtx)) {
1866 if (count >= tc_tick) {
1870 mtx_unlock_spin(&tc_setclock_mtx);
1874 static void __inline
1875 tc_adjprecision(void)
1879 if (tc_timepercentage > 0) {
1880 t = (99 + tc_timepercentage) / tc_timepercentage;
1881 tc_precexp = fls(t + (t >> 1)) - 1;
1882 FREQ2BT(hz / tc_tick, &bt_timethreshold);
1883 FREQ2BT(hz, &bt_tickthreshold);
1884 bintime_shift(&bt_timethreshold, tc_precexp);
1885 bintime_shift(&bt_tickthreshold, tc_precexp);
1888 bt_timethreshold.sec = INT_MAX;
1889 bt_timethreshold.frac = ~(uint64_t)0;
1890 bt_tickthreshold = bt_timethreshold;
1892 sbt_timethreshold = bttosbt(bt_timethreshold);
1893 sbt_tickthreshold = bttosbt(bt_tickthreshold);
1897 sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS)
1901 val = tc_timepercentage;
1902 error = sysctl_handle_int(oidp, &val, 0, req);
1903 if (error != 0 || req->newptr == NULL)
1905 tc_timepercentage = val;
1913 /* Set up the requested number of timehands. */
1915 inittimehands(void *dummy)
1917 struct timehands *thp;
1920 TUNABLE_INT_FETCH("kern.timecounter.timehands_count",
1922 if (timehands_count < 1)
1923 timehands_count = 1;
1924 if (timehands_count > nitems(ths))
1925 timehands_count = nitems(ths);
1926 for (i = 1, thp = &ths[0]; i < timehands_count; thp = &ths[i++])
1927 thp->th_next = &ths[i];
1928 thp->th_next = &ths[0];
1930 TUNABLE_STR_FETCH("kern.timecounter.hardware", tc_from_tunable,
1931 sizeof(tc_from_tunable));
1933 SYSINIT(timehands, SI_SUB_TUNABLES, SI_ORDER_ANY, inittimehands, NULL);
1936 inittimecounter(void *dummy)
1942 * Set the initial timeout to
1943 * max(1, <approx. number of hardclock ticks in a millisecond>).
1944 * People should probably not use the sysctl to set the timeout
1945 * to smaller than its initial value, since that value is the
1946 * smallest reasonable one. If they want better timestamps they
1947 * should use the non-"get"* functions.
1950 tc_tick = (hz + 500) / 1000;
1954 FREQ2BT(hz, &tick_bt);
1955 tick_sbt = bttosbt(tick_bt);
1956 tick_rate = hz / tc_tick;
1957 FREQ2BT(tick_rate, &tc_tick_bt);
1958 tc_tick_sbt = bttosbt(tc_tick_bt);
1959 p = (tc_tick * 1000000) / hz;
1960 printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
1966 /* warm up new timecounter (again) and get rolling. */
1967 (void)timecounter->tc_get_timecount(timecounter);
1968 mtx_lock_spin(&tc_setclock_mtx);
1970 mtx_unlock_spin(&tc_setclock_mtx);
1973 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL);
1975 /* Cpu tick handling -------------------------------------------------*/
1977 static int cpu_tick_variable;
1978 static uint64_t cpu_tick_frequency;
1980 DPCPU_DEFINE_STATIC(uint64_t, tc_cpu_ticks_base);
1981 DPCPU_DEFINE_STATIC(unsigned, tc_cpu_ticks_last);
1986 struct timecounter *tc;
1987 uint64_t res, *base;
1991 base = DPCPU_PTR(tc_cpu_ticks_base);
1992 last = DPCPU_PTR(tc_cpu_ticks_last);
1993 tc = timehands->th_counter;
1994 u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
1996 *base += (uint64_t)tc->tc_counter_mask + 1;
2004 cpu_tick_calibration(void)
2006 static time_t last_calib;
2008 if (time_uptime != last_calib && !(time_uptime & 0xf)) {
2009 cpu_tick_calibrate(0);
2010 last_calib = time_uptime;
2015 * This function gets called every 16 seconds on only one designated
2016 * CPU in the system from hardclock() via cpu_tick_calibration()().
2018 * Whenever the real time clock is stepped we get called with reset=1
2019 * to make sure we handle suspend/resume and similar events correctly.
2023 cpu_tick_calibrate(int reset)
2025 static uint64_t c_last;
2026 uint64_t c_this, c_delta;
2027 static struct bintime t_last;
2028 struct bintime t_this, t_delta;
2032 /* The clock was stepped, abort & reset */
2037 /* we don't calibrate fixed rate cputicks */
2038 if (!cpu_tick_variable)
2041 getbinuptime(&t_this);
2042 c_this = cpu_ticks();
2043 if (t_last.sec != 0) {
2044 c_delta = c_this - c_last;
2046 bintime_sub(&t_delta, &t_last);
2049 * 2^(64-20) / 16[s] =
2051 * 17.592.186.044.416 / 16 =
2052 * 1.099.511.627.776 [Hz]
2054 divi = t_delta.sec << 20;
2055 divi |= t_delta.frac >> (64 - 20);
2058 if (c_delta > cpu_tick_frequency) {
2059 if (0 && bootverbose)
2060 printf("cpu_tick increased to %ju Hz\n",
2062 cpu_tick_frequency = c_delta;
2070 set_cputicker(cpu_tick_f *func, uint64_t freq, unsigned var)
2074 cpu_ticks = tc_cpu_ticks;
2076 cpu_tick_frequency = freq;
2077 cpu_tick_variable = var;
2086 if (cpu_ticks == tc_cpu_ticks)
2087 return (tc_getfrequency());
2088 return (cpu_tick_frequency);
2092 * We need to be slightly careful converting cputicks to microseconds.
2093 * There is plenty of margin in 64 bits of microseconds (half a million
2094 * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
2095 * before divide conversion (to retain precision) we find that the
2096 * margin shrinks to 1.5 hours (one millionth of 146y).
2097 * With a three prong approach we never lose significant bits, no
2098 * matter what the cputick rate and length of timeinterval is.
2102 cputick2usec(uint64_t tick)
2105 if (tick > 18446744073709551LL) /* floor(2^64 / 1000) */
2106 return (tick / (cpu_tickrate() / 1000000LL));
2107 else if (tick > 18446744073709LL) /* floor(2^64 / 1000000) */
2108 return ((tick * 1000LL) / (cpu_tickrate() / 1000LL));
2110 return ((tick * 1000000LL) / cpu_tickrate());
2113 cpu_tick_f *cpu_ticks = tc_cpu_ticks;
2115 static int vdso_th_enable = 1;
2117 sysctl_fast_gettime(SYSCTL_HANDLER_ARGS)
2119 int old_vdso_th_enable, error;
2121 old_vdso_th_enable = vdso_th_enable;
2122 error = sysctl_handle_int(oidp, &old_vdso_th_enable, 0, req);
2125 vdso_th_enable = old_vdso_th_enable;
2128 SYSCTL_PROC(_kern_timecounter, OID_AUTO, fast_gettime,
2129 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
2130 NULL, 0, sysctl_fast_gettime, "I", "Enable fast time of day");
2133 tc_fill_vdso_timehands(struct vdso_timehands *vdso_th)
2135 struct timehands *th;
2139 vdso_th->th_scale = th->th_scale;
2140 vdso_th->th_offset_count = th->th_offset_count;
2141 vdso_th->th_counter_mask = th->th_counter->tc_counter_mask;
2142 vdso_th->th_offset = th->th_offset;
2143 vdso_th->th_boottime = th->th_boottime;
2144 if (th->th_counter->tc_fill_vdso_timehands != NULL) {
2145 enabled = th->th_counter->tc_fill_vdso_timehands(vdso_th,
2149 if (!vdso_th_enable)
2154 #ifdef COMPAT_FREEBSD32
2156 tc_fill_vdso_timehands32(struct vdso_timehands32 *vdso_th32)
2158 struct timehands *th;
2162 *(uint64_t *)&vdso_th32->th_scale[0] = th->th_scale;
2163 vdso_th32->th_offset_count = th->th_offset_count;
2164 vdso_th32->th_counter_mask = th->th_counter->tc_counter_mask;
2165 vdso_th32->th_offset.sec = th->th_offset.sec;
2166 *(uint64_t *)&vdso_th32->th_offset.frac[0] = th->th_offset.frac;
2167 vdso_th32->th_boottime.sec = th->th_boottime.sec;
2168 *(uint64_t *)&vdso_th32->th_boottime.frac[0] = th->th_boottime.frac;
2169 if (th->th_counter->tc_fill_vdso_timehands32 != NULL) {
2170 enabled = th->th_counter->tc_fill_vdso_timehands32(vdso_th32,
2174 if (!vdso_th_enable)
2180 #include "opt_ddb.h"
2182 #include <ddb/ddb.h>
2184 DB_SHOW_COMMAND(timecounter, db_show_timecounter)
2186 struct timehands *th;
2187 struct timecounter *tc;
2191 tc = th->th_counter;
2192 val1 = tc->tc_get_timecount(tc);
2193 __compiler_membar();
2194 val2 = tc->tc_get_timecount(tc);
2196 db_printf("timecounter %p %s\n", tc, tc->tc_name);
2197 db_printf(" mask %#x freq %ju qual %d flags %#x priv %p\n",
2198 tc->tc_counter_mask, (uintmax_t)tc->tc_frequency, tc->tc_quality,
2199 tc->tc_flags, tc->tc_priv);
2200 db_printf(" val %#x %#x\n", val1, val2);
2201 db_printf("timehands adj %#jx scale %#jx ldelta %d off_cnt %d gen %d\n",
2202 (uintmax_t)th->th_adjustment, (uintmax_t)th->th_scale,
2203 th->th_large_delta, th->th_offset_count, th->th_generation);
2204 db_printf(" offset %jd %jd boottime %jd %jd\n",
2205 (intmax_t)th->th_offset.sec, (uintmax_t)th->th_offset.frac,
2206 (intmax_t)th->th_boottime.sec, (uintmax_t)th->th_boottime.frac);