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
13 * Portions of this software were developed by Julien Ridoux at the University
14 * of Melbourne under sponsorship from the FreeBSD Foundation.
16 * Portions of this software were developed by Konstantin Belousov
17 * under sponsorship from the FreeBSD Foundation.
20 #include <sys/cdefs.h>
21 __FBSDID("$FreeBSD$");
24 #include "opt_ffclock.h"
26 #include <sys/param.h>
27 #include <sys/kernel.h>
28 #include <sys/limits.h>
30 #include <sys/mutex.h>
33 #include <sys/sleepqueue.h>
34 #include <sys/sysctl.h>
35 #include <sys/syslog.h>
36 #include <sys/systm.h>
37 #include <sys/timeffc.h>
38 #include <sys/timepps.h>
39 #include <sys/timetc.h>
40 #include <sys/timex.h>
44 * A large step happens on boot. This constant detects such steps.
45 * It is relatively small so that ntp_update_second gets called enough
46 * in the typical 'missed a couple of seconds' case, but doesn't loop
47 * forever when the time step is large.
49 #define LARGE_STEP 200
52 * Implement a dummy timecounter which we can use until we get a real one
53 * in the air. This allows the console and other early stuff to use
58 dummy_get_timecount(struct timecounter *tc)
65 static struct timecounter dummy_timecounter = {
66 dummy_get_timecount, 0, ~0u, 1000000, "dummy", -1000000
70 /* These fields must be initialized by the driver. */
71 struct timecounter *th_counter;
72 int64_t th_adjustment;
75 u_int th_offset_count;
76 struct bintime th_offset;
77 struct bintime th_bintime;
78 struct timeval th_microtime;
79 struct timespec th_nanotime;
80 struct bintime th_boottime;
81 /* Fields not to be copied in tc_windup start with th_generation. */
83 struct timehands *th_next;
86 static struct timehands ths[16] = {
88 .th_counter = &dummy_timecounter,
89 .th_scale = (uint64_t)-1 / 1000000,
90 .th_large_delta = 1000000,
91 .th_offset = { .sec = 1 },
96 static struct timehands *volatile timehands = &ths[0];
97 struct timecounter *timecounter = &dummy_timecounter;
98 static struct timecounter *timecounters = &dummy_timecounter;
100 /* Mutex to protect the timecounter list. */
101 static struct mtx tc_lock;
103 int tc_min_ticktock_freq = 1;
105 volatile time_t time_second = 1;
106 volatile time_t time_uptime = 1;
108 static int sysctl_kern_boottime(SYSCTL_HANDLER_ARGS);
109 SYSCTL_PROC(_kern, KERN_BOOTTIME, boottime,
110 CTLTYPE_STRUCT | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
111 sysctl_kern_boottime, "S,timeval",
114 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
116 static SYSCTL_NODE(_kern_timecounter, OID_AUTO, tc,
117 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
120 static int timestepwarnings;
121 SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RWTUN,
122 ×tepwarnings, 0, "Log time steps");
124 static int timehands_count = 2;
125 SYSCTL_INT(_kern_timecounter, OID_AUTO, timehands_count,
126 CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
127 &timehands_count, 0, "Count of timehands in rotation");
129 struct bintime bt_timethreshold;
130 struct bintime bt_tickthreshold;
131 sbintime_t sbt_timethreshold;
132 sbintime_t sbt_tickthreshold;
133 struct bintime tc_tick_bt;
134 sbintime_t tc_tick_sbt;
136 int tc_timepercentage = TC_DEFAULTPERC;
137 static int sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS);
138 SYSCTL_PROC(_kern_timecounter, OID_AUTO, alloweddeviation,
139 CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, 0, 0,
140 sysctl_kern_timecounter_adjprecision, "I",
141 "Allowed time interval deviation in percents");
143 volatile int rtc_generation = 1;
145 static int tc_chosen; /* Non-zero if a specific tc was chosen via sysctl. */
146 static char tc_from_tunable[16];
148 static void tc_windup(struct bintime *new_boottimebin);
149 static void cpu_tick_calibrate(int);
151 void dtrace_getnanotime(struct timespec *tsp);
152 void dtrace_getnanouptime(struct timespec *tsp);
155 sysctl_kern_boottime(SYSCTL_HANDLER_ARGS)
157 struct timeval boottime;
159 getboottime(&boottime);
161 /* i386 is the only arch which uses a 32bits time_t */
166 if (req->flags & SCTL_MASK32) {
167 tv[0] = boottime.tv_sec;
168 tv[1] = boottime.tv_usec;
169 return (SYSCTL_OUT(req, tv, sizeof(tv)));
173 return (SYSCTL_OUT(req, &boottime, sizeof(boottime)));
177 sysctl_kern_timecounter_get(SYSCTL_HANDLER_ARGS)
180 struct timecounter *tc = arg1;
182 ncount = tc->tc_get_timecount(tc);
183 return (sysctl_handle_int(oidp, &ncount, 0, req));
187 sysctl_kern_timecounter_freq(SYSCTL_HANDLER_ARGS)
190 struct timecounter *tc = arg1;
192 freq = tc->tc_frequency;
193 return (sysctl_handle_64(oidp, &freq, 0, req));
197 * Return the difference between the timehands' counter value now and what
198 * was when we copied it to the timehands' offset_count.
200 static __inline u_int
201 tc_delta(struct timehands *th)
203 struct timecounter *tc;
206 return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
207 tc->tc_counter_mask);
211 bintime_add_tc_delta(struct bintime *bt, uint64_t scale,
212 uint64_t large_delta, uint64_t delta)
216 if (__predict_false(delta >= large_delta)) {
217 /* Avoid overflow for scale * delta. */
218 x = (scale >> 32) * delta;
220 bintime_addx(bt, x << 32);
221 bintime_addx(bt, (scale & 0xffffffff) * delta);
223 bintime_addx(bt, scale * delta);
228 * Functions for reading the time. We have to loop until we are sure that
229 * the timehands that we operated on was not updated under our feet. See
230 * the comment in <sys/time.h> for a description of these 12 functions.
234 bintime_off(struct bintime *bt, u_int off)
236 struct timehands *th;
239 u_int delta, gen, large_delta;
243 gen = atomic_load_acq_int(&th->th_generation);
244 btp = (struct bintime *)((vm_offset_t)th + off);
246 scale = th->th_scale;
247 delta = tc_delta(th);
248 large_delta = th->th_large_delta;
249 atomic_thread_fence_acq();
250 } while (gen == 0 || gen != th->th_generation);
252 bintime_add_tc_delta(bt, scale, large_delta, delta);
254 #define GETTHBINTIME(dst, member) \
256 _Static_assert(_Generic(((struct timehands *)NULL)->member, \
257 struct bintime: 1, default: 0) == 1, \
258 "struct timehands member is not of struct bintime type"); \
259 bintime_off(dst, __offsetof(struct timehands, member)); \
263 getthmember(void *out, size_t out_size, u_int off)
265 struct timehands *th;
270 gen = atomic_load_acq_int(&th->th_generation);
271 memcpy(out, (char *)th + off, out_size);
272 atomic_thread_fence_acq();
273 } while (gen == 0 || gen != th->th_generation);
275 #define GETTHMEMBER(dst, member) \
277 _Static_assert(_Generic(*dst, \
278 __typeof(((struct timehands *)NULL)->member): 1, \
280 "*dst and struct timehands member have different types"); \
281 getthmember(dst, sizeof(*dst), __offsetof(struct timehands, \
287 fbclock_binuptime(struct bintime *bt)
290 GETTHBINTIME(bt, th_offset);
294 fbclock_nanouptime(struct timespec *tsp)
298 fbclock_binuptime(&bt);
299 bintime2timespec(&bt, tsp);
303 fbclock_microuptime(struct timeval *tvp)
307 fbclock_binuptime(&bt);
308 bintime2timeval(&bt, tvp);
312 fbclock_bintime(struct bintime *bt)
315 GETTHBINTIME(bt, th_bintime);
319 fbclock_nanotime(struct timespec *tsp)
323 fbclock_bintime(&bt);
324 bintime2timespec(&bt, tsp);
328 fbclock_microtime(struct timeval *tvp)
332 fbclock_bintime(&bt);
333 bintime2timeval(&bt, tvp);
337 fbclock_getbinuptime(struct bintime *bt)
340 GETTHMEMBER(bt, th_offset);
344 fbclock_getnanouptime(struct timespec *tsp)
348 GETTHMEMBER(&bt, th_offset);
349 bintime2timespec(&bt, tsp);
353 fbclock_getmicrouptime(struct timeval *tvp)
357 GETTHMEMBER(&bt, th_offset);
358 bintime2timeval(&bt, tvp);
362 fbclock_getbintime(struct bintime *bt)
365 GETTHMEMBER(bt, th_bintime);
369 fbclock_getnanotime(struct timespec *tsp)
372 GETTHMEMBER(tsp, th_nanotime);
376 fbclock_getmicrotime(struct timeval *tvp)
379 GETTHMEMBER(tvp, th_microtime);
384 binuptime(struct bintime *bt)
387 GETTHBINTIME(bt, th_offset);
391 nanouptime(struct timespec *tsp)
396 bintime2timespec(&bt, tsp);
400 microuptime(struct timeval *tvp)
405 bintime2timeval(&bt, tvp);
409 bintime(struct bintime *bt)
412 GETTHBINTIME(bt, th_bintime);
416 nanotime(struct timespec *tsp)
421 bintime2timespec(&bt, tsp);
425 microtime(struct timeval *tvp)
430 bintime2timeval(&bt, tvp);
434 getbinuptime(struct bintime *bt)
437 GETTHMEMBER(bt, th_offset);
441 getnanouptime(struct timespec *tsp)
445 GETTHMEMBER(&bt, th_offset);
446 bintime2timespec(&bt, tsp);
450 getmicrouptime(struct timeval *tvp)
454 GETTHMEMBER(&bt, th_offset);
455 bintime2timeval(&bt, tvp);
459 getbintime(struct bintime *bt)
462 GETTHMEMBER(bt, th_bintime);
466 getnanotime(struct timespec *tsp)
469 GETTHMEMBER(tsp, th_nanotime);
473 getmicrotime(struct timeval *tvp)
476 GETTHMEMBER(tvp, th_microtime);
481 getboottime(struct timeval *boottime)
483 struct bintime boottimebin;
485 getboottimebin(&boottimebin);
486 bintime2timeval(&boottimebin, boottime);
490 getboottimebin(struct bintime *boottimebin)
493 GETTHMEMBER(boottimebin, th_boottime);
498 * Support for feed-forward synchronization algorithms. This is heavily inspired
499 * by the timehands mechanism but kept independent from it. *_windup() functions
500 * have some connection to avoid accessing the timecounter hardware more than
504 /* Feed-forward clock estimates kept updated by the synchronization daemon. */
505 struct ffclock_estimate ffclock_estimate;
506 struct bintime ffclock_boottime; /* Feed-forward boot time estimate. */
507 uint32_t ffclock_status; /* Feed-forward clock status. */
508 int8_t ffclock_updated; /* New estimates are available. */
509 struct mtx ffclock_mtx; /* Mutex on ffclock_estimate. */
512 struct ffclock_estimate cest;
513 struct bintime tick_time;
514 struct bintime tick_time_lerp;
515 ffcounter tick_ffcount;
516 uint64_t period_lerp;
517 volatile uint8_t gen;
518 struct fftimehands *next;
521 #define NUM_ELEMENTS(x) (sizeof(x) / sizeof(*x))
523 static struct fftimehands ffth[10];
524 static struct fftimehands *volatile fftimehands = ffth;
529 struct fftimehands *cur;
530 struct fftimehands *last;
532 memset(ffth, 0, sizeof(ffth));
534 last = ffth + NUM_ELEMENTS(ffth) - 1;
535 for (cur = ffth; cur < last; cur++)
540 ffclock_status = FFCLOCK_STA_UNSYNC;
541 mtx_init(&ffclock_mtx, "ffclock lock", NULL, MTX_DEF);
545 * Reset the feed-forward clock estimates. Called from inittodr() to get things
546 * kick started and uses the timecounter nominal frequency as a first period
547 * estimate. Note: this function may be called several time just after boot.
548 * Note: this is the only function that sets the value of boot time for the
549 * monotonic (i.e. uptime) version of the feed-forward clock.
552 ffclock_reset_clock(struct timespec *ts)
554 struct timecounter *tc;
555 struct ffclock_estimate cest;
557 tc = timehands->th_counter;
558 memset(&cest, 0, sizeof(struct ffclock_estimate));
560 timespec2bintime(ts, &ffclock_boottime);
561 timespec2bintime(ts, &(cest.update_time));
562 ffclock_read_counter(&cest.update_ffcount);
563 cest.leapsec_next = 0;
564 cest.period = ((1ULL << 63) / tc->tc_frequency) << 1;
567 cest.status = FFCLOCK_STA_UNSYNC;
568 cest.leapsec_total = 0;
571 mtx_lock(&ffclock_mtx);
572 bcopy(&cest, &ffclock_estimate, sizeof(struct ffclock_estimate));
573 ffclock_updated = INT8_MAX;
574 mtx_unlock(&ffclock_mtx);
576 printf("ffclock reset: %s (%llu Hz), time = %ld.%09lu\n", tc->tc_name,
577 (unsigned long long)tc->tc_frequency, (long)ts->tv_sec,
578 (unsigned long)ts->tv_nsec);
582 * Sub-routine to convert a time interval measured in RAW counter units to time
583 * in seconds stored in bintime format.
584 * NOTE: bintime_mul requires u_int, but the value of the ffcounter may be
585 * larger than the max value of u_int (on 32 bit architecture). Loop to consume
589 ffclock_convert_delta(ffcounter ffdelta, uint64_t period, struct bintime *bt)
592 ffcounter delta, delta_max;
594 delta_max = (1ULL << (8 * sizeof(unsigned int))) - 1;
597 if (ffdelta > delta_max)
603 bintime_mul(&bt2, (unsigned int)delta);
604 bintime_add(bt, &bt2);
606 } while (ffdelta > 0);
610 * Update the fftimehands.
611 * Push the tick ffcount and time(s) forward based on current clock estimate.
612 * The conversion from ffcounter to bintime relies on the difference clock
613 * principle, whose accuracy relies on computing small time intervals. If a new
614 * clock estimate has been passed by the synchronisation daemon, make it
615 * current, and compute the linear interpolation for monotonic time if needed.
618 ffclock_windup(unsigned int delta)
620 struct ffclock_estimate *cest;
621 struct fftimehands *ffth;
622 struct bintime bt, gap_lerp;
625 unsigned int polling;
626 uint8_t forward_jump, ogen;
629 * Pick the next timehand, copy current ffclock estimates and move tick
630 * times and counter forward.
633 ffth = fftimehands->next;
637 bcopy(&fftimehands->cest, cest, sizeof(struct ffclock_estimate));
638 ffdelta = (ffcounter)delta;
639 ffth->period_lerp = fftimehands->period_lerp;
641 ffth->tick_time = fftimehands->tick_time;
642 ffclock_convert_delta(ffdelta, cest->period, &bt);
643 bintime_add(&ffth->tick_time, &bt);
645 ffth->tick_time_lerp = fftimehands->tick_time_lerp;
646 ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt);
647 bintime_add(&ffth->tick_time_lerp, &bt);
649 ffth->tick_ffcount = fftimehands->tick_ffcount + ffdelta;
652 * Assess the status of the clock, if the last update is too old, it is
653 * likely the synchronisation daemon is dead and the clock is free
656 if (ffclock_updated == 0) {
657 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
658 ffclock_convert_delta(ffdelta, cest->period, &bt);
659 if (bt.sec > 2 * FFCLOCK_SKM_SCALE)
660 ffclock_status |= FFCLOCK_STA_UNSYNC;
664 * If available, grab updated clock estimates and make them current.
665 * Recompute time at this tick using the updated estimates. The clock
666 * estimates passed the feed-forward synchronisation daemon may result
667 * in time conversion that is not monotonically increasing (just after
668 * the update). time_lerp is a particular linear interpolation over the
669 * synchronisation algo polling period that ensures monotonicity for the
670 * clock ids requesting it.
672 if (ffclock_updated > 0) {
673 bcopy(&ffclock_estimate, cest, sizeof(struct ffclock_estimate));
674 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
675 ffth->tick_time = cest->update_time;
676 ffclock_convert_delta(ffdelta, cest->period, &bt);
677 bintime_add(&ffth->tick_time, &bt);
679 /* ffclock_reset sets ffclock_updated to INT8_MAX */
680 if (ffclock_updated == INT8_MAX)
681 ffth->tick_time_lerp = ffth->tick_time;
683 if (bintime_cmp(&ffth->tick_time, &ffth->tick_time_lerp, >))
688 bintime_clear(&gap_lerp);
690 gap_lerp = ffth->tick_time;
691 bintime_sub(&gap_lerp, &ffth->tick_time_lerp);
693 gap_lerp = ffth->tick_time_lerp;
694 bintime_sub(&gap_lerp, &ffth->tick_time);
698 * The reset from the RTC clock may be far from accurate, and
699 * reducing the gap between real time and interpolated time
700 * could take a very long time if the interpolated clock insists
701 * on strict monotonicity. The clock is reset under very strict
702 * conditions (kernel time is known to be wrong and
703 * synchronization daemon has been restarted recently.
704 * ffclock_boottime absorbs the jump to ensure boot time is
705 * correct and uptime functions stay consistent.
707 if (((ffclock_status & FFCLOCK_STA_UNSYNC) == FFCLOCK_STA_UNSYNC) &&
708 ((cest->status & FFCLOCK_STA_UNSYNC) == 0) &&
709 ((cest->status & FFCLOCK_STA_WARMUP) == FFCLOCK_STA_WARMUP)) {
711 bintime_add(&ffclock_boottime, &gap_lerp);
713 bintime_sub(&ffclock_boottime, &gap_lerp);
714 ffth->tick_time_lerp = ffth->tick_time;
715 bintime_clear(&gap_lerp);
718 ffclock_status = cest->status;
719 ffth->period_lerp = cest->period;
722 * Compute corrected period used for the linear interpolation of
723 * time. The rate of linear interpolation is capped to 5000PPM
726 if (bintime_isset(&gap_lerp)) {
727 ffdelta = cest->update_ffcount;
728 ffdelta -= fftimehands->cest.update_ffcount;
729 ffclock_convert_delta(ffdelta, cest->period, &bt);
732 bt.frac = 5000000 * (uint64_t)18446744073LL;
733 bintime_mul(&bt, polling);
734 if (bintime_cmp(&gap_lerp, &bt, >))
737 /* Approximate 1 sec by 1-(1/2^64) to ease arithmetic */
739 if (gap_lerp.sec > 0) {
741 frac /= ffdelta / gap_lerp.sec;
743 frac += gap_lerp.frac / ffdelta;
746 ffth->period_lerp += frac;
748 ffth->period_lerp -= frac;
760 * Adjust the fftimehands when the timecounter is changed. Stating the obvious,
761 * the old and new hardware counter cannot be read simultaneously. tc_windup()
762 * does read the two counters 'back to back', but a few cycles are effectively
763 * lost, and not accumulated in tick_ffcount. This is a fairly radical
764 * operation for a feed-forward synchronization daemon, and it is its job to not
765 * pushing irrelevant data to the kernel. Because there is no locking here,
766 * simply force to ignore pending or next update to give daemon a chance to
767 * realize the counter has changed.
770 ffclock_change_tc(struct timehands *th)
772 struct fftimehands *ffth;
773 struct ffclock_estimate *cest;
774 struct timecounter *tc;
778 ffth = fftimehands->next;
783 bcopy(&(fftimehands->cest), cest, sizeof(struct ffclock_estimate));
784 cest->period = ((1ULL << 63) / tc->tc_frequency ) << 1;
787 cest->status |= FFCLOCK_STA_UNSYNC;
789 ffth->tick_ffcount = fftimehands->tick_ffcount;
790 ffth->tick_time_lerp = fftimehands->tick_time_lerp;
791 ffth->tick_time = fftimehands->tick_time;
792 ffth->period_lerp = cest->period;
794 /* Do not lock but ignore next update from synchronization daemon. */
804 * Retrieve feed-forward counter and time of last kernel tick.
807 ffclock_last_tick(ffcounter *ffcount, struct bintime *bt, uint32_t flags)
809 struct fftimehands *ffth;
813 * No locking but check generation has not changed. Also need to make
814 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
819 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP)
820 *bt = ffth->tick_time_lerp;
822 *bt = ffth->tick_time;
823 *ffcount = ffth->tick_ffcount;
824 } while (gen == 0 || gen != ffth->gen);
828 * Absolute clock conversion. Low level function to convert ffcounter to
829 * bintime. The ffcounter is converted using the current ffclock period estimate
830 * or the "interpolated period" to ensure monotonicity.
831 * NOTE: this conversion may have been deferred, and the clock updated since the
832 * hardware counter has been read.
835 ffclock_convert_abs(ffcounter ffcount, struct bintime *bt, uint32_t flags)
837 struct fftimehands *ffth;
843 * No locking but check generation has not changed. Also need to make
844 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
849 if (ffcount > ffth->tick_ffcount)
850 ffdelta = ffcount - ffth->tick_ffcount;
852 ffdelta = ffth->tick_ffcount - ffcount;
854 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP) {
855 *bt = ffth->tick_time_lerp;
856 ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt2);
858 *bt = ffth->tick_time;
859 ffclock_convert_delta(ffdelta, ffth->cest.period, &bt2);
862 if (ffcount > ffth->tick_ffcount)
863 bintime_add(bt, &bt2);
865 bintime_sub(bt, &bt2);
866 } while (gen == 0 || gen != ffth->gen);
870 * Difference clock conversion.
871 * Low level function to Convert a time interval measured in RAW counter units
872 * into bintime. The difference clock allows measuring small intervals much more
873 * reliably than the absolute clock.
876 ffclock_convert_diff(ffcounter ffdelta, struct bintime *bt)
878 struct fftimehands *ffth;
881 /* No locking but check generation has not changed. */
885 ffclock_convert_delta(ffdelta, ffth->cest.period, bt);
886 } while (gen == 0 || gen != ffth->gen);
890 * Access to current ffcounter value.
893 ffclock_read_counter(ffcounter *ffcount)
895 struct timehands *th;
896 struct fftimehands *ffth;
897 unsigned int gen, delta;
900 * ffclock_windup() called from tc_windup(), safe to rely on
901 * th->th_generation only, for correct delta and ffcounter.
905 gen = atomic_load_acq_int(&th->th_generation);
907 delta = tc_delta(th);
908 *ffcount = ffth->tick_ffcount;
909 atomic_thread_fence_acq();
910 } while (gen == 0 || gen != th->th_generation);
916 binuptime(struct bintime *bt)
919 binuptime_fromclock(bt, sysclock_active);
923 nanouptime(struct timespec *tsp)
926 nanouptime_fromclock(tsp, sysclock_active);
930 microuptime(struct timeval *tvp)
933 microuptime_fromclock(tvp, sysclock_active);
937 bintime(struct bintime *bt)
940 bintime_fromclock(bt, sysclock_active);
944 nanotime(struct timespec *tsp)
947 nanotime_fromclock(tsp, sysclock_active);
951 microtime(struct timeval *tvp)
954 microtime_fromclock(tvp, sysclock_active);
958 getbinuptime(struct bintime *bt)
961 getbinuptime_fromclock(bt, sysclock_active);
965 getnanouptime(struct timespec *tsp)
968 getnanouptime_fromclock(tsp, sysclock_active);
972 getmicrouptime(struct timeval *tvp)
975 getmicrouptime_fromclock(tvp, sysclock_active);
979 getbintime(struct bintime *bt)
982 getbintime_fromclock(bt, sysclock_active);
986 getnanotime(struct timespec *tsp)
989 getnanotime_fromclock(tsp, sysclock_active);
993 getmicrotime(struct timeval *tvp)
996 getmicrouptime_fromclock(tvp, sysclock_active);
1002 * This is a clone of getnanotime and used for walltimestamps.
1003 * The dtrace_ prefix prevents fbt from creating probes for
1004 * it so walltimestamp can be safely used in all fbt probes.
1007 dtrace_getnanotime(struct timespec *tsp)
1010 GETTHMEMBER(tsp, th_nanotime);
1014 * This is a clone of getnanouptime used for time since boot.
1015 * The dtrace_ prefix prevents fbt from creating probes for
1016 * it so an uptime that can be safely used in all fbt probes.
1019 dtrace_getnanouptime(struct timespec *tsp)
1023 GETTHMEMBER(&bt, th_offset);
1024 bintime2timespec(&bt, tsp);
1028 * System clock currently providing time to the system. Modifiable via sysctl
1029 * when the FFCLOCK option is defined.
1031 int sysclock_active = SYSCLOCK_FBCK;
1033 /* Internal NTP status and error estimates. */
1034 extern int time_status;
1035 extern long time_esterror;
1038 * Take a snapshot of sysclock data which can be used to compare system clocks
1039 * and generate timestamps after the fact.
1042 sysclock_getsnapshot(struct sysclock_snap *clock_snap, int fast)
1044 struct fbclock_info *fbi;
1045 struct timehands *th;
1047 unsigned int delta, gen;
1050 struct fftimehands *ffth;
1051 struct ffclock_info *ffi;
1052 struct ffclock_estimate cest;
1054 ffi = &clock_snap->ff_info;
1057 fbi = &clock_snap->fb_info;
1062 gen = atomic_load_acq_int(&th->th_generation);
1063 fbi->th_scale = th->th_scale;
1064 fbi->tick_time = th->th_offset;
1067 ffi->tick_time = ffth->tick_time_lerp;
1068 ffi->tick_time_lerp = ffth->tick_time_lerp;
1069 ffi->period = ffth->cest.period;
1070 ffi->period_lerp = ffth->period_lerp;
1071 clock_snap->ffcount = ffth->tick_ffcount;
1075 delta = tc_delta(th);
1076 atomic_thread_fence_acq();
1077 } while (gen == 0 || gen != th->th_generation);
1079 clock_snap->delta = delta;
1080 clock_snap->sysclock_active = sysclock_active;
1082 /* Record feedback clock status and error. */
1083 clock_snap->fb_info.status = time_status;
1084 /* XXX: Very crude estimate of feedback clock error. */
1085 bt.sec = time_esterror / 1000000;
1086 bt.frac = ((time_esterror - bt.sec) * 1000000) *
1087 (uint64_t)18446744073709ULL;
1088 clock_snap->fb_info.error = bt;
1092 clock_snap->ffcount += delta;
1094 /* Record feed-forward clock leap second adjustment. */
1095 ffi->leapsec_adjustment = cest.leapsec_total;
1096 if (clock_snap->ffcount > cest.leapsec_next)
1097 ffi->leapsec_adjustment -= cest.leapsec;
1099 /* Record feed-forward clock status and error. */
1100 clock_snap->ff_info.status = cest.status;
1101 ffcount = clock_snap->ffcount - cest.update_ffcount;
1102 ffclock_convert_delta(ffcount, cest.period, &bt);
1103 /* 18446744073709 = int(2^64/1e12), err_bound_rate in [ps/s]. */
1104 bintime_mul(&bt, cest.errb_rate * (uint64_t)18446744073709ULL);
1105 /* 18446744073 = int(2^64 / 1e9), since err_abs in [ns]. */
1106 bintime_addx(&bt, cest.errb_abs * (uint64_t)18446744073ULL);
1107 clock_snap->ff_info.error = bt;
1112 * Convert a sysclock snapshot into a struct bintime based on the specified
1113 * clock source and flags.
1116 sysclock_snap2bintime(struct sysclock_snap *cs, struct bintime *bt,
1117 int whichclock, uint32_t flags)
1119 struct bintime boottimebin;
1125 switch (whichclock) {
1127 *bt = cs->fb_info.tick_time;
1129 /* If snapshot was created with !fast, delta will be >0. */
1131 bintime_addx(bt, cs->fb_info.th_scale * cs->delta);
1133 if ((flags & FBCLOCK_UPTIME) == 0) {
1134 getboottimebin(&boottimebin);
1135 bintime_add(bt, &boottimebin);
1140 if (flags & FFCLOCK_LERP) {
1141 *bt = cs->ff_info.tick_time_lerp;
1142 period = cs->ff_info.period_lerp;
1144 *bt = cs->ff_info.tick_time;
1145 period = cs->ff_info.period;
1148 /* If snapshot was created with !fast, delta will be >0. */
1149 if (cs->delta > 0) {
1150 ffclock_convert_delta(cs->delta, period, &bt2);
1151 bintime_add(bt, &bt2);
1154 /* Leap second adjustment. */
1155 if (flags & FFCLOCK_LEAPSEC)
1156 bt->sec -= cs->ff_info.leapsec_adjustment;
1158 /* Boot time adjustment, for uptime/monotonic clocks. */
1159 if (flags & FFCLOCK_UPTIME)
1160 bintime_sub(bt, &ffclock_boottime);
1172 * Initialize a new timecounter and possibly use it.
1175 tc_init(struct timecounter *tc)
1178 struct sysctl_oid *tc_root;
1180 u = tc->tc_frequency / tc->tc_counter_mask;
1181 /* XXX: We need some margin here, 10% is a guess */
1184 if (u > hz && tc->tc_quality >= 0) {
1185 tc->tc_quality = -2000;
1187 printf("Timecounter \"%s\" frequency %ju Hz",
1188 tc->tc_name, (uintmax_t)tc->tc_frequency);
1189 printf(" -- Insufficient hz, needs at least %u\n", u);
1191 } else if (tc->tc_quality >= 0 || bootverbose) {
1192 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
1193 tc->tc_name, (uintmax_t)tc->tc_frequency,
1198 * Set up sysctl tree for this counter.
1200 tc_root = SYSCTL_ADD_NODE_WITH_LABEL(NULL,
1201 SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name,
1202 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1203 "timecounter description", "timecounter");
1204 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1205 "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0,
1206 "mask for implemented bits");
1207 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1208 "counter", CTLTYPE_UINT | CTLFLAG_RD | CTLFLAG_MPSAFE, tc,
1209 sizeof(*tc), sysctl_kern_timecounter_get, "IU",
1210 "current timecounter value");
1211 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1212 "frequency", CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, tc,
1213 sizeof(*tc), sysctl_kern_timecounter_freq, "QU",
1214 "timecounter frequency");
1215 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1216 "quality", CTLFLAG_RD, &(tc->tc_quality), 0,
1217 "goodness of time counter");
1220 tc->tc_next = timecounters;
1224 * Do not automatically switch if the current tc was specifically
1225 * chosen. Never automatically use a timecounter with negative quality.
1226 * Even though we run on the dummy counter, switching here may be
1227 * worse since this timecounter may not be monotonic.
1231 if (tc->tc_quality < 0)
1233 if (tc_from_tunable[0] != '\0' &&
1234 strcmp(tc->tc_name, tc_from_tunable) == 0) {
1236 tc_from_tunable[0] = '\0';
1238 if (tc->tc_quality < timecounter->tc_quality)
1240 if (tc->tc_quality == timecounter->tc_quality &&
1241 tc->tc_frequency < timecounter->tc_frequency)
1244 (void)tc->tc_get_timecount(tc);
1247 mtx_unlock(&tc_lock);
1250 /* Report the frequency of the current timecounter. */
1252 tc_getfrequency(void)
1255 return (timehands->th_counter->tc_frequency);
1259 sleeping_on_old_rtc(struct thread *td)
1263 * td_rtcgen is modified by curthread when it is running,
1264 * and by other threads in this function. By finding the thread
1265 * on a sleepqueue and holding the lock on the sleepqueue
1266 * chain, we guarantee that the thread is not running and that
1267 * modifying td_rtcgen is safe. Setting td_rtcgen to zero informs
1268 * the thread that it was woken due to a real-time clock adjustment.
1269 * (The declaration of td_rtcgen refers to this comment.)
1271 if (td->td_rtcgen != 0 && td->td_rtcgen != rtc_generation) {
1278 static struct mtx tc_setclock_mtx;
1279 MTX_SYSINIT(tc_setclock_init, &tc_setclock_mtx, "tcsetc", MTX_SPIN);
1282 * Step our concept of UTC. This is done by modifying our estimate of
1286 tc_setclock(struct timespec *ts)
1288 struct timespec tbef, taft;
1289 struct bintime bt, bt2;
1291 timespec2bintime(ts, &bt);
1293 mtx_lock_spin(&tc_setclock_mtx);
1294 cpu_tick_calibrate(1);
1296 bintime_sub(&bt, &bt2);
1298 /* XXX fiddle all the little crinkly bits around the fiords... */
1300 mtx_unlock_spin(&tc_setclock_mtx);
1302 /* Avoid rtc_generation == 0, since td_rtcgen == 0 is special. */
1303 atomic_add_rel_int(&rtc_generation, 2);
1304 sleepq_chains_remove_matching(sleeping_on_old_rtc);
1305 if (timestepwarnings) {
1308 "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
1309 (intmax_t)tbef.tv_sec, tbef.tv_nsec,
1310 (intmax_t)taft.tv_sec, taft.tv_nsec,
1311 (intmax_t)ts->tv_sec, ts->tv_nsec);
1316 * Recalculate the scaling factor. We want the number of 1/2^64
1317 * fractions of a second per period of the hardware counter, taking
1318 * into account the th_adjustment factor which the NTP PLL/adjtime(2)
1319 * processing provides us with.
1321 * The th_adjustment is nanoseconds per second with 32 bit binary
1322 * fraction and we want 64 bit binary fraction of second:
1324 * x = a * 2^32 / 10^9 = a * 4.294967296
1326 * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
1327 * we can only multiply by about 850 without overflowing, that
1328 * leaves no suitably precise fractions for multiply before divide.
1330 * Divide before multiply with a fraction of 2199/512 results in a
1331 * systematic undercompensation of 10PPM of th_adjustment. On a
1332 * 5000PPM adjustment this is a 0.05PPM error. This is acceptable.
1334 * We happily sacrifice the lowest of the 64 bits of our result
1335 * to the goddess of code clarity.
1338 recalculate_scaling_factor_and_large_delta(struct timehands *th)
1342 scale = (uint64_t)1 << 63;
1343 scale += (th->th_adjustment / 1024) * 2199;
1344 scale /= th->th_counter->tc_frequency;
1345 th->th_scale = scale * 2;
1346 th->th_large_delta = MIN(((uint64_t)1 << 63) / scale, UINT_MAX);
1350 * Initialize the next struct timehands in the ring and make
1351 * it the active timehands. Along the way we might switch to a different
1352 * timecounter and/or do seconds processing in NTP. Slightly magic.
1355 tc_windup(struct bintime *new_boottimebin)
1358 struct timecounter *tc;
1359 struct timehands *th, *tho;
1360 u_int delta, ncount, ogen;
1365 * Make the next timehands a copy of the current one, but do
1366 * not overwrite the generation or next pointer. While we
1367 * update the contents, the generation must be zero. We need
1368 * to ensure that the zero generation is visible before the
1369 * data updates become visible, which requires release fence.
1370 * For similar reasons, re-reading of the generation after the
1371 * data is read should use acquire fence.
1375 ogen = th->th_generation;
1376 th->th_generation = 0;
1377 atomic_thread_fence_rel();
1378 memcpy(th, tho, offsetof(struct timehands, th_generation));
1379 if (new_boottimebin != NULL)
1380 th->th_boottime = *new_boottimebin;
1383 * Capture a timecounter delta on the current timecounter and if
1384 * changing timecounters, a counter value from the new timecounter.
1385 * Update the offset fields accordingly.
1387 tc = atomic_load_ptr(&timecounter);
1388 delta = tc_delta(th);
1389 if (th->th_counter != tc)
1390 ncount = tc->tc_get_timecount(tc);
1394 ffclock_windup(delta);
1396 th->th_offset_count += delta;
1397 th->th_offset_count &= th->th_counter->tc_counter_mask;
1398 bintime_add_tc_delta(&th->th_offset, th->th_scale,
1399 th->th_large_delta, delta);
1402 * Hardware latching timecounters may not generate interrupts on
1403 * PPS events, so instead we poll them. There is a finite risk that
1404 * the hardware might capture a count which is later than the one we
1405 * got above, and therefore possibly in the next NTP second which might
1406 * have a different rate than the current NTP second. It doesn't
1407 * matter in practice.
1409 if (tho->th_counter->tc_poll_pps)
1410 tho->th_counter->tc_poll_pps(tho->th_counter);
1413 * Deal with NTP second processing. The loop normally
1414 * iterates at most once, but in extreme situations it might
1415 * keep NTP sane if timeouts are not run for several seconds.
1416 * At boot, the time step can be large when the TOD hardware
1417 * has been read, so on really large steps, we call
1418 * ntp_update_second only twice. We need to call it twice in
1419 * case we missed a leap second.
1422 bintime_add(&bt, &th->th_boottime);
1423 i = bt.sec - tho->th_microtime.tv_sec;
1430 ntp_update_second(&th->th_adjustment, &bt.sec);
1432 th->th_boottime.sec += bt.sec - t;
1436 recalculate_scaling_factor_and_large_delta(th);
1439 /* Update the UTC timestamps used by the get*() functions. */
1440 th->th_bintime = bt;
1441 bintime2timeval(&bt, &th->th_microtime);
1442 bintime2timespec(&bt, &th->th_nanotime);
1444 /* Now is a good time to change timecounters. */
1445 if (th->th_counter != tc) {
1447 if ((tc->tc_flags & TC_FLAGS_C2STOP) != 0)
1448 cpu_disable_c2_sleep++;
1449 if ((th->th_counter->tc_flags & TC_FLAGS_C2STOP) != 0)
1450 cpu_disable_c2_sleep--;
1452 th->th_counter = tc;
1453 th->th_offset_count = ncount;
1454 tc_min_ticktock_freq = max(1, tc->tc_frequency /
1455 (((uint64_t)tc->tc_counter_mask + 1) / 3));
1456 recalculate_scaling_factor_and_large_delta(th);
1458 ffclock_change_tc(th);
1463 * Now that the struct timehands is again consistent, set the new
1464 * generation number, making sure to not make it zero.
1468 atomic_store_rel_int(&th->th_generation, ogen);
1470 /* Go live with the new struct timehands. */
1472 switch (sysclock_active) {
1475 time_second = th->th_microtime.tv_sec;
1476 time_uptime = th->th_offset.sec;
1480 time_second = fftimehands->tick_time_lerp.sec;
1481 time_uptime = fftimehands->tick_time_lerp.sec - ffclock_boottime.sec;
1487 timekeep_push_vdso();
1490 /* Report or change the active timecounter hardware. */
1492 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
1495 struct timecounter *newtc, *tc;
1500 strlcpy(newname, tc->tc_name, sizeof(newname));
1501 mtx_unlock(&tc_lock);
1503 error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
1504 if (error != 0 || req->newptr == NULL)
1508 /* Record that the tc in use now was specifically chosen. */
1510 if (strcmp(newname, tc->tc_name) == 0) {
1511 mtx_unlock(&tc_lock);
1514 for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
1515 if (strcmp(newname, newtc->tc_name) != 0)
1518 /* Warm up new timecounter. */
1519 (void)newtc->tc_get_timecount(newtc);
1521 timecounter = newtc;
1524 * The vdso timehands update is deferred until the next
1527 * This is prudent given that 'timekeep_push_vdso()' does not
1528 * use any locking and that it can be called in hard interrupt
1529 * context via 'tc_windup()'.
1533 mtx_unlock(&tc_lock);
1534 return (newtc != NULL ? 0 : EINVAL);
1536 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware,
1537 CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, 0, 0,
1538 sysctl_kern_timecounter_hardware, "A",
1539 "Timecounter hardware selected");
1541 /* Report the available timecounter hardware. */
1543 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
1546 struct timecounter *tc;
1549 error = sysctl_wire_old_buffer(req, 0);
1552 sbuf_new_for_sysctl(&sb, NULL, 0, req);
1554 for (tc = timecounters; tc != NULL; tc = tc->tc_next) {
1555 if (tc != timecounters)
1556 sbuf_putc(&sb, ' ');
1557 sbuf_printf(&sb, "%s(%d)", tc->tc_name, tc->tc_quality);
1559 mtx_unlock(&tc_lock);
1560 error = sbuf_finish(&sb);
1565 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice,
1566 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0,
1567 sysctl_kern_timecounter_choice, "A",
1568 "Timecounter hardware detected");
1571 * RFC 2783 PPS-API implementation.
1575 * Return true if the driver is aware of the abi version extensions in the
1576 * pps_state structure, and it supports at least the given abi version number.
1579 abi_aware(struct pps_state *pps, int vers)
1582 return ((pps->kcmode & KCMODE_ABIFLAG) && pps->driver_abi >= vers);
1586 pps_fetch(struct pps_fetch_args *fapi, struct pps_state *pps)
1589 pps_seq_t aseq, cseq;
1592 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1596 * If no timeout is requested, immediately return whatever values were
1597 * most recently captured. If timeout seconds is -1, that's a request
1598 * to block without a timeout. WITNESS won't let us sleep forever
1599 * without a lock (we really don't need a lock), so just repeatedly
1600 * sleep a long time.
1602 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) {
1603 if (fapi->timeout.tv_sec == -1)
1606 tv.tv_sec = fapi->timeout.tv_sec;
1607 tv.tv_usec = fapi->timeout.tv_nsec / 1000;
1610 aseq = atomic_load_int(&pps->ppsinfo.assert_sequence);
1611 cseq = atomic_load_int(&pps->ppsinfo.clear_sequence);
1612 while (aseq == atomic_load_int(&pps->ppsinfo.assert_sequence) &&
1613 cseq == atomic_load_int(&pps->ppsinfo.clear_sequence)) {
1614 if (abi_aware(pps, 1) && pps->driver_mtx != NULL) {
1615 if (pps->flags & PPSFLAG_MTX_SPIN) {
1616 err = msleep_spin(pps, pps->driver_mtx,
1619 err = msleep(pps, pps->driver_mtx, PCATCH,
1623 err = tsleep(pps, PCATCH, "ppsfch", timo);
1625 if (err == EWOULDBLOCK) {
1626 if (fapi->timeout.tv_sec == -1) {
1631 } else if (err != 0) {
1637 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1638 fapi->pps_info_buf = pps->ppsinfo;
1644 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1647 struct pps_fetch_args *fapi;
1649 struct pps_fetch_ffc_args *fapi_ffc;
1652 struct pps_kcbind_args *kapi;
1655 KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
1657 case PPS_IOC_CREATE:
1659 case PPS_IOC_DESTROY:
1661 case PPS_IOC_SETPARAMS:
1662 app = (pps_params_t *)data;
1663 if (app->mode & ~pps->ppscap)
1666 /* Ensure only a single clock is selected for ffc timestamp. */
1667 if ((app->mode & PPS_TSCLK_MASK) == PPS_TSCLK_MASK)
1670 pps->ppsparam = *app;
1672 case PPS_IOC_GETPARAMS:
1673 app = (pps_params_t *)data;
1674 *app = pps->ppsparam;
1675 app->api_version = PPS_API_VERS_1;
1677 case PPS_IOC_GETCAP:
1678 *(int*)data = pps->ppscap;
1681 fapi = (struct pps_fetch_args *)data;
1682 return (pps_fetch(fapi, pps));
1684 case PPS_IOC_FETCH_FFCOUNTER:
1685 fapi_ffc = (struct pps_fetch_ffc_args *)data;
1686 if (fapi_ffc->tsformat && fapi_ffc->tsformat !=
1689 if (fapi_ffc->timeout.tv_sec || fapi_ffc->timeout.tv_nsec)
1690 return (EOPNOTSUPP);
1691 pps->ppsinfo_ffc.current_mode = pps->ppsparam.mode;
1692 fapi_ffc->pps_info_buf_ffc = pps->ppsinfo_ffc;
1693 /* Overwrite timestamps if feedback clock selected. */
1694 switch (pps->ppsparam.mode & PPS_TSCLK_MASK) {
1695 case PPS_TSCLK_FBCK:
1696 fapi_ffc->pps_info_buf_ffc.assert_timestamp =
1697 pps->ppsinfo.assert_timestamp;
1698 fapi_ffc->pps_info_buf_ffc.clear_timestamp =
1699 pps->ppsinfo.clear_timestamp;
1701 case PPS_TSCLK_FFWD:
1707 #endif /* FFCLOCK */
1708 case PPS_IOC_KCBIND:
1710 kapi = (struct pps_kcbind_args *)data;
1711 /* XXX Only root should be able to do this */
1712 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1714 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1716 if (kapi->edge & ~pps->ppscap)
1718 pps->kcmode = (kapi->edge & KCMODE_EDGEMASK) |
1719 (pps->kcmode & KCMODE_ABIFLAG);
1722 return (EOPNOTSUPP);
1730 pps_init(struct pps_state *pps)
1732 pps->ppscap |= PPS_TSFMT_TSPEC | PPS_CANWAIT;
1733 if (pps->ppscap & PPS_CAPTUREASSERT)
1734 pps->ppscap |= PPS_OFFSETASSERT;
1735 if (pps->ppscap & PPS_CAPTURECLEAR)
1736 pps->ppscap |= PPS_OFFSETCLEAR;
1738 pps->ppscap |= PPS_TSCLK_MASK;
1740 pps->kcmode &= ~KCMODE_ABIFLAG;
1744 pps_init_abi(struct pps_state *pps)
1748 if (pps->driver_abi > 0) {
1749 pps->kcmode |= KCMODE_ABIFLAG;
1750 pps->kernel_abi = PPS_ABI_VERSION;
1755 pps_capture(struct pps_state *pps)
1757 struct timehands *th;
1759 KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
1761 pps->capgen = atomic_load_acq_int(&th->th_generation);
1764 pps->capffth = fftimehands;
1766 pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
1767 atomic_thread_fence_acq();
1768 if (pps->capgen != th->th_generation)
1773 pps_event(struct pps_state *pps, int event)
1776 struct timespec ts, *tsp, *osp;
1777 u_int tcount, *pcount;
1781 struct timespec *tsp_ffc;
1782 pps_seq_t *pseq_ffc;
1789 KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
1790 /* Nothing to do if not currently set to capture this event type. */
1791 if ((event & pps->ppsparam.mode) == 0)
1793 /* If the timecounter was wound up underneath us, bail out. */
1794 if (pps->capgen == 0 || pps->capgen !=
1795 atomic_load_acq_int(&pps->capth->th_generation))
1798 /* Things would be easier with arrays. */
1799 if (event == PPS_CAPTUREASSERT) {
1800 tsp = &pps->ppsinfo.assert_timestamp;
1801 osp = &pps->ppsparam.assert_offset;
1802 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1804 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1806 pcount = &pps->ppscount[0];
1807 pseq = &pps->ppsinfo.assert_sequence;
1809 ffcount = &pps->ppsinfo_ffc.assert_ffcount;
1810 tsp_ffc = &pps->ppsinfo_ffc.assert_timestamp;
1811 pseq_ffc = &pps->ppsinfo_ffc.assert_sequence;
1814 tsp = &pps->ppsinfo.clear_timestamp;
1815 osp = &pps->ppsparam.clear_offset;
1816 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1818 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1820 pcount = &pps->ppscount[1];
1821 pseq = &pps->ppsinfo.clear_sequence;
1823 ffcount = &pps->ppsinfo_ffc.clear_ffcount;
1824 tsp_ffc = &pps->ppsinfo_ffc.clear_timestamp;
1825 pseq_ffc = &pps->ppsinfo_ffc.clear_sequence;
1830 * If the timecounter changed, we cannot compare the count values, so
1831 * we have to drop the rest of the PPS-stuff until the next event.
1833 if (pps->ppstc != pps->capth->th_counter) {
1834 pps->ppstc = pps->capth->th_counter;
1835 *pcount = pps->capcount;
1836 pps->ppscount[2] = pps->capcount;
1840 /* Convert the count to a timespec. */
1841 tcount = pps->capcount - pps->capth->th_offset_count;
1842 tcount &= pps->capth->th_counter->tc_counter_mask;
1843 bt = pps->capth->th_bintime;
1844 bintime_addx(&bt, pps->capth->th_scale * tcount);
1845 bintime2timespec(&bt, &ts);
1847 /* If the timecounter was wound up underneath us, bail out. */
1848 atomic_thread_fence_acq();
1849 if (pps->capgen != pps->capth->th_generation)
1852 *pcount = pps->capcount;
1857 timespecadd(tsp, osp, tsp);
1858 if (tsp->tv_nsec < 0) {
1859 tsp->tv_nsec += 1000000000;
1865 *ffcount = pps->capffth->tick_ffcount + tcount;
1866 bt = pps->capffth->tick_time;
1867 ffclock_convert_delta(tcount, pps->capffth->cest.period, &bt);
1868 bintime_add(&bt, &pps->capffth->tick_time);
1869 bintime2timespec(&bt, &ts);
1879 * Feed the NTP PLL/FLL.
1880 * The FLL wants to know how many (hardware) nanoseconds
1881 * elapsed since the previous event.
1883 tcount = pps->capcount - pps->ppscount[2];
1884 pps->ppscount[2] = pps->capcount;
1885 tcount &= pps->capth->th_counter->tc_counter_mask;
1886 scale = (uint64_t)1 << 63;
1887 scale /= pps->capth->th_counter->tc_frequency;
1891 bintime_addx(&bt, scale * tcount);
1892 bintime2timespec(&bt, &ts);
1893 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
1897 /* Wakeup anyone sleeping in pps_fetch(). */
1902 * Timecounters need to be updated every so often to prevent the hardware
1903 * counter from overflowing. Updating also recalculates the cached values
1904 * used by the get*() family of functions, so their precision depends on
1905 * the update frequency.
1909 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0,
1910 "Approximate number of hardclock ticks in a millisecond");
1913 tc_ticktock(int cnt)
1917 if (mtx_trylock_spin(&tc_setclock_mtx)) {
1919 if (count >= tc_tick) {
1923 mtx_unlock_spin(&tc_setclock_mtx);
1927 static void __inline
1928 tc_adjprecision(void)
1932 if (tc_timepercentage > 0) {
1933 t = (99 + tc_timepercentage) / tc_timepercentage;
1934 tc_precexp = fls(t + (t >> 1)) - 1;
1935 FREQ2BT(hz / tc_tick, &bt_timethreshold);
1936 FREQ2BT(hz, &bt_tickthreshold);
1937 bintime_shift(&bt_timethreshold, tc_precexp);
1938 bintime_shift(&bt_tickthreshold, tc_precexp);
1941 bt_timethreshold.sec = INT_MAX;
1942 bt_timethreshold.frac = ~(uint64_t)0;
1943 bt_tickthreshold = bt_timethreshold;
1945 sbt_timethreshold = bttosbt(bt_timethreshold);
1946 sbt_tickthreshold = bttosbt(bt_tickthreshold);
1950 sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS)
1954 val = tc_timepercentage;
1955 error = sysctl_handle_int(oidp, &val, 0, req);
1956 if (error != 0 || req->newptr == NULL)
1958 tc_timepercentage = val;
1966 /* Set up the requested number of timehands. */
1968 inittimehands(void *dummy)
1970 struct timehands *thp;
1973 TUNABLE_INT_FETCH("kern.timecounter.timehands_count",
1975 if (timehands_count < 1)
1976 timehands_count = 1;
1977 if (timehands_count > nitems(ths))
1978 timehands_count = nitems(ths);
1979 for (i = 1, thp = &ths[0]; i < timehands_count; thp = &ths[i++])
1980 thp->th_next = &ths[i];
1981 thp->th_next = &ths[0];
1983 TUNABLE_STR_FETCH("kern.timecounter.hardware", tc_from_tunable,
1984 sizeof(tc_from_tunable));
1986 mtx_init(&tc_lock, "tc", NULL, MTX_DEF);
1988 SYSINIT(timehands, SI_SUB_TUNABLES, SI_ORDER_ANY, inittimehands, NULL);
1991 inittimecounter(void *dummy)
1997 * Set the initial timeout to
1998 * max(1, <approx. number of hardclock ticks in a millisecond>).
1999 * People should probably not use the sysctl to set the timeout
2000 * to smaller than its initial value, since that value is the
2001 * smallest reasonable one. If they want better timestamps they
2002 * should use the non-"get"* functions.
2005 tc_tick = (hz + 500) / 1000;
2009 FREQ2BT(hz, &tick_bt);
2010 tick_sbt = bttosbt(tick_bt);
2011 tick_rate = hz / tc_tick;
2012 FREQ2BT(tick_rate, &tc_tick_bt);
2013 tc_tick_sbt = bttosbt(tc_tick_bt);
2014 p = (tc_tick * 1000000) / hz;
2015 printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
2021 /* warm up new timecounter (again) and get rolling. */
2022 (void)timecounter->tc_get_timecount(timecounter);
2023 mtx_lock_spin(&tc_setclock_mtx);
2025 mtx_unlock_spin(&tc_setclock_mtx);
2028 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL);
2030 /* Cpu tick handling -------------------------------------------------*/
2032 static bool cpu_tick_variable;
2033 static uint64_t cpu_tick_frequency;
2035 DPCPU_DEFINE_STATIC(uint64_t, tc_cpu_ticks_base);
2036 DPCPU_DEFINE_STATIC(unsigned, tc_cpu_ticks_last);
2041 struct timecounter *tc;
2042 uint64_t res, *base;
2046 base = DPCPU_PTR(tc_cpu_ticks_base);
2047 last = DPCPU_PTR(tc_cpu_ticks_last);
2048 tc = timehands->th_counter;
2049 u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
2051 *base += (uint64_t)tc->tc_counter_mask + 1;
2059 cpu_tick_calibration(void)
2061 static time_t last_calib;
2063 if (time_uptime != last_calib && !(time_uptime & 0xf)) {
2064 cpu_tick_calibrate(0);
2065 last_calib = time_uptime;
2070 * This function gets called every 16 seconds on only one designated
2071 * CPU in the system from hardclock() via cpu_tick_calibration()().
2073 * Whenever the real time clock is stepped we get called with reset=1
2074 * to make sure we handle suspend/resume and similar events correctly.
2078 cpu_tick_calibrate(int reset)
2080 static uint64_t c_last;
2081 uint64_t c_this, c_delta;
2082 static struct bintime t_last;
2083 struct bintime t_this, t_delta;
2087 /* The clock was stepped, abort & reset */
2092 /* we don't calibrate fixed rate cputicks */
2093 if (!cpu_tick_variable)
2096 getbinuptime(&t_this);
2097 c_this = cpu_ticks();
2098 if (t_last.sec != 0) {
2099 c_delta = c_this - c_last;
2101 bintime_sub(&t_delta, &t_last);
2104 * 2^(64-20) / 16[s] =
2106 * 17.592.186.044.416 / 16 =
2107 * 1.099.511.627.776 [Hz]
2109 divi = t_delta.sec << 20;
2110 divi |= t_delta.frac >> (64 - 20);
2113 if (c_delta > cpu_tick_frequency) {
2114 if (0 && bootverbose)
2115 printf("cpu_tick increased to %ju Hz\n",
2117 cpu_tick_frequency = c_delta;
2125 set_cputicker(cpu_tick_f *func, uint64_t freq, bool isvariable)
2129 cpu_ticks = tc_cpu_ticks;
2131 cpu_tick_frequency = freq;
2132 cpu_tick_variable = isvariable;
2141 if (cpu_ticks == tc_cpu_ticks)
2142 return (tc_getfrequency());
2143 return (cpu_tick_frequency);
2147 * We need to be slightly careful converting cputicks to microseconds.
2148 * There is plenty of margin in 64 bits of microseconds (half a million
2149 * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
2150 * before divide conversion (to retain precision) we find that the
2151 * margin shrinks to 1.5 hours (one millionth of 146y).
2152 * With a three prong approach we never lose significant bits, no
2153 * matter what the cputick rate and length of timeinterval is.
2157 cputick2usec(uint64_t tick)
2160 if (tick > 18446744073709551LL) /* floor(2^64 / 1000) */
2161 return (tick / (cpu_tickrate() / 1000000LL));
2162 else if (tick > 18446744073709LL) /* floor(2^64 / 1000000) */
2163 return ((tick * 1000LL) / (cpu_tickrate() / 1000LL));
2165 return ((tick * 1000000LL) / cpu_tickrate());
2168 cpu_tick_f *cpu_ticks = tc_cpu_ticks;
2170 static int vdso_th_enable = 1;
2172 sysctl_fast_gettime(SYSCTL_HANDLER_ARGS)
2174 int old_vdso_th_enable, error;
2176 old_vdso_th_enable = vdso_th_enable;
2177 error = sysctl_handle_int(oidp, &old_vdso_th_enable, 0, req);
2180 vdso_th_enable = old_vdso_th_enable;
2183 SYSCTL_PROC(_kern_timecounter, OID_AUTO, fast_gettime,
2184 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
2185 NULL, 0, sysctl_fast_gettime, "I", "Enable fast time of day");
2188 tc_fill_vdso_timehands(struct vdso_timehands *vdso_th)
2190 struct timehands *th;
2194 vdso_th->th_scale = th->th_scale;
2195 vdso_th->th_offset_count = th->th_offset_count;
2196 vdso_th->th_counter_mask = th->th_counter->tc_counter_mask;
2197 vdso_th->th_offset = th->th_offset;
2198 vdso_th->th_boottime = th->th_boottime;
2199 if (th->th_counter->tc_fill_vdso_timehands != NULL) {
2200 enabled = th->th_counter->tc_fill_vdso_timehands(vdso_th,
2204 if (!vdso_th_enable)
2209 #ifdef COMPAT_FREEBSD32
2211 tc_fill_vdso_timehands32(struct vdso_timehands32 *vdso_th32)
2213 struct timehands *th;
2217 *(uint64_t *)&vdso_th32->th_scale[0] = th->th_scale;
2218 vdso_th32->th_offset_count = th->th_offset_count;
2219 vdso_th32->th_counter_mask = th->th_counter->tc_counter_mask;
2220 vdso_th32->th_offset.sec = th->th_offset.sec;
2221 *(uint64_t *)&vdso_th32->th_offset.frac[0] = th->th_offset.frac;
2222 vdso_th32->th_boottime.sec = th->th_boottime.sec;
2223 *(uint64_t *)&vdso_th32->th_boottime.frac[0] = th->th_boottime.frac;
2224 if (th->th_counter->tc_fill_vdso_timehands32 != NULL) {
2225 enabled = th->th_counter->tc_fill_vdso_timehands32(vdso_th32,
2229 if (!vdso_th_enable)
2235 #include "opt_ddb.h"
2237 #include <ddb/ddb.h>
2239 DB_SHOW_COMMAND(timecounter, db_show_timecounter)
2241 struct timehands *th;
2242 struct timecounter *tc;
2246 tc = th->th_counter;
2247 val1 = tc->tc_get_timecount(tc);
2248 __compiler_membar();
2249 val2 = tc->tc_get_timecount(tc);
2251 db_printf("timecounter %p %s\n", tc, tc->tc_name);
2252 db_printf(" mask %#x freq %ju qual %d flags %#x priv %p\n",
2253 tc->tc_counter_mask, (uintmax_t)tc->tc_frequency, tc->tc_quality,
2254 tc->tc_flags, tc->tc_priv);
2255 db_printf(" val %#x %#x\n", val1, val2);
2256 db_printf("timehands adj %#jx scale %#jx ldelta %d off_cnt %d gen %d\n",
2257 (uintmax_t)th->th_adjustment, (uintmax_t)th->th_scale,
2258 th->th_large_delta, th->th_offset_count, th->th_generation);
2259 db_printf(" offset %jd %jd boottime %jd %jd\n",
2260 (intmax_t)th->th_offset.sec, (uintmax_t)th->th_offset.frac,
2261 (intmax_t)th->th_boottime.sec, (uintmax_t)th->th_boottime.frac);