The Design and Implementation of the FreeBSD Operating System, Second Edition
Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)


[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]

FreeBSD/Linux Kernel Cross Reference
sys/kern/kern_tc.c

Version: -  FREEBSD  -  FREEBSD-13-STABLE  -  FREEBSD-13-0  -  FREEBSD-12-STABLE  -  FREEBSD-12-0  -  FREEBSD-11-STABLE  -  FREEBSD-11-0  -  FREEBSD-10-STABLE  -  FREEBSD-10-0  -  FREEBSD-9-STABLE  -  FREEBSD-9-0  -  FREEBSD-8-STABLE  -  FREEBSD-8-0  -  FREEBSD-7-STABLE  -  FREEBSD-7-0  -  FREEBSD-6-STABLE  -  FREEBSD-6-0  -  FREEBSD-5-STABLE  -  FREEBSD-5-0  -  FREEBSD-4-STABLE  -  FREEBSD-3-STABLE  -  FREEBSD22  -  l41  -  OPENBSD  -  linux-2.6  -  MK84  -  PLAN9  -  xnu-8792 
SearchContext: -  none  -  3  -  10 

    1 /*-
    2  * SPDX-License-Identifier: Beerware
    3  *
    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  * ----------------------------------------------------------------------------
   10  *
   11  * Copyright (c) 2011, 2015, 2016 The FreeBSD Foundation
   12  * All rights reserved.
   13  *
   14  * Portions of this software were developed by Julien Ridoux at the University
   15  * of Melbourne under sponsorship from the FreeBSD Foundation.
   16  *
   17  * Portions of this software were developed by Konstantin Belousov
   18  * under sponsorship from the FreeBSD Foundation.
   19  */
   20 
   21 #include <sys/cdefs.h>
   22 __FBSDID("$FreeBSD$");
   23 
   24 #include "opt_ntp.h"
   25 #include "opt_ffclock.h"
   26 
   27 #include <sys/param.h>
   28 #include <sys/kernel.h>
   29 #include <sys/limits.h>
   30 #include <sys/lock.h>
   31 #include <sys/mutex.h>
   32 #include <sys/proc.h>
   33 #include <sys/sbuf.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>
   42 #include <sys/vdso.h>
   43 
   44 /*
   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.
   49  */
   50 #define LARGE_STEP      200
   51 
   52 /*
   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
   55  * time services.
   56  */
   57 
   58 static u_int
   59 dummy_get_timecount(struct timecounter *tc)
   60 {
   61         static u_int now;
   62 
   63         return (++now);
   64 }
   65 
   66 static struct timecounter dummy_timecounter = {
   67         dummy_get_timecount, 0, ~0u, 1000000, "dummy", -1000000
   68 };
   69 
   70 struct timehands {
   71         /* These fields must be initialized by the driver. */
   72         struct timecounter      *th_counter;
   73         int64_t                 th_adjustment;
   74         uint64_t                th_scale;
   75         u_int                   th_large_delta;
   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. */
   83         u_int                   th_generation;
   84         struct timehands        *th_next;
   85 };
   86 
   87 static struct timehands ths[16] = {
   88     [0] =  {
   89         .th_counter = &dummy_timecounter,
   90         .th_scale = (uint64_t)-1 / 1000000,
   91         .th_large_delta = 1000000,
   92         .th_offset = { .sec = 1 },
   93         .th_generation = 1,
   94     },
   95 };
   96 
   97 static struct timehands *volatile timehands = &ths[0];
   98 struct timecounter *timecounter = &dummy_timecounter;
   99 static struct timecounter *timecounters = &dummy_timecounter;
  100 
  101 /* Mutex to protect the timecounter list. */
  102 static struct mtx tc_lock;
  103 
  104 int tc_min_ticktock_freq = 1;
  105 
  106 volatile time_t time_second = 1;
  107 volatile time_t time_uptime = 1;
  108 
  109 static int sysctl_kern_boottime(SYSCTL_HANDLER_ARGS);
  110 SYSCTL_PROC(_kern, KERN_BOOTTIME, boottime,
  111     CTLTYPE_STRUCT | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
  112     sysctl_kern_boottime, "S,timeval",
  113     "System boottime");
  114 
  115 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
  116     "");
  117 static SYSCTL_NODE(_kern_timecounter, OID_AUTO, tc,
  118     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
  119     "");
  120 
  121 static int timestepwarnings;
  122 SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RWTUN,
  123     &timestepwarnings, 0, "Log time steps");
  124 
  125 static int timehands_count = 2;
  126 SYSCTL_INT(_kern_timecounter, OID_AUTO, timehands_count,
  127     CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
  128     &timehands_count, 0, "Count of timehands in rotation");
  129 
  130 struct bintime bt_timethreshold;
  131 struct bintime bt_tickthreshold;
  132 sbintime_t sbt_timethreshold;
  133 sbintime_t sbt_tickthreshold;
  134 struct bintime tc_tick_bt;
  135 sbintime_t tc_tick_sbt;
  136 int tc_precexp;
  137 int tc_timepercentage = TC_DEFAULTPERC;
  138 static int sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS);
  139 SYSCTL_PROC(_kern_timecounter, OID_AUTO, alloweddeviation,
  140     CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, 0, 0,
  141     sysctl_kern_timecounter_adjprecision, "I",
  142     "Allowed time interval deviation in percents");
  143 
  144 volatile int rtc_generation = 1;
  145 
  146 static int tc_chosen;   /* Non-zero if a specific tc was chosen via sysctl. */
  147 static char tc_from_tunable[16];
  148 
  149 static void tc_windup(struct bintime *new_boottimebin);
  150 static void cpu_tick_calibrate(int);
  151 
  152 void dtrace_getnanotime(struct timespec *tsp);
  153 void dtrace_getnanouptime(struct timespec *tsp);
  154 
  155 static int
  156 sysctl_kern_boottime(SYSCTL_HANDLER_ARGS)
  157 {
  158         struct timeval boottime;
  159 
  160         getboottime(&boottime);
  161 
  162 /* i386 is the only arch which uses a 32bits time_t */
  163 #ifdef __amd64__
  164 #ifdef SCTL_MASK32
  165         int tv[2];
  166 
  167         if (req->flags & SCTL_MASK32) {
  168                 tv[0] = boottime.tv_sec;
  169                 tv[1] = boottime.tv_usec;
  170                 return (SYSCTL_OUT(req, tv, sizeof(tv)));
  171         }
  172 #endif
  173 #endif
  174         return (SYSCTL_OUT(req, &boottime, sizeof(boottime)));
  175 }
  176 
  177 static int
  178 sysctl_kern_timecounter_get(SYSCTL_HANDLER_ARGS)
  179 {
  180         u_int ncount;
  181         struct timecounter *tc = arg1;
  182 
  183         ncount = tc->tc_get_timecount(tc);
  184         return (sysctl_handle_int(oidp, &ncount, 0, req));
  185 }
  186 
  187 static int
  188 sysctl_kern_timecounter_freq(SYSCTL_HANDLER_ARGS)
  189 {
  190         uint64_t freq;
  191         struct timecounter *tc = arg1;
  192 
  193         freq = tc->tc_frequency;
  194         return (sysctl_handle_64(oidp, &freq, 0, req));
  195 }
  196 
  197 /*
  198  * Return the difference between the timehands' counter value now and what
  199  * was when we copied it to the timehands' offset_count.
  200  */
  201 static __inline u_int
  202 tc_delta(struct timehands *th)
  203 {
  204         struct timecounter *tc;
  205 
  206         tc = th->th_counter;
  207         return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
  208             tc->tc_counter_mask);
  209 }
  210 
  211 /*
  212  * Functions for reading the time.  We have to loop until we are sure that
  213  * the timehands that we operated on was not updated under our feet.  See
  214  * the comment in <sys/time.h> for a description of these 12 functions.
  215  */
  216 
  217 static __inline void
  218 bintime_off(struct bintime *bt, u_int off)
  219 {
  220         struct timehands *th;
  221         struct bintime *btp;
  222         uint64_t scale, x;
  223         u_int delta, gen, large_delta;
  224 
  225         do {
  226                 th = timehands;
  227                 gen = atomic_load_acq_int(&th->th_generation);
  228                 btp = (struct bintime *)((vm_offset_t)th + off);
  229                 *bt = *btp;
  230                 scale = th->th_scale;
  231                 delta = tc_delta(th);
  232                 large_delta = th->th_large_delta;
  233                 atomic_thread_fence_acq();
  234         } while (gen == 0 || gen != th->th_generation);
  235 
  236         if (__predict_false(delta >= large_delta)) {
  237                 /* Avoid overflow for scale * delta. */
  238                 x = (scale >> 32) * delta;
  239                 bt->sec += x >> 32;
  240                 bintime_addx(bt, x << 32);
  241                 bintime_addx(bt, (scale & 0xffffffff) * delta);
  242         } else {
  243                 bintime_addx(bt, scale * delta);
  244         }
  245 }
  246 #define GETTHBINTIME(dst, member)                                       \
  247 do {                                                                    \
  248         _Static_assert(_Generic(((struct timehands *)NULL)->member,     \
  249             struct bintime: 1, default: 0) == 1,                        \
  250             "struct timehands member is not of struct bintime type");   \
  251         bintime_off(dst, __offsetof(struct timehands, member));         \
  252 } while (0)
  253 
  254 static __inline void
  255 getthmember(void *out, size_t out_size, u_int off)
  256 {
  257         struct timehands *th;
  258         u_int gen;
  259 
  260         do {
  261                 th = timehands;
  262                 gen = atomic_load_acq_int(&th->th_generation);
  263                 memcpy(out, (char *)th + off, out_size);
  264                 atomic_thread_fence_acq();
  265         } while (gen == 0 || gen != th->th_generation);
  266 }
  267 #define GETTHMEMBER(dst, member)                                        \
  268 do {                                                                    \
  269         _Static_assert(_Generic(*dst,                                   \
  270             __typeof(((struct timehands *)NULL)->member): 1,            \
  271             default: 0) == 1,                                           \
  272             "*dst and struct timehands member have different types");   \
  273         getthmember(dst, sizeof(*dst), __offsetof(struct timehands,     \
  274             member));                                                   \
  275 } while (0)
  276 
  277 #ifdef FFCLOCK
  278 void
  279 fbclock_binuptime(struct bintime *bt)
  280 {
  281 
  282         GETTHBINTIME(bt, th_offset);
  283 }
  284 
  285 void
  286 fbclock_nanouptime(struct timespec *tsp)
  287 {
  288         struct bintime bt;
  289 
  290         fbclock_binuptime(&bt);
  291         bintime2timespec(&bt, tsp);
  292 }
  293 
  294 void
  295 fbclock_microuptime(struct timeval *tvp)
  296 {
  297         struct bintime bt;
  298 
  299         fbclock_binuptime(&bt);
  300         bintime2timeval(&bt, tvp);
  301 }
  302 
  303 void
  304 fbclock_bintime(struct bintime *bt)
  305 {
  306 
  307         GETTHBINTIME(bt, th_bintime);
  308 }
  309 
  310 void
  311 fbclock_nanotime(struct timespec *tsp)
  312 {
  313         struct bintime bt;
  314 
  315         fbclock_bintime(&bt);
  316         bintime2timespec(&bt, tsp);
  317 }
  318 
  319 void
  320 fbclock_microtime(struct timeval *tvp)
  321 {
  322         struct bintime bt;
  323 
  324         fbclock_bintime(&bt);
  325         bintime2timeval(&bt, tvp);
  326 }
  327 
  328 void
  329 fbclock_getbinuptime(struct bintime *bt)
  330 {
  331 
  332         GETTHMEMBER(bt, th_offset);
  333 }
  334 
  335 void
  336 fbclock_getnanouptime(struct timespec *tsp)
  337 {
  338         struct bintime bt;
  339 
  340         GETTHMEMBER(&bt, th_offset);
  341         bintime2timespec(&bt, tsp);
  342 }
  343 
  344 void
  345 fbclock_getmicrouptime(struct timeval *tvp)
  346 {
  347         struct bintime bt;
  348 
  349         GETTHMEMBER(&bt, th_offset);
  350         bintime2timeval(&bt, tvp);
  351 }
  352 
  353 void
  354 fbclock_getbintime(struct bintime *bt)
  355 {
  356 
  357         GETTHMEMBER(bt, th_bintime);
  358 }
  359 
  360 void
  361 fbclock_getnanotime(struct timespec *tsp)
  362 {
  363 
  364         GETTHMEMBER(tsp, th_nanotime);
  365 }
  366 
  367 void
  368 fbclock_getmicrotime(struct timeval *tvp)
  369 {
  370 
  371         GETTHMEMBER(tvp, th_microtime);
  372 }
  373 #else /* !FFCLOCK */
  374 
  375 void
  376 binuptime(struct bintime *bt)
  377 {
  378 
  379         GETTHBINTIME(bt, th_offset);
  380 }
  381 
  382 void
  383 nanouptime(struct timespec *tsp)
  384 {
  385         struct bintime bt;
  386 
  387         binuptime(&bt);
  388         bintime2timespec(&bt, tsp);
  389 }
  390 
  391 void
  392 microuptime(struct timeval *tvp)
  393 {
  394         struct bintime bt;
  395 
  396         binuptime(&bt);
  397         bintime2timeval(&bt, tvp);
  398 }
  399 
  400 void
  401 bintime(struct bintime *bt)
  402 {
  403 
  404         GETTHBINTIME(bt, th_bintime);
  405 }
  406 
  407 void
  408 nanotime(struct timespec *tsp)
  409 {
  410         struct bintime bt;
  411 
  412         bintime(&bt);
  413         bintime2timespec(&bt, tsp);
  414 }
  415 
  416 void
  417 microtime(struct timeval *tvp)
  418 {
  419         struct bintime bt;
  420 
  421         bintime(&bt);
  422         bintime2timeval(&bt, tvp);
  423 }
  424 
  425 void
  426 getbinuptime(struct bintime *bt)
  427 {
  428 
  429         GETTHMEMBER(bt, th_offset);
  430 }
  431 
  432 void
  433 getnanouptime(struct timespec *tsp)
  434 {
  435         struct bintime bt;
  436 
  437         GETTHMEMBER(&bt, th_offset);
  438         bintime2timespec(&bt, tsp);
  439 }
  440 
  441 void
  442 getmicrouptime(struct timeval *tvp)
  443 {
  444         struct bintime bt;
  445 
  446         GETTHMEMBER(&bt, th_offset);
  447         bintime2timeval(&bt, tvp);
  448 }
  449 
  450 void
  451 getbintime(struct bintime *bt)
  452 {
  453 
  454         GETTHMEMBER(bt, th_bintime);
  455 }
  456 
  457 void
  458 getnanotime(struct timespec *tsp)
  459 {
  460 
  461         GETTHMEMBER(tsp, th_nanotime);
  462 }
  463 
  464 void
  465 getmicrotime(struct timeval *tvp)
  466 {
  467 
  468         GETTHMEMBER(tvp, th_microtime);
  469 }
  470 #endif /* FFCLOCK */
  471 
  472 void
  473 getboottime(struct timeval *boottime)
  474 {
  475         struct bintime boottimebin;
  476 
  477         getboottimebin(&boottimebin);
  478         bintime2timeval(&boottimebin, boottime);
  479 }
  480 
  481 void
  482 getboottimebin(struct bintime *boottimebin)
  483 {
  484 
  485         GETTHMEMBER(boottimebin, th_boottime);
  486 }
  487 
  488 #ifdef FFCLOCK
  489 /*
  490  * Support for feed-forward synchronization algorithms. This is heavily inspired
  491  * by the timehands mechanism but kept independent from it. *_windup() functions
  492  * have some connection to avoid accessing the timecounter hardware more than
  493  * necessary.
  494  */
  495 
  496 /* Feed-forward clock estimates kept updated by the synchronization daemon. */
  497 struct ffclock_estimate ffclock_estimate;
  498 struct bintime ffclock_boottime;        /* Feed-forward boot time estimate. */
  499 uint32_t ffclock_status;                /* Feed-forward clock status. */
  500 int8_t ffclock_updated;                 /* New estimates are available. */
  501 struct mtx ffclock_mtx;                 /* Mutex on ffclock_estimate. */
  502 
  503 struct fftimehands {
  504         struct ffclock_estimate cest;
  505         struct bintime          tick_time;
  506         struct bintime          tick_time_lerp;
  507         ffcounter               tick_ffcount;
  508         uint64_t                period_lerp;
  509         volatile uint8_t        gen;
  510         struct fftimehands      *next;
  511 };
  512 
  513 #define NUM_ELEMENTS(x) (sizeof(x) / sizeof(*x))
  514 
  515 static struct fftimehands ffth[10];
  516 static struct fftimehands *volatile fftimehands = ffth;
  517 
  518 static void
  519 ffclock_init(void)
  520 {
  521         struct fftimehands *cur;
  522         struct fftimehands *last;
  523 
  524         memset(ffth, 0, sizeof(ffth));
  525 
  526         last = ffth + NUM_ELEMENTS(ffth) - 1;
  527         for (cur = ffth; cur < last; cur++)
  528                 cur->next = cur + 1;
  529         last->next = ffth;
  530 
  531         ffclock_updated = 0;
  532         ffclock_status = FFCLOCK_STA_UNSYNC;
  533         mtx_init(&ffclock_mtx, "ffclock lock", NULL, MTX_DEF);
  534 }
  535 
  536 /*
  537  * Reset the feed-forward clock estimates. Called from inittodr() to get things
  538  * kick started and uses the timecounter nominal frequency as a first period
  539  * estimate. Note: this function may be called several time just after boot.
  540  * Note: this is the only function that sets the value of boot time for the
  541  * monotonic (i.e. uptime) version of the feed-forward clock.
  542  */
  543 void
  544 ffclock_reset_clock(struct timespec *ts)
  545 {
  546         struct timecounter *tc;
  547         struct ffclock_estimate cest;
  548 
  549         tc = timehands->th_counter;
  550         memset(&cest, 0, sizeof(struct ffclock_estimate));
  551 
  552         timespec2bintime(ts, &ffclock_boottime);
  553         timespec2bintime(ts, &(cest.update_time));
  554         ffclock_read_counter(&cest.update_ffcount);
  555         cest.leapsec_next = 0;
  556         cest.period = ((1ULL << 63) / tc->tc_frequency) << 1;
  557         cest.errb_abs = 0;
  558         cest.errb_rate = 0;
  559         cest.status = FFCLOCK_STA_UNSYNC;
  560         cest.leapsec_total = 0;
  561         cest.leapsec = 0;
  562 
  563         mtx_lock(&ffclock_mtx);
  564         bcopy(&cest, &ffclock_estimate, sizeof(struct ffclock_estimate));
  565         ffclock_updated = INT8_MAX;
  566         mtx_unlock(&ffclock_mtx);
  567 
  568         printf("ffclock reset: %s (%llu Hz), time = %ld.%09lu\n", tc->tc_name,
  569             (unsigned long long)tc->tc_frequency, (long)ts->tv_sec,
  570             (unsigned long)ts->tv_nsec);
  571 }
  572 
  573 /*
  574  * Sub-routine to convert a time interval measured in RAW counter units to time
  575  * in seconds stored in bintime format.
  576  * NOTE: bintime_mul requires u_int, but the value of the ffcounter may be
  577  * larger than the max value of u_int (on 32 bit architecture). Loop to consume
  578  * extra cycles.
  579  */
  580 static void
  581 ffclock_convert_delta(ffcounter ffdelta, uint64_t period, struct bintime *bt)
  582 {
  583         struct bintime bt2;
  584         ffcounter delta, delta_max;
  585 
  586         delta_max = (1ULL << (8 * sizeof(unsigned int))) - 1;
  587         bintime_clear(bt);
  588         do {
  589                 if (ffdelta > delta_max)
  590                         delta = delta_max;
  591                 else
  592                         delta = ffdelta;
  593                 bt2.sec = 0;
  594                 bt2.frac = period;
  595                 bintime_mul(&bt2, (unsigned int)delta);
  596                 bintime_add(bt, &bt2);
  597                 ffdelta -= delta;
  598         } while (ffdelta > 0);
  599 }
  600 
  601 /*
  602  * Update the fftimehands.
  603  * Push the tick ffcount and time(s) forward based on current clock estimate.
  604  * The conversion from ffcounter to bintime relies on the difference clock
  605  * principle, whose accuracy relies on computing small time intervals. If a new
  606  * clock estimate has been passed by the synchronisation daemon, make it
  607  * current, and compute the linear interpolation for monotonic time if needed.
  608  */
  609 static void
  610 ffclock_windup(unsigned int delta)
  611 {
  612         struct ffclock_estimate *cest;
  613         struct fftimehands *ffth;
  614         struct bintime bt, gap_lerp;
  615         ffcounter ffdelta;
  616         uint64_t frac;
  617         unsigned int polling;
  618         uint8_t forward_jump, ogen;
  619 
  620         /*
  621          * Pick the next timehand, copy current ffclock estimates and move tick
  622          * times and counter forward.
  623          */
  624         forward_jump = 0;
  625         ffth = fftimehands->next;
  626         ogen = ffth->gen;
  627         ffth->gen = 0;
  628         cest = &ffth->cest;
  629         bcopy(&fftimehands->cest, cest, sizeof(struct ffclock_estimate));
  630         ffdelta = (ffcounter)delta;
  631         ffth->period_lerp = fftimehands->period_lerp;
  632 
  633         ffth->tick_time = fftimehands->tick_time;
  634         ffclock_convert_delta(ffdelta, cest->period, &bt);
  635         bintime_add(&ffth->tick_time, &bt);
  636 
  637         ffth->tick_time_lerp = fftimehands->tick_time_lerp;
  638         ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt);
  639         bintime_add(&ffth->tick_time_lerp, &bt);
  640 
  641         ffth->tick_ffcount = fftimehands->tick_ffcount + ffdelta;
  642 
  643         /*
  644          * Assess the status of the clock, if the last update is too old, it is
  645          * likely the synchronisation daemon is dead and the clock is free
  646          * running.
  647          */
  648         if (ffclock_updated == 0) {
  649                 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
  650                 ffclock_convert_delta(ffdelta, cest->period, &bt);
  651                 if (bt.sec > 2 * FFCLOCK_SKM_SCALE)
  652                         ffclock_status |= FFCLOCK_STA_UNSYNC;
  653         }
  654 
  655         /*
  656          * If available, grab updated clock estimates and make them current.
  657          * Recompute time at this tick using the updated estimates. The clock
  658          * estimates passed the feed-forward synchronisation daemon may result
  659          * in time conversion that is not monotonically increasing (just after
  660          * the update). time_lerp is a particular linear interpolation over the
  661          * synchronisation algo polling period that ensures monotonicity for the
  662          * clock ids requesting it.
  663          */
  664         if (ffclock_updated > 0) {
  665                 bcopy(&ffclock_estimate, cest, sizeof(struct ffclock_estimate));
  666                 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
  667                 ffth->tick_time = cest->update_time;
  668                 ffclock_convert_delta(ffdelta, cest->period, &bt);
  669                 bintime_add(&ffth->tick_time, &bt);
  670 
  671                 /* ffclock_reset sets ffclock_updated to INT8_MAX */
  672                 if (ffclock_updated == INT8_MAX)
  673                         ffth->tick_time_lerp = ffth->tick_time;
  674 
  675                 if (bintime_cmp(&ffth->tick_time, &ffth->tick_time_lerp, >))
  676                         forward_jump = 1;
  677                 else
  678                         forward_jump = 0;
  679 
  680                 bintime_clear(&gap_lerp);
  681                 if (forward_jump) {
  682                         gap_lerp = ffth->tick_time;
  683                         bintime_sub(&gap_lerp, &ffth->tick_time_lerp);
  684                 } else {
  685                         gap_lerp = ffth->tick_time_lerp;
  686                         bintime_sub(&gap_lerp, &ffth->tick_time);
  687                 }
  688 
  689                 /*
  690                  * The reset from the RTC clock may be far from accurate, and
  691                  * reducing the gap between real time and interpolated time
  692                  * could take a very long time if the interpolated clock insists
  693                  * on strict monotonicity. The clock is reset under very strict
  694                  * conditions (kernel time is known to be wrong and
  695                  * synchronization daemon has been restarted recently.
  696                  * ffclock_boottime absorbs the jump to ensure boot time is
  697                  * correct and uptime functions stay consistent.
  698                  */
  699                 if (((ffclock_status & FFCLOCK_STA_UNSYNC) == FFCLOCK_STA_UNSYNC) &&
  700                     ((cest->status & FFCLOCK_STA_UNSYNC) == 0) &&
  701                     ((cest->status & FFCLOCK_STA_WARMUP) == FFCLOCK_STA_WARMUP)) {
  702                         if (forward_jump)
  703                                 bintime_add(&ffclock_boottime, &gap_lerp);
  704                         else
  705                                 bintime_sub(&ffclock_boottime, &gap_lerp);
  706                         ffth->tick_time_lerp = ffth->tick_time;
  707                         bintime_clear(&gap_lerp);
  708                 }
  709 
  710                 ffclock_status = cest->status;
  711                 ffth->period_lerp = cest->period;
  712 
  713                 /*
  714                  * Compute corrected period used for the linear interpolation of
  715                  * time. The rate of linear interpolation is capped to 5000PPM
  716                  * (5ms/s).
  717                  */
  718                 if (bintime_isset(&gap_lerp)) {
  719                         ffdelta = cest->update_ffcount;
  720                         ffdelta -= fftimehands->cest.update_ffcount;
  721                         ffclock_convert_delta(ffdelta, cest->period, &bt);
  722                         polling = bt.sec;
  723                         bt.sec = 0;
  724                         bt.frac = 5000000 * (uint64_t)18446744073LL;
  725                         bintime_mul(&bt, polling);
  726                         if (bintime_cmp(&gap_lerp, &bt, >))
  727                                 gap_lerp = bt;
  728 
  729                         /* Approximate 1 sec by 1-(1/2^64) to ease arithmetic */
  730                         frac = 0;
  731                         if (gap_lerp.sec > 0) {
  732                                 frac -= 1;
  733                                 frac /= ffdelta / gap_lerp.sec;
  734                         }
  735                         frac += gap_lerp.frac / ffdelta;
  736 
  737                         if (forward_jump)
  738                                 ffth->period_lerp += frac;
  739                         else
  740                                 ffth->period_lerp -= frac;
  741                 }
  742 
  743                 ffclock_updated = 0;
  744         }
  745         if (++ogen == 0)
  746                 ogen = 1;
  747         ffth->gen = ogen;
  748         fftimehands = ffth;
  749 }
  750 
  751 /*
  752  * Adjust the fftimehands when the timecounter is changed. Stating the obvious,
  753  * the old and new hardware counter cannot be read simultaneously. tc_windup()
  754  * does read the two counters 'back to back', but a few cycles are effectively
  755  * lost, and not accumulated in tick_ffcount. This is a fairly radical
  756  * operation for a feed-forward synchronization daemon, and it is its job to not
  757  * pushing irrelevant data to the kernel. Because there is no locking here,
  758  * simply force to ignore pending or next update to give daemon a chance to
  759  * realize the counter has changed.
  760  */
  761 static void
  762 ffclock_change_tc(struct timehands *th)
  763 {
  764         struct fftimehands *ffth;
  765         struct ffclock_estimate *cest;
  766         struct timecounter *tc;
  767         uint8_t ogen;
  768 
  769         tc = th->th_counter;
  770         ffth = fftimehands->next;
  771         ogen = ffth->gen;
  772         ffth->gen = 0;
  773 
  774         cest = &ffth->cest;
  775         bcopy(&(fftimehands->cest), cest, sizeof(struct ffclock_estimate));
  776         cest->period = ((1ULL << 63) / tc->tc_frequency ) << 1;
  777         cest->errb_abs = 0;
  778         cest->errb_rate = 0;
  779         cest->status |= FFCLOCK_STA_UNSYNC;
  780 
  781         ffth->tick_ffcount = fftimehands->tick_ffcount;
  782         ffth->tick_time_lerp = fftimehands->tick_time_lerp;
  783         ffth->tick_time = fftimehands->tick_time;
  784         ffth->period_lerp = cest->period;
  785 
  786         /* Do not lock but ignore next update from synchronization daemon. */
  787         ffclock_updated--;
  788 
  789         if (++ogen == 0)
  790                 ogen = 1;
  791         ffth->gen = ogen;
  792         fftimehands = ffth;
  793 }
  794 
  795 /*
  796  * Retrieve feed-forward counter and time of last kernel tick.
  797  */
  798 void
  799 ffclock_last_tick(ffcounter *ffcount, struct bintime *bt, uint32_t flags)
  800 {
  801         struct fftimehands *ffth;
  802         uint8_t gen;
  803 
  804         /*
  805          * No locking but check generation has not changed. Also need to make
  806          * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
  807          */
  808         do {
  809                 ffth = fftimehands;
  810                 gen = ffth->gen;
  811                 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP)
  812                         *bt = ffth->tick_time_lerp;
  813                 else
  814                         *bt = ffth->tick_time;
  815                 *ffcount = ffth->tick_ffcount;
  816         } while (gen == 0 || gen != ffth->gen);
  817 }
  818 
  819 /*
  820  * Absolute clock conversion. Low level function to convert ffcounter to
  821  * bintime. The ffcounter is converted using the current ffclock period estimate
  822  * or the "interpolated period" to ensure monotonicity.
  823  * NOTE: this conversion may have been deferred, and the clock updated since the
  824  * hardware counter has been read.
  825  */
  826 void
  827 ffclock_convert_abs(ffcounter ffcount, struct bintime *bt, uint32_t flags)
  828 {
  829         struct fftimehands *ffth;
  830         struct bintime bt2;
  831         ffcounter ffdelta;
  832         uint8_t gen;
  833 
  834         /*
  835          * No locking but check generation has not changed. Also need to make
  836          * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
  837          */
  838         do {
  839                 ffth = fftimehands;
  840                 gen = ffth->gen;
  841                 if (ffcount > ffth->tick_ffcount)
  842                         ffdelta = ffcount - ffth->tick_ffcount;
  843                 else
  844                         ffdelta = ffth->tick_ffcount - ffcount;
  845 
  846                 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP) {
  847                         *bt = ffth->tick_time_lerp;
  848                         ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt2);
  849                 } else {
  850                         *bt = ffth->tick_time;
  851                         ffclock_convert_delta(ffdelta, ffth->cest.period, &bt2);
  852                 }
  853 
  854                 if (ffcount > ffth->tick_ffcount)
  855                         bintime_add(bt, &bt2);
  856                 else
  857                         bintime_sub(bt, &bt2);
  858         } while (gen == 0 || gen != ffth->gen);
  859 }
  860 
  861 /*
  862  * Difference clock conversion.
  863  * Low level function to Convert a time interval measured in RAW counter units
  864  * into bintime. The difference clock allows measuring small intervals much more
  865  * reliably than the absolute clock.
  866  */
  867 void
  868 ffclock_convert_diff(ffcounter ffdelta, struct bintime *bt)
  869 {
  870         struct fftimehands *ffth;
  871         uint8_t gen;
  872 
  873         /* No locking but check generation has not changed. */
  874         do {
  875                 ffth = fftimehands;
  876                 gen = ffth->gen;
  877                 ffclock_convert_delta(ffdelta, ffth->cest.period, bt);
  878         } while (gen == 0 || gen != ffth->gen);
  879 }
  880 
  881 /*
  882  * Access to current ffcounter value.
  883  */
  884 void
  885 ffclock_read_counter(ffcounter *ffcount)
  886 {
  887         struct timehands *th;
  888         struct fftimehands *ffth;
  889         unsigned int gen, delta;
  890 
  891         /*
  892          * ffclock_windup() called from tc_windup(), safe to rely on
  893          * th->th_generation only, for correct delta and ffcounter.
  894          */
  895         do {
  896                 th = timehands;
  897                 gen = atomic_load_acq_int(&th->th_generation);
  898                 ffth = fftimehands;
  899                 delta = tc_delta(th);
  900                 *ffcount = ffth->tick_ffcount;
  901                 atomic_thread_fence_acq();
  902         } while (gen == 0 || gen != th->th_generation);
  903 
  904         *ffcount += delta;
  905 }
  906 
  907 void
  908 binuptime(struct bintime *bt)
  909 {
  910 
  911         binuptime_fromclock(bt, sysclock_active);
  912 }
  913 
  914 void
  915 nanouptime(struct timespec *tsp)
  916 {
  917 
  918         nanouptime_fromclock(tsp, sysclock_active);
  919 }
  920 
  921 void
  922 microuptime(struct timeval *tvp)
  923 {
  924 
  925         microuptime_fromclock(tvp, sysclock_active);
  926 }
  927 
  928 void
  929 bintime(struct bintime *bt)
  930 {
  931 
  932         bintime_fromclock(bt, sysclock_active);
  933 }
  934 
  935 void
  936 nanotime(struct timespec *tsp)
  937 {
  938 
  939         nanotime_fromclock(tsp, sysclock_active);
  940 }
  941 
  942 void
  943 microtime(struct timeval *tvp)
  944 {
  945 
  946         microtime_fromclock(tvp, sysclock_active);
  947 }
  948 
  949 void
  950 getbinuptime(struct bintime *bt)
  951 {
  952 
  953         getbinuptime_fromclock(bt, sysclock_active);
  954 }
  955 
  956 void
  957 getnanouptime(struct timespec *tsp)
  958 {
  959 
  960         getnanouptime_fromclock(tsp, sysclock_active);
  961 }
  962 
  963 void
  964 getmicrouptime(struct timeval *tvp)
  965 {
  966 
  967         getmicrouptime_fromclock(tvp, sysclock_active);
  968 }
  969 
  970 void
  971 getbintime(struct bintime *bt)
  972 {
  973 
  974         getbintime_fromclock(bt, sysclock_active);
  975 }
  976 
  977 void
  978 getnanotime(struct timespec *tsp)
  979 {
  980 
  981         getnanotime_fromclock(tsp, sysclock_active);
  982 }
  983 
  984 void
  985 getmicrotime(struct timeval *tvp)
  986 {
  987 
  988         getmicrouptime_fromclock(tvp, sysclock_active);
  989 }
  990 
  991 #endif /* FFCLOCK */
  992 
  993 /*
  994  * This is a clone of getnanotime and used for walltimestamps.
  995  * The dtrace_ prefix prevents fbt from creating probes for
  996  * it so walltimestamp can be safely used in all fbt probes.
  997  */
  998 void
  999 dtrace_getnanotime(struct timespec *tsp)
 1000 {
 1001 
 1002         GETTHMEMBER(tsp, th_nanotime);
 1003 }
 1004 
 1005 /*
 1006  * This is a clone of getnanouptime used for time since boot.
 1007  * The dtrace_ prefix prevents fbt from creating probes for
 1008  * it so an uptime that can be safely used in all fbt probes.
 1009  */
 1010 void
 1011 dtrace_getnanouptime(struct timespec *tsp)
 1012 {
 1013         struct bintime bt;
 1014 
 1015         GETTHMEMBER(&bt, th_offset);
 1016         bintime2timespec(&bt, tsp);
 1017 }
 1018 
 1019 /*
 1020  * System clock currently providing time to the system. Modifiable via sysctl
 1021  * when the FFCLOCK option is defined.
 1022  */
 1023 int sysclock_active = SYSCLOCK_FBCK;
 1024 
 1025 /* Internal NTP status and error estimates. */
 1026 extern int time_status;
 1027 extern long time_esterror;
 1028 
 1029 /*
 1030  * Take a snapshot of sysclock data which can be used to compare system clocks
 1031  * and generate timestamps after the fact.
 1032  */
 1033 void
 1034 sysclock_getsnapshot(struct sysclock_snap *clock_snap, int fast)
 1035 {
 1036         struct fbclock_info *fbi;
 1037         struct timehands *th;
 1038         struct bintime bt;
 1039         unsigned int delta, gen;
 1040 #ifdef FFCLOCK
 1041         ffcounter ffcount;
 1042         struct fftimehands *ffth;
 1043         struct ffclock_info *ffi;
 1044         struct ffclock_estimate cest;
 1045 
 1046         ffi = &clock_snap->ff_info;
 1047 #endif
 1048 
 1049         fbi = &clock_snap->fb_info;
 1050         delta = 0;
 1051 
 1052         do {
 1053                 th = timehands;
 1054                 gen = atomic_load_acq_int(&th->th_generation);
 1055                 fbi->th_scale = th->th_scale;
 1056                 fbi->tick_time = th->th_offset;
 1057 #ifdef FFCLOCK
 1058                 ffth = fftimehands;
 1059                 ffi->tick_time = ffth->tick_time_lerp;
 1060                 ffi->tick_time_lerp = ffth->tick_time_lerp;
 1061                 ffi->period = ffth->cest.period;
 1062                 ffi->period_lerp = ffth->period_lerp;
 1063                 clock_snap->ffcount = ffth->tick_ffcount;
 1064                 cest = ffth->cest;
 1065 #endif
 1066                 if (!fast)
 1067                         delta = tc_delta(th);
 1068                 atomic_thread_fence_acq();
 1069         } while (gen == 0 || gen != th->th_generation);
 1070 
 1071         clock_snap->delta = delta;
 1072         clock_snap->sysclock_active = sysclock_active;
 1073 
 1074         /* Record feedback clock status and error. */
 1075         clock_snap->fb_info.status = time_status;
 1076         /* XXX: Very crude estimate of feedback clock error. */
 1077         bt.sec = time_esterror / 1000000;
 1078         bt.frac = ((time_esterror - bt.sec) * 1000000) *
 1079             (uint64_t)18446744073709ULL;
 1080         clock_snap->fb_info.error = bt;
 1081 
 1082 #ifdef FFCLOCK
 1083         if (!fast)
 1084                 clock_snap->ffcount += delta;
 1085 
 1086         /* Record feed-forward clock leap second adjustment. */
 1087         ffi->leapsec_adjustment = cest.leapsec_total;
 1088         if (clock_snap->ffcount > cest.leapsec_next)
 1089                 ffi->leapsec_adjustment -= cest.leapsec;
 1090 
 1091         /* Record feed-forward clock status and error. */
 1092         clock_snap->ff_info.status = cest.status;
 1093         ffcount = clock_snap->ffcount - cest.update_ffcount;
 1094         ffclock_convert_delta(ffcount, cest.period, &bt);
 1095         /* 18446744073709 = int(2^64/1e12), err_bound_rate in [ps/s]. */
 1096         bintime_mul(&bt, cest.errb_rate * (uint64_t)18446744073709ULL);
 1097         /* 18446744073 = int(2^64 / 1e9), since err_abs in [ns]. */
 1098         bintime_addx(&bt, cest.errb_abs * (uint64_t)18446744073ULL);
 1099         clock_snap->ff_info.error = bt;
 1100 #endif
 1101 }
 1102 
 1103 /*
 1104  * Convert a sysclock snapshot into a struct bintime based on the specified
 1105  * clock source and flags.
 1106  */
 1107 int
 1108 sysclock_snap2bintime(struct sysclock_snap *cs, struct bintime *bt,
 1109     int whichclock, uint32_t flags)
 1110 {
 1111         struct bintime boottimebin;
 1112 #ifdef FFCLOCK
 1113         struct bintime bt2;
 1114         uint64_t period;
 1115 #endif
 1116 
 1117         switch (whichclock) {
 1118         case SYSCLOCK_FBCK:
 1119                 *bt = cs->fb_info.tick_time;
 1120 
 1121                 /* If snapshot was created with !fast, delta will be >0. */
 1122                 if (cs->delta > 0)
 1123                         bintime_addx(bt, cs->fb_info.th_scale * cs->delta);
 1124 
 1125                 if ((flags & FBCLOCK_UPTIME) == 0) {
 1126                         getboottimebin(&boottimebin);
 1127                         bintime_add(bt, &boottimebin);
 1128                 }
 1129                 break;
 1130 #ifdef FFCLOCK
 1131         case SYSCLOCK_FFWD:
 1132                 if (flags & FFCLOCK_LERP) {
 1133                         *bt = cs->ff_info.tick_time_lerp;
 1134                         period = cs->ff_info.period_lerp;
 1135                 } else {
 1136                         *bt = cs->ff_info.tick_time;
 1137                         period = cs->ff_info.period;
 1138                 }
 1139 
 1140                 /* If snapshot was created with !fast, delta will be >0. */
 1141                 if (cs->delta > 0) {
 1142                         ffclock_convert_delta(cs->delta, period, &bt2);
 1143                         bintime_add(bt, &bt2);
 1144                 }
 1145 
 1146                 /* Leap second adjustment. */
 1147                 if (flags & FFCLOCK_LEAPSEC)
 1148                         bt->sec -= cs->ff_info.leapsec_adjustment;
 1149 
 1150                 /* Boot time adjustment, for uptime/monotonic clocks. */
 1151                 if (flags & FFCLOCK_UPTIME)
 1152                         bintime_sub(bt, &ffclock_boottime);
 1153                 break;
 1154 #endif
 1155         default:
 1156                 return (EINVAL);
 1157                 break;
 1158         }
 1159 
 1160         return (0);
 1161 }
 1162 
 1163 /*
 1164  * Initialize a new timecounter and possibly use it.
 1165  */
 1166 void
 1167 tc_init(struct timecounter *tc)
 1168 {
 1169         u_int u;
 1170         struct sysctl_oid *tc_root;
 1171 
 1172         u = tc->tc_frequency / tc->tc_counter_mask;
 1173         /* XXX: We need some margin here, 10% is a guess */
 1174         u *= 11;
 1175         u /= 10;
 1176         if (u > hz && tc->tc_quality >= 0) {
 1177                 tc->tc_quality = -2000;
 1178                 if (bootverbose) {
 1179                         printf("Timecounter \"%s\" frequency %ju Hz",
 1180                             tc->tc_name, (uintmax_t)tc->tc_frequency);
 1181                         printf(" -- Insufficient hz, needs at least %u\n", u);
 1182                 }
 1183         } else if (tc->tc_quality >= 0 || bootverbose) {
 1184                 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
 1185                     tc->tc_name, (uintmax_t)tc->tc_frequency,
 1186                     tc->tc_quality);
 1187         }
 1188 
 1189         /*
 1190          * Set up sysctl tree for this counter.
 1191          */
 1192         tc_root = SYSCTL_ADD_NODE_WITH_LABEL(NULL,
 1193             SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name,
 1194             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 1195             "timecounter description", "timecounter");
 1196         SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 1197             "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0,
 1198             "mask for implemented bits");
 1199         SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 1200             "counter", CTLTYPE_UINT | CTLFLAG_RD | CTLFLAG_MPSAFE, tc,
 1201             sizeof(*tc), sysctl_kern_timecounter_get, "IU",
 1202             "current timecounter value");
 1203         SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 1204             "frequency", CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, tc,
 1205             sizeof(*tc), sysctl_kern_timecounter_freq, "QU",
 1206             "timecounter frequency");
 1207         SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 1208             "quality", CTLFLAG_RD, &(tc->tc_quality), 0,
 1209             "goodness of time counter");
 1210 
 1211         mtx_lock(&tc_lock);
 1212         tc->tc_next = timecounters;
 1213         timecounters = tc;
 1214 
 1215         /*
 1216          * Do not automatically switch if the current tc was specifically
 1217          * chosen.  Never automatically use a timecounter with negative quality.
 1218          * Even though we run on the dummy counter, switching here may be
 1219          * worse since this timecounter may not be monotonic.
 1220          */
 1221         if (tc_chosen)
 1222                 goto unlock;
 1223         if (tc->tc_quality < 0)
 1224                 goto unlock;
 1225         if (tc_from_tunable[0] != '\0' &&
 1226             strcmp(tc->tc_name, tc_from_tunable) == 0) {
 1227                 tc_chosen = 1;
 1228                 tc_from_tunable[0] = '\0';
 1229         } else {
 1230                 if (tc->tc_quality < timecounter->tc_quality)
 1231                         goto unlock;
 1232                 if (tc->tc_quality == timecounter->tc_quality &&
 1233                     tc->tc_frequency < timecounter->tc_frequency)
 1234                         goto unlock;
 1235         }
 1236         (void)tc->tc_get_timecount(tc);
 1237         timecounter = tc;
 1238 unlock:
 1239         mtx_unlock(&tc_lock);
 1240 }
 1241 
 1242 /* Report the frequency of the current timecounter. */
 1243 uint64_t
 1244 tc_getfrequency(void)
 1245 {
 1246 
 1247         return (timehands->th_counter->tc_frequency);
 1248 }
 1249 
 1250 static bool
 1251 sleeping_on_old_rtc(struct thread *td)
 1252 {
 1253 
 1254         /*
 1255          * td_rtcgen is modified by curthread when it is running,
 1256          * and by other threads in this function.  By finding the thread
 1257          * on a sleepqueue and holding the lock on the sleepqueue
 1258          * chain, we guarantee that the thread is not running and that
 1259          * modifying td_rtcgen is safe.  Setting td_rtcgen to zero informs
 1260          * the thread that it was woken due to a real-time clock adjustment.
 1261          * (The declaration of td_rtcgen refers to this comment.)
 1262          */
 1263         if (td->td_rtcgen != 0 && td->td_rtcgen != rtc_generation) {
 1264                 td->td_rtcgen = 0;
 1265                 return (true);
 1266         }
 1267         return (false);
 1268 }
 1269 
 1270 static struct mtx tc_setclock_mtx;
 1271 MTX_SYSINIT(tc_setclock_init, &tc_setclock_mtx, "tcsetc", MTX_SPIN);
 1272 
 1273 /*
 1274  * Step our concept of UTC.  This is done by modifying our estimate of
 1275  * when we booted.
 1276  */
 1277 void
 1278 tc_setclock(struct timespec *ts)
 1279 {
 1280         struct timespec tbef, taft;
 1281         struct bintime bt, bt2;
 1282 
 1283         timespec2bintime(ts, &bt);
 1284         nanotime(&tbef);
 1285         mtx_lock_spin(&tc_setclock_mtx);
 1286         cpu_tick_calibrate(1);
 1287         binuptime(&bt2);
 1288         bintime_sub(&bt, &bt2);
 1289 
 1290         /* XXX fiddle all the little crinkly bits around the fiords... */
 1291         tc_windup(&bt);
 1292         mtx_unlock_spin(&tc_setclock_mtx);
 1293 
 1294         /* Avoid rtc_generation == 0, since td_rtcgen == 0 is special. */
 1295         atomic_add_rel_int(&rtc_generation, 2);
 1296         sleepq_chains_remove_matching(sleeping_on_old_rtc);
 1297         if (timestepwarnings) {
 1298                 nanotime(&taft);
 1299                 log(LOG_INFO,
 1300                     "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
 1301                     (intmax_t)tbef.tv_sec, tbef.tv_nsec,
 1302                     (intmax_t)taft.tv_sec, taft.tv_nsec,
 1303                     (intmax_t)ts->tv_sec, ts->tv_nsec);
 1304         }
 1305 }
 1306 
 1307 /*
 1308  * Recalculate the scaling factor.  We want the number of 1/2^64
 1309  * fractions of a second per period of the hardware counter, taking
 1310  * into account the th_adjustment factor which the NTP PLL/adjtime(2)
 1311  * processing provides us with.
 1312  *
 1313  * The th_adjustment is nanoseconds per second with 32 bit binary
 1314  * fraction and we want 64 bit binary fraction of second:
 1315  *
 1316  *       x = a * 2^32 / 10^9 = a * 4.294967296
 1317  *
 1318  * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
 1319  * we can only multiply by about 850 without overflowing, that
 1320  * leaves no suitably precise fractions for multiply before divide.
 1321  *
 1322  * Divide before multiply with a fraction of 2199/512 results in a
 1323  * systematic undercompensation of 10PPM of th_adjustment.  On a
 1324  * 5000PPM adjustment this is a 0.05PPM error.  This is acceptable.
 1325  *
 1326  * We happily sacrifice the lowest of the 64 bits of our result
 1327  * to the goddess of code clarity.
 1328  */
 1329 static void
 1330 recalculate_scaling_factor_and_large_delta(struct timehands *th)
 1331 {
 1332         uint64_t scale;
 1333 
 1334         scale = (uint64_t)1 << 63;
 1335         scale += (th->th_adjustment / 1024) * 2199;
 1336         scale /= th->th_counter->tc_frequency;
 1337         th->th_scale = scale * 2;
 1338         th->th_large_delta = MIN(((uint64_t)1 << 63) / scale, UINT_MAX);
 1339 }
 1340 
 1341 /*
 1342  * Initialize the next struct timehands in the ring and make
 1343  * it the active timehands.  Along the way we might switch to a different
 1344  * timecounter and/or do seconds processing in NTP.  Slightly magic.
 1345  */
 1346 static void
 1347 tc_windup(struct bintime *new_boottimebin)
 1348 {
 1349         struct bintime bt;
 1350         struct timehands *th, *tho;
 1351         u_int delta, ncount, ogen;
 1352         int i;
 1353         time_t t;
 1354 
 1355         /*
 1356          * Make the next timehands a copy of the current one, but do
 1357          * not overwrite the generation or next pointer.  While we
 1358          * update the contents, the generation must be zero.  We need
 1359          * to ensure that the zero generation is visible before the
 1360          * data updates become visible, which requires release fence.
 1361          * For similar reasons, re-reading of the generation after the
 1362          * data is read should use acquire fence.
 1363          */
 1364         tho = timehands;
 1365         th = tho->th_next;
 1366         ogen = th->th_generation;
 1367         th->th_generation = 0;
 1368         atomic_thread_fence_rel();
 1369         memcpy(th, tho, offsetof(struct timehands, th_generation));
 1370         if (new_boottimebin != NULL)
 1371                 th->th_boottime = *new_boottimebin;
 1372 
 1373         /*
 1374          * Capture a timecounter delta on the current timecounter and if
 1375          * changing timecounters, a counter value from the new timecounter.
 1376          * Update the offset fields accordingly.
 1377          */
 1378         delta = tc_delta(th);
 1379         if (th->th_counter != timecounter)
 1380                 ncount = timecounter->tc_get_timecount(timecounter);
 1381         else
 1382                 ncount = 0;
 1383 #ifdef FFCLOCK
 1384         ffclock_windup(delta);
 1385 #endif
 1386         th->th_offset_count += delta;
 1387         th->th_offset_count &= th->th_counter->tc_counter_mask;
 1388         while (delta > th->th_counter->tc_frequency) {
 1389                 /* Eat complete unadjusted seconds. */
 1390                 delta -= th->th_counter->tc_frequency;
 1391                 th->th_offset.sec++;
 1392         }
 1393         if ((delta > th->th_counter->tc_frequency / 2) &&
 1394             (th->th_scale * delta < ((uint64_t)1 << 63))) {
 1395                 /* The product th_scale * delta just barely overflows. */
 1396                 th->th_offset.sec++;
 1397         }
 1398         bintime_addx(&th->th_offset, th->th_scale * delta);
 1399 
 1400         /*
 1401          * Hardware latching timecounters may not generate interrupts on
 1402          * PPS events, so instead we poll them.  There is a finite risk that
 1403          * the hardware might capture a count which is later than the one we
 1404          * got above, and therefore possibly in the next NTP second which might
 1405          * have a different rate than the current NTP second.  It doesn't
 1406          * matter in practice.
 1407          */
 1408         if (tho->th_counter->tc_poll_pps)
 1409                 tho->th_counter->tc_poll_pps(tho->th_counter);
 1410 
 1411         /*
 1412          * Deal with NTP second processing.  The loop normally
 1413          * iterates at most once, but in extreme situations it might
 1414          * keep NTP sane if timeouts are not run for several seconds.
 1415          * At boot, the time step can be large when the TOD hardware
 1416          * has been read, so on really large steps, we call
 1417          * ntp_update_second only twice.  We need to call it twice in
 1418          * case we missed a leap second.
 1419          */
 1420         bt = th->th_offset;
 1421         bintime_add(&bt, &th->th_boottime);
 1422         i = bt.sec - tho->th_microtime.tv_sec;
 1423         if (i > 0) {
 1424                 if (i > LARGE_STEP)
 1425                         i = 2;
 1426 
 1427                 do {
 1428                         t = bt.sec;
 1429                         ntp_update_second(&th->th_adjustment, &bt.sec);
 1430                         if (bt.sec != t)
 1431                                 th->th_boottime.sec += bt.sec - t;
 1432                         --i;
 1433                 } while (i > 0);
 1434 
 1435                 recalculate_scaling_factor_and_large_delta(th);
 1436         }
 1437 
 1438         /* Update the UTC timestamps used by the get*() functions. */
 1439         th->th_bintime = bt;
 1440         bintime2timeval(&bt, &th->th_microtime);
 1441         bintime2timespec(&bt, &th->th_nanotime);
 1442 
 1443         /* Now is a good time to change timecounters. */
 1444         if (th->th_counter != timecounter) {
 1445 #ifndef __arm__
 1446                 if ((timecounter->tc_flags & TC_FLAGS_C2STOP) != 0)
 1447                         cpu_disable_c2_sleep++;
 1448                 if ((th->th_counter->tc_flags & TC_FLAGS_C2STOP) != 0)
 1449                         cpu_disable_c2_sleep--;
 1450 #endif
 1451                 th->th_counter = timecounter;
 1452                 th->th_offset_count = ncount;
 1453                 tc_min_ticktock_freq = max(1, timecounter->tc_frequency /
 1454                     (((uint64_t)timecounter->tc_counter_mask + 1) / 3));
 1455                 recalculate_scaling_factor_and_large_delta(th);
 1456 #ifdef FFCLOCK
 1457                 ffclock_change_tc(th);
 1458 #endif
 1459         }
 1460 
 1461         /*
 1462          * Now that the struct timehands is again consistent, set the new
 1463          * generation number, making sure to not make it zero.
 1464          */
 1465         if (++ogen == 0)
 1466                 ogen = 1;
 1467         atomic_store_rel_int(&th->th_generation, ogen);
 1468 
 1469         /* Go live with the new struct timehands. */
 1470 #ifdef FFCLOCK
 1471         switch (sysclock_active) {
 1472         case SYSCLOCK_FBCK:
 1473 #endif
 1474                 time_second = th->th_microtime.tv_sec;
 1475                 time_uptime = th->th_offset.sec;
 1476 #ifdef FFCLOCK
 1477                 break;
 1478         case SYSCLOCK_FFWD:
 1479                 time_second = fftimehands->tick_time_lerp.sec;
 1480                 time_uptime = fftimehands->tick_time_lerp.sec - ffclock_boottime.sec;
 1481                 break;
 1482         }
 1483 #endif
 1484 
 1485         timehands = th;
 1486         timekeep_push_vdso();
 1487 }
 1488 
 1489 /* Report or change the active timecounter hardware. */
 1490 static int
 1491 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
 1492 {
 1493         char newname[32];
 1494         struct timecounter *newtc, *tc;
 1495         int error;
 1496 
 1497         mtx_lock(&tc_lock);
 1498         tc = timecounter;
 1499         strlcpy(newname, tc->tc_name, sizeof(newname));
 1500         mtx_unlock(&tc_lock);
 1501 
 1502         error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
 1503         if (error != 0 || req->newptr == NULL)
 1504                 return (error);
 1505 
 1506         mtx_lock(&tc_lock);
 1507         /* Record that the tc in use now was specifically chosen. */
 1508         tc_chosen = 1;
 1509         if (strcmp(newname, tc->tc_name) == 0) {
 1510                 mtx_unlock(&tc_lock);
 1511                 return (0);
 1512         }
 1513         for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
 1514                 if (strcmp(newname, newtc->tc_name) != 0)
 1515                         continue;
 1516 
 1517                 /* Warm up new timecounter. */
 1518                 (void)newtc->tc_get_timecount(newtc);
 1519 
 1520                 timecounter = newtc;
 1521 
 1522                 /*
 1523                  * The vdso timehands update is deferred until the next
 1524                  * 'tc_windup()'.
 1525                  *
 1526                  * This is prudent given that 'timekeep_push_vdso()' does not
 1527                  * use any locking and that it can be called in hard interrupt
 1528                  * context via 'tc_windup()'.
 1529                  */
 1530                 break;
 1531         }
 1532         mtx_unlock(&tc_lock);
 1533         return (newtc != NULL ? 0 : EINVAL);
 1534 }
 1535 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware,
 1536     CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, 0, 0,
 1537     sysctl_kern_timecounter_hardware, "A",
 1538     "Timecounter hardware selected");
 1539 
 1540 /* Report the available timecounter hardware. */
 1541 static int
 1542 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
 1543 {
 1544         struct sbuf sb;
 1545         struct timecounter *tc;
 1546         int error;
 1547 
 1548         error = sysctl_wire_old_buffer(req, 0);
 1549         if (error != 0)
 1550                 return (error);
 1551         sbuf_new_for_sysctl(&sb, NULL, 0, req);
 1552         mtx_lock(&tc_lock);
 1553         for (tc = timecounters; tc != NULL; tc = tc->tc_next) {
 1554                 if (tc != timecounters)
 1555                         sbuf_putc(&sb, ' ');
 1556                 sbuf_printf(&sb, "%s(%d)", tc->tc_name, tc->tc_quality);
 1557         }
 1558         mtx_unlock(&tc_lock);
 1559         error = sbuf_finish(&sb);
 1560         sbuf_delete(&sb);
 1561         return (error);
 1562 }
 1563 
 1564 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice,
 1565     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0,
 1566     sysctl_kern_timecounter_choice, "A",
 1567     "Timecounter hardware detected");
 1568 
 1569 /*
 1570  * RFC 2783 PPS-API implementation.
 1571  */
 1572 
 1573 /*
 1574  *  Return true if the driver is aware of the abi version extensions in the
 1575  *  pps_state structure, and it supports at least the given abi version number.
 1576  */
 1577 static inline int
 1578 abi_aware(struct pps_state *pps, int vers)
 1579 {
 1580 
 1581         return ((pps->kcmode & KCMODE_ABIFLAG) && pps->driver_abi >= vers);
 1582 }
 1583 
 1584 static int
 1585 pps_fetch(struct pps_fetch_args *fapi, struct pps_state *pps)
 1586 {
 1587         int err, timo;
 1588         pps_seq_t aseq, cseq;
 1589         struct timeval tv;
 1590 
 1591         if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
 1592                 return (EINVAL);
 1593 
 1594         /*
 1595          * If no timeout is requested, immediately return whatever values were
 1596          * most recently captured.  If timeout seconds is -1, that's a request
 1597          * to block without a timeout.  WITNESS won't let us sleep forever
 1598          * without a lock (we really don't need a lock), so just repeatedly
 1599          * sleep a long time.
 1600          */
 1601         if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) {
 1602                 if (fapi->timeout.tv_sec == -1)
 1603                         timo = 0x7fffffff;
 1604                 else {
 1605                         tv.tv_sec = fapi->timeout.tv_sec;
 1606                         tv.tv_usec = fapi->timeout.tv_nsec / 1000;
 1607                         timo = tvtohz(&tv);
 1608                 }
 1609                 aseq = atomic_load_int(&pps->ppsinfo.assert_sequence);
 1610                 cseq = atomic_load_int(&pps->ppsinfo.clear_sequence);
 1611                 while (aseq == atomic_load_int(&pps->ppsinfo.assert_sequence) &&
 1612                     cseq == atomic_load_int(&pps->ppsinfo.clear_sequence)) {
 1613                         if (abi_aware(pps, 1) && pps->driver_mtx != NULL) {
 1614                                 if (pps->flags & PPSFLAG_MTX_SPIN) {
 1615                                         err = msleep_spin(pps, pps->driver_mtx,
 1616                                             "ppsfch", timo);
 1617                                 } else {
 1618                                         err = msleep(pps, pps->driver_mtx, PCATCH,
 1619                                             "ppsfch", timo);
 1620                                 }
 1621                         } else {
 1622                                 err = tsleep(pps, PCATCH, "ppsfch", timo);
 1623                         }
 1624                         if (err == EWOULDBLOCK) {
 1625                                 if (fapi->timeout.tv_sec == -1) {
 1626                                         continue;
 1627                                 } else {
 1628                                         return (ETIMEDOUT);
 1629                                 }
 1630                         } else if (err != 0) {
 1631                                 return (err);
 1632                         }
 1633                 }
 1634         }
 1635 
 1636         pps->ppsinfo.current_mode = pps->ppsparam.mode;
 1637         fapi->pps_info_buf = pps->ppsinfo;
 1638 
 1639         return (0);
 1640 }
 1641 
 1642 int
 1643 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
 1644 {
 1645         pps_params_t *app;
 1646         struct pps_fetch_args *fapi;
 1647 #ifdef FFCLOCK
 1648         struct pps_fetch_ffc_args *fapi_ffc;
 1649 #endif
 1650 #ifdef PPS_SYNC
 1651         struct pps_kcbind_args *kapi;
 1652 #endif
 1653 
 1654         KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
 1655         switch (cmd) {
 1656         case PPS_IOC_CREATE:
 1657                 return (0);
 1658         case PPS_IOC_DESTROY:
 1659                 return (0);
 1660         case PPS_IOC_SETPARAMS:
 1661                 app = (pps_params_t *)data;
 1662                 if (app->mode & ~pps->ppscap)
 1663                         return (EINVAL);
 1664 #ifdef FFCLOCK
 1665                 /* Ensure only a single clock is selected for ffc timestamp. */
 1666                 if ((app->mode & PPS_TSCLK_MASK) == PPS_TSCLK_MASK)
 1667                         return (EINVAL);
 1668 #endif
 1669                 pps->ppsparam = *app;
 1670                 return (0);
 1671         case PPS_IOC_GETPARAMS:
 1672                 app = (pps_params_t *)data;
 1673                 *app = pps->ppsparam;
 1674                 app->api_version = PPS_API_VERS_1;
 1675                 return (0);
 1676         case PPS_IOC_GETCAP:
 1677                 *(int*)data = pps->ppscap;
 1678                 return (0);
 1679         case PPS_IOC_FETCH:
 1680                 fapi = (struct pps_fetch_args *)data;
 1681                 return (pps_fetch(fapi, pps));
 1682 #ifdef FFCLOCK
 1683         case PPS_IOC_FETCH_FFCOUNTER:
 1684                 fapi_ffc = (struct pps_fetch_ffc_args *)data;
 1685                 if (fapi_ffc->tsformat && fapi_ffc->tsformat !=
 1686                     PPS_TSFMT_TSPEC)
 1687                         return (EINVAL);
 1688                 if (fapi_ffc->timeout.tv_sec || fapi_ffc->timeout.tv_nsec)
 1689                         return (EOPNOTSUPP);
 1690                 pps->ppsinfo_ffc.current_mode = pps->ppsparam.mode;
 1691                 fapi_ffc->pps_info_buf_ffc = pps->ppsinfo_ffc;
 1692                 /* Overwrite timestamps if feedback clock selected. */
 1693                 switch (pps->ppsparam.mode & PPS_TSCLK_MASK) {
 1694                 case PPS_TSCLK_FBCK:
 1695                         fapi_ffc->pps_info_buf_ffc.assert_timestamp =
 1696                             pps->ppsinfo.assert_timestamp;
 1697                         fapi_ffc->pps_info_buf_ffc.clear_timestamp =
 1698                             pps->ppsinfo.clear_timestamp;
 1699                         break;
 1700                 case PPS_TSCLK_FFWD:
 1701                         break;
 1702                 default:
 1703                         break;
 1704                 }
 1705                 return (0);
 1706 #endif /* FFCLOCK */
 1707         case PPS_IOC_KCBIND:
 1708 #ifdef PPS_SYNC
 1709                 kapi = (struct pps_kcbind_args *)data;
 1710                 /* XXX Only root should be able to do this */
 1711                 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
 1712                         return (EINVAL);
 1713                 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
 1714                         return (EINVAL);
 1715                 if (kapi->edge & ~pps->ppscap)
 1716                         return (EINVAL);
 1717                 pps->kcmode = (kapi->edge & KCMODE_EDGEMASK) |
 1718                     (pps->kcmode & KCMODE_ABIFLAG);
 1719                 return (0);
 1720 #else
 1721                 return (EOPNOTSUPP);
 1722 #endif
 1723         default:
 1724                 return (ENOIOCTL);
 1725         }
 1726 }
 1727 
 1728 void
 1729 pps_init(struct pps_state *pps)
 1730 {
 1731         pps->ppscap |= PPS_TSFMT_TSPEC | PPS_CANWAIT;
 1732         if (pps->ppscap & PPS_CAPTUREASSERT)
 1733                 pps->ppscap |= PPS_OFFSETASSERT;
 1734         if (pps->ppscap & PPS_CAPTURECLEAR)
 1735                 pps->ppscap |= PPS_OFFSETCLEAR;
 1736 #ifdef FFCLOCK
 1737         pps->ppscap |= PPS_TSCLK_MASK;
 1738 #endif
 1739         pps->kcmode &= ~KCMODE_ABIFLAG;
 1740 }
 1741 
 1742 void
 1743 pps_init_abi(struct pps_state *pps)
 1744 {
 1745 
 1746         pps_init(pps);
 1747         if (pps->driver_abi > 0) {
 1748                 pps->kcmode |= KCMODE_ABIFLAG;
 1749                 pps->kernel_abi = PPS_ABI_VERSION;
 1750         }
 1751 }
 1752 
 1753 void
 1754 pps_capture(struct pps_state *pps)
 1755 {
 1756         struct timehands *th;
 1757 
 1758         KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
 1759         th = timehands;
 1760         pps->capgen = atomic_load_acq_int(&th->th_generation);
 1761         pps->capth = th;
 1762 #ifdef FFCLOCK
 1763         pps->capffth = fftimehands;
 1764 #endif
 1765         pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
 1766         atomic_thread_fence_acq();
 1767         if (pps->capgen != th->th_generation)
 1768                 pps->capgen = 0;
 1769 }
 1770 
 1771 void
 1772 pps_event(struct pps_state *pps, int event)
 1773 {
 1774         struct bintime bt;
 1775         struct timespec ts, *tsp, *osp;
 1776         u_int tcount, *pcount;
 1777         int foff;
 1778         pps_seq_t *pseq;
 1779 #ifdef FFCLOCK
 1780         struct timespec *tsp_ffc;
 1781         pps_seq_t *pseq_ffc;
 1782         ffcounter *ffcount;
 1783 #endif
 1784 #ifdef PPS_SYNC
 1785         int fhard;
 1786 #endif
 1787 
 1788         KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
 1789         /* Nothing to do if not currently set to capture this event type. */
 1790         if ((event & pps->ppsparam.mode) == 0)
 1791                 return;
 1792         /* If the timecounter was wound up underneath us, bail out. */
 1793         if (pps->capgen == 0 || pps->capgen !=
 1794             atomic_load_acq_int(&pps->capth->th_generation))
 1795                 return;
 1796 
 1797         /* Things would be easier with arrays. */
 1798         if (event == PPS_CAPTUREASSERT) {
 1799                 tsp = &pps->ppsinfo.assert_timestamp;
 1800                 osp = &pps->ppsparam.assert_offset;
 1801                 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
 1802 #ifdef PPS_SYNC
 1803                 fhard = pps->kcmode & PPS_CAPTUREASSERT;
 1804 #endif
 1805                 pcount = &pps->ppscount[0];
 1806                 pseq = &pps->ppsinfo.assert_sequence;
 1807 #ifdef FFCLOCK
 1808                 ffcount = &pps->ppsinfo_ffc.assert_ffcount;
 1809                 tsp_ffc = &pps->ppsinfo_ffc.assert_timestamp;
 1810                 pseq_ffc = &pps->ppsinfo_ffc.assert_sequence;
 1811 #endif
 1812         } else {
 1813                 tsp = &pps->ppsinfo.clear_timestamp;
 1814                 osp = &pps->ppsparam.clear_offset;
 1815                 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
 1816 #ifdef PPS_SYNC
 1817                 fhard = pps->kcmode & PPS_CAPTURECLEAR;
 1818 #endif
 1819                 pcount = &pps->ppscount[1];
 1820                 pseq = &pps->ppsinfo.clear_sequence;
 1821 #ifdef FFCLOCK
 1822                 ffcount = &pps->ppsinfo_ffc.clear_ffcount;
 1823                 tsp_ffc = &pps->ppsinfo_ffc.clear_timestamp;
 1824                 pseq_ffc = &pps->ppsinfo_ffc.clear_sequence;
 1825 #endif
 1826         }
 1827 
 1828         /*
 1829          * If the timecounter changed, we cannot compare the count values, so
 1830          * we have to drop the rest of the PPS-stuff until the next event.
 1831          */
 1832         if (pps->ppstc != pps->capth->th_counter) {
 1833                 pps->ppstc = pps->capth->th_counter;
 1834                 *pcount = pps->capcount;
 1835                 pps->ppscount[2] = pps->capcount;
 1836                 return;
 1837         }
 1838 
 1839         /* Convert the count to a timespec. */
 1840         tcount = pps->capcount - pps->capth->th_offset_count;
 1841         tcount &= pps->capth->th_counter->tc_counter_mask;
 1842         bt = pps->capth->th_bintime;
 1843         bintime_addx(&bt, pps->capth->th_scale * tcount);
 1844         bintime2timespec(&bt, &ts);
 1845 
 1846         /* If the timecounter was wound up underneath us, bail out. */
 1847         atomic_thread_fence_acq();
 1848         if (pps->capgen != pps->capth->th_generation)
 1849                 return;
 1850 
 1851         *pcount = pps->capcount;
 1852         (*pseq)++;
 1853         *tsp = ts;
 1854 
 1855         if (foff) {
 1856                 timespecadd(tsp, osp, tsp);
 1857                 if (tsp->tv_nsec < 0) {
 1858                         tsp->tv_nsec += 1000000000;
 1859                         tsp->tv_sec -= 1;
 1860                 }
 1861         }
 1862 
 1863 #ifdef FFCLOCK
 1864         *ffcount = pps->capffth->tick_ffcount + tcount;
 1865         bt = pps->capffth->tick_time;
 1866         ffclock_convert_delta(tcount, pps->capffth->cest.period, &bt);
 1867         bintime_add(&bt, &pps->capffth->tick_time);
 1868         bintime2timespec(&bt, &ts);
 1869         (*pseq_ffc)++;
 1870         *tsp_ffc = ts;
 1871 #endif
 1872 
 1873 #ifdef PPS_SYNC
 1874         if (fhard) {
 1875                 uint64_t scale;
 1876 
 1877                 /*
 1878                  * Feed the NTP PLL/FLL.
 1879                  * The FLL wants to know how many (hardware) nanoseconds
 1880                  * elapsed since the previous event.
 1881                  */
 1882                 tcount = pps->capcount - pps->ppscount[2];
 1883                 pps->ppscount[2] = pps->capcount;
 1884                 tcount &= pps->capth->th_counter->tc_counter_mask;
 1885                 scale = (uint64_t)1 << 63;
 1886                 scale /= pps->capth->th_counter->tc_frequency;
 1887                 scale *= 2;
 1888                 bt.sec = 0;
 1889                 bt.frac = 0;
 1890                 bintime_addx(&bt, scale * tcount);
 1891                 bintime2timespec(&bt, &ts);
 1892                 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
 1893         }
 1894 #endif
 1895 
 1896         /* Wakeup anyone sleeping in pps_fetch().  */
 1897         wakeup(pps);
 1898 }
 1899 
 1900 /*
 1901  * Timecounters need to be updated every so often to prevent the hardware
 1902  * counter from overflowing.  Updating also recalculates the cached values
 1903  * used by the get*() family of functions, so their precision depends on
 1904  * the update frequency.
 1905  */
 1906 
 1907 static int tc_tick;
 1908 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0,
 1909     "Approximate number of hardclock ticks in a millisecond");
 1910 
 1911 void
 1912 tc_ticktock(int cnt)
 1913 {
 1914         static int count;
 1915 
 1916         if (mtx_trylock_spin(&tc_setclock_mtx)) {
 1917                 count += cnt;
 1918                 if (count >= tc_tick) {
 1919                         count = 0;
 1920                         tc_windup(NULL);
 1921                 }
 1922                 mtx_unlock_spin(&tc_setclock_mtx);
 1923         }
 1924 }
 1925 
 1926 static void __inline
 1927 tc_adjprecision(void)
 1928 {
 1929         int t;
 1930 
 1931         if (tc_timepercentage > 0) {
 1932                 t = (99 + tc_timepercentage) / tc_timepercentage;
 1933                 tc_precexp = fls(t + (t >> 1)) - 1;
 1934                 FREQ2BT(hz / tc_tick, &bt_timethreshold);
 1935                 FREQ2BT(hz, &bt_tickthreshold);
 1936                 bintime_shift(&bt_timethreshold, tc_precexp);
 1937                 bintime_shift(&bt_tickthreshold, tc_precexp);
 1938         } else {
 1939                 tc_precexp = 31;
 1940                 bt_timethreshold.sec = INT_MAX;
 1941                 bt_timethreshold.frac = ~(uint64_t)0;
 1942                 bt_tickthreshold = bt_timethreshold;
 1943         }
 1944         sbt_timethreshold = bttosbt(bt_timethreshold);
 1945         sbt_tickthreshold = bttosbt(bt_tickthreshold);
 1946 }
 1947 
 1948 static int
 1949 sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS)
 1950 {
 1951         int error, val;
 1952 
 1953         val = tc_timepercentage;
 1954         error = sysctl_handle_int(oidp, &val, 0, req);
 1955         if (error != 0 || req->newptr == NULL)
 1956                 return (error);
 1957         tc_timepercentage = val;
 1958         if (cold)
 1959                 goto done;
 1960         tc_adjprecision();
 1961 done:
 1962         return (0);
 1963 }
 1964 
 1965 /* Set up the requested number of timehands. */
 1966 static void
 1967 inittimehands(void *dummy)
 1968 {
 1969         struct timehands *thp;
 1970         int i;
 1971 
 1972         TUNABLE_INT_FETCH("kern.timecounter.timehands_count",
 1973             &timehands_count);
 1974         if (timehands_count < 1)
 1975                 timehands_count = 1;
 1976         if (timehands_count > nitems(ths))
 1977                 timehands_count = nitems(ths);
 1978         for (i = 1, thp = &ths[0]; i < timehands_count;  thp = &ths[i++])
 1979                 thp->th_next = &ths[i];
 1980         thp->th_next = &ths[0];
 1981 
 1982         TUNABLE_STR_FETCH("kern.timecounter.hardware", tc_from_tunable,
 1983             sizeof(tc_from_tunable));
 1984 
 1985         mtx_init(&tc_lock, "tc", NULL, MTX_DEF);
 1986 }
 1987 SYSINIT(timehands, SI_SUB_TUNABLES, SI_ORDER_ANY, inittimehands, NULL);
 1988 
 1989 static void
 1990 inittimecounter(void *dummy)
 1991 {
 1992         u_int p;
 1993         int tick_rate;
 1994 
 1995         /*
 1996          * Set the initial timeout to
 1997          * max(1, <approx. number of hardclock ticks in a millisecond>).
 1998          * People should probably not use the sysctl to set the timeout
 1999          * to smaller than its initial value, since that value is the
 2000          * smallest reasonable one.  If they want better timestamps they
 2001          * should use the non-"get"* functions.
 2002          */
 2003         if (hz > 1000)
 2004                 tc_tick = (hz + 500) / 1000;
 2005         else
 2006                 tc_tick = 1;
 2007         tc_adjprecision();
 2008         FREQ2BT(hz, &tick_bt);
 2009         tick_sbt = bttosbt(tick_bt);
 2010         tick_rate = hz / tc_tick;
 2011         FREQ2BT(tick_rate, &tc_tick_bt);
 2012         tc_tick_sbt = bttosbt(tc_tick_bt);
 2013         p = (tc_tick * 1000000) / hz;
 2014         printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
 2015 
 2016 #ifdef FFCLOCK
 2017         ffclock_init();
 2018 #endif
 2019 
 2020         /* warm up new timecounter (again) and get rolling. */
 2021         (void)timecounter->tc_get_timecount(timecounter);
 2022         mtx_lock_spin(&tc_setclock_mtx);
 2023         tc_windup(NULL);
 2024         mtx_unlock_spin(&tc_setclock_mtx);
 2025 }
 2026 
 2027 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL);
 2028 
 2029 /* Cpu tick handling -------------------------------------------------*/
 2030 
 2031 static int cpu_tick_variable;
 2032 static uint64_t cpu_tick_frequency;
 2033 
 2034 DPCPU_DEFINE_STATIC(uint64_t, tc_cpu_ticks_base);
 2035 DPCPU_DEFINE_STATIC(unsigned, tc_cpu_ticks_last);
 2036 
 2037 static uint64_t
 2038 tc_cpu_ticks(void)
 2039 {
 2040         struct timecounter *tc;
 2041         uint64_t res, *base;
 2042         unsigned u, *last;
 2043 
 2044         critical_enter();
 2045         base = DPCPU_PTR(tc_cpu_ticks_base);
 2046         last = DPCPU_PTR(tc_cpu_ticks_last);
 2047         tc = timehands->th_counter;
 2048         u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
 2049         if (u < *last)
 2050                 *base += (uint64_t)tc->tc_counter_mask + 1;
 2051         *last = u;
 2052         res = u + *base;
 2053         critical_exit();
 2054         return (res);
 2055 }
 2056 
 2057 void
 2058 cpu_tick_calibration(void)
 2059 {
 2060         static time_t last_calib;
 2061 
 2062         if (time_uptime != last_calib && !(time_uptime & 0xf)) {
 2063                 cpu_tick_calibrate(0);
 2064                 last_calib = time_uptime;
 2065         }
 2066 }
 2067 
 2068 /*
 2069  * This function gets called every 16 seconds on only one designated
 2070  * CPU in the system from hardclock() via cpu_tick_calibration()().
 2071  *
 2072  * Whenever the real time clock is stepped we get called with reset=1
 2073  * to make sure we handle suspend/resume and similar events correctly.
 2074  */
 2075 
 2076 static void
 2077 cpu_tick_calibrate(int reset)
 2078 {
 2079         static uint64_t c_last;
 2080         uint64_t c_this, c_delta;
 2081         static struct bintime  t_last;
 2082         struct bintime t_this, t_delta;
 2083         uint32_t divi;
 2084 
 2085         if (reset) {
 2086                 /* The clock was stepped, abort & reset */
 2087                 t_last.sec = 0;
 2088                 return;
 2089         }
 2090 
 2091         /* we don't calibrate fixed rate cputicks */
 2092         if (!cpu_tick_variable)
 2093                 return;
 2094 
 2095         getbinuptime(&t_this);
 2096         c_this = cpu_ticks();
 2097         if (t_last.sec != 0) {
 2098                 c_delta = c_this - c_last;
 2099                 t_delta = t_this;
 2100                 bintime_sub(&t_delta, &t_last);
 2101                 /*
 2102                  * Headroom:
 2103                  *      2^(64-20) / 16[s] =
 2104                  *      2^(44) / 16[s] =
 2105                  *      17.592.186.044.416 / 16 =
 2106                  *      1.099.511.627.776 [Hz]
 2107                  */
 2108                 divi = t_delta.sec << 20;
 2109                 divi |= t_delta.frac >> (64 - 20);
 2110                 c_delta <<= 20;
 2111                 c_delta /= divi;
 2112                 if (c_delta > cpu_tick_frequency) {
 2113                         if (0 && bootverbose)
 2114                                 printf("cpu_tick increased to %ju Hz\n",
 2115                                     c_delta);
 2116                         cpu_tick_frequency = c_delta;
 2117                 }
 2118         }
 2119         c_last = c_this;
 2120         t_last = t_this;
 2121 }
 2122 
 2123 void
 2124 set_cputicker(cpu_tick_f *func, uint64_t freq, unsigned var)
 2125 {
 2126 
 2127         if (func == NULL) {
 2128                 cpu_ticks = tc_cpu_ticks;
 2129         } else {
 2130                 cpu_tick_frequency = freq;
 2131                 cpu_tick_variable = var;
 2132                 cpu_ticks = func;
 2133         }
 2134 }
 2135 
 2136 uint64_t
 2137 cpu_tickrate(void)
 2138 {
 2139 
 2140         if (cpu_ticks == tc_cpu_ticks) 
 2141                 return (tc_getfrequency());
 2142         return (cpu_tick_frequency);
 2143 }
 2144 
 2145 /*
 2146  * We need to be slightly careful converting cputicks to microseconds.
 2147  * There is plenty of margin in 64 bits of microseconds (half a million
 2148  * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
 2149  * before divide conversion (to retain precision) we find that the
 2150  * margin shrinks to 1.5 hours (one millionth of 146y).
 2151  * With a three prong approach we never lose significant bits, no
 2152  * matter what the cputick rate and length of timeinterval is.
 2153  */
 2154 
 2155 uint64_t
 2156 cputick2usec(uint64_t tick)
 2157 {
 2158 
 2159         if (tick > 18446744073709551LL)         /* floor(2^64 / 1000) */
 2160                 return (tick / (cpu_tickrate() / 1000000LL));
 2161         else if (tick > 18446744073709LL)       /* floor(2^64 / 1000000) */
 2162                 return ((tick * 1000LL) / (cpu_tickrate() / 1000LL));
 2163         else
 2164                 return ((tick * 1000000LL) / cpu_tickrate());
 2165 }
 2166 
 2167 cpu_tick_f      *cpu_ticks = tc_cpu_ticks;
 2168 
 2169 static int vdso_th_enable = 1;
 2170 static int
 2171 sysctl_fast_gettime(SYSCTL_HANDLER_ARGS)
 2172 {
 2173         int old_vdso_th_enable, error;
 2174 
 2175         old_vdso_th_enable = vdso_th_enable;
 2176         error = sysctl_handle_int(oidp, &old_vdso_th_enable, 0, req);
 2177         if (error != 0)
 2178                 return (error);
 2179         vdso_th_enable = old_vdso_th_enable;
 2180         return (0);
 2181 }
 2182 SYSCTL_PROC(_kern_timecounter, OID_AUTO, fast_gettime,
 2183     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
 2184     NULL, 0, sysctl_fast_gettime, "I", "Enable fast time of day");
 2185 
 2186 uint32_t
 2187 tc_fill_vdso_timehands(struct vdso_timehands *vdso_th)
 2188 {
 2189         struct timehands *th;
 2190         uint32_t enabled;
 2191 
 2192         th = timehands;
 2193         vdso_th->th_scale = th->th_scale;
 2194         vdso_th->th_offset_count = th->th_offset_count;
 2195         vdso_th->th_counter_mask = th->th_counter->tc_counter_mask;
 2196         vdso_th->th_offset = th->th_offset;
 2197         vdso_th->th_boottime = th->th_boottime;
 2198         if (th->th_counter->tc_fill_vdso_timehands != NULL) {
 2199                 enabled = th->th_counter->tc_fill_vdso_timehands(vdso_th,
 2200                     th->th_counter);
 2201         } else
 2202                 enabled = 0;
 2203         if (!vdso_th_enable)
 2204                 enabled = 0;
 2205         return (enabled);
 2206 }
 2207 
 2208 #ifdef COMPAT_FREEBSD32
 2209 uint32_t
 2210 tc_fill_vdso_timehands32(struct vdso_timehands32 *vdso_th32)
 2211 {
 2212         struct timehands *th;
 2213         uint32_t enabled;
 2214 
 2215         th = timehands;
 2216         *(uint64_t *)&vdso_th32->th_scale[0] = th->th_scale;
 2217         vdso_th32->th_offset_count = th->th_offset_count;
 2218         vdso_th32->th_counter_mask = th->th_counter->tc_counter_mask;
 2219         vdso_th32->th_offset.sec = th->th_offset.sec;
 2220         *(uint64_t *)&vdso_th32->th_offset.frac[0] = th->th_offset.frac;
 2221         vdso_th32->th_boottime.sec = th->th_boottime.sec;
 2222         *(uint64_t *)&vdso_th32->th_boottime.frac[0] = th->th_boottime.frac;
 2223         if (th->th_counter->tc_fill_vdso_timehands32 != NULL) {
 2224                 enabled = th->th_counter->tc_fill_vdso_timehands32(vdso_th32,
 2225                     th->th_counter);
 2226         } else
 2227                 enabled = 0;
 2228         if (!vdso_th_enable)
 2229                 enabled = 0;
 2230         return (enabled);
 2231 }
 2232 #endif
 2233 
 2234 #include "opt_ddb.h"
 2235 #ifdef DDB
 2236 #include <ddb/ddb.h>
 2237 
 2238 DB_SHOW_COMMAND(timecounter, db_show_timecounter)
 2239 {
 2240         struct timehands *th;
 2241         struct timecounter *tc;
 2242         u_int val1, val2;
 2243 
 2244         th = timehands;
 2245         tc = th->th_counter;
 2246         val1 = tc->tc_get_timecount(tc);
 2247         __compiler_membar();
 2248         val2 = tc->tc_get_timecount(tc);
 2249 
 2250         db_printf("timecounter %p %s\n", tc, tc->tc_name);
 2251         db_printf("  mask %#x freq %ju qual %d flags %#x priv %p\n",
 2252             tc->tc_counter_mask, (uintmax_t)tc->tc_frequency, tc->tc_quality,
 2253             tc->tc_flags, tc->tc_priv);
 2254         db_printf("  val %#x %#x\n", val1, val2);
 2255         db_printf("timehands adj %#jx scale %#jx ldelta %d off_cnt %d gen %d\n",
 2256             (uintmax_t)th->th_adjustment, (uintmax_t)th->th_scale,
 2257             th->th_large_delta, th->th_offset_count, th->th_generation);
 2258         db_printf("  offset %jd %jd boottime %jd %jd\n",
 2259             (intmax_t)th->th_offset.sec, (uintmax_t)th->th_offset.frac,
 2260             (intmax_t)th->th_boottime.sec, (uintmax_t)th->th_boottime.frac);
 2261 }
 2262 #endif

Cache object: bfab18346890232382d670ac3429cf4a


[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]


This page is part of the FreeBSD/Linux Linux Kernel Cross-Reference, and was automatically generated using a modified version of the LXR engine.