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

Cache object: b0408ac30029151289bd173914c0b70b


[ 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.