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  * ----------------------------------------------------------------------------
    3  * "THE BEER-WARE LICENSE" (Revision 42):
    4  * <phk@FreeBSD.ORG> wrote this file.  As long as you retain this notice you
    5  * can do whatever you want with this stuff. If we meet some day, and you think
    6  * this stuff is worth it, you can buy me a beer in return.   Poul-Henning Kamp
    7  * ----------------------------------------------------------------------------
    8  *
    9  * Copyright (c) 2011, 2015, 2016 The FreeBSD Foundation
   10  * All rights reserved.
   11  *
   12  * Portions of this software were developed by Julien Ridoux at the University
   13  * of Melbourne under sponsorship from the FreeBSD Foundation.
   14  *
   15  * Portions of this software were developed by Konstantin Belousov
   16  * under sponsorship from the FreeBSD Foundation.
   17  */
   18 
   19 #include <sys/cdefs.h>
   20 __FBSDID("$FreeBSD$");
   21 
   22 #include "opt_compat.h"
   23 #include "opt_ntp.h"
   24 #include "opt_ffclock.h"
   25 
   26 #include <sys/param.h>
   27 #include <sys/kernel.h>
   28 #include <sys/limits.h>
   29 #include <sys/lock.h>
   30 #include <sys/mutex.h>
   31 #include <sys/proc.h>
   32 #include <sys/sbuf.h>
   33 #include <sys/sleepqueue.h>
   34 #include <sys/sysctl.h>
   35 #include <sys/syslog.h>
   36 #include <sys/systm.h>
   37 #include <sys/timeffc.h>
   38 #include <sys/timepps.h>
   39 #include <sys/timetc.h>
   40 #include <sys/timex.h>
   41 #include <sys/vdso.h>
   42 
   43 /*
   44  * A large step happens on boot.  This constant detects such steps.
   45  * It is relatively small so that ntp_update_second gets called enough
   46  * in the typical 'missed a couple of seconds' case, but doesn't loop
   47  * forever when the time step is large.
   48  */
   49 #define LARGE_STEP      200
   50 
   51 /*
   52  * Implement a dummy timecounter which we can use until we get a real one
   53  * in the air.  This allows the console and other early stuff to use
   54  * time services.
   55  */
   56 
   57 static u_int
   58 dummy_get_timecount(struct timecounter *tc)
   59 {
   60         static u_int now;
   61 
   62         return (++now);
   63 }
   64 
   65 static struct timecounter dummy_timecounter = {
   66         dummy_get_timecount, 0, ~0u, 1000000, "dummy", -1000000
   67 };
   68 
   69 struct timehands {
   70         /* These fields must be initialized by the driver. */
   71         struct timecounter      *th_counter;
   72         int64_t                 th_adjustment;
   73         uint64_t                th_scale;
   74         u_int                   th_offset_count;
   75         struct bintime          th_offset;
   76         struct bintime          th_bintime;
   77         struct timeval          th_microtime;
   78         struct timespec         th_nanotime;
   79         struct bintime          th_boottime;
   80         /* Fields not to be copied in tc_windup start with th_generation. */
   81         u_int                   th_generation;
   82         struct timehands        *th_next;
   83 };
   84 
   85 static struct timehands ths[16] = {
   86     [0] =  {
   87         .th_counter = &dummy_timecounter,
   88         .th_scale = (uint64_t)-1 / 1000000,
   89         .th_offset = { .sec = 1 },
   90         .th_generation = 1,
   91     },
   92 };
   93 
   94 static struct timehands *volatile timehands = &ths[0];
   95 struct timecounter *timecounter = &dummy_timecounter;
   96 static struct timecounter *timecounters = &dummy_timecounter;
   97 
   98 int tc_min_ticktock_freq = 1;
   99 
  100 volatile time_t time_second = 1;
  101 volatile time_t time_uptime = 1;
  102 
  103 struct bintime boottimebin;
  104 struct timeval boottime;
  105 static int sysctl_kern_boottime(SYSCTL_HANDLER_ARGS);
  106 SYSCTL_PROC(_kern, KERN_BOOTTIME, boottime, CTLTYPE_STRUCT|CTLFLAG_RD,
  107     NULL, 0, sysctl_kern_boottime, "S,timeval", "System boottime");
  108 
  109 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW, 0, "");
  110 static SYSCTL_NODE(_kern_timecounter, OID_AUTO, tc, CTLFLAG_RW, 0, "");
  111 
  112 static int timestepwarnings;
  113 SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RW,
  114     &timestepwarnings, 0, "Log time steps");
  115 
  116 static int timehands_count = 2;
  117 SYSCTL_INT(_kern_timecounter, OID_AUTO, timehands_count,
  118     CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
  119     &timehands_count, 0, "Count of timehands in rotation");
  120 
  121 struct bintime bt_timethreshold;
  122 struct bintime bt_tickthreshold;
  123 sbintime_t sbt_timethreshold;
  124 sbintime_t sbt_tickthreshold;
  125 struct bintime tc_tick_bt;
  126 sbintime_t tc_tick_sbt;
  127 int tc_precexp;
  128 int tc_timepercentage = TC_DEFAULTPERC;
  129 static int sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS);
  130 SYSCTL_PROC(_kern_timecounter, OID_AUTO, alloweddeviation,
  131     CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, 0, 0,
  132     sysctl_kern_timecounter_adjprecision, "I",
  133     "Allowed time interval deviation in percents");
  134 
  135 volatile int rtc_generation = 1;
  136 
  137 static int tc_chosen;   /* Non-zero if a specific tc was chosen via sysctl. */
  138 
  139 static void tc_windup(struct bintime *new_boottimebin);
  140 static void cpu_tick_calibrate(int);
  141 
  142 void dtrace_getnanotime(struct timespec *tsp);
  143 
  144 static int
  145 sysctl_kern_boottime(SYSCTL_HANDLER_ARGS)
  146 {
  147         struct timeval boottime_x;
  148 
  149         getboottime(&boottime_x);
  150 
  151 #ifndef __mips__
  152 #ifdef SCTL_MASK32
  153         int tv[2];
  154 
  155         if (req->flags & SCTL_MASK32) {
  156                 tv[0] = boottime_x.tv_sec;
  157                 tv[1] = boottime_x.tv_usec;
  158                 return (SYSCTL_OUT(req, tv, sizeof(tv)));
  159         }
  160 #endif
  161 #endif
  162         return (SYSCTL_OUT(req, &boottime_x, sizeof(boottime_x)));
  163 }
  164 
  165 static int
  166 sysctl_kern_timecounter_get(SYSCTL_HANDLER_ARGS)
  167 {
  168         u_int ncount;
  169         struct timecounter *tc = arg1;
  170 
  171         ncount = tc->tc_get_timecount(tc);
  172         return (sysctl_handle_int(oidp, &ncount, 0, req));
  173 }
  174 
  175 static int
  176 sysctl_kern_timecounter_freq(SYSCTL_HANDLER_ARGS)
  177 {
  178         uint64_t freq;
  179         struct timecounter *tc = arg1;
  180 
  181         freq = tc->tc_frequency;
  182         return (sysctl_handle_64(oidp, &freq, 0, req));
  183 }
  184 
  185 /*
  186  * Return the difference between the timehands' counter value now and what
  187  * was when we copied it to the timehands' offset_count.
  188  */
  189 static __inline u_int
  190 tc_delta(struct timehands *th)
  191 {
  192         struct timecounter *tc;
  193 
  194         tc = th->th_counter;
  195         return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
  196             tc->tc_counter_mask);
  197 }
  198 
  199 /*
  200  * Functions for reading the time.  We have to loop until we are sure that
  201  * the timehands that we operated on was not updated under our feet.  See
  202  * the comment in <sys/time.h> for a description of these 12 functions.
  203  */
  204 
  205 #ifdef FFCLOCK
  206 void
  207 fbclock_binuptime(struct bintime *bt)
  208 {
  209         struct timehands *th;
  210         unsigned int gen;
  211 
  212         do {
  213                 th = timehands;
  214                 gen = atomic_load_acq_int(&th->th_generation);
  215                 *bt = th->th_offset;
  216                 bintime_addx(bt, th->th_scale * tc_delta(th));
  217                 atomic_thread_fence_acq();
  218         } while (gen == 0 || gen != th->th_generation);
  219 }
  220 
  221 void
  222 fbclock_nanouptime(struct timespec *tsp)
  223 {
  224         struct bintime bt;
  225 
  226         fbclock_binuptime(&bt);
  227         bintime2timespec(&bt, tsp);
  228 }
  229 
  230 void
  231 fbclock_microuptime(struct timeval *tvp)
  232 {
  233         struct bintime bt;
  234 
  235         fbclock_binuptime(&bt);
  236         bintime2timeval(&bt, tvp);
  237 }
  238 
  239 void
  240 fbclock_bintime(struct bintime *bt)
  241 {
  242         struct timehands *th;
  243         unsigned int gen;
  244 
  245         do {
  246                 th = timehands;
  247                 gen = atomic_load_acq_int(&th->th_generation);
  248                 *bt = th->th_bintime;
  249                 bintime_addx(bt, th->th_scale * tc_delta(th));
  250                 atomic_thread_fence_acq();
  251         } while (gen == 0 || gen != th->th_generation);
  252 }
  253 
  254 void
  255 fbclock_nanotime(struct timespec *tsp)
  256 {
  257         struct bintime bt;
  258 
  259         fbclock_bintime(&bt);
  260         bintime2timespec(&bt, tsp);
  261 }
  262 
  263 void
  264 fbclock_microtime(struct timeval *tvp)
  265 {
  266         struct bintime bt;
  267 
  268         fbclock_bintime(&bt);
  269         bintime2timeval(&bt, tvp);
  270 }
  271 
  272 void
  273 fbclock_getbinuptime(struct bintime *bt)
  274 {
  275         struct timehands *th;
  276         unsigned int gen;
  277 
  278         do {
  279                 th = timehands;
  280                 gen = atomic_load_acq_int(&th->th_generation);
  281                 *bt = th->th_offset;
  282                 atomic_thread_fence_acq();
  283         } while (gen == 0 || gen != th->th_generation);
  284 }
  285 
  286 void
  287 fbclock_getnanouptime(struct timespec *tsp)
  288 {
  289         struct timehands *th;
  290         unsigned int gen;
  291 
  292         do {
  293                 th = timehands;
  294                 gen = atomic_load_acq_int(&th->th_generation);
  295                 bintime2timespec(&th->th_offset, tsp);
  296                 atomic_thread_fence_acq();
  297         } while (gen == 0 || gen != th->th_generation);
  298 }
  299 
  300 void
  301 fbclock_getmicrouptime(struct timeval *tvp)
  302 {
  303         struct timehands *th;
  304         unsigned int gen;
  305 
  306         do {
  307                 th = timehands;
  308                 gen = atomic_load_acq_int(&th->th_generation);
  309                 bintime2timeval(&th->th_offset, tvp);
  310                 atomic_thread_fence_acq();
  311         } while (gen == 0 || gen != th->th_generation);
  312 }
  313 
  314 void
  315 fbclock_getbintime(struct bintime *bt)
  316 {
  317         struct timehands *th;
  318         unsigned int gen;
  319 
  320         do {
  321                 th = timehands;
  322                 gen = atomic_load_acq_int(&th->th_generation);
  323                 *bt = th->th_bintime;
  324                 atomic_thread_fence_acq();
  325         } while (gen == 0 || gen != th->th_generation);
  326 }
  327 
  328 void
  329 fbclock_getnanotime(struct timespec *tsp)
  330 {
  331         struct timehands *th;
  332         unsigned int gen;
  333 
  334         do {
  335                 th = timehands;
  336                 gen = atomic_load_acq_int(&th->th_generation);
  337                 *tsp = th->th_nanotime;
  338                 atomic_thread_fence_acq();
  339         } while (gen == 0 || gen != th->th_generation);
  340 }
  341 
  342 void
  343 fbclock_getmicrotime(struct timeval *tvp)
  344 {
  345         struct timehands *th;
  346         unsigned int gen;
  347 
  348         do {
  349                 th = timehands;
  350                 gen = atomic_load_acq_int(&th->th_generation);
  351                 *tvp = th->th_microtime;
  352                 atomic_thread_fence_acq();
  353         } while (gen == 0 || gen != th->th_generation);
  354 }
  355 #else /* !FFCLOCK */
  356 void
  357 binuptime(struct bintime *bt)
  358 {
  359         struct timehands *th;
  360         u_int gen;
  361 
  362         do {
  363                 th = timehands;
  364                 gen = atomic_load_acq_int(&th->th_generation);
  365                 *bt = th->th_offset;
  366                 bintime_addx(bt, th->th_scale * tc_delta(th));
  367                 atomic_thread_fence_acq();
  368         } while (gen == 0 || gen != th->th_generation);
  369 }
  370 
  371 void
  372 nanouptime(struct timespec *tsp)
  373 {
  374         struct bintime bt;
  375 
  376         binuptime(&bt);
  377         bintime2timespec(&bt, tsp);
  378 }
  379 
  380 void
  381 microuptime(struct timeval *tvp)
  382 {
  383         struct bintime bt;
  384 
  385         binuptime(&bt);
  386         bintime2timeval(&bt, tvp);
  387 }
  388 
  389 void
  390 bintime(struct bintime *bt)
  391 {
  392         struct timehands *th;
  393         u_int gen;
  394 
  395         do {
  396                 th = timehands;
  397                 gen = atomic_load_acq_int(&th->th_generation);
  398                 *bt = th->th_bintime;
  399                 bintime_addx(bt, th->th_scale * tc_delta(th));
  400                 atomic_thread_fence_acq();
  401         } while (gen == 0 || gen != th->th_generation);
  402 }
  403 
  404 void
  405 nanotime(struct timespec *tsp)
  406 {
  407         struct bintime bt;
  408 
  409         bintime(&bt);
  410         bintime2timespec(&bt, tsp);
  411 }
  412 
  413 void
  414 microtime(struct timeval *tvp)
  415 {
  416         struct bintime bt;
  417 
  418         bintime(&bt);
  419         bintime2timeval(&bt, tvp);
  420 }
  421 
  422 void
  423 getbinuptime(struct bintime *bt)
  424 {
  425         struct timehands *th;
  426         u_int gen;
  427 
  428         do {
  429                 th = timehands;
  430                 gen = atomic_load_acq_int(&th->th_generation);
  431                 *bt = th->th_offset;
  432                 atomic_thread_fence_acq();
  433         } while (gen == 0 || gen != th->th_generation);
  434 }
  435 
  436 void
  437 getnanouptime(struct timespec *tsp)
  438 {
  439         struct timehands *th;
  440         u_int gen;
  441 
  442         do {
  443                 th = timehands;
  444                 gen = atomic_load_acq_int(&th->th_generation);
  445                 bintime2timespec(&th->th_offset, tsp);
  446                 atomic_thread_fence_acq();
  447         } while (gen == 0 || gen != th->th_generation);
  448 }
  449 
  450 void
  451 getmicrouptime(struct timeval *tvp)
  452 {
  453         struct timehands *th;
  454         u_int gen;
  455 
  456         do {
  457                 th = timehands;
  458                 gen = atomic_load_acq_int(&th->th_generation);
  459                 bintime2timeval(&th->th_offset, tvp);
  460                 atomic_thread_fence_acq();
  461         } while (gen == 0 || gen != th->th_generation);
  462 }
  463 
  464 void
  465 getbintime(struct bintime *bt)
  466 {
  467         struct timehands *th;
  468         u_int gen;
  469 
  470         do {
  471                 th = timehands;
  472                 gen = atomic_load_acq_int(&th->th_generation);
  473                 *bt = th->th_bintime;
  474                 atomic_thread_fence_acq();
  475         } while (gen == 0 || gen != th->th_generation);
  476 }
  477 
  478 void
  479 getnanotime(struct timespec *tsp)
  480 {
  481         struct timehands *th;
  482         u_int gen;
  483 
  484         do {
  485                 th = timehands;
  486                 gen = atomic_load_acq_int(&th->th_generation);
  487                 *tsp = th->th_nanotime;
  488                 atomic_thread_fence_acq();
  489         } while (gen == 0 || gen != th->th_generation);
  490 }
  491 
  492 void
  493 getmicrotime(struct timeval *tvp)
  494 {
  495         struct timehands *th;
  496         u_int gen;
  497 
  498         do {
  499                 th = timehands;
  500                 gen = atomic_load_acq_int(&th->th_generation);
  501                 *tvp = th->th_microtime;
  502                 atomic_thread_fence_acq();
  503         } while (gen == 0 || gen != th->th_generation);
  504 }
  505 #endif /* FFCLOCK */
  506 
  507 void
  508 getboottime(struct timeval *boottime_x)
  509 {
  510         struct bintime boottimebin_x;
  511 
  512         getboottimebin(&boottimebin_x);
  513         bintime2timeval(&boottimebin_x, boottime_x);
  514 }
  515 
  516 void
  517 getboottimebin(struct bintime *boottimebin_x)
  518 {
  519         struct timehands *th;
  520         u_int gen;
  521 
  522         do {
  523                 th = timehands;
  524                 gen = atomic_load_acq_int(&th->th_generation);
  525                 *boottimebin_x = th->th_boottime;
  526                 atomic_thread_fence_acq();
  527         } while (gen == 0 || gen != th->th_generation);
  528 }
  529 
  530 #ifdef FFCLOCK
  531 /*
  532  * Support for feed-forward synchronization algorithms. This is heavily inspired
  533  * by the timehands mechanism but kept independent from it. *_windup() functions
  534  * have some connection to avoid accessing the timecounter hardware more than
  535  * necessary.
  536  */
  537 
  538 /* Feed-forward clock estimates kept updated by the synchronization daemon. */
  539 struct ffclock_estimate ffclock_estimate;
  540 struct bintime ffclock_boottime;        /* Feed-forward boot time estimate. */
  541 uint32_t ffclock_status;                /* Feed-forward clock status. */
  542 int8_t ffclock_updated;                 /* New estimates are available. */
  543 struct mtx ffclock_mtx;                 /* Mutex on ffclock_estimate. */
  544 
  545 struct fftimehands {
  546         struct ffclock_estimate cest;
  547         struct bintime          tick_time;
  548         struct bintime          tick_time_lerp;
  549         ffcounter               tick_ffcount;
  550         uint64_t                period_lerp;
  551         volatile uint8_t        gen;
  552         struct fftimehands      *next;
  553 };
  554 
  555 #define NUM_ELEMENTS(x) (sizeof(x) / sizeof(*x))
  556 
  557 static struct fftimehands ffth[10];
  558 static struct fftimehands *volatile fftimehands = ffth;
  559 
  560 static void
  561 ffclock_init(void)
  562 {
  563         struct fftimehands *cur;
  564         struct fftimehands *last;
  565 
  566         memset(ffth, 0, sizeof(ffth));
  567 
  568         last = ffth + NUM_ELEMENTS(ffth) - 1;
  569         for (cur = ffth; cur < last; cur++)
  570                 cur->next = cur + 1;
  571         last->next = ffth;
  572 
  573         ffclock_updated = 0;
  574         ffclock_status = FFCLOCK_STA_UNSYNC;
  575         mtx_init(&ffclock_mtx, "ffclock lock", NULL, MTX_DEF);
  576 }
  577 
  578 /*
  579  * Reset the feed-forward clock estimates. Called from inittodr() to get things
  580  * kick started and uses the timecounter nominal frequency as a first period
  581  * estimate. Note: this function may be called several time just after boot.
  582  * Note: this is the only function that sets the value of boot time for the
  583  * monotonic (i.e. uptime) version of the feed-forward clock.
  584  */
  585 void
  586 ffclock_reset_clock(struct timespec *ts)
  587 {
  588         struct timecounter *tc;
  589         struct ffclock_estimate cest;
  590 
  591         tc = timehands->th_counter;
  592         memset(&cest, 0, sizeof(struct ffclock_estimate));
  593 
  594         timespec2bintime(ts, &ffclock_boottime);
  595         timespec2bintime(ts, &(cest.update_time));
  596         ffclock_read_counter(&cest.update_ffcount);
  597         cest.leapsec_next = 0;
  598         cest.period = ((1ULL << 63) / tc->tc_frequency) << 1;
  599         cest.errb_abs = 0;
  600         cest.errb_rate = 0;
  601         cest.status = FFCLOCK_STA_UNSYNC;
  602         cest.leapsec_total = 0;
  603         cest.leapsec = 0;
  604 
  605         mtx_lock(&ffclock_mtx);
  606         bcopy(&cest, &ffclock_estimate, sizeof(struct ffclock_estimate));
  607         ffclock_updated = INT8_MAX;
  608         mtx_unlock(&ffclock_mtx);
  609 
  610         printf("ffclock reset: %s (%llu Hz), time = %ld.%09lu\n", tc->tc_name,
  611             (unsigned long long)tc->tc_frequency, (long)ts->tv_sec,
  612             (unsigned long)ts->tv_nsec);
  613 }
  614 
  615 /*
  616  * Sub-routine to convert a time interval measured in RAW counter units to time
  617  * in seconds stored in bintime format.
  618  * NOTE: bintime_mul requires u_int, but the value of the ffcounter may be
  619  * larger than the max value of u_int (on 32 bit architecture). Loop to consume
  620  * extra cycles.
  621  */
  622 static void
  623 ffclock_convert_delta(ffcounter ffdelta, uint64_t period, struct bintime *bt)
  624 {
  625         struct bintime bt2;
  626         ffcounter delta, delta_max;
  627 
  628         delta_max = (1ULL << (8 * sizeof(unsigned int))) - 1;
  629         bintime_clear(bt);
  630         do {
  631                 if (ffdelta > delta_max)
  632                         delta = delta_max;
  633                 else
  634                         delta = ffdelta;
  635                 bt2.sec = 0;
  636                 bt2.frac = period;
  637                 bintime_mul(&bt2, (unsigned int)delta);
  638                 bintime_add(bt, &bt2);
  639                 ffdelta -= delta;
  640         } while (ffdelta > 0);
  641 }
  642 
  643 /*
  644  * Update the fftimehands.
  645  * Push the tick ffcount and time(s) forward based on current clock estimate.
  646  * The conversion from ffcounter to bintime relies on the difference clock
  647  * principle, whose accuracy relies on computing small time intervals. If a new
  648  * clock estimate has been passed by the synchronisation daemon, make it
  649  * current, and compute the linear interpolation for monotonic time if needed.
  650  */
  651 static void
  652 ffclock_windup(unsigned int delta)
  653 {
  654         struct ffclock_estimate *cest;
  655         struct fftimehands *ffth;
  656         struct bintime bt, gap_lerp;
  657         ffcounter ffdelta;
  658         uint64_t frac;
  659         unsigned int polling;
  660         uint8_t forward_jump, ogen;
  661 
  662         /*
  663          * Pick the next timehand, copy current ffclock estimates and move tick
  664          * times and counter forward.
  665          */
  666         forward_jump = 0;
  667         ffth = fftimehands->next;
  668         ogen = ffth->gen;
  669         ffth->gen = 0;
  670         cest = &ffth->cest;
  671         bcopy(&fftimehands->cest, cest, sizeof(struct ffclock_estimate));
  672         ffdelta = (ffcounter)delta;
  673         ffth->period_lerp = fftimehands->period_lerp;
  674 
  675         ffth->tick_time = fftimehands->tick_time;
  676         ffclock_convert_delta(ffdelta, cest->period, &bt);
  677         bintime_add(&ffth->tick_time, &bt);
  678 
  679         ffth->tick_time_lerp = fftimehands->tick_time_lerp;
  680         ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt);
  681         bintime_add(&ffth->tick_time_lerp, &bt);
  682 
  683         ffth->tick_ffcount = fftimehands->tick_ffcount + ffdelta;
  684 
  685         /*
  686          * Assess the status of the clock, if the last update is too old, it is
  687          * likely the synchronisation daemon is dead and the clock is free
  688          * running.
  689          */
  690         if (ffclock_updated == 0) {
  691                 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
  692                 ffclock_convert_delta(ffdelta, cest->period, &bt);
  693                 if (bt.sec > 2 * FFCLOCK_SKM_SCALE)
  694                         ffclock_status |= FFCLOCK_STA_UNSYNC;
  695         }
  696 
  697         /*
  698          * If available, grab updated clock estimates and make them current.
  699          * Recompute time at this tick using the updated estimates. The clock
  700          * estimates passed the feed-forward synchronisation daemon may result
  701          * in time conversion that is not monotonically increasing (just after
  702          * the update). time_lerp is a particular linear interpolation over the
  703          * synchronisation algo polling period that ensures monotonicity for the
  704          * clock ids requesting it.
  705          */
  706         if (ffclock_updated > 0) {
  707                 bcopy(&ffclock_estimate, cest, sizeof(struct ffclock_estimate));
  708                 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
  709                 ffth->tick_time = cest->update_time;
  710                 ffclock_convert_delta(ffdelta, cest->period, &bt);
  711                 bintime_add(&ffth->tick_time, &bt);
  712 
  713                 /* ffclock_reset sets ffclock_updated to INT8_MAX */
  714                 if (ffclock_updated == INT8_MAX)
  715                         ffth->tick_time_lerp = ffth->tick_time;
  716 
  717                 if (bintime_cmp(&ffth->tick_time, &ffth->tick_time_lerp, >))
  718                         forward_jump = 1;
  719                 else
  720                         forward_jump = 0;
  721 
  722                 bintime_clear(&gap_lerp);
  723                 if (forward_jump) {
  724                         gap_lerp = ffth->tick_time;
  725                         bintime_sub(&gap_lerp, &ffth->tick_time_lerp);
  726                 } else {
  727                         gap_lerp = ffth->tick_time_lerp;
  728                         bintime_sub(&gap_lerp, &ffth->tick_time);
  729                 }
  730 
  731                 /*
  732                  * The reset from the RTC clock may be far from accurate, and
  733                  * reducing the gap between real time and interpolated time
  734                  * could take a very long time if the interpolated clock insists
  735                  * on strict monotonicity. The clock is reset under very strict
  736                  * conditions (kernel time is known to be wrong and
  737                  * synchronization daemon has been restarted recently.
  738                  * ffclock_boottime absorbs the jump to ensure boot time is
  739                  * correct and uptime functions stay consistent.
  740                  */
  741                 if (((ffclock_status & FFCLOCK_STA_UNSYNC) == FFCLOCK_STA_UNSYNC) &&
  742                     ((cest->status & FFCLOCK_STA_UNSYNC) == 0) &&
  743                     ((cest->status & FFCLOCK_STA_WARMUP) == FFCLOCK_STA_WARMUP)) {
  744                         if (forward_jump)
  745                                 bintime_add(&ffclock_boottime, &gap_lerp);
  746                         else
  747                                 bintime_sub(&ffclock_boottime, &gap_lerp);
  748                         ffth->tick_time_lerp = ffth->tick_time;
  749                         bintime_clear(&gap_lerp);
  750                 }
  751 
  752                 ffclock_status = cest->status;
  753                 ffth->period_lerp = cest->period;
  754 
  755                 /*
  756                  * Compute corrected period used for the linear interpolation of
  757                  * time. The rate of linear interpolation is capped to 5000PPM
  758                  * (5ms/s).
  759                  */
  760                 if (bintime_isset(&gap_lerp)) {
  761                         ffdelta = cest->update_ffcount;
  762                         ffdelta -= fftimehands->cest.update_ffcount;
  763                         ffclock_convert_delta(ffdelta, cest->period, &bt);
  764                         polling = bt.sec;
  765                         bt.sec = 0;
  766                         bt.frac = 5000000 * (uint64_t)18446744073LL;
  767                         bintime_mul(&bt, polling);
  768                         if (bintime_cmp(&gap_lerp, &bt, >))
  769                                 gap_lerp = bt;
  770 
  771                         /* Approximate 1 sec by 1-(1/2^64) to ease arithmetic */
  772                         frac = 0;
  773                         if (gap_lerp.sec > 0) {
  774                                 frac -= 1;
  775                                 frac /= ffdelta / gap_lerp.sec;
  776                         }
  777                         frac += gap_lerp.frac / ffdelta;
  778 
  779                         if (forward_jump)
  780                                 ffth->period_lerp += frac;
  781                         else
  782                                 ffth->period_lerp -= frac;
  783                 }
  784 
  785                 ffclock_updated = 0;
  786         }
  787         if (++ogen == 0)
  788                 ogen = 1;
  789         ffth->gen = ogen;
  790         fftimehands = ffth;
  791 }
  792 
  793 /*
  794  * Adjust the fftimehands when the timecounter is changed. Stating the obvious,
  795  * the old and new hardware counter cannot be read simultaneously. tc_windup()
  796  * does read the two counters 'back to back', but a few cycles are effectively
  797  * lost, and not accumulated in tick_ffcount. This is a fairly radical
  798  * operation for a feed-forward synchronization daemon, and it is its job to not
  799  * pushing irrelevant data to the kernel. Because there is no locking here,
  800  * simply force to ignore pending or next update to give daemon a chance to
  801  * realize the counter has changed.
  802  */
  803 static void
  804 ffclock_change_tc(struct timehands *th)
  805 {
  806         struct fftimehands *ffth;
  807         struct ffclock_estimate *cest;
  808         struct timecounter *tc;
  809         uint8_t ogen;
  810 
  811         tc = th->th_counter;
  812         ffth = fftimehands->next;
  813         ogen = ffth->gen;
  814         ffth->gen = 0;
  815 
  816         cest = &ffth->cest;
  817         bcopy(&(fftimehands->cest), cest, sizeof(struct ffclock_estimate));
  818         cest->period = ((1ULL << 63) / tc->tc_frequency ) << 1;
  819         cest->errb_abs = 0;
  820         cest->errb_rate = 0;
  821         cest->status |= FFCLOCK_STA_UNSYNC;
  822 
  823         ffth->tick_ffcount = fftimehands->tick_ffcount;
  824         ffth->tick_time_lerp = fftimehands->tick_time_lerp;
  825         ffth->tick_time = fftimehands->tick_time;
  826         ffth->period_lerp = cest->period;
  827 
  828         /* Do not lock but ignore next update from synchronization daemon. */
  829         ffclock_updated--;
  830 
  831         if (++ogen == 0)
  832                 ogen = 1;
  833         ffth->gen = ogen;
  834         fftimehands = ffth;
  835 }
  836 
  837 /*
  838  * Retrieve feed-forward counter and time of last kernel tick.
  839  */
  840 void
  841 ffclock_last_tick(ffcounter *ffcount, struct bintime *bt, uint32_t flags)
  842 {
  843         struct fftimehands *ffth;
  844         uint8_t gen;
  845 
  846         /*
  847          * No locking but check generation has not changed. Also need to make
  848          * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
  849          */
  850         do {
  851                 ffth = fftimehands;
  852                 gen = ffth->gen;
  853                 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP)
  854                         *bt = ffth->tick_time_lerp;
  855                 else
  856                         *bt = ffth->tick_time;
  857                 *ffcount = ffth->tick_ffcount;
  858         } while (gen == 0 || gen != ffth->gen);
  859 }
  860 
  861 /*
  862  * Absolute clock conversion. Low level function to convert ffcounter to
  863  * bintime. The ffcounter is converted using the current ffclock period estimate
  864  * or the "interpolated period" to ensure monotonicity.
  865  * NOTE: this conversion may have been deferred, and the clock updated since the
  866  * hardware counter has been read.
  867  */
  868 void
  869 ffclock_convert_abs(ffcounter ffcount, struct bintime *bt, uint32_t flags)
  870 {
  871         struct fftimehands *ffth;
  872         struct bintime bt2;
  873         ffcounter ffdelta;
  874         uint8_t gen;
  875 
  876         /*
  877          * No locking but check generation has not changed. Also need to make
  878          * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
  879          */
  880         do {
  881                 ffth = fftimehands;
  882                 gen = ffth->gen;
  883                 if (ffcount > ffth->tick_ffcount)
  884                         ffdelta = ffcount - ffth->tick_ffcount;
  885                 else
  886                         ffdelta = ffth->tick_ffcount - ffcount;
  887 
  888                 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP) {
  889                         *bt = ffth->tick_time_lerp;
  890                         ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt2);
  891                 } else {
  892                         *bt = ffth->tick_time;
  893                         ffclock_convert_delta(ffdelta, ffth->cest.period, &bt2);
  894                 }
  895 
  896                 if (ffcount > ffth->tick_ffcount)
  897                         bintime_add(bt, &bt2);
  898                 else
  899                         bintime_sub(bt, &bt2);
  900         } while (gen == 0 || gen != ffth->gen);
  901 }
  902 
  903 /*
  904  * Difference clock conversion.
  905  * Low level function to Convert a time interval measured in RAW counter units
  906  * into bintime. The difference clock allows measuring small intervals much more
  907  * reliably than the absolute clock.
  908  */
  909 void
  910 ffclock_convert_diff(ffcounter ffdelta, struct bintime *bt)
  911 {
  912         struct fftimehands *ffth;
  913         uint8_t gen;
  914 
  915         /* No locking but check generation has not changed. */
  916         do {
  917                 ffth = fftimehands;
  918                 gen = ffth->gen;
  919                 ffclock_convert_delta(ffdelta, ffth->cest.period, bt);
  920         } while (gen == 0 || gen != ffth->gen);
  921 }
  922 
  923 /*
  924  * Access to current ffcounter value.
  925  */
  926 void
  927 ffclock_read_counter(ffcounter *ffcount)
  928 {
  929         struct timehands *th;
  930         struct fftimehands *ffth;
  931         unsigned int gen, delta;
  932 
  933         /*
  934          * ffclock_windup() called from tc_windup(), safe to rely on
  935          * th->th_generation only, for correct delta and ffcounter.
  936          */
  937         do {
  938                 th = timehands;
  939                 gen = atomic_load_acq_int(&th->th_generation);
  940                 ffth = fftimehands;
  941                 delta = tc_delta(th);
  942                 *ffcount = ffth->tick_ffcount;
  943                 atomic_thread_fence_acq();
  944         } while (gen == 0 || gen != th->th_generation);
  945 
  946         *ffcount += delta;
  947 }
  948 
  949 void
  950 binuptime(struct bintime *bt)
  951 {
  952 
  953         binuptime_fromclock(bt, sysclock_active);
  954 }
  955 
  956 void
  957 nanouptime(struct timespec *tsp)
  958 {
  959 
  960         nanouptime_fromclock(tsp, sysclock_active);
  961 }
  962 
  963 void
  964 microuptime(struct timeval *tvp)
  965 {
  966 
  967         microuptime_fromclock(tvp, sysclock_active);
  968 }
  969 
  970 void
  971 bintime(struct bintime *bt)
  972 {
  973 
  974         bintime_fromclock(bt, sysclock_active);
  975 }
  976 
  977 void
  978 nanotime(struct timespec *tsp)
  979 {
  980 
  981         nanotime_fromclock(tsp, sysclock_active);
  982 }
  983 
  984 void
  985 microtime(struct timeval *tvp)
  986 {
  987 
  988         microtime_fromclock(tvp, sysclock_active);
  989 }
  990 
  991 void
  992 getbinuptime(struct bintime *bt)
  993 {
  994 
  995         getbinuptime_fromclock(bt, sysclock_active);
  996 }
  997 
  998 void
  999 getnanouptime(struct timespec *tsp)
 1000 {
 1001 
 1002         getnanouptime_fromclock(tsp, sysclock_active);
 1003 }
 1004 
 1005 void
 1006 getmicrouptime(struct timeval *tvp)
 1007 {
 1008 
 1009         getmicrouptime_fromclock(tvp, sysclock_active);
 1010 }
 1011 
 1012 void
 1013 getbintime(struct bintime *bt)
 1014 {
 1015 
 1016         getbintime_fromclock(bt, sysclock_active);
 1017 }
 1018 
 1019 void
 1020 getnanotime(struct timespec *tsp)
 1021 {
 1022 
 1023         getnanotime_fromclock(tsp, sysclock_active);
 1024 }
 1025 
 1026 void
 1027 getmicrotime(struct timeval *tvp)
 1028 {
 1029 
 1030         getmicrouptime_fromclock(tvp, sysclock_active);
 1031 }
 1032 
 1033 #endif /* FFCLOCK */
 1034 
 1035 /*
 1036  * This is a clone of getnanotime and used for walltimestamps.
 1037  * The dtrace_ prefix prevents fbt from creating probes for
 1038  * it so walltimestamp can be safely used in all fbt probes.
 1039  */
 1040 void
 1041 dtrace_getnanotime(struct timespec *tsp)
 1042 {
 1043         struct timehands *th;
 1044         u_int gen;
 1045 
 1046         do {
 1047                 th = timehands;
 1048                 gen = atomic_load_acq_int(&th->th_generation);
 1049                 *tsp = th->th_nanotime;
 1050                 atomic_thread_fence_acq();
 1051         } while (gen == 0 || gen != th->th_generation);
 1052 }
 1053 
 1054 /*
 1055  * System clock currently providing time to the system. Modifiable via sysctl
 1056  * when the FFCLOCK option is defined.
 1057  */
 1058 int sysclock_active = SYSCLOCK_FBCK;
 1059 
 1060 /* Internal NTP status and error estimates. */
 1061 extern int time_status;
 1062 extern long time_esterror;
 1063 
 1064 /*
 1065  * Take a snapshot of sysclock data which can be used to compare system clocks
 1066  * and generate timestamps after the fact.
 1067  */
 1068 void
 1069 sysclock_getsnapshot(struct sysclock_snap *clock_snap, int fast)
 1070 {
 1071         struct fbclock_info *fbi;
 1072         struct timehands *th;
 1073         struct bintime bt;
 1074         unsigned int delta, gen;
 1075 #ifdef FFCLOCK
 1076         ffcounter ffcount;
 1077         struct fftimehands *ffth;
 1078         struct ffclock_info *ffi;
 1079         struct ffclock_estimate cest;
 1080 
 1081         ffi = &clock_snap->ff_info;
 1082 #endif
 1083 
 1084         fbi = &clock_snap->fb_info;
 1085         delta = 0;
 1086 
 1087         do {
 1088                 th = timehands;
 1089                 gen = atomic_load_acq_int(&th->th_generation);
 1090                 fbi->th_scale = th->th_scale;
 1091                 fbi->tick_time = th->th_offset;
 1092 #ifdef FFCLOCK
 1093                 ffth = fftimehands;
 1094                 ffi->tick_time = ffth->tick_time_lerp;
 1095                 ffi->tick_time_lerp = ffth->tick_time_lerp;
 1096                 ffi->period = ffth->cest.period;
 1097                 ffi->period_lerp = ffth->period_lerp;
 1098                 clock_snap->ffcount = ffth->tick_ffcount;
 1099                 cest = ffth->cest;
 1100 #endif
 1101                 if (!fast)
 1102                         delta = tc_delta(th);
 1103                 atomic_thread_fence_acq();
 1104         } while (gen == 0 || gen != th->th_generation);
 1105 
 1106         clock_snap->delta = delta;
 1107         clock_snap->sysclock_active = sysclock_active;
 1108 
 1109         /* Record feedback clock status and error. */
 1110         clock_snap->fb_info.status = time_status;
 1111         /* XXX: Very crude estimate of feedback clock error. */
 1112         bt.sec = time_esterror / 1000000;
 1113         bt.frac = ((time_esterror - bt.sec) * 1000000) *
 1114             (uint64_t)18446744073709ULL;
 1115         clock_snap->fb_info.error = bt;
 1116 
 1117 #ifdef FFCLOCK
 1118         if (!fast)
 1119                 clock_snap->ffcount += delta;
 1120 
 1121         /* Record feed-forward clock leap second adjustment. */
 1122         ffi->leapsec_adjustment = cest.leapsec_total;
 1123         if (clock_snap->ffcount > cest.leapsec_next)
 1124                 ffi->leapsec_adjustment -= cest.leapsec;
 1125 
 1126         /* Record feed-forward clock status and error. */
 1127         clock_snap->ff_info.status = cest.status;
 1128         ffcount = clock_snap->ffcount - cest.update_ffcount;
 1129         ffclock_convert_delta(ffcount, cest.period, &bt);
 1130         /* 18446744073709 = int(2^64/1e12), err_bound_rate in [ps/s]. */
 1131         bintime_mul(&bt, cest.errb_rate * (uint64_t)18446744073709ULL);
 1132         /* 18446744073 = int(2^64 / 1e9), since err_abs in [ns]. */
 1133         bintime_addx(&bt, cest.errb_abs * (uint64_t)18446744073ULL);
 1134         clock_snap->ff_info.error = bt;
 1135 #endif
 1136 }
 1137 
 1138 /*
 1139  * Convert a sysclock snapshot into a struct bintime based on the specified
 1140  * clock source and flags.
 1141  */
 1142 int
 1143 sysclock_snap2bintime(struct sysclock_snap *cs, struct bintime *bt,
 1144     int whichclock, uint32_t flags)
 1145 {
 1146         struct bintime boottimebin_x;
 1147 #ifdef FFCLOCK
 1148         struct bintime bt2;
 1149         uint64_t period;
 1150 #endif
 1151 
 1152         switch (whichclock) {
 1153         case SYSCLOCK_FBCK:
 1154                 *bt = cs->fb_info.tick_time;
 1155 
 1156                 /* If snapshot was created with !fast, delta will be >0. */
 1157                 if (cs->delta > 0)
 1158                         bintime_addx(bt, cs->fb_info.th_scale * cs->delta);
 1159 
 1160                 if ((flags & FBCLOCK_UPTIME) == 0) {
 1161                         getboottimebin(&boottimebin_x);
 1162                         bintime_add(bt, &boottimebin_x);
 1163                 }
 1164                 break;
 1165 #ifdef FFCLOCK
 1166         case SYSCLOCK_FFWD:
 1167                 if (flags & FFCLOCK_LERP) {
 1168                         *bt = cs->ff_info.tick_time_lerp;
 1169                         period = cs->ff_info.period_lerp;
 1170                 } else {
 1171                         *bt = cs->ff_info.tick_time;
 1172                         period = cs->ff_info.period;
 1173                 }
 1174 
 1175                 /* If snapshot was created with !fast, delta will be >0. */
 1176                 if (cs->delta > 0) {
 1177                         ffclock_convert_delta(cs->delta, period, &bt2);
 1178                         bintime_add(bt, &bt2);
 1179                 }
 1180 
 1181                 /* Leap second adjustment. */
 1182                 if (flags & FFCLOCK_LEAPSEC)
 1183                         bt->sec -= cs->ff_info.leapsec_adjustment;
 1184 
 1185                 /* Boot time adjustment, for uptime/monotonic clocks. */
 1186                 if (flags & FFCLOCK_UPTIME)
 1187                         bintime_sub(bt, &ffclock_boottime);
 1188                 break;
 1189 #endif
 1190         default:
 1191                 return (EINVAL);
 1192                 break;
 1193         }
 1194 
 1195         return (0);
 1196 }
 1197 
 1198 /*
 1199  * Initialize a new timecounter and possibly use it.
 1200  */
 1201 void
 1202 tc_init(struct timecounter *tc)
 1203 {
 1204         u_int u;
 1205         struct sysctl_oid *tc_root;
 1206 
 1207         u = tc->tc_frequency / tc->tc_counter_mask;
 1208         /* XXX: We need some margin here, 10% is a guess */
 1209         u *= 11;
 1210         u /= 10;
 1211         if (u > hz && tc->tc_quality >= 0) {
 1212                 tc->tc_quality = -2000;
 1213                 if (bootverbose) {
 1214                         printf("Timecounter \"%s\" frequency %ju Hz",
 1215                             tc->tc_name, (uintmax_t)tc->tc_frequency);
 1216                         printf(" -- Insufficient hz, needs at least %u\n", u);
 1217                 }
 1218         } else if (tc->tc_quality >= 0 || bootverbose) {
 1219                 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
 1220                     tc->tc_name, (uintmax_t)tc->tc_frequency,
 1221                     tc->tc_quality);
 1222         }
 1223 
 1224         tc->tc_next = timecounters;
 1225         timecounters = tc;
 1226         /*
 1227          * Set up sysctl tree for this counter.
 1228          */
 1229         tc_root = SYSCTL_ADD_NODE(NULL,
 1230             SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name,
 1231             CTLFLAG_RW, 0, "timecounter description");
 1232         SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 1233             "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0,
 1234             "mask for implemented bits");
 1235         SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 1236             "counter", CTLTYPE_UINT | CTLFLAG_RD, tc, sizeof(*tc),
 1237             sysctl_kern_timecounter_get, "IU", "current timecounter value");
 1238         SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 1239             "frequency", CTLTYPE_U64 | CTLFLAG_RD, tc, sizeof(*tc),
 1240              sysctl_kern_timecounter_freq, "QU", "timecounter frequency");
 1241         SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
 1242             "quality", CTLFLAG_RD, &(tc->tc_quality), 0,
 1243             "goodness of time counter");
 1244         /*
 1245          * Do not automatically switch if the current tc was specifically
 1246          * chosen.  Never automatically use a timecounter with negative quality.
 1247          * Even though we run on the dummy counter, switching here may be
 1248          * worse since this timecounter may not be monotonic.
 1249          */
 1250         if (tc_chosen)
 1251                 return;
 1252         if (tc->tc_quality < 0)
 1253                 return;
 1254         if (tc->tc_quality < timecounter->tc_quality)
 1255                 return;
 1256         if (tc->tc_quality == timecounter->tc_quality &&
 1257             tc->tc_frequency < timecounter->tc_frequency)
 1258                 return;
 1259         (void)tc->tc_get_timecount(tc);
 1260         (void)tc->tc_get_timecount(tc);
 1261         timecounter = tc;
 1262 }
 1263 
 1264 /* Report the frequency of the current timecounter. */
 1265 uint64_t
 1266 tc_getfrequency(void)
 1267 {
 1268 
 1269         return (timehands->th_counter->tc_frequency);
 1270 }
 1271 
 1272 static bool
 1273 sleeping_on_old_rtc(struct thread *td)
 1274 {
 1275 
 1276         /*
 1277          * td_rtcgen is modified by curthread when it is running,
 1278          * and by other threads in this function.  By finding the thread
 1279          * on a sleepqueue and holding the lock on the sleepqueue
 1280          * chain, we guarantee that the thread is not running and that
 1281          * modifying td_rtcgen is safe.  Setting td_rtcgen to zero informs
 1282          * the thread that it was woken due to a real-time clock adjustment.
 1283          * (The declaration of td_rtcgen refers to this comment.)
 1284          */
 1285         if (td->td_rtcgen != 0 && td->td_rtcgen != rtc_generation) {
 1286                 td->td_rtcgen = 0;
 1287                 return (true);
 1288         }
 1289         return (false);
 1290 }
 1291 
 1292 static struct mtx tc_setclock_mtx;
 1293 MTX_SYSINIT(tc_setclock_init, &tc_setclock_mtx, "tcsetc", MTX_SPIN);
 1294 
 1295 /*
 1296  * Step our concept of UTC.  This is done by modifying our estimate of
 1297  * when we booted.
 1298  */
 1299 void
 1300 tc_setclock(struct timespec *ts)
 1301 {
 1302         struct timespec tbef, taft;
 1303         struct bintime bt, bt2;
 1304 
 1305         timespec2bintime(ts, &bt);
 1306         nanotime(&tbef);
 1307         mtx_lock_spin(&tc_setclock_mtx);
 1308         cpu_tick_calibrate(1);
 1309         binuptime(&bt2);
 1310         bintime_sub(&bt, &bt2);
 1311 
 1312         /* XXX fiddle all the little crinkly bits around the fiords... */
 1313         tc_windup(&bt);
 1314         mtx_unlock_spin(&tc_setclock_mtx);
 1315         getboottimebin(&boottimebin);
 1316         bintime2timeval(&boottimebin, &boottime);
 1317 
 1318         /* Avoid rtc_generation == 0, since td_rtcgen == 0 is special. */
 1319         atomic_add_rel_int(&rtc_generation, 2);
 1320         sleepq_chains_remove_matching(sleeping_on_old_rtc);
 1321         if (timestepwarnings) {
 1322                 nanotime(&taft);
 1323                 log(LOG_INFO,
 1324                     "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
 1325                     (intmax_t)tbef.tv_sec, tbef.tv_nsec,
 1326                     (intmax_t)taft.tv_sec, taft.tv_nsec,
 1327                     (intmax_t)ts->tv_sec, ts->tv_nsec);
 1328         }
 1329 }
 1330 
 1331 /*
 1332  * Initialize the next struct timehands in the ring and make
 1333  * it the active timehands.  Along the way we might switch to a different
 1334  * timecounter and/or do seconds processing in NTP.  Slightly magic.
 1335  */
 1336 static void
 1337 tc_windup(struct bintime *new_boottimebin)
 1338 {
 1339         struct bintime bt;
 1340         struct timehands *th, *tho;
 1341         uint64_t scale;
 1342         u_int delta, ncount, ogen;
 1343         int i;
 1344         time_t t;
 1345 
 1346         /*
 1347          * Make the next timehands a copy of the current one, but do
 1348          * not overwrite the generation or next pointer.  While we
 1349          * update the contents, the generation must be zero.  We need
 1350          * to ensure that the zero generation is visible before the
 1351          * data updates become visible, which requires release fence.
 1352          * For similar reasons, re-reading of the generation after the
 1353          * data is read should use acquire fence.
 1354          */
 1355         tho = timehands;
 1356         th = tho->th_next;
 1357         ogen = th->th_generation;
 1358         th->th_generation = 0;
 1359         atomic_thread_fence_rel();
 1360         bcopy(tho, th, offsetof(struct timehands, th_generation));
 1361         if (new_boottimebin != NULL)
 1362                 th->th_boottime = *new_boottimebin;
 1363 
 1364         /*
 1365          * Capture a timecounter delta on the current timecounter and if
 1366          * changing timecounters, a counter value from the new timecounter.
 1367          * Update the offset fields accordingly.
 1368          */
 1369         delta = tc_delta(th);
 1370         if (th->th_counter != timecounter)
 1371                 ncount = timecounter->tc_get_timecount(timecounter);
 1372         else
 1373                 ncount = 0;
 1374 #ifdef FFCLOCK
 1375         ffclock_windup(delta);
 1376 #endif
 1377         th->th_offset_count += delta;
 1378         th->th_offset_count &= th->th_counter->tc_counter_mask;
 1379         while (delta > th->th_counter->tc_frequency) {
 1380                 /* Eat complete unadjusted seconds. */
 1381                 delta -= th->th_counter->tc_frequency;
 1382                 th->th_offset.sec++;
 1383         }
 1384         if ((delta > th->th_counter->tc_frequency / 2) &&
 1385             (th->th_scale * delta < ((uint64_t)1 << 63))) {
 1386                 /* The product th_scale * delta just barely overflows. */
 1387                 th->th_offset.sec++;
 1388         }
 1389         bintime_addx(&th->th_offset, th->th_scale * delta);
 1390 
 1391         /*
 1392          * Hardware latching timecounters may not generate interrupts on
 1393          * PPS events, so instead we poll them.  There is a finite risk that
 1394          * the hardware might capture a count which is later than the one we
 1395          * got above, and therefore possibly in the next NTP second which might
 1396          * have a different rate than the current NTP second.  It doesn't
 1397          * matter in practice.
 1398          */
 1399         if (tho->th_counter->tc_poll_pps)
 1400                 tho->th_counter->tc_poll_pps(tho->th_counter);
 1401 
 1402         /*
 1403          * Deal with NTP second processing.  The for loop normally
 1404          * iterates at most once, but in extreme situations it might
 1405          * keep NTP sane if timeouts are not run for several seconds.
 1406          * At boot, the time step can be large when the TOD hardware
 1407          * has been read, so on really large steps, we call
 1408          * ntp_update_second only twice.  We need to call it twice in
 1409          * case we missed a leap second.
 1410          */
 1411         bt = th->th_offset;
 1412         bintime_add(&bt, &th->th_boottime);
 1413         i = bt.sec - tho->th_microtime.tv_sec;
 1414         if (i > LARGE_STEP)
 1415                 i = 2;
 1416         for (; i > 0; i--) {
 1417                 t = bt.sec;
 1418                 ntp_update_second(&th->th_adjustment, &bt.sec);
 1419                 if (bt.sec != t)
 1420                         th->th_boottime.sec += bt.sec - t;
 1421         }
 1422         /* Update the UTC timestamps used by the get*() functions. */
 1423         th->th_bintime = bt;
 1424         bintime2timeval(&bt, &th->th_microtime);
 1425         bintime2timespec(&bt, &th->th_nanotime);
 1426 
 1427         /* Now is a good time to change timecounters. */
 1428         if (th->th_counter != timecounter) {
 1429 #ifndef __arm__
 1430                 if ((timecounter->tc_flags & TC_FLAGS_C2STOP) != 0)
 1431                         cpu_disable_c2_sleep++;
 1432                 if ((th->th_counter->tc_flags & TC_FLAGS_C2STOP) != 0)
 1433                         cpu_disable_c2_sleep--;
 1434 #endif
 1435                 th->th_counter = timecounter;
 1436                 th->th_offset_count = ncount;
 1437                 tc_min_ticktock_freq = max(1, timecounter->tc_frequency /
 1438                     (((uint64_t)timecounter->tc_counter_mask + 1) / 3));
 1439 #ifdef FFCLOCK
 1440                 ffclock_change_tc(th);
 1441 #endif
 1442         }
 1443 
 1444         /*-
 1445          * Recalculate the scaling factor.  We want the number of 1/2^64
 1446          * fractions of a second per period of the hardware counter, taking
 1447          * into account the th_adjustment factor which the NTP PLL/adjtime(2)
 1448          * processing provides us with.
 1449          *
 1450          * The th_adjustment is nanoseconds per second with 32 bit binary
 1451          * fraction and we want 64 bit binary fraction of second:
 1452          *
 1453          *       x = a * 2^32 / 10^9 = a * 4.294967296
 1454          *
 1455          * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
 1456          * we can only multiply by about 850 without overflowing, that
 1457          * leaves no suitably precise fractions for multiply before divide.
 1458          *
 1459          * Divide before multiply with a fraction of 2199/512 results in a
 1460          * systematic undercompensation of 10PPM of th_adjustment.  On a
 1461          * 5000PPM adjustment this is a 0.05PPM error.  This is acceptable.
 1462          *
 1463          * We happily sacrifice the lowest of the 64 bits of our result
 1464          * to the goddess of code clarity.
 1465          *
 1466          */
 1467         scale = (uint64_t)1 << 63;
 1468         scale += (th->th_adjustment / 1024) * 2199;
 1469         scale /= th->th_counter->tc_frequency;
 1470         th->th_scale = scale * 2;
 1471 
 1472         /*
 1473          * Now that the struct timehands is again consistent, set the new
 1474          * generation number, making sure to not make it zero.
 1475          */
 1476         if (++ogen == 0)
 1477                 ogen = 1;
 1478         atomic_store_rel_int(&th->th_generation, ogen);
 1479 
 1480         /* Go live with the new struct timehands. */
 1481 #ifdef FFCLOCK
 1482         switch (sysclock_active) {
 1483         case SYSCLOCK_FBCK:
 1484 #endif
 1485                 time_second = th->th_microtime.tv_sec;
 1486                 time_uptime = th->th_offset.sec;
 1487 #ifdef FFCLOCK
 1488                 break;
 1489         case SYSCLOCK_FFWD:
 1490                 time_second = fftimehands->tick_time_lerp.sec;
 1491                 time_uptime = fftimehands->tick_time_lerp.sec - ffclock_boottime.sec;
 1492                 break;
 1493         }
 1494 #endif
 1495 
 1496         timehands = th;
 1497         timekeep_push_vdso();
 1498 }
 1499 
 1500 /* Report or change the active timecounter hardware. */
 1501 static int
 1502 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
 1503 {
 1504         char newname[32];
 1505         struct timecounter *newtc, *tc;
 1506         int error;
 1507 
 1508         tc = timecounter;
 1509         strlcpy(newname, tc->tc_name, sizeof(newname));
 1510 
 1511         error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
 1512         if (error != 0 || req->newptr == NULL)
 1513                 return (error);
 1514         /* Record that the tc in use now was specifically chosen. */
 1515         tc_chosen = 1;
 1516         if (strcmp(newname, tc->tc_name) == 0)
 1517                 return (0);
 1518         for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
 1519                 if (strcmp(newname, newtc->tc_name) != 0)
 1520                         continue;
 1521 
 1522                 /* Warm up new timecounter. */
 1523                 (void)newtc->tc_get_timecount(newtc);
 1524                 (void)newtc->tc_get_timecount(newtc);
 1525 
 1526                 timecounter = newtc;
 1527 
 1528                 /*
 1529                  * The vdso timehands update is deferred until the next
 1530                  * 'tc_windup()'.
 1531                  *
 1532                  * This is prudent given that 'timekeep_push_vdso()' does not
 1533                  * use any locking and that it can be called in hard interrupt
 1534                  * context via 'tc_windup()'.
 1535                  */
 1536                 return (0);
 1537         }
 1538         return (EINVAL);
 1539 }
 1540 
 1541 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW,
 1542     0, 0, sysctl_kern_timecounter_hardware, "A",
 1543     "Timecounter hardware selected");
 1544 
 1545 
 1546 /* Report the available timecounter hardware. */
 1547 static int
 1548 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
 1549 {
 1550         struct sbuf sb;
 1551         struct timecounter *tc;
 1552         int error;
 1553 
 1554         sbuf_new_for_sysctl(&sb, NULL, 0, req);
 1555         for (tc = timecounters; tc != NULL; tc = tc->tc_next) {
 1556                 if (tc != timecounters)
 1557                         sbuf_putc(&sb, ' ');
 1558                 sbuf_printf(&sb, "%s(%d)", tc->tc_name, tc->tc_quality);
 1559         }
 1560         error = sbuf_finish(&sb);
 1561         sbuf_delete(&sb);
 1562         return (error);
 1563 }
 1564 
 1565 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice, CTLTYPE_STRING | CTLFLAG_RD,
 1566     0, 0, sysctl_kern_timecounter_choice, "A", "Timecounter hardware detected");
 1567 
 1568 /*
 1569  * RFC 2783 PPS-API implementation.
 1570  */
 1571 
 1572 /*
 1573  *  Return true if the driver is aware of the abi version extensions in the
 1574  *  pps_state structure, and it supports at least the given abi version number.
 1575  */
 1576 static inline int
 1577 abi_aware(struct pps_state *pps, int vers)
 1578 {
 1579 
 1580         return ((pps->kcmode & KCMODE_ABIFLAG) && pps->driver_abi >= vers);
 1581 }
 1582 
 1583 static int
 1584 pps_fetch(struct pps_fetch_args *fapi, struct pps_state *pps)
 1585 {
 1586         int err, timo;
 1587         pps_seq_t aseq, cseq;
 1588         struct timeval tv;
 1589 
 1590         if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
 1591                 return (EINVAL);
 1592 
 1593         /*
 1594          * If no timeout is requested, immediately return whatever values were
 1595          * most recently captured.  If timeout seconds is -1, that's a request
 1596          * to block without a timeout.  WITNESS won't let us sleep forever
 1597          * without a lock (we really don't need a lock), so just repeatedly
 1598          * sleep a long time.
 1599          */
 1600         if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) {
 1601                 if (fapi->timeout.tv_sec == -1)
 1602                         timo = 0x7fffffff;
 1603                 else {
 1604                         tv.tv_sec = fapi->timeout.tv_sec;
 1605                         tv.tv_usec = fapi->timeout.tv_nsec / 1000;
 1606                         timo = tvtohz(&tv);
 1607                 }
 1608                 aseq = atomic_load_int(&pps->ppsinfo.assert_sequence);
 1609                 cseq = atomic_load_int(&pps->ppsinfo.clear_sequence);
 1610                 while (aseq == atomic_load_int(&pps->ppsinfo.assert_sequence) &&
 1611                     cseq == atomic_load_int(&pps->ppsinfo.clear_sequence)) {
 1612                         if (abi_aware(pps, 1) && pps->driver_mtx != NULL) {
 1613                                 if (pps->flags & PPSFLAG_MTX_SPIN) {
 1614                                         err = msleep_spin(pps, pps->driver_mtx,
 1615                                             "ppsfch", timo);
 1616                                 } else {
 1617                                         err = msleep(pps, pps->driver_mtx, PCATCH,
 1618                                             "ppsfch", timo);
 1619                                 }
 1620                         } else {
 1621                                 err = tsleep(pps, PCATCH, "ppsfch", timo);
 1622                         }
 1623                         if (err == EWOULDBLOCK) {
 1624                                 if (fapi->timeout.tv_sec == -1) {
 1625                                         continue;
 1626                                 } else {
 1627                                         return (ETIMEDOUT);
 1628                                 }
 1629                         } else if (err != 0) {
 1630                                 return (err);
 1631                         }
 1632                 }
 1633         }
 1634 
 1635         pps->ppsinfo.current_mode = pps->ppsparam.mode;
 1636         fapi->pps_info_buf = pps->ppsinfo;
 1637 
 1638         return (0);
 1639 }
 1640 
 1641 int
 1642 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
 1643 {
 1644         pps_params_t *app;
 1645         struct pps_fetch_args *fapi;
 1646 #ifdef FFCLOCK
 1647         struct pps_fetch_ffc_args *fapi_ffc;
 1648 #endif
 1649 #ifdef PPS_SYNC
 1650         struct pps_kcbind_args *kapi;
 1651 #endif
 1652 
 1653         KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
 1654         switch (cmd) {
 1655         case PPS_IOC_CREATE:
 1656                 return (0);
 1657         case PPS_IOC_DESTROY:
 1658                 return (0);
 1659         case PPS_IOC_SETPARAMS:
 1660                 app = (pps_params_t *)data;
 1661                 if (app->mode & ~pps->ppscap)
 1662                         return (EINVAL);
 1663 #ifdef FFCLOCK
 1664                 /* Ensure only a single clock is selected for ffc timestamp. */
 1665                 if ((app->mode & PPS_TSCLK_MASK) == PPS_TSCLK_MASK)
 1666                         return (EINVAL);
 1667 #endif
 1668                 pps->ppsparam = *app;
 1669                 return (0);
 1670         case PPS_IOC_GETPARAMS:
 1671                 app = (pps_params_t *)data;
 1672                 *app = pps->ppsparam;
 1673                 app->api_version = PPS_API_VERS_1;
 1674                 return (0);
 1675         case PPS_IOC_GETCAP:
 1676                 *(int*)data = pps->ppscap;
 1677                 return (0);
 1678         case PPS_IOC_FETCH:
 1679                 fapi = (struct pps_fetch_args *)data;
 1680                 return (pps_fetch(fapi, pps));
 1681 #ifdef FFCLOCK
 1682         case PPS_IOC_FETCH_FFCOUNTER:
 1683                 fapi_ffc = (struct pps_fetch_ffc_args *)data;
 1684                 if (fapi_ffc->tsformat && fapi_ffc->tsformat !=
 1685                     PPS_TSFMT_TSPEC)
 1686                         return (EINVAL);
 1687                 if (fapi_ffc->timeout.tv_sec || fapi_ffc->timeout.tv_nsec)
 1688                         return (EOPNOTSUPP);
 1689                 pps->ppsinfo_ffc.current_mode = pps->ppsparam.mode;
 1690                 fapi_ffc->pps_info_buf_ffc = pps->ppsinfo_ffc;
 1691                 /* Overwrite timestamps if feedback clock selected. */
 1692                 switch (pps->ppsparam.mode & PPS_TSCLK_MASK) {
 1693                 case PPS_TSCLK_FBCK:
 1694                         fapi_ffc->pps_info_buf_ffc.assert_timestamp =
 1695                             pps->ppsinfo.assert_timestamp;
 1696                         fapi_ffc->pps_info_buf_ffc.clear_timestamp =
 1697                             pps->ppsinfo.clear_timestamp;
 1698                         break;
 1699                 case PPS_TSCLK_FFWD:
 1700                         break;
 1701                 default:
 1702                         break;
 1703                 }
 1704                 return (0);
 1705 #endif /* FFCLOCK */
 1706         case PPS_IOC_KCBIND:
 1707 #ifdef PPS_SYNC
 1708                 kapi = (struct pps_kcbind_args *)data;
 1709                 /* XXX Only root should be able to do this */
 1710                 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
 1711                         return (EINVAL);
 1712                 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
 1713                         return (EINVAL);
 1714                 if (kapi->edge & ~pps->ppscap)
 1715                         return (EINVAL);
 1716                 pps->kcmode = (kapi->edge & KCMODE_EDGEMASK) |
 1717                     (pps->kcmode & KCMODE_ABIFLAG);
 1718                 return (0);
 1719 #else
 1720                 return (EOPNOTSUPP);
 1721 #endif
 1722         default:
 1723                 return (ENOIOCTL);
 1724         }
 1725 }
 1726 
 1727 void
 1728 pps_init(struct pps_state *pps)
 1729 {
 1730         pps->ppscap |= PPS_TSFMT_TSPEC | PPS_CANWAIT;
 1731         if (pps->ppscap & PPS_CAPTUREASSERT)
 1732                 pps->ppscap |= PPS_OFFSETASSERT;
 1733         if (pps->ppscap & PPS_CAPTURECLEAR)
 1734                 pps->ppscap |= PPS_OFFSETCLEAR;
 1735 #ifdef FFCLOCK
 1736         pps->ppscap |= PPS_TSCLK_MASK;
 1737 #endif
 1738         pps->kcmode &= ~KCMODE_ABIFLAG;
 1739 }
 1740 
 1741 void
 1742 pps_init_abi(struct pps_state *pps)
 1743 {
 1744 
 1745         pps_init(pps);
 1746         if (pps->driver_abi > 0) {
 1747                 pps->kcmode |= KCMODE_ABIFLAG;
 1748                 pps->kernel_abi = PPS_ABI_VERSION;
 1749         }
 1750 }
 1751 
 1752 void
 1753 pps_capture(struct pps_state *pps)
 1754 {
 1755         struct timehands *th;
 1756 
 1757         KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
 1758         th = timehands;
 1759         pps->capgen = atomic_load_acq_int(&th->th_generation);
 1760         pps->capth = th;
 1761 #ifdef FFCLOCK
 1762         pps->capffth = fftimehands;
 1763 #endif
 1764         pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
 1765         atomic_thread_fence_acq();
 1766         if (pps->capgen != th->th_generation)
 1767                 pps->capgen = 0;
 1768 }
 1769 
 1770 void
 1771 pps_event(struct pps_state *pps, int event)
 1772 {
 1773         struct bintime bt;
 1774         struct timespec ts, *tsp, *osp;
 1775         u_int tcount, *pcount;
 1776         int foff;
 1777         pps_seq_t *pseq;
 1778 #ifdef FFCLOCK
 1779         struct timespec *tsp_ffc;
 1780         pps_seq_t *pseq_ffc;
 1781         ffcounter *ffcount;
 1782 #endif
 1783 #ifdef PPS_SYNC
 1784         int fhard;
 1785 #endif
 1786 
 1787         KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
 1788         /* Nothing to do if not currently set to capture this event type. */
 1789         if ((event & pps->ppsparam.mode) == 0)
 1790                 return;
 1791         /* If the timecounter was wound up underneath us, bail out. */
 1792         if (pps->capgen == 0 || pps->capgen !=
 1793             atomic_load_acq_int(&pps->capth->th_generation))
 1794                 return;
 1795 
 1796         /* Things would be easier with arrays. */
 1797         if (event == PPS_CAPTUREASSERT) {
 1798                 tsp = &pps->ppsinfo.assert_timestamp;
 1799                 osp = &pps->ppsparam.assert_offset;
 1800                 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
 1801 #ifdef PPS_SYNC
 1802                 fhard = pps->kcmode & PPS_CAPTUREASSERT;
 1803 #endif
 1804                 pcount = &pps->ppscount[0];
 1805                 pseq = &pps->ppsinfo.assert_sequence;
 1806 #ifdef FFCLOCK
 1807                 ffcount = &pps->ppsinfo_ffc.assert_ffcount;
 1808                 tsp_ffc = &pps->ppsinfo_ffc.assert_timestamp;
 1809                 pseq_ffc = &pps->ppsinfo_ffc.assert_sequence;
 1810 #endif
 1811         } else {
 1812                 tsp = &pps->ppsinfo.clear_timestamp;
 1813                 osp = &pps->ppsparam.clear_offset;
 1814                 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
 1815 #ifdef PPS_SYNC
 1816                 fhard = pps->kcmode & PPS_CAPTURECLEAR;
 1817 #endif
 1818                 pcount = &pps->ppscount[1];
 1819                 pseq = &pps->ppsinfo.clear_sequence;
 1820 #ifdef FFCLOCK
 1821                 ffcount = &pps->ppsinfo_ffc.clear_ffcount;
 1822                 tsp_ffc = &pps->ppsinfo_ffc.clear_timestamp;
 1823                 pseq_ffc = &pps->ppsinfo_ffc.clear_sequence;
 1824 #endif
 1825         }
 1826 
 1827         /*
 1828          * If the timecounter changed, we cannot compare the count values, so
 1829          * we have to drop the rest of the PPS-stuff until the next event.
 1830          */
 1831         if (pps->ppstc != pps->capth->th_counter) {
 1832                 pps->ppstc = pps->capth->th_counter;
 1833                 *pcount = pps->capcount;
 1834                 pps->ppscount[2] = pps->capcount;
 1835                 return;
 1836         }
 1837 
 1838         /* Convert the count to a timespec. */
 1839         tcount = pps->capcount - pps->capth->th_offset_count;
 1840         tcount &= pps->capth->th_counter->tc_counter_mask;
 1841         bt = pps->capth->th_bintime;
 1842         bintime_addx(&bt, pps->capth->th_scale * tcount);
 1843         bintime2timespec(&bt, &ts);
 1844 
 1845         /* If the timecounter was wound up underneath us, bail out. */
 1846         atomic_thread_fence_acq();
 1847         if (pps->capgen != pps->capth->th_generation)
 1848                 return;
 1849 
 1850         *pcount = pps->capcount;
 1851         (*pseq)++;
 1852         *tsp = ts;
 1853 
 1854         if (foff) {
 1855                 timespecadd(tsp, osp);
 1856                 if (tsp->tv_nsec < 0) {
 1857                         tsp->tv_nsec += 1000000000;
 1858                         tsp->tv_sec -= 1;
 1859                 }
 1860         }
 1861 
 1862 #ifdef FFCLOCK
 1863         *ffcount = pps->capffth->tick_ffcount + tcount;
 1864         bt = pps->capffth->tick_time;
 1865         ffclock_convert_delta(tcount, pps->capffth->cest.period, &bt);
 1866         bintime_add(&bt, &pps->capffth->tick_time);
 1867         bintime2timespec(&bt, &ts);
 1868         (*pseq_ffc)++;
 1869         *tsp_ffc = ts;
 1870 #endif
 1871 
 1872 #ifdef PPS_SYNC
 1873         if (fhard) {
 1874                 uint64_t scale;
 1875 
 1876                 /*
 1877                  * Feed the NTP PLL/FLL.
 1878                  * The FLL wants to know how many (hardware) nanoseconds
 1879                  * elapsed since the previous event.
 1880                  */
 1881                 tcount = pps->capcount - pps->ppscount[2];
 1882                 pps->ppscount[2] = pps->capcount;
 1883                 tcount &= pps->capth->th_counter->tc_counter_mask;
 1884                 scale = (uint64_t)1 << 63;
 1885                 scale /= pps->capth->th_counter->tc_frequency;
 1886                 scale *= 2;
 1887                 bt.sec = 0;
 1888                 bt.frac = 0;
 1889                 bintime_addx(&bt, scale * tcount);
 1890                 bintime2timespec(&bt, &ts);
 1891                 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
 1892         }
 1893 #endif
 1894 
 1895         /* Wakeup anyone sleeping in pps_fetch().  */
 1896         wakeup(pps);
 1897 }
 1898 
 1899 /*
 1900  * Timecounters need to be updated every so often to prevent the hardware
 1901  * counter from overflowing.  Updating also recalculates the cached values
 1902  * used by the get*() family of functions, so their precision depends on
 1903  * the update frequency.
 1904  */
 1905 
 1906 static int tc_tick;
 1907 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0,
 1908     "Approximate number of hardclock ticks in a millisecond");
 1909 
 1910 void
 1911 tc_ticktock(int cnt)
 1912 {
 1913         static int count;
 1914 
 1915         if (mtx_trylock_spin(&tc_setclock_mtx)) {
 1916                 count += cnt;
 1917                 if (count >= tc_tick) {
 1918                         count = 0;
 1919                         tc_windup(NULL);
 1920                 }
 1921                 mtx_unlock_spin(&tc_setclock_mtx);
 1922         }
 1923 }
 1924 
 1925 static void __inline
 1926 tc_adjprecision(void)
 1927 {
 1928         int t;
 1929 
 1930         if (tc_timepercentage > 0) {
 1931                 t = (99 + tc_timepercentage) / tc_timepercentage;
 1932                 tc_precexp = fls(t + (t >> 1)) - 1;
 1933                 FREQ2BT(hz / tc_tick, &bt_timethreshold);
 1934                 FREQ2BT(hz, &bt_tickthreshold);
 1935                 bintime_shift(&bt_timethreshold, tc_precexp);
 1936                 bintime_shift(&bt_tickthreshold, tc_precexp);
 1937         } else {
 1938                 tc_precexp = 31;
 1939                 bt_timethreshold.sec = INT_MAX;
 1940                 bt_timethreshold.frac = ~(uint64_t)0;
 1941                 bt_tickthreshold = bt_timethreshold;
 1942         }
 1943         sbt_timethreshold = bttosbt(bt_timethreshold);
 1944         sbt_tickthreshold = bttosbt(bt_tickthreshold);
 1945 }
 1946 
 1947 static int
 1948 sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS)
 1949 {
 1950         int error, val;
 1951 
 1952         val = tc_timepercentage;
 1953         error = sysctl_handle_int(oidp, &val, 0, req);
 1954         if (error != 0 || req->newptr == NULL)
 1955                 return (error);
 1956         tc_timepercentage = val;
 1957         if (cold)
 1958                 goto done;
 1959         tc_adjprecision();
 1960 done:
 1961         return (0);
 1962 }
 1963 
 1964 /* Set up the requested number of timehands. */
 1965 static void
 1966 inittimehands(void *dummy)
 1967 {
 1968         struct timehands *thp;
 1969         int i;
 1970 
 1971         TUNABLE_INT_FETCH("kern.timecounter.timehands_count",
 1972             &timehands_count);
 1973         if (timehands_count < 1)
 1974                 timehands_count = 1;
 1975         if (timehands_count > nitems(ths))
 1976                 timehands_count = nitems(ths);
 1977         for (i = 1, thp = &ths[0]; i < timehands_count;  thp = &ths[i++])
 1978                 thp->th_next = &ths[i];
 1979         thp->th_next = &ths[0];
 1980 }
 1981 SYSINIT(timehands, SI_SUB_TUNABLES, SI_ORDER_ANY, inittimehands, NULL);
 1982 
 1983 static void
 1984 inittimecounter(void *dummy)
 1985 {
 1986         u_int p;
 1987         int tick_rate;
 1988 
 1989         /*
 1990          * Set the initial timeout to
 1991          * max(1, <approx. number of hardclock ticks in a millisecond>).
 1992          * People should probably not use the sysctl to set the timeout
 1993          * to smaller than its initial value, since that value is the
 1994          * smallest reasonable one.  If they want better timestamps they
 1995          * should use the non-"get"* functions.
 1996          */
 1997         if (hz > 1000)
 1998                 tc_tick = (hz + 500) / 1000;
 1999         else
 2000                 tc_tick = 1;
 2001         tc_adjprecision();
 2002         FREQ2BT(hz, &tick_bt);
 2003         tick_sbt = bttosbt(tick_bt);
 2004         tick_rate = hz / tc_tick;
 2005         FREQ2BT(tick_rate, &tc_tick_bt);
 2006         tc_tick_sbt = bttosbt(tc_tick_bt);
 2007         p = (tc_tick * 1000000) / hz;
 2008         printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
 2009 
 2010 #ifdef FFCLOCK
 2011         ffclock_init();
 2012 #endif
 2013 
 2014         /* warm up new timecounter (again) and get rolling. */
 2015         (void)timecounter->tc_get_timecount(timecounter);
 2016         (void)timecounter->tc_get_timecount(timecounter);
 2017         mtx_lock_spin(&tc_setclock_mtx);
 2018         tc_windup(NULL);
 2019         mtx_unlock_spin(&tc_setclock_mtx);
 2020 }
 2021 
 2022 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL);
 2023 
 2024 /* Cpu tick handling -------------------------------------------------*/
 2025 
 2026 static int cpu_tick_variable;
 2027 static uint64_t cpu_tick_frequency;
 2028 
 2029 static DPCPU_DEFINE(uint64_t, tc_cpu_ticks_base);
 2030 static DPCPU_DEFINE(unsigned, tc_cpu_ticks_last);
 2031 
 2032 static uint64_t
 2033 tc_cpu_ticks(void)
 2034 {
 2035         struct timecounter *tc;
 2036         uint64_t res, *base;
 2037         unsigned u, *last;
 2038 
 2039         critical_enter();
 2040         base = DPCPU_PTR(tc_cpu_ticks_base);
 2041         last = DPCPU_PTR(tc_cpu_ticks_last);
 2042         tc = timehands->th_counter;
 2043         u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
 2044         if (u < *last)
 2045                 *base += (uint64_t)tc->tc_counter_mask + 1;
 2046         *last = u;
 2047         res = u + *base;
 2048         critical_exit();
 2049         return (res);
 2050 }
 2051 
 2052 void
 2053 cpu_tick_calibration(void)
 2054 {
 2055         static time_t last_calib;
 2056 
 2057         if (time_uptime != last_calib && !(time_uptime & 0xf)) {
 2058                 cpu_tick_calibrate(0);
 2059                 last_calib = time_uptime;
 2060         }
 2061 }
 2062 
 2063 /*
 2064  * This function gets called every 16 seconds on only one designated
 2065  * CPU in the system from hardclock() via cpu_tick_calibration()().
 2066  *
 2067  * Whenever the real time clock is stepped we get called with reset=1
 2068  * to make sure we handle suspend/resume and similar events correctly.
 2069  */
 2070 
 2071 static void
 2072 cpu_tick_calibrate(int reset)
 2073 {
 2074         static uint64_t c_last;
 2075         uint64_t c_this, c_delta;
 2076         static struct bintime  t_last;
 2077         struct bintime t_this, t_delta;
 2078         uint32_t divi;
 2079 
 2080         if (reset) {
 2081                 /* The clock was stepped, abort & reset */
 2082                 t_last.sec = 0;
 2083                 return;
 2084         }
 2085 
 2086         /* we don't calibrate fixed rate cputicks */
 2087         if (!cpu_tick_variable)
 2088                 return;
 2089 
 2090         getbinuptime(&t_this);
 2091         c_this = cpu_ticks();
 2092         if (t_last.sec != 0) {
 2093                 c_delta = c_this - c_last;
 2094                 t_delta = t_this;
 2095                 bintime_sub(&t_delta, &t_last);
 2096                 /*
 2097                  * Headroom:
 2098                  *      2^(64-20) / 16[s] =
 2099                  *      2^(44) / 16[s] =
 2100                  *      17.592.186.044.416 / 16 =
 2101                  *      1.099.511.627.776 [Hz]
 2102                  */
 2103                 divi = t_delta.sec << 20;
 2104                 divi |= t_delta.frac >> (64 - 20);
 2105                 c_delta <<= 20;
 2106                 c_delta /= divi;
 2107                 if (c_delta > cpu_tick_frequency) {
 2108                         if (0 && bootverbose)
 2109                                 printf("cpu_tick increased to %ju Hz\n",
 2110                                     c_delta);
 2111                         cpu_tick_frequency = c_delta;
 2112                 }
 2113         }
 2114         c_last = c_this;
 2115         t_last = t_this;
 2116 }
 2117 
 2118 void
 2119 set_cputicker(cpu_tick_f *func, uint64_t freq, unsigned var)
 2120 {
 2121 
 2122         if (func == NULL) {
 2123                 cpu_ticks = tc_cpu_ticks;
 2124         } else {
 2125                 cpu_tick_frequency = freq;
 2126                 cpu_tick_variable = var;
 2127                 cpu_ticks = func;
 2128         }
 2129 }
 2130 
 2131 uint64_t
 2132 cpu_tickrate(void)
 2133 {
 2134 
 2135         if (cpu_ticks == tc_cpu_ticks) 
 2136                 return (tc_getfrequency());
 2137         return (cpu_tick_frequency);
 2138 }
 2139 
 2140 /*
 2141  * We need to be slightly careful converting cputicks to microseconds.
 2142  * There is plenty of margin in 64 bits of microseconds (half a million
 2143  * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
 2144  * before divide conversion (to retain precision) we find that the
 2145  * margin shrinks to 1.5 hours (one millionth of 146y).
 2146  * With a three prong approach we never lose significant bits, no
 2147  * matter what the cputick rate and length of timeinterval is.
 2148  */
 2149 
 2150 uint64_t
 2151 cputick2usec(uint64_t tick)
 2152 {
 2153 
 2154         if (tick > 18446744073709551LL)         /* floor(2^64 / 1000) */
 2155                 return (tick / (cpu_tickrate() / 1000000LL));
 2156         else if (tick > 18446744073709LL)       /* floor(2^64 / 1000000) */
 2157                 return ((tick * 1000LL) / (cpu_tickrate() / 1000LL));
 2158         else
 2159                 return ((tick * 1000000LL) / cpu_tickrate());
 2160 }
 2161 
 2162 cpu_tick_f      *cpu_ticks = tc_cpu_ticks;
 2163 
 2164 static int vdso_th_enable = 1;
 2165 static int
 2166 sysctl_fast_gettime(SYSCTL_HANDLER_ARGS)
 2167 {
 2168         int old_vdso_th_enable, error;
 2169 
 2170         old_vdso_th_enable = vdso_th_enable;
 2171         error = sysctl_handle_int(oidp, &old_vdso_th_enable, 0, req);
 2172         if (error != 0)
 2173                 return (error);
 2174         vdso_th_enable = old_vdso_th_enable;
 2175         return (0);
 2176 }
 2177 SYSCTL_PROC(_kern_timecounter, OID_AUTO, fast_gettime,
 2178     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
 2179     NULL, 0, sysctl_fast_gettime, "I", "Enable fast time of day");
 2180 
 2181 uint32_t
 2182 tc_fill_vdso_timehands(struct vdso_timehands *vdso_th)
 2183 {
 2184         struct timehands *th;
 2185         uint32_t enabled;
 2186 
 2187         th = timehands;
 2188         vdso_th->th_scale = th->th_scale;
 2189         vdso_th->th_offset_count = th->th_offset_count;
 2190         vdso_th->th_counter_mask = th->th_counter->tc_counter_mask;
 2191         vdso_th->th_offset = th->th_offset;
 2192         vdso_th->th_boottime = th->th_boottime;
 2193         if (th->th_counter->tc_fill_vdso_timehands != NULL) {
 2194                 enabled = th->th_counter->tc_fill_vdso_timehands(vdso_th,
 2195                     th->th_counter);
 2196         } else
 2197                 enabled = 0;
 2198         if (!vdso_th_enable)
 2199                 enabled = 0;
 2200         return (enabled);
 2201 }
 2202 
 2203 #ifdef COMPAT_FREEBSD32
 2204 uint32_t
 2205 tc_fill_vdso_timehands32(struct vdso_timehands32 *vdso_th32)
 2206 {
 2207         struct timehands *th;
 2208         uint32_t enabled;
 2209 
 2210         th = timehands;
 2211         *(uint64_t *)&vdso_th32->th_scale[0] = th->th_scale;
 2212         vdso_th32->th_offset_count = th->th_offset_count;
 2213         vdso_th32->th_counter_mask = th->th_counter->tc_counter_mask;
 2214         vdso_th32->th_offset.sec = th->th_offset.sec;
 2215         *(uint64_t *)&vdso_th32->th_offset.frac[0] = th->th_offset.frac;
 2216         vdso_th32->th_boottime.sec = th->th_boottime.sec;
 2217         *(uint64_t *)&vdso_th32->th_boottime.frac[0] = th->th_boottime.frac;
 2218         if (th->th_counter->tc_fill_vdso_timehands32 != NULL) {
 2219                 enabled = th->th_counter->tc_fill_vdso_timehands32(vdso_th32,
 2220                     th->th_counter);
 2221         } else
 2222                 enabled = 0;
 2223         if (!vdso_th_enable)
 2224                 enabled = 0;
 2225         return (enabled);
 2226 }
 2227 #endif

Cache object: 5766453edea28dbcd8e65980168270dd


[ 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.