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-12-STABLE  -  FREEBSD-12-0  -  FREEBSD-11-STABLE  -  FREEBSD-11-2  -  FREEBSD-11-1  -  FREEBSD-11-0  -  FREEBSD-10-STABLE  -  FREEBSD-10-4  -  FREEBSD-10-3  -  FREEBSD-10-2  -  FREEBSD-10-1  -  FREEBSD-10-0  -  FREEBSD-9-STABLE  -  FREEBSD-9-3  -  FREEBSD-9-2  -  FREEBSD-9-1  -  FREEBSD-9-0  -  FREEBSD-8-STABLE  -  FREEBSD-8-4  -  FREEBSD-8-3  -  FREEBSD-8-2  -  FREEBSD-8-1  -  FREEBSD-8-0  -  FREEBSD-7-STABLE  -  FREEBSD-7-4  -  FREEBSD-7-3  -  FREEBSD-7-2  -  FREEBSD-7-1  -  FREEBSD-7-0  -  FREEBSD-6-STABLE  -  FREEBSD-6-4  -  FREEBSD-6-3  -  FREEBSD-6-2  -  FREEBSD-6-1  -  FREEBSD-6-0  -  FREEBSD-5-STABLE  -  FREEBSD-5-5  -  FREEBSD-5-4  -  FREEBSD-5-3  -  FREEBSD-5-2  -  FREEBSD-5-1  -  FREEBSD-5-0  -  FREEBSD-4-STABLE  -  FREEBSD-3-STABLE  -  FREEBSD22  -  linux-2.6  -  linux-2.4.22  -  MK83  -  MK84  -  PLAN9  -  DFBSD  -  NETBSD  -  NETBSD5  -  NETBSD4  -  NETBSD3  -  NETBSD20  -  OPENBSD  -  xnu-517  -  xnu-792  -  xnu-792.6.70  -  xnu-1228  -  xnu-1456.1.26  -  xnu-1699.24.8  -  xnu-2050.18.24  -  OPENSOLARIS  -  minix-3-1-1 
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 
   10 #include <sys/cdefs.h>
   11 __FBSDID("$FreeBSD: releng/8.3/sys/kern/kern_tc.c 216015 2010-11-28 18:59:52Z cperciva $");
   12 
   13 #include "opt_ntp.h"
   14 
   15 #include <sys/param.h>
   16 #include <sys/kernel.h>
   17 #include <sys/sysctl.h>
   18 #include <sys/syslog.h>
   19 #include <sys/systm.h>
   20 #include <sys/timepps.h>
   21 #include <sys/timetc.h>
   22 #include <sys/timex.h>
   23 
   24 /*
   25  * A large step happens on boot.  This constant detects such steps.
   26  * It is relatively small so that ntp_update_second gets called enough
   27  * in the typical 'missed a couple of seconds' case, but doesn't loop
   28  * forever when the time step is large.
   29  */
   30 #define LARGE_STEP      200
   31 
   32 /*
   33  * Implement a dummy timecounter which we can use until we get a real one
   34  * in the air.  This allows the console and other early stuff to use
   35  * time services.
   36  */
   37 
   38 static u_int
   39 dummy_get_timecount(struct timecounter *tc)
   40 {
   41         static u_int now;
   42 
   43         return (++now);
   44 }
   45 
   46 static struct timecounter dummy_timecounter = {
   47         dummy_get_timecount, 0, ~0u, 1000000, "dummy", -1000000
   48 };
   49 
   50 struct timehands {
   51         /* These fields must be initialized by the driver. */
   52         struct timecounter      *th_counter;
   53         int64_t                 th_adjustment;
   54         u_int64_t               th_scale;
   55         u_int                   th_offset_count;
   56         struct bintime          th_offset;
   57         struct timeval          th_microtime;
   58         struct timespec         th_nanotime;
   59         /* Fields not to be copied in tc_windup start with th_generation. */
   60         volatile u_int          th_generation;
   61         struct timehands        *th_next;
   62 };
   63 
   64 static struct timehands th0;
   65 static struct timehands th9 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th0};
   66 static struct timehands th8 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th9};
   67 static struct timehands th7 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th8};
   68 static struct timehands th6 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th7};
   69 static struct timehands th5 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th6};
   70 static struct timehands th4 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th5};
   71 static struct timehands th3 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th4};
   72 static struct timehands th2 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th3};
   73 static struct timehands th1 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th2};
   74 static struct timehands th0 = {
   75         &dummy_timecounter,
   76         0,
   77         (uint64_t)-1 / 1000000,
   78         0,
   79         {1, 0},
   80         {0, 0},
   81         {0, 0},
   82         1,
   83         &th1
   84 };
   85 
   86 static struct timehands *volatile timehands = &th0;
   87 struct timecounter *timecounter = &dummy_timecounter;
   88 static struct timecounter *timecounters = &dummy_timecounter;
   89 
   90 time_t time_second = 1;
   91 time_t time_uptime = 1;
   92 
   93 static struct bintime boottimebin;
   94 struct timeval boottime;
   95 static int sysctl_kern_boottime(SYSCTL_HANDLER_ARGS);
   96 SYSCTL_PROC(_kern, KERN_BOOTTIME, boottime, CTLTYPE_STRUCT|CTLFLAG_RD,
   97     NULL, 0, sysctl_kern_boottime, "S,timeval", "System boottime");
   98 
   99 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW, 0, "");
  100 SYSCTL_NODE(_kern_timecounter, OID_AUTO, tc, CTLFLAG_RW, 0, "");
  101 
  102 static int timestepwarnings;
  103 SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RW,
  104     &timestepwarnings, 0, "Log time steps");
  105 
  106 static void tc_windup(void);
  107 static void cpu_tick_calibrate(int);
  108 
  109 static int
  110 sysctl_kern_boottime(SYSCTL_HANDLER_ARGS)
  111 {
  112 #ifdef SCTL_MASK32
  113         int tv[2];
  114 
  115         if (req->flags & SCTL_MASK32) {
  116                 tv[0] = boottime.tv_sec;
  117                 tv[1] = boottime.tv_usec;
  118                 return SYSCTL_OUT(req, tv, sizeof(tv));
  119         } else
  120 #endif
  121                 return SYSCTL_OUT(req, &boottime, sizeof(boottime));
  122 }
  123 
  124 static int
  125 sysctl_kern_timecounter_get(SYSCTL_HANDLER_ARGS)
  126 {
  127         u_int ncount;
  128         struct timecounter *tc = arg1;
  129 
  130         ncount = tc->tc_get_timecount(tc);
  131         return sysctl_handle_int(oidp, &ncount, 0, req);
  132 }
  133 
  134 static int
  135 sysctl_kern_timecounter_freq(SYSCTL_HANDLER_ARGS)
  136 {
  137         u_int64_t freq;
  138         struct timecounter *tc = arg1;
  139 
  140         freq = tc->tc_frequency;
  141         return sysctl_handle_quad(oidp, &freq, 0, req);
  142 }
  143 
  144 /*
  145  * Return the difference between the timehands' counter value now and what
  146  * was when we copied it to the timehands' offset_count.
  147  */
  148 static __inline u_int
  149 tc_delta(struct timehands *th)
  150 {
  151         struct timecounter *tc;
  152 
  153         tc = th->th_counter;
  154         return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
  155             tc->tc_counter_mask);
  156 }
  157 
  158 /*
  159  * Functions for reading the time.  We have to loop until we are sure that
  160  * the timehands that we operated on was not updated under our feet.  See
  161  * the comment in <sys/time.h> for a description of these 12 functions.
  162  */
  163 
  164 void
  165 binuptime(struct bintime *bt)
  166 {
  167         struct timehands *th;
  168         u_int gen;
  169 
  170         do {
  171                 th = timehands;
  172                 gen = th->th_generation;
  173                 *bt = th->th_offset;
  174                 bintime_addx(bt, th->th_scale * tc_delta(th));
  175         } while (gen == 0 || gen != th->th_generation);
  176 }
  177 
  178 void
  179 nanouptime(struct timespec *tsp)
  180 {
  181         struct bintime bt;
  182 
  183         binuptime(&bt);
  184         bintime2timespec(&bt, tsp);
  185 }
  186 
  187 void
  188 microuptime(struct timeval *tvp)
  189 {
  190         struct bintime bt;
  191 
  192         binuptime(&bt);
  193         bintime2timeval(&bt, tvp);
  194 }
  195 
  196 void
  197 bintime(struct bintime *bt)
  198 {
  199 
  200         binuptime(bt);
  201         bintime_add(bt, &boottimebin);
  202 }
  203 
  204 void
  205 nanotime(struct timespec *tsp)
  206 {
  207         struct bintime bt;
  208 
  209         bintime(&bt);
  210         bintime2timespec(&bt, tsp);
  211 }
  212 
  213 void
  214 microtime(struct timeval *tvp)
  215 {
  216         struct bintime bt;
  217 
  218         bintime(&bt);
  219         bintime2timeval(&bt, tvp);
  220 }
  221 
  222 void
  223 getbinuptime(struct bintime *bt)
  224 {
  225         struct timehands *th;
  226         u_int gen;
  227 
  228         do {
  229                 th = timehands;
  230                 gen = th->th_generation;
  231                 *bt = th->th_offset;
  232         } while (gen == 0 || gen != th->th_generation);
  233 }
  234 
  235 void
  236 getnanouptime(struct timespec *tsp)
  237 {
  238         struct timehands *th;
  239         u_int gen;
  240 
  241         do {
  242                 th = timehands;
  243                 gen = th->th_generation;
  244                 bintime2timespec(&th->th_offset, tsp);
  245         } while (gen == 0 || gen != th->th_generation);
  246 }
  247 
  248 void
  249 getmicrouptime(struct timeval *tvp)
  250 {
  251         struct timehands *th;
  252         u_int gen;
  253 
  254         do {
  255                 th = timehands;
  256                 gen = th->th_generation;
  257                 bintime2timeval(&th->th_offset, tvp);
  258         } while (gen == 0 || gen != th->th_generation);
  259 }
  260 
  261 void
  262 getbintime(struct bintime *bt)
  263 {
  264         struct timehands *th;
  265         u_int gen;
  266 
  267         do {
  268                 th = timehands;
  269                 gen = th->th_generation;
  270                 *bt = th->th_offset;
  271         } while (gen == 0 || gen != th->th_generation);
  272         bintime_add(bt, &boottimebin);
  273 }
  274 
  275 void
  276 getnanotime(struct timespec *tsp)
  277 {
  278         struct timehands *th;
  279         u_int gen;
  280 
  281         do {
  282                 th = timehands;
  283                 gen = th->th_generation;
  284                 *tsp = th->th_nanotime;
  285         } while (gen == 0 || gen != th->th_generation);
  286 }
  287 
  288 void
  289 getmicrotime(struct timeval *tvp)
  290 {
  291         struct timehands *th;
  292         u_int gen;
  293 
  294         do {
  295                 th = timehands;
  296                 gen = th->th_generation;
  297                 *tvp = th->th_microtime;
  298         } while (gen == 0 || gen != th->th_generation);
  299 }
  300 
  301 /*
  302  * Initialize a new timecounter and possibly use it.
  303  */
  304 void
  305 tc_init(struct timecounter *tc)
  306 {
  307         u_int u;
  308         struct sysctl_oid *tc_root;
  309 
  310         u = tc->tc_frequency / tc->tc_counter_mask;
  311         /* XXX: We need some margin here, 10% is a guess */
  312         u *= 11;
  313         u /= 10;
  314         if (u > hz && tc->tc_quality >= 0) {
  315                 tc->tc_quality = -2000;
  316                 if (bootverbose) {
  317                         printf("Timecounter \"%s\" frequency %ju Hz",
  318                             tc->tc_name, (uintmax_t)tc->tc_frequency);
  319                         printf(" -- Insufficient hz, needs at least %u\n", u);
  320                 }
  321         } else if (tc->tc_quality >= 0 || bootverbose) {
  322                 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
  323                     tc->tc_name, (uintmax_t)tc->tc_frequency,
  324                     tc->tc_quality);
  325         }
  326 
  327         tc->tc_next = timecounters;
  328         timecounters = tc;
  329         /*
  330          * Set up sysctl tree for this counter.
  331          */
  332         tc_root = SYSCTL_ADD_NODE(NULL,
  333             SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name,
  334             CTLFLAG_RW, 0, "timecounter description");
  335         SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
  336             "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0,
  337             "mask for implemented bits");
  338         SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
  339             "counter", CTLTYPE_UINT | CTLFLAG_RD, tc, sizeof(*tc),
  340             sysctl_kern_timecounter_get, "IU", "current timecounter value");
  341         SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
  342             "frequency", CTLTYPE_QUAD | CTLFLAG_RD, tc, sizeof(*tc),
  343              sysctl_kern_timecounter_freq, "QU", "timecounter frequency");
  344         SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
  345             "quality", CTLFLAG_RD, &(tc->tc_quality), 0,
  346             "goodness of time counter");
  347         /*
  348          * Never automatically use a timecounter with negative quality.
  349          * Even though we run on the dummy counter, switching here may be
  350          * worse since this timecounter may not be monotonous.
  351          */
  352         if (tc->tc_quality < 0)
  353                 return;
  354         if (tc->tc_quality < timecounter->tc_quality)
  355                 return;
  356         if (tc->tc_quality == timecounter->tc_quality &&
  357             tc->tc_frequency < timecounter->tc_frequency)
  358                 return;
  359         (void)tc->tc_get_timecount(tc);
  360         (void)tc->tc_get_timecount(tc);
  361         timecounter = tc;
  362 }
  363 
  364 /* Report the frequency of the current timecounter. */
  365 u_int64_t
  366 tc_getfrequency(void)
  367 {
  368 
  369         return (timehands->th_counter->tc_frequency);
  370 }
  371 
  372 /*
  373  * Step our concept of UTC.  This is done by modifying our estimate of
  374  * when we booted.
  375  * XXX: not locked.
  376  */
  377 void
  378 tc_setclock(struct timespec *ts)
  379 {
  380         struct timespec tbef, taft;
  381         struct bintime bt, bt2;
  382 
  383         cpu_tick_calibrate(1);
  384         nanotime(&tbef);
  385         timespec2bintime(ts, &bt);
  386         binuptime(&bt2);
  387         bintime_sub(&bt, &bt2);
  388         bintime_add(&bt2, &boottimebin);
  389         boottimebin = bt;
  390         bintime2timeval(&bt, &boottime);
  391 
  392         /* XXX fiddle all the little crinkly bits around the fiords... */
  393         tc_windup();
  394         nanotime(&taft);
  395         if (timestepwarnings) {
  396                 log(LOG_INFO,
  397                     "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
  398                     (intmax_t)tbef.tv_sec, tbef.tv_nsec,
  399                     (intmax_t)taft.tv_sec, taft.tv_nsec,
  400                     (intmax_t)ts->tv_sec, ts->tv_nsec);
  401         }
  402         cpu_tick_calibrate(1);
  403 }
  404 
  405 /*
  406  * Initialize the next struct timehands in the ring and make
  407  * it the active timehands.  Along the way we might switch to a different
  408  * timecounter and/or do seconds processing in NTP.  Slightly magic.
  409  */
  410 static void
  411 tc_windup(void)
  412 {
  413         struct bintime bt;
  414         struct timehands *th, *tho;
  415         u_int64_t scale;
  416         u_int delta, ncount, ogen;
  417         int i;
  418         time_t t;
  419 
  420         /*
  421          * Make the next timehands a copy of the current one, but do not
  422          * overwrite the generation or next pointer.  While we update
  423          * the contents, the generation must be zero.
  424          */
  425         tho = timehands;
  426         th = tho->th_next;
  427         ogen = th->th_generation;
  428         th->th_generation = 0;
  429         bcopy(tho, th, offsetof(struct timehands, th_generation));
  430 
  431         /*
  432          * Capture a timecounter delta on the current timecounter and if
  433          * changing timecounters, a counter value from the new timecounter.
  434          * Update the offset fields accordingly.
  435          */
  436         delta = tc_delta(th);
  437         if (th->th_counter != timecounter)
  438                 ncount = timecounter->tc_get_timecount(timecounter);
  439         else
  440                 ncount = 0;
  441         th->th_offset_count += delta;
  442         th->th_offset_count &= th->th_counter->tc_counter_mask;
  443         while (delta > th->th_counter->tc_frequency) {
  444                 /* Eat complete unadjusted seconds. */
  445                 delta -= th->th_counter->tc_frequency;
  446                 th->th_offset.sec++;
  447         }
  448         if ((delta > th->th_counter->tc_frequency / 2) &&
  449             (th->th_scale * delta < ((uint64_t)1 << 63))) {
  450                 /* The product th_scale * delta just barely overflows. */
  451                 th->th_offset.sec++;
  452         }
  453         bintime_addx(&th->th_offset, th->th_scale * delta);
  454 
  455         /*
  456          * Hardware latching timecounters may not generate interrupts on
  457          * PPS events, so instead we poll them.  There is a finite risk that
  458          * the hardware might capture a count which is later than the one we
  459          * got above, and therefore possibly in the next NTP second which might
  460          * have a different rate than the current NTP second.  It doesn't
  461          * matter in practice.
  462          */
  463         if (tho->th_counter->tc_poll_pps)
  464                 tho->th_counter->tc_poll_pps(tho->th_counter);
  465 
  466         /*
  467          * Deal with NTP second processing.  The for loop normally
  468          * iterates at most once, but in extreme situations it might
  469          * keep NTP sane if timeouts are not run for several seconds.
  470          * At boot, the time step can be large when the TOD hardware
  471          * has been read, so on really large steps, we call
  472          * ntp_update_second only twice.  We need to call it twice in
  473          * case we missed a leap second.
  474          */
  475         bt = th->th_offset;
  476         bintime_add(&bt, &boottimebin);
  477         i = bt.sec - tho->th_microtime.tv_sec;
  478         if (i > LARGE_STEP)
  479                 i = 2;
  480         for (; i > 0; i--) {
  481                 t = bt.sec;
  482                 ntp_update_second(&th->th_adjustment, &bt.sec);
  483                 if (bt.sec != t)
  484                         boottimebin.sec += bt.sec - t;
  485         }
  486         /* Update the UTC timestamps used by the get*() functions. */
  487         /* XXX shouldn't do this here.  Should force non-`get' versions. */
  488         bintime2timeval(&bt, &th->th_microtime);
  489         bintime2timespec(&bt, &th->th_nanotime);
  490 
  491         /* Now is a good time to change timecounters. */
  492         if (th->th_counter != timecounter) {
  493                 th->th_counter = timecounter;
  494                 th->th_offset_count = ncount;
  495         }
  496 
  497         /*-
  498          * Recalculate the scaling factor.  We want the number of 1/2^64
  499          * fractions of a second per period of the hardware counter, taking
  500          * into account the th_adjustment factor which the NTP PLL/adjtime(2)
  501          * processing provides us with.
  502          *
  503          * The th_adjustment is nanoseconds per second with 32 bit binary
  504          * fraction and we want 64 bit binary fraction of second:
  505          *
  506          *       x = a * 2^32 / 10^9 = a * 4.294967296
  507          *
  508          * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
  509          * we can only multiply by about 850 without overflowing, that
  510          * leaves no suitably precise fractions for multiply before divide.
  511          *
  512          * Divide before multiply with a fraction of 2199/512 results in a
  513          * systematic undercompensation of 10PPM of th_adjustment.  On a
  514          * 5000PPM adjustment this is a 0.05PPM error.  This is acceptable.
  515          *
  516          * We happily sacrifice the lowest of the 64 bits of our result
  517          * to the goddess of code clarity.
  518          *
  519          */
  520         scale = (u_int64_t)1 << 63;
  521         scale += (th->th_adjustment / 1024) * 2199;
  522         scale /= th->th_counter->tc_frequency;
  523         th->th_scale = scale * 2;
  524 
  525         /*
  526          * Now that the struct timehands is again consistent, set the new
  527          * generation number, making sure to not make it zero.
  528          */
  529         if (++ogen == 0)
  530                 ogen = 1;
  531         th->th_generation = ogen;
  532 
  533         /* Go live with the new struct timehands. */
  534         time_second = th->th_microtime.tv_sec;
  535         time_uptime = th->th_offset.sec;
  536         timehands = th;
  537 }
  538 
  539 /* Report or change the active timecounter hardware. */
  540 static int
  541 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
  542 {
  543         char newname[32];
  544         struct timecounter *newtc, *tc;
  545         int error;
  546 
  547         tc = timecounter;
  548         strlcpy(newname, tc->tc_name, sizeof(newname));
  549 
  550         error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
  551         if (error != 0 || req->newptr == NULL ||
  552             strcmp(newname, tc->tc_name) == 0)
  553                 return (error);
  554         for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
  555                 if (strcmp(newname, newtc->tc_name) != 0)
  556                         continue;
  557 
  558                 /* Warm up new timecounter. */
  559                 (void)newtc->tc_get_timecount(newtc);
  560                 (void)newtc->tc_get_timecount(newtc);
  561 
  562                 timecounter = newtc;
  563                 return (0);
  564         }
  565         return (EINVAL);
  566 }
  567 
  568 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW,
  569     0, 0, sysctl_kern_timecounter_hardware, "A",
  570     "Timecounter hardware selected");
  571 
  572 
  573 /* Report or change the active timecounter hardware. */
  574 static int
  575 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
  576 {
  577         char buf[32], *spc;
  578         struct timecounter *tc;
  579         int error;
  580 
  581         spc = "";
  582         error = 0;
  583         for (tc = timecounters; error == 0 && tc != NULL; tc = tc->tc_next) {
  584                 sprintf(buf, "%s%s(%d)",
  585                     spc, tc->tc_name, tc->tc_quality);
  586                 error = SYSCTL_OUT(req, buf, strlen(buf));
  587                 spc = " ";
  588         }
  589         return (error);
  590 }
  591 
  592 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice, CTLTYPE_STRING | CTLFLAG_RD,
  593     0, 0, sysctl_kern_timecounter_choice, "A", "Timecounter hardware detected");
  594 
  595 /*
  596  * RFC 2783 PPS-API implementation.
  597  */
  598 
  599 int
  600 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
  601 {
  602         pps_params_t *app;
  603         struct pps_fetch_args *fapi;
  604 #ifdef PPS_SYNC
  605         struct pps_kcbind_args *kapi;
  606 #endif
  607 
  608         KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
  609         switch (cmd) {
  610         case PPS_IOC_CREATE:
  611                 return (0);
  612         case PPS_IOC_DESTROY:
  613                 return (0);
  614         case PPS_IOC_SETPARAMS:
  615                 app = (pps_params_t *)data;
  616                 if (app->mode & ~pps->ppscap)
  617                         return (EINVAL);
  618                 pps->ppsparam = *app;
  619                 return (0);
  620         case PPS_IOC_GETPARAMS:
  621                 app = (pps_params_t *)data;
  622                 *app = pps->ppsparam;
  623                 app->api_version = PPS_API_VERS_1;
  624                 return (0);
  625         case PPS_IOC_GETCAP:
  626                 *(int*)data = pps->ppscap;
  627                 return (0);
  628         case PPS_IOC_FETCH:
  629                 fapi = (struct pps_fetch_args *)data;
  630                 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
  631                         return (EINVAL);
  632                 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
  633                         return (EOPNOTSUPP);
  634                 pps->ppsinfo.current_mode = pps->ppsparam.mode;
  635                 fapi->pps_info_buf = pps->ppsinfo;
  636                 return (0);
  637         case PPS_IOC_KCBIND:
  638 #ifdef PPS_SYNC
  639                 kapi = (struct pps_kcbind_args *)data;
  640                 /* XXX Only root should be able to do this */
  641                 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
  642                         return (EINVAL);
  643                 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
  644                         return (EINVAL);
  645                 if (kapi->edge & ~pps->ppscap)
  646                         return (EINVAL);
  647                 pps->kcmode = kapi->edge;
  648                 return (0);
  649 #else
  650                 return (EOPNOTSUPP);
  651 #endif
  652         default:
  653                 return (ENOIOCTL);
  654         }
  655 }
  656 
  657 void
  658 pps_init(struct pps_state *pps)
  659 {
  660         pps->ppscap |= PPS_TSFMT_TSPEC;
  661         if (pps->ppscap & PPS_CAPTUREASSERT)
  662                 pps->ppscap |= PPS_OFFSETASSERT;
  663         if (pps->ppscap & PPS_CAPTURECLEAR)
  664                 pps->ppscap |= PPS_OFFSETCLEAR;
  665 }
  666 
  667 void
  668 pps_capture(struct pps_state *pps)
  669 {
  670         struct timehands *th;
  671 
  672         KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
  673         th = timehands;
  674         pps->capgen = th->th_generation;
  675         pps->capth = th;
  676         pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
  677         if (pps->capgen != th->th_generation)
  678                 pps->capgen = 0;
  679 }
  680 
  681 void
  682 pps_event(struct pps_state *pps, int event)
  683 {
  684         struct bintime bt;
  685         struct timespec ts, *tsp, *osp;
  686         u_int tcount, *pcount;
  687         int foff, fhard;
  688         pps_seq_t *pseq;
  689 
  690         KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
  691         /* If the timecounter was wound up underneath us, bail out. */
  692         if (pps->capgen == 0 || pps->capgen != pps->capth->th_generation)
  693                 return;
  694 
  695         /* Things would be easier with arrays. */
  696         if (event == PPS_CAPTUREASSERT) {
  697                 tsp = &pps->ppsinfo.assert_timestamp;
  698                 osp = &pps->ppsparam.assert_offset;
  699                 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
  700                 fhard = pps->kcmode & PPS_CAPTUREASSERT;
  701                 pcount = &pps->ppscount[0];
  702                 pseq = &pps->ppsinfo.assert_sequence;
  703         } else {
  704                 tsp = &pps->ppsinfo.clear_timestamp;
  705                 osp = &pps->ppsparam.clear_offset;
  706                 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
  707                 fhard = pps->kcmode & PPS_CAPTURECLEAR;
  708                 pcount = &pps->ppscount[1];
  709                 pseq = &pps->ppsinfo.clear_sequence;
  710         }
  711 
  712         /*
  713          * If the timecounter changed, we cannot compare the count values, so
  714          * we have to drop the rest of the PPS-stuff until the next event.
  715          */
  716         if (pps->ppstc != pps->capth->th_counter) {
  717                 pps->ppstc = pps->capth->th_counter;
  718                 *pcount = pps->capcount;
  719                 pps->ppscount[2] = pps->capcount;
  720                 return;
  721         }
  722 
  723         /* Convert the count to a timespec. */
  724         tcount = pps->capcount - pps->capth->th_offset_count;
  725         tcount &= pps->capth->th_counter->tc_counter_mask;
  726         bt = pps->capth->th_offset;
  727         bintime_addx(&bt, pps->capth->th_scale * tcount);
  728         bintime_add(&bt, &boottimebin);
  729         bintime2timespec(&bt, &ts);
  730 
  731         /* If the timecounter was wound up underneath us, bail out. */
  732         if (pps->capgen != pps->capth->th_generation)
  733                 return;
  734 
  735         *pcount = pps->capcount;
  736         (*pseq)++;
  737         *tsp = ts;
  738 
  739         if (foff) {
  740                 timespecadd(tsp, osp);
  741                 if (tsp->tv_nsec < 0) {
  742                         tsp->tv_nsec += 1000000000;
  743                         tsp->tv_sec -= 1;
  744                 }
  745         }
  746 #ifdef PPS_SYNC
  747         if (fhard) {
  748                 u_int64_t scale;
  749 
  750                 /*
  751                  * Feed the NTP PLL/FLL.
  752                  * The FLL wants to know how many (hardware) nanoseconds
  753                  * elapsed since the previous event.
  754                  */
  755                 tcount = pps->capcount - pps->ppscount[2];
  756                 pps->ppscount[2] = pps->capcount;
  757                 tcount &= pps->capth->th_counter->tc_counter_mask;
  758                 scale = (u_int64_t)1 << 63;
  759                 scale /= pps->capth->th_counter->tc_frequency;
  760                 scale *= 2;
  761                 bt.sec = 0;
  762                 bt.frac = 0;
  763                 bintime_addx(&bt, scale * tcount);
  764                 bintime2timespec(&bt, &ts);
  765                 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
  766         }
  767 #endif
  768 }
  769 
  770 /*
  771  * Timecounters need to be updated every so often to prevent the hardware
  772  * counter from overflowing.  Updating also recalculates the cached values
  773  * used by the get*() family of functions, so their precision depends on
  774  * the update frequency.
  775  */
  776 
  777 static int tc_tick;
  778 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0,
  779     "Approximate number of hardclock ticks in a millisecond");
  780 
  781 void
  782 tc_ticktock(void)
  783 {
  784         static int count;
  785         static time_t last_calib;
  786 
  787         if (++count < tc_tick)
  788                 return;
  789         count = 0;
  790         tc_windup();
  791         if (time_uptime != last_calib && !(time_uptime & 0xf)) {
  792                 cpu_tick_calibrate(0);
  793                 last_calib = time_uptime;
  794         }
  795 }
  796 
  797 static void
  798 inittimecounter(void *dummy)
  799 {
  800         u_int p;
  801 
  802         /*
  803          * Set the initial timeout to
  804          * max(1, <approx. number of hardclock ticks in a millisecond>).
  805          * People should probably not use the sysctl to set the timeout
  806          * to smaller than its inital value, since that value is the
  807          * smallest reasonable one.  If they want better timestamps they
  808          * should use the non-"get"* functions.
  809          */
  810         if (hz > 1000)
  811                 tc_tick = (hz + 500) / 1000;
  812         else
  813                 tc_tick = 1;
  814         p = (tc_tick * 1000000) / hz;
  815         printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
  816 
  817         /* warm up new timecounter (again) and get rolling. */
  818         (void)timecounter->tc_get_timecount(timecounter);
  819         (void)timecounter->tc_get_timecount(timecounter);
  820 }
  821 
  822 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL);
  823 
  824 /* Cpu tick handling -------------------------------------------------*/
  825 
  826 static int cpu_tick_variable;
  827 static uint64_t cpu_tick_frequency;
  828 
  829 static uint64_t
  830 tc_cpu_ticks(void)
  831 {
  832         static uint64_t base;
  833         static unsigned last;
  834         unsigned u;
  835         struct timecounter *tc;
  836 
  837         tc = timehands->th_counter;
  838         u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
  839         if (u < last)
  840                 base += (uint64_t)tc->tc_counter_mask + 1;
  841         last = u;
  842         return (u + base);
  843 }
  844 
  845 /*
  846  * This function gets called every 16 seconds on only one designated
  847  * CPU in the system from hardclock() via tc_ticktock().
  848  *
  849  * Whenever the real time clock is stepped we get called with reset=1
  850  * to make sure we handle suspend/resume and similar events correctly.
  851  */
  852 
  853 static void
  854 cpu_tick_calibrate(int reset)
  855 {
  856         static uint64_t c_last;
  857         uint64_t c_this, c_delta;
  858         static struct bintime  t_last;
  859         struct bintime t_this, t_delta;
  860         uint32_t divi;
  861 
  862         if (reset) {
  863                 /* The clock was stepped, abort & reset */
  864                 t_last.sec = 0;
  865                 return;
  866         }
  867 
  868         /* we don't calibrate fixed rate cputicks */
  869         if (!cpu_tick_variable)
  870                 return;
  871 
  872         getbinuptime(&t_this);
  873         c_this = cpu_ticks();
  874         if (t_last.sec != 0) {
  875                 c_delta = c_this - c_last;
  876                 t_delta = t_this;
  877                 bintime_sub(&t_delta, &t_last);
  878                 /*
  879                  * Validate that 16 +/- 1/256 seconds passed. 
  880                  * After division by 16 this gives us a precision of
  881                  * roughly 250PPM which is sufficient
  882                  */
  883                 if (t_delta.sec > 16 || (
  884                     t_delta.sec == 16 && t_delta.frac >= (0x01LL << 56))) {
  885                         /* too long */
  886                         if (bootverbose)
  887                                 printf("t_delta %ju.%016jx too long\n",
  888                                     (uintmax_t)t_delta.sec,
  889                                     (uintmax_t)t_delta.frac);
  890                 } else if (t_delta.sec < 15 ||
  891                     (t_delta.sec == 15 && t_delta.frac <= (0xffLL << 56))) {
  892                         /* too short */
  893                         if (bootverbose)
  894                                 printf("t_delta %ju.%016jx too short\n",
  895                                     (uintmax_t)t_delta.sec,
  896                                     (uintmax_t)t_delta.frac);
  897                 } else {
  898                         /* just right */
  899                         /*
  900                          * Headroom:
  901                          *      2^(64-20) / 16[s] =
  902                          *      2^(44) / 16[s] =
  903                          *      17.592.186.044.416 / 16 =
  904                          *      1.099.511.627.776 [Hz]
  905                          */
  906                         divi = t_delta.sec << 20;
  907                         divi |= t_delta.frac >> (64 - 20);
  908                         c_delta <<= 20;
  909                         c_delta /= divi;
  910                         if (c_delta  > cpu_tick_frequency) {
  911                                 if (0 && bootverbose)
  912                                         printf("cpu_tick increased to %ju Hz\n",
  913                                             c_delta);
  914                                 cpu_tick_frequency = c_delta;
  915                         }
  916                 }
  917         }
  918         c_last = c_this;
  919         t_last = t_this;
  920 }
  921 
  922 void
  923 set_cputicker(cpu_tick_f *func, uint64_t freq, unsigned var)
  924 {
  925 
  926         if (func == NULL) {
  927                 cpu_ticks = tc_cpu_ticks;
  928         } else {
  929                 cpu_tick_frequency = freq;
  930                 cpu_tick_variable = var;
  931                 cpu_ticks = func;
  932         }
  933 }
  934 
  935 uint64_t
  936 cpu_tickrate(void)
  937 {
  938 
  939         if (cpu_ticks == tc_cpu_ticks) 
  940                 return (tc_getfrequency());
  941         return (cpu_tick_frequency);
  942 }
  943 
  944 /*
  945  * We need to be slightly careful converting cputicks to microseconds.
  946  * There is plenty of margin in 64 bits of microseconds (half a million
  947  * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
  948  * before divide conversion (to retain precision) we find that the
  949  * margin shrinks to 1.5 hours (one millionth of 146y).
  950  * With a three prong approach we never lose significant bits, no
  951  * matter what the cputick rate and length of timeinterval is.
  952  */
  953 
  954 uint64_t
  955 cputick2usec(uint64_t tick)
  956 {
  957 
  958         if (tick > 18446744073709551LL)         /* floor(2^64 / 1000) */
  959                 return (tick / (cpu_tickrate() / 1000000LL));
  960         else if (tick > 18446744073709LL)       /* floor(2^64 / 1000000) */
  961                 return ((tick * 1000LL) / (cpu_tickrate() / 1000LL));
  962         else
  963                 return ((tick * 1000000LL) / cpu_tickrate());
  964 }
  965 
  966 cpu_tick_f      *cpu_ticks = tc_cpu_ticks;

Cache object: 76fa6af16fabd0b8545d4ac26c458cf0


[ 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.