1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*      Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T     */
  22 /*        All Rights Reserved   */
  23 
  24 /*
  25  * Copyright (c) 1988, 2010, Oracle and/or its affiliates. All rights reserved.
  26  * Copyright (c) 2013, Joyent, Inc.  All rights reserved.
  27  */
  28 
  29 #include <sys/param.h>
  30 #include <sys/t_lock.h>
  31 #include <sys/types.h>
  32 #include <sys/tuneable.h>
  33 #include <sys/sysmacros.h>
  34 #include <sys/systm.h>
  35 #include <sys/cpuvar.h>
  36 #include <sys/lgrp.h>
  37 #include <sys/user.h>
  38 #include <sys/proc.h>
  39 #include <sys/callo.h>
  40 #include <sys/kmem.h>
  41 #include <sys/var.h>
  42 #include <sys/cmn_err.h>
  43 #include <sys/swap.h>
  44 #include <sys/vmsystm.h>
  45 #include <sys/class.h>
  46 #include <sys/time.h>
  47 #include <sys/debug.h>
  48 #include <sys/vtrace.h>
  49 #include <sys/spl.h>
  50 #include <sys/atomic.h>
  51 #include <sys/dumphdr.h>
  52 #include <sys/archsystm.h>
  53 #include <sys/fs/swapnode.h>
  54 #include <sys/panic.h>
  55 #include <sys/disp.h>
  56 #include <sys/msacct.h>
  57 #include <sys/mem_cage.h>
  58 
  59 #include <vm/page.h>
  60 #include <vm/anon.h>
  61 #include <vm/rm.h>
  62 #include <sys/cyclic.h>
  63 #include <sys/cpupart.h>
  64 #include <sys/rctl.h>
  65 #include <sys/task.h>
  66 #include <sys/sdt.h>
  67 #include <sys/ddi_periodic.h>
  68 #include <sys/random.h>
  69 #include <sys/modctl.h>
  70 #include <sys/zone.h>
  71 
  72 /*
  73  * for NTP support
  74  */
  75 #include <sys/timex.h>
  76 #include <sys/inttypes.h>
  77 
  78 #include <sys/sunddi.h>
  79 #include <sys/clock_impl.h>
  80 
  81 /*
  82  * clock() is called straight from the clock cyclic; see clock_init().
  83  *
  84  * Functions:
  85  *      reprime clock
  86  *      maintain date
  87  *      jab the scheduler
  88  */
  89 
  90 extern kcondvar_t       fsflush_cv;
  91 extern sysinfo_t        sysinfo;
  92 extern vminfo_t vminfo;
  93 extern int      idleswtch;      /* flag set while idle in pswtch() */
  94 extern hrtime_t volatile devinfo_freeze;
  95 
  96 /*
  97  * high-precision avenrun values.  These are needed to make the
  98  * regular avenrun values accurate.
  99  */
 100 static uint64_t hp_avenrun[3];
 101 int     avenrun[3];             /* FSCALED average run queue lengths */
 102 time_t  time;   /* time in seconds since 1970 - for compatibility only */
 103 
 104 static struct loadavg_s loadavg;
 105 /*
 106  * Phase/frequency-lock loop (PLL/FLL) definitions
 107  *
 108  * The following variables are read and set by the ntp_adjtime() system
 109  * call.
 110  *
 111  * time_state shows the state of the system clock, with values defined
 112  * in the timex.h header file.
 113  *
 114  * time_status shows the status of the system clock, with bits defined
 115  * in the timex.h header file.
 116  *
 117  * time_offset is used by the PLL/FLL to adjust the system time in small
 118  * increments.
 119  *
 120  * time_constant determines the bandwidth or "stiffness" of the PLL.
 121  *
 122  * time_tolerance determines maximum frequency error or tolerance of the
 123  * CPU clock oscillator and is a property of the architecture; however,
 124  * in principle it could change as result of the presence of external
 125  * discipline signals, for instance.
 126  *
 127  * time_precision is usually equal to the kernel tick variable; however,
 128  * in cases where a precision clock counter or external clock is
 129  * available, the resolution can be much less than this and depend on
 130  * whether the external clock is working or not.
 131  *
 132  * time_maxerror is initialized by a ntp_adjtime() call and increased by
 133  * the kernel once each second to reflect the maximum error bound
 134  * growth.
 135  *
 136  * time_esterror is set and read by the ntp_adjtime() call, but
 137  * otherwise not used by the kernel.
 138  */
 139 int32_t time_state = TIME_OK;   /* clock state */
 140 int32_t time_status = STA_UNSYNC;       /* clock status bits */
 141 int32_t time_offset = 0;                /* time offset (us) */
 142 int32_t time_constant = 0;              /* pll time constant */
 143 int32_t time_tolerance = MAXFREQ;       /* frequency tolerance (scaled ppm) */
 144 int32_t time_precision = 1;     /* clock precision (us) */
 145 int32_t time_maxerror = MAXPHASE;       /* maximum error (us) */
 146 int32_t time_esterror = MAXPHASE;       /* estimated error (us) */
 147 
 148 /*
 149  * The following variables establish the state of the PLL/FLL and the
 150  * residual time and frequency offset of the local clock. The scale
 151  * factors are defined in the timex.h header file.
 152  *
 153  * time_phase and time_freq are the phase increment and the frequency
 154  * increment, respectively, of the kernel time variable.
 155  *
 156  * time_freq is set via ntp_adjtime() from a value stored in a file when
 157  * the synchronization daemon is first started. Its value is retrieved
 158  * via ntp_adjtime() and written to the file about once per hour by the
 159  * daemon.
 160  *
 161  * time_adj is the adjustment added to the value of tick at each timer
 162  * interrupt and is recomputed from time_phase and time_freq at each
 163  * seconds rollover.
 164  *
 165  * time_reftime is the second's portion of the system time at the last
 166  * call to ntp_adjtime(). It is used to adjust the time_freq variable
 167  * and to increase the time_maxerror as the time since last update
 168  * increases.
 169  */
 170 int32_t time_phase = 0;         /* phase offset (scaled us) */
 171 int32_t time_freq = 0;          /* frequency offset (scaled ppm) */
 172 int32_t time_adj = 0;           /* tick adjust (scaled 1 / hz) */
 173 int32_t time_reftime = 0;               /* time at last adjustment (s) */
 174 
 175 /*
 176  * The scale factors of the following variables are defined in the
 177  * timex.h header file.
 178  *
 179  * pps_time contains the time at each calibration interval, as read by
 180  * microtime(). pps_count counts the seconds of the calibration
 181  * interval, the duration of which is nominally pps_shift in powers of
 182  * two.
 183  *
 184  * pps_offset is the time offset produced by the time median filter
 185  * pps_tf[], while pps_jitter is the dispersion (jitter) measured by
 186  * this filter.
 187  *
 188  * pps_freq is the frequency offset produced by the frequency median
 189  * filter pps_ff[], while pps_stabil is the dispersion (wander) measured
 190  * by this filter.
 191  *
 192  * pps_usec is latched from a high resolution counter or external clock
 193  * at pps_time. Here we want the hardware counter contents only, not the
 194  * contents plus the time_tv.usec as usual.
 195  *
 196  * pps_valid counts the number of seconds since the last PPS update. It
 197  * is used as a watchdog timer to disable the PPS discipline should the
 198  * PPS signal be lost.
 199  *
 200  * pps_glitch counts the number of seconds since the beginning of an
 201  * offset burst more than tick/2 from current nominal offset. It is used
 202  * mainly to suppress error bursts due to priority conflicts between the
 203  * PPS interrupt and timer interrupt.
 204  *
 205  * pps_intcnt counts the calibration intervals for use in the interval-
 206  * adaptation algorithm. It's just too complicated for words.
 207  */
 208 struct timeval pps_time;        /* kernel time at last interval */
 209 int32_t pps_tf[] = {0, 0, 0};   /* pps time offset median filter (us) */
 210 int32_t pps_offset = 0;         /* pps time offset (us) */
 211 int32_t pps_jitter = MAXTIME;   /* time dispersion (jitter) (us) */
 212 int32_t pps_ff[] = {0, 0, 0};   /* pps frequency offset median filter */
 213 int32_t pps_freq = 0;           /* frequency offset (scaled ppm) */
 214 int32_t pps_stabil = MAXFREQ;   /* frequency dispersion (scaled ppm) */
 215 int32_t pps_usec = 0;           /* microsec counter at last interval */
 216 int32_t pps_valid = PPS_VALID;  /* pps signal watchdog counter */
 217 int32_t pps_glitch = 0;         /* pps signal glitch counter */
 218 int32_t pps_count = 0;          /* calibration interval counter (s) */
 219 int32_t pps_shift = PPS_SHIFT;  /* interval duration (s) (shift) */
 220 int32_t pps_intcnt = 0;         /* intervals at current duration */
 221 
 222 /*
 223  * PPS signal quality monitors
 224  *
 225  * pps_jitcnt counts the seconds that have been discarded because the
 226  * jitter measured by the time median filter exceeds the limit MAXTIME
 227  * (100 us).
 228  *
 229  * pps_calcnt counts the frequency calibration intervals, which are
 230  * variable from 4 s to 256 s.
 231  *
 232  * pps_errcnt counts the calibration intervals which have been discarded
 233  * because the wander exceeds the limit MAXFREQ (100 ppm) or where the
 234  * calibration interval jitter exceeds two ticks.
 235  *
 236  * pps_stbcnt counts the calibration intervals that have been discarded
 237  * because the frequency wander exceeds the limit MAXFREQ / 4 (25 us).
 238  */
 239 int32_t pps_jitcnt = 0;         /* jitter limit exceeded */
 240 int32_t pps_calcnt = 0;         /* calibration intervals */
 241 int32_t pps_errcnt = 0;         /* calibration errors */
 242 int32_t pps_stbcnt = 0;         /* stability limit exceeded */
 243 
 244 kcondvar_t lbolt_cv;
 245 
 246 /*
 247  * Hybrid lbolt implementation:
 248  *
 249  * The service historically provided by the lbolt and lbolt64 variables has
 250  * been replaced by the ddi_get_lbolt() and ddi_get_lbolt64() routines, and the
 251  * original symbols removed from the system. The once clock driven variables are
 252  * now implemented in an event driven fashion, backed by gethrtime() coarsed to
 253  * the appropriate clock resolution. The default event driven implementation is
 254  * complemented by a cyclic driven one, active only during periods of intense
 255  * activity around the DDI lbolt routines, when a lbolt specific cyclic is
 256  * reprogramed to fire at a clock tick interval to serve consumers of lbolt who
 257  * rely on the original low cost of consulting a memory position.
 258  *
 259  * The implementation uses the number of calls to these routines and the
 260  * frequency of these to determine when to transition from event to cyclic
 261  * driven and vice-versa. These values are kept on a per CPU basis for
 262  * scalability reasons and to prevent CPUs from constantly invalidating a single
 263  * cache line when modifying a global variable. The transition from event to
 264  * cyclic mode happens once the thresholds are crossed, and activity on any CPU
 265  * can cause such transition.
 266  *
 267  * The lbolt_hybrid function pointer is called by ddi_get_lbolt() and
 268  * ddi_get_lbolt64(), and will point to lbolt_event_driven() or
 269  * lbolt_cyclic_driven() according to the current mode. When the thresholds
 270  * are exceeded, lbolt_event_driven() will reprogram the lbolt cyclic to
 271  * fire at a nsec_per_tick interval and increment an internal variable at
 272  * each firing. lbolt_hybrid will then point to lbolt_cyclic_driven(), which
 273  * will simply return the value of such variable. lbolt_cyclic() will attempt
 274  * to shut itself off at each threshold interval (sampling period for calls
 275  * to the DDI lbolt routines), and return to the event driven mode, but will
 276  * be prevented from doing so if lbolt_cyclic_driven() is being heavily used.
 277  *
 278  * lbolt_bootstrap is used during boot to serve lbolt consumers who don't wait
 279  * for the cyclic subsystem to be intialized.
 280  *
 281  */
 282 int64_t lbolt_bootstrap(void);
 283 int64_t lbolt_event_driven(void);
 284 int64_t lbolt_cyclic_driven(void);
 285 int64_t (*lbolt_hybrid)(void) = lbolt_bootstrap;
 286 uint_t lbolt_ev_to_cyclic(caddr_t, caddr_t);
 287 
 288 /*
 289  * lbolt's cyclic, installed by clock_init().
 290  */
 291 static void lbolt_cyclic(void);
 292 
 293 /*
 294  * Tunable to keep lbolt in cyclic driven mode. This will prevent the system
 295  * from switching back to event driven, once it reaches cyclic mode.
 296  */
 297 static boolean_t lbolt_cyc_only = B_FALSE;
 298 
 299 /*
 300  * Cache aligned, per CPU structure with lbolt usage statistics.
 301  */
 302 static lbolt_cpu_t *lb_cpu;
 303 
 304 /*
 305  * Single, cache aligned, structure with all the information required by
 306  * the lbolt implementation.
 307  */
 308 lbolt_info_t *lb_info;
 309 
 310 
 311 int one_sec = 1; /* turned on once every second */
 312 static int fsflushcnt;  /* counter for t_fsflushr */
 313 int     dosynctodr = 1; /* patchable; enable/disable sync to TOD chip */
 314 int     tod_needsync = 0;       /* need to sync tod chip with software time */
 315 static int tod_broken = 0;      /* clock chip doesn't work */
 316 time_t  boot_time = 0;          /* Boot time in seconds since 1970 */
 317 cyclic_id_t clock_cyclic;       /* clock()'s cyclic_id */
 318 cyclic_id_t deadman_cyclic;     /* deadman()'s cyclic_id */
 319 
 320 extern void     clock_tick_schedule(int);
 321 
 322 static int lgrp_ticks;          /* counter to schedule lgrp load calcs */
 323 
 324 /*
 325  * for tod fault detection
 326  */
 327 #define TOD_REF_FREQ            ((longlong_t)(NANOSEC))
 328 #define TOD_STALL_THRESHOLD     (TOD_REF_FREQ * 3 / 2)
 329 #define TOD_JUMP_THRESHOLD      (TOD_REF_FREQ / 2)
 330 #define TOD_FILTER_N            4
 331 #define TOD_FILTER_SETTLE       (4 * TOD_FILTER_N)
 332 static enum tod_fault_type tod_faulted = TOD_NOFAULT;
 333 
 334 static int tod_status_flag = 0;         /* used by tod_validate() */
 335 
 336 static hrtime_t prev_set_tick = 0;      /* gethrtime() prior to tod_set() */
 337 static time_t prev_set_tod = 0;         /* tv_sec value passed to tod_set() */
 338 
 339 /* patchable via /etc/system */
 340 int tod_validate_enable = 1;
 341 
 342 /* Diagnose/Limit messages about delay(9F) called from interrupt context */
 343 int                     delay_from_interrupt_diagnose = 0;
 344 volatile uint32_t       delay_from_interrupt_msg = 20;
 345 
 346 /*
 347  * On non-SPARC systems, TOD validation must be deferred until gethrtime
 348  * returns non-zero values (after mach_clkinit's execution).
 349  * On SPARC systems, it must be deferred until after hrtime_base
 350  * and hres_last_tick are set (in the first invocation of hres_tick).
 351  * Since in both cases the prerequisites occur before the invocation of
 352  * tod_get() in clock(), the deferment is lifted there.
 353  */
 354 static boolean_t tod_validate_deferred = B_TRUE;
 355 
 356 /*
 357  * tod_fault_table[] must be aligned with
 358  * enum tod_fault_type in systm.h
 359  */
 360 static char *tod_fault_table[] = {
 361         "Reversed",                     /* TOD_REVERSED */
 362         "Stalled",                      /* TOD_STALLED */
 363         "Jumped",                       /* TOD_JUMPED */
 364         "Changed in Clock Rate",        /* TOD_RATECHANGED */
 365         "Is Read-Only"                  /* TOD_RDONLY */
 366         /*
 367          * no strings needed for TOD_NOFAULT
 368          */
 369 };
 370 
 371 /*
 372  * test hook for tod broken detection in tod_validate
 373  */
 374 int tod_unit_test = 0;
 375 time_t tod_test_injector;
 376 
 377 #define CLOCK_ADJ_HIST_SIZE     4
 378 
 379 static int      adj_hist_entry;
 380 
 381 int64_t clock_adj_hist[CLOCK_ADJ_HIST_SIZE];
 382 
 383 static void calcloadavg(int, uint64_t *);
 384 static int genloadavg(struct loadavg_s *);
 385 static void loadavg_update();
 386 
 387 void (*cmm_clock_callout)() = NULL;
 388 void (*cpucaps_clock_callout)() = NULL;
 389 
 390 extern clock_t clock_tick_proc_max;
 391 
 392 static int64_t deadman_counter = 0;
 393 
 394 static void
 395 clock(void)
 396 {
 397         kthread_t       *t;
 398         uint_t  nrunnable;
 399         uint_t  w_io;
 400         cpu_t   *cp;
 401         cpupart_t *cpupart;
 402         extern  void    set_freemem();
 403         void    (*funcp)();
 404         int32_t ltemp;
 405         int64_t lltemp;
 406         int s;
 407         int do_lgrp_load;
 408         int i;
 409         clock_t now = LBOLT_NO_ACCOUNT; /* current tick */
 410 
 411         if (panicstr)
 412                 return;
 413 
 414         /*
 415          * Make sure that 'freemem' do not drift too far from the truth
 416          */
 417         set_freemem();
 418 
 419 
 420         /*
 421          * Before the section which is repeated is executed, we do
 422          * the time delta processing which occurs every clock tick
 423          *
 424          * There is additional processing which happens every time
 425          * the nanosecond counter rolls over which is described
 426          * below - see the section which begins with : if (one_sec)
 427          *
 428          * This section marks the beginning of the precision-kernel
 429          * code fragment.
 430          *
 431          * First, compute the phase adjustment. If the low-order bits
 432          * (time_phase) of the update overflow, bump the higher order
 433          * bits (time_update).
 434          */
 435         time_phase += time_adj;
 436         if (time_phase <= -FINEUSEC) {
 437                 ltemp = -time_phase / SCALE_PHASE;
 438                 time_phase += ltemp * SCALE_PHASE;
 439                 s = hr_clock_lock();
 440                 timedelta -= ltemp * (NANOSEC/MICROSEC);
 441                 hr_clock_unlock(s);
 442         } else if (time_phase >= FINEUSEC) {
 443                 ltemp = time_phase / SCALE_PHASE;
 444                 time_phase -= ltemp * SCALE_PHASE;
 445                 s = hr_clock_lock();
 446                 timedelta += ltemp * (NANOSEC/MICROSEC);
 447                 hr_clock_unlock(s);
 448         }
 449 
 450         /*
 451          * End of precision-kernel code fragment which is processed
 452          * every timer interrupt.
 453          *
 454          * Continue with the interrupt processing as scheduled.
 455          */
 456         /*
 457          * Count the number of runnable threads and the number waiting
 458          * for some form of I/O to complete -- gets added to
 459          * sysinfo.waiting.  To know the state of the system, must add
 460          * wait counts from all CPUs.  Also add up the per-partition
 461          * statistics.
 462          */
 463         w_io = 0;
 464         nrunnable = 0;
 465 
 466         /*
 467          * keep track of when to update lgrp/part loads
 468          */
 469 
 470         do_lgrp_load = 0;
 471         if (lgrp_ticks++ >= hz / 10) {
 472                 lgrp_ticks = 0;
 473                 do_lgrp_load = 1;
 474         }
 475 
 476         if (one_sec) {
 477                 loadavg_update();
 478                 deadman_counter++;
 479         }
 480 
 481         /*
 482          * First count the threads waiting on kpreempt queues in each
 483          * CPU partition.
 484          */
 485 
 486         cpupart = cp_list_head;
 487         do {
 488                 uint_t cpupart_nrunnable = cpupart->cp_kp_queue.disp_nrunnable;
 489 
 490                 cpupart->cp_updates++;
 491                 nrunnable += cpupart_nrunnable;
 492                 cpupart->cp_nrunnable_cum += cpupart_nrunnable;
 493                 if (one_sec) {
 494                         cpupart->cp_nrunning = 0;
 495                         cpupart->cp_nrunnable = cpupart_nrunnable;
 496                 }
 497         } while ((cpupart = cpupart->cp_next) != cp_list_head);
 498 
 499 
 500         /* Now count the per-CPU statistics. */
 501         cp = cpu_list;
 502         do {
 503                 uint_t cpu_nrunnable = cp->cpu_disp->disp_nrunnable;
 504 
 505                 nrunnable += cpu_nrunnable;
 506                 cpupart = cp->cpu_part;
 507                 cpupart->cp_nrunnable_cum += cpu_nrunnable;
 508                 if (one_sec) {
 509                         cpupart->cp_nrunnable += cpu_nrunnable;
 510                         /*
 511                          * Update user, system, and idle cpu times.
 512                          */
 513                         cpupart->cp_nrunning++;
 514                         /*
 515                          * w_io is used to update sysinfo.waiting during
 516                          * one_second processing below.  Only gather w_io
 517                          * information when we walk the list of cpus if we're
 518                          * going to perform one_second processing.
 519                          */
 520                         w_io += CPU_STATS(cp, sys.iowait);
 521                 }
 522 
 523                 if (one_sec && (cp->cpu_flags & CPU_EXISTS)) {
 524                         int i, load, change;
 525                         hrtime_t intracct, intrused;
 526                         const hrtime_t maxnsec = 1000000000;
 527                         const int precision = 100;
 528 
 529                         /*
 530                          * Estimate interrupt load on this cpu each second.
 531                          * Computes cpu_intrload as %utilization (0-99).
 532                          */
 533 
 534                         /* add up interrupt time from all micro states */
 535                         for (intracct = 0, i = 0; i < NCMSTATES; i++)
 536                                 intracct += cp->cpu_intracct[i];
 537                         scalehrtime(&intracct);
 538 
 539                         /* compute nsec used in the past second */
 540                         intrused = intracct - cp->cpu_intrlast;
 541                         cp->cpu_intrlast = intracct;
 542 
 543                         /* limit the value for safety (and the first pass) */
 544                         if (intrused >= maxnsec)
 545                                 intrused = maxnsec - 1;
 546 
 547                         /* calculate %time in interrupt */
 548                         load = (precision * intrused) / maxnsec;
 549                         ASSERT(load >= 0 && load < precision);
 550                         change = cp->cpu_intrload - load;
 551 
 552                         /* jump to new max, or decay the old max */
 553                         if (change < 0)
 554                                 cp->cpu_intrload = load;
 555                         else if (change > 0)
 556                                 cp->cpu_intrload -= (change + 3) / 4;
 557 
 558                         DTRACE_PROBE3(cpu_intrload,
 559                             cpu_t *, cp,
 560                             hrtime_t, intracct,
 561                             hrtime_t, intrused);
 562                 }
 563 
 564                 if (do_lgrp_load &&
 565                     (cp->cpu_flags & CPU_EXISTS)) {
 566                         /*
 567                          * When updating the lgroup's load average,
 568                          * account for the thread running on the CPU.
 569                          * If the CPU is the current one, then we need
 570                          * to account for the underlying thread which
 571                          * got the clock interrupt not the thread that is
 572                          * handling the interrupt and caculating the load
 573                          * average
 574                          */
 575                         t = cp->cpu_thread;
 576                         if (CPU == cp)
 577                                 t = t->t_intr;
 578 
 579                         /*
 580                          * Account for the load average for this thread if
 581                          * it isn't the idle thread or it is on the interrupt
 582                          * stack and not the current CPU handling the clock
 583                          * interrupt
 584                          */
 585                         if ((t && t != cp->cpu_idle_thread) || (CPU != cp &&
 586                             CPU_ON_INTR(cp))) {
 587                                 if (t->t_lpl == cp->cpu_lpl) {
 588                                         /* local thread */
 589                                         cpu_nrunnable++;
 590                                 } else {
 591                                         /*
 592                                          * This is a remote thread, charge it
 593                                          * against its home lgroup.  Note that
 594                                          * we notice that a thread is remote
 595                                          * only if it's currently executing.
 596                                          * This is a reasonable approximation,
 597                                          * since queued remote threads are rare.
 598                                          * Note also that if we didn't charge
 599                                          * it to its home lgroup, remote
 600                                          * execution would often make a system
 601                                          * appear balanced even though it was
 602                                          * not, and thread placement/migration
 603                                          * would often not be done correctly.
 604                                          */
 605                                         lgrp_loadavg(t->t_lpl,
 606                                             LGRP_LOADAVG_IN_THREAD_MAX, 0);
 607                                 }
 608                         }
 609                         lgrp_loadavg(cp->cpu_lpl,
 610                             cpu_nrunnable * LGRP_LOADAVG_IN_THREAD_MAX, 1);
 611                 }
 612         } while ((cp = cp->cpu_next) != cpu_list);
 613 
 614         clock_tick_schedule(one_sec);
 615 
 616         /*
 617          * Check for a callout that needs be called from the clock
 618          * thread to support the membership protocol in a clustered
 619          * system.  Copy the function pointer so that we can reset
 620          * this to NULL if needed.
 621          */
 622         if ((funcp = cmm_clock_callout) != NULL)
 623                 (*funcp)();
 624 
 625         if ((funcp = cpucaps_clock_callout) != NULL)
 626                 (*funcp)();
 627 
 628         /*
 629          * Wakeup the cageout thread waiters once per second.
 630          */
 631         if (one_sec)
 632                 kcage_tick();
 633 
 634         if (one_sec) {
 635 
 636                 int drift, absdrift;
 637                 timestruc_t tod;
 638                 int s;
 639 
 640                 /*
 641                  * Beginning of precision-kernel code fragment executed
 642                  * every second.
 643                  *
 644                  * On rollover of the second the phase adjustment to be
 645                  * used for the next second is calculated.  Also, the
 646                  * maximum error is increased by the tolerance.  If the
 647                  * PPS frequency discipline code is present, the phase is
 648                  * increased to compensate for the CPU clock oscillator
 649                  * frequency error.
 650                  *
 651                  * On a 32-bit machine and given parameters in the timex.h
 652                  * header file, the maximum phase adjustment is +-512 ms
 653                  * and maximum frequency offset is (a tad less than)
 654                  * +-512 ppm. On a 64-bit machine, you shouldn't need to ask.
 655                  */
 656                 time_maxerror += time_tolerance / SCALE_USEC;
 657 
 658                 /*
 659                  * Leap second processing. If in leap-insert state at
 660                  * the end of the day, the system clock is set back one
 661                  * second; if in leap-delete state, the system clock is
 662                  * set ahead one second. The microtime() routine or
 663                  * external clock driver will insure that reported time
 664                  * is always monotonic. The ugly divides should be
 665                  * replaced.
 666                  */
 667                 switch (time_state) {
 668 
 669                 case TIME_OK:
 670                         if (time_status & STA_INS)
 671                                 time_state = TIME_INS;
 672                         else if (time_status & STA_DEL)
 673                                 time_state = TIME_DEL;
 674                         break;
 675 
 676                 case TIME_INS:
 677                         if (hrestime.tv_sec % 86400 == 0) {
 678                                 s = hr_clock_lock();
 679                                 hrestime.tv_sec--;
 680                                 hr_clock_unlock(s);
 681                                 time_state = TIME_OOP;
 682                         }
 683                         break;
 684 
 685                 case TIME_DEL:
 686                         if ((hrestime.tv_sec + 1) % 86400 == 0) {
 687                                 s = hr_clock_lock();
 688                                 hrestime.tv_sec++;
 689                                 hr_clock_unlock(s);
 690                                 time_state = TIME_WAIT;
 691                         }
 692                         break;
 693 
 694                 case TIME_OOP:
 695                         time_state = TIME_WAIT;
 696                         break;
 697 
 698                 case TIME_WAIT:
 699                         if (!(time_status & (STA_INS | STA_DEL)))
 700                                 time_state = TIME_OK;
 701                 default:
 702                         break;
 703                 }
 704 
 705                 /*
 706                  * Compute the phase adjustment for the next second. In
 707                  * PLL mode, the offset is reduced by a fixed factor
 708                  * times the time constant. In FLL mode the offset is
 709                  * used directly. In either mode, the maximum phase
 710                  * adjustment for each second is clamped so as to spread
 711                  * the adjustment over not more than the number of
 712                  * seconds between updates.
 713                  */
 714                 if (time_offset == 0)
 715                         time_adj = 0;
 716                 else if (time_offset < 0) {
 717                         lltemp = -time_offset;
 718                         if (!(time_status & STA_FLL)) {
 719                                 if ((1 << time_constant) >= SCALE_KG)
 720                                         lltemp *= (1 << time_constant) /
 721                                             SCALE_KG;
 722                                 else
 723                                         lltemp = (lltemp / SCALE_KG) >>
 724                                             time_constant;
 725                         }
 726                         if (lltemp > (MAXPHASE / MINSEC) * SCALE_UPDATE)
 727                                 lltemp = (MAXPHASE / MINSEC) * SCALE_UPDATE;
 728                         time_offset += lltemp;
 729                         time_adj = -(lltemp * SCALE_PHASE) / hz / SCALE_UPDATE;
 730                 } else {
 731                         lltemp = time_offset;
 732                         if (!(time_status & STA_FLL)) {
 733                                 if ((1 << time_constant) >= SCALE_KG)
 734                                         lltemp *= (1 << time_constant) /
 735                                             SCALE_KG;
 736                                 else
 737                                         lltemp = (lltemp / SCALE_KG) >>
 738                                             time_constant;
 739                         }
 740                         if (lltemp > (MAXPHASE / MINSEC) * SCALE_UPDATE)
 741                                 lltemp = (MAXPHASE / MINSEC) * SCALE_UPDATE;
 742                         time_offset -= lltemp;
 743                         time_adj = (lltemp * SCALE_PHASE) / hz / SCALE_UPDATE;
 744                 }
 745 
 746                 /*
 747                  * Compute the frequency estimate and additional phase
 748                  * adjustment due to frequency error for the next
 749                  * second. When the PPS signal is engaged, gnaw on the
 750                  * watchdog counter and update the frequency computed by
 751                  * the pll and the PPS signal.
 752                  */
 753                 pps_valid++;
 754                 if (pps_valid == PPS_VALID) {
 755                         pps_jitter = MAXTIME;
 756                         pps_stabil = MAXFREQ;
 757                         time_status &= ~(STA_PPSSIGNAL | STA_PPSJITTER |
 758                             STA_PPSWANDER | STA_PPSERROR);
 759                 }
 760                 lltemp = time_freq + pps_freq;
 761 
 762                 if (lltemp)
 763                         time_adj += (lltemp * SCALE_PHASE) / (SCALE_USEC * hz);
 764 
 765                 /*
 766                  * End of precision kernel-code fragment
 767                  *
 768                  * The section below should be modified if we are planning
 769                  * to use NTP for synchronization.
 770                  *
 771                  * Note: the clock synchronization code now assumes
 772                  * the following:
 773                  *   - if dosynctodr is 1, then compute the drift between
 774                  *      the tod chip and software time and adjust one or
 775                  *      the other depending on the circumstances
 776                  *
 777                  *   - if dosynctodr is 0, then the tod chip is independent
 778                  *      of the software clock and should not be adjusted,
 779                  *      but allowed to free run.  this allows NTP to sync.
 780                  *      hrestime without any interference from the tod chip.
 781                  */
 782 
 783                 tod_validate_deferred = B_FALSE;
 784                 mutex_enter(&tod_lock);
 785                 tod = tod_get();
 786                 drift = tod.tv_sec - hrestime.tv_sec;
 787                 absdrift = (drift >= 0) ? drift : -drift;
 788                 if (tod_needsync || absdrift > 1) {
 789                         int s;
 790                         if (absdrift > 2) {
 791                                 if (!tod_broken && tod_faulted == TOD_NOFAULT) {
 792                                         s = hr_clock_lock();
 793                                         hrestime = tod;
 794                                         membar_enter(); /* hrestime visible */
 795                                         timedelta = 0;
 796                                         timechanged++;
 797                                         tod_needsync = 0;
 798                                         hr_clock_unlock(s);
 799                                         callout_hrestime();
 800 
 801                                 }
 802                         } else {
 803                                 if (tod_needsync || !dosynctodr) {
 804                                         gethrestime(&tod);
 805                                         tod_set(tod);
 806                                         s = hr_clock_lock();
 807                                         if (timedelta == 0)
 808                                                 tod_needsync = 0;
 809                                         hr_clock_unlock(s);
 810                                 } else {
 811                                         /*
 812                                          * If the drift is 2 seconds on the
 813                                          * money, then the TOD is adjusting
 814                                          * the clock;  record that.
 815                                          */
 816                                         clock_adj_hist[adj_hist_entry++ %
 817                                             CLOCK_ADJ_HIST_SIZE] = now;
 818                                         s = hr_clock_lock();
 819                                         timedelta = (int64_t)drift*NANOSEC;
 820                                         hr_clock_unlock(s);
 821                                 }
 822                         }
 823                 }
 824                 one_sec = 0;
 825                 time = gethrestime_sec();  /* for crusty old kmem readers */
 826                 mutex_exit(&tod_lock);
 827 
 828                 /*
 829                  * Some drivers still depend on this... XXX
 830                  */
 831                 cv_broadcast(&lbolt_cv);
 832 
 833                 vminfo.freemem += freemem;
 834                 {
 835                         pgcnt_t maxswap, resv, free;
 836                         pgcnt_t avail =
 837                             MAX((spgcnt_t)(availrmem - swapfs_minfree), 0);
 838 
 839                         maxswap = k_anoninfo.ani_mem_resv +
 840                             k_anoninfo.ani_max +avail;
 841                         /* Update ani_free */
 842                         set_anoninfo();
 843                         free = k_anoninfo.ani_free + avail;
 844                         resv = k_anoninfo.ani_phys_resv +
 845                             k_anoninfo.ani_mem_resv;
 846 
 847                         vminfo.swap_resv += resv;
 848                         /* number of reserved and allocated pages */
 849 #ifdef  DEBUG
 850                         if (maxswap < free)
 851                                 cmn_err(CE_WARN, "clock: maxswap < free");
 852                         if (maxswap < resv)
 853                                 cmn_err(CE_WARN, "clock: maxswap < resv");
 854 #endif
 855                         vminfo.swap_alloc += maxswap - free;
 856                         vminfo.swap_avail += maxswap - resv;
 857                         vminfo.swap_free += free;
 858                 }
 859                 vminfo.updates++;
 860                 if (nrunnable) {
 861                         sysinfo.runque += nrunnable;
 862                         sysinfo.runocc++;
 863                 }
 864                 if (nswapped) {
 865                         sysinfo.swpque += nswapped;
 866                         sysinfo.swpocc++;
 867                 }
 868                 sysinfo.waiting += w_io;
 869                 sysinfo.updates++;
 870 
 871                 /*
 872                  * Wake up fsflush to write out DELWRI
 873                  * buffers, dirty pages and other cached
 874                  * administrative data, e.g. inodes.
 875                  */
 876                 if (--fsflushcnt <= 0) {
 877                         fsflushcnt = tune.t_fsflushr;
 878                         cv_signal(&fsflush_cv);
 879                 }
 880 
 881                 vmmeter();
 882                 calcloadavg(genloadavg(&loadavg), hp_avenrun);
 883                 for (i = 0; i < 3; i++)
 884                         /*
 885                          * At the moment avenrun[] can only hold 31
 886                          * bits of load average as it is a signed
 887                          * int in the API. We need to ensure that
 888                          * hp_avenrun[i] >> (16 - FSHIFT) will not be
 889                          * too large. If it is, we put the largest value
 890                          * that we can use into avenrun[i]. This is
 891                          * kludgey, but about all we can do until we
 892                          * avenrun[] is declared as an array of uint64[]
 893                          */
 894                         if (hp_avenrun[i] < ((uint64_t)1<<(31+16-FSHIFT)))
 895                                 avenrun[i] = (int32_t)(hp_avenrun[i] >>
 896                                     (16 - FSHIFT));
 897                         else
 898                                 avenrun[i] = 0x7fffffff;
 899 
 900                 cpupart = cp_list_head;
 901                 do {
 902                         calcloadavg(genloadavg(&cpupart->cp_loadavg),
 903                             cpupart->cp_hp_avenrun);
 904                 } while ((cpupart = cpupart->cp_next) != cp_list_head);
 905         }
 906 }
 907 
 908 void
 909 clock_init(void)
 910 {
 911         cyc_handler_t clk_hdlr, lbolt_hdlr;
 912         cyc_time_t clk_when, lbolt_when;
 913         int i, sz;
 914         intptr_t buf;
 915 
 916         /*
 917          * Setup handler and timer for the clock cyclic.
 918          */
 919         clk_hdlr.cyh_func = (cyc_func_t)clock;
 920         clk_hdlr.cyh_level = CY_LOCK_LEVEL;
 921         clk_hdlr.cyh_arg = NULL;
 922 
 923         clk_when.cyt_when = 0;
 924         clk_when.cyt_interval = nsec_per_tick;
 925 
 926         /*
 927          * The lbolt cyclic will be reprogramed to fire at a nsec_per_tick
 928          * interval to satisfy performance needs of the DDI lbolt consumers.
 929          * It is off by default.
 930          */
 931         lbolt_hdlr.cyh_func = (cyc_func_t)lbolt_cyclic;
 932         lbolt_hdlr.cyh_level = CY_LOCK_LEVEL;
 933         lbolt_hdlr.cyh_arg = NULL;
 934 
 935         lbolt_when.cyt_interval = nsec_per_tick;
 936 
 937         /*
 938          * Allocate cache line aligned space for the per CPU lbolt data and
 939          * lbolt info structures, and initialize them with their default
 940          * values. Note that these structures are also cache line sized.
 941          */
 942         sz = sizeof (lbolt_info_t) + CPU_CACHE_COHERENCE_SIZE;
 943         buf = (intptr_t)kmem_zalloc(sz, KM_SLEEP);
 944         lb_info = (lbolt_info_t *)P2ROUNDUP(buf, CPU_CACHE_COHERENCE_SIZE);
 945 
 946         if (hz != HZ_DEFAULT)
 947                 lb_info->lbi_thresh_interval = LBOLT_THRESH_INTERVAL *
 948                     hz/HZ_DEFAULT;
 949         else
 950                 lb_info->lbi_thresh_interval = LBOLT_THRESH_INTERVAL;
 951 
 952         lb_info->lbi_thresh_calls = LBOLT_THRESH_CALLS;
 953 
 954         sz = (sizeof (lbolt_cpu_t) * max_ncpus) + CPU_CACHE_COHERENCE_SIZE;
 955         buf = (intptr_t)kmem_zalloc(sz, KM_SLEEP);
 956         lb_cpu = (lbolt_cpu_t *)P2ROUNDUP(buf, CPU_CACHE_COHERENCE_SIZE);
 957 
 958         for (i = 0; i < max_ncpus; i++)
 959                 lb_cpu[i].lbc_counter = lb_info->lbi_thresh_calls;
 960 
 961         /*
 962          * Install the softint used to switch between event and cyclic driven
 963          * lbolt. We use a soft interrupt to make sure the context of the
 964          * cyclic reprogram call is safe.
 965          */
 966         lbolt_softint_add();
 967 
 968         /*
 969          * Since the hybrid lbolt implementation is based on a hardware counter
 970          * that is reset at every hardware reboot and that we'd like to have
 971          * the lbolt value starting at zero after both a hardware and a fast
 972          * reboot, we calculate the number of clock ticks the system's been up
 973          * and store it in the lbi_debug_time field of the lbolt info structure.
 974          * The value of this field will be subtracted from lbolt before
 975          * returning it.
 976          */
 977         lb_info->lbi_internal = lb_info->lbi_debug_time =
 978             (gethrtime()/nsec_per_tick);
 979 
 980         /*
 981          * lbolt_hybrid points at lbolt_bootstrap until now. The LBOLT_* macros
 982          * and lbolt_debug_{enter,return} use this value as an indication that
 983          * the initializaion above hasn't been completed. Setting lbolt_hybrid
 984          * to either lbolt_{cyclic,event}_driven here signals those code paths
 985          * that the lbolt related structures can be used.
 986          */
 987         if (lbolt_cyc_only) {
 988                 lbolt_when.cyt_when = 0;
 989                 lbolt_hybrid = lbolt_cyclic_driven;
 990         } else {
 991                 lbolt_when.cyt_when = CY_INFINITY;
 992                 lbolt_hybrid = lbolt_event_driven;
 993         }
 994 
 995         /*
 996          * Grab cpu_lock and install all three cyclics.
 997          */
 998         mutex_enter(&cpu_lock);
 999 
1000         clock_cyclic = cyclic_add(&clk_hdlr, &clk_when);
1001         lb_info->id.lbi_cyclic_id = cyclic_add(&lbolt_hdlr, &lbolt_when);
1002 
1003         mutex_exit(&cpu_lock);
1004 }
1005 
1006 /*
1007  * Called before calcloadavg to get 10-sec moving loadavg together
1008  */
1009 
1010 static int
1011 genloadavg(struct loadavg_s *avgs)
1012 {
1013         int avg;
1014         int spos; /* starting position */
1015         int cpos; /* moving current position */
1016         int i;
1017         int slen;
1018         hrtime_t hr_avg;
1019 
1020         /* 10-second snapshot, calculate first positon */
1021         if (avgs->lg_len == 0) {
1022                 return (0);
1023         }
1024         slen = avgs->lg_len < S_MOVAVG_SZ ? avgs->lg_len : S_MOVAVG_SZ;
1025 
1026         spos = (avgs->lg_cur - 1) >= 0 ? avgs->lg_cur - 1 :
1027             S_LOADAVG_SZ + (avgs->lg_cur - 1);
1028         for (i = hr_avg = 0; i < slen; i++) {
1029                 cpos = (spos - i) >= 0 ? spos - i : S_LOADAVG_SZ + (spos - i);
1030                 hr_avg += avgs->lg_loads[cpos];
1031         }
1032 
1033         hr_avg = hr_avg / slen;
1034         avg = hr_avg / (NANOSEC / LGRP_LOADAVG_IN_THREAD_MAX);
1035 
1036         return (avg);
1037 }
1038 
1039 /*
1040  * Run every second from clock () to update the loadavg count available to the
1041  * system and cpu-partitions.
1042  *
1043  * This works by sampling the previous usr, sys, wait time elapsed,
1044  * computing a delta, and adding that delta to the elapsed usr, sys,
1045  * wait increase.
1046  */
1047 
1048 static void
1049 loadavg_update()
1050 {
1051         cpu_t *cp;
1052         cpupart_t *cpupart;
1053         hrtime_t cpu_total;
1054         int prev;
1055 
1056         cp = cpu_list;
1057         loadavg.lg_total = 0;
1058 
1059         /*
1060          * first pass totals up per-cpu statistics for system and cpu
1061          * partitions
1062          */
1063 
1064         do {
1065                 struct loadavg_s *lavg;
1066 
1067                 lavg = &cp->cpu_loadavg;
1068 
1069                 cpu_total = cp->cpu_acct[CMS_USER] +
1070                     cp->cpu_acct[CMS_SYSTEM] + cp->cpu_waitrq;
1071                 /* compute delta against last total */
1072                 scalehrtime(&cpu_total);
1073                 prev = (lavg->lg_cur - 1) >= 0 ? lavg->lg_cur - 1 :
1074                     S_LOADAVG_SZ + (lavg->lg_cur - 1);
1075                 if (lavg->lg_loads[prev] <= 0) {
1076                         lavg->lg_loads[lavg->lg_cur] = cpu_total;
1077                         cpu_total = 0;
1078                 } else {
1079                         lavg->lg_loads[lavg->lg_cur] = cpu_total;
1080                         cpu_total = cpu_total - lavg->lg_loads[prev];
1081                         if (cpu_total < 0)
1082                                 cpu_total = 0;
1083                 }
1084 
1085                 lavg->lg_cur = (lavg->lg_cur + 1) % S_LOADAVG_SZ;
1086                 lavg->lg_len = (lavg->lg_len + 1) < S_LOADAVG_SZ ?
1087                     lavg->lg_len + 1 : S_LOADAVG_SZ;
1088 
1089                 loadavg.lg_total += cpu_total;
1090                 cp->cpu_part->cp_loadavg.lg_total += cpu_total;
1091 
1092         } while ((cp = cp->cpu_next) != cpu_list);
1093 
1094         loadavg.lg_loads[loadavg.lg_cur] = loadavg.lg_total;
1095         loadavg.lg_cur = (loadavg.lg_cur + 1) % S_LOADAVG_SZ;
1096         loadavg.lg_len = (loadavg.lg_len + 1) < S_LOADAVG_SZ ?
1097             loadavg.lg_len + 1 : S_LOADAVG_SZ;
1098         /*
1099          * Second pass updates counts
1100          */
1101         cpupart = cp_list_head;
1102 
1103         do {
1104                 struct loadavg_s *lavg;
1105 
1106                 lavg = &cpupart->cp_loadavg;
1107                 lavg->lg_loads[lavg->lg_cur] = lavg->lg_total;
1108                 lavg->lg_total = 0;
1109                 lavg->lg_cur = (lavg->lg_cur + 1) % S_LOADAVG_SZ;
1110                 lavg->lg_len = (lavg->lg_len + 1) < S_LOADAVG_SZ ?
1111                     lavg->lg_len + 1 : S_LOADAVG_SZ;
1112 
1113         } while ((cpupart = cpupart->cp_next) != cp_list_head);
1114 
1115         /*
1116          * Third pass totals up per-zone statistics.
1117          */
1118         zone_loadavg_update();
1119 }
1120 
1121 /*
1122  * clock_update() - local clock update
1123  *
1124  * This routine is called by ntp_adjtime() to update the local clock
1125  * phase and frequency. The implementation is of an
1126  * adaptive-parameter, hybrid phase/frequency-lock loop (PLL/FLL). The
1127  * routine computes new time and frequency offset estimates for each
1128  * call.  The PPS signal itself determines the new time offset,
1129  * instead of the calling argument.  Presumably, calls to
1130  * ntp_adjtime() occur only when the caller believes the local clock
1131  * is valid within some bound (+-128 ms with NTP). If the caller's
1132  * time is far different than the PPS time, an argument will ensue,
1133  * and it's not clear who will lose.
1134  *
1135  * For uncompensated quartz crystal oscillatores and nominal update
1136  * intervals less than 1024 s, operation should be in phase-lock mode
1137  * (STA_FLL = 0), where the loop is disciplined to phase. For update
1138  * intervals greater than this, operation should be in frequency-lock
1139  * mode (STA_FLL = 1), where the loop is disciplined to frequency.
1140  *
1141  * Note: mutex(&tod_lock) is in effect.
1142  */
1143 void
1144 clock_update(int offset)
1145 {
1146         int ltemp, mtemp, s;
1147 
1148         ASSERT(MUTEX_HELD(&tod_lock));
1149 
1150         if (!(time_status & STA_PLL) && !(time_status & STA_PPSTIME))
1151                 return;
1152         ltemp = offset;
1153         if ((time_status & STA_PPSTIME) && (time_status & STA_PPSSIGNAL))
1154                 ltemp = pps_offset;
1155 
1156         /*
1157          * Scale the phase adjustment and clamp to the operating range.
1158          */
1159         if (ltemp > MAXPHASE)
1160                 time_offset = MAXPHASE * SCALE_UPDATE;
1161         else if (ltemp < -MAXPHASE)
1162                 time_offset = -(MAXPHASE * SCALE_UPDATE);
1163         else
1164                 time_offset = ltemp * SCALE_UPDATE;
1165 
1166         /*
1167          * Select whether the frequency is to be controlled and in which
1168          * mode (PLL or FLL). Clamp to the operating range. Ugly
1169          * multiply/divide should be replaced someday.
1170          */
1171         if (time_status & STA_FREQHOLD || time_reftime == 0)
1172                 time_reftime = hrestime.tv_sec;
1173 
1174         mtemp = hrestime.tv_sec - time_reftime;
1175         time_reftime = hrestime.tv_sec;
1176 
1177         if (time_status & STA_FLL) {
1178                 if (mtemp >= MINSEC) {
1179                         ltemp = ((time_offset / mtemp) * (SCALE_USEC /
1180                             SCALE_UPDATE));
1181                         if (ltemp)
1182                                 time_freq += ltemp / SCALE_KH;
1183                 }
1184         } else {
1185                 if (mtemp < MAXSEC) {
1186                         ltemp *= mtemp;
1187                         if (ltemp)
1188                                 time_freq += (int)(((int64_t)ltemp *
1189                                     SCALE_USEC) / SCALE_KF)
1190                                     / (1 << (time_constant * 2));
1191                 }
1192         }
1193         if (time_freq > time_tolerance)
1194                 time_freq = time_tolerance;
1195         else if (time_freq < -time_tolerance)
1196                 time_freq = -time_tolerance;
1197 
1198         s = hr_clock_lock();
1199         tod_needsync = 1;
1200         hr_clock_unlock(s);
1201 }
1202 
1203 /*
1204  * ddi_hardpps() - discipline CPU clock oscillator to external PPS signal
1205  *
1206  * This routine is called at each PPS interrupt in order to discipline
1207  * the CPU clock oscillator to the PPS signal. It measures the PPS phase
1208  * and leaves it in a handy spot for the clock() routine. It
1209  * integrates successive PPS phase differences and calculates the
1210  * frequency offset. This is used in clock() to discipline the CPU
1211  * clock oscillator so that intrinsic frequency error is cancelled out.
1212  * The code requires the caller to capture the time and hardware counter
1213  * value at the on-time PPS signal transition.
1214  *
1215  * Note that, on some Unix systems, this routine runs at an interrupt
1216  * priority level higher than the timer interrupt routine clock().
1217  * Therefore, the variables used are distinct from the clock()
1218  * variables, except for certain exceptions: The PPS frequency pps_freq
1219  * and phase pps_offset variables are determined by this routine and
1220  * updated atomically. The time_tolerance variable can be considered a
1221  * constant, since it is infrequently changed, and then only when the
1222  * PPS signal is disabled. The watchdog counter pps_valid is updated
1223  * once per second by clock() and is atomically cleared in this
1224  * routine.
1225  *
1226  * tvp is the time of the last tick; usec is a microsecond count since the
1227  * last tick.
1228  *
1229  * Note: In Solaris systems, the tick value is actually given by
1230  *       usec_per_tick.  This is called from the serial driver cdintr(),
1231  *       or equivalent, at a high PIL.  Because the kernel keeps a
1232  *       highresolution time, the following code can accept either
1233  *       the traditional argument pair, or the current highres timestamp
1234  *       in tvp and zero in usec.
1235  */
1236 void
1237 ddi_hardpps(struct timeval *tvp, int usec)
1238 {
1239         int u_usec, v_usec, bigtick;
1240         time_t cal_sec;
1241         int cal_usec;
1242 
1243         /*
1244          * An occasional glitch can be produced when the PPS interrupt
1245          * occurs in the clock() routine before the time variable is
1246          * updated. Here the offset is discarded when the difference
1247          * between it and the last one is greater than tick/2, but not
1248          * if the interval since the first discard exceeds 30 s.
1249          */
1250         time_status |= STA_PPSSIGNAL;
1251         time_status &= ~(STA_PPSJITTER | STA_PPSWANDER | STA_PPSERROR);
1252         pps_valid = 0;
1253         u_usec = -tvp->tv_usec;
1254         if (u_usec < -(MICROSEC/2))
1255                 u_usec += MICROSEC;
1256         v_usec = pps_offset - u_usec;
1257         if (v_usec < 0)
1258                 v_usec = -v_usec;
1259         if (v_usec > (usec_per_tick >> 1)) {
1260                 if (pps_glitch > MAXGLITCH) {
1261                         pps_glitch = 0;
1262                         pps_tf[2] = u_usec;
1263                         pps_tf[1] = u_usec;
1264                 } else {
1265                         pps_glitch++;
1266                         u_usec = pps_offset;
1267                 }
1268         } else
1269                 pps_glitch = 0;
1270 
1271         /*
1272          * A three-stage median filter is used to help deglitch the pps
1273          * time. The median sample becomes the time offset estimate; the
1274          * difference between the other two samples becomes the time
1275          * dispersion (jitter) estimate.
1276          */
1277         pps_tf[2] = pps_tf[1];
1278         pps_tf[1] = pps_tf[0];
1279         pps_tf[0] = u_usec;
1280         if (pps_tf[0] > pps_tf[1]) {
1281                 if (pps_tf[1] > pps_tf[2]) {
1282                         pps_offset = pps_tf[1];         /* 0 1 2 */
1283                         v_usec = pps_tf[0] - pps_tf[2];
1284                 } else if (pps_tf[2] > pps_tf[0]) {
1285                         pps_offset = pps_tf[0];         /* 2 0 1 */
1286                         v_usec = pps_tf[2] - pps_tf[1];
1287                 } else {
1288                         pps_offset = pps_tf[2];         /* 0 2 1 */
1289                         v_usec = pps_tf[0] - pps_tf[1];
1290                 }
1291         } else {
1292                 if (pps_tf[1] < pps_tf[2]) {
1293                         pps_offset = pps_tf[1];         /* 2 1 0 */
1294                         v_usec = pps_tf[2] - pps_tf[0];
1295                 } else  if (pps_tf[2] < pps_tf[0]) {
1296                         pps_offset = pps_tf[0];         /* 1 0 2 */
1297                         v_usec = pps_tf[1] - pps_tf[2];
1298                 } else {
1299                         pps_offset = pps_tf[2];         /* 1 2 0 */
1300                         v_usec = pps_tf[1] - pps_tf[0];
1301                 }
1302         }
1303         if (v_usec > MAXTIME)
1304                 pps_jitcnt++;
1305         v_usec = (v_usec << PPS_AVG) - pps_jitter;
1306         pps_jitter += v_usec / (1 << PPS_AVG);
1307         if (pps_jitter > (MAXTIME >> 1))
1308                 time_status |= STA_PPSJITTER;
1309 
1310         /*
1311          * During the calibration interval adjust the starting time when
1312          * the tick overflows. At the end of the interval compute the
1313          * duration of the interval and the difference of the hardware
1314          * counters at the beginning and end of the interval. This code
1315          * is deliciously complicated by the fact valid differences may
1316          * exceed the value of tick when using long calibration
1317          * intervals and small ticks. Note that the counter can be
1318          * greater than tick if caught at just the wrong instant, but
1319          * the values returned and used here are correct.
1320          */
1321         bigtick = (int)usec_per_tick * SCALE_USEC;
1322         pps_usec -= pps_freq;
1323         if (pps_usec >= bigtick)
1324                 pps_usec -= bigtick;
1325         if (pps_usec < 0)
1326                 pps_usec += bigtick;
1327         pps_time.tv_sec++;
1328         pps_count++;
1329         if (pps_count < (1 << pps_shift))
1330                 return;
1331         pps_count = 0;
1332         pps_calcnt++;
1333         u_usec = usec * SCALE_USEC;
1334         v_usec = pps_usec - u_usec;
1335         if (v_usec >= bigtick >> 1)
1336                 v_usec -= bigtick;
1337         if (v_usec < -(bigtick >> 1))
1338                 v_usec += bigtick;
1339         if (v_usec < 0)
1340                 v_usec = -(-v_usec >> pps_shift);
1341         else
1342                 v_usec = v_usec >> pps_shift;
1343         pps_usec = u_usec;
1344         cal_sec = tvp->tv_sec;
1345         cal_usec = tvp->tv_usec;
1346         cal_sec -= pps_time.tv_sec;
1347         cal_usec -= pps_time.tv_usec;
1348         if (cal_usec < 0) {
1349                 cal_usec += MICROSEC;
1350                 cal_sec--;
1351         }
1352         pps_time = *tvp;
1353 
1354         /*
1355          * Check for lost interrupts, noise, excessive jitter and
1356          * excessive frequency error. The number of timer ticks during
1357          * the interval may vary +-1 tick. Add to this a margin of one
1358          * tick for the PPS signal jitter and maximum frequency
1359          * deviation. If the limits are exceeded, the calibration
1360          * interval is reset to the minimum and we start over.
1361          */
1362         u_usec = (int)usec_per_tick << 1;
1363         if (!((cal_sec == -1 && cal_usec > (MICROSEC - u_usec)) ||
1364             (cal_sec == 0 && cal_usec < u_usec)) ||
1365             v_usec > time_tolerance || v_usec < -time_tolerance) {
1366                 pps_errcnt++;
1367                 pps_shift = PPS_SHIFT;
1368                 pps_intcnt = 0;
1369                 time_status |= STA_PPSERROR;
1370                 return;
1371         }
1372 
1373         /*
1374          * A three-stage median filter is used to help deglitch the pps
1375          * frequency. The median sample becomes the frequency offset
1376          * estimate; the difference between the other two samples
1377          * becomes the frequency dispersion (stability) estimate.
1378          */
1379         pps_ff[2] = pps_ff[1];
1380         pps_ff[1] = pps_ff[0];
1381         pps_ff[0] = v_usec;
1382         if (pps_ff[0] > pps_ff[1]) {
1383                 if (pps_ff[1] > pps_ff[2]) {
1384                         u_usec = pps_ff[1];             /* 0 1 2 */
1385                         v_usec = pps_ff[0] - pps_ff[2];
1386                 } else if (pps_ff[2] > pps_ff[0]) {
1387                         u_usec = pps_ff[0];             /* 2 0 1 */
1388                         v_usec = pps_ff[2] - pps_ff[1];
1389                 } else {
1390                         u_usec = pps_ff[2];             /* 0 2 1 */
1391                         v_usec = pps_ff[0] - pps_ff[1];
1392                 }
1393         } else {
1394                 if (pps_ff[1] < pps_ff[2]) {
1395                         u_usec = pps_ff[1];             /* 2 1 0 */
1396                         v_usec = pps_ff[2] - pps_ff[0];
1397                 } else  if (pps_ff[2] < pps_ff[0]) {
1398                         u_usec = pps_ff[0];             /* 1 0 2 */
1399                         v_usec = pps_ff[1] - pps_ff[2];
1400                 } else {
1401                         u_usec = pps_ff[2];             /* 1 2 0 */
1402                         v_usec = pps_ff[1] - pps_ff[0];
1403                 }
1404         }
1405 
1406         /*
1407          * Here the frequency dispersion (stability) is updated. If it
1408          * is less than one-fourth the maximum (MAXFREQ), the frequency
1409          * offset is updated as well, but clamped to the tolerance. It
1410          * will be processed later by the clock() routine.
1411          */
1412         v_usec = (v_usec >> 1) - pps_stabil;
1413         if (v_usec < 0)
1414                 pps_stabil -= -v_usec >> PPS_AVG;
1415         else
1416                 pps_stabil += v_usec >> PPS_AVG;
1417         if (pps_stabil > MAXFREQ >> 2) {
1418                 pps_stbcnt++;
1419                 time_status |= STA_PPSWANDER;
1420                 return;
1421         }
1422         if (time_status & STA_PPSFREQ) {
1423                 if (u_usec < 0) {
1424                         pps_freq -= -u_usec >> PPS_AVG;
1425                         if (pps_freq < -time_tolerance)
1426                                 pps_freq = -time_tolerance;
1427                         u_usec = -u_usec;
1428                 } else {
1429                         pps_freq += u_usec >> PPS_AVG;
1430                         if (pps_freq > time_tolerance)
1431                                 pps_freq = time_tolerance;
1432                 }
1433         }
1434 
1435         /*
1436          * Here the calibration interval is adjusted. If the maximum
1437          * time difference is greater than tick / 4, reduce the interval
1438          * by half. If this is not the case for four consecutive
1439          * intervals, double the interval.
1440          */
1441         if (u_usec << pps_shift > bigtick >> 2) {
1442                 pps_intcnt = 0;
1443                 if (pps_shift > PPS_SHIFT)
1444                         pps_shift--;
1445         } else if (pps_intcnt >= 4) {
1446                 pps_intcnt = 0;
1447                 if (pps_shift < PPS_SHIFTMAX)
1448                         pps_shift++;
1449         } else
1450                 pps_intcnt++;
1451 
1452         /*
1453          * If recovering from kmdb, then make sure the tod chip gets resynced.
1454          * If we took an early exit above, then we don't yet have a stable
1455          * calibration signal to lock onto, so don't mark the tod for sync
1456          * until we get all the way here.
1457          */
1458         {
1459                 int s = hr_clock_lock();
1460 
1461                 tod_needsync = 1;
1462                 hr_clock_unlock(s);
1463         }
1464 }
1465 
1466 /*
1467  * Handle clock tick processing for a thread.
1468  * Check for timer action, enforce CPU rlimit, do profiling etc.
1469  */
1470 void
1471 clock_tick(kthread_t *t, int pending)
1472 {
1473         struct proc *pp;
1474         klwp_id_t    lwp;
1475         struct as *as;
1476         clock_t ticks;
1477         int     poke = 0;               /* notify another CPU */
1478         int     user_mode;
1479         size_t   rss;
1480         int i, total_usec, usec;
1481         rctl_qty_t secs;
1482 
1483         ASSERT(pending > 0);
1484 
1485         /* Must be operating on a lwp/thread */
1486         if ((lwp = ttolwp(t)) == NULL) {
1487                 panic("clock_tick: no lwp");
1488                 /*NOTREACHED*/
1489         }
1490 
1491         for (i = 0; i < pending; i++) {
1492                 CL_TICK(t);     /* Class specific tick processing */
1493                 DTRACE_SCHED1(tick, kthread_t *, t);
1494         }
1495 
1496         pp = ttoproc(t);
1497 
1498         /* pp->p_lock makes sure that the thread does not exit */
1499         ASSERT(MUTEX_HELD(&pp->p_lock));
1500 
1501         user_mode = (lwp->lwp_state == LWP_USER);
1502 
1503         ticks = (pp->p_utime + pp->p_stime) % hz;
1504         /*
1505          * Update process times. Should use high res clock and state
1506          * changes instead of statistical sampling method. XXX
1507          */
1508         if (user_mode) {
1509                 pp->p_utime += pending;
1510         } else {
1511                 pp->p_stime += pending;
1512         }
1513 
1514         pp->p_ttime += pending;
1515         as = pp->p_as;
1516 
1517         /*
1518          * Update user profiling statistics. Get the pc from the
1519          * lwp when the AST happens.
1520          */
1521         if (pp->p_prof.pr_scale) {
1522                 atomic_add_32(&lwp->lwp_oweupc, (int32_t)pending);
1523                 if (user_mode) {
1524                         poke = 1;
1525                         aston(t);
1526                 }
1527         }
1528 
1529         /*
1530          * If CPU was in user state, process lwp-virtual time
1531          * interval timer. The value passed to itimerdecr() has to be
1532          * in microseconds and has to be less than one second. Hence
1533          * this loop.
1534          */
1535         total_usec = usec_per_tick * pending;
1536         while (total_usec > 0) {
1537                 usec = MIN(total_usec, (MICROSEC - 1));
1538                 if (user_mode &&
1539                     timerisset(&lwp->lwp_timer[ITIMER_VIRTUAL].it_value) &&
1540                     itimerdecr(&lwp->lwp_timer[ITIMER_VIRTUAL], usec) == 0) {
1541                         poke = 1;
1542                         sigtoproc(pp, t, SIGVTALRM);
1543                 }
1544                 total_usec -= usec;
1545         }
1546 
1547         /*
1548          * If CPU was in user state, process lwp-profile
1549          * interval timer.
1550          */
1551         total_usec = usec_per_tick * pending;
1552         while (total_usec > 0) {
1553                 usec = MIN(total_usec, (MICROSEC - 1));
1554                 if (timerisset(&lwp->lwp_timer[ITIMER_PROF].it_value) &&
1555                     itimerdecr(&lwp->lwp_timer[ITIMER_PROF], usec) == 0) {
1556                         poke = 1;
1557                         sigtoproc(pp, t, SIGPROF);
1558                 }
1559                 total_usec -= usec;
1560         }
1561 
1562         /*
1563          * Enforce CPU resource controls:
1564          *   (a) process.max-cpu-time resource control
1565          *
1566          * Perform the check only if we have accumulated more a second.
1567          */
1568         if ((ticks + pending) >= hz) {
1569                 (void) rctl_test(rctlproc_legacy[RLIMIT_CPU], pp->p_rctls, pp,
1570                     (pp->p_utime + pp->p_stime)/hz, RCA_UNSAFE_SIGINFO);
1571         }
1572 
1573         /*
1574          *   (b) task.max-cpu-time resource control
1575          *
1576          * If we have accumulated enough ticks, increment the task CPU
1577          * time usage and test for the resource limit. This minimizes the
1578          * number of calls to the rct_test(). The task CPU time mutex
1579          * is highly contentious as many processes can be sharing a task.
1580          */
1581         if (pp->p_ttime >= clock_tick_proc_max) {
1582                 secs = task_cpu_time_incr(pp->p_task, pp->p_ttime);
1583                 pp->p_ttime = 0;
1584                 if (secs) {
1585                         (void) rctl_test(rc_task_cpu_time, pp->p_task->tk_rctls,
1586                             pp, secs, RCA_UNSAFE_SIGINFO);
1587                 }
1588         }
1589 
1590         /*
1591          * Update memory usage for the currently running process.
1592          */
1593         rss = rm_asrss(as);
1594         PTOU(pp)->u_mem += rss;
1595         if (rss > PTOU(pp)->u_mem_max)
1596                 PTOU(pp)->u_mem_max = rss;
1597 
1598         /*
1599          * Notify the CPU the thread is running on.
1600          */
1601         if (poke && t->t_cpu != CPU)
1602                 poke_cpu(t->t_cpu->cpu_id);
1603 }
1604 
1605 void
1606 profil_tick(uintptr_t upc)
1607 {
1608         int ticks;
1609         proc_t *p = ttoproc(curthread);
1610         klwp_t *lwp = ttolwp(curthread);
1611         struct prof *pr = &p->p_prof;
1612 
1613         do {
1614                 ticks = lwp->lwp_oweupc;
1615         } while (atomic_cas_32(&lwp->lwp_oweupc, ticks, 0) != ticks);
1616 
1617         mutex_enter(&p->p_pflock);
1618         if (pr->pr_scale >= 2 && upc >= pr->pr_off) {
1619                 /*
1620                  * Old-style profiling
1621                  */
1622                 uint16_t *slot = pr->pr_base;
1623                 uint16_t old, new;
1624                 if (pr->pr_scale != 2) {
1625                         uintptr_t delta = upc - pr->pr_off;
1626                         uintptr_t byteoff = ((delta >> 16) * pr->pr_scale) +
1627                             (((delta & 0xffff) * pr->pr_scale) >> 16);
1628                         if (byteoff >= (uintptr_t)pr->pr_size) {
1629                                 mutex_exit(&p->p_pflock);
1630                                 return;
1631                         }
1632                         slot += byteoff / sizeof (uint16_t);
1633                 }
1634                 if (fuword16(slot, &old) < 0 ||
1635                     (new = old + ticks) > SHRT_MAX ||
1636                     suword16(slot, new) < 0) {
1637                         pr->pr_scale = 0;
1638                 }
1639         } else if (pr->pr_scale == 1) {
1640                 /*
1641                  * PC Sampling
1642                  */
1643                 model_t model = lwp_getdatamodel(lwp);
1644                 int result;
1645 #ifdef __lint
1646                 model = model;
1647 #endif
1648                 while (ticks-- > 0) {
1649                         if (pr->pr_samples == pr->pr_size) {
1650                                 /* buffer full, turn off sampling */
1651                                 pr->pr_scale = 0;
1652                                 break;
1653                         }
1654                         switch (SIZEOF_PTR(model)) {
1655                         case sizeof (uint32_t):
1656                                 result = suword32(pr->pr_base, (uint32_t)upc);
1657                                 break;
1658 #ifdef _LP64
1659                         case sizeof (uint64_t):
1660                                 result = suword64(pr->pr_base, (uint64_t)upc);
1661                                 break;
1662 #endif
1663                         default:
1664                                 cmn_err(CE_WARN, "profil_tick: unexpected "
1665                                     "data model");
1666                                 result = -1;
1667                                 break;
1668                         }
1669                         if (result != 0) {
1670                                 pr->pr_scale = 0;
1671                                 break;
1672                         }
1673                         pr->pr_base = (caddr_t)pr->pr_base + SIZEOF_PTR(model);
1674                         pr->pr_samples++;
1675                 }
1676         }
1677         mutex_exit(&p->p_pflock);
1678 }
1679 
1680 static void
1681 delay_wakeup(void *arg)
1682 {
1683         kthread_t       *t = arg;
1684 
1685         mutex_enter(&t->t_delay_lock);
1686         cv_signal(&t->t_delay_cv);
1687         mutex_exit(&t->t_delay_lock);
1688 }
1689 
1690 /*
1691  * The delay(9F) man page indicates that it can only be called from user or
1692  * kernel context - detect and diagnose bad calls. The following macro will
1693  * produce a limited number of messages identifying bad callers.  This is done
1694  * in a macro so that caller() is meaningful. When a bad caller is identified,
1695  * switching to 'drv_usecwait(TICK_TO_USEC(ticks));' may be appropriate.
1696  */
1697 #define DELAY_CONTEXT_CHECK()   {                                       \
1698         uint32_t        m;                                              \
1699         char            *f;                                             \
1700         ulong_t         off;                                            \
1701                                                                         \
1702         m = delay_from_interrupt_msg;                                   \
1703         if (delay_from_interrupt_diagnose && servicing_interrupt() &&   \
1704             !panicstr && !devinfo_freeze &&                             \
1705             atomic_cas_32(&delay_from_interrupt_msg, m ? m : 1, m-1)) {     \
1706                 f = modgetsymname((uintptr_t)caller(), &off);               \
1707                 cmn_err(CE_WARN, "delay(9F) called from "               \
1708                     "interrupt context: %s`%s",                         \
1709                     mod_containing_pc(caller()), f ? f : "...");        \
1710         }                                                               \
1711 }
1712 
1713 /*
1714  * delay_common: common delay code.
1715  */
1716 static void
1717 delay_common(clock_t ticks)
1718 {
1719         kthread_t       *t = curthread;
1720         clock_t         deadline;
1721         clock_t         timeleft;
1722         callout_id_t    id;
1723 
1724         /* If timeouts aren't running all we can do is spin. */
1725         if (panicstr || devinfo_freeze) {
1726                 /* Convert delay(9F) call into drv_usecwait(9F) call. */
1727                 if (ticks > 0)
1728                         drv_usecwait(TICK_TO_USEC(ticks));
1729                 return;
1730         }
1731 
1732         deadline = ddi_get_lbolt() + ticks;
1733         while ((timeleft = deadline - ddi_get_lbolt()) > 0) {
1734                 mutex_enter(&t->t_delay_lock);
1735                 id = timeout_default(delay_wakeup, t, timeleft);
1736                 cv_wait(&t->t_delay_cv, &t->t_delay_lock);
1737                 mutex_exit(&t->t_delay_lock);
1738                 (void) untimeout_default(id, 0);
1739         }
1740 }
1741 
1742 /*
1743  * Delay specified number of clock ticks.
1744  */
1745 void
1746 delay(clock_t ticks)
1747 {
1748         DELAY_CONTEXT_CHECK();
1749 
1750         delay_common(ticks);
1751 }
1752 
1753 /*
1754  * Delay a random number of clock ticks between 1 and ticks.
1755  */
1756 void
1757 delay_random(clock_t ticks)
1758 {
1759         int     r;
1760 
1761         DELAY_CONTEXT_CHECK();
1762 
1763         (void) random_get_pseudo_bytes((void *)&r, sizeof (r));
1764         if (ticks == 0)
1765                 ticks = 1;
1766         ticks = (r % ticks) + 1;
1767         delay_common(ticks);
1768 }
1769 
1770 /*
1771  * Like delay, but interruptible by a signal.
1772  */
1773 int
1774 delay_sig(clock_t ticks)
1775 {
1776         kthread_t       *t = curthread;
1777         clock_t         deadline;
1778         clock_t         rc;
1779 
1780         /* If timeouts aren't running all we can do is spin. */
1781         if (panicstr || devinfo_freeze) {
1782                 if (ticks > 0)
1783                         drv_usecwait(TICK_TO_USEC(ticks));
1784                 return (0);
1785         }
1786 
1787         deadline = ddi_get_lbolt() + ticks;
1788         mutex_enter(&t->t_delay_lock);
1789         do {
1790                 rc = cv_timedwait_sig(&t->t_delay_cv,
1791                     &t->t_delay_lock, deadline);
1792                 /* loop until past deadline or signaled */
1793         } while (rc > 0);
1794         mutex_exit(&t->t_delay_lock);
1795         if (rc == 0)
1796                 return (EINTR);
1797         return (0);
1798 }
1799 
1800 
1801 #define SECONDS_PER_DAY 86400
1802 
1803 /*
1804  * Initialize the system time based on the TOD chip.  approx is used as
1805  * an approximation of time (e.g. from the filesystem) in the event that
1806  * the TOD chip has been cleared or is unresponsive.  An approx of -1
1807  * means the filesystem doesn't keep time.
1808  */
1809 void
1810 clkset(time_t approx)
1811 {
1812         timestruc_t ts;
1813         int spl;
1814         int set_clock = 0;
1815 
1816         mutex_enter(&tod_lock);
1817         ts = tod_get();
1818 
1819         if (ts.tv_sec > 365 * SECONDS_PER_DAY) {
1820                 /*
1821                  * If the TOD chip is reporting some time after 1971,
1822                  * then it probably didn't lose power or become otherwise
1823                  * cleared in the recent past;  check to assure that
1824                  * the time coming from the filesystem isn't in the future
1825                  * according to the TOD chip.
1826                  */
1827                 if (approx != -1 && approx > ts.tv_sec) {
1828                         cmn_err(CE_WARN, "Last shutdown is later "
1829                             "than time on time-of-day chip; check date.");
1830                 }
1831         } else {
1832                 /*
1833                  * If the TOD chip isn't giving correct time, set it to the
1834                  * greater of i) approx and ii) 1987. That way if approx
1835                  * is negative or is earlier than 1987, we set the clock
1836                  * back to a time when Oliver North, ALF and Dire Straits
1837                  * were all on the collective brain:  1987.
1838                  */
1839                 timestruc_t tmp;
1840                 time_t diagnose_date = (1987 - 1970) * 365 * SECONDS_PER_DAY;
1841                 ts.tv_sec = (approx > diagnose_date ? approx : diagnose_date);
1842                 ts.tv_nsec = 0;
1843 
1844                 /*
1845                  * Attempt to write the new time to the TOD chip.  Set spl high
1846                  * to avoid getting preempted between the tod_set and tod_get.
1847                  */
1848                 spl = splhi();
1849                 tod_set(ts);
1850                 tmp = tod_get();
1851                 splx(spl);
1852 
1853                 if (tmp.tv_sec != ts.tv_sec && tmp.tv_sec != ts.tv_sec + 1) {
1854                         tod_broken = 1;
1855                         dosynctodr = 0;
1856                         cmn_err(CE_WARN, "Time-of-day chip unresponsive.");
1857                 } else {
1858                         cmn_err(CE_WARN, "Time-of-day chip had "
1859                             "incorrect date; check and reset.");
1860                 }
1861                 set_clock = 1;
1862         }
1863 
1864         if (!boot_time) {
1865                 boot_time = ts.tv_sec;
1866                 set_clock = 1;
1867         }
1868 
1869         if (set_clock)
1870                 set_hrestime(&ts);
1871 
1872         mutex_exit(&tod_lock);
1873 }
1874 
1875 int     timechanged;    /* for testing if the system time has been reset */
1876 
1877 void
1878 set_hrestime(timestruc_t *ts)
1879 {
1880         int spl = hr_clock_lock();
1881         hrestime = *ts;
1882         membar_enter(); /* hrestime must be visible before timechanged++ */
1883         timedelta = 0;
1884         timechanged++;
1885         hr_clock_unlock(spl);
1886         callout_hrestime();
1887 }
1888 
1889 static uint_t deadman_seconds;
1890 static uint32_t deadman_panics;
1891 static int deadman_enabled = 0;
1892 static int deadman_panic_timers = 1;
1893 
1894 static void
1895 deadman(void)
1896 {
1897         if (panicstr) {
1898                 /*
1899                  * During panic, other CPUs besides the panic
1900                  * master continue to handle cyclics and some other
1901                  * interrupts.  The code below is intended to be
1902                  * single threaded, so any CPU other than the master
1903                  * must keep out.
1904                  */
1905                 if (CPU->cpu_id != panic_cpu.cpu_id)
1906                         return;
1907 
1908                 if (!deadman_panic_timers)
1909                         return; /* allow all timers to be manually disabled */
1910 
1911                 /*
1912                  * If we are generating a crash dump or syncing filesystems and
1913                  * the corresponding timer is set, decrement it and re-enter
1914                  * the panic code to abort it and advance to the next state.
1915                  * The panic states and triggers are explained in panic.c.
1916                  */
1917                 if (panic_dump) {
1918                         if (dump_timeleft && (--dump_timeleft == 0)) {
1919                                 panic("panic dump timeout");
1920                                 /*NOTREACHED*/
1921                         }
1922                 } else if (panic_sync) {
1923                         if (sync_timeleft && (--sync_timeleft == 0)) {
1924                                 panic("panic sync timeout");
1925                                 /*NOTREACHED*/
1926                         }
1927                 }
1928 
1929                 return;
1930         }
1931 
1932         if (deadman_counter != CPU->cpu_deadman_counter) {
1933                 CPU->cpu_deadman_counter = deadman_counter;
1934                 CPU->cpu_deadman_countdown = deadman_seconds;
1935                 return;
1936         }
1937 
1938         if (--CPU->cpu_deadman_countdown > 0)
1939                 return;
1940 
1941         /*
1942          * Regardless of whether or not we actually bring the system down,
1943          * bump the deadman_panics variable.
1944          *
1945          * N.B. deadman_panics is incremented once for each CPU that
1946          * passes through here.  It's expected that all the CPUs will
1947          * detect this condition within one second of each other, so
1948          * when deadman_enabled is off, deadman_panics will
1949          * typically be a multiple of the total number of CPUs in
1950          * the system.
1951          */
1952         atomic_inc_32(&deadman_panics);
1953 
1954         if (!deadman_enabled) {
1955                 CPU->cpu_deadman_countdown = deadman_seconds;
1956                 return;
1957         }
1958 
1959         /*
1960          * If we're here, we want to bring the system down.
1961          */
1962         panic("deadman: timed out after %d seconds of clock "
1963             "inactivity", deadman_seconds);
1964         /*NOTREACHED*/
1965 }
1966 
1967 /*ARGSUSED*/
1968 static void
1969 deadman_online(void *arg, cpu_t *cpu, cyc_handler_t *hdlr, cyc_time_t *when)
1970 {
1971         cpu->cpu_deadman_counter = 0;
1972         cpu->cpu_deadman_countdown = deadman_seconds;
1973 
1974         hdlr->cyh_func = (cyc_func_t)deadman;
1975         hdlr->cyh_level = CY_HIGH_LEVEL;
1976         hdlr->cyh_arg = NULL;
1977 
1978         /*
1979          * Stagger the CPUs so that they don't all run deadman() at
1980          * the same time.  Simplest reason to do this is to make it
1981          * more likely that only one CPU will panic in case of a
1982          * timeout.  This is (strictly speaking) an aesthetic, not a
1983          * technical consideration.
1984          */
1985         when->cyt_when = cpu->cpu_id * (NANOSEC / NCPU);
1986         when->cyt_interval = NANOSEC;
1987 }
1988 
1989 
1990 void
1991 deadman_init(void)
1992 {
1993         cyc_omni_handler_t hdlr;
1994 
1995         if (deadman_seconds == 0)
1996                 deadman_seconds = snoop_interval / MICROSEC;
1997 
1998         if (snooping)
1999                 deadman_enabled = 1;
2000 
2001         hdlr.cyo_online = deadman_online;
2002         hdlr.cyo_offline = NULL;
2003         hdlr.cyo_arg = NULL;
2004 
2005         mutex_enter(&cpu_lock);
2006         deadman_cyclic = cyclic_add_omni(&hdlr);
2007         mutex_exit(&cpu_lock);
2008 }
2009 
2010 /*
2011  * tod_fault() is for updating tod validate mechanism state:
2012  * (1) TOD_NOFAULT: for resetting the state to 'normal'.
2013  *     currently used for debugging only
2014  * (2) The following four cases detected by tod validate mechanism:
2015  *       TOD_REVERSED: current tod value is less than previous value.
2016  *       TOD_STALLED: current tod value hasn't advanced.
2017  *       TOD_JUMPED: current tod value advanced too far from previous value.
2018  *       TOD_RATECHANGED: the ratio between average tod delta and
2019  *       average tick delta has changed.
2020  * (3) TOD_RDONLY: when the TOD clock is not writeable e.g. because it is
2021  *     a virtual TOD provided by a hypervisor.
2022  */
2023 enum tod_fault_type
2024 tod_fault(enum tod_fault_type ftype, int off)
2025 {
2026         ASSERT(MUTEX_HELD(&tod_lock));
2027 
2028         if (tod_faulted != ftype) {
2029                 switch (ftype) {
2030                 case TOD_NOFAULT:
2031                         plat_tod_fault(TOD_NOFAULT);
2032                         cmn_err(CE_NOTE, "Restarted tracking "
2033                             "Time of Day clock.");
2034                         tod_faulted = ftype;
2035                         break;
2036                 case TOD_REVERSED:
2037                 case TOD_JUMPED:
2038                         if (tod_faulted == TOD_NOFAULT) {
2039                                 plat_tod_fault(ftype);
2040                                 cmn_err(CE_WARN, "Time of Day clock error: "
2041                                     "reason [%s by 0x%x]. -- "
2042                                     " Stopped tracking Time Of Day clock.",
2043                                     tod_fault_table[ftype], off);
2044                                 tod_faulted = ftype;
2045                         }
2046                         break;
2047                 case TOD_STALLED:
2048                 case TOD_RATECHANGED:
2049                         if (tod_faulted == TOD_NOFAULT) {
2050                                 plat_tod_fault(ftype);
2051                                 cmn_err(CE_WARN, "Time of Day clock error: "
2052                                     "reason [%s]. -- "
2053                                     " Stopped tracking Time Of Day clock.",
2054                                     tod_fault_table[ftype]);
2055                                 tod_faulted = ftype;
2056                         }
2057                         break;
2058                 case TOD_RDONLY:
2059                         if (tod_faulted == TOD_NOFAULT) {
2060                                 plat_tod_fault(ftype);
2061                                 cmn_err(CE_NOTE, "!Time of Day clock is "
2062                                     "Read-Only; set of Date/Time will not "
2063                                     "persist across reboot.");
2064                                 tod_faulted = ftype;
2065                         }
2066                         break;
2067                 default:
2068                         break;
2069                 }
2070         }
2071         return (tod_faulted);
2072 }
2073 
2074 /*
2075  * Two functions that allow tod_status_flag to be manipulated by functions
2076  * external to this file.
2077  */
2078 
2079 void
2080 tod_status_set(int tod_flag)
2081 {
2082         tod_status_flag |= tod_flag;
2083 }
2084 
2085 void
2086 tod_status_clear(int tod_flag)
2087 {
2088         tod_status_flag &= ~tod_flag;
2089 }
2090 
2091 /*
2092  * Record a timestamp and the value passed to tod_set().  The next call to
2093  * tod_validate() can use these values, prev_set_tick and prev_set_tod,
2094  * when checking the timestruc_t returned by tod_get().  Ordinarily,
2095  * tod_validate() will use prev_tick and prev_tod for this task but these
2096  * become obsolete, and will be re-assigned with the prev_set_* values,
2097  * in the case when the TOD is re-written.
2098  */
2099 void
2100 tod_set_prev(timestruc_t ts)
2101 {
2102         if ((tod_validate_enable == 0) || (tod_faulted != TOD_NOFAULT) ||
2103             tod_validate_deferred) {
2104                 return;
2105         }
2106         prev_set_tick = gethrtime();
2107         /*
2108          * A negative value will be set to zero in utc_to_tod() so we fake
2109          * a zero here in such a case.  This would need to change if the
2110          * behavior of utc_to_tod() changes.
2111          */
2112         prev_set_tod = ts.tv_sec < 0 ? 0 : ts.tv_sec;
2113 }
2114 
2115 /*
2116  * tod_validate() is used for checking values returned by tod_get().
2117  * Four error cases can be detected by this routine:
2118  *   TOD_REVERSED: current tod value is less than previous.
2119  *   TOD_STALLED: current tod value hasn't advanced.
2120  *   TOD_JUMPED: current tod value advanced too far from previous value.
2121  *   TOD_RATECHANGED: the ratio between average tod delta and
2122  *   average tick delta has changed.
2123  */
2124 time_t
2125 tod_validate(time_t tod)
2126 {
2127         time_t diff_tod;
2128         hrtime_t diff_tick;
2129 
2130         long dtick;
2131         int dtick_delta;
2132 
2133         int off = 0;
2134         enum tod_fault_type tod_bad = TOD_NOFAULT;
2135 
2136         static int firsttime = 1;
2137 
2138         static time_t prev_tod = 0;
2139         static hrtime_t prev_tick = 0;
2140         static long dtick_avg = TOD_REF_FREQ;
2141 
2142         int cpr_resume_done = 0;
2143         int dr_resume_done = 0;
2144 
2145         hrtime_t tick = gethrtime();
2146 
2147         ASSERT(MUTEX_HELD(&tod_lock));
2148 
2149         /*
2150          * tod_validate_enable is patchable via /etc/system.
2151          * If TOD is already faulted, or if TOD validation is deferred,
2152          * there is nothing to do.
2153          */
2154         if ((tod_validate_enable == 0) || (tod_faulted != TOD_NOFAULT) ||
2155             tod_validate_deferred) {
2156                 return (tod);
2157         }
2158 
2159         /*
2160          * If this is the first time through, we just need to save the tod
2161          * we were called with and hrtime so we can use them next time to
2162          * validate tod_get().
2163          */
2164         if (firsttime) {
2165                 firsttime = 0;
2166                 prev_tod = tod;
2167                 prev_tick = tick;
2168                 return (tod);
2169         }
2170 
2171         /*
2172          * Handle any flags that have been turned on by tod_status_set().
2173          * In the case where a tod_set() is done and then a subsequent
2174          * tod_get() fails (ie, both TOD_SET_DONE and TOD_GET_FAILED are
2175          * true), we treat the TOD_GET_FAILED with precedence by switching
2176          * off the flag, returning tod and leaving TOD_SET_DONE asserted
2177          * until such time as tod_get() completes successfully.
2178          */
2179         if (tod_status_flag & TOD_GET_FAILED) {
2180                 /*
2181                  * tod_get() has encountered an issue, possibly transitory,
2182                  * when reading TOD.  We'll just return the incoming tod
2183                  * value (which is actually hrestime.tv_sec in this case)
2184                  * and when we get a genuine tod, following a successful
2185                  * tod_get(), we can validate using prev_tod and prev_tick.
2186                  */
2187                 tod_status_flag &= ~TOD_GET_FAILED;
2188                 return (tod);
2189         } else if (tod_status_flag & TOD_SET_DONE) {
2190                 /*
2191                  * TOD has been modified.  Just before the TOD was written,
2192                  * tod_set_prev() saved tod and hrtime; we can now use
2193                  * those values, prev_set_tod and prev_set_tick, to validate
2194                  * the incoming tod that's just been read.
2195                  */
2196                 prev_tod = prev_set_tod;
2197                 prev_tick = prev_set_tick;
2198                 dtick_avg = TOD_REF_FREQ;
2199                 tod_status_flag &= ~TOD_SET_DONE;
2200                 /*
2201                  * If a tod_set() preceded a cpr_suspend() without an
2202                  * intervening tod_validate(), we need to ensure that a
2203                  * TOD_JUMPED condition is ignored.
2204                  * Note this isn't a concern in the case of DR as we've
2205                  * just reassigned dtick_avg, above.
2206                  */
2207                 if (tod_status_flag & TOD_CPR_RESUME_DONE) {
2208                         cpr_resume_done = 1;
2209                         tod_status_flag &= ~TOD_CPR_RESUME_DONE;
2210                 }
2211         } else if (tod_status_flag & TOD_CPR_RESUME_DONE) {
2212                 /*
2213                  * The system's coming back from a checkpoint resume.
2214                  */
2215                 cpr_resume_done = 1;
2216                 tod_status_flag &= ~TOD_CPR_RESUME_DONE;
2217                 /*
2218                  * We need to handle the possibility of a CPR suspend
2219                  * operation having been initiated whilst a DR event was
2220                  * in-flight.
2221                  */
2222                 if (tod_status_flag & TOD_DR_RESUME_DONE) {
2223                         dr_resume_done = 1;
2224                         tod_status_flag &= ~TOD_DR_RESUME_DONE;
2225                 }
2226         } else if (tod_status_flag & TOD_DR_RESUME_DONE) {
2227                 /*
2228                  * A Dynamic Reconfiguration event has taken place.
2229                  */
2230                 dr_resume_done = 1;
2231                 tod_status_flag &= ~TOD_DR_RESUME_DONE;
2232         }
2233 
2234         /* test hook */
2235         switch (tod_unit_test) {
2236         case 1: /* for testing jumping tod */
2237                 tod += tod_test_injector;
2238                 tod_unit_test = 0;
2239                 break;
2240         case 2: /* for testing stuck tod bit */
2241                 tod |= 1 << tod_test_injector;
2242                 tod_unit_test = 0;
2243                 break;
2244         case 3: /* for testing stalled tod */
2245                 tod = prev_tod;
2246                 tod_unit_test = 0;
2247                 break;
2248         case 4: /* reset tod fault status */
2249                 (void) tod_fault(TOD_NOFAULT, 0);
2250                 tod_unit_test = 0;
2251                 break;
2252         default:
2253                 break;
2254         }
2255 
2256         diff_tod = tod - prev_tod;
2257         diff_tick = tick - prev_tick;
2258 
2259         ASSERT(diff_tick >= 0);
2260 
2261         if (diff_tod < 0) {
2262                 /* ERROR - tod reversed */
2263                 tod_bad = TOD_REVERSED;
2264                 off = (int)(prev_tod - tod);
2265         } else if (diff_tod == 0) {
2266                 /* tod did not advance */
2267                 if (diff_tick > TOD_STALL_THRESHOLD) {
2268                         /* ERROR - tod stalled */
2269                         tod_bad = TOD_STALLED;
2270                 } else {
2271                         /*
2272                          * Make sure we don't update prev_tick
2273                          * so that diff_tick is calculated since
2274                          * the first diff_tod == 0
2275                          */
2276                         return (tod);
2277                 }
2278         } else {
2279                 /* calculate dtick */
2280                 dtick = diff_tick / diff_tod;
2281 
2282                 /* update dtick averages */
2283                 dtick_avg += ((dtick - dtick_avg) / TOD_FILTER_N);
2284 
2285                 /*
2286                  * Calculate dtick_delta as
2287                  * variation from reference freq in quartiles
2288                  */
2289                 dtick_delta = (dtick_avg - TOD_REF_FREQ) /
2290                     (TOD_REF_FREQ >> 2);
2291 
2292                 /*
2293                  * Even with a perfectly functioning TOD device,
2294                  * when the number of elapsed seconds is low the
2295                  * algorithm can calculate a rate that is beyond
2296                  * tolerance, causing an error.  The algorithm is
2297                  * inaccurate when elapsed time is low (less than
2298                  * 5 seconds).
2299                  */
2300                 if (diff_tod > 4) {
2301                         if (dtick < TOD_JUMP_THRESHOLD) {
2302                                 /*
2303                                  * If we've just done a CPR resume, we detect
2304                                  * a jump in the TOD but, actually, what's
2305                                  * happened is that the TOD has been increasing
2306                                  * whilst the system was suspended and the tick
2307                                  * count hasn't kept up.  We consider the first
2308                                  * occurrence of this after a resume as normal
2309                                  * and ignore it; otherwise, in a non-resume
2310                                  * case, we regard it as a TOD problem.
2311                                  */
2312                                 if (!cpr_resume_done) {
2313                                         /* ERROR - tod jumped */
2314                                         tod_bad = TOD_JUMPED;
2315                                         off = (int)diff_tod;
2316                                 }
2317                         }
2318                         if (dtick_delta) {
2319                                 /*
2320                                  * If we've just done a DR resume, dtick_avg
2321                                  * can go a bit askew so we reset it and carry
2322                                  * on; otherwise, the TOD is in error.
2323                                  */
2324                                 if (dr_resume_done) {
2325                                         dtick_avg = TOD_REF_FREQ;
2326                                 } else {
2327                                         /* ERROR - change in clock rate */
2328                                         tod_bad = TOD_RATECHANGED;
2329                                 }
2330                         }
2331                 }
2332         }
2333 
2334         if (tod_bad != TOD_NOFAULT) {
2335                 (void) tod_fault(tod_bad, off);
2336 
2337                 /*
2338                  * Disable dosynctodr since we are going to fault
2339                  * the TOD chip anyway here
2340                  */
2341                 dosynctodr = 0;
2342 
2343                 /*
2344                  * Set tod to the correct value from hrestime
2345                  */
2346                 tod = hrestime.tv_sec;
2347         }
2348 
2349         prev_tod = tod;
2350         prev_tick = tick;
2351         return (tod);
2352 }
2353 
2354 static void
2355 calcloadavg(int nrun, uint64_t *hp_ave)
2356 {
2357         static int64_t f[3] = { 135, 27, 9 };
2358         uint_t i;
2359         int64_t q, r;
2360 
2361         /*
2362          * Compute load average over the last 1, 5, and 15 minutes
2363          * (60, 300, and 900 seconds).  The constants in f[3] are for
2364          * exponential decay:
2365          * (1 - exp(-1/60)) << 13 = 135,
2366          * (1 - exp(-1/300)) << 13 = 27,
2367          * (1 - exp(-1/900)) << 13 = 9.
2368          */
2369 
2370         /*
2371          * a little hoop-jumping to avoid integer overflow
2372          */
2373         for (i = 0; i < 3; i++) {
2374                 q = (hp_ave[i]  >> 16) << 7;
2375                 r = (hp_ave[i]  & 0xffff) << 7;
2376                 hp_ave[i] += ((nrun - q) * f[i] - ((r * f[i]) >> 16)) >> 4;
2377         }
2378 }
2379 
2380 /*
2381  * lbolt_hybrid() is used by ddi_get_lbolt() and ddi_get_lbolt64() to
2382  * calculate the value of lbolt according to the current mode. In the event
2383  * driven mode (the default), lbolt is calculated by dividing the current hires
2384  * time by the number of nanoseconds per clock tick. In the cyclic driven mode
2385  * an internal variable is incremented at each firing of the lbolt cyclic
2386  * and returned by lbolt_cyclic_driven().
2387  *
2388  * The system will transition from event to cyclic driven mode when the number
2389  * of calls to lbolt_event_driven() exceeds the (per CPU) threshold within a
2390  * window of time. It does so by reprograming lbolt_cyclic from CY_INFINITY to
2391  * nsec_per_tick. The lbolt cyclic will remain ON while at least one CPU is
2392  * causing enough activity to cross the thresholds.
2393  */
2394 int64_t
2395 lbolt_bootstrap(void)
2396 {
2397         return (0);
2398 }
2399 
2400 /* ARGSUSED */
2401 uint_t
2402 lbolt_ev_to_cyclic(caddr_t arg1, caddr_t arg2)
2403 {
2404         hrtime_t ts, exp;
2405         int ret;
2406 
2407         ASSERT(lbolt_hybrid != lbolt_cyclic_driven);
2408 
2409         kpreempt_disable();
2410 
2411         ts = gethrtime();
2412         lb_info->lbi_internal = (ts/nsec_per_tick);
2413 
2414         /*
2415          * Align the next expiration to a clock tick boundary.
2416          */
2417         exp = ts + nsec_per_tick - 1;
2418         exp = (exp/nsec_per_tick) * nsec_per_tick;
2419 
2420         ret = cyclic_reprogram(lb_info->id.lbi_cyclic_id, exp);
2421         ASSERT(ret);
2422 
2423         lbolt_hybrid = lbolt_cyclic_driven;
2424         lb_info->lbi_cyc_deactivate = B_FALSE;
2425         lb_info->lbi_cyc_deac_start = lb_info->lbi_internal;
2426 
2427         kpreempt_enable();
2428 
2429         ret = atomic_dec_32_nv(&lb_info->lbi_token);
2430         ASSERT(ret == 0);
2431 
2432         return (1);
2433 }
2434 
2435 int64_t
2436 lbolt_event_driven(void)
2437 {
2438         hrtime_t ts;
2439         int64_t lb;
2440         int ret, cpu = CPU->cpu_seqid;
2441 
2442         ts = gethrtime();
2443         ASSERT(ts > 0);
2444 
2445         ASSERT(nsec_per_tick > 0);
2446         lb = (ts/nsec_per_tick);
2447 
2448         /*
2449          * Switch to cyclic mode if the number of calls to this routine
2450          * has reached the threshold within the interval.
2451          */
2452         if ((lb - lb_cpu[cpu].lbc_cnt_start) < lb_info->lbi_thresh_interval) {
2453 
2454                 if (--lb_cpu[cpu].lbc_counter == 0) {
2455                         /*
2456                          * Reached the threshold within the interval, reset
2457                          * the usage statistics.
2458                          */
2459                         lb_cpu[cpu].lbc_counter = lb_info->lbi_thresh_calls;
2460                         lb_cpu[cpu].lbc_cnt_start = lb;
2461 
2462                         /*
2463                          * Make sure only one thread reprograms the
2464                          * lbolt cyclic and changes the mode.
2465                          */
2466                         if (panicstr == NULL &&
2467                             atomic_cas_32(&lb_info->lbi_token, 0, 1) == 0) {
2468 
2469                                 if (lbolt_hybrid == lbolt_cyclic_driven) {
2470                                         ret = atomic_dec_32_nv(
2471                                             &lb_info->lbi_token);
2472                                         ASSERT(ret == 0);
2473                                 } else {
2474                                         lbolt_softint_post();
2475                                 }
2476                         }
2477                 }
2478         } else {
2479                 /*
2480                  * Exceeded the interval, reset the usage statistics.
2481                  */
2482                 lb_cpu[cpu].lbc_counter = lb_info->lbi_thresh_calls;
2483                 lb_cpu[cpu].lbc_cnt_start = lb;
2484         }
2485 
2486         ASSERT(lb >= lb_info->lbi_debug_time);
2487 
2488         return (lb - lb_info->lbi_debug_time);
2489 }
2490 
2491 int64_t
2492 lbolt_cyclic_driven(void)
2493 {
2494         int64_t lb = lb_info->lbi_internal;
2495         int cpu;
2496 
2497         /*
2498          * If a CPU has already prevented the lbolt cyclic from deactivating
2499          * itself, don't bother tracking the usage. Otherwise check if we're
2500          * within the interval and how the per CPU counter is doing.
2501          */
2502         if (lb_info->lbi_cyc_deactivate) {
2503                 cpu = CPU->cpu_seqid;
2504                 if ((lb - lb_cpu[cpu].lbc_cnt_start) <
2505                     lb_info->lbi_thresh_interval) {
2506 
2507                         if (lb_cpu[cpu].lbc_counter == 0)
2508                                 /*
2509                                  * Reached the threshold within the interval,
2510                                  * prevent the lbolt cyclic from turning itself
2511                                  * off.
2512                                  */
2513                                 lb_info->lbi_cyc_deactivate = B_FALSE;
2514                         else
2515                                 lb_cpu[cpu].lbc_counter--;
2516                 } else {
2517                         /*
2518                          * Only reset the usage statistics when we have
2519                          * exceeded the interval.
2520                          */
2521                         lb_cpu[cpu].lbc_counter = lb_info->lbi_thresh_calls;
2522                         lb_cpu[cpu].lbc_cnt_start = lb;
2523                 }
2524         }
2525 
2526         ASSERT(lb >= lb_info->lbi_debug_time);
2527 
2528         return (lb - lb_info->lbi_debug_time);
2529 }
2530 
2531 /*
2532  * The lbolt_cyclic() routine will fire at a nsec_per_tick interval to satisfy
2533  * performance needs of ddi_get_lbolt() and ddi_get_lbolt64() consumers.
2534  * It is inactive by default, and will be activated when switching from event
2535  * to cyclic driven lbolt. The cyclic will turn itself off unless signaled
2536  * by lbolt_cyclic_driven().
2537  */
2538 static void
2539 lbolt_cyclic(void)
2540 {
2541         int ret;
2542 
2543         lb_info->lbi_internal++;
2544 
2545         if (!lbolt_cyc_only) {
2546 
2547                 if (lb_info->lbi_cyc_deactivate) {
2548                         /*
2549                          * Switching from cyclic to event driven mode.
2550                          */
2551                         if (panicstr == NULL &&
2552                             atomic_cas_32(&lb_info->lbi_token, 0, 1) == 0) {
2553 
2554                                 if (lbolt_hybrid == lbolt_event_driven) {
2555                                         ret = atomic_dec_32_nv(
2556                                             &lb_info->lbi_token);
2557                                         ASSERT(ret == 0);
2558                                         return;
2559                                 }
2560 
2561                                 kpreempt_disable();
2562 
2563                                 lbolt_hybrid = lbolt_event_driven;
2564                                 ret = cyclic_reprogram(
2565                                     lb_info->id.lbi_cyclic_id,
2566                                     CY_INFINITY);
2567                                 ASSERT(ret);
2568 
2569                                 kpreempt_enable();
2570 
2571                                 ret = atomic_dec_32_nv(&lb_info->lbi_token);
2572                                 ASSERT(ret == 0);
2573                         }
2574                 }
2575 
2576                 /*
2577                  * The lbolt cyclic should not try to deactivate itself before
2578                  * the sampling period has elapsed.
2579                  */
2580                 if (lb_info->lbi_internal - lb_info->lbi_cyc_deac_start >=
2581                     lb_info->lbi_thresh_interval) {
2582                         lb_info->lbi_cyc_deactivate = B_TRUE;
2583                         lb_info->lbi_cyc_deac_start = lb_info->lbi_internal;
2584                 }
2585         }
2586 }
2587 
2588 /*
2589  * Since the lbolt service was historically cyclic driven, it must be 'stopped'
2590  * when the system drops into the kernel debugger. lbolt_debug_entry() is
2591  * called by the KDI system claim callbacks to record a hires timestamp at
2592  * debug enter time. lbolt_debug_return() is called by the sistem release
2593  * callbacks to account for the time spent in the debugger. The value is then
2594  * accumulated in the lb_info structure and used by lbolt_event_driven() and
2595  * lbolt_cyclic_driven(), as well as the mdb_get_lbolt() routine.
2596  */
2597 void
2598 lbolt_debug_entry(void)
2599 {
2600         if (lbolt_hybrid != lbolt_bootstrap) {
2601                 ASSERT(lb_info != NULL);
2602                 lb_info->lbi_debug_ts = gethrtime();
2603         }
2604 }
2605 
2606 /*
2607  * Calculate the time spent in the debugger and add it to the lbolt info
2608  * structure. We also update the internal lbolt value in case we were in
2609  * cyclic driven mode going in.
2610  */
2611 void
2612 lbolt_debug_return(void)
2613 {
2614         hrtime_t ts;
2615 
2616         if (lbolt_hybrid != lbolt_bootstrap) {
2617                 ASSERT(lb_info != NULL);
2618                 ASSERT(nsec_per_tick > 0);
2619 
2620                 ts = gethrtime();
2621                 lb_info->lbi_internal = (ts/nsec_per_tick);
2622                 lb_info->lbi_debug_time +=
2623                     ((ts - lb_info->lbi_debug_ts)/nsec_per_tick);
2624 
2625                 lb_info->lbi_debug_ts = 0;
2626         }
2627 }