xref: /illumos-gate/usr/src/uts/common/os/clock.c (revision aa8590303de9350e3cbe3b4eb3c3d8f951d77f46)
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  * Copyright (c) 2016 by Delphix. All rights reserved.
28  */
29 
30 #include <sys/param.h>
31 #include <sys/t_lock.h>
32 #include <sys/types.h>
33 #include <sys/tuneable.h>
34 #include <sys/sysmacros.h>
35 #include <sys/systm.h>
36 #include <sys/cpuvar.h>
37 #include <sys/lgrp.h>
38 #include <sys/user.h>
39 #include <sys/proc.h>
40 #include <sys/callo.h>
41 #include <sys/kmem.h>
42 #include <sys/var.h>
43 #include <sys/cmn_err.h>
44 #include <sys/swap.h>
45 #include <sys/vmsystm.h>
46 #include <sys/class.h>
47 #include <sys/time.h>
48 #include <sys/debug.h>
49 #include <sys/vtrace.h>
50 #include <sys/spl.h>
51 #include <sys/atomic.h>
52 #include <sys/dumphdr.h>
53 #include <sys/archsystm.h>
54 #include <sys/fs/swapnode.h>
55 #include <sys/panic.h>
56 #include <sys/disp.h>
57 #include <sys/msacct.h>
58 #include <sys/mem_cage.h>
59 
60 #include <vm/page.h>
61 #include <vm/anon.h>
62 #include <vm/rm.h>
63 #include <sys/cyclic.h>
64 #include <sys/cpupart.h>
65 #include <sys/rctl.h>
66 #include <sys/task.h>
67 #include <sys/sdt.h>
68 #include <sys/ddi_periodic.h>
69 #include <sys/random.h>
70 #include <sys/modctl.h>
71 #include <sys/zone.h>
72 
73 /*
74  * for NTP support
75  */
76 #include <sys/timex.h>
77 #include <sys/inttypes.h>
78 
79 #include <sys/sunddi.h>
80 #include <sys/clock_impl.h>
81 
82 /*
83  * clock() is called straight from the clock cyclic; see clock_init().
84  *
85  * Functions:
86  *	reprime clock
87  *	maintain date
88  *	jab the scheduler
89  */
90 
91 extern kcondvar_t	fsflush_cv;
92 extern sysinfo_t	sysinfo;
93 extern vminfo_t	vminfo;
94 extern int	idleswtch;	/* flag set while idle in pswtch() */
95 extern hrtime_t volatile devinfo_freeze;
96 
97 /*
98  * high-precision avenrun values.  These are needed to make the
99  * regular avenrun values accurate.
100  */
101 static uint64_t hp_avenrun[3];
102 int	avenrun[3];		/* FSCALED average run queue lengths */
103 time_t	time;	/* time in seconds since 1970 - for compatibility only */
104 
105 static struct loadavg_s loadavg;
106 /*
107  * Phase/frequency-lock loop (PLL/FLL) definitions
108  *
109  * The following variables are read and set by the ntp_adjtime() system
110  * call.
111  *
112  * time_state shows the state of the system clock, with values defined
113  * in the timex.h header file.
114  *
115  * time_status shows the status of the system clock, with bits defined
116  * in the timex.h header file.
117  *
118  * time_offset is used by the PLL/FLL to adjust the system time in small
119  * increments.
120  *
121  * time_constant determines the bandwidth or "stiffness" of the PLL.
122  *
123  * time_tolerance determines maximum frequency error or tolerance of the
124  * CPU clock oscillator and is a property of the architecture; however,
125  * in principle it could change as result of the presence of external
126  * discipline signals, for instance.
127  *
128  * time_precision is usually equal to the kernel tick variable; however,
129  * in cases where a precision clock counter or external clock is
130  * available, the resolution can be much less than this and depend on
131  * whether the external clock is working or not.
132  *
133  * time_maxerror is initialized by a ntp_adjtime() call and increased by
134  * the kernel once each second to reflect the maximum error bound
135  * growth.
136  *
137  * time_esterror is set and read by the ntp_adjtime() call, but
138  * otherwise not used by the kernel.
139  */
140 int32_t time_state = TIME_OK;	/* clock state */
141 int32_t time_status = STA_UNSYNC;	/* clock status bits */
142 int32_t time_offset = 0;		/* time offset (us) */
143 int32_t time_constant = 0;		/* pll time constant */
144 int32_t time_tolerance = MAXFREQ;	/* frequency tolerance (scaled ppm) */
145 int32_t time_precision = 1;	/* clock precision (us) */
146 int32_t time_maxerror = MAXPHASE;	/* maximum error (us) */
147 int32_t time_esterror = MAXPHASE;	/* estimated error (us) */
148 
149 /*
150  * The following variables establish the state of the PLL/FLL and the
151  * residual time and frequency offset of the local clock. The scale
152  * factors are defined in the timex.h header file.
153  *
154  * time_phase and time_freq are the phase increment and the frequency
155  * increment, respectively, of the kernel time variable.
156  *
157  * time_freq is set via ntp_adjtime() from a value stored in a file when
158  * the synchronization daemon is first started. Its value is retrieved
159  * via ntp_adjtime() and written to the file about once per hour by the
160  * daemon.
161  *
162  * time_adj is the adjustment added to the value of tick at each timer
163  * interrupt and is recomputed from time_phase and time_freq at each
164  * seconds rollover.
165  *
166  * time_reftime is the second's portion of the system time at the last
167  * call to ntp_adjtime(). It is used to adjust the time_freq variable
168  * and to increase the time_maxerror as the time since last update
169  * increases.
170  */
171 int32_t time_phase = 0;		/* phase offset (scaled us) */
172 int32_t time_freq = 0;		/* frequency offset (scaled ppm) */
173 int32_t time_adj = 0;		/* tick adjust (scaled 1 / hz) */
174 int32_t time_reftime = 0;		/* time at last adjustment (s) */
175 
176 /*
177  * The scale factors of the following variables are defined in the
178  * timex.h header file.
179  *
180  * pps_time contains the time at each calibration interval, as read by
181  * microtime(). pps_count counts the seconds of the calibration
182  * interval, the duration of which is nominally pps_shift in powers of
183  * two.
184  *
185  * pps_offset is the time offset produced by the time median filter
186  * pps_tf[], while pps_jitter is the dispersion (jitter) measured by
187  * this filter.
188  *
189  * pps_freq is the frequency offset produced by the frequency median
190  * filter pps_ff[], while pps_stabil is the dispersion (wander) measured
191  * by this filter.
192  *
193  * pps_usec is latched from a high resolution counter or external clock
194  * at pps_time. Here we want the hardware counter contents only, not the
195  * contents plus the time_tv.usec as usual.
196  *
197  * pps_valid counts the number of seconds since the last PPS update. It
198  * is used as a watchdog timer to disable the PPS discipline should the
199  * PPS signal be lost.
200  *
201  * pps_glitch counts the number of seconds since the beginning of an
202  * offset burst more than tick/2 from current nominal offset. It is used
203  * mainly to suppress error bursts due to priority conflicts between the
204  * PPS interrupt and timer interrupt.
205  *
206  * pps_intcnt counts the calibration intervals for use in the interval-
207  * adaptation algorithm. It's just too complicated for words.
208  */
209 struct timeval pps_time;	/* kernel time at last interval */
210 int32_t pps_tf[] = {0, 0, 0};	/* pps time offset median filter (us) */
211 int32_t pps_offset = 0;		/* pps time offset (us) */
212 int32_t pps_jitter = MAXTIME;	/* time dispersion (jitter) (us) */
213 int32_t pps_ff[] = {0, 0, 0};	/* pps frequency offset median filter */
214 int32_t pps_freq = 0;		/* frequency offset (scaled ppm) */
215 int32_t pps_stabil = MAXFREQ;	/* frequency dispersion (scaled ppm) */
216 int32_t pps_usec = 0;		/* microsec counter at last interval */
217 int32_t pps_valid = PPS_VALID;	/* pps signal watchdog counter */
218 int32_t pps_glitch = 0;		/* pps signal glitch counter */
219 int32_t pps_count = 0;		/* calibration interval counter (s) */
220 int32_t pps_shift = PPS_SHIFT;	/* interval duration (s) (shift) */
221 int32_t pps_intcnt = 0;		/* intervals at current duration */
222 
223 /*
224  * PPS signal quality monitors
225  *
226  * pps_jitcnt counts the seconds that have been discarded because the
227  * jitter measured by the time median filter exceeds the limit MAXTIME
228  * (100 us).
229  *
230  * pps_calcnt counts the frequency calibration intervals, which are
231  * variable from 4 s to 256 s.
232  *
233  * pps_errcnt counts the calibration intervals which have been discarded
234  * because the wander exceeds the limit MAXFREQ (100 ppm) or where the
235  * calibration interval jitter exceeds two ticks.
236  *
237  * pps_stbcnt counts the calibration intervals that have been discarded
238  * because the frequency wander exceeds the limit MAXFREQ / 4 (25 us).
239  */
240 int32_t pps_jitcnt = 0;		/* jitter limit exceeded */
241 int32_t pps_calcnt = 0;		/* calibration intervals */
242 int32_t pps_errcnt = 0;		/* calibration errors */
243 int32_t pps_stbcnt = 0;		/* stability limit exceeded */
244 
245 kcondvar_t lbolt_cv;
246 
247 /*
248  * Hybrid lbolt implementation:
249  *
250  * The service historically provided by the lbolt and lbolt64 variables has
251  * been replaced by the ddi_get_lbolt() and ddi_get_lbolt64() routines, and the
252  * original symbols removed from the system. The once clock driven variables are
253  * now implemented in an event driven fashion, backed by gethrtime() coarsed to
254  * the appropriate clock resolution. The default event driven implementation is
255  * complemented by a cyclic driven one, active only during periods of intense
256  * activity around the DDI lbolt routines, when a lbolt specific cyclic is
257  * reprogramed to fire at a clock tick interval to serve consumers of lbolt who
258  * rely on the original low cost of consulting a memory position.
259  *
260  * The implementation uses the number of calls to these routines and the
261  * frequency of these to determine when to transition from event to cyclic
262  * driven and vice-versa. These values are kept on a per CPU basis for
263  * scalability reasons and to prevent CPUs from constantly invalidating a single
264  * cache line when modifying a global variable. The transition from event to
265  * cyclic mode happens once the thresholds are crossed, and activity on any CPU
266  * can cause such transition.
267  *
268  * The lbolt_hybrid function pointer is called by ddi_get_lbolt() and
269  * ddi_get_lbolt64(), and will point to lbolt_event_driven() or
270  * lbolt_cyclic_driven() according to the current mode. When the thresholds
271  * are exceeded, lbolt_event_driven() will reprogram the lbolt cyclic to
272  * fire at a nsec_per_tick interval and increment an internal variable at
273  * each firing. lbolt_hybrid will then point to lbolt_cyclic_driven(), which
274  * will simply return the value of such variable. lbolt_cyclic() will attempt
275  * to shut itself off at each threshold interval (sampling period for calls
276  * to the DDI lbolt routines), and return to the event driven mode, but will
277  * be prevented from doing so if lbolt_cyclic_driven() is being heavily used.
278  *
279  * lbolt_bootstrap is used during boot to serve lbolt consumers who don't wait
280  * for the cyclic subsystem to be intialized.
281  *
282  */
283 int64_t lbolt_bootstrap(void);
284 int64_t lbolt_event_driven(void);
285 int64_t lbolt_cyclic_driven(void);
286 int64_t (*lbolt_hybrid)(void) = lbolt_bootstrap;
287 uint_t lbolt_ev_to_cyclic(caddr_t, caddr_t);
288 
289 /*
290  * lbolt's cyclic, installed by clock_init().
291  */
292 static void lbolt_cyclic(void);
293 
294 /*
295  * Tunable to keep lbolt in cyclic driven mode. This will prevent the system
296  * from switching back to event driven, once it reaches cyclic mode.
297  */
298 static boolean_t lbolt_cyc_only = B_FALSE;
299 
300 /*
301  * Cache aligned, per CPU structure with lbolt usage statistics.
302  */
303 static lbolt_cpu_t *lb_cpu;
304 
305 /*
306  * Single, cache aligned, structure with all the information required by
307  * the lbolt implementation.
308  */
309 lbolt_info_t *lb_info;
310 
311 
312 int one_sec = 1; /* turned on once every second */
313 static int fsflushcnt;	/* counter for t_fsflushr */
314 int	dosynctodr = 1;	/* patchable; enable/disable sync to TOD chip */
315 int	tod_needsync = 0;	/* need to sync tod chip with software time */
316 static int tod_broken = 0;	/* clock chip doesn't work */
317 time_t	boot_time = 0;		/* Boot time in seconds since 1970 */
318 cyclic_id_t clock_cyclic;	/* clock()'s cyclic_id */
319 cyclic_id_t deadman_cyclic;	/* deadman()'s cyclic_id */
320 
321 extern void clock_tick_schedule(int);
322 extern void set_freemem(void);
323 extern void pageout_deadman(void);
324 
325 static int lgrp_ticks;		/* counter to schedule lgrp load calcs */
326 
327 /*
328  * for tod fault detection
329  */
330 #define	TOD_REF_FREQ		((longlong_t)(NANOSEC))
331 #define	TOD_STALL_THRESHOLD	(TOD_REF_FREQ * 3 / 2)
332 #define	TOD_JUMP_THRESHOLD	(TOD_REF_FREQ / 2)
333 #define	TOD_FILTER_N		4
334 #define	TOD_FILTER_SETTLE	(4 * TOD_FILTER_N)
335 static enum tod_fault_type tod_faulted = TOD_NOFAULT;
336 
337 static int tod_status_flag = 0;		/* used by tod_validate() */
338 
339 static hrtime_t prev_set_tick = 0;	/* gethrtime() prior to tod_set() */
340 static time_t prev_set_tod = 0;		/* tv_sec value passed to tod_set() */
341 
342 /* patchable via /etc/system */
343 int tod_validate_enable = 1;
344 
345 /* Diagnose/Limit messages about delay(9F) called from interrupt context */
346 int			delay_from_interrupt_diagnose = 0;
347 volatile uint32_t	delay_from_interrupt_msg = 20;
348 
349 /*
350  * On non-SPARC systems, TOD validation must be deferred until gethrtime
351  * returns non-zero values (after mach_clkinit's execution).
352  * On SPARC systems, it must be deferred until after hrtime_base
353  * and hres_last_tick are set (in the first invocation of hres_tick).
354  * Since in both cases the prerequisites occur before the invocation of
355  * tod_get() in clock(), the deferment is lifted there.
356  */
357 static boolean_t tod_validate_deferred = B_TRUE;
358 
359 /*
360  * tod_fault_table[] must be aligned with
361  * enum tod_fault_type in systm.h
362  */
363 static char *tod_fault_table[] = {
364 	"Reversed",			/* TOD_REVERSED */
365 	"Stalled",			/* TOD_STALLED */
366 	"Jumped",			/* TOD_JUMPED */
367 	"Changed in Clock Rate",	/* TOD_RATECHANGED */
368 	"Is Read-Only"			/* TOD_RDONLY */
369 	/*
370 	 * no strings needed for TOD_NOFAULT
371 	 */
372 };
373 
374 /*
375  * test hook for tod broken detection in tod_validate
376  */
377 int tod_unit_test = 0;
378 time_t tod_test_injector;
379 
380 #define	CLOCK_ADJ_HIST_SIZE	4
381 
382 static int	adj_hist_entry;
383 
384 int64_t clock_adj_hist[CLOCK_ADJ_HIST_SIZE];
385 
386 static void calcloadavg(int, uint64_t *);
387 static int genloadavg(struct loadavg_s *);
388 static void loadavg_update();
389 
390 void (*cmm_clock_callout)() = NULL;
391 void (*cpucaps_clock_callout)() = NULL;
392 
393 extern clock_t clock_tick_proc_max;
394 
395 static int64_t deadman_counter = 0;
396 
397 static void
398 clock(void)
399 {
400 	kthread_t	*t;
401 	uint_t	nrunnable;
402 	uint_t	w_io;
403 	cpu_t	*cp;
404 	cpupart_t *cpupart;
405 	void	(*funcp)();
406 	int32_t ltemp;
407 	int64_t lltemp;
408 	int s;
409 	int do_lgrp_load;
410 	int i;
411 	clock_t now = LBOLT_NO_ACCOUNT;	/* current tick */
412 
413 	if (panicstr)
414 		return;
415 
416 	/*
417 	 * Make sure that 'freemem' do not drift too far from the truth
418 	 */
419 	set_freemem();
420 
421 
422 	/*
423 	 * Before the section which is repeated is executed, we do
424 	 * the time delta processing which occurs every clock tick
425 	 *
426 	 * There is additional processing which happens every time
427 	 * the nanosecond counter rolls over which is described
428 	 * below - see the section which begins with : if (one_sec)
429 	 *
430 	 * This section marks the beginning of the precision-kernel
431 	 * code fragment.
432 	 *
433 	 * First, compute the phase adjustment. If the low-order bits
434 	 * (time_phase) of the update overflow, bump the higher order
435 	 * bits (time_update).
436 	 */
437 	time_phase += time_adj;
438 	if (time_phase <= -FINEUSEC) {
439 		ltemp = -time_phase / SCALE_PHASE;
440 		time_phase += ltemp * SCALE_PHASE;
441 		s = hr_clock_lock();
442 		timedelta -= ltemp * (NANOSEC/MICROSEC);
443 		hr_clock_unlock(s);
444 	} else if (time_phase >= FINEUSEC) {
445 		ltemp = time_phase / SCALE_PHASE;
446 		time_phase -= ltemp * SCALE_PHASE;
447 		s = hr_clock_lock();
448 		timedelta += ltemp * (NANOSEC/MICROSEC);
449 		hr_clock_unlock(s);
450 	}
451 
452 	/*
453 	 * End of precision-kernel code fragment which is processed
454 	 * every timer interrupt.
455 	 *
456 	 * Continue with the interrupt processing as scheduled.
457 	 */
458 	/*
459 	 * Count the number of runnable threads and the number waiting
460 	 * for some form of I/O to complete -- gets added to
461 	 * sysinfo.waiting.  To know the state of the system, must add
462 	 * wait counts from all CPUs.  Also add up the per-partition
463 	 * statistics.
464 	 */
465 	w_io = 0;
466 	nrunnable = 0;
467 
468 	/*
469 	 * keep track of when to update lgrp/part loads
470 	 */
471 
472 	do_lgrp_load = 0;
473 	if (lgrp_ticks++ >= hz / 10) {
474 		lgrp_ticks = 0;
475 		do_lgrp_load = 1;
476 	}
477 
478 	if (one_sec) {
479 		loadavg_update();
480 		deadman_counter++;
481 		pageout_deadman();
482 	}
483 
484 	/*
485 	 * First count the threads waiting on kpreempt queues in each
486 	 * CPU partition.
487 	 */
488 
489 	cpupart = cp_list_head;
490 	do {
491 		uint_t cpupart_nrunnable = cpupart->cp_kp_queue.disp_nrunnable;
492 
493 		cpupart->cp_updates++;
494 		nrunnable += cpupart_nrunnable;
495 		cpupart->cp_nrunnable_cum += cpupart_nrunnable;
496 		if (one_sec) {
497 			cpupart->cp_nrunning = 0;
498 			cpupart->cp_nrunnable = cpupart_nrunnable;
499 		}
500 	} while ((cpupart = cpupart->cp_next) != cp_list_head);
501 
502 
503 	/* Now count the per-CPU statistics. */
504 	cp = cpu_list;
505 	do {
506 		uint_t cpu_nrunnable = cp->cpu_disp->disp_nrunnable;
507 
508 		nrunnable += cpu_nrunnable;
509 		cpupart = cp->cpu_part;
510 		cpupart->cp_nrunnable_cum += cpu_nrunnable;
511 		if (one_sec) {
512 			cpupart->cp_nrunnable += cpu_nrunnable;
513 			/*
514 			 * Update user, system, and idle cpu times.
515 			 */
516 			cpupart->cp_nrunning++;
517 			/*
518 			 * w_io is used to update sysinfo.waiting during
519 			 * one_second processing below.  Only gather w_io
520 			 * information when we walk the list of cpus if we're
521 			 * going to perform one_second processing.
522 			 */
523 			w_io += CPU_STATS(cp, sys.iowait);
524 		}
525 
526 		if (one_sec && (cp->cpu_flags & CPU_EXISTS)) {
527 			int i, load, change;
528 			hrtime_t intracct, intrused;
529 			const hrtime_t maxnsec = 1000000000;
530 			const int precision = 100;
531 
532 			/*
533 			 * Estimate interrupt load on this cpu each second.
534 			 * Computes cpu_intrload as %utilization (0-99).
535 			 */
536 
537 			/* add up interrupt time from all micro states */
538 			for (intracct = 0, i = 0; i < NCMSTATES; i++)
539 				intracct += cp->cpu_intracct[i];
540 			scalehrtime(&intracct);
541 
542 			/* compute nsec used in the past second */
543 			intrused = intracct - cp->cpu_intrlast;
544 			cp->cpu_intrlast = intracct;
545 
546 			/* limit the value for safety (and the first pass) */
547 			if (intrused >= maxnsec)
548 				intrused = maxnsec - 1;
549 
550 			/* calculate %time in interrupt */
551 			load = (precision * intrused) / maxnsec;
552 			ASSERT(load >= 0 && load < precision);
553 			change = cp->cpu_intrload - load;
554 
555 			/* jump to new max, or decay the old max */
556 			if (change < 0)
557 				cp->cpu_intrload = load;
558 			else if (change > 0)
559 				cp->cpu_intrload -= (change + 3) / 4;
560 
561 			DTRACE_PROBE3(cpu_intrload,
562 			    cpu_t *, cp,
563 			    hrtime_t, intracct,
564 			    hrtime_t, intrused);
565 		}
566 
567 		if (do_lgrp_load &&
568 		    (cp->cpu_flags & CPU_EXISTS)) {
569 			/*
570 			 * When updating the lgroup's load average,
571 			 * account for the thread running on the CPU.
572 			 * If the CPU is the current one, then we need
573 			 * to account for the underlying thread which
574 			 * got the clock interrupt not the thread that is
575 			 * handling the interrupt and caculating the load
576 			 * average
577 			 */
578 			t = cp->cpu_thread;
579 			if (CPU == cp)
580 				t = t->t_intr;
581 
582 			/*
583 			 * Account for the load average for this thread if
584 			 * it isn't the idle thread or it is on the interrupt
585 			 * stack and not the current CPU handling the clock
586 			 * interrupt
587 			 */
588 			if ((t && t != cp->cpu_idle_thread) || (CPU != cp &&
589 			    CPU_ON_INTR(cp))) {
590 				if (t->t_lpl == cp->cpu_lpl) {
591 					/* local thread */
592 					cpu_nrunnable++;
593 				} else {
594 					/*
595 					 * This is a remote thread, charge it
596 					 * against its home lgroup.  Note that
597 					 * we notice that a thread is remote
598 					 * only if it's currently executing.
599 					 * This is a reasonable approximation,
600 					 * since queued remote threads are rare.
601 					 * Note also that if we didn't charge
602 					 * it to its home lgroup, remote
603 					 * execution would often make a system
604 					 * appear balanced even though it was
605 					 * not, and thread placement/migration
606 					 * would often not be done correctly.
607 					 */
608 					lgrp_loadavg(t->t_lpl,
609 					    LGRP_LOADAVG_IN_THREAD_MAX, 0);
610 				}
611 			}
612 			lgrp_loadavg(cp->cpu_lpl,
613 			    cpu_nrunnable * LGRP_LOADAVG_IN_THREAD_MAX, 1);
614 		}
615 	} while ((cp = cp->cpu_next) != cpu_list);
616 
617 	clock_tick_schedule(one_sec);
618 
619 	/*
620 	 * Check for a callout that needs be called from the clock
621 	 * thread to support the membership protocol in a clustered
622 	 * system.  Copy the function pointer so that we can reset
623 	 * this to NULL if needed.
624 	 */
625 	if ((funcp = cmm_clock_callout) != NULL)
626 		(*funcp)();
627 
628 	if ((funcp = cpucaps_clock_callout) != NULL)
629 		(*funcp)();
630 
631 	/*
632 	 * Wakeup the cageout thread waiters once per second.
633 	 */
634 	if (one_sec)
635 		kcage_tick();
636 
637 	if (one_sec) {
638 
639 		int drift, absdrift;
640 		timestruc_t tod;
641 		int s;
642 
643 		/*
644 		 * Beginning of precision-kernel code fragment executed
645 		 * every second.
646 		 *
647 		 * On rollover of the second the phase adjustment to be
648 		 * used for the next second is calculated.  Also, the
649 		 * maximum error is increased by the tolerance.  If the
650 		 * PPS frequency discipline code is present, the phase is
651 		 * increased to compensate for the CPU clock oscillator
652 		 * frequency error.
653 		 *
654 		 * On a 32-bit machine and given parameters in the timex.h
655 		 * header file, the maximum phase adjustment is +-512 ms
656 		 * and maximum frequency offset is (a tad less than)
657 		 * +-512 ppm. On a 64-bit machine, you shouldn't need to ask.
658 		 */
659 		time_maxerror += time_tolerance / SCALE_USEC;
660 
661 		/*
662 		 * Leap second processing. If in leap-insert state at
663 		 * the end of the day, the system clock is set back one
664 		 * second; if in leap-delete state, the system clock is
665 		 * set ahead one second. The microtime() routine or
666 		 * external clock driver will insure that reported time
667 		 * is always monotonic. The ugly divides should be
668 		 * replaced.
669 		 */
670 		switch (time_state) {
671 
672 		case TIME_OK:
673 			if (time_status & STA_INS)
674 				time_state = TIME_INS;
675 			else if (time_status & STA_DEL)
676 				time_state = TIME_DEL;
677 			break;
678 
679 		case TIME_INS:
680 			if (hrestime.tv_sec % 86400 == 0) {
681 				s = hr_clock_lock();
682 				hrestime.tv_sec--;
683 				hr_clock_unlock(s);
684 				time_state = TIME_OOP;
685 			}
686 			break;
687 
688 		case TIME_DEL:
689 			if ((hrestime.tv_sec + 1) % 86400 == 0) {
690 				s = hr_clock_lock();
691 				hrestime.tv_sec++;
692 				hr_clock_unlock(s);
693 				time_state = TIME_WAIT;
694 			}
695 			break;
696 
697 		case TIME_OOP:
698 			time_state = TIME_WAIT;
699 			break;
700 
701 		case TIME_WAIT:
702 			if (!(time_status & (STA_INS | STA_DEL)))
703 				time_state = TIME_OK;
704 		default:
705 			break;
706 		}
707 
708 		/*
709 		 * Compute the phase adjustment for the next second. In
710 		 * PLL mode, the offset is reduced by a fixed factor
711 		 * times the time constant. In FLL mode the offset is
712 		 * used directly. In either mode, the maximum phase
713 		 * adjustment for each second is clamped so as to spread
714 		 * the adjustment over not more than the number of
715 		 * seconds between updates.
716 		 */
717 		if (time_offset == 0)
718 			time_adj = 0;
719 		else if (time_offset < 0) {
720 			lltemp = -time_offset;
721 			if (!(time_status & STA_FLL)) {
722 				if ((1 << time_constant) >= SCALE_KG)
723 					lltemp *= (1 << time_constant) /
724 					    SCALE_KG;
725 				else
726 					lltemp = (lltemp / SCALE_KG) >>
727 					    time_constant;
728 			}
729 			if (lltemp > (MAXPHASE / MINSEC) * SCALE_UPDATE)
730 				lltemp = (MAXPHASE / MINSEC) * SCALE_UPDATE;
731 			time_offset += lltemp;
732 			time_adj = -(lltemp * SCALE_PHASE) / hz / SCALE_UPDATE;
733 		} else {
734 			lltemp = time_offset;
735 			if (!(time_status & STA_FLL)) {
736 				if ((1 << time_constant) >= SCALE_KG)
737 					lltemp *= (1 << time_constant) /
738 					    SCALE_KG;
739 				else
740 					lltemp = (lltemp / SCALE_KG) >>
741 					    time_constant;
742 			}
743 			if (lltemp > (MAXPHASE / MINSEC) * SCALE_UPDATE)
744 				lltemp = (MAXPHASE / MINSEC) * SCALE_UPDATE;
745 			time_offset -= lltemp;
746 			time_adj = (lltemp * SCALE_PHASE) / hz / SCALE_UPDATE;
747 		}
748 
749 		/*
750 		 * Compute the frequency estimate and additional phase
751 		 * adjustment due to frequency error for the next
752 		 * second. When the PPS signal is engaged, gnaw on the
753 		 * watchdog counter and update the frequency computed by
754 		 * the pll and the PPS signal.
755 		 */
756 		pps_valid++;
757 		if (pps_valid == PPS_VALID) {
758 			pps_jitter = MAXTIME;
759 			pps_stabil = MAXFREQ;
760 			time_status &= ~(STA_PPSSIGNAL | STA_PPSJITTER |
761 			    STA_PPSWANDER | STA_PPSERROR);
762 		}
763 		lltemp = time_freq + pps_freq;
764 
765 		if (lltemp)
766 			time_adj += (lltemp * SCALE_PHASE) / (SCALE_USEC * hz);
767 
768 		/*
769 		 * End of precision kernel-code fragment
770 		 *
771 		 * The section below should be modified if we are planning
772 		 * to use NTP for synchronization.
773 		 *
774 		 * Note: the clock synchronization code now assumes
775 		 * the following:
776 		 *   - if dosynctodr is 1, then compute the drift between
777 		 *	the tod chip and software time and adjust one or
778 		 *	the other depending on the circumstances
779 		 *
780 		 *   - if dosynctodr is 0, then the tod chip is independent
781 		 *	of the software clock and should not be adjusted,
782 		 *	but allowed to free run.  this allows NTP to sync.
783 		 *	hrestime without any interference from the tod chip.
784 		 */
785 
786 		tod_validate_deferred = B_FALSE;
787 		mutex_enter(&tod_lock);
788 		tod = tod_get();
789 		drift = tod.tv_sec - hrestime.tv_sec;
790 		absdrift = (drift >= 0) ? drift : -drift;
791 		if (tod_needsync || absdrift > 1) {
792 			int s;
793 			if (absdrift > 2) {
794 				if (!tod_broken && tod_faulted == TOD_NOFAULT) {
795 					s = hr_clock_lock();
796 					hrestime = tod;
797 					membar_enter();	/* hrestime visible */
798 					timedelta = 0;
799 					timechanged++;
800 					tod_needsync = 0;
801 					hr_clock_unlock(s);
802 					callout_hrestime();
803 
804 				}
805 			} else {
806 				if (tod_needsync || !dosynctodr) {
807 					gethrestime(&tod);
808 					tod_set(tod);
809 					s = hr_clock_lock();
810 					if (timedelta == 0)
811 						tod_needsync = 0;
812 					hr_clock_unlock(s);
813 				} else {
814 					/*
815 					 * If the drift is 2 seconds on the
816 					 * money, then the TOD is adjusting
817 					 * the clock;  record that.
818 					 */
819 					clock_adj_hist[adj_hist_entry++ %
820 					    CLOCK_ADJ_HIST_SIZE] = now;
821 					s = hr_clock_lock();
822 					timedelta = (int64_t)drift*NANOSEC;
823 					hr_clock_unlock(s);
824 				}
825 			}
826 		}
827 		one_sec = 0;
828 		time = gethrestime_sec();  /* for crusty old kmem readers */
829 		mutex_exit(&tod_lock);
830 
831 		/*
832 		 * Some drivers still depend on this... XXX
833 		 */
834 		cv_broadcast(&lbolt_cv);
835 
836 		vminfo.freemem += freemem;
837 		{
838 			pgcnt_t maxswap, resv, free;
839 			pgcnt_t avail =
840 			    MAX((spgcnt_t)(availrmem - swapfs_minfree), 0);
841 
842 			maxswap = k_anoninfo.ani_mem_resv +
843 			    k_anoninfo.ani_max +avail;
844 			/* Update ani_free */
845 			set_anoninfo();
846 			free = k_anoninfo.ani_free + avail;
847 			resv = k_anoninfo.ani_phys_resv +
848 			    k_anoninfo.ani_mem_resv;
849 
850 			vminfo.swap_resv += resv;
851 			/* number of reserved and allocated pages */
852 #ifdef	DEBUG
853 			if (maxswap < free)
854 				cmn_err(CE_WARN, "clock: maxswap < free");
855 			if (maxswap < resv)
856 				cmn_err(CE_WARN, "clock: maxswap < resv");
857 #endif
858 			vminfo.swap_alloc += maxswap - free;
859 			vminfo.swap_avail += maxswap - resv;
860 			vminfo.swap_free += free;
861 		}
862 		vminfo.updates++;
863 		if (nrunnable) {
864 			sysinfo.runque += nrunnable;
865 			sysinfo.runocc++;
866 		}
867 		if (nswapped) {
868 			sysinfo.swpque += nswapped;
869 			sysinfo.swpocc++;
870 		}
871 		sysinfo.waiting += w_io;
872 		sysinfo.updates++;
873 
874 		/*
875 		 * Wake up fsflush to write out DELWRI
876 		 * buffers, dirty pages and other cached
877 		 * administrative data, e.g. inodes.
878 		 */
879 		if (--fsflushcnt <= 0) {
880 			fsflushcnt = tune.t_fsflushr;
881 			cv_signal(&fsflush_cv);
882 		}
883 
884 		vmmeter();
885 		calcloadavg(genloadavg(&loadavg), hp_avenrun);
886 		for (i = 0; i < 3; i++)
887 			/*
888 			 * At the moment avenrun[] can only hold 31
889 			 * bits of load average as it is a signed
890 			 * int in the API. We need to ensure that
891 			 * hp_avenrun[i] >> (16 - FSHIFT) will not be
892 			 * too large. If it is, we put the largest value
893 			 * that we can use into avenrun[i]. This is
894 			 * kludgey, but about all we can do until we
895 			 * avenrun[] is declared as an array of uint64[]
896 			 */
897 			if (hp_avenrun[i] < ((uint64_t)1<<(31+16-FSHIFT)))
898 				avenrun[i] = (int32_t)(hp_avenrun[i] >>
899 				    (16 - FSHIFT));
900 			else
901 				avenrun[i] = 0x7fffffff;
902 
903 		cpupart = cp_list_head;
904 		do {
905 			calcloadavg(genloadavg(&cpupart->cp_loadavg),
906 			    cpupart->cp_hp_avenrun);
907 		} while ((cpupart = cpupart->cp_next) != cp_list_head);
908 
909 		/*
910 		 * Wake up the swapper thread if necessary.
911 		 */
912 		if (runin ||
913 		    (runout && (avefree < desfree || wake_sched_sec))) {
914 			t = &t0;
915 			thread_lock(t);
916 			if (t->t_state == TS_STOPPED) {
917 				runin = runout = 0;
918 				wake_sched_sec = 0;
919 				t->t_whystop = 0;
920 				t->t_whatstop = 0;
921 				t->t_schedflag &= ~TS_ALLSTART;
922 				THREAD_TRANSITION(t);
923 				setfrontdq(t);
924 			}
925 			thread_unlock(t);
926 		}
927 	}
928 
929 	/*
930 	 * Wake up the swapper if any high priority swapped-out threads
931 	 * became runable during the last tick.
932 	 */
933 	if (wake_sched) {
934 		t = &t0;
935 		thread_lock(t);
936 		if (t->t_state == TS_STOPPED) {
937 			runin = runout = 0;
938 			wake_sched = 0;
939 			t->t_whystop = 0;
940 			t->t_whatstop = 0;
941 			t->t_schedflag &= ~TS_ALLSTART;
942 			THREAD_TRANSITION(t);
943 			setfrontdq(t);
944 		}
945 		thread_unlock(t);
946 	}
947 }
948 
949 void
950 clock_init(void)
951 {
952 	cyc_handler_t clk_hdlr, lbolt_hdlr;
953 	cyc_time_t clk_when, lbolt_when;
954 	int i, sz;
955 	intptr_t buf;
956 
957 	/*
958 	 * Setup handler and timer for the clock cyclic.
959 	 */
960 	clk_hdlr.cyh_func = (cyc_func_t)clock;
961 	clk_hdlr.cyh_level = CY_LOCK_LEVEL;
962 	clk_hdlr.cyh_arg = NULL;
963 
964 	clk_when.cyt_when = 0;
965 	clk_when.cyt_interval = nsec_per_tick;
966 
967 	/*
968 	 * The lbolt cyclic will be reprogramed to fire at a nsec_per_tick
969 	 * interval to satisfy performance needs of the DDI lbolt consumers.
970 	 * It is off by default.
971 	 */
972 	lbolt_hdlr.cyh_func = (cyc_func_t)lbolt_cyclic;
973 	lbolt_hdlr.cyh_level = CY_LOCK_LEVEL;
974 	lbolt_hdlr.cyh_arg = NULL;
975 
976 	lbolt_when.cyt_interval = nsec_per_tick;
977 
978 	/*
979 	 * Allocate cache line aligned space for the per CPU lbolt data and
980 	 * lbolt info structures, and initialize them with their default
981 	 * values. Note that these structures are also cache line sized.
982 	 */
983 	sz = sizeof (lbolt_info_t) + CPU_CACHE_COHERENCE_SIZE;
984 	buf = (intptr_t)kmem_zalloc(sz, KM_SLEEP);
985 	lb_info = (lbolt_info_t *)P2ROUNDUP(buf, CPU_CACHE_COHERENCE_SIZE);
986 
987 	if (hz != HZ_DEFAULT)
988 		lb_info->lbi_thresh_interval = LBOLT_THRESH_INTERVAL *
989 		    hz/HZ_DEFAULT;
990 	else
991 		lb_info->lbi_thresh_interval = LBOLT_THRESH_INTERVAL;
992 
993 	lb_info->lbi_thresh_calls = LBOLT_THRESH_CALLS;
994 
995 	sz = (sizeof (lbolt_cpu_t) * max_ncpus) + CPU_CACHE_COHERENCE_SIZE;
996 	buf = (intptr_t)kmem_zalloc(sz, KM_SLEEP);
997 	lb_cpu = (lbolt_cpu_t *)P2ROUNDUP(buf, CPU_CACHE_COHERENCE_SIZE);
998 
999 	for (i = 0; i < max_ncpus; i++)
1000 		lb_cpu[i].lbc_counter = lb_info->lbi_thresh_calls;
1001 
1002 	/*
1003 	 * Install the softint used to switch between event and cyclic driven
1004 	 * lbolt. We use a soft interrupt to make sure the context of the
1005 	 * cyclic reprogram call is safe.
1006 	 */
1007 	lbolt_softint_add();
1008 
1009 	/*
1010 	 * Since the hybrid lbolt implementation is based on a hardware counter
1011 	 * that is reset at every hardware reboot and that we'd like to have
1012 	 * the lbolt value starting at zero after both a hardware and a fast
1013 	 * reboot, we calculate the number of clock ticks the system's been up
1014 	 * and store it in the lbi_debug_time field of the lbolt info structure.
1015 	 * The value of this field will be subtracted from lbolt before
1016 	 * returning it.
1017 	 */
1018 	lb_info->lbi_internal = lb_info->lbi_debug_time =
1019 	    (gethrtime()/nsec_per_tick);
1020 
1021 	/*
1022 	 * lbolt_hybrid points at lbolt_bootstrap until now. The LBOLT_* macros
1023 	 * and lbolt_debug_{enter,return} use this value as an indication that
1024 	 * the initializaion above hasn't been completed. Setting lbolt_hybrid
1025 	 * to either lbolt_{cyclic,event}_driven here signals those code paths
1026 	 * that the lbolt related structures can be used.
1027 	 */
1028 	if (lbolt_cyc_only) {
1029 		lbolt_when.cyt_when = 0;
1030 		lbolt_hybrid = lbolt_cyclic_driven;
1031 	} else {
1032 		lbolt_when.cyt_when = CY_INFINITY;
1033 		lbolt_hybrid = lbolt_event_driven;
1034 	}
1035 
1036 	/*
1037 	 * Grab cpu_lock and install all three cyclics.
1038 	 */
1039 	mutex_enter(&cpu_lock);
1040 
1041 	clock_cyclic = cyclic_add(&clk_hdlr, &clk_when);
1042 	lb_info->id.lbi_cyclic_id = cyclic_add(&lbolt_hdlr, &lbolt_when);
1043 
1044 	mutex_exit(&cpu_lock);
1045 }
1046 
1047 /*
1048  * Called before calcloadavg to get 10-sec moving loadavg together
1049  */
1050 
1051 static int
1052 genloadavg(struct loadavg_s *avgs)
1053 {
1054 	int avg;
1055 	int spos; /* starting position */
1056 	int cpos; /* moving current position */
1057 	int i;
1058 	int slen;
1059 	hrtime_t hr_avg;
1060 
1061 	/* 10-second snapshot, calculate first positon */
1062 	if (avgs->lg_len == 0) {
1063 		return (0);
1064 	}
1065 	slen = avgs->lg_len < S_MOVAVG_SZ ? avgs->lg_len : S_MOVAVG_SZ;
1066 
1067 	spos = (avgs->lg_cur - 1) >= 0 ? avgs->lg_cur - 1 :
1068 	    S_LOADAVG_SZ + (avgs->lg_cur - 1);
1069 	for (i = hr_avg = 0; i < slen; i++) {
1070 		cpos = (spos - i) >= 0 ? spos - i : S_LOADAVG_SZ + (spos - i);
1071 		hr_avg += avgs->lg_loads[cpos];
1072 	}
1073 
1074 	hr_avg = hr_avg / slen;
1075 	avg = hr_avg / (NANOSEC / LGRP_LOADAVG_IN_THREAD_MAX);
1076 
1077 	return (avg);
1078 }
1079 
1080 /*
1081  * Run every second from clock () to update the loadavg count available to the
1082  * system and cpu-partitions.
1083  *
1084  * This works by sampling the previous usr, sys, wait time elapsed,
1085  * computing a delta, and adding that delta to the elapsed usr, sys,
1086  * wait increase.
1087  */
1088 
1089 static void
1090 loadavg_update()
1091 {
1092 	cpu_t *cp;
1093 	cpupart_t *cpupart;
1094 	hrtime_t cpu_total;
1095 	int prev;
1096 
1097 	cp = cpu_list;
1098 	loadavg.lg_total = 0;
1099 
1100 	/*
1101 	 * first pass totals up per-cpu statistics for system and cpu
1102 	 * partitions
1103 	 */
1104 
1105 	do {
1106 		struct loadavg_s *lavg;
1107 
1108 		lavg = &cp->cpu_loadavg;
1109 
1110 		cpu_total = cp->cpu_acct[CMS_USER] +
1111 		    cp->cpu_acct[CMS_SYSTEM] + cp->cpu_waitrq;
1112 		/* compute delta against last total */
1113 		scalehrtime(&cpu_total);
1114 		prev = (lavg->lg_cur - 1) >= 0 ? lavg->lg_cur - 1 :
1115 		    S_LOADAVG_SZ + (lavg->lg_cur - 1);
1116 		if (lavg->lg_loads[prev] <= 0) {
1117 			lavg->lg_loads[lavg->lg_cur] = cpu_total;
1118 			cpu_total = 0;
1119 		} else {
1120 			lavg->lg_loads[lavg->lg_cur] = cpu_total;
1121 			cpu_total = cpu_total - lavg->lg_loads[prev];
1122 			if (cpu_total < 0)
1123 				cpu_total = 0;
1124 		}
1125 
1126 		lavg->lg_cur = (lavg->lg_cur + 1) % S_LOADAVG_SZ;
1127 		lavg->lg_len = (lavg->lg_len + 1) < S_LOADAVG_SZ ?
1128 		    lavg->lg_len + 1 : S_LOADAVG_SZ;
1129 
1130 		loadavg.lg_total += cpu_total;
1131 		cp->cpu_part->cp_loadavg.lg_total += cpu_total;
1132 
1133 	} while ((cp = cp->cpu_next) != cpu_list);
1134 
1135 	loadavg.lg_loads[loadavg.lg_cur] = loadavg.lg_total;
1136 	loadavg.lg_cur = (loadavg.lg_cur + 1) % S_LOADAVG_SZ;
1137 	loadavg.lg_len = (loadavg.lg_len + 1) < S_LOADAVG_SZ ?
1138 	    loadavg.lg_len + 1 : S_LOADAVG_SZ;
1139 	/*
1140 	 * Second pass updates counts
1141 	 */
1142 	cpupart = cp_list_head;
1143 
1144 	do {
1145 		struct loadavg_s *lavg;
1146 
1147 		lavg = &cpupart->cp_loadavg;
1148 		lavg->lg_loads[lavg->lg_cur] = lavg->lg_total;
1149 		lavg->lg_total = 0;
1150 		lavg->lg_cur = (lavg->lg_cur + 1) % S_LOADAVG_SZ;
1151 		lavg->lg_len = (lavg->lg_len + 1) < S_LOADAVG_SZ ?
1152 		    lavg->lg_len + 1 : S_LOADAVG_SZ;
1153 
1154 	} while ((cpupart = cpupart->cp_next) != cp_list_head);
1155 
1156 	/*
1157 	 * Third pass totals up per-zone statistics.
1158 	 */
1159 	zone_loadavg_update();
1160 }
1161 
1162 /*
1163  * clock_update() - local clock update
1164  *
1165  * This routine is called by ntp_adjtime() to update the local clock
1166  * phase and frequency. The implementation is of an
1167  * adaptive-parameter, hybrid phase/frequency-lock loop (PLL/FLL). The
1168  * routine computes new time and frequency offset estimates for each
1169  * call.  The PPS signal itself determines the new time offset,
1170  * instead of the calling argument.  Presumably, calls to
1171  * ntp_adjtime() occur only when the caller believes the local clock
1172  * is valid within some bound (+-128 ms with NTP). If the caller's
1173  * time is far different than the PPS time, an argument will ensue,
1174  * and it's not clear who will lose.
1175  *
1176  * For uncompensated quartz crystal oscillatores and nominal update
1177  * intervals less than 1024 s, operation should be in phase-lock mode
1178  * (STA_FLL = 0), where the loop is disciplined to phase. For update
1179  * intervals greater than this, operation should be in frequency-lock
1180  * mode (STA_FLL = 1), where the loop is disciplined to frequency.
1181  *
1182  * Note: mutex(&tod_lock) is in effect.
1183  */
1184 void
1185 clock_update(int offset)
1186 {
1187 	int ltemp, mtemp, s;
1188 
1189 	ASSERT(MUTEX_HELD(&tod_lock));
1190 
1191 	if (!(time_status & STA_PLL) && !(time_status & STA_PPSTIME))
1192 		return;
1193 	ltemp = offset;
1194 	if ((time_status & STA_PPSTIME) && (time_status & STA_PPSSIGNAL))
1195 		ltemp = pps_offset;
1196 
1197 	/*
1198 	 * Scale the phase adjustment and clamp to the operating range.
1199 	 */
1200 	if (ltemp > MAXPHASE)
1201 		time_offset = MAXPHASE * SCALE_UPDATE;
1202 	else if (ltemp < -MAXPHASE)
1203 		time_offset = -(MAXPHASE * SCALE_UPDATE);
1204 	else
1205 		time_offset = ltemp * SCALE_UPDATE;
1206 
1207 	/*
1208 	 * Select whether the frequency is to be controlled and in which
1209 	 * mode (PLL or FLL). Clamp to the operating range. Ugly
1210 	 * multiply/divide should be replaced someday.
1211 	 */
1212 	if (time_status & STA_FREQHOLD || time_reftime == 0)
1213 		time_reftime = hrestime.tv_sec;
1214 
1215 	mtemp = hrestime.tv_sec - time_reftime;
1216 	time_reftime = hrestime.tv_sec;
1217 
1218 	if (time_status & STA_FLL) {
1219 		if (mtemp >= MINSEC) {
1220 			ltemp = ((time_offset / mtemp) * (SCALE_USEC /
1221 			    SCALE_UPDATE));
1222 			if (ltemp)
1223 				time_freq += ltemp / SCALE_KH;
1224 		}
1225 	} else {
1226 		if (mtemp < MAXSEC) {
1227 			ltemp *= mtemp;
1228 			if (ltemp)
1229 				time_freq += (int)(((int64_t)ltemp *
1230 				    SCALE_USEC) / SCALE_KF)
1231 				    / (1 << (time_constant * 2));
1232 		}
1233 	}
1234 	if (time_freq > time_tolerance)
1235 		time_freq = time_tolerance;
1236 	else if (time_freq < -time_tolerance)
1237 		time_freq = -time_tolerance;
1238 
1239 	s = hr_clock_lock();
1240 	tod_needsync = 1;
1241 	hr_clock_unlock(s);
1242 }
1243 
1244 /*
1245  * ddi_hardpps() - discipline CPU clock oscillator to external PPS signal
1246  *
1247  * This routine is called at each PPS interrupt in order to discipline
1248  * the CPU clock oscillator to the PPS signal. It measures the PPS phase
1249  * and leaves it in a handy spot for the clock() routine. It
1250  * integrates successive PPS phase differences and calculates the
1251  * frequency offset. This is used in clock() to discipline the CPU
1252  * clock oscillator so that intrinsic frequency error is cancelled out.
1253  * The code requires the caller to capture the time and hardware counter
1254  * value at the on-time PPS signal transition.
1255  *
1256  * Note that, on some Unix systems, this routine runs at an interrupt
1257  * priority level higher than the timer interrupt routine clock().
1258  * Therefore, the variables used are distinct from the clock()
1259  * variables, except for certain exceptions: The PPS frequency pps_freq
1260  * and phase pps_offset variables are determined by this routine and
1261  * updated atomically. The time_tolerance variable can be considered a
1262  * constant, since it is infrequently changed, and then only when the
1263  * PPS signal is disabled. The watchdog counter pps_valid is updated
1264  * once per second by clock() and is atomically cleared in this
1265  * routine.
1266  *
1267  * tvp is the time of the last tick; usec is a microsecond count since the
1268  * last tick.
1269  *
1270  * Note: In Solaris systems, the tick value is actually given by
1271  *       usec_per_tick.  This is called from the serial driver cdintr(),
1272  *	 or equivalent, at a high PIL.  Because the kernel keeps a
1273  *	 highresolution time, the following code can accept either
1274  *	 the traditional argument pair, or the current highres timestamp
1275  *       in tvp and zero in usec.
1276  */
1277 void
1278 ddi_hardpps(struct timeval *tvp, int usec)
1279 {
1280 	int u_usec, v_usec, bigtick;
1281 	time_t cal_sec;
1282 	int cal_usec;
1283 
1284 	/*
1285 	 * An occasional glitch can be produced when the PPS interrupt
1286 	 * occurs in the clock() routine before the time variable is
1287 	 * updated. Here the offset is discarded when the difference
1288 	 * between it and the last one is greater than tick/2, but not
1289 	 * if the interval since the first discard exceeds 30 s.
1290 	 */
1291 	time_status |= STA_PPSSIGNAL;
1292 	time_status &= ~(STA_PPSJITTER | STA_PPSWANDER | STA_PPSERROR);
1293 	pps_valid = 0;
1294 	u_usec = -tvp->tv_usec;
1295 	if (u_usec < -(MICROSEC/2))
1296 		u_usec += MICROSEC;
1297 	v_usec = pps_offset - u_usec;
1298 	if (v_usec < 0)
1299 		v_usec = -v_usec;
1300 	if (v_usec > (usec_per_tick >> 1)) {
1301 		if (pps_glitch > MAXGLITCH) {
1302 			pps_glitch = 0;
1303 			pps_tf[2] = u_usec;
1304 			pps_tf[1] = u_usec;
1305 		} else {
1306 			pps_glitch++;
1307 			u_usec = pps_offset;
1308 		}
1309 	} else
1310 		pps_glitch = 0;
1311 
1312 	/*
1313 	 * A three-stage median filter is used to help deglitch the pps
1314 	 * time. The median sample becomes the time offset estimate; the
1315 	 * difference between the other two samples becomes the time
1316 	 * dispersion (jitter) estimate.
1317 	 */
1318 	pps_tf[2] = pps_tf[1];
1319 	pps_tf[1] = pps_tf[0];
1320 	pps_tf[0] = u_usec;
1321 	if (pps_tf[0] > pps_tf[1]) {
1322 		if (pps_tf[1] > pps_tf[2]) {
1323 			pps_offset = pps_tf[1];		/* 0 1 2 */
1324 			v_usec = pps_tf[0] - pps_tf[2];
1325 		} else if (pps_tf[2] > pps_tf[0]) {
1326 			pps_offset = pps_tf[0];		/* 2 0 1 */
1327 			v_usec = pps_tf[2] - pps_tf[1];
1328 		} else {
1329 			pps_offset = pps_tf[2];		/* 0 2 1 */
1330 			v_usec = pps_tf[0] - pps_tf[1];
1331 		}
1332 	} else {
1333 		if (pps_tf[1] < pps_tf[2]) {
1334 			pps_offset = pps_tf[1];		/* 2 1 0 */
1335 			v_usec = pps_tf[2] - pps_tf[0];
1336 		} else  if (pps_tf[2] < pps_tf[0]) {
1337 			pps_offset = pps_tf[0];		/* 1 0 2 */
1338 			v_usec = pps_tf[1] - pps_tf[2];
1339 		} else {
1340 			pps_offset = pps_tf[2];		/* 1 2 0 */
1341 			v_usec = pps_tf[1] - pps_tf[0];
1342 		}
1343 	}
1344 	if (v_usec > MAXTIME)
1345 		pps_jitcnt++;
1346 	v_usec = (v_usec << PPS_AVG) - pps_jitter;
1347 	pps_jitter += v_usec / (1 << PPS_AVG);
1348 	if (pps_jitter > (MAXTIME >> 1))
1349 		time_status |= STA_PPSJITTER;
1350 
1351 	/*
1352 	 * During the calibration interval adjust the starting time when
1353 	 * the tick overflows. At the end of the interval compute the
1354 	 * duration of the interval and the difference of the hardware
1355 	 * counters at the beginning and end of the interval. This code
1356 	 * is deliciously complicated by the fact valid differences may
1357 	 * exceed the value of tick when using long calibration
1358 	 * intervals and small ticks. Note that the counter can be
1359 	 * greater than tick if caught at just the wrong instant, but
1360 	 * the values returned and used here are correct.
1361 	 */
1362 	bigtick = (int)usec_per_tick * SCALE_USEC;
1363 	pps_usec -= pps_freq;
1364 	if (pps_usec >= bigtick)
1365 		pps_usec -= bigtick;
1366 	if (pps_usec < 0)
1367 		pps_usec += bigtick;
1368 	pps_time.tv_sec++;
1369 	pps_count++;
1370 	if (pps_count < (1 << pps_shift))
1371 		return;
1372 	pps_count = 0;
1373 	pps_calcnt++;
1374 	u_usec = usec * SCALE_USEC;
1375 	v_usec = pps_usec - u_usec;
1376 	if (v_usec >= bigtick >> 1)
1377 		v_usec -= bigtick;
1378 	if (v_usec < -(bigtick >> 1))
1379 		v_usec += bigtick;
1380 	if (v_usec < 0)
1381 		v_usec = -(-v_usec >> pps_shift);
1382 	else
1383 		v_usec = v_usec >> pps_shift;
1384 	pps_usec = u_usec;
1385 	cal_sec = tvp->tv_sec;
1386 	cal_usec = tvp->tv_usec;
1387 	cal_sec -= pps_time.tv_sec;
1388 	cal_usec -= pps_time.tv_usec;
1389 	if (cal_usec < 0) {
1390 		cal_usec += MICROSEC;
1391 		cal_sec--;
1392 	}
1393 	pps_time = *tvp;
1394 
1395 	/*
1396 	 * Check for lost interrupts, noise, excessive jitter and
1397 	 * excessive frequency error. The number of timer ticks during
1398 	 * the interval may vary +-1 tick. Add to this a margin of one
1399 	 * tick for the PPS signal jitter and maximum frequency
1400 	 * deviation. If the limits are exceeded, the calibration
1401 	 * interval is reset to the minimum and we start over.
1402 	 */
1403 	u_usec = (int)usec_per_tick << 1;
1404 	if (!((cal_sec == -1 && cal_usec > (MICROSEC - u_usec)) ||
1405 	    (cal_sec == 0 && cal_usec < u_usec)) ||
1406 	    v_usec > time_tolerance || v_usec < -time_tolerance) {
1407 		pps_errcnt++;
1408 		pps_shift = PPS_SHIFT;
1409 		pps_intcnt = 0;
1410 		time_status |= STA_PPSERROR;
1411 		return;
1412 	}
1413 
1414 	/*
1415 	 * A three-stage median filter is used to help deglitch the pps
1416 	 * frequency. The median sample becomes the frequency offset
1417 	 * estimate; the difference between the other two samples
1418 	 * becomes the frequency dispersion (stability) estimate.
1419 	 */
1420 	pps_ff[2] = pps_ff[1];
1421 	pps_ff[1] = pps_ff[0];
1422 	pps_ff[0] = v_usec;
1423 	if (pps_ff[0] > pps_ff[1]) {
1424 		if (pps_ff[1] > pps_ff[2]) {
1425 			u_usec = pps_ff[1];		/* 0 1 2 */
1426 			v_usec = pps_ff[0] - pps_ff[2];
1427 		} else if (pps_ff[2] > pps_ff[0]) {
1428 			u_usec = pps_ff[0];		/* 2 0 1 */
1429 			v_usec = pps_ff[2] - pps_ff[1];
1430 		} else {
1431 			u_usec = pps_ff[2];		/* 0 2 1 */
1432 			v_usec = pps_ff[0] - pps_ff[1];
1433 		}
1434 	} else {
1435 		if (pps_ff[1] < pps_ff[2]) {
1436 			u_usec = pps_ff[1];		/* 2 1 0 */
1437 			v_usec = pps_ff[2] - pps_ff[0];
1438 		} else  if (pps_ff[2] < pps_ff[0]) {
1439 			u_usec = pps_ff[0];		/* 1 0 2 */
1440 			v_usec = pps_ff[1] - pps_ff[2];
1441 		} else {
1442 			u_usec = pps_ff[2];		/* 1 2 0 */
1443 			v_usec = pps_ff[1] - pps_ff[0];
1444 		}
1445 	}
1446 
1447 	/*
1448 	 * Here the frequency dispersion (stability) is updated. If it
1449 	 * is less than one-fourth the maximum (MAXFREQ), the frequency
1450 	 * offset is updated as well, but clamped to the tolerance. It
1451 	 * will be processed later by the clock() routine.
1452 	 */
1453 	v_usec = (v_usec >> 1) - pps_stabil;
1454 	if (v_usec < 0)
1455 		pps_stabil -= -v_usec >> PPS_AVG;
1456 	else
1457 		pps_stabil += v_usec >> PPS_AVG;
1458 	if (pps_stabil > MAXFREQ >> 2) {
1459 		pps_stbcnt++;
1460 		time_status |= STA_PPSWANDER;
1461 		return;
1462 	}
1463 	if (time_status & STA_PPSFREQ) {
1464 		if (u_usec < 0) {
1465 			pps_freq -= -u_usec >> PPS_AVG;
1466 			if (pps_freq < -time_tolerance)
1467 				pps_freq = -time_tolerance;
1468 			u_usec = -u_usec;
1469 		} else {
1470 			pps_freq += u_usec >> PPS_AVG;
1471 			if (pps_freq > time_tolerance)
1472 				pps_freq = time_tolerance;
1473 		}
1474 	}
1475 
1476 	/*
1477 	 * Here the calibration interval is adjusted. If the maximum
1478 	 * time difference is greater than tick / 4, reduce the interval
1479 	 * by half. If this is not the case for four consecutive
1480 	 * intervals, double the interval.
1481 	 */
1482 	if (u_usec << pps_shift > bigtick >> 2) {
1483 		pps_intcnt = 0;
1484 		if (pps_shift > PPS_SHIFT)
1485 			pps_shift--;
1486 	} else if (pps_intcnt >= 4) {
1487 		pps_intcnt = 0;
1488 		if (pps_shift < PPS_SHIFTMAX)
1489 			pps_shift++;
1490 	} else
1491 		pps_intcnt++;
1492 
1493 	/*
1494 	 * If recovering from kmdb, then make sure the tod chip gets resynced.
1495 	 * If we took an early exit above, then we don't yet have a stable
1496 	 * calibration signal to lock onto, so don't mark the tod for sync
1497 	 * until we get all the way here.
1498 	 */
1499 	{
1500 		int s = hr_clock_lock();
1501 
1502 		tod_needsync = 1;
1503 		hr_clock_unlock(s);
1504 	}
1505 }
1506 
1507 /*
1508  * Handle clock tick processing for a thread.
1509  * Check for timer action, enforce CPU rlimit, do profiling etc.
1510  */
1511 void
1512 clock_tick(kthread_t *t, int pending)
1513 {
1514 	struct proc *pp;
1515 	klwp_id_t    lwp;
1516 	struct as *as;
1517 	clock_t	ticks;
1518 	int	poke = 0;		/* notify another CPU */
1519 	int	user_mode;
1520 	size_t	 rss;
1521 	int i, total_usec, usec;
1522 	rctl_qty_t secs;
1523 
1524 	ASSERT(pending > 0);
1525 
1526 	/* Must be operating on a lwp/thread */
1527 	if ((lwp = ttolwp(t)) == NULL) {
1528 		panic("clock_tick: no lwp");
1529 		/*NOTREACHED*/
1530 	}
1531 
1532 	for (i = 0; i < pending; i++) {
1533 		CL_TICK(t);	/* Class specific tick processing */
1534 		DTRACE_SCHED1(tick, kthread_t *, t);
1535 	}
1536 
1537 	pp = ttoproc(t);
1538 
1539 	/* pp->p_lock makes sure that the thread does not exit */
1540 	ASSERT(MUTEX_HELD(&pp->p_lock));
1541 
1542 	user_mode = (lwp->lwp_state == LWP_USER);
1543 
1544 	ticks = (pp->p_utime + pp->p_stime) % hz;
1545 	/*
1546 	 * Update process times. Should use high res clock and state
1547 	 * changes instead of statistical sampling method. XXX
1548 	 */
1549 	if (user_mode) {
1550 		pp->p_utime += pending;
1551 	} else {
1552 		pp->p_stime += pending;
1553 	}
1554 
1555 	pp->p_ttime += pending;
1556 	as = pp->p_as;
1557 
1558 	/*
1559 	 * Update user profiling statistics. Get the pc from the
1560 	 * lwp when the AST happens.
1561 	 */
1562 	if (pp->p_prof.pr_scale) {
1563 		atomic_add_32(&lwp->lwp_oweupc, (int32_t)pending);
1564 		if (user_mode) {
1565 			poke = 1;
1566 			aston(t);
1567 		}
1568 	}
1569 
1570 	/*
1571 	 * If CPU was in user state, process lwp-virtual time
1572 	 * interval timer. The value passed to itimerdecr() has to be
1573 	 * in microseconds and has to be less than one second. Hence
1574 	 * this loop.
1575 	 */
1576 	total_usec = usec_per_tick * pending;
1577 	while (total_usec > 0) {
1578 		usec = MIN(total_usec, (MICROSEC - 1));
1579 		if (user_mode &&
1580 		    timerisset(&lwp->lwp_timer[ITIMER_VIRTUAL].it_value) &&
1581 		    itimerdecr(&lwp->lwp_timer[ITIMER_VIRTUAL], usec) == 0) {
1582 			poke = 1;
1583 			sigtoproc(pp, t, SIGVTALRM);
1584 		}
1585 		total_usec -= usec;
1586 	}
1587 
1588 	/*
1589 	 * If CPU was in user state, process lwp-profile
1590 	 * interval timer.
1591 	 */
1592 	total_usec = usec_per_tick * pending;
1593 	while (total_usec > 0) {
1594 		usec = MIN(total_usec, (MICROSEC - 1));
1595 		if (timerisset(&lwp->lwp_timer[ITIMER_PROF].it_value) &&
1596 		    itimerdecr(&lwp->lwp_timer[ITIMER_PROF], usec) == 0) {
1597 			poke = 1;
1598 			sigtoproc(pp, t, SIGPROF);
1599 		}
1600 		total_usec -= usec;
1601 	}
1602 
1603 	/*
1604 	 * Enforce CPU resource controls:
1605 	 *   (a) process.max-cpu-time resource control
1606 	 *
1607 	 * Perform the check only if we have accumulated more a second.
1608 	 */
1609 	if ((ticks + pending) >= hz) {
1610 		(void) rctl_test(rctlproc_legacy[RLIMIT_CPU], pp->p_rctls, pp,
1611 		    (pp->p_utime + pp->p_stime)/hz, RCA_UNSAFE_SIGINFO);
1612 	}
1613 
1614 	/*
1615 	 *   (b) task.max-cpu-time resource control
1616 	 *
1617 	 * If we have accumulated enough ticks, increment the task CPU
1618 	 * time usage and test for the resource limit. This minimizes the
1619 	 * number of calls to the rct_test(). The task CPU time mutex
1620 	 * is highly contentious as many processes can be sharing a task.
1621 	 */
1622 	if (pp->p_ttime >= clock_tick_proc_max) {
1623 		secs = task_cpu_time_incr(pp->p_task, pp->p_ttime);
1624 		pp->p_ttime = 0;
1625 		if (secs) {
1626 			(void) rctl_test(rc_task_cpu_time, pp->p_task->tk_rctls,
1627 			    pp, secs, RCA_UNSAFE_SIGINFO);
1628 		}
1629 	}
1630 
1631 	/*
1632 	 * Update memory usage for the currently running process.
1633 	 */
1634 	rss = rm_asrss(as);
1635 	PTOU(pp)->u_mem += rss;
1636 	if (rss > PTOU(pp)->u_mem_max)
1637 		PTOU(pp)->u_mem_max = rss;
1638 
1639 	/*
1640 	 * Notify the CPU the thread is running on.
1641 	 */
1642 	if (poke && t->t_cpu != CPU)
1643 		poke_cpu(t->t_cpu->cpu_id);
1644 }
1645 
1646 void
1647 profil_tick(uintptr_t upc)
1648 {
1649 	int ticks;
1650 	proc_t *p = ttoproc(curthread);
1651 	klwp_t *lwp = ttolwp(curthread);
1652 	struct prof *pr = &p->p_prof;
1653 
1654 	do {
1655 		ticks = lwp->lwp_oweupc;
1656 	} while (atomic_cas_32(&lwp->lwp_oweupc, ticks, 0) != ticks);
1657 
1658 	mutex_enter(&p->p_pflock);
1659 	if (pr->pr_scale >= 2 && upc >= pr->pr_off) {
1660 		/*
1661 		 * Old-style profiling
1662 		 */
1663 		uint16_t *slot = pr->pr_base;
1664 		uint16_t old, new;
1665 		if (pr->pr_scale != 2) {
1666 			uintptr_t delta = upc - pr->pr_off;
1667 			uintptr_t byteoff = ((delta >> 16) * pr->pr_scale) +
1668 			    (((delta & 0xffff) * pr->pr_scale) >> 16);
1669 			if (byteoff >= (uintptr_t)pr->pr_size) {
1670 				mutex_exit(&p->p_pflock);
1671 				return;
1672 			}
1673 			slot += byteoff / sizeof (uint16_t);
1674 		}
1675 		if (fuword16(slot, &old) < 0 ||
1676 		    (new = old + ticks) > SHRT_MAX ||
1677 		    suword16(slot, new) < 0) {
1678 			pr->pr_scale = 0;
1679 		}
1680 	} else if (pr->pr_scale == 1) {
1681 		/*
1682 		 * PC Sampling
1683 		 */
1684 		model_t model = lwp_getdatamodel(lwp);
1685 		int result;
1686 #ifdef __lint
1687 		model = model;
1688 #endif
1689 		while (ticks-- > 0) {
1690 			if (pr->pr_samples == pr->pr_size) {
1691 				/* buffer full, turn off sampling */
1692 				pr->pr_scale = 0;
1693 				break;
1694 			}
1695 			switch (SIZEOF_PTR(model)) {
1696 			case sizeof (uint32_t):
1697 				result = suword32(pr->pr_base, (uint32_t)upc);
1698 				break;
1699 #ifdef _LP64
1700 			case sizeof (uint64_t):
1701 				result = suword64(pr->pr_base, (uint64_t)upc);
1702 				break;
1703 #endif
1704 			default:
1705 				cmn_err(CE_WARN, "profil_tick: unexpected "
1706 				    "data model");
1707 				result = -1;
1708 				break;
1709 			}
1710 			if (result != 0) {
1711 				pr->pr_scale = 0;
1712 				break;
1713 			}
1714 			pr->pr_base = (caddr_t)pr->pr_base + SIZEOF_PTR(model);
1715 			pr->pr_samples++;
1716 		}
1717 	}
1718 	mutex_exit(&p->p_pflock);
1719 }
1720 
1721 static void
1722 delay_wakeup(void *arg)
1723 {
1724 	kthread_t	*t = arg;
1725 
1726 	mutex_enter(&t->t_delay_lock);
1727 	cv_signal(&t->t_delay_cv);
1728 	mutex_exit(&t->t_delay_lock);
1729 }
1730 
1731 /*
1732  * The delay(9F) man page indicates that it can only be called from user or
1733  * kernel context - detect and diagnose bad calls. The following macro will
1734  * produce a limited number of messages identifying bad callers.  This is done
1735  * in a macro so that caller() is meaningful. When a bad caller is identified,
1736  * switching to 'drv_usecwait(TICK_TO_USEC(ticks));' may be appropriate.
1737  */
1738 #define	DELAY_CONTEXT_CHECK()	{					\
1739 	uint32_t	m;						\
1740 	char		*f;						\
1741 	ulong_t		off;						\
1742 									\
1743 	m = delay_from_interrupt_msg;					\
1744 	if (delay_from_interrupt_diagnose && servicing_interrupt() &&	\
1745 	    !panicstr && !devinfo_freeze &&				\
1746 	    atomic_cas_32(&delay_from_interrupt_msg, m ? m : 1, m-1)) {	\
1747 		f = modgetsymname((uintptr_t)caller(), &off);		\
1748 		cmn_err(CE_WARN, "delay(9F) called from "		\
1749 		    "interrupt context: %s`%s",				\
1750 		    mod_containing_pc(caller()), f ? f : "...");	\
1751 	}								\
1752 }
1753 
1754 /*
1755  * delay_common: common delay code.
1756  */
1757 static void
1758 delay_common(clock_t ticks)
1759 {
1760 	kthread_t	*t = curthread;
1761 	clock_t		deadline;
1762 	clock_t		timeleft;
1763 	callout_id_t	id;
1764 
1765 	/* If timeouts aren't running all we can do is spin. */
1766 	if (panicstr || devinfo_freeze) {
1767 		/* Convert delay(9F) call into drv_usecwait(9F) call. */
1768 		if (ticks > 0)
1769 			drv_usecwait(TICK_TO_USEC(ticks));
1770 		return;
1771 	}
1772 
1773 	deadline = ddi_get_lbolt() + ticks;
1774 	while ((timeleft = deadline - ddi_get_lbolt()) > 0) {
1775 		mutex_enter(&t->t_delay_lock);
1776 		id = timeout_default(delay_wakeup, t, timeleft);
1777 		cv_wait(&t->t_delay_cv, &t->t_delay_lock);
1778 		mutex_exit(&t->t_delay_lock);
1779 		(void) untimeout_default(id, 0);
1780 	}
1781 }
1782 
1783 /*
1784  * Delay specified number of clock ticks.
1785  */
1786 void
1787 delay(clock_t ticks)
1788 {
1789 	DELAY_CONTEXT_CHECK();
1790 
1791 	delay_common(ticks);
1792 }
1793 
1794 /*
1795  * Delay a random number of clock ticks between 1 and ticks.
1796  */
1797 void
1798 delay_random(clock_t ticks)
1799 {
1800 	int	r;
1801 
1802 	DELAY_CONTEXT_CHECK();
1803 
1804 	(void) random_get_pseudo_bytes((void *)&r, sizeof (r));
1805 	if (ticks == 0)
1806 		ticks = 1;
1807 	ticks = (r % ticks) + 1;
1808 	delay_common(ticks);
1809 }
1810 
1811 /*
1812  * Like delay, but interruptible by a signal.
1813  */
1814 int
1815 delay_sig(clock_t ticks)
1816 {
1817 	kthread_t	*t = curthread;
1818 	clock_t		deadline;
1819 	clock_t		rc;
1820 
1821 	/* If timeouts aren't running all we can do is spin. */
1822 	if (panicstr || devinfo_freeze) {
1823 		if (ticks > 0)
1824 			drv_usecwait(TICK_TO_USEC(ticks));
1825 		return (0);
1826 	}
1827 
1828 	deadline = ddi_get_lbolt() + ticks;
1829 	mutex_enter(&t->t_delay_lock);
1830 	do {
1831 		rc = cv_timedwait_sig(&t->t_delay_cv,
1832 		    &t->t_delay_lock, deadline);
1833 		/* loop until past deadline or signaled */
1834 	} while (rc > 0);
1835 	mutex_exit(&t->t_delay_lock);
1836 	if (rc == 0)
1837 		return (EINTR);
1838 	return (0);
1839 }
1840 
1841 
1842 #define	SECONDS_PER_DAY 86400
1843 
1844 /*
1845  * Initialize the system time based on the TOD chip.  approx is used as
1846  * an approximation of time (e.g. from the filesystem) in the event that
1847  * the TOD chip has been cleared or is unresponsive.  An approx of -1
1848  * means the filesystem doesn't keep time.
1849  */
1850 void
1851 clkset(time_t approx)
1852 {
1853 	timestruc_t ts;
1854 	int spl;
1855 	int set_clock = 0;
1856 
1857 	mutex_enter(&tod_lock);
1858 	ts = tod_get();
1859 
1860 	if (ts.tv_sec > 365 * SECONDS_PER_DAY) {
1861 		/*
1862 		 * If the TOD chip is reporting some time after 1971,
1863 		 * then it probably didn't lose power or become otherwise
1864 		 * cleared in the recent past;  check to assure that
1865 		 * the time coming from the filesystem isn't in the future
1866 		 * according to the TOD chip.
1867 		 */
1868 		if (approx != -1 && approx > ts.tv_sec) {
1869 			cmn_err(CE_WARN, "Last shutdown is later "
1870 			    "than time on time-of-day chip; check date.");
1871 		}
1872 	} else {
1873 		/*
1874 		 * If the TOD chip isn't giving correct time, set it to the
1875 		 * greater of i) approx and ii) 1987. That way if approx
1876 		 * is negative or is earlier than 1987, we set the clock
1877 		 * back to a time when Oliver North, ALF and Dire Straits
1878 		 * were all on the collective brain:  1987.
1879 		 */
1880 		timestruc_t tmp;
1881 		time_t diagnose_date = (1987 - 1970) * 365 * SECONDS_PER_DAY;
1882 		ts.tv_sec = (approx > diagnose_date ? approx : diagnose_date);
1883 		ts.tv_nsec = 0;
1884 
1885 		/*
1886 		 * Attempt to write the new time to the TOD chip.  Set spl high
1887 		 * to avoid getting preempted between the tod_set and tod_get.
1888 		 */
1889 		spl = splhi();
1890 		tod_set(ts);
1891 		tmp = tod_get();
1892 		splx(spl);
1893 
1894 		if (tmp.tv_sec != ts.tv_sec && tmp.tv_sec != ts.tv_sec + 1) {
1895 			tod_broken = 1;
1896 			dosynctodr = 0;
1897 			cmn_err(CE_WARN, "Time-of-day chip unresponsive.");
1898 		} else {
1899 			cmn_err(CE_WARN, "Time-of-day chip had "
1900 			    "incorrect date; check and reset.");
1901 		}
1902 		set_clock = 1;
1903 	}
1904 
1905 	if (!boot_time) {
1906 		boot_time = ts.tv_sec;
1907 		set_clock = 1;
1908 	}
1909 
1910 	if (set_clock)
1911 		set_hrestime(&ts);
1912 
1913 	mutex_exit(&tod_lock);
1914 }
1915 
1916 int	timechanged;	/* for testing if the system time has been reset */
1917 
1918 void
1919 set_hrestime(timestruc_t *ts)
1920 {
1921 	int spl = hr_clock_lock();
1922 	hrestime = *ts;
1923 	membar_enter();	/* hrestime must be visible before timechanged++ */
1924 	timedelta = 0;
1925 	timechanged++;
1926 	hr_clock_unlock(spl);
1927 	callout_hrestime();
1928 }
1929 
1930 static uint_t deadman_seconds;
1931 static uint32_t deadman_panics;
1932 static int deadman_enabled = 0;
1933 static int deadman_panic_timers = 1;
1934 
1935 static void
1936 deadman(void)
1937 {
1938 	if (panicstr) {
1939 		/*
1940 		 * During panic, other CPUs besides the panic
1941 		 * master continue to handle cyclics and some other
1942 		 * interrupts.  The code below is intended to be
1943 		 * single threaded, so any CPU other than the master
1944 		 * must keep out.
1945 		 */
1946 		if (CPU->cpu_id != panic_cpu.cpu_id)
1947 			return;
1948 
1949 		if (!deadman_panic_timers)
1950 			return; /* allow all timers to be manually disabled */
1951 
1952 		/*
1953 		 * If we are generating a crash dump or syncing filesystems and
1954 		 * the corresponding timer is set, decrement it and re-enter
1955 		 * the panic code to abort it and advance to the next state.
1956 		 * The panic states and triggers are explained in panic.c.
1957 		 */
1958 		if (panic_dump) {
1959 			if (dump_timeleft && (--dump_timeleft == 0)) {
1960 				panic("panic dump timeout");
1961 				/*NOTREACHED*/
1962 			}
1963 		}
1964 		return;
1965 	}
1966 
1967 	if (deadman_counter != CPU->cpu_deadman_counter) {
1968 		CPU->cpu_deadman_counter = deadman_counter;
1969 		CPU->cpu_deadman_countdown = deadman_seconds;
1970 		return;
1971 	}
1972 
1973 	if (--CPU->cpu_deadman_countdown > 0)
1974 		return;
1975 
1976 	/*
1977 	 * Regardless of whether or not we actually bring the system down,
1978 	 * bump the deadman_panics variable.
1979 	 *
1980 	 * N.B. deadman_panics is incremented once for each CPU that
1981 	 * passes through here.  It's expected that all the CPUs will
1982 	 * detect this condition within one second of each other, so
1983 	 * when deadman_enabled is off, deadman_panics will
1984 	 * typically be a multiple of the total number of CPUs in
1985 	 * the system.
1986 	 */
1987 	atomic_inc_32(&deadman_panics);
1988 
1989 	if (!deadman_enabled) {
1990 		CPU->cpu_deadman_countdown = deadman_seconds;
1991 		return;
1992 	}
1993 
1994 	/*
1995 	 * If we're here, we want to bring the system down.
1996 	 */
1997 	panic("deadman: timed out after %d seconds of clock "
1998 	    "inactivity", deadman_seconds);
1999 	/*NOTREACHED*/
2000 }
2001 
2002 /*ARGSUSED*/
2003 static void
2004 deadman_online(void *arg, cpu_t *cpu, cyc_handler_t *hdlr, cyc_time_t *when)
2005 {
2006 	cpu->cpu_deadman_counter = 0;
2007 	cpu->cpu_deadman_countdown = deadman_seconds;
2008 
2009 	hdlr->cyh_func = (cyc_func_t)deadman;
2010 	hdlr->cyh_level = CY_HIGH_LEVEL;
2011 	hdlr->cyh_arg = NULL;
2012 
2013 	/*
2014 	 * Stagger the CPUs so that they don't all run deadman() at
2015 	 * the same time.  Simplest reason to do this is to make it
2016 	 * more likely that only one CPU will panic in case of a
2017 	 * timeout.  This is (strictly speaking) an aesthetic, not a
2018 	 * technical consideration.
2019 	 */
2020 	when->cyt_when = cpu->cpu_id * (NANOSEC / NCPU);
2021 	when->cyt_interval = NANOSEC;
2022 }
2023 
2024 
2025 void
2026 deadman_init(void)
2027 {
2028 	cyc_omni_handler_t hdlr;
2029 
2030 	if (deadman_seconds == 0)
2031 		deadman_seconds = snoop_interval / MICROSEC;
2032 
2033 	if (snooping)
2034 		deadman_enabled = 1;
2035 
2036 	hdlr.cyo_online = deadman_online;
2037 	hdlr.cyo_offline = NULL;
2038 	hdlr.cyo_arg = NULL;
2039 
2040 	mutex_enter(&cpu_lock);
2041 	deadman_cyclic = cyclic_add_omni(&hdlr);
2042 	mutex_exit(&cpu_lock);
2043 }
2044 
2045 /*
2046  * tod_fault() is for updating tod validate mechanism state:
2047  * (1) TOD_NOFAULT: for resetting the state to 'normal'.
2048  *     currently used for debugging only
2049  * (2) The following four cases detected by tod validate mechanism:
2050  *       TOD_REVERSED: current tod value is less than previous value.
2051  *       TOD_STALLED: current tod value hasn't advanced.
2052  *       TOD_JUMPED: current tod value advanced too far from previous value.
2053  *       TOD_RATECHANGED: the ratio between average tod delta and
2054  *       average tick delta has changed.
2055  * (3) TOD_RDONLY: when the TOD clock is not writeable e.g. because it is
2056  *     a virtual TOD provided by a hypervisor.
2057  */
2058 enum tod_fault_type
2059 tod_fault(enum tod_fault_type ftype, int off)
2060 {
2061 	ASSERT(MUTEX_HELD(&tod_lock));
2062 
2063 	if (tod_faulted != ftype) {
2064 		switch (ftype) {
2065 		case TOD_NOFAULT:
2066 			plat_tod_fault(TOD_NOFAULT);
2067 			cmn_err(CE_NOTE, "Restarted tracking "
2068 			    "Time of Day clock.");
2069 			tod_faulted = ftype;
2070 			break;
2071 		case TOD_REVERSED:
2072 		case TOD_JUMPED:
2073 			if (tod_faulted == TOD_NOFAULT) {
2074 				plat_tod_fault(ftype);
2075 				cmn_err(CE_WARN, "Time of Day clock error: "
2076 				    "reason [%s by 0x%x]. -- "
2077 				    " Stopped tracking Time Of Day clock.",
2078 				    tod_fault_table[ftype], off);
2079 				tod_faulted = ftype;
2080 			}
2081 			break;
2082 		case TOD_STALLED:
2083 		case TOD_RATECHANGED:
2084 			if (tod_faulted == TOD_NOFAULT) {
2085 				plat_tod_fault(ftype);
2086 				cmn_err(CE_WARN, "Time of Day clock error: "
2087 				    "reason [%s]. -- "
2088 				    " Stopped tracking Time Of Day clock.",
2089 				    tod_fault_table[ftype]);
2090 				tod_faulted = ftype;
2091 			}
2092 			break;
2093 		case TOD_RDONLY:
2094 			if (tod_faulted == TOD_NOFAULT) {
2095 				plat_tod_fault(ftype);
2096 				cmn_err(CE_NOTE, "!Time of Day clock is "
2097 				    "Read-Only; set of Date/Time will not "
2098 				    "persist across reboot.");
2099 				tod_faulted = ftype;
2100 			}
2101 			break;
2102 		default:
2103 			break;
2104 		}
2105 	}
2106 	return (tod_faulted);
2107 }
2108 
2109 /*
2110  * Two functions that allow tod_status_flag to be manipulated by functions
2111  * external to this file.
2112  */
2113 
2114 void
2115 tod_status_set(int tod_flag)
2116 {
2117 	tod_status_flag |= tod_flag;
2118 }
2119 
2120 void
2121 tod_status_clear(int tod_flag)
2122 {
2123 	tod_status_flag &= ~tod_flag;
2124 }
2125 
2126 /*
2127  * Record a timestamp and the value passed to tod_set().  The next call to
2128  * tod_validate() can use these values, prev_set_tick and prev_set_tod,
2129  * when checking the timestruc_t returned by tod_get().  Ordinarily,
2130  * tod_validate() will use prev_tick and prev_tod for this task but these
2131  * become obsolete, and will be re-assigned with the prev_set_* values,
2132  * in the case when the TOD is re-written.
2133  */
2134 void
2135 tod_set_prev(timestruc_t ts)
2136 {
2137 	if ((tod_validate_enable == 0) || (tod_faulted != TOD_NOFAULT) ||
2138 	    tod_validate_deferred) {
2139 		return;
2140 	}
2141 	prev_set_tick = gethrtime();
2142 	/*
2143 	 * A negative value will be set to zero in utc_to_tod() so we fake
2144 	 * a zero here in such a case.  This would need to change if the
2145 	 * behavior of utc_to_tod() changes.
2146 	 */
2147 	prev_set_tod = ts.tv_sec < 0 ? 0 : ts.tv_sec;
2148 }
2149 
2150 /*
2151  * tod_validate() is used for checking values returned by tod_get().
2152  * Four error cases can be detected by this routine:
2153  *   TOD_REVERSED: current tod value is less than previous.
2154  *   TOD_STALLED: current tod value hasn't advanced.
2155  *   TOD_JUMPED: current tod value advanced too far from previous value.
2156  *   TOD_RATECHANGED: the ratio between average tod delta and
2157  *   average tick delta has changed.
2158  */
2159 time_t
2160 tod_validate(time_t tod)
2161 {
2162 	time_t diff_tod;
2163 	hrtime_t diff_tick;
2164 
2165 	long dtick;
2166 	int dtick_delta;
2167 
2168 	int off = 0;
2169 	enum tod_fault_type tod_bad = TOD_NOFAULT;
2170 
2171 	static int firsttime = 1;
2172 
2173 	static time_t prev_tod = 0;
2174 	static hrtime_t prev_tick = 0;
2175 	static long dtick_avg = TOD_REF_FREQ;
2176 
2177 	int cpr_resume_done = 0;
2178 	int dr_resume_done = 0;
2179 
2180 	hrtime_t tick = gethrtime();
2181 
2182 	ASSERT(MUTEX_HELD(&tod_lock));
2183 
2184 	/*
2185 	 * tod_validate_enable is patchable via /etc/system.
2186 	 * If TOD is already faulted, or if TOD validation is deferred,
2187 	 * there is nothing to do.
2188 	 */
2189 	if ((tod_validate_enable == 0) || (tod_faulted != TOD_NOFAULT) ||
2190 	    tod_validate_deferred) {
2191 		return (tod);
2192 	}
2193 
2194 	/*
2195 	 * If this is the first time through, we just need to save the tod
2196 	 * we were called with and hrtime so we can use them next time to
2197 	 * validate tod_get().
2198 	 */
2199 	if (firsttime) {
2200 		firsttime = 0;
2201 		prev_tod = tod;
2202 		prev_tick = tick;
2203 		return (tod);
2204 	}
2205 
2206 	/*
2207 	 * Handle any flags that have been turned on by tod_status_set().
2208 	 * In the case where a tod_set() is done and then a subsequent
2209 	 * tod_get() fails (ie, both TOD_SET_DONE and TOD_GET_FAILED are
2210 	 * true), we treat the TOD_GET_FAILED with precedence by switching
2211 	 * off the flag, returning tod and leaving TOD_SET_DONE asserted
2212 	 * until such time as tod_get() completes successfully.
2213 	 */
2214 	if (tod_status_flag & TOD_GET_FAILED) {
2215 		/*
2216 		 * tod_get() has encountered an issue, possibly transitory,
2217 		 * when reading TOD.  We'll just return the incoming tod
2218 		 * value (which is actually hrestime.tv_sec in this case)
2219 		 * and when we get a genuine tod, following a successful
2220 		 * tod_get(), we can validate using prev_tod and prev_tick.
2221 		 */
2222 		tod_status_flag &= ~TOD_GET_FAILED;
2223 		return (tod);
2224 	} else if (tod_status_flag & TOD_SET_DONE) {
2225 		/*
2226 		 * TOD has been modified.  Just before the TOD was written,
2227 		 * tod_set_prev() saved tod and hrtime; we can now use
2228 		 * those values, prev_set_tod and prev_set_tick, to validate
2229 		 * the incoming tod that's just been read.
2230 		 */
2231 		prev_tod = prev_set_tod;
2232 		prev_tick = prev_set_tick;
2233 		dtick_avg = TOD_REF_FREQ;
2234 		tod_status_flag &= ~TOD_SET_DONE;
2235 		/*
2236 		 * If a tod_set() preceded a cpr_suspend() without an
2237 		 * intervening tod_validate(), we need to ensure that a
2238 		 * TOD_JUMPED condition is ignored.
2239 		 * Note this isn't a concern in the case of DR as we've
2240 		 * just reassigned dtick_avg, above.
2241 		 */
2242 		if (tod_status_flag & TOD_CPR_RESUME_DONE) {
2243 			cpr_resume_done = 1;
2244 			tod_status_flag &= ~TOD_CPR_RESUME_DONE;
2245 		}
2246 	} else if (tod_status_flag & TOD_CPR_RESUME_DONE) {
2247 		/*
2248 		 * The system's coming back from a checkpoint resume.
2249 		 */
2250 		cpr_resume_done = 1;
2251 		tod_status_flag &= ~TOD_CPR_RESUME_DONE;
2252 		/*
2253 		 * We need to handle the possibility of a CPR suspend
2254 		 * operation having been initiated whilst a DR event was
2255 		 * in-flight.
2256 		 */
2257 		if (tod_status_flag & TOD_DR_RESUME_DONE) {
2258 			dr_resume_done = 1;
2259 			tod_status_flag &= ~TOD_DR_RESUME_DONE;
2260 		}
2261 	} else if (tod_status_flag & TOD_DR_RESUME_DONE) {
2262 		/*
2263 		 * A Dynamic Reconfiguration event has taken place.
2264 		 */
2265 		dr_resume_done = 1;
2266 		tod_status_flag &= ~TOD_DR_RESUME_DONE;
2267 	}
2268 
2269 	/* test hook */
2270 	switch (tod_unit_test) {
2271 	case 1: /* for testing jumping tod */
2272 		tod += tod_test_injector;
2273 		tod_unit_test = 0;
2274 		break;
2275 	case 2:	/* for testing stuck tod bit */
2276 		tod |= 1 << tod_test_injector;
2277 		tod_unit_test = 0;
2278 		break;
2279 	case 3:	/* for testing stalled tod */
2280 		tod = prev_tod;
2281 		tod_unit_test = 0;
2282 		break;
2283 	case 4:	/* reset tod fault status */
2284 		(void) tod_fault(TOD_NOFAULT, 0);
2285 		tod_unit_test = 0;
2286 		break;
2287 	default:
2288 		break;
2289 	}
2290 
2291 	diff_tod = tod - prev_tod;
2292 	diff_tick = tick - prev_tick;
2293 
2294 	ASSERT(diff_tick >= 0);
2295 
2296 	if (diff_tod < 0) {
2297 		/* ERROR - tod reversed */
2298 		tod_bad = TOD_REVERSED;
2299 		off = (int)(prev_tod - tod);
2300 	} else if (diff_tod == 0) {
2301 		/* tod did not advance */
2302 		if (diff_tick > TOD_STALL_THRESHOLD) {
2303 			/* ERROR - tod stalled */
2304 			tod_bad = TOD_STALLED;
2305 		} else {
2306 			/*
2307 			 * Make sure we don't update prev_tick
2308 			 * so that diff_tick is calculated since
2309 			 * the first diff_tod == 0
2310 			 */
2311 			return (tod);
2312 		}
2313 	} else {
2314 		/* calculate dtick */
2315 		dtick = diff_tick / diff_tod;
2316 
2317 		/* update dtick averages */
2318 		dtick_avg += ((dtick - dtick_avg) / TOD_FILTER_N);
2319 
2320 		/*
2321 		 * Calculate dtick_delta as
2322 		 * variation from reference freq in quartiles
2323 		 */
2324 		dtick_delta = (dtick_avg - TOD_REF_FREQ) /
2325 		    (TOD_REF_FREQ >> 2);
2326 
2327 		/*
2328 		 * Even with a perfectly functioning TOD device,
2329 		 * when the number of elapsed seconds is low the
2330 		 * algorithm can calculate a rate that is beyond
2331 		 * tolerance, causing an error.  The algorithm is
2332 		 * inaccurate when elapsed time is low (less than
2333 		 * 5 seconds).
2334 		 */
2335 		if (diff_tod > 4) {
2336 			if (dtick < TOD_JUMP_THRESHOLD) {
2337 				/*
2338 				 * If we've just done a CPR resume, we detect
2339 				 * a jump in the TOD but, actually, what's
2340 				 * happened is that the TOD has been increasing
2341 				 * whilst the system was suspended and the tick
2342 				 * count hasn't kept up.  We consider the first
2343 				 * occurrence of this after a resume as normal
2344 				 * and ignore it; otherwise, in a non-resume
2345 				 * case, we regard it as a TOD problem.
2346 				 */
2347 				if (!cpr_resume_done) {
2348 					/* ERROR - tod jumped */
2349 					tod_bad = TOD_JUMPED;
2350 					off = (int)diff_tod;
2351 				}
2352 			}
2353 			if (dtick_delta) {
2354 				/*
2355 				 * If we've just done a DR resume, dtick_avg
2356 				 * can go a bit askew so we reset it and carry
2357 				 * on; otherwise, the TOD is in error.
2358 				 */
2359 				if (dr_resume_done) {
2360 					dtick_avg = TOD_REF_FREQ;
2361 				} else {
2362 					/* ERROR - change in clock rate */
2363 					tod_bad = TOD_RATECHANGED;
2364 				}
2365 			}
2366 		}
2367 	}
2368 
2369 	if (tod_bad != TOD_NOFAULT) {
2370 		(void) tod_fault(tod_bad, off);
2371 
2372 		/*
2373 		 * Disable dosynctodr since we are going to fault
2374 		 * the TOD chip anyway here
2375 		 */
2376 		dosynctodr = 0;
2377 
2378 		/*
2379 		 * Set tod to the correct value from hrestime
2380 		 */
2381 		tod = hrestime.tv_sec;
2382 	}
2383 
2384 	prev_tod = tod;
2385 	prev_tick = tick;
2386 	return (tod);
2387 }
2388 
2389 static void
2390 calcloadavg(int nrun, uint64_t *hp_ave)
2391 {
2392 	static int64_t f[3] = { 135, 27, 9 };
2393 	uint_t i;
2394 	int64_t q, r;
2395 
2396 	/*
2397 	 * Compute load average over the last 1, 5, and 15 minutes
2398 	 * (60, 300, and 900 seconds).  The constants in f[3] are for
2399 	 * exponential decay:
2400 	 * (1 - exp(-1/60)) << 13 = 135,
2401 	 * (1 - exp(-1/300)) << 13 = 27,
2402 	 * (1 - exp(-1/900)) << 13 = 9.
2403 	 */
2404 
2405 	/*
2406 	 * a little hoop-jumping to avoid integer overflow
2407 	 */
2408 	for (i = 0; i < 3; i++) {
2409 		q = (hp_ave[i]  >> 16) << 7;
2410 		r = (hp_ave[i]  & 0xffff) << 7;
2411 		hp_ave[i] += ((nrun - q) * f[i] - ((r * f[i]) >> 16)) >> 4;
2412 	}
2413 }
2414 
2415 /*
2416  * lbolt_hybrid() is used by ddi_get_lbolt() and ddi_get_lbolt64() to
2417  * calculate the value of lbolt according to the current mode. In the event
2418  * driven mode (the default), lbolt is calculated by dividing the current hires
2419  * time by the number of nanoseconds per clock tick. In the cyclic driven mode
2420  * an internal variable is incremented at each firing of the lbolt cyclic
2421  * and returned by lbolt_cyclic_driven().
2422  *
2423  * The system will transition from event to cyclic driven mode when the number
2424  * of calls to lbolt_event_driven() exceeds the (per CPU) threshold within a
2425  * window of time. It does so by reprograming lbolt_cyclic from CY_INFINITY to
2426  * nsec_per_tick. The lbolt cyclic will remain ON while at least one CPU is
2427  * causing enough activity to cross the thresholds.
2428  */
2429 int64_t
2430 lbolt_bootstrap(void)
2431 {
2432 	return (0);
2433 }
2434 
2435 /* ARGSUSED */
2436 uint_t
2437 lbolt_ev_to_cyclic(caddr_t arg1, caddr_t arg2)
2438 {
2439 	hrtime_t ts, exp;
2440 	int ret;
2441 
2442 	ASSERT(lbolt_hybrid != lbolt_cyclic_driven);
2443 
2444 	kpreempt_disable();
2445 
2446 	ts = gethrtime();
2447 	lb_info->lbi_internal = (ts/nsec_per_tick);
2448 
2449 	/*
2450 	 * Align the next expiration to a clock tick boundary.
2451 	 */
2452 	exp = ts + nsec_per_tick - 1;
2453 	exp = (exp/nsec_per_tick) * nsec_per_tick;
2454 
2455 	ret = cyclic_reprogram(lb_info->id.lbi_cyclic_id, exp);
2456 	ASSERT(ret);
2457 
2458 	lbolt_hybrid = lbolt_cyclic_driven;
2459 	lb_info->lbi_cyc_deactivate = B_FALSE;
2460 	lb_info->lbi_cyc_deac_start = lb_info->lbi_internal;
2461 
2462 	kpreempt_enable();
2463 
2464 	ret = atomic_dec_32_nv(&lb_info->lbi_token);
2465 	ASSERT(ret == 0);
2466 
2467 	return (1);
2468 }
2469 
2470 int64_t
2471 lbolt_event_driven(void)
2472 {
2473 	hrtime_t ts;
2474 	int64_t lb;
2475 	int ret, cpu = CPU->cpu_seqid;
2476 
2477 	ts = gethrtime();
2478 	ASSERT(ts > 0);
2479 
2480 	ASSERT(nsec_per_tick > 0);
2481 	lb = (ts/nsec_per_tick);
2482 
2483 	/*
2484 	 * Switch to cyclic mode if the number of calls to this routine
2485 	 * has reached the threshold within the interval.
2486 	 */
2487 	if ((lb - lb_cpu[cpu].lbc_cnt_start) < lb_info->lbi_thresh_interval) {
2488 
2489 		if (--lb_cpu[cpu].lbc_counter == 0) {
2490 			/*
2491 			 * Reached the threshold within the interval, reset
2492 			 * the usage statistics.
2493 			 */
2494 			lb_cpu[cpu].lbc_counter = lb_info->lbi_thresh_calls;
2495 			lb_cpu[cpu].lbc_cnt_start = lb;
2496 
2497 			/*
2498 			 * Make sure only one thread reprograms the
2499 			 * lbolt cyclic and changes the mode.
2500 			 */
2501 			if (panicstr == NULL &&
2502 			    atomic_cas_32(&lb_info->lbi_token, 0, 1) == 0) {
2503 
2504 				if (lbolt_hybrid == lbolt_cyclic_driven) {
2505 					ret = atomic_dec_32_nv(
2506 					    &lb_info->lbi_token);
2507 					ASSERT(ret == 0);
2508 				} else {
2509 					lbolt_softint_post();
2510 				}
2511 			}
2512 		}
2513 	} else {
2514 		/*
2515 		 * Exceeded the interval, reset the usage statistics.
2516 		 */
2517 		lb_cpu[cpu].lbc_counter = lb_info->lbi_thresh_calls;
2518 		lb_cpu[cpu].lbc_cnt_start = lb;
2519 	}
2520 
2521 	ASSERT(lb >= lb_info->lbi_debug_time);
2522 
2523 	return (lb - lb_info->lbi_debug_time);
2524 }
2525 
2526 int64_t
2527 lbolt_cyclic_driven(void)
2528 {
2529 	int64_t lb = lb_info->lbi_internal;
2530 	int cpu;
2531 
2532 	/*
2533 	 * If a CPU has already prevented the lbolt cyclic from deactivating
2534 	 * itself, don't bother tracking the usage. Otherwise check if we're
2535 	 * within the interval and how the per CPU counter is doing.
2536 	 */
2537 	if (lb_info->lbi_cyc_deactivate) {
2538 		cpu = CPU->cpu_seqid;
2539 		if ((lb - lb_cpu[cpu].lbc_cnt_start) <
2540 		    lb_info->lbi_thresh_interval) {
2541 
2542 			if (lb_cpu[cpu].lbc_counter == 0)
2543 				/*
2544 				 * Reached the threshold within the interval,
2545 				 * prevent the lbolt cyclic from turning itself
2546 				 * off.
2547 				 */
2548 				lb_info->lbi_cyc_deactivate = B_FALSE;
2549 			else
2550 				lb_cpu[cpu].lbc_counter--;
2551 		} else {
2552 			/*
2553 			 * Only reset the usage statistics when we have
2554 			 * exceeded the interval.
2555 			 */
2556 			lb_cpu[cpu].lbc_counter = lb_info->lbi_thresh_calls;
2557 			lb_cpu[cpu].lbc_cnt_start = lb;
2558 		}
2559 	}
2560 
2561 	ASSERT(lb >= lb_info->lbi_debug_time);
2562 
2563 	return (lb - lb_info->lbi_debug_time);
2564 }
2565 
2566 /*
2567  * The lbolt_cyclic() routine will fire at a nsec_per_tick interval to satisfy
2568  * performance needs of ddi_get_lbolt() and ddi_get_lbolt64() consumers.
2569  * It is inactive by default, and will be activated when switching from event
2570  * to cyclic driven lbolt. The cyclic will turn itself off unless signaled
2571  * by lbolt_cyclic_driven().
2572  */
2573 static void
2574 lbolt_cyclic(void)
2575 {
2576 	int ret;
2577 
2578 	lb_info->lbi_internal++;
2579 
2580 	if (!lbolt_cyc_only) {
2581 
2582 		if (lb_info->lbi_cyc_deactivate) {
2583 			/*
2584 			 * Switching from cyclic to event driven mode.
2585 			 */
2586 			if (panicstr == NULL &&
2587 			    atomic_cas_32(&lb_info->lbi_token, 0, 1) == 0) {
2588 
2589 				if (lbolt_hybrid == lbolt_event_driven) {
2590 					ret = atomic_dec_32_nv(
2591 					    &lb_info->lbi_token);
2592 					ASSERT(ret == 0);
2593 					return;
2594 				}
2595 
2596 				kpreempt_disable();
2597 
2598 				lbolt_hybrid = lbolt_event_driven;
2599 				ret = cyclic_reprogram(
2600 				    lb_info->id.lbi_cyclic_id,
2601 				    CY_INFINITY);
2602 				ASSERT(ret);
2603 
2604 				kpreempt_enable();
2605 
2606 				ret = atomic_dec_32_nv(&lb_info->lbi_token);
2607 				ASSERT(ret == 0);
2608 			}
2609 		}
2610 
2611 		/*
2612 		 * The lbolt cyclic should not try to deactivate itself before
2613 		 * the sampling period has elapsed.
2614 		 */
2615 		if (lb_info->lbi_internal - lb_info->lbi_cyc_deac_start >=
2616 		    lb_info->lbi_thresh_interval) {
2617 			lb_info->lbi_cyc_deactivate = B_TRUE;
2618 			lb_info->lbi_cyc_deac_start = lb_info->lbi_internal;
2619 		}
2620 	}
2621 }
2622 
2623 /*
2624  * Since the lbolt service was historically cyclic driven, it must be 'stopped'
2625  * when the system drops into the kernel debugger. lbolt_debug_entry() is
2626  * called by the KDI system claim callbacks to record a hires timestamp at
2627  * debug enter time. lbolt_debug_return() is called by the sistem release
2628  * callbacks to account for the time spent in the debugger. The value is then
2629  * accumulated in the lb_info structure and used by lbolt_event_driven() and
2630  * lbolt_cyclic_driven(), as well as the mdb_get_lbolt() routine.
2631  */
2632 void
2633 lbolt_debug_entry(void)
2634 {
2635 	if (lbolt_hybrid != lbolt_bootstrap) {
2636 		ASSERT(lb_info != NULL);
2637 		lb_info->lbi_debug_ts = gethrtime();
2638 	}
2639 }
2640 
2641 /*
2642  * Calculate the time spent in the debugger and add it to the lbolt info
2643  * structure. We also update the internal lbolt value in case we were in
2644  * cyclic driven mode going in.
2645  */
2646 void
2647 lbolt_debug_return(void)
2648 {
2649 	hrtime_t ts;
2650 
2651 	if (lbolt_hybrid != lbolt_bootstrap) {
2652 		ASSERT(lb_info != NULL);
2653 		ASSERT(nsec_per_tick > 0);
2654 
2655 		ts = gethrtime();
2656 		lb_info->lbi_internal = (ts/nsec_per_tick);
2657 		lb_info->lbi_debug_time +=
2658 		    ((ts - lb_info->lbi_debug_ts)/nsec_per_tick);
2659 
2660 		lb_info->lbi_debug_ts = 0;
2661 	}
2662 }
2663