xref: /titanic_41/usr/src/uts/common/os/clock.c (revision 5aefb6555731130ca4fd295960123d71f2d21fe8)
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, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
23 /*	  All Rights Reserved	*/
24 
25 
26 /*
27  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
28  * Use is subject to license terms.
29  */
30 
31 #pragma ident	"%Z%%M%	%I%	%E% SMI"
32 
33 #include <sys/param.h>
34 #include <sys/t_lock.h>
35 #include <sys/types.h>
36 #include <sys/tuneable.h>
37 #include <sys/sysmacros.h>
38 #include <sys/systm.h>
39 #include <sys/cpuvar.h>
40 #include <sys/lgrp.h>
41 #include <sys/user.h>
42 #include <sys/proc.h>
43 #include <sys/callo.h>
44 #include <sys/kmem.h>
45 #include <sys/var.h>
46 #include <sys/cmn_err.h>
47 #include <sys/swap.h>
48 #include <sys/vmsystm.h>
49 #include <sys/class.h>
50 #include <sys/time.h>
51 #include <sys/debug.h>
52 #include <sys/vtrace.h>
53 #include <sys/spl.h>
54 #include <sys/atomic.h>
55 #include <sys/dumphdr.h>
56 #include <sys/archsystm.h>
57 #include <sys/fs/swapnode.h>
58 #include <sys/panic.h>
59 #include <sys/disp.h>
60 #include <sys/msacct.h>
61 #include <sys/mem_cage.h>
62 
63 #include <vm/page.h>
64 #include <vm/anon.h>
65 #include <vm/rm.h>
66 #include <sys/cyclic.h>
67 #include <sys/cpupart.h>
68 #include <sys/rctl.h>
69 #include <sys/task.h>
70 #include <sys/chip.h>
71 #include <sys/sdt.h>
72 
73 /*
74  * for NTP support
75  */
76 #include <sys/timex.h>
77 #include <sys/inttypes.h>
78 
79 /*
80  * clock is called straight from
81  * the real time clock interrupt.
82  *
83  * Functions:
84  *	reprime clock
85  *	schedule callouts
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 
95 /*
96  * high-precision avenrun values.  These are needed to make the
97  * regular avenrun values accurate.
98  */
99 static uint64_t hp_avenrun[3];
100 int	avenrun[3];		/* FSCALED average run queue lengths */
101 time_t	time;	/* time in seconds since 1970 - for compatibility only */
102 
103 static struct loadavg_s loadavg;
104 /*
105  * Phase/frequency-lock loop (PLL/FLL) definitions
106  *
107  * The following variables are read and set by the ntp_adjtime() system
108  * call.
109  *
110  * time_state shows the state of the system clock, with values defined
111  * in the timex.h header file.
112  *
113  * time_status shows the status of the system clock, with bits defined
114  * in the timex.h header file.
115  *
116  * time_offset is used by the PLL/FLL to adjust the system time in small
117  * increments.
118  *
119  * time_constant determines the bandwidth or "stiffness" of the PLL.
120  *
121  * time_tolerance determines maximum frequency error or tolerance of the
122  * CPU clock oscillator and is a property of the architecture; however,
123  * in principle it could change as result of the presence of external
124  * discipline signals, for instance.
125  *
126  * time_precision is usually equal to the kernel tick variable; however,
127  * in cases where a precision clock counter or external clock is
128  * available, the resolution can be much less than this and depend on
129  * whether the external clock is working or not.
130  *
131  * time_maxerror is initialized by a ntp_adjtime() call and increased by
132  * the kernel once each second to reflect the maximum error bound
133  * growth.
134  *
135  * time_esterror is set and read by the ntp_adjtime() call, but
136  * otherwise not used by the kernel.
137  */
138 int32_t time_state = TIME_OK;	/* clock state */
139 int32_t time_status = STA_UNSYNC;	/* clock status bits */
140 int32_t time_offset = 0;		/* time offset (us) */
141 int32_t time_constant = 0;		/* pll time constant */
142 int32_t time_tolerance = MAXFREQ;	/* frequency tolerance (scaled ppm) */
143 int32_t time_precision = 1;	/* clock precision (us) */
144 int32_t time_maxerror = MAXPHASE;	/* maximum error (us) */
145 int32_t time_esterror = MAXPHASE;	/* estimated error (us) */
146 
147 /*
148  * The following variables establish the state of the PLL/FLL and the
149  * residual time and frequency offset of the local clock. The scale
150  * factors are defined in the timex.h header file.
151  *
152  * time_phase and time_freq are the phase increment and the frequency
153  * increment, respectively, of the kernel time variable.
154  *
155  * time_freq is set via ntp_adjtime() from a value stored in a file when
156  * the synchronization daemon is first started. Its value is retrieved
157  * via ntp_adjtime() and written to the file about once per hour by the
158  * daemon.
159  *
160  * time_adj is the adjustment added to the value of tick at each timer
161  * interrupt and is recomputed from time_phase and time_freq at each
162  * seconds rollover.
163  *
164  * time_reftime is the second's portion of the system time at the last
165  * call to ntp_adjtime(). It is used to adjust the time_freq variable
166  * and to increase the time_maxerror as the time since last update
167  * increases.
168  */
169 int32_t time_phase = 0;		/* phase offset (scaled us) */
170 int32_t time_freq = 0;		/* frequency offset (scaled ppm) */
171 int32_t time_adj = 0;		/* tick adjust (scaled 1 / hz) */
172 int32_t time_reftime = 0;		/* time at last adjustment (s) */
173 
174 /*
175  * The scale factors of the following variables are defined in the
176  * timex.h header file.
177  *
178  * pps_time contains the time at each calibration interval, as read by
179  * microtime(). pps_count counts the seconds of the calibration
180  * interval, the duration of which is nominally pps_shift in powers of
181  * two.
182  *
183  * pps_offset is the time offset produced by the time median filter
184  * pps_tf[], while pps_jitter is the dispersion (jitter) measured by
185  * this filter.
186  *
187  * pps_freq is the frequency offset produced by the frequency median
188  * filter pps_ff[], while pps_stabil is the dispersion (wander) measured
189  * by this filter.
190  *
191  * pps_usec is latched from a high resolution counter or external clock
192  * at pps_time. Here we want the hardware counter contents only, not the
193  * contents plus the time_tv.usec as usual.
194  *
195  * pps_valid counts the number of seconds since the last PPS update. It
196  * is used as a watchdog timer to disable the PPS discipline should the
197  * PPS signal be lost.
198  *
199  * pps_glitch counts the number of seconds since the beginning of an
200  * offset burst more than tick/2 from current nominal offset. It is used
201  * mainly to suppress error bursts due to priority conflicts between the
202  * PPS interrupt and timer interrupt.
203  *
204  * pps_intcnt counts the calibration intervals for use in the interval-
205  * adaptation algorithm. It's just too complicated for words.
206  */
207 struct timeval pps_time;	/* kernel time at last interval */
208 int32_t pps_tf[] = {0, 0, 0};	/* pps time offset median filter (us) */
209 int32_t pps_offset = 0;		/* pps time offset (us) */
210 int32_t pps_jitter = MAXTIME;	/* time dispersion (jitter) (us) */
211 int32_t pps_ff[] = {0, 0, 0};	/* pps frequency offset median filter */
212 int32_t pps_freq = 0;		/* frequency offset (scaled ppm) */
213 int32_t pps_stabil = MAXFREQ;	/* frequency dispersion (scaled ppm) */
214 int32_t pps_usec = 0;		/* microsec counter at last interval */
215 int32_t pps_valid = PPS_VALID;	/* pps signal watchdog counter */
216 int32_t pps_glitch = 0;		/* pps signal glitch counter */
217 int32_t pps_count = 0;		/* calibration interval counter (s) */
218 int32_t pps_shift = PPS_SHIFT;	/* interval duration (s) (shift) */
219 int32_t pps_intcnt = 0;		/* intervals at current duration */
220 
221 /*
222  * PPS signal quality monitors
223  *
224  * pps_jitcnt counts the seconds that have been discarded because the
225  * jitter measured by the time median filter exceeds the limit MAXTIME
226  * (100 us).
227  *
228  * pps_calcnt counts the frequency calibration intervals, which are
229  * variable from 4 s to 256 s.
230  *
231  * pps_errcnt counts the calibration intervals which have been discarded
232  * because the wander exceeds the limit MAXFREQ (100 ppm) or where the
233  * calibration interval jitter exceeds two ticks.
234  *
235  * pps_stbcnt counts the calibration intervals that have been discarded
236  * because the frequency wander exceeds the limit MAXFREQ / 4 (25 us).
237  */
238 int32_t pps_jitcnt = 0;		/* jitter limit exceeded */
239 int32_t pps_calcnt = 0;		/* calibration intervals */
240 int32_t pps_errcnt = 0;		/* calibration errors */
241 int32_t pps_stbcnt = 0;		/* stability limit exceeded */
242 
243 /* The following variables require no explicit locking */
244 volatile clock_t lbolt;		/* time in Hz since last boot */
245 volatile int64_t lbolt64;	/* lbolt64 won't wrap for 2.9 billion yrs */
246 
247 kcondvar_t lbolt_cv;
248 int one_sec = 1; /* turned on once every second */
249 static int fsflushcnt;	/* counter for t_fsflushr */
250 int	dosynctodr = 1;	/* patchable; enable/disable sync to TOD chip */
251 int	tod_needsync = 0;	/* need to sync tod chip with software time */
252 static int tod_broken = 0;	/* clock chip doesn't work */
253 time_t	boot_time = 0;		/* Boot time in seconds since 1970 */
254 cyclic_id_t clock_cyclic;	/* clock()'s cyclic_id */
255 cyclic_id_t deadman_cyclic;	/* deadman()'s cyclic_id */
256 
257 static int lgrp_ticks;		/* counter to schedule lgrp load calcs */
258 
259 /*
260  * rechoose_interval_history is used to detect when rechoose_interval's
261  * value has changed (via hotpatching for example), so that the
262  * cached values in the cpu structures may be updated.
263  */
264 static int rechoose_interval_history = RECHOOSE_INTERVAL;
265 
266 /*
267  * for tod fault detection
268  */
269 #define	TOD_REF_FREQ		((longlong_t)(NANOSEC))
270 #define	TOD_STALL_THRESHOLD	(TOD_REF_FREQ * 3 / 2)
271 #define	TOD_JUMP_THRESHOLD	(TOD_REF_FREQ / 2)
272 #define	TOD_FILTER_N		4
273 #define	TOD_FILTER_SETTLE	(4 * TOD_FILTER_N)
274 static int tod_faulted = TOD_NOFAULT;
275 static int tod_fault_reset_flag = 0;
276 
277 /* patchable via /etc/system */
278 int tod_validate_enable = 1;
279 
280 /*
281  * On non-SPARC systems, TOD validation must be deferred until gethrtime
282  * returns non-zero values (after mach_clkinit's execution).
283  * On SPARC systems, it must be deferred until after hrtime_base
284  * and hres_last_tick are set (in the first invocation of hres_tick).
285  * Since in both cases the prerequisites occur before the invocation of
286  * tod_get() in clock(), the deferment is lifted there.
287  */
288 static boolean_t tod_validate_deferred = B_TRUE;
289 
290 /*
291  * tod_fault_table[] must be aligned with
292  * enum tod_fault_type in systm.h
293  */
294 static char *tod_fault_table[] = {
295 	"Reversed",			/* TOD_REVERSED */
296 	"Stalled",			/* TOD_STALLED */
297 	"Jumped",			/* TOD_JUMPED */
298 	"Changed in Clock Rate"		/* TOD_RATECHANGED */
299 	/*
300 	 * no strings needed for TOD_NOFAULT
301 	 */
302 };
303 
304 /*
305  * test hook for tod broken detection in tod_validate
306  */
307 int tod_unit_test = 0;
308 time_t tod_test_injector;
309 
310 #define	CLOCK_ADJ_HIST_SIZE	4
311 
312 static int	adj_hist_entry;
313 
314 int64_t clock_adj_hist[CLOCK_ADJ_HIST_SIZE];
315 
316 static void clock_tick(kthread_t *);
317 static void calcloadavg(int, uint64_t *);
318 static int genloadavg(struct loadavg_s *);
319 static void loadavg_update();
320 
321 void (*cmm_clock_callout)() = NULL;
322 
323 #ifdef	KSLICE
324 int kslice = KSLICE;
325 #endif
326 
327 static void
328 clock(void)
329 {
330 	kthread_t	*t;
331 	kmutex_t	*plockp;	/* pointer to thread's process lock */
332 	int	pinned_intr = 0;
333 	uint_t	nrunnable, nrunning;
334 	uint_t	w_io;
335 	cpu_t	*cp;
336 	cpupart_t *cpupart;
337 	int	exiting;
338 	extern void set_anoninfo();
339 	extern	void	set_freemem();
340 	void	(*funcp)();
341 	int32_t ltemp;
342 	int64_t lltemp;
343 	int s;
344 	int do_lgrp_load;
345 	int rechoose_update = 0;
346 	int rechoose;
347 	int i;
348 
349 	if (panicstr)
350 		return;
351 
352 	set_anoninfo();
353 	/*
354 	 * Make sure that 'freemem' do not drift too far from the truth
355 	 */
356 	set_freemem();
357 
358 
359 	/*
360 	 * Before the section which is repeated is executed, we do
361 	 * the time delta processing which occurs every clock tick
362 	 *
363 	 * There is additional processing which happens every time
364 	 * the nanosecond counter rolls over which is described
365 	 * below - see the section which begins with : if (one_sec)
366 	 *
367 	 * This section marks the beginning of the precision-kernel
368 	 * code fragment.
369 	 *
370 	 * First, compute the phase adjustment. If the low-order bits
371 	 * (time_phase) of the update overflow, bump the higher order
372 	 * bits (time_update).
373 	 */
374 	time_phase += time_adj;
375 	if (time_phase <= -FINEUSEC) {
376 		ltemp = -time_phase / SCALE_PHASE;
377 		time_phase += ltemp * SCALE_PHASE;
378 		s = hr_clock_lock();
379 		timedelta -= ltemp * (NANOSEC/MICROSEC);
380 		hr_clock_unlock(s);
381 	} else if (time_phase >= FINEUSEC) {
382 		ltemp = time_phase / SCALE_PHASE;
383 		time_phase -= ltemp * SCALE_PHASE;
384 		s = hr_clock_lock();
385 		timedelta += ltemp * (NANOSEC/MICROSEC);
386 		hr_clock_unlock(s);
387 	}
388 
389 	/*
390 	 * End of precision-kernel code fragment which is processed
391 	 * every timer interrupt.
392 	 *
393 	 * Continue with the interrupt processing as scheduled.
394 	 *
395 	 * Did we pin another interrupt thread?  Need to check this before
396 	 * grabbing any adaptive locks, since if we block on a lock the
397 	 * pinned thread could escape.  Note that this is just a heuristic;
398 	 * if we take multiple laps though clock() without returning from
399 	 * the interrupt because we have another clock tick pending, then
400 	 * the pinned interrupt could be released by one of the previous
401 	 * laps.  The only consequence is that the CPU will be counted as
402 	 * in idle (or wait) state once the pinned interrupt is released.
403 	 * Since this accounting is inaccurate by nature, this isn't a big
404 	 * deal --- but we should try to get it right in the common case
405 	 * where we only call clock() once per interrupt.
406 	 */
407 	if (curthread->t_intr != NULL)
408 		pinned_intr = (curthread->t_intr->t_flag & T_INTR_THREAD);
409 
410 	/*
411 	 * Count the number of runnable threads and the number waiting
412 	 * for some form of I/O to complete -- gets added to
413 	 * sysinfo.waiting.  To know the state of the system, must add
414 	 * wait counts from all CPUs.  Also add up the per-partition
415 	 * statistics.
416 	 */
417 	w_io = 0;
418 	nrunnable = 0;
419 
420 	/*
421 	 * keep track of when to update lgrp/part loads
422 	 */
423 
424 	do_lgrp_load = 0;
425 	if (lgrp_ticks++ >= hz / 10) {
426 		lgrp_ticks = 0;
427 		do_lgrp_load = 1;
428 	}
429 
430 	/*
431 	 * The dispatcher tunable rechoose_interval may be hot-patched.
432 	 * Note if it has a new value. If so, the effective rechoose_interval
433 	 * cached in the cpu structures needs to be updated.
434 	 * If needed we'll do this during the walk of the cpu_list below.
435 	 */
436 	if (rechoose_interval != rechoose_interval_history) {
437 		rechoose_interval_history = rechoose_interval;
438 		rechoose_update = 1;
439 	}
440 
441 	if (one_sec)
442 		loadavg_update();
443 
444 
445 	/*
446 	 * First count the threads waiting on kpreempt queues in each
447 	 * CPU partition.
448 	 */
449 
450 	cpupart = cp_list_head;
451 	do {
452 		uint_t cpupart_nrunnable = cpupart->cp_kp_queue.disp_nrunnable;
453 
454 		cpupart->cp_updates++;
455 		nrunnable += cpupart_nrunnable;
456 		cpupart->cp_nrunnable_cum += cpupart_nrunnable;
457 		if (one_sec) {
458 			cpupart->cp_nrunning = 0;
459 			cpupart->cp_nrunnable = cpupart_nrunnable;
460 		}
461 	} while ((cpupart = cpupart->cp_next) != cp_list_head);
462 
463 
464 	/* Now count the per-CPU statistics. */
465 	cp = cpu_list;
466 	do {
467 		uint_t cpu_nrunnable = cp->cpu_disp->disp_nrunnable;
468 
469 		nrunnable += cpu_nrunnable;
470 		cpupart = cp->cpu_part;
471 		cpupart->cp_nrunnable_cum += cpu_nrunnable;
472 		if (one_sec)
473 			cpupart->cp_nrunnable += cpu_nrunnable;
474 		if (do_lgrp_load &&
475 		    (cp->cpu_flags & CPU_EXISTS)) {
476 			/*
477 			 * When updating the lgroup's load average,
478 			 * account for the thread running on the CPU.
479 			 * If the CPU is the current one, then we need
480 			 * to account for the underlying thread which
481 			 * got the clock interrupt not the thread that is
482 			 * handling the interrupt and caculating the load
483 			 * average
484 			 */
485 			t = cp->cpu_thread;
486 			if (CPU == cp)
487 				t = t->t_intr;
488 
489 			/*
490 			 * Account for the load average for this thread if
491 			 * it isn't the idle thread or it is on the interrupt
492 			 * stack and not the current CPU handling the clock
493 			 * interrupt
494 			 */
495 			if ((t && t != cp->cpu_idle_thread) || (CPU != cp &&
496 			    CPU_ON_INTR(cp))) {
497 				if (t->t_lpl == cp->cpu_lpl) {
498 					/* local thread */
499 					cpu_nrunnable++;
500 				} else {
501 					/*
502 					 * This is a remote thread, charge it
503 					 * against its home lgroup.  Note that
504 					 * we notice that a thread is remote
505 					 * only if it's currently executing.
506 					 * This is a reasonable approximation,
507 					 * since queued remote threads are rare.
508 					 * Note also that if we didn't charge
509 					 * it to its home lgroup, remote
510 					 * execution would often make a system
511 					 * appear balanced even though it was
512 					 * not, and thread placement/migration
513 					 * would often not be done correctly.
514 					 */
515 					lgrp_loadavg(t->t_lpl,
516 					    LGRP_LOADAVG_IN_THREAD_MAX, 0);
517 				}
518 			}
519 			lgrp_loadavg(cp->cpu_lpl,
520 			    cpu_nrunnable * LGRP_LOADAVG_IN_THREAD_MAX, 1);
521 		}
522 		/*
523 		 * The platform may define a per physical processor
524 		 * adjustment of rechoose_interval. The effective
525 		 * (base + adjustment) rechoose_interval is cached
526 		 * in the cpu structures for efficiency. Above we detect
527 		 * if the cached values need updating, and here is where
528 		 * the update happens.
529 		 */
530 		if (rechoose_update) {
531 			rechoose = rechoose_interval +
532 				cp->cpu_chip->chip_rechoose_adj;
533 			cp->cpu_rechoose = (rechoose < 0) ? 0 : rechoose;
534 		}
535 	} while ((cp = cp->cpu_next) != cpu_list);
536 
537 	/*
538 	 * Do tick processing for all the active threads running in
539 	 * the system.
540 	 */
541 	cp = cpu_list;
542 	nrunning = 0;
543 	do {
544 		klwp_id_t lwp;
545 		int intr;
546 		int thread_away;
547 
548 		/*
549 		 * Don't do any tick processing on CPUs that
550 		 * aren't even in the system or aren't up yet.
551 		 */
552 		if ((cp->cpu_flags & CPU_EXISTS) == 0) {
553 			continue;
554 		}
555 
556 		/*
557 		 * The locking here is rather tricky.  We use
558 		 * thread_free_lock to keep the currently running
559 		 * thread from being freed or recycled while we're
560 		 * looking at it.  We can then check if the thread
561 		 * is exiting and get the appropriate p_lock if it
562 		 * is not.  We have to be careful, though, because
563 		 * the _process_ can still be freed while we're
564 		 * holding thread_free_lock.  To avoid touching the
565 		 * proc structure we put a pointer to the p_lock in the
566 		 * thread structure.  The p_lock is persistent so we
567 		 * can acquire it even if the process is gone.  At that
568 		 * point we can check (again) if the thread is exiting
569 		 * and either drop the lock or do the tick processing.
570 		 */
571 		mutex_enter(&thread_free_lock);
572 		/*
573 		 * We cannot hold the cpu_lock to prevent the
574 		 * cpu_list from changing in the clock interrupt.
575 		 * As long as we don't block (or don't get pre-empted)
576 		 * the cpu_list will not change (all threads are paused
577 		 * before list modification). If the list does change
578 		 * any deleted cpu structures will remain with cpu_next
579 		 * set to NULL, hence the following test.
580 		 */
581 		if (cp->cpu_next == NULL) {
582 			mutex_exit(&thread_free_lock);
583 			break;
584 		}
585 		t = cp->cpu_thread;	/* Current running thread */
586 		if (CPU == cp) {
587 			/*
588 			 * 't' will be the clock interrupt thread on this
589 			 * CPU.  Use the pinned thread (if any) on this CPU
590 			 * as the target of the clock tick.  If we pinned
591 			 * an interrupt, though, just keep using the clock
592 			 * interrupt thread since the formerly pinned one
593 			 * may have gone away.  One interrupt thread is as
594 			 * good as another, and this means we don't have
595 			 * to continue to check pinned_intr in subsequent
596 			 * code.
597 			 */
598 			ASSERT(t == curthread);
599 			if (t->t_intr != NULL && !pinned_intr)
600 				t = t->t_intr;
601 		}
602 
603 		intr = t->t_flag & T_INTR_THREAD;
604 		lwp = ttolwp(t);
605 		if (lwp == NULL || (t->t_proc_flag & TP_LWPEXIT) || intr) {
606 			/*
607 			 * Thread is exiting (or uninteresting) so don't
608 			 * do tick processing or grab p_lock.  Once we
609 			 * drop thread_free_lock we can't look inside the
610 			 * thread or lwp structure, since the thread may
611 			 * have gone away.
612 			 */
613 			exiting = 1;
614 		} else {
615 			/*
616 			 * OK, try to grab the process lock.  See
617 			 * comments above for why we're not using
618 			 * ttoproc(t)->p_lockp here.
619 			 */
620 			plockp = t->t_plockp;
621 			mutex_enter(plockp);
622 			/* See above comment. */
623 			if (cp->cpu_next == NULL) {
624 				mutex_exit(plockp);
625 				mutex_exit(&thread_free_lock);
626 				break;
627 			}
628 			/*
629 			 * The thread may have exited between when we
630 			 * checked above, and when we got the p_lock.
631 			 */
632 			if (t->t_proc_flag & TP_LWPEXIT) {
633 				mutex_exit(plockp);
634 				exiting = 1;
635 			} else {
636 				exiting = 0;
637 			}
638 		}
639 		/*
640 		 * Either we have the p_lock for the thread's process,
641 		 * or we don't care about the thread structure any more.
642 		 * Either way we can drop thread_free_lock.
643 		 */
644 		mutex_exit(&thread_free_lock);
645 
646 		/*
647 		 * Update user, system, and idle cpu times.
648 		 */
649 		if (one_sec) {
650 			nrunning++;
651 			cp->cpu_part->cp_nrunning++;
652 		}
653 		/*
654 		 * If we haven't done tick processing for this
655 		 * lwp, then do it now. Since we don't hold the
656 		 * lwp down on a CPU it can migrate and show up
657 		 * more than once, hence the lbolt check.
658 		 *
659 		 * Also, make sure that it's okay to perform the
660 		 * tick processing before calling clock_tick.
661 		 * Setting thread_away to a TRUE value (ie. not 0)
662 		 * results in tick processing not being performed for
663 		 * that thread.  Or, in other words, keeps the thread
664 		 * away from clock_tick processing.
665 		 */
666 		thread_away = ((cp->cpu_flags & CPU_QUIESCED) ||
667 		    CPU_ON_INTR(cp) || intr ||
668 		    (cp->cpu_dispthread == cp->cpu_idle_thread) || exiting);
669 
670 		if ((!thread_away) && (lbolt - t->t_lbolt != 0)) {
671 			t->t_lbolt = lbolt;
672 			clock_tick(t);
673 		}
674 
675 #ifdef KSLICE
676 		/*
677 		 * Ah what the heck, give this kid a taste of the real
678 		 * world and yank the rug out from under it.
679 		 * But, only if we are running UniProcessor.
680 		 */
681 		if ((kslice) && (ncpus == 1)) {
682 			aston(t);
683 			cp->cpu_runrun = 1;
684 			cp->cpu_kprunrun = 1;
685 		}
686 #endif
687 		if (!exiting)
688 			mutex_exit(plockp);
689 	} while ((cp = cp->cpu_next) != cpu_list);
690 
691 	/*
692 	 * bump time in ticks
693 	 *
694 	 * We rely on there being only one clock thread and hence
695 	 * don't need a lock to protect lbolt.
696 	 */
697 	lbolt++;
698 	atomic_add_64((uint64_t *)&lbolt64, (int64_t)1);
699 
700 	/*
701 	 * Check for a callout that needs be called from the clock
702 	 * thread to support the membership protocol in a clustered
703 	 * system.  Copy the function pointer so that we can reset
704 	 * this to NULL if needed.
705 	 */
706 	if ((funcp = cmm_clock_callout) != NULL)
707 		(*funcp)();
708 
709 	/*
710 	 * Wakeup the cageout thread waiters once per second.
711 	 */
712 	if (one_sec)
713 		kcage_tick();
714 
715 	/*
716 	 * Schedule timeout() requests if any are due at this time.
717 	 */
718 	callout_schedule();
719 
720 	if (one_sec) {
721 
722 		int drift, absdrift;
723 		timestruc_t tod;
724 		int s;
725 
726 		/*
727 		 * Beginning of precision-kernel code fragment executed
728 		 * every second.
729 		 *
730 		 * On rollover of the second the phase adjustment to be
731 		 * used for the next second is calculated.  Also, the
732 		 * maximum error is increased by the tolerance.  If the
733 		 * PPS frequency discipline code is present, the phase is
734 		 * increased to compensate for the CPU clock oscillator
735 		 * frequency error.
736 		 *
737 		 * On a 32-bit machine and given parameters in the timex.h
738 		 * header file, the maximum phase adjustment is +-512 ms
739 		 * and maximum frequency offset is (a tad less than)
740 		 * +-512 ppm. On a 64-bit machine, you shouldn't need to ask.
741 		 */
742 		time_maxerror += time_tolerance / SCALE_USEC;
743 
744 		/*
745 		 * Leap second processing. If in leap-insert state at
746 		 * the end of the day, the system clock is set back one
747 		 * second; if in leap-delete state, the system clock is
748 		 * set ahead one second. The microtime() routine or
749 		 * external clock driver will insure that reported time
750 		 * is always monotonic. The ugly divides should be
751 		 * replaced.
752 		 */
753 		switch (time_state) {
754 
755 		case TIME_OK:
756 			if (time_status & STA_INS)
757 				time_state = TIME_INS;
758 			else if (time_status & STA_DEL)
759 				time_state = TIME_DEL;
760 			break;
761 
762 		case TIME_INS:
763 			if (hrestime.tv_sec % 86400 == 0) {
764 				s = hr_clock_lock();
765 				hrestime.tv_sec--;
766 				hr_clock_unlock(s);
767 				time_state = TIME_OOP;
768 			}
769 			break;
770 
771 		case TIME_DEL:
772 			if ((hrestime.tv_sec + 1) % 86400 == 0) {
773 				s = hr_clock_lock();
774 				hrestime.tv_sec++;
775 				hr_clock_unlock(s);
776 				time_state = TIME_WAIT;
777 			}
778 			break;
779 
780 		case TIME_OOP:
781 			time_state = TIME_WAIT;
782 			break;
783 
784 		case TIME_WAIT:
785 			if (!(time_status & (STA_INS | STA_DEL)))
786 				time_state = TIME_OK;
787 		default:
788 			break;
789 		}
790 
791 		/*
792 		 * Compute the phase adjustment for the next second. In
793 		 * PLL mode, the offset is reduced by a fixed factor
794 		 * times the time constant. In FLL mode the offset is
795 		 * used directly. In either mode, the maximum phase
796 		 * adjustment for each second is clamped so as to spread
797 		 * the adjustment over not more than the number of
798 		 * seconds between updates.
799 		 */
800 		if (time_offset == 0)
801 			time_adj = 0;
802 		else if (time_offset < 0) {
803 			lltemp = -time_offset;
804 			if (!(time_status & STA_FLL)) {
805 				if ((1 << time_constant) >= SCALE_KG)
806 					lltemp *= (1 << time_constant) /
807 					    SCALE_KG;
808 				else
809 					lltemp = (lltemp / SCALE_KG) >>
810 					    time_constant;
811 			}
812 			if (lltemp > (MAXPHASE / MINSEC) * SCALE_UPDATE)
813 				lltemp = (MAXPHASE / MINSEC) * SCALE_UPDATE;
814 			time_offset += lltemp;
815 			time_adj = -(lltemp * SCALE_PHASE) / hz / SCALE_UPDATE;
816 		} else {
817 			lltemp = time_offset;
818 			if (!(time_status & STA_FLL)) {
819 				if ((1 << time_constant) >= SCALE_KG)
820 					lltemp *= (1 << time_constant) /
821 					    SCALE_KG;
822 				else
823 					lltemp = (lltemp / SCALE_KG) >>
824 					    time_constant;
825 			}
826 			if (lltemp > (MAXPHASE / MINSEC) * SCALE_UPDATE)
827 				lltemp = (MAXPHASE / MINSEC) * SCALE_UPDATE;
828 			time_offset -= lltemp;
829 			time_adj = (lltemp * SCALE_PHASE) / hz / SCALE_UPDATE;
830 		}
831 
832 		/*
833 		 * Compute the frequency estimate and additional phase
834 		 * adjustment due to frequency error for the next
835 		 * second. When the PPS signal is engaged, gnaw on the
836 		 * watchdog counter and update the frequency computed by
837 		 * the pll and the PPS signal.
838 		 */
839 		pps_valid++;
840 		if (pps_valid == PPS_VALID) {
841 			pps_jitter = MAXTIME;
842 			pps_stabil = MAXFREQ;
843 			time_status &= ~(STA_PPSSIGNAL | STA_PPSJITTER |
844 			    STA_PPSWANDER | STA_PPSERROR);
845 		}
846 		lltemp = time_freq + pps_freq;
847 
848 		if (lltemp)
849 			time_adj += (lltemp * SCALE_PHASE) / (SCALE_USEC * hz);
850 
851 		/*
852 		 * End of precision kernel-code fragment
853 		 *
854 		 * The section below should be modified if we are planning
855 		 * to use NTP for synchronization.
856 		 *
857 		 * Note: the clock synchronization code now assumes
858 		 * the following:
859 		 *   - if dosynctodr is 1, then compute the drift between
860 		 *	the tod chip and software time and adjust one or
861 		 *	the other depending on the circumstances
862 		 *
863 		 *   - if dosynctodr is 0, then the tod chip is independent
864 		 *	of the software clock and should not be adjusted,
865 		 *	but allowed to free run.  this allows NTP to sync.
866 		 *	hrestime without any interference from the tod chip.
867 		 */
868 
869 		tod_validate_deferred = B_FALSE;
870 		mutex_enter(&tod_lock);
871 		tod = tod_get();
872 		drift = tod.tv_sec - hrestime.tv_sec;
873 		absdrift = (drift >= 0) ? drift : -drift;
874 		if (tod_needsync || absdrift > 1) {
875 			int s;
876 			if (absdrift > 2) {
877 				if (!tod_broken && tod_faulted == TOD_NOFAULT) {
878 					s = hr_clock_lock();
879 					hrestime = tod;
880 					membar_enter();	/* hrestime visible */
881 					timedelta = 0;
882 					timechanged++;
883 					tod_needsync = 0;
884 					hr_clock_unlock(s);
885 				}
886 			} else {
887 				if (tod_needsync || !dosynctodr) {
888 					gethrestime(&tod);
889 					tod_set(tod);
890 					s = hr_clock_lock();
891 					if (timedelta == 0)
892 						tod_needsync = 0;
893 					hr_clock_unlock(s);
894 				} else {
895 					/*
896 					 * If the drift is 2 seconds on the
897 					 * money, then the TOD is adjusting
898 					 * the clock;  record that.
899 					 */
900 					clock_adj_hist[adj_hist_entry++ %
901 					    CLOCK_ADJ_HIST_SIZE] = lbolt64;
902 					s = hr_clock_lock();
903 					timedelta = (int64_t)drift*NANOSEC;
904 					hr_clock_unlock(s);
905 				}
906 			}
907 		}
908 		one_sec = 0;
909 		time = gethrestime_sec();  /* for crusty old kmem readers */
910 		mutex_exit(&tod_lock);
911 
912 		/*
913 		 * Some drivers still depend on this... XXX
914 		 */
915 		cv_broadcast(&lbolt_cv);
916 
917 		sysinfo.updates++;
918 		vminfo.freemem += freemem;
919 		{
920 			pgcnt_t maxswap, resv, free;
921 			pgcnt_t avail =
922 			    MAX((spgcnt_t)(availrmem - swapfs_minfree), 0);
923 
924 			maxswap = k_anoninfo.ani_mem_resv
925 					+ k_anoninfo.ani_max +avail;
926 			free = k_anoninfo.ani_free + avail;
927 			resv = k_anoninfo.ani_phys_resv +
928 			    k_anoninfo.ani_mem_resv;
929 
930 			vminfo.swap_resv += resv;
931 			/* number of reserved and allocated pages */
932 #ifdef	DEBUG
933 			if (maxswap < free)
934 				cmn_err(CE_WARN, "clock: maxswap < free");
935 			if (maxswap < resv)
936 				cmn_err(CE_WARN, "clock: maxswap < resv");
937 #endif
938 			vminfo.swap_alloc += maxswap - free;
939 			vminfo.swap_avail += maxswap - resv;
940 			vminfo.swap_free += free;
941 		}
942 		if (nrunnable) {
943 			sysinfo.runque += nrunnable;
944 			sysinfo.runocc++;
945 		}
946 		if (nswapped) {
947 			sysinfo.swpque += nswapped;
948 			sysinfo.swpocc++;
949 		}
950 		sysinfo.waiting += w_io;
951 
952 		/*
953 		 * Wake up fsflush to write out DELWRI
954 		 * buffers, dirty pages and other cached
955 		 * administrative data, e.g. inodes.
956 		 */
957 		if (--fsflushcnt <= 0) {
958 			fsflushcnt = tune.t_fsflushr;
959 			cv_signal(&fsflush_cv);
960 		}
961 
962 		vmmeter();
963 		calcloadavg(genloadavg(&loadavg), hp_avenrun);
964 		for (i = 0; i < 3; i++)
965 			/*
966 			 * At the moment avenrun[] can only hold 31
967 			 * bits of load average as it is a signed
968 			 * int in the API. We need to ensure that
969 			 * hp_avenrun[i] >> (16 - FSHIFT) will not be
970 			 * too large. If it is, we put the largest value
971 			 * that we can use into avenrun[i]. This is
972 			 * kludgey, but about all we can do until we
973 			 * avenrun[] is declared as an array of uint64[]
974 			 */
975 			if (hp_avenrun[i] < ((uint64_t)1<<(31+16-FSHIFT)))
976 				avenrun[i] = (int32_t)(hp_avenrun[i] >>
977 				    (16 - FSHIFT));
978 			else
979 				avenrun[i] = 0x7fffffff;
980 
981 		cpupart = cp_list_head;
982 		do {
983 			calcloadavg(genloadavg(&cpupart->cp_loadavg),
984 			    cpupart->cp_hp_avenrun);
985 		} while ((cpupart = cpupart->cp_next) != cp_list_head);
986 
987 		/*
988 		 * Wake up the swapper thread if necessary.
989 		 */
990 		if (runin ||
991 		    (runout && (avefree < desfree || wake_sched_sec))) {
992 			t = &t0;
993 			thread_lock(t);
994 			if (t->t_state == TS_STOPPED) {
995 				runin = runout = 0;
996 				wake_sched_sec = 0;
997 				t->t_whystop = 0;
998 				t->t_whatstop = 0;
999 				t->t_schedflag &= ~TS_ALLSTART;
1000 				THREAD_TRANSITION(t);
1001 				setfrontdq(t);
1002 			}
1003 			thread_unlock(t);
1004 		}
1005 	}
1006 
1007 	/*
1008 	 * Wake up the swapper if any high priority swapped-out threads
1009 	 * became runable during the last tick.
1010 	 */
1011 	if (wake_sched) {
1012 		t = &t0;
1013 		thread_lock(t);
1014 		if (t->t_state == TS_STOPPED) {
1015 			runin = runout = 0;
1016 			wake_sched = 0;
1017 			t->t_whystop = 0;
1018 			t->t_whatstop = 0;
1019 			t->t_schedflag &= ~TS_ALLSTART;
1020 			THREAD_TRANSITION(t);
1021 			setfrontdq(t);
1022 		}
1023 		thread_unlock(t);
1024 	}
1025 }
1026 
1027 void
1028 clock_init(void)
1029 {
1030 	cyc_handler_t hdlr;
1031 	cyc_time_t when;
1032 
1033 	hdlr.cyh_func = (cyc_func_t)clock;
1034 	hdlr.cyh_level = CY_LOCK_LEVEL;
1035 	hdlr.cyh_arg = NULL;
1036 
1037 	when.cyt_when = 0;
1038 	when.cyt_interval = nsec_per_tick;
1039 
1040 	mutex_enter(&cpu_lock);
1041 	clock_cyclic = cyclic_add(&hdlr, &when);
1042 	mutex_exit(&cpu_lock);
1043 }
1044 
1045 /*
1046  * Called before calcloadavg to get 10-sec moving loadavg together
1047  */
1048 
1049 static int
1050 genloadavg(struct loadavg_s *avgs)
1051 {
1052 	int avg;
1053 	int spos; /* starting position */
1054 	int cpos; /* moving current position */
1055 	int i;
1056 	int slen;
1057 	hrtime_t hr_avg;
1058 
1059 	/* 10-second snapshot, calculate first positon */
1060 	if (avgs->lg_len == 0) {
1061 		return (0);
1062 	}
1063 	slen = avgs->lg_len < S_MOVAVG_SZ ? avgs->lg_len : S_MOVAVG_SZ;
1064 
1065 	spos = (avgs->lg_cur - 1) >= 0 ? avgs->lg_cur - 1 :
1066 	    S_LOADAVG_SZ + (avgs->lg_cur - 1);
1067 	for (i = hr_avg = 0; i < slen; i++) {
1068 		cpos = (spos - i) >= 0 ? spos - i : S_LOADAVG_SZ + (spos - i);
1069 		hr_avg += avgs->lg_loads[cpos];
1070 	}
1071 
1072 	hr_avg = hr_avg / slen;
1073 	avg = hr_avg / (NANOSEC / LGRP_LOADAVG_IN_THREAD_MAX);
1074 
1075 	return (avg);
1076 }
1077 
1078 /*
1079  * Run every second from clock () to update the loadavg count available to the
1080  * system and cpu-partitions.
1081  *
1082  * This works by sampling the previous usr, sys, wait time elapsed,
1083  * computing a delta, and adding that delta to the elapsed usr, sys,
1084  * wait increase.
1085  */
1086 
1087 static void
1088 loadavg_update()
1089 {
1090 	cpu_t *cp;
1091 	cpupart_t *cpupart;
1092 	hrtime_t cpu_total;
1093 	int prev;
1094 
1095 	cp = cpu_list;
1096 	loadavg.lg_total = 0;
1097 
1098 	/*
1099 	 * first pass totals up per-cpu statistics for system and cpu
1100 	 * partitions
1101 	 */
1102 
1103 	do {
1104 		struct loadavg_s *lavg;
1105 
1106 		lavg = &cp->cpu_loadavg;
1107 
1108 		cpu_total = cp->cpu_acct[CMS_USER] +
1109 		    cp->cpu_acct[CMS_SYSTEM] + cp->cpu_waitrq;
1110 		/* compute delta against last total */
1111 		scalehrtime(&cpu_total);
1112 		prev = (lavg->lg_cur - 1) >= 0 ? lavg->lg_cur - 1 :
1113 		    S_LOADAVG_SZ + (lavg->lg_cur - 1);
1114 		if (lavg->lg_loads[prev] <= 0) {
1115 			lavg->lg_loads[lavg->lg_cur] = cpu_total;
1116 			cpu_total = 0;
1117 		} else {
1118 			lavg->lg_loads[lavg->lg_cur] = cpu_total;
1119 			cpu_total = cpu_total - lavg->lg_loads[prev];
1120 			if (cpu_total < 0)
1121 				cpu_total = 0;
1122 		}
1123 
1124 		lavg->lg_cur = (lavg->lg_cur + 1) % S_LOADAVG_SZ;
1125 		lavg->lg_len = (lavg->lg_len + 1) < S_LOADAVG_SZ ?
1126 		    lavg->lg_len + 1 : S_LOADAVG_SZ;
1127 
1128 		loadavg.lg_total += cpu_total;
1129 		cp->cpu_part->cp_loadavg.lg_total += cpu_total;
1130 
1131 	} while ((cp = cp->cpu_next) != cpu_list);
1132 
1133 	loadavg.lg_loads[loadavg.lg_cur] = loadavg.lg_total;
1134 	loadavg.lg_cur = (loadavg.lg_cur + 1) % S_LOADAVG_SZ;
1135 	loadavg.lg_len = (loadavg.lg_len + 1) < S_LOADAVG_SZ ?
1136 	    loadavg.lg_len + 1 : S_LOADAVG_SZ;
1137 	/*
1138 	 * Second pass updates counts
1139 	 */
1140 	cpupart = cp_list_head;
1141 
1142 	do {
1143 		struct loadavg_s *lavg;
1144 
1145 		lavg = &cpupart->cp_loadavg;
1146 		lavg->lg_loads[lavg->lg_cur] = lavg->lg_total;
1147 		lavg->lg_total = 0;
1148 		lavg->lg_cur = (lavg->lg_cur + 1) % S_LOADAVG_SZ;
1149 		lavg->lg_len = (lavg->lg_len + 1) < S_LOADAVG_SZ ?
1150 		    lavg->lg_len + 1 : S_LOADAVG_SZ;
1151 
1152 	} while ((cpupart = cpupart->cp_next) != cp_list_head);
1153 
1154 }
1155 
1156 /*
1157  * clock_update() - local clock update
1158  *
1159  * This routine is called by ntp_adjtime() to update the local clock
1160  * phase and frequency. The implementation is of an
1161  * adaptive-parameter, hybrid phase/frequency-lock loop (PLL/FLL). The
1162  * routine computes new time and frequency offset estimates for each
1163  * call.  The PPS signal itself determines the new time offset,
1164  * instead of the calling argument.  Presumably, calls to
1165  * ntp_adjtime() occur only when the caller believes the local clock
1166  * is valid within some bound (+-128 ms with NTP). If the caller's
1167  * time is far different than the PPS time, an argument will ensue,
1168  * and it's not clear who will lose.
1169  *
1170  * For uncompensated quartz crystal oscillatores and nominal update
1171  * intervals less than 1024 s, operation should be in phase-lock mode
1172  * (STA_FLL = 0), where the loop is disciplined to phase. For update
1173  * intervals greater than this, operation should be in frequency-lock
1174  * mode (STA_FLL = 1), where the loop is disciplined to frequency.
1175  *
1176  * Note: mutex(&tod_lock) is in effect.
1177  */
1178 void
1179 clock_update(int offset)
1180 {
1181 	int ltemp, mtemp, s;
1182 
1183 	ASSERT(MUTEX_HELD(&tod_lock));
1184 
1185 	if (!(time_status & STA_PLL) && !(time_status & STA_PPSTIME))
1186 		return;
1187 	ltemp = offset;
1188 	if ((time_status & STA_PPSTIME) && (time_status & STA_PPSSIGNAL))
1189 		ltemp = pps_offset;
1190 
1191 	/*
1192 	 * Scale the phase adjustment and clamp to the operating range.
1193 	 */
1194 	if (ltemp > MAXPHASE)
1195 		time_offset = MAXPHASE * SCALE_UPDATE;
1196 	else if (ltemp < -MAXPHASE)
1197 		time_offset = -(MAXPHASE * SCALE_UPDATE);
1198 	else
1199 		time_offset = ltemp * SCALE_UPDATE;
1200 
1201 	/*
1202 	 * Select whether the frequency is to be controlled and in which
1203 	 * mode (PLL or FLL). Clamp to the operating range. Ugly
1204 	 * multiply/divide should be replaced someday.
1205 	 */
1206 	if (time_status & STA_FREQHOLD || time_reftime == 0)
1207 		time_reftime = hrestime.tv_sec;
1208 
1209 	mtemp = hrestime.tv_sec - time_reftime;
1210 	time_reftime = hrestime.tv_sec;
1211 
1212 	if (time_status & STA_FLL) {
1213 		if (mtemp >= MINSEC) {
1214 			ltemp = ((time_offset / mtemp) * (SCALE_USEC /
1215 			    SCALE_UPDATE));
1216 			if (ltemp)
1217 				time_freq += ltemp / SCALE_KH;
1218 		}
1219 	} else {
1220 		if (mtemp < MAXSEC) {
1221 			ltemp *= mtemp;
1222 			if (ltemp)
1223 				time_freq += (int)(((int64_t)ltemp *
1224 				    SCALE_USEC) / SCALE_KF)
1225 				    / (1 << (time_constant * 2));
1226 		}
1227 	}
1228 	if (time_freq > time_tolerance)
1229 		time_freq = time_tolerance;
1230 	else if (time_freq < -time_tolerance)
1231 		time_freq = -time_tolerance;
1232 
1233 	s = hr_clock_lock();
1234 	tod_needsync = 1;
1235 	hr_clock_unlock(s);
1236 }
1237 
1238 /*
1239  * ddi_hardpps() - discipline CPU clock oscillator to external PPS signal
1240  *
1241  * This routine is called at each PPS interrupt in order to discipline
1242  * the CPU clock oscillator to the PPS signal. It measures the PPS phase
1243  * and leaves it in a handy spot for the clock() routine. It
1244  * integrates successive PPS phase differences and calculates the
1245  * frequency offset. This is used in clock() to discipline the CPU
1246  * clock oscillator so that intrinsic frequency error is cancelled out.
1247  * The code requires the caller to capture the time and hardware counter
1248  * value at the on-time PPS signal transition.
1249  *
1250  * Note that, on some Unix systems, this routine runs at an interrupt
1251  * priority level higher than the timer interrupt routine clock().
1252  * Therefore, the variables used are distinct from the clock()
1253  * variables, except for certain exceptions: The PPS frequency pps_freq
1254  * and phase pps_offset variables are determined by this routine and
1255  * updated atomically. The time_tolerance variable can be considered a
1256  * constant, since it is infrequently changed, and then only when the
1257  * PPS signal is disabled. The watchdog counter pps_valid is updated
1258  * once per second by clock() and is atomically cleared in this
1259  * routine.
1260  *
1261  * tvp is the time of the last tick; usec is a microsecond count since the
1262  * last tick.
1263  *
1264  * Note: In Solaris systems, the tick value is actually given by
1265  *       usec_per_tick.  This is called from the serial driver cdintr(),
1266  *	 or equivalent, at a high PIL.  Because the kernel keeps a
1267  *	 highresolution time, the following code can accept either
1268  *	 the traditional argument pair, or the current highres timestamp
1269  *       in tvp and zero in usec.
1270  */
1271 void
1272 ddi_hardpps(struct timeval *tvp, int usec)
1273 {
1274 	int u_usec, v_usec, bigtick;
1275 	time_t cal_sec;
1276 	int cal_usec;
1277 
1278 	/*
1279 	 * An occasional glitch can be produced when the PPS interrupt
1280 	 * occurs in the clock() routine before the time variable is
1281 	 * updated. Here the offset is discarded when the difference
1282 	 * between it and the last one is greater than tick/2, but not
1283 	 * if the interval since the first discard exceeds 30 s.
1284 	 */
1285 	time_status |= STA_PPSSIGNAL;
1286 	time_status &= ~(STA_PPSJITTER | STA_PPSWANDER | STA_PPSERROR);
1287 	pps_valid = 0;
1288 	u_usec = -tvp->tv_usec;
1289 	if (u_usec < -(MICROSEC/2))
1290 		u_usec += MICROSEC;
1291 	v_usec = pps_offset - u_usec;
1292 	if (v_usec < 0)
1293 		v_usec = -v_usec;
1294 	if (v_usec > (usec_per_tick >> 1)) {
1295 		if (pps_glitch > MAXGLITCH) {
1296 			pps_glitch = 0;
1297 			pps_tf[2] = u_usec;
1298 			pps_tf[1] = u_usec;
1299 		} else {
1300 			pps_glitch++;
1301 			u_usec = pps_offset;
1302 		}
1303 	} else
1304 		pps_glitch = 0;
1305 
1306 	/*
1307 	 * A three-stage median filter is used to help deglitch the pps
1308 	 * time. The median sample becomes the time offset estimate; the
1309 	 * difference between the other two samples becomes the time
1310 	 * dispersion (jitter) estimate.
1311 	 */
1312 	pps_tf[2] = pps_tf[1];
1313 	pps_tf[1] = pps_tf[0];
1314 	pps_tf[0] = u_usec;
1315 	if (pps_tf[0] > pps_tf[1]) {
1316 		if (pps_tf[1] > pps_tf[2]) {
1317 			pps_offset = pps_tf[1];		/* 0 1 2 */
1318 			v_usec = pps_tf[0] - pps_tf[2];
1319 		} else if (pps_tf[2] > pps_tf[0]) {
1320 			pps_offset = pps_tf[0];		/* 2 0 1 */
1321 			v_usec = pps_tf[2] - pps_tf[1];
1322 		} else {
1323 			pps_offset = pps_tf[2];		/* 0 2 1 */
1324 			v_usec = pps_tf[0] - pps_tf[1];
1325 		}
1326 	} else {
1327 		if (pps_tf[1] < pps_tf[2]) {
1328 			pps_offset = pps_tf[1];		/* 2 1 0 */
1329 			v_usec = pps_tf[2] - pps_tf[0];
1330 		} else  if (pps_tf[2] < pps_tf[0]) {
1331 			pps_offset = pps_tf[0];		/* 1 0 2 */
1332 			v_usec = pps_tf[1] - pps_tf[2];
1333 		} else {
1334 			pps_offset = pps_tf[2];		/* 1 2 0 */
1335 			v_usec = pps_tf[1] - pps_tf[0];
1336 		}
1337 	}
1338 	if (v_usec > MAXTIME)
1339 		pps_jitcnt++;
1340 	v_usec = (v_usec << PPS_AVG) - pps_jitter;
1341 	pps_jitter += v_usec / (1 << PPS_AVG);
1342 	if (pps_jitter > (MAXTIME >> 1))
1343 		time_status |= STA_PPSJITTER;
1344 
1345 	/*
1346 	 * During the calibration interval adjust the starting time when
1347 	 * the tick overflows. At the end of the interval compute the
1348 	 * duration of the interval and the difference of the hardware
1349 	 * counters at the beginning and end of the interval. This code
1350 	 * is deliciously complicated by the fact valid differences may
1351 	 * exceed the value of tick when using long calibration
1352 	 * intervals and small ticks. Note that the counter can be
1353 	 * greater than tick if caught at just the wrong instant, but
1354 	 * the values returned and used here are correct.
1355 	 */
1356 	bigtick = (int)usec_per_tick * SCALE_USEC;
1357 	pps_usec -= pps_freq;
1358 	if (pps_usec >= bigtick)
1359 		pps_usec -= bigtick;
1360 	if (pps_usec < 0)
1361 		pps_usec += bigtick;
1362 	pps_time.tv_sec++;
1363 	pps_count++;
1364 	if (pps_count < (1 << pps_shift))
1365 		return;
1366 	pps_count = 0;
1367 	pps_calcnt++;
1368 	u_usec = usec * SCALE_USEC;
1369 	v_usec = pps_usec - u_usec;
1370 	if (v_usec >= bigtick >> 1)
1371 		v_usec -= bigtick;
1372 	if (v_usec < -(bigtick >> 1))
1373 		v_usec += bigtick;
1374 	if (v_usec < 0)
1375 		v_usec = -(-v_usec >> pps_shift);
1376 	else
1377 		v_usec = v_usec >> pps_shift;
1378 	pps_usec = u_usec;
1379 	cal_sec = tvp->tv_sec;
1380 	cal_usec = tvp->tv_usec;
1381 	cal_sec -= pps_time.tv_sec;
1382 	cal_usec -= pps_time.tv_usec;
1383 	if (cal_usec < 0) {
1384 		cal_usec += MICROSEC;
1385 		cal_sec--;
1386 	}
1387 	pps_time = *tvp;
1388 
1389 	/*
1390 	 * Check for lost interrupts, noise, excessive jitter and
1391 	 * excessive frequency error. The number of timer ticks during
1392 	 * the interval may vary +-1 tick. Add to this a margin of one
1393 	 * tick for the PPS signal jitter and maximum frequency
1394 	 * deviation. If the limits are exceeded, the calibration
1395 	 * interval is reset to the minimum and we start over.
1396 	 */
1397 	u_usec = (int)usec_per_tick << 1;
1398 	if (!((cal_sec == -1 && cal_usec > (MICROSEC - u_usec)) ||
1399 	    (cal_sec == 0 && cal_usec < u_usec)) ||
1400 	    v_usec > time_tolerance || v_usec < -time_tolerance) {
1401 		pps_errcnt++;
1402 		pps_shift = PPS_SHIFT;
1403 		pps_intcnt = 0;
1404 		time_status |= STA_PPSERROR;
1405 		return;
1406 	}
1407 
1408 	/*
1409 	 * A three-stage median filter is used to help deglitch the pps
1410 	 * frequency. The median sample becomes the frequency offset
1411 	 * estimate; the difference between the other two samples
1412 	 * becomes the frequency dispersion (stability) estimate.
1413 	 */
1414 	pps_ff[2] = pps_ff[1];
1415 	pps_ff[1] = pps_ff[0];
1416 	pps_ff[0] = v_usec;
1417 	if (pps_ff[0] > pps_ff[1]) {
1418 		if (pps_ff[1] > pps_ff[2]) {
1419 			u_usec = pps_ff[1];		/* 0 1 2 */
1420 			v_usec = pps_ff[0] - pps_ff[2];
1421 		} else if (pps_ff[2] > pps_ff[0]) {
1422 			u_usec = pps_ff[0];		/* 2 0 1 */
1423 			v_usec = pps_ff[2] - pps_ff[1];
1424 		} else {
1425 			u_usec = pps_ff[2];		/* 0 2 1 */
1426 			v_usec = pps_ff[0] - pps_ff[1];
1427 		}
1428 	} else {
1429 		if (pps_ff[1] < pps_ff[2]) {
1430 			u_usec = pps_ff[1];		/* 2 1 0 */
1431 			v_usec = pps_ff[2] - pps_ff[0];
1432 		} else  if (pps_ff[2] < pps_ff[0]) {
1433 			u_usec = pps_ff[0];		/* 1 0 2 */
1434 			v_usec = pps_ff[1] - pps_ff[2];
1435 		} else {
1436 			u_usec = pps_ff[2];		/* 1 2 0 */
1437 			v_usec = pps_ff[1] - pps_ff[0];
1438 		}
1439 	}
1440 
1441 	/*
1442 	 * Here the frequency dispersion (stability) is updated. If it
1443 	 * is less than one-fourth the maximum (MAXFREQ), the frequency
1444 	 * offset is updated as well, but clamped to the tolerance. It
1445 	 * will be processed later by the clock() routine.
1446 	 */
1447 	v_usec = (v_usec >> 1) - pps_stabil;
1448 	if (v_usec < 0)
1449 		pps_stabil -= -v_usec >> PPS_AVG;
1450 	else
1451 		pps_stabil += v_usec >> PPS_AVG;
1452 	if (pps_stabil > MAXFREQ >> 2) {
1453 		pps_stbcnt++;
1454 		time_status |= STA_PPSWANDER;
1455 		return;
1456 	}
1457 	if (time_status & STA_PPSFREQ) {
1458 		if (u_usec < 0) {
1459 			pps_freq -= -u_usec >> PPS_AVG;
1460 			if (pps_freq < -time_tolerance)
1461 				pps_freq = -time_tolerance;
1462 			u_usec = -u_usec;
1463 		} else {
1464 			pps_freq += u_usec >> PPS_AVG;
1465 			if (pps_freq > time_tolerance)
1466 				pps_freq = time_tolerance;
1467 		}
1468 	}
1469 
1470 	/*
1471 	 * Here the calibration interval is adjusted. If the maximum
1472 	 * time difference is greater than tick / 4, reduce the interval
1473 	 * by half. If this is not the case for four consecutive
1474 	 * intervals, double the interval.
1475 	 */
1476 	if (u_usec << pps_shift > bigtick >> 2) {
1477 		pps_intcnt = 0;
1478 		if (pps_shift > PPS_SHIFT)
1479 			pps_shift--;
1480 	} else if (pps_intcnt >= 4) {
1481 		pps_intcnt = 0;
1482 		if (pps_shift < PPS_SHIFTMAX)
1483 			pps_shift++;
1484 	} else
1485 		pps_intcnt++;
1486 
1487 	/*
1488 	 * If recovering from kmdb, then make sure the tod chip gets resynced.
1489 	 * If we took an early exit above, then we don't yet have a stable
1490 	 * calibration signal to lock onto, so don't mark the tod for sync
1491 	 * until we get all the way here.
1492 	 */
1493 	{
1494 		int s = hr_clock_lock();
1495 
1496 		tod_needsync = 1;
1497 		hr_clock_unlock(s);
1498 	}
1499 }
1500 
1501 /*
1502  * Handle clock tick processing for a thread.
1503  * Check for timer action, enforce CPU rlimit, do profiling etc.
1504  */
1505 void
1506 clock_tick(kthread_t *t)
1507 {
1508 	struct proc *pp;
1509 	klwp_id_t    lwp;
1510 	struct as *as;
1511 	clock_t	utime;
1512 	clock_t	stime;
1513 	int	poke = 0;		/* notify another CPU */
1514 	int	user_mode;
1515 	size_t	 rss;
1516 
1517 	/* Must be operating on a lwp/thread */
1518 	if ((lwp = ttolwp(t)) == NULL) {
1519 		panic("clock_tick: no lwp");
1520 		/*NOTREACHED*/
1521 	}
1522 
1523 	CL_TICK(t);	/* Class specific tick processing */
1524 	DTRACE_SCHED1(tick, kthread_t *, t);
1525 
1526 	pp = ttoproc(t);
1527 
1528 	/* pp->p_lock makes sure that the thread does not exit */
1529 	ASSERT(MUTEX_HELD(&pp->p_lock));
1530 
1531 	user_mode = (lwp->lwp_state == LWP_USER);
1532 
1533 	/*
1534 	 * Update process times. Should use high res clock and state
1535 	 * changes instead of statistical sampling method. XXX
1536 	 */
1537 	if (user_mode) {
1538 		pp->p_utime++;
1539 		pp->p_task->tk_cpu_time++;
1540 	} else {
1541 		pp->p_stime++;
1542 		pp->p_task->tk_cpu_time++;
1543 	}
1544 	as = pp->p_as;
1545 
1546 	/*
1547 	 * Update user profiling statistics. Get the pc from the
1548 	 * lwp when the AST happens.
1549 	 */
1550 	if (pp->p_prof.pr_scale) {
1551 		atomic_add_32(&lwp->lwp_oweupc, 1);
1552 		if (user_mode) {
1553 			poke = 1;
1554 			aston(t);
1555 		}
1556 	}
1557 
1558 	utime = pp->p_utime;
1559 	stime = pp->p_stime;
1560 
1561 	/*
1562 	 * If CPU was in user state, process lwp-virtual time
1563 	 * interval timer.
1564 	 */
1565 	if (user_mode &&
1566 	    timerisset(&lwp->lwp_timer[ITIMER_VIRTUAL].it_value) &&
1567 	    itimerdecr(&lwp->lwp_timer[ITIMER_VIRTUAL], usec_per_tick) == 0) {
1568 		poke = 1;
1569 		sigtoproc(pp, t, SIGVTALRM);
1570 	}
1571 
1572 	if (timerisset(&lwp->lwp_timer[ITIMER_PROF].it_value) &&
1573 	    itimerdecr(&lwp->lwp_timer[ITIMER_PROF], usec_per_tick) == 0) {
1574 		poke = 1;
1575 		sigtoproc(pp, t, SIGPROF);
1576 	}
1577 
1578 	/*
1579 	 * Enforce CPU resource controls:
1580 	 *   (a) process.max-cpu-time resource control
1581 	 */
1582 	(void) rctl_test(rctlproc_legacy[RLIMIT_CPU], pp->p_rctls, pp,
1583 	    (utime + stime)/hz, RCA_UNSAFE_SIGINFO);
1584 
1585 	/*
1586 	 *   (b) task.max-cpu-time resource control
1587 	 */
1588 	(void) rctl_test(rc_task_cpu_time, pp->p_task->tk_rctls, pp, 1,
1589 	    RCA_UNSAFE_SIGINFO);
1590 
1591 	/*
1592 	 * Update memory usage for the currently running process.
1593 	 */
1594 	rss = rm_asrss(as);
1595 	PTOU(pp)->u_mem += rss;
1596 	if (rss > PTOU(pp)->u_mem_max)
1597 		PTOU(pp)->u_mem_max = rss;
1598 
1599 	/*
1600 	 * Notify the CPU the thread is running on.
1601 	 */
1602 	if (poke && t->t_cpu != CPU)
1603 		poke_cpu(t->t_cpu->cpu_id);
1604 }
1605 
1606 void
1607 profil_tick(uintptr_t upc)
1608 {
1609 	int ticks;
1610 	proc_t *p = ttoproc(curthread);
1611 	klwp_t *lwp = ttolwp(curthread);
1612 	struct prof *pr = &p->p_prof;
1613 
1614 	do {
1615 		ticks = lwp->lwp_oweupc;
1616 	} while (cas32(&lwp->lwp_oweupc, ticks, 0) != ticks);
1617 
1618 	mutex_enter(&p->p_pflock);
1619 	if (pr->pr_scale >= 2 && upc >= pr->pr_off) {
1620 		/*
1621 		 * Old-style profiling
1622 		 */
1623 		uint16_t *slot = pr->pr_base;
1624 		uint16_t old, new;
1625 		if (pr->pr_scale != 2) {
1626 			uintptr_t delta = upc - pr->pr_off;
1627 			uintptr_t byteoff = ((delta >> 16) * pr->pr_scale) +
1628 			    (((delta & 0xffff) * pr->pr_scale) >> 16);
1629 			if (byteoff >= (uintptr_t)pr->pr_size) {
1630 				mutex_exit(&p->p_pflock);
1631 				return;
1632 			}
1633 			slot += byteoff / sizeof (uint16_t);
1634 		}
1635 		if (fuword16(slot, &old) < 0 ||
1636 		    (new = old + ticks) > SHRT_MAX ||
1637 		    suword16(slot, new) < 0) {
1638 			pr->pr_scale = 0;
1639 		}
1640 	} else if (pr->pr_scale == 1) {
1641 		/*
1642 		 * PC Sampling
1643 		 */
1644 		model_t model = lwp_getdatamodel(lwp);
1645 		int result;
1646 #ifdef __lint
1647 		model = model;
1648 #endif
1649 		while (ticks-- > 0) {
1650 			if (pr->pr_samples == pr->pr_size) {
1651 				/* buffer full, turn off sampling */
1652 				pr->pr_scale = 0;
1653 				break;
1654 			}
1655 			switch (SIZEOF_PTR(model)) {
1656 			case sizeof (uint32_t):
1657 				result = suword32(pr->pr_base, (uint32_t)upc);
1658 				break;
1659 #ifdef _LP64
1660 			case sizeof (uint64_t):
1661 				result = suword64(pr->pr_base, (uint64_t)upc);
1662 				break;
1663 #endif
1664 			default:
1665 				cmn_err(CE_WARN, "profil_tick: unexpected "
1666 				    "data model");
1667 				result = -1;
1668 				break;
1669 			}
1670 			if (result != 0) {
1671 				pr->pr_scale = 0;
1672 				break;
1673 			}
1674 			pr->pr_base = (caddr_t)pr->pr_base + SIZEOF_PTR(model);
1675 			pr->pr_samples++;
1676 		}
1677 	}
1678 	mutex_exit(&p->p_pflock);
1679 }
1680 
1681 static void
1682 delay_wakeup(void *arg)
1683 {
1684 	kthread_t *t = arg;
1685 
1686 	mutex_enter(&t->t_delay_lock);
1687 	cv_signal(&t->t_delay_cv);
1688 	mutex_exit(&t->t_delay_lock);
1689 }
1690 
1691 void
1692 delay(clock_t ticks)
1693 {
1694 	kthread_t *t = curthread;
1695 	clock_t deadline = lbolt + ticks;
1696 	clock_t timeleft;
1697 	timeout_id_t id;
1698 
1699 	if (panicstr && ticks > 0) {
1700 		/*
1701 		 * Timeouts aren't running, so all we can do is spin.
1702 		 */
1703 		drv_usecwait(TICK_TO_USEC(ticks));
1704 		return;
1705 	}
1706 
1707 	while ((timeleft = deadline - lbolt) > 0) {
1708 		mutex_enter(&t->t_delay_lock);
1709 		id = timeout(delay_wakeup, t, timeleft);
1710 		cv_wait(&t->t_delay_cv, &t->t_delay_lock);
1711 		mutex_exit(&t->t_delay_lock);
1712 		(void) untimeout(id);
1713 	}
1714 }
1715 
1716 /*
1717  * Like delay, but interruptible by a signal.
1718  */
1719 int
1720 delay_sig(clock_t ticks)
1721 {
1722 	clock_t deadline = lbolt + ticks;
1723 	clock_t rc;
1724 
1725 	mutex_enter(&curthread->t_delay_lock);
1726 	do {
1727 		rc = cv_timedwait_sig(&curthread->t_delay_cv,
1728 		    &curthread->t_delay_lock, deadline);
1729 	} while (rc > 0);
1730 	mutex_exit(&curthread->t_delay_lock);
1731 	if (rc == 0)
1732 		return (EINTR);
1733 	return (0);
1734 }
1735 
1736 #define	SECONDS_PER_DAY 86400
1737 
1738 /*
1739  * Initialize the system time based on the TOD chip.  approx is used as
1740  * an approximation of time (e.g. from the filesystem) in the event that
1741  * the TOD chip has been cleared or is unresponsive.  An approx of -1
1742  * means the filesystem doesn't keep time.
1743  */
1744 void
1745 clkset(time_t approx)
1746 {
1747 	timestruc_t ts;
1748 	int spl;
1749 	int set_clock = 0;
1750 
1751 	mutex_enter(&tod_lock);
1752 	ts = tod_get();
1753 
1754 	if (ts.tv_sec > 365 * SECONDS_PER_DAY) {
1755 		/*
1756 		 * If the TOD chip is reporting some time after 1971,
1757 		 * then it probably didn't lose power or become otherwise
1758 		 * cleared in the recent past;  check to assure that
1759 		 * the time coming from the filesystem isn't in the future
1760 		 * according to the TOD chip.
1761 		 */
1762 		if (approx != -1 && approx > ts.tv_sec) {
1763 			cmn_err(CE_WARN, "Last shutdown is later "
1764 			    "than time on time-of-day chip; check date.");
1765 		}
1766 	} else {
1767 		/*
1768 		 * If the TOD chip isn't giving correct time, then set it to
1769 		 * the time that was passed in as a rough estimate.  If we
1770 		 * don't have an estimate, then set the clock back to a time
1771 		 * when Oliver North, ALF and Dire Straits were all on the
1772 		 * collective brain:  1987.
1773 		 */
1774 		timestruc_t tmp;
1775 		if (approx == -1)
1776 			ts.tv_sec = (1987 - 1970) * 365 * SECONDS_PER_DAY;
1777 		else
1778 			ts.tv_sec = approx;
1779 		ts.tv_nsec = 0;
1780 
1781 		/*
1782 		 * Attempt to write the new time to the TOD chip.  Set spl high
1783 		 * to avoid getting preempted between the tod_set and tod_get.
1784 		 */
1785 		spl = splhi();
1786 		tod_set(ts);
1787 		tmp = tod_get();
1788 		splx(spl);
1789 
1790 		if (tmp.tv_sec != ts.tv_sec && tmp.tv_sec != ts.tv_sec + 1) {
1791 			tod_broken = 1;
1792 			dosynctodr = 0;
1793 			cmn_err(CE_WARN, "Time-of-day chip unresponsive;"
1794 			    " dead batteries?");
1795 		} else {
1796 			cmn_err(CE_WARN, "Time-of-day chip had "
1797 			    "incorrect date; check and reset.");
1798 		}
1799 		set_clock = 1;
1800 	}
1801 
1802 	if (!boot_time) {
1803 		boot_time = ts.tv_sec;
1804 		set_clock = 1;
1805 	}
1806 
1807 	if (set_clock)
1808 		set_hrestime(&ts);
1809 
1810 	mutex_exit(&tod_lock);
1811 }
1812 
1813 int	timechanged;	/* for testing if the system time has been reset */
1814 
1815 void
1816 set_hrestime(timestruc_t *ts)
1817 {
1818 	int spl = hr_clock_lock();
1819 	hrestime = *ts;
1820 	membar_enter();	/* hrestime must be visible before timechanged++ */
1821 	timedelta = 0;
1822 	timechanged++;
1823 	hr_clock_unlock(spl);
1824 }
1825 
1826 static uint_t deadman_seconds;
1827 static uint32_t deadman_panics;
1828 static int deadman_enabled = 0;
1829 static int deadman_panic_timers = 1;
1830 
1831 static void
1832 deadman(void)
1833 {
1834 	if (panicstr) {
1835 		/*
1836 		 * During panic, other CPUs besides the panic
1837 		 * master continue to handle cyclics and some other
1838 		 * interrupts.  The code below is intended to be
1839 		 * single threaded, so any CPU other than the master
1840 		 * must keep out.
1841 		 */
1842 		if (CPU->cpu_id != panic_cpu.cpu_id)
1843 			return;
1844 
1845 		/*
1846 		 * If we're panicking, the deadman cyclic continues to increase
1847 		 * lbolt in case the dump device driver relies on this for
1848 		 * timeouts.  Note that we rely on deadman() being invoked once
1849 		 * per second, and credit lbolt and lbolt64 with hz ticks each.
1850 		 */
1851 		lbolt += hz;
1852 		lbolt64 += hz;
1853 
1854 		if (!deadman_panic_timers)
1855 			return; /* allow all timers to be manually disabled */
1856 
1857 		/*
1858 		 * If we are generating a crash dump or syncing filesystems and
1859 		 * the corresponding timer is set, decrement it and re-enter
1860 		 * the panic code to abort it and advance to the next state.
1861 		 * The panic states and triggers are explained in panic.c.
1862 		 */
1863 		if (panic_dump) {
1864 			if (dump_timeleft && (--dump_timeleft == 0)) {
1865 				panic("panic dump timeout");
1866 				/*NOTREACHED*/
1867 			}
1868 		} else if (panic_sync) {
1869 			if (sync_timeleft && (--sync_timeleft == 0)) {
1870 				panic("panic sync timeout");
1871 				/*NOTREACHED*/
1872 			}
1873 		}
1874 
1875 		return;
1876 	}
1877 
1878 	if (lbolt != CPU->cpu_deadman_lbolt) {
1879 		CPU->cpu_deadman_lbolt = lbolt;
1880 		CPU->cpu_deadman_countdown = deadman_seconds;
1881 		return;
1882 	}
1883 
1884 	if (CPU->cpu_deadman_countdown-- > 0)
1885 		return;
1886 
1887 	/*
1888 	 * Regardless of whether or not we actually bring the system down,
1889 	 * bump the deadman_panics variable.
1890 	 *
1891 	 * N.B. deadman_panics is incremented once for each CPU that
1892 	 * passes through here.  It's expected that all the CPUs will
1893 	 * detect this condition within one second of each other, so
1894 	 * when deadman_enabled is off, deadman_panics will
1895 	 * typically be a multiple of the total number of CPUs in
1896 	 * the system.
1897 	 */
1898 	atomic_add_32(&deadman_panics, 1);
1899 
1900 	if (!deadman_enabled) {
1901 		CPU->cpu_deadman_countdown = deadman_seconds;
1902 		return;
1903 	}
1904 
1905 	/*
1906 	 * If we're here, we want to bring the system down.
1907 	 */
1908 	panic("deadman: timed out after %d seconds of clock "
1909 	    "inactivity", deadman_seconds);
1910 	/*NOTREACHED*/
1911 }
1912 
1913 /*ARGSUSED*/
1914 static void
1915 deadman_online(void *arg, cpu_t *cpu, cyc_handler_t *hdlr, cyc_time_t *when)
1916 {
1917 	cpu->cpu_deadman_lbolt = 0;
1918 	cpu->cpu_deadman_countdown = deadman_seconds;
1919 
1920 	hdlr->cyh_func = (cyc_func_t)deadman;
1921 	hdlr->cyh_level = CY_HIGH_LEVEL;
1922 	hdlr->cyh_arg = NULL;
1923 
1924 	/*
1925 	 * Stagger the CPUs so that they don't all run deadman() at
1926 	 * the same time.  Simplest reason to do this is to make it
1927 	 * more likely that only one CPU will panic in case of a
1928 	 * timeout.  This is (strictly speaking) an aesthetic, not a
1929 	 * technical consideration.
1930 	 *
1931 	 * The interval must be one second in accordance with the
1932 	 * code in deadman() above to increase lbolt during panic.
1933 	 */
1934 	when->cyt_when = cpu->cpu_id * (NANOSEC / NCPU);
1935 	when->cyt_interval = NANOSEC;
1936 }
1937 
1938 
1939 void
1940 deadman_init(void)
1941 {
1942 	cyc_omni_handler_t hdlr;
1943 
1944 	if (deadman_seconds == 0)
1945 		deadman_seconds = snoop_interval / MICROSEC;
1946 
1947 	if (snooping)
1948 		deadman_enabled = 1;
1949 
1950 	hdlr.cyo_online = deadman_online;
1951 	hdlr.cyo_offline = NULL;
1952 	hdlr.cyo_arg = NULL;
1953 
1954 	mutex_enter(&cpu_lock);
1955 	deadman_cyclic = cyclic_add_omni(&hdlr);
1956 	mutex_exit(&cpu_lock);
1957 }
1958 
1959 /*
1960  * tod_fault() is for updating tod validate mechanism state:
1961  * (1) TOD_NOFAULT: for resetting the state to 'normal'.
1962  *     currently used for debugging only
1963  * (2) The following four cases detected by tod validate mechanism:
1964  *       TOD_REVERSED: current tod value is less than previous value.
1965  *       TOD_STALLED: current tod value hasn't advanced.
1966  *       TOD_JUMPED: current tod value advanced too far from previous value.
1967  *       TOD_RATECHANGED: the ratio between average tod delta and
1968  *       average tick delta has changed.
1969  */
1970 enum tod_fault_type
1971 tod_fault(enum tod_fault_type ftype, int off)
1972 {
1973 	ASSERT(MUTEX_HELD(&tod_lock));
1974 
1975 	if (tod_faulted != ftype) {
1976 		switch (ftype) {
1977 		case TOD_NOFAULT:
1978 			plat_tod_fault(TOD_NOFAULT);
1979 			cmn_err(CE_NOTE, "Restarted tracking "
1980 					"Time of Day clock.");
1981 			tod_faulted = ftype;
1982 			break;
1983 		case TOD_REVERSED:
1984 		case TOD_JUMPED:
1985 			if (tod_faulted == TOD_NOFAULT) {
1986 				plat_tod_fault(ftype);
1987 				cmn_err(CE_WARN, "Time of Day clock error: "
1988 				    "reason [%s by 0x%x]. -- "
1989 				    " Stopped tracking Time Of Day clock.",
1990 				    tod_fault_table[ftype], off);
1991 				tod_faulted = ftype;
1992 			}
1993 			break;
1994 		case TOD_STALLED:
1995 		case TOD_RATECHANGED:
1996 			if (tod_faulted == TOD_NOFAULT) {
1997 				plat_tod_fault(ftype);
1998 				cmn_err(CE_WARN, "Time of Day clock error: "
1999 				    "reason [%s]. -- "
2000 				    " Stopped tracking Time Of Day clock.",
2001 				    tod_fault_table[ftype]);
2002 				tod_faulted = ftype;
2003 			}
2004 			break;
2005 		default:
2006 			break;
2007 		}
2008 	}
2009 	return (tod_faulted);
2010 }
2011 
2012 void
2013 tod_fault_reset()
2014 {
2015 	tod_fault_reset_flag = 1;
2016 }
2017 
2018 
2019 /*
2020  * tod_validate() is used for checking values returned by tod_get().
2021  * Four error cases can be detected by this routine:
2022  *   TOD_REVERSED: current tod value is less than previous.
2023  *   TOD_STALLED: current tod value hasn't advanced.
2024  *   TOD_JUMPED: current tod value advanced too far from previous value.
2025  *   TOD_RATECHANGED: the ratio between average tod delta and
2026  *   average tick delta has changed.
2027  */
2028 time_t
2029 tod_validate(time_t tod)
2030 {
2031 	time_t diff_tod;
2032 	hrtime_t diff_tick;
2033 
2034 	long dtick;
2035 	int dtick_delta;
2036 
2037 	int off = 0;
2038 	enum tod_fault_type tod_bad = TOD_NOFAULT;
2039 
2040 	static int firsttime = 1;
2041 
2042 	static time_t prev_tod = 0;
2043 	static hrtime_t prev_tick = 0;
2044 	static long dtick_avg = TOD_REF_FREQ;
2045 
2046 	hrtime_t tick = gethrtime();
2047 
2048 	ASSERT(MUTEX_HELD(&tod_lock));
2049 
2050 	/*
2051 	 * tod_validate_enable is patchable via /etc/system.
2052 	 * If TOD is already faulted, or if TOD validation is deferred,
2053 	 * there is nothing to do.
2054 	 */
2055 	if ((tod_validate_enable == 0) || (tod_faulted != TOD_NOFAULT) ||
2056 	    tod_validate_deferred) {
2057 		return (tod);
2058 	}
2059 
2060 	/*
2061 	 * Update prev_tod and prev_tick values for first run
2062 	 */
2063 	if (firsttime) {
2064 		firsttime = 0;
2065 		prev_tod = tod;
2066 		prev_tick = tick;
2067 		return (tod);
2068 	}
2069 
2070 	/*
2071 	 * For either of these conditions, we need to reset ourself
2072 	 * and start validation from zero since each condition
2073 	 * indicates that the TOD will be updated with new value
2074 	 * Also, note that tod_needsync will be reset in clock()
2075 	 */
2076 	if (tod_needsync || tod_fault_reset_flag) {
2077 		firsttime = 1;
2078 		prev_tod = 0;
2079 		prev_tick = 0;
2080 		dtick_avg = TOD_REF_FREQ;
2081 
2082 		if (tod_fault_reset_flag)
2083 			tod_fault_reset_flag = 0;
2084 
2085 		return (tod);
2086 	}
2087 
2088 	/* test hook */
2089 	switch (tod_unit_test) {
2090 	case 1: /* for testing jumping tod */
2091 		tod += tod_test_injector;
2092 		tod_unit_test = 0;
2093 		break;
2094 	case 2:	/* for testing stuck tod bit */
2095 		tod |= 1 << tod_test_injector;
2096 		tod_unit_test = 0;
2097 		break;
2098 	case 3:	/* for testing stalled tod */
2099 		tod = prev_tod;
2100 		tod_unit_test = 0;
2101 		break;
2102 	case 4:	/* reset tod fault status */
2103 		(void) tod_fault(TOD_NOFAULT, 0);
2104 		tod_unit_test = 0;
2105 		break;
2106 	default:
2107 		break;
2108 	}
2109 
2110 	diff_tod = tod - prev_tod;
2111 	diff_tick = tick - prev_tick;
2112 
2113 	ASSERT(diff_tick >= 0);
2114 
2115 	if (diff_tod < 0) {
2116 		/* ERROR - tod reversed */
2117 		tod_bad = TOD_REVERSED;
2118 		off = (int)(prev_tod - tod);
2119 	} else if (diff_tod == 0) {
2120 		/* tod did not advance */
2121 		if (diff_tick > TOD_STALL_THRESHOLD) {
2122 			/* ERROR - tod stalled */
2123 			tod_bad = TOD_STALLED;
2124 		} else {
2125 			/*
2126 			 * Make sure we don't update prev_tick
2127 			 * so that diff_tick is calculated since
2128 			 * the first diff_tod == 0
2129 			 */
2130 			return (tod);
2131 		}
2132 	} else {
2133 		/* calculate dtick */
2134 		dtick = diff_tick / diff_tod;
2135 
2136 		/* update dtick averages */
2137 		dtick_avg += ((dtick - dtick_avg) / TOD_FILTER_N);
2138 
2139 		/*
2140 		 * Calculate dtick_delta as
2141 		 * variation from reference freq in quartiles
2142 		 */
2143 		dtick_delta = (dtick_avg - TOD_REF_FREQ) /
2144 			(TOD_REF_FREQ >> 2);
2145 
2146 		/*
2147 		 * Even with a perfectly functioning TOD device,
2148 		 * when the number of elapsed seconds is low the
2149 		 * algorithm can calculate a rate that is beyond
2150 		 * tolerance, causing an error.  The algorithm is
2151 		 * inaccurate when elapsed time is low (less than
2152 		 * 5 seconds).
2153 		 */
2154 		if (diff_tod > 4) {
2155 			if (dtick < TOD_JUMP_THRESHOLD) {
2156 				/* ERROR - tod jumped */
2157 				tod_bad = TOD_JUMPED;
2158 				off = (int)diff_tod;
2159 			} else if (dtick_delta) {
2160 				/* ERROR - change in clock rate */
2161 				tod_bad = TOD_RATECHANGED;
2162 			}
2163 		}
2164 	}
2165 
2166 	if (tod_bad != TOD_NOFAULT) {
2167 		(void) tod_fault(tod_bad, off);
2168 
2169 		/*
2170 		 * Disable dosynctodr since we are going to fault
2171 		 * the TOD chip anyway here
2172 		 */
2173 		dosynctodr = 0;
2174 
2175 		/*
2176 		 * Set tod to the correct value from hrestime
2177 		 */
2178 		tod = hrestime.tv_sec;
2179 	}
2180 
2181 	prev_tod = tod;
2182 	prev_tick = tick;
2183 	return (tod);
2184 }
2185 
2186 static void
2187 calcloadavg(int nrun, uint64_t *hp_ave)
2188 {
2189 	static int64_t f[3] = { 135, 27, 9 };
2190 	uint_t i;
2191 	int64_t q, r;
2192 
2193 	/*
2194 	 * Compute load average over the last 1, 5, and 15 minutes
2195 	 * (60, 300, and 900 seconds).  The constants in f[3] are for
2196 	 * exponential decay:
2197 	 * (1 - exp(-1/60)) << 13 = 135,
2198 	 * (1 - exp(-1/300)) << 13 = 27,
2199 	 * (1 - exp(-1/900)) << 13 = 9.
2200 	 */
2201 
2202 	/*
2203 	 * a little hoop-jumping to avoid integer overflow
2204 	 */
2205 	for (i = 0; i < 3; i++) {
2206 		q = (hp_ave[i]  >> 16) << 7;
2207 		r = (hp_ave[i]  & 0xffff) << 7;
2208 		hp_ave[i] += ((nrun - q) * f[i] - ((r * f[i]) >> 16)) >> 4;
2209 	}
2210 }
2211