xref: /titanic_41/usr/src/uts/common/os/clock.c (revision 0d6bb4c6728fd20087fe25f4028a3838250e6e9c)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
22 /*	  All Rights Reserved	*/
23 
24 /*
25  * Copyright (c) 1988, 2010, Oracle and/or its affiliates. All rights reserved.
26  * Copyright (c) 2013, Joyent, Inc.  All rights reserved.
27  */
28 
29 #include <sys/param.h>
30 #include <sys/t_lock.h>
31 #include <sys/types.h>
32 #include <sys/tuneable.h>
33 #include <sys/sysmacros.h>
34 #include <sys/systm.h>
35 #include <sys/cpuvar.h>
36 #include <sys/lgrp.h>
37 #include <sys/user.h>
38 #include <sys/proc.h>
39 #include <sys/callo.h>
40 #include <sys/kmem.h>
41 #include <sys/var.h>
42 #include <sys/cmn_err.h>
43 #include <sys/swap.h>
44 #include <sys/vmsystm.h>
45 #include <sys/class.h>
46 #include <sys/time.h>
47 #include <sys/debug.h>
48 #include <sys/vtrace.h>
49 #include <sys/spl.h>
50 #include <sys/atomic.h>
51 #include <sys/dumphdr.h>
52 #include <sys/archsystm.h>
53 #include <sys/fs/swapnode.h>
54 #include <sys/panic.h>
55 #include <sys/disp.h>
56 #include <sys/msacct.h>
57 #include <sys/mem_cage.h>
58 
59 #include <vm/page.h>
60 #include <vm/anon.h>
61 #include <vm/rm.h>
62 #include <sys/cyclic.h>
63 #include <sys/cpupart.h>
64 #include <sys/rctl.h>
65 #include <sys/task.h>
66 #include <sys/sdt.h>
67 #include <sys/ddi_periodic.h>
68 #include <sys/random.h>
69 #include <sys/modctl.h>
70 #include <sys/zone.h>
71 
72 /*
73  * for NTP support
74  */
75 #include <sys/timex.h>
76 #include <sys/inttypes.h>
77 
78 #include <sys/sunddi.h>
79 #include <sys/clock_impl.h>
80 
81 /*
82  * clock() is called straight from the clock cyclic; see clock_init().
83  *
84  * Functions:
85  *	reprime clock
86  *	maintain date
87  *	jab the scheduler
88  */
89 
90 extern kcondvar_t	fsflush_cv;
91 extern sysinfo_t	sysinfo;
92 extern vminfo_t	vminfo;
93 extern int	idleswtch;	/* flag set while idle in pswtch() */
94 extern hrtime_t volatile devinfo_freeze;
95 
96 /*
97  * high-precision avenrun values.  These are needed to make the
98  * regular avenrun values accurate.
99  */
100 static uint64_t hp_avenrun[3];
101 int	avenrun[3];		/* FSCALED average run queue lengths */
102 time_t	time;	/* time in seconds since 1970 - for compatibility only */
103 
104 static struct loadavg_s loadavg;
105 /*
106  * Phase/frequency-lock loop (PLL/FLL) definitions
107  *
108  * The following variables are read and set by the ntp_adjtime() system
109  * call.
110  *
111  * time_state shows the state of the system clock, with values defined
112  * in the timex.h header file.
113  *
114  * time_status shows the status of the system clock, with bits defined
115  * in the timex.h header file.
116  *
117  * time_offset is used by the PLL/FLL to adjust the system time in small
118  * increments.
119  *
120  * time_constant determines the bandwidth or "stiffness" of the PLL.
121  *
122  * time_tolerance determines maximum frequency error or tolerance of the
123  * CPU clock oscillator and is a property of the architecture; however,
124  * in principle it could change as result of the presence of external
125  * discipline signals, for instance.
126  *
127  * time_precision is usually equal to the kernel tick variable; however,
128  * in cases where a precision clock counter or external clock is
129  * available, the resolution can be much less than this and depend on
130  * whether the external clock is working or not.
131  *
132  * time_maxerror is initialized by a ntp_adjtime() call and increased by
133  * the kernel once each second to reflect the maximum error bound
134  * growth.
135  *
136  * time_esterror is set and read by the ntp_adjtime() call, but
137  * otherwise not used by the kernel.
138  */
139 int32_t time_state = TIME_OK;	/* clock state */
140 int32_t time_status = STA_UNSYNC;	/* clock status bits */
141 int32_t time_offset = 0;		/* time offset (us) */
142 int32_t time_constant = 0;		/* pll time constant */
143 int32_t time_tolerance = MAXFREQ;	/* frequency tolerance (scaled ppm) */
144 int32_t time_precision = 1;	/* clock precision (us) */
145 int32_t time_maxerror = MAXPHASE;	/* maximum error (us) */
146 int32_t time_esterror = MAXPHASE;	/* estimated error (us) */
147 
148 /*
149  * The following variables establish the state of the PLL/FLL and the
150  * residual time and frequency offset of the local clock. The scale
151  * factors are defined in the timex.h header file.
152  *
153  * time_phase and time_freq are the phase increment and the frequency
154  * increment, respectively, of the kernel time variable.
155  *
156  * time_freq is set via ntp_adjtime() from a value stored in a file when
157  * the synchronization daemon is first started. Its value is retrieved
158  * via ntp_adjtime() and written to the file about once per hour by the
159  * daemon.
160  *
161  * time_adj is the adjustment added to the value of tick at each timer
162  * interrupt and is recomputed from time_phase and time_freq at each
163  * seconds rollover.
164  *
165  * time_reftime is the second's portion of the system time at the last
166  * call to ntp_adjtime(). It is used to adjust the time_freq variable
167  * and to increase the time_maxerror as the time since last update
168  * increases.
169  */
170 int32_t time_phase = 0;		/* phase offset (scaled us) */
171 int32_t time_freq = 0;		/* frequency offset (scaled ppm) */
172 int32_t time_adj = 0;		/* tick adjust (scaled 1 / hz) */
173 int32_t time_reftime = 0;		/* time at last adjustment (s) */
174 
175 /*
176  * The scale factors of the following variables are defined in the
177  * timex.h header file.
178  *
179  * pps_time contains the time at each calibration interval, as read by
180  * microtime(). pps_count counts the seconds of the calibration
181  * interval, the duration of which is nominally pps_shift in powers of
182  * two.
183  *
184  * pps_offset is the time offset produced by the time median filter
185  * pps_tf[], while pps_jitter is the dispersion (jitter) measured by
186  * this filter.
187  *
188  * pps_freq is the frequency offset produced by the frequency median
189  * filter pps_ff[], while pps_stabil is the dispersion (wander) measured
190  * by this filter.
191  *
192  * pps_usec is latched from a high resolution counter or external clock
193  * at pps_time. Here we want the hardware counter contents only, not the
194  * contents plus the time_tv.usec as usual.
195  *
196  * pps_valid counts the number of seconds since the last PPS update. It
197  * is used as a watchdog timer to disable the PPS discipline should the
198  * PPS signal be lost.
199  *
200  * pps_glitch counts the number of seconds since the beginning of an
201  * offset burst more than tick/2 from current nominal offset. It is used
202  * mainly to suppress error bursts due to priority conflicts between the
203  * PPS interrupt and timer interrupt.
204  *
205  * pps_intcnt counts the calibration intervals for use in the interval-
206  * adaptation algorithm. It's just too complicated for words.
207  */
208 struct timeval pps_time;	/* kernel time at last interval */
209 int32_t pps_tf[] = {0, 0, 0};	/* pps time offset median filter (us) */
210 int32_t pps_offset = 0;		/* pps time offset (us) */
211 int32_t pps_jitter = MAXTIME;	/* time dispersion (jitter) (us) */
212 int32_t pps_ff[] = {0, 0, 0};	/* pps frequency offset median filter */
213 int32_t pps_freq = 0;		/* frequency offset (scaled ppm) */
214 int32_t pps_stabil = MAXFREQ;	/* frequency dispersion (scaled ppm) */
215 int32_t pps_usec = 0;		/* microsec counter at last interval */
216 int32_t pps_valid = PPS_VALID;	/* pps signal watchdog counter */
217 int32_t pps_glitch = 0;		/* pps signal glitch counter */
218 int32_t pps_count = 0;		/* calibration interval counter (s) */
219 int32_t pps_shift = PPS_SHIFT;	/* interval duration (s) (shift) */
220 int32_t pps_intcnt = 0;		/* intervals at current duration */
221 
222 /*
223  * PPS signal quality monitors
224  *
225  * pps_jitcnt counts the seconds that have been discarded because the
226  * jitter measured by the time median filter exceeds the limit MAXTIME
227  * (100 us).
228  *
229  * pps_calcnt counts the frequency calibration intervals, which are
230  * variable from 4 s to 256 s.
231  *
232  * pps_errcnt counts the calibration intervals which have been discarded
233  * because the wander exceeds the limit MAXFREQ (100 ppm) or where the
234  * calibration interval jitter exceeds two ticks.
235  *
236  * pps_stbcnt counts the calibration intervals that have been discarded
237  * because the frequency wander exceeds the limit MAXFREQ / 4 (25 us).
238  */
239 int32_t pps_jitcnt = 0;		/* jitter limit exceeded */
240 int32_t pps_calcnt = 0;		/* calibration intervals */
241 int32_t pps_errcnt = 0;		/* calibration errors */
242 int32_t pps_stbcnt = 0;		/* stability limit exceeded */
243 
244 kcondvar_t lbolt_cv;
245 
246 /*
247  * Hybrid lbolt implementation:
248  *
249  * The service historically provided by the lbolt and lbolt64 variables has
250  * been replaced by the ddi_get_lbolt() and ddi_get_lbolt64() routines, and the
251  * original symbols removed from the system. The once clock driven variables are
252  * now implemented in an event driven fashion, backed by gethrtime() coarsed to
253  * the appropriate clock resolution. The default event driven implementation is
254  * complemented by a cyclic driven one, active only during periods of intense
255  * activity around the DDI lbolt routines, when a lbolt specific cyclic is
256  * reprogramed to fire at a clock tick interval to serve consumers of lbolt who
257  * rely on the original low cost of consulting a memory position.
258  *
259  * The implementation uses the number of calls to these routines and the
260  * frequency of these to determine when to transition from event to cyclic
261  * driven and vice-versa. These values are kept on a per CPU basis for
262  * scalability reasons and to prevent CPUs from constantly invalidating a single
263  * cache line when modifying a global variable. The transition from event to
264  * cyclic mode happens once the thresholds are crossed, and activity on any CPU
265  * can cause such transition.
266  *
267  * The lbolt_hybrid function pointer is called by ddi_get_lbolt() and
268  * ddi_get_lbolt64(), and will point to lbolt_event_driven() or
269  * lbolt_cyclic_driven() according to the current mode. When the thresholds
270  * are exceeded, lbolt_event_driven() will reprogram the lbolt cyclic to
271  * fire at a nsec_per_tick interval and increment an internal variable at
272  * each firing. lbolt_hybrid will then point to lbolt_cyclic_driven(), which
273  * will simply return the value of such variable. lbolt_cyclic() will attempt
274  * to shut itself off at each threshold interval (sampling period for calls
275  * to the DDI lbolt routines), and return to the event driven mode, but will
276  * be prevented from doing so if lbolt_cyclic_driven() is being heavily used.
277  *
278  * lbolt_bootstrap is used during boot to serve lbolt consumers who don't wait
279  * for the cyclic subsystem to be intialized.
280  *
281  */
282 int64_t lbolt_bootstrap(void);
283 int64_t lbolt_event_driven(void);
284 int64_t lbolt_cyclic_driven(void);
285 int64_t (*lbolt_hybrid)(void) = lbolt_bootstrap;
286 uint_t lbolt_ev_to_cyclic(caddr_t, caddr_t);
287 
288 /*
289  * lbolt's cyclic, installed by clock_init().
290  */
291 static void lbolt_cyclic(void);
292 
293 /*
294  * Tunable to keep lbolt in cyclic driven mode. This will prevent the system
295  * from switching back to event driven, once it reaches cyclic mode.
296  */
297 static boolean_t lbolt_cyc_only = B_FALSE;
298 
299 /*
300  * Cache aligned, per CPU structure with lbolt usage statistics.
301  */
302 static lbolt_cpu_t *lb_cpu;
303 
304 /*
305  * Single, cache aligned, structure with all the information required by
306  * the lbolt implementation.
307  */
308 lbolt_info_t *lb_info;
309 
310 
311 int one_sec = 1; /* turned on once every second */
312 static int fsflushcnt;	/* counter for t_fsflushr */
313 int	dosynctodr = 1;	/* patchable; enable/disable sync to TOD chip */
314 int	tod_needsync = 0;	/* need to sync tod chip with software time */
315 static int tod_broken = 0;	/* clock chip doesn't work */
316 time_t	boot_time = 0;		/* Boot time in seconds since 1970 */
317 cyclic_id_t clock_cyclic;	/* clock()'s cyclic_id */
318 cyclic_id_t deadman_cyclic;	/* deadman()'s cyclic_id */
319 
320 extern void	clock_tick_schedule(int);
321 
322 static int lgrp_ticks;		/* counter to schedule lgrp load calcs */
323 
324 /*
325  * for tod fault detection
326  */
327 #define	TOD_REF_FREQ		((longlong_t)(NANOSEC))
328 #define	TOD_STALL_THRESHOLD	(TOD_REF_FREQ * 3 / 2)
329 #define	TOD_JUMP_THRESHOLD	(TOD_REF_FREQ / 2)
330 #define	TOD_FILTER_N		4
331 #define	TOD_FILTER_SETTLE	(4 * TOD_FILTER_N)
332 static enum tod_fault_type tod_faulted = TOD_NOFAULT;
333 
334 static int tod_status_flag = 0;		/* used by tod_validate() */
335 
336 static hrtime_t prev_set_tick = 0;	/* gethrtime() prior to tod_set() */
337 static time_t prev_set_tod = 0;		/* tv_sec value passed to tod_set() */
338 
339 /* patchable via /etc/system */
340 int tod_validate_enable = 1;
341 
342 /* Diagnose/Limit messages about delay(9F) called from interrupt context */
343 int			delay_from_interrupt_diagnose = 0;
344 volatile uint32_t	delay_from_interrupt_msg = 20;
345 
346 /*
347  * On non-SPARC systems, TOD validation must be deferred until gethrtime
348  * returns non-zero values (after mach_clkinit's execution).
349  * On SPARC systems, it must be deferred until after hrtime_base
350  * and hres_last_tick are set (in the first invocation of hres_tick).
351  * Since in both cases the prerequisites occur before the invocation of
352  * tod_get() in clock(), the deferment is lifted there.
353  */
354 static boolean_t tod_validate_deferred = B_TRUE;
355 
356 /*
357  * tod_fault_table[] must be aligned with
358  * enum tod_fault_type in systm.h
359  */
360 static char *tod_fault_table[] = {
361 	"Reversed",			/* TOD_REVERSED */
362 	"Stalled",			/* TOD_STALLED */
363 	"Jumped",			/* TOD_JUMPED */
364 	"Changed in Clock Rate",	/* TOD_RATECHANGED */
365 	"Is Read-Only"			/* TOD_RDONLY */
366 	/*
367 	 * no strings needed for TOD_NOFAULT
368 	 */
369 };
370 
371 /*
372  * test hook for tod broken detection in tod_validate
373  */
374 int tod_unit_test = 0;
375 time_t tod_test_injector;
376 
377 #define	CLOCK_ADJ_HIST_SIZE	4
378 
379 static int	adj_hist_entry;
380 
381 int64_t clock_adj_hist[CLOCK_ADJ_HIST_SIZE];
382 
383 static void calcloadavg(int, uint64_t *);
384 static int genloadavg(struct loadavg_s *);
385 static void loadavg_update();
386 
387 void (*cmm_clock_callout)() = NULL;
388 void (*cpucaps_clock_callout)() = NULL;
389 
390 extern clock_t clock_tick_proc_max;
391 
392 static int64_t deadman_counter = 0;
393 
394 static void
clock(void)395 clock(void)
396 {
397 	kthread_t	*t;
398 	uint_t	nrunnable;
399 	uint_t	w_io;
400 	cpu_t	*cp;
401 	cpupart_t *cpupart;
402 	extern	void	set_freemem();
403 	void	(*funcp)();
404 	int32_t ltemp;
405 	int64_t lltemp;
406 	int s;
407 	int do_lgrp_load;
408 	int i;
409 	clock_t now = LBOLT_NO_ACCOUNT;	/* current tick */
410 
411 	if (panicstr)
412 		return;
413 
414 	/*
415 	 * Make sure that 'freemem' do not drift too far from the truth
416 	 */
417 	set_freemem();
418 
419 
420 	/*
421 	 * Before the section which is repeated is executed, we do
422 	 * the time delta processing which occurs every clock tick
423 	 *
424 	 * There is additional processing which happens every time
425 	 * the nanosecond counter rolls over which is described
426 	 * below - see the section which begins with : if (one_sec)
427 	 *
428 	 * This section marks the beginning of the precision-kernel
429 	 * code fragment.
430 	 *
431 	 * First, compute the phase adjustment. If the low-order bits
432 	 * (time_phase) of the update overflow, bump the higher order
433 	 * bits (time_update).
434 	 */
435 	time_phase += time_adj;
436 	if (time_phase <= -FINEUSEC) {
437 		ltemp = -time_phase / SCALE_PHASE;
438 		time_phase += ltemp * SCALE_PHASE;
439 		s = hr_clock_lock();
440 		timedelta -= ltemp * (NANOSEC/MICROSEC);
441 		hr_clock_unlock(s);
442 	} else if (time_phase >= FINEUSEC) {
443 		ltemp = time_phase / SCALE_PHASE;
444 		time_phase -= ltemp * SCALE_PHASE;
445 		s = hr_clock_lock();
446 		timedelta += ltemp * (NANOSEC/MICROSEC);
447 		hr_clock_unlock(s);
448 	}
449 
450 	/*
451 	 * End of precision-kernel code fragment which is processed
452 	 * every timer interrupt.
453 	 *
454 	 * Continue with the interrupt processing as scheduled.
455 	 */
456 	/*
457 	 * Count the number of runnable threads and the number waiting
458 	 * for some form of I/O to complete -- gets added to
459 	 * sysinfo.waiting.  To know the state of the system, must add
460 	 * wait counts from all CPUs.  Also add up the per-partition
461 	 * statistics.
462 	 */
463 	w_io = 0;
464 	nrunnable = 0;
465 
466 	/*
467 	 * keep track of when to update lgrp/part loads
468 	 */
469 
470 	do_lgrp_load = 0;
471 	if (lgrp_ticks++ >= hz / 10) {
472 		lgrp_ticks = 0;
473 		do_lgrp_load = 1;
474 	}
475 
476 	if (one_sec) {
477 		loadavg_update();
478 		deadman_counter++;
479 	}
480 
481 	/*
482 	 * First count the threads waiting on kpreempt queues in each
483 	 * CPU partition.
484 	 */
485 
486 	cpupart = cp_list_head;
487 	do {
488 		uint_t cpupart_nrunnable = cpupart->cp_kp_queue.disp_nrunnable;
489 
490 		cpupart->cp_updates++;
491 		nrunnable += cpupart_nrunnable;
492 		cpupart->cp_nrunnable_cum += cpupart_nrunnable;
493 		if (one_sec) {
494 			cpupart->cp_nrunning = 0;
495 			cpupart->cp_nrunnable = cpupart_nrunnable;
496 		}
497 	} while ((cpupart = cpupart->cp_next) != cp_list_head);
498 
499 
500 	/* Now count the per-CPU statistics. */
501 	cp = cpu_list;
502 	do {
503 		uint_t cpu_nrunnable = cp->cpu_disp->disp_nrunnable;
504 
505 		nrunnable += cpu_nrunnable;
506 		cpupart = cp->cpu_part;
507 		cpupart->cp_nrunnable_cum += cpu_nrunnable;
508 		if (one_sec) {
509 			cpupart->cp_nrunnable += cpu_nrunnable;
510 			/*
511 			 * Update user, system, and idle cpu times.
512 			 */
513 			cpupart->cp_nrunning++;
514 			/*
515 			 * w_io is used to update sysinfo.waiting during
516 			 * one_second processing below.  Only gather w_io
517 			 * information when we walk the list of cpus if we're
518 			 * going to perform one_second processing.
519 			 */
520 			w_io += CPU_STATS(cp, sys.iowait);
521 		}
522 
523 		if (one_sec && (cp->cpu_flags & CPU_EXISTS)) {
524 			int i, load, change;
525 			hrtime_t intracct, intrused;
526 			const hrtime_t maxnsec = 1000000000;
527 			const int precision = 100;
528 
529 			/*
530 			 * Estimate interrupt load on this cpu each second.
531 			 * Computes cpu_intrload as %utilization (0-99).
532 			 */
533 
534 			/* add up interrupt time from all micro states */
535 			for (intracct = 0, i = 0; i < NCMSTATES; i++)
536 				intracct += cp->cpu_intracct[i];
537 			scalehrtime(&intracct);
538 
539 			/* compute nsec used in the past second */
540 			intrused = intracct - cp->cpu_intrlast;
541 			cp->cpu_intrlast = intracct;
542 
543 			/* limit the value for safety (and the first pass) */
544 			if (intrused >= maxnsec)
545 				intrused = maxnsec - 1;
546 
547 			/* calculate %time in interrupt */
548 			load = (precision * intrused) / maxnsec;
549 			ASSERT(load >= 0 && load < precision);
550 			change = cp->cpu_intrload - load;
551 
552 			/* jump to new max, or decay the old max */
553 			if (change < 0)
554 				cp->cpu_intrload = load;
555 			else if (change > 0)
556 				cp->cpu_intrload -= (change + 3) / 4;
557 
558 			DTRACE_PROBE3(cpu_intrload,
559 			    cpu_t *, cp,
560 			    hrtime_t, intracct,
561 			    hrtime_t, intrused);
562 		}
563 
564 		if (do_lgrp_load &&
565 		    (cp->cpu_flags & CPU_EXISTS)) {
566 			/*
567 			 * When updating the lgroup's load average,
568 			 * account for the thread running on the CPU.
569 			 * If the CPU is the current one, then we need
570 			 * to account for the underlying thread which
571 			 * got the clock interrupt not the thread that is
572 			 * handling the interrupt and caculating the load
573 			 * average
574 			 */
575 			t = cp->cpu_thread;
576 			if (CPU == cp)
577 				t = t->t_intr;
578 
579 			/*
580 			 * Account for the load average for this thread if
581 			 * it isn't the idle thread or it is on the interrupt
582 			 * stack and not the current CPU handling the clock
583 			 * interrupt
584 			 */
585 			if ((t && t != cp->cpu_idle_thread) || (CPU != cp &&
586 			    CPU_ON_INTR(cp))) {
587 				if (t->t_lpl == cp->cpu_lpl) {
588 					/* local thread */
589 					cpu_nrunnable++;
590 				} else {
591 					/*
592 					 * This is a remote thread, charge it
593 					 * against its home lgroup.  Note that
594 					 * we notice that a thread is remote
595 					 * only if it's currently executing.
596 					 * This is a reasonable approximation,
597 					 * since queued remote threads are rare.
598 					 * Note also that if we didn't charge
599 					 * it to its home lgroup, remote
600 					 * execution would often make a system
601 					 * appear balanced even though it was
602 					 * not, and thread placement/migration
603 					 * would often not be done correctly.
604 					 */
605 					lgrp_loadavg(t->t_lpl,
606 					    LGRP_LOADAVG_IN_THREAD_MAX, 0);
607 				}
608 			}
609 			lgrp_loadavg(cp->cpu_lpl,
610 			    cpu_nrunnable * LGRP_LOADAVG_IN_THREAD_MAX, 1);
611 		}
612 	} while ((cp = cp->cpu_next) != cpu_list);
613 
614 	clock_tick_schedule(one_sec);
615 
616 	/*
617 	 * Check for a callout that needs be called from the clock
618 	 * thread to support the membership protocol in a clustered
619 	 * system.  Copy the function pointer so that we can reset
620 	 * this to NULL if needed.
621 	 */
622 	if ((funcp = cmm_clock_callout) != NULL)
623 		(*funcp)();
624 
625 	if ((funcp = cpucaps_clock_callout) != NULL)
626 		(*funcp)();
627 
628 	/*
629 	 * Wakeup the cageout thread waiters once per second.
630 	 */
631 	if (one_sec)
632 		kcage_tick();
633 
634 	if (one_sec) {
635 
636 		int drift, absdrift;
637 		timestruc_t tod;
638 		int s;
639 
640 		/*
641 		 * Beginning of precision-kernel code fragment executed
642 		 * every second.
643 		 *
644 		 * On rollover of the second the phase adjustment to be
645 		 * used for the next second is calculated.  Also, the
646 		 * maximum error is increased by the tolerance.  If the
647 		 * PPS frequency discipline code is present, the phase is
648 		 * increased to compensate for the CPU clock oscillator
649 		 * frequency error.
650 		 *
651 		 * On a 32-bit machine and given parameters in the timex.h
652 		 * header file, the maximum phase adjustment is +-512 ms
653 		 * and maximum frequency offset is (a tad less than)
654 		 * +-512 ppm. On a 64-bit machine, you shouldn't need to ask.
655 		 */
656 		time_maxerror += time_tolerance / SCALE_USEC;
657 
658 		/*
659 		 * Leap second processing. If in leap-insert state at
660 		 * the end of the day, the system clock is set back one
661 		 * second; if in leap-delete state, the system clock is
662 		 * set ahead one second. The microtime() routine or
663 		 * external clock driver will insure that reported time
664 		 * is always monotonic. The ugly divides should be
665 		 * replaced.
666 		 */
667 		switch (time_state) {
668 
669 		case TIME_OK:
670 			if (time_status & STA_INS)
671 				time_state = TIME_INS;
672 			else if (time_status & STA_DEL)
673 				time_state = TIME_DEL;
674 			break;
675 
676 		case TIME_INS:
677 			if (hrestime.tv_sec % 86400 == 0) {
678 				s = hr_clock_lock();
679 				hrestime.tv_sec--;
680 				hr_clock_unlock(s);
681 				time_state = TIME_OOP;
682 			}
683 			break;
684 
685 		case TIME_DEL:
686 			if ((hrestime.tv_sec + 1) % 86400 == 0) {
687 				s = hr_clock_lock();
688 				hrestime.tv_sec++;
689 				hr_clock_unlock(s);
690 				time_state = TIME_WAIT;
691 			}
692 			break;
693 
694 		case TIME_OOP:
695 			time_state = TIME_WAIT;
696 			break;
697 
698 		case TIME_WAIT:
699 			if (!(time_status & (STA_INS | STA_DEL)))
700 				time_state = TIME_OK;
701 		default:
702 			break;
703 		}
704 
705 		/*
706 		 * Compute the phase adjustment for the next second. In
707 		 * PLL mode, the offset is reduced by a fixed factor
708 		 * times the time constant. In FLL mode the offset is
709 		 * used directly. In either mode, the maximum phase
710 		 * adjustment for each second is clamped so as to spread
711 		 * the adjustment over not more than the number of
712 		 * seconds between updates.
713 		 */
714 		if (time_offset == 0)
715 			time_adj = 0;
716 		else if (time_offset < 0) {
717 			lltemp = -time_offset;
718 			if (!(time_status & STA_FLL)) {
719 				if ((1 << time_constant) >= SCALE_KG)
720 					lltemp *= (1 << time_constant) /
721 					    SCALE_KG;
722 				else
723 					lltemp = (lltemp / SCALE_KG) >>
724 					    time_constant;
725 			}
726 			if (lltemp > (MAXPHASE / MINSEC) * SCALE_UPDATE)
727 				lltemp = (MAXPHASE / MINSEC) * SCALE_UPDATE;
728 			time_offset += lltemp;
729 			time_adj = -(lltemp * SCALE_PHASE) / hz / SCALE_UPDATE;
730 		} else {
731 			lltemp = time_offset;
732 			if (!(time_status & STA_FLL)) {
733 				if ((1 << time_constant) >= SCALE_KG)
734 					lltemp *= (1 << time_constant) /
735 					    SCALE_KG;
736 				else
737 					lltemp = (lltemp / SCALE_KG) >>
738 					    time_constant;
739 			}
740 			if (lltemp > (MAXPHASE / MINSEC) * SCALE_UPDATE)
741 				lltemp = (MAXPHASE / MINSEC) * SCALE_UPDATE;
742 			time_offset -= lltemp;
743 			time_adj = (lltemp * SCALE_PHASE) / hz / SCALE_UPDATE;
744 		}
745 
746 		/*
747 		 * Compute the frequency estimate and additional phase
748 		 * adjustment due to frequency error for the next
749 		 * second. When the PPS signal is engaged, gnaw on the
750 		 * watchdog counter and update the frequency computed by
751 		 * the pll and the PPS signal.
752 		 */
753 		pps_valid++;
754 		if (pps_valid == PPS_VALID) {
755 			pps_jitter = MAXTIME;
756 			pps_stabil = MAXFREQ;
757 			time_status &= ~(STA_PPSSIGNAL | STA_PPSJITTER |
758 			    STA_PPSWANDER | STA_PPSERROR);
759 		}
760 		lltemp = time_freq + pps_freq;
761 
762 		if (lltemp)
763 			time_adj += (lltemp * SCALE_PHASE) / (SCALE_USEC * hz);
764 
765 		/*
766 		 * End of precision kernel-code fragment
767 		 *
768 		 * The section below should be modified if we are planning
769 		 * to use NTP for synchronization.
770 		 *
771 		 * Note: the clock synchronization code now assumes
772 		 * the following:
773 		 *   - if dosynctodr is 1, then compute the drift between
774 		 *	the tod chip and software time and adjust one or
775 		 *	the other depending on the circumstances
776 		 *
777 		 *   - if dosynctodr is 0, then the tod chip is independent
778 		 *	of the software clock and should not be adjusted,
779 		 *	but allowed to free run.  this allows NTP to sync.
780 		 *	hrestime without any interference from the tod chip.
781 		 */
782 
783 		tod_validate_deferred = B_FALSE;
784 		mutex_enter(&tod_lock);
785 		tod = tod_get();
786 		drift = tod.tv_sec - hrestime.tv_sec;
787 		absdrift = (drift >= 0) ? drift : -drift;
788 		if (tod_needsync || absdrift > 1) {
789 			int s;
790 			if (absdrift > 2) {
791 				if (!tod_broken && tod_faulted == TOD_NOFAULT) {
792 					s = hr_clock_lock();
793 					hrestime = tod;
794 					membar_enter();	/* hrestime visible */
795 					timedelta = 0;
796 					timechanged++;
797 					tod_needsync = 0;
798 					hr_clock_unlock(s);
799 					callout_hrestime();
800 
801 				}
802 			} else {
803 				if (tod_needsync || !dosynctodr) {
804 					gethrestime(&tod);
805 					tod_set(tod);
806 					s = hr_clock_lock();
807 					if (timedelta == 0)
808 						tod_needsync = 0;
809 					hr_clock_unlock(s);
810 				} else {
811 					/*
812 					 * If the drift is 2 seconds on the
813 					 * money, then the TOD is adjusting
814 					 * the clock;  record that.
815 					 */
816 					clock_adj_hist[adj_hist_entry++ %
817 					    CLOCK_ADJ_HIST_SIZE] = now;
818 					s = hr_clock_lock();
819 					timedelta = (int64_t)drift*NANOSEC;
820 					hr_clock_unlock(s);
821 				}
822 			}
823 		}
824 		one_sec = 0;
825 		time = gethrestime_sec();  /* for crusty old kmem readers */
826 		mutex_exit(&tod_lock);
827 
828 		/*
829 		 * Some drivers still depend on this... XXX
830 		 */
831 		cv_broadcast(&lbolt_cv);
832 
833 		vminfo.freemem += freemem;
834 		{
835 			pgcnt_t maxswap, resv, free;
836 			pgcnt_t avail =
837 			    MAX((spgcnt_t)(availrmem - swapfs_minfree), 0);
838 
839 			maxswap = k_anoninfo.ani_mem_resv +
840 			    k_anoninfo.ani_max +avail;
841 			/* Update ani_free */
842 			set_anoninfo();
843 			free = k_anoninfo.ani_free + avail;
844 			resv = k_anoninfo.ani_phys_resv +
845 			    k_anoninfo.ani_mem_resv;
846 
847 			vminfo.swap_resv += resv;
848 			/* number of reserved and allocated pages */
849 #ifdef	DEBUG
850 			if (maxswap < free)
851 				cmn_err(CE_WARN, "clock: maxswap < free");
852 			if (maxswap < resv)
853 				cmn_err(CE_WARN, "clock: maxswap < resv");
854 #endif
855 			vminfo.swap_alloc += maxswap - free;
856 			vminfo.swap_avail += maxswap - resv;
857 			vminfo.swap_free += free;
858 		}
859 		vminfo.updates++;
860 		if (nrunnable) {
861 			sysinfo.runque += nrunnable;
862 			sysinfo.runocc++;
863 		}
864 		if (nswapped) {
865 			sysinfo.swpque += nswapped;
866 			sysinfo.swpocc++;
867 		}
868 		sysinfo.waiting += w_io;
869 		sysinfo.updates++;
870 
871 		/*
872 		 * Wake up fsflush to write out DELWRI
873 		 * buffers, dirty pages and other cached
874 		 * administrative data, e.g. inodes.
875 		 */
876 		if (--fsflushcnt <= 0) {
877 			fsflushcnt = tune.t_fsflushr;
878 			cv_signal(&fsflush_cv);
879 		}
880 
881 		vmmeter();
882 		calcloadavg(genloadavg(&loadavg), hp_avenrun);
883 		for (i = 0; i < 3; i++)
884 			/*
885 			 * At the moment avenrun[] can only hold 31
886 			 * bits of load average as it is a signed
887 			 * int in the API. We need to ensure that
888 			 * hp_avenrun[i] >> (16 - FSHIFT) will not be
889 			 * too large. If it is, we put the largest value
890 			 * that we can use into avenrun[i]. This is
891 			 * kludgey, but about all we can do until we
892 			 * avenrun[] is declared as an array of uint64[]
893 			 */
894 			if (hp_avenrun[i] < ((uint64_t)1<<(31+16-FSHIFT)))
895 				avenrun[i] = (int32_t)(hp_avenrun[i] >>
896 				    (16 - FSHIFT));
897 			else
898 				avenrun[i] = 0x7fffffff;
899 
900 		cpupart = cp_list_head;
901 		do {
902 			calcloadavg(genloadavg(&cpupart->cp_loadavg),
903 			    cpupart->cp_hp_avenrun);
904 		} while ((cpupart = cpupart->cp_next) != cp_list_head);
905 
906 		/*
907 		 * Wake up the swapper thread if necessary.
908 		 */
909 		if (runin ||
910 		    (runout && (avefree < desfree || wake_sched_sec))) {
911 			t = &t0;
912 			thread_lock(t);
913 			if (t->t_state == TS_STOPPED) {
914 				runin = runout = 0;
915 				wake_sched_sec = 0;
916 				t->t_whystop = 0;
917 				t->t_whatstop = 0;
918 				t->t_schedflag &= ~TS_ALLSTART;
919 				THREAD_TRANSITION(t);
920 				setfrontdq(t);
921 			}
922 			thread_unlock(t);
923 		}
924 	}
925 
926 	/*
927 	 * Wake up the swapper if any high priority swapped-out threads
928 	 * became runable during the last tick.
929 	 */
930 	if (wake_sched) {
931 		t = &t0;
932 		thread_lock(t);
933 		if (t->t_state == TS_STOPPED) {
934 			runin = runout = 0;
935 			wake_sched = 0;
936 			t->t_whystop = 0;
937 			t->t_whatstop = 0;
938 			t->t_schedflag &= ~TS_ALLSTART;
939 			THREAD_TRANSITION(t);
940 			setfrontdq(t);
941 		}
942 		thread_unlock(t);
943 	}
944 }
945 
946 void
clock_init(void)947 clock_init(void)
948 {
949 	cyc_handler_t clk_hdlr, lbolt_hdlr;
950 	cyc_time_t clk_when, lbolt_when;
951 	int i, sz;
952 	intptr_t buf;
953 
954 	/*
955 	 * Setup handler and timer for the clock cyclic.
956 	 */
957 	clk_hdlr.cyh_func = (cyc_func_t)clock;
958 	clk_hdlr.cyh_level = CY_LOCK_LEVEL;
959 	clk_hdlr.cyh_arg = NULL;
960 
961 	clk_when.cyt_when = 0;
962 	clk_when.cyt_interval = nsec_per_tick;
963 
964 	/*
965 	 * The lbolt cyclic will be reprogramed to fire at a nsec_per_tick
966 	 * interval to satisfy performance needs of the DDI lbolt consumers.
967 	 * It is off by default.
968 	 */
969 	lbolt_hdlr.cyh_func = (cyc_func_t)lbolt_cyclic;
970 	lbolt_hdlr.cyh_level = CY_LOCK_LEVEL;
971 	lbolt_hdlr.cyh_arg = NULL;
972 
973 	lbolt_when.cyt_interval = nsec_per_tick;
974 
975 	/*
976 	 * Allocate cache line aligned space for the per CPU lbolt data and
977 	 * lbolt info structures, and initialize them with their default
978 	 * values. Note that these structures are also cache line sized.
979 	 */
980 	sz = sizeof (lbolt_info_t) + CPU_CACHE_COHERENCE_SIZE;
981 	buf = (intptr_t)kmem_zalloc(sz, KM_SLEEP);
982 	lb_info = (lbolt_info_t *)P2ROUNDUP(buf, CPU_CACHE_COHERENCE_SIZE);
983 
984 	if (hz != HZ_DEFAULT)
985 		lb_info->lbi_thresh_interval = LBOLT_THRESH_INTERVAL *
986 		    hz/HZ_DEFAULT;
987 	else
988 		lb_info->lbi_thresh_interval = LBOLT_THRESH_INTERVAL;
989 
990 	lb_info->lbi_thresh_calls = LBOLT_THRESH_CALLS;
991 
992 	sz = (sizeof (lbolt_cpu_t) * max_ncpus) + CPU_CACHE_COHERENCE_SIZE;
993 	buf = (intptr_t)kmem_zalloc(sz, KM_SLEEP);
994 	lb_cpu = (lbolt_cpu_t *)P2ROUNDUP(buf, CPU_CACHE_COHERENCE_SIZE);
995 
996 	for (i = 0; i < max_ncpus; i++)
997 		lb_cpu[i].lbc_counter = lb_info->lbi_thresh_calls;
998 
999 	/*
1000 	 * Install the softint used to switch between event and cyclic driven
1001 	 * lbolt. We use a soft interrupt to make sure the context of the
1002 	 * cyclic reprogram call is safe.
1003 	 */
1004 	lbolt_softint_add();
1005 
1006 	/*
1007 	 * Since the hybrid lbolt implementation is based on a hardware counter
1008 	 * that is reset at every hardware reboot and that we'd like to have
1009 	 * the lbolt value starting at zero after both a hardware and a fast
1010 	 * reboot, we calculate the number of clock ticks the system's been up
1011 	 * and store it in the lbi_debug_time field of the lbolt info structure.
1012 	 * The value of this field will be subtracted from lbolt before
1013 	 * returning it.
1014 	 */
1015 	lb_info->lbi_internal = lb_info->lbi_debug_time =
1016 	    (gethrtime()/nsec_per_tick);
1017 
1018 	/*
1019 	 * lbolt_hybrid points at lbolt_bootstrap until now. The LBOLT_* macros
1020 	 * and lbolt_debug_{enter,return} use this value as an indication that
1021 	 * the initializaion above hasn't been completed. Setting lbolt_hybrid
1022 	 * to either lbolt_{cyclic,event}_driven here signals those code paths
1023 	 * that the lbolt related structures can be used.
1024 	 */
1025 	if (lbolt_cyc_only) {
1026 		lbolt_when.cyt_when = 0;
1027 		lbolt_hybrid = lbolt_cyclic_driven;
1028 	} else {
1029 		lbolt_when.cyt_when = CY_INFINITY;
1030 		lbolt_hybrid = lbolt_event_driven;
1031 	}
1032 
1033 	/*
1034 	 * Grab cpu_lock and install all three cyclics.
1035 	 */
1036 	mutex_enter(&cpu_lock);
1037 
1038 	clock_cyclic = cyclic_add(&clk_hdlr, &clk_when);
1039 	lb_info->id.lbi_cyclic_id = cyclic_add(&lbolt_hdlr, &lbolt_when);
1040 
1041 	mutex_exit(&cpu_lock);
1042 }
1043 
1044 /*
1045  * Called before calcloadavg to get 10-sec moving loadavg together
1046  */
1047 
1048 static int
genloadavg(struct loadavg_s * avgs)1049 genloadavg(struct loadavg_s *avgs)
1050 {
1051 	int avg;
1052 	int spos; /* starting position */
1053 	int cpos; /* moving current position */
1054 	int i;
1055 	int slen;
1056 	hrtime_t hr_avg;
1057 
1058 	/* 10-second snapshot, calculate first positon */
1059 	if (avgs->lg_len == 0) {
1060 		return (0);
1061 	}
1062 	slen = avgs->lg_len < S_MOVAVG_SZ ? avgs->lg_len : S_MOVAVG_SZ;
1063 
1064 	spos = (avgs->lg_cur - 1) >= 0 ? avgs->lg_cur - 1 :
1065 	    S_LOADAVG_SZ + (avgs->lg_cur - 1);
1066 	for (i = hr_avg = 0; i < slen; i++) {
1067 		cpos = (spos - i) >= 0 ? spos - i : S_LOADAVG_SZ + (spos - i);
1068 		hr_avg += avgs->lg_loads[cpos];
1069 	}
1070 
1071 	hr_avg = hr_avg / slen;
1072 	avg = hr_avg / (NANOSEC / LGRP_LOADAVG_IN_THREAD_MAX);
1073 
1074 	return (avg);
1075 }
1076 
1077 /*
1078  * Run every second from clock () to update the loadavg count available to the
1079  * system and cpu-partitions.
1080  *
1081  * This works by sampling the previous usr, sys, wait time elapsed,
1082  * computing a delta, and adding that delta to the elapsed usr, sys,
1083  * wait increase.
1084  */
1085 
1086 static void
loadavg_update()1087 loadavg_update()
1088 {
1089 	cpu_t *cp;
1090 	cpupart_t *cpupart;
1091 	hrtime_t cpu_total;
1092 	int prev;
1093 
1094 	cp = cpu_list;
1095 	loadavg.lg_total = 0;
1096 
1097 	/*
1098 	 * first pass totals up per-cpu statistics for system and cpu
1099 	 * partitions
1100 	 */
1101 
1102 	do {
1103 		struct loadavg_s *lavg;
1104 
1105 		lavg = &cp->cpu_loadavg;
1106 
1107 		cpu_total = cp->cpu_acct[CMS_USER] +
1108 		    cp->cpu_acct[CMS_SYSTEM] + cp->cpu_waitrq;
1109 		/* compute delta against last total */
1110 		scalehrtime(&cpu_total);
1111 		prev = (lavg->lg_cur - 1) >= 0 ? lavg->lg_cur - 1 :
1112 		    S_LOADAVG_SZ + (lavg->lg_cur - 1);
1113 		if (lavg->lg_loads[prev] <= 0) {
1114 			lavg->lg_loads[lavg->lg_cur] = cpu_total;
1115 			cpu_total = 0;
1116 		} else {
1117 			lavg->lg_loads[lavg->lg_cur] = cpu_total;
1118 			cpu_total = cpu_total - lavg->lg_loads[prev];
1119 			if (cpu_total < 0)
1120 				cpu_total = 0;
1121 		}
1122 
1123 		lavg->lg_cur = (lavg->lg_cur + 1) % S_LOADAVG_SZ;
1124 		lavg->lg_len = (lavg->lg_len + 1) < S_LOADAVG_SZ ?
1125 		    lavg->lg_len + 1 : S_LOADAVG_SZ;
1126 
1127 		loadavg.lg_total += cpu_total;
1128 		cp->cpu_part->cp_loadavg.lg_total += cpu_total;
1129 
1130 	} while ((cp = cp->cpu_next) != cpu_list);
1131 
1132 	loadavg.lg_loads[loadavg.lg_cur] = loadavg.lg_total;
1133 	loadavg.lg_cur = (loadavg.lg_cur + 1) % S_LOADAVG_SZ;
1134 	loadavg.lg_len = (loadavg.lg_len + 1) < S_LOADAVG_SZ ?
1135 	    loadavg.lg_len + 1 : S_LOADAVG_SZ;
1136 	/*
1137 	 * Second pass updates counts
1138 	 */
1139 	cpupart = cp_list_head;
1140 
1141 	do {
1142 		struct loadavg_s *lavg;
1143 
1144 		lavg = &cpupart->cp_loadavg;
1145 		lavg->lg_loads[lavg->lg_cur] = lavg->lg_total;
1146 		lavg->lg_total = 0;
1147 		lavg->lg_cur = (lavg->lg_cur + 1) % S_LOADAVG_SZ;
1148 		lavg->lg_len = (lavg->lg_len + 1) < S_LOADAVG_SZ ?
1149 		    lavg->lg_len + 1 : S_LOADAVG_SZ;
1150 
1151 	} while ((cpupart = cpupart->cp_next) != cp_list_head);
1152 
1153 	/*
1154 	 * Third pass totals up per-zone statistics.
1155 	 */
1156 	zone_loadavg_update();
1157 }
1158 
1159 /*
1160  * clock_update() - local clock update
1161  *
1162  * This routine is called by ntp_adjtime() to update the local clock
1163  * phase and frequency. The implementation is of an
1164  * adaptive-parameter, hybrid phase/frequency-lock loop (PLL/FLL). The
1165  * routine computes new time and frequency offset estimates for each
1166  * call.  The PPS signal itself determines the new time offset,
1167  * instead of the calling argument.  Presumably, calls to
1168  * ntp_adjtime() occur only when the caller believes the local clock
1169  * is valid within some bound (+-128 ms with NTP). If the caller's
1170  * time is far different than the PPS time, an argument will ensue,
1171  * and it's not clear who will lose.
1172  *
1173  * For uncompensated quartz crystal oscillatores and nominal update
1174  * intervals less than 1024 s, operation should be in phase-lock mode
1175  * (STA_FLL = 0), where the loop is disciplined to phase. For update
1176  * intervals greater than this, operation should be in frequency-lock
1177  * mode (STA_FLL = 1), where the loop is disciplined to frequency.
1178  *
1179  * Note: mutex(&tod_lock) is in effect.
1180  */
1181 void
clock_update(int offset)1182 clock_update(int offset)
1183 {
1184 	int ltemp, mtemp, s;
1185 
1186 	ASSERT(MUTEX_HELD(&tod_lock));
1187 
1188 	if (!(time_status & STA_PLL) && !(time_status & STA_PPSTIME))
1189 		return;
1190 	ltemp = offset;
1191 	if ((time_status & STA_PPSTIME) && (time_status & STA_PPSSIGNAL))
1192 		ltemp = pps_offset;
1193 
1194 	/*
1195 	 * Scale the phase adjustment and clamp to the operating range.
1196 	 */
1197 	if (ltemp > MAXPHASE)
1198 		time_offset = MAXPHASE * SCALE_UPDATE;
1199 	else if (ltemp < -MAXPHASE)
1200 		time_offset = -(MAXPHASE * SCALE_UPDATE);
1201 	else
1202 		time_offset = ltemp * SCALE_UPDATE;
1203 
1204 	/*
1205 	 * Select whether the frequency is to be controlled and in which
1206 	 * mode (PLL or FLL). Clamp to the operating range. Ugly
1207 	 * multiply/divide should be replaced someday.
1208 	 */
1209 	if (time_status & STA_FREQHOLD || time_reftime == 0)
1210 		time_reftime = hrestime.tv_sec;
1211 
1212 	mtemp = hrestime.tv_sec - time_reftime;
1213 	time_reftime = hrestime.tv_sec;
1214 
1215 	if (time_status & STA_FLL) {
1216 		if (mtemp >= MINSEC) {
1217 			ltemp = ((time_offset / mtemp) * (SCALE_USEC /
1218 			    SCALE_UPDATE));
1219 			if (ltemp)
1220 				time_freq += ltemp / SCALE_KH;
1221 		}
1222 	} else {
1223 		if (mtemp < MAXSEC) {
1224 			ltemp *= mtemp;
1225 			if (ltemp)
1226 				time_freq += (int)(((int64_t)ltemp *
1227 				    SCALE_USEC) / SCALE_KF)
1228 				    / (1 << (time_constant * 2));
1229 		}
1230 	}
1231 	if (time_freq > time_tolerance)
1232 		time_freq = time_tolerance;
1233 	else if (time_freq < -time_tolerance)
1234 		time_freq = -time_tolerance;
1235 
1236 	s = hr_clock_lock();
1237 	tod_needsync = 1;
1238 	hr_clock_unlock(s);
1239 }
1240 
1241 /*
1242  * ddi_hardpps() - discipline CPU clock oscillator to external PPS signal
1243  *
1244  * This routine is called at each PPS interrupt in order to discipline
1245  * the CPU clock oscillator to the PPS signal. It measures the PPS phase
1246  * and leaves it in a handy spot for the clock() routine. It
1247  * integrates successive PPS phase differences and calculates the
1248  * frequency offset. This is used in clock() to discipline the CPU
1249  * clock oscillator so that intrinsic frequency error is cancelled out.
1250  * The code requires the caller to capture the time and hardware counter
1251  * value at the on-time PPS signal transition.
1252  *
1253  * Note that, on some Unix systems, this routine runs at an interrupt
1254  * priority level higher than the timer interrupt routine clock().
1255  * Therefore, the variables used are distinct from the clock()
1256  * variables, except for certain exceptions: The PPS frequency pps_freq
1257  * and phase pps_offset variables are determined by this routine and
1258  * updated atomically. The time_tolerance variable can be considered a
1259  * constant, since it is infrequently changed, and then only when the
1260  * PPS signal is disabled. The watchdog counter pps_valid is updated
1261  * once per second by clock() and is atomically cleared in this
1262  * routine.
1263  *
1264  * tvp is the time of the last tick; usec is a microsecond count since the
1265  * last tick.
1266  *
1267  * Note: In Solaris systems, the tick value is actually given by
1268  *       usec_per_tick.  This is called from the serial driver cdintr(),
1269  *	 or equivalent, at a high PIL.  Because the kernel keeps a
1270  *	 highresolution time, the following code can accept either
1271  *	 the traditional argument pair, or the current highres timestamp
1272  *       in tvp and zero in usec.
1273  */
1274 void
ddi_hardpps(struct timeval * tvp,int usec)1275 ddi_hardpps(struct timeval *tvp, int usec)
1276 {
1277 	int u_usec, v_usec, bigtick;
1278 	time_t cal_sec;
1279 	int cal_usec;
1280 
1281 	/*
1282 	 * An occasional glitch can be produced when the PPS interrupt
1283 	 * occurs in the clock() routine before the time variable is
1284 	 * updated. Here the offset is discarded when the difference
1285 	 * between it and the last one is greater than tick/2, but not
1286 	 * if the interval since the first discard exceeds 30 s.
1287 	 */
1288 	time_status |= STA_PPSSIGNAL;
1289 	time_status &= ~(STA_PPSJITTER | STA_PPSWANDER | STA_PPSERROR);
1290 	pps_valid = 0;
1291 	u_usec = -tvp->tv_usec;
1292 	if (u_usec < -(MICROSEC/2))
1293 		u_usec += MICROSEC;
1294 	v_usec = pps_offset - u_usec;
1295 	if (v_usec < 0)
1296 		v_usec = -v_usec;
1297 	if (v_usec > (usec_per_tick >> 1)) {
1298 		if (pps_glitch > MAXGLITCH) {
1299 			pps_glitch = 0;
1300 			pps_tf[2] = u_usec;
1301 			pps_tf[1] = u_usec;
1302 		} else {
1303 			pps_glitch++;
1304 			u_usec = pps_offset;
1305 		}
1306 	} else
1307 		pps_glitch = 0;
1308 
1309 	/*
1310 	 * A three-stage median filter is used to help deglitch the pps
1311 	 * time. The median sample becomes the time offset estimate; the
1312 	 * difference between the other two samples becomes the time
1313 	 * dispersion (jitter) estimate.
1314 	 */
1315 	pps_tf[2] = pps_tf[1];
1316 	pps_tf[1] = pps_tf[0];
1317 	pps_tf[0] = u_usec;
1318 	if (pps_tf[0] > pps_tf[1]) {
1319 		if (pps_tf[1] > pps_tf[2]) {
1320 			pps_offset = pps_tf[1];		/* 0 1 2 */
1321 			v_usec = pps_tf[0] - pps_tf[2];
1322 		} else if (pps_tf[2] > pps_tf[0]) {
1323 			pps_offset = pps_tf[0];		/* 2 0 1 */
1324 			v_usec = pps_tf[2] - pps_tf[1];
1325 		} else {
1326 			pps_offset = pps_tf[2];		/* 0 2 1 */
1327 			v_usec = pps_tf[0] - pps_tf[1];
1328 		}
1329 	} else {
1330 		if (pps_tf[1] < pps_tf[2]) {
1331 			pps_offset = pps_tf[1];		/* 2 1 0 */
1332 			v_usec = pps_tf[2] - pps_tf[0];
1333 		} else  if (pps_tf[2] < pps_tf[0]) {
1334 			pps_offset = pps_tf[0];		/* 1 0 2 */
1335 			v_usec = pps_tf[1] - pps_tf[2];
1336 		} else {
1337 			pps_offset = pps_tf[2];		/* 1 2 0 */
1338 			v_usec = pps_tf[1] - pps_tf[0];
1339 		}
1340 	}
1341 	if (v_usec > MAXTIME)
1342 		pps_jitcnt++;
1343 	v_usec = (v_usec << PPS_AVG) - pps_jitter;
1344 	pps_jitter += v_usec / (1 << PPS_AVG);
1345 	if (pps_jitter > (MAXTIME >> 1))
1346 		time_status |= STA_PPSJITTER;
1347 
1348 	/*
1349 	 * During the calibration interval adjust the starting time when
1350 	 * the tick overflows. At the end of the interval compute the
1351 	 * duration of the interval and the difference of the hardware
1352 	 * counters at the beginning and end of the interval. This code
1353 	 * is deliciously complicated by the fact valid differences may
1354 	 * exceed the value of tick when using long calibration
1355 	 * intervals and small ticks. Note that the counter can be
1356 	 * greater than tick if caught at just the wrong instant, but
1357 	 * the values returned and used here are correct.
1358 	 */
1359 	bigtick = (int)usec_per_tick * SCALE_USEC;
1360 	pps_usec -= pps_freq;
1361 	if (pps_usec >= bigtick)
1362 		pps_usec -= bigtick;
1363 	if (pps_usec < 0)
1364 		pps_usec += bigtick;
1365 	pps_time.tv_sec++;
1366 	pps_count++;
1367 	if (pps_count < (1 << pps_shift))
1368 		return;
1369 	pps_count = 0;
1370 	pps_calcnt++;
1371 	u_usec = usec * SCALE_USEC;
1372 	v_usec = pps_usec - u_usec;
1373 	if (v_usec >= bigtick >> 1)
1374 		v_usec -= bigtick;
1375 	if (v_usec < -(bigtick >> 1))
1376 		v_usec += bigtick;
1377 	if (v_usec < 0)
1378 		v_usec = -(-v_usec >> pps_shift);
1379 	else
1380 		v_usec = v_usec >> pps_shift;
1381 	pps_usec = u_usec;
1382 	cal_sec = tvp->tv_sec;
1383 	cal_usec = tvp->tv_usec;
1384 	cal_sec -= pps_time.tv_sec;
1385 	cal_usec -= pps_time.tv_usec;
1386 	if (cal_usec < 0) {
1387 		cal_usec += MICROSEC;
1388 		cal_sec--;
1389 	}
1390 	pps_time = *tvp;
1391 
1392 	/*
1393 	 * Check for lost interrupts, noise, excessive jitter and
1394 	 * excessive frequency error. The number of timer ticks during
1395 	 * the interval may vary +-1 tick. Add to this a margin of one
1396 	 * tick for the PPS signal jitter and maximum frequency
1397 	 * deviation. If the limits are exceeded, the calibration
1398 	 * interval is reset to the minimum and we start over.
1399 	 */
1400 	u_usec = (int)usec_per_tick << 1;
1401 	if (!((cal_sec == -1 && cal_usec > (MICROSEC - u_usec)) ||
1402 	    (cal_sec == 0 && cal_usec < u_usec)) ||
1403 	    v_usec > time_tolerance || v_usec < -time_tolerance) {
1404 		pps_errcnt++;
1405 		pps_shift = PPS_SHIFT;
1406 		pps_intcnt = 0;
1407 		time_status |= STA_PPSERROR;
1408 		return;
1409 	}
1410 
1411 	/*
1412 	 * A three-stage median filter is used to help deglitch the pps
1413 	 * frequency. The median sample becomes the frequency offset
1414 	 * estimate; the difference between the other two samples
1415 	 * becomes the frequency dispersion (stability) estimate.
1416 	 */
1417 	pps_ff[2] = pps_ff[1];
1418 	pps_ff[1] = pps_ff[0];
1419 	pps_ff[0] = v_usec;
1420 	if (pps_ff[0] > pps_ff[1]) {
1421 		if (pps_ff[1] > pps_ff[2]) {
1422 			u_usec = pps_ff[1];		/* 0 1 2 */
1423 			v_usec = pps_ff[0] - pps_ff[2];
1424 		} else if (pps_ff[2] > pps_ff[0]) {
1425 			u_usec = pps_ff[0];		/* 2 0 1 */
1426 			v_usec = pps_ff[2] - pps_ff[1];
1427 		} else {
1428 			u_usec = pps_ff[2];		/* 0 2 1 */
1429 			v_usec = pps_ff[0] - pps_ff[1];
1430 		}
1431 	} else {
1432 		if (pps_ff[1] < pps_ff[2]) {
1433 			u_usec = pps_ff[1];		/* 2 1 0 */
1434 			v_usec = pps_ff[2] - pps_ff[0];
1435 		} else  if (pps_ff[2] < pps_ff[0]) {
1436 			u_usec = pps_ff[0];		/* 1 0 2 */
1437 			v_usec = pps_ff[1] - pps_ff[2];
1438 		} else {
1439 			u_usec = pps_ff[2];		/* 1 2 0 */
1440 			v_usec = pps_ff[1] - pps_ff[0];
1441 		}
1442 	}
1443 
1444 	/*
1445 	 * Here the frequency dispersion (stability) is updated. If it
1446 	 * is less than one-fourth the maximum (MAXFREQ), the frequency
1447 	 * offset is updated as well, but clamped to the tolerance. It
1448 	 * will be processed later by the clock() routine.
1449 	 */
1450 	v_usec = (v_usec >> 1) - pps_stabil;
1451 	if (v_usec < 0)
1452 		pps_stabil -= -v_usec >> PPS_AVG;
1453 	else
1454 		pps_stabil += v_usec >> PPS_AVG;
1455 	if (pps_stabil > MAXFREQ >> 2) {
1456 		pps_stbcnt++;
1457 		time_status |= STA_PPSWANDER;
1458 		return;
1459 	}
1460 	if (time_status & STA_PPSFREQ) {
1461 		if (u_usec < 0) {
1462 			pps_freq -= -u_usec >> PPS_AVG;
1463 			if (pps_freq < -time_tolerance)
1464 				pps_freq = -time_tolerance;
1465 			u_usec = -u_usec;
1466 		} else {
1467 			pps_freq += u_usec >> PPS_AVG;
1468 			if (pps_freq > time_tolerance)
1469 				pps_freq = time_tolerance;
1470 		}
1471 	}
1472 
1473 	/*
1474 	 * Here the calibration interval is adjusted. If the maximum
1475 	 * time difference is greater than tick / 4, reduce the interval
1476 	 * by half. If this is not the case for four consecutive
1477 	 * intervals, double the interval.
1478 	 */
1479 	if (u_usec << pps_shift > bigtick >> 2) {
1480 		pps_intcnt = 0;
1481 		if (pps_shift > PPS_SHIFT)
1482 			pps_shift--;
1483 	} else if (pps_intcnt >= 4) {
1484 		pps_intcnt = 0;
1485 		if (pps_shift < PPS_SHIFTMAX)
1486 			pps_shift++;
1487 	} else
1488 		pps_intcnt++;
1489 
1490 	/*
1491 	 * If recovering from kmdb, then make sure the tod chip gets resynced.
1492 	 * If we took an early exit above, then we don't yet have a stable
1493 	 * calibration signal to lock onto, so don't mark the tod for sync
1494 	 * until we get all the way here.
1495 	 */
1496 	{
1497 		int s = hr_clock_lock();
1498 
1499 		tod_needsync = 1;
1500 		hr_clock_unlock(s);
1501 	}
1502 }
1503 
1504 /*
1505  * Handle clock tick processing for a thread.
1506  * Check for timer action, enforce CPU rlimit, do profiling etc.
1507  */
1508 void
clock_tick(kthread_t * t,int pending)1509 clock_tick(kthread_t *t, int pending)
1510 {
1511 	struct proc *pp;
1512 	klwp_id_t    lwp;
1513 	struct as *as;
1514 	clock_t	ticks;
1515 	int	poke = 0;		/* notify another CPU */
1516 	int	user_mode;
1517 	size_t	 rss;
1518 	int i, total_usec, usec;
1519 	rctl_qty_t secs;
1520 
1521 	ASSERT(pending > 0);
1522 
1523 	/* Must be operating on a lwp/thread */
1524 	if ((lwp = ttolwp(t)) == NULL) {
1525 		panic("clock_tick: no lwp");
1526 		/*NOTREACHED*/
1527 	}
1528 
1529 	for (i = 0; i < pending; i++) {
1530 		CL_TICK(t);	/* Class specific tick processing */
1531 		DTRACE_SCHED1(tick, kthread_t *, t);
1532 	}
1533 
1534 	pp = ttoproc(t);
1535 
1536 	/* pp->p_lock makes sure that the thread does not exit */
1537 	ASSERT(MUTEX_HELD(&pp->p_lock));
1538 
1539 	user_mode = (lwp->lwp_state == LWP_USER);
1540 
1541 	ticks = (pp->p_utime + pp->p_stime) % hz;
1542 	/*
1543 	 * Update process times. Should use high res clock and state
1544 	 * changes instead of statistical sampling method. XXX
1545 	 */
1546 	if (user_mode) {
1547 		pp->p_utime += pending;
1548 	} else {
1549 		pp->p_stime += pending;
1550 	}
1551 
1552 	pp->p_ttime += pending;
1553 	as = pp->p_as;
1554 
1555 	/*
1556 	 * Update user profiling statistics. Get the pc from the
1557 	 * lwp when the AST happens.
1558 	 */
1559 	if (pp->p_prof.pr_scale) {
1560 		atomic_add_32(&lwp->lwp_oweupc, (int32_t)pending);
1561 		if (user_mode) {
1562 			poke = 1;
1563 			aston(t);
1564 		}
1565 	}
1566 
1567 	/*
1568 	 * If CPU was in user state, process lwp-virtual time
1569 	 * interval timer. The value passed to itimerdecr() has to be
1570 	 * in microseconds and has to be less than one second. Hence
1571 	 * this loop.
1572 	 */
1573 	total_usec = usec_per_tick * pending;
1574 	while (total_usec > 0) {
1575 		usec = MIN(total_usec, (MICROSEC - 1));
1576 		if (user_mode &&
1577 		    timerisset(&lwp->lwp_timer[ITIMER_VIRTUAL].it_value) &&
1578 		    itimerdecr(&lwp->lwp_timer[ITIMER_VIRTUAL], usec) == 0) {
1579 			poke = 1;
1580 			sigtoproc(pp, t, SIGVTALRM);
1581 		}
1582 		total_usec -= usec;
1583 	}
1584 
1585 	/*
1586 	 * If CPU was in user state, process lwp-profile
1587 	 * interval timer.
1588 	 */
1589 	total_usec = usec_per_tick * pending;
1590 	while (total_usec > 0) {
1591 		usec = MIN(total_usec, (MICROSEC - 1));
1592 		if (timerisset(&lwp->lwp_timer[ITIMER_PROF].it_value) &&
1593 		    itimerdecr(&lwp->lwp_timer[ITIMER_PROF], usec) == 0) {
1594 			poke = 1;
1595 			sigtoproc(pp, t, SIGPROF);
1596 		}
1597 		total_usec -= usec;
1598 	}
1599 
1600 	/*
1601 	 * Enforce CPU resource controls:
1602 	 *   (a) process.max-cpu-time resource control
1603 	 *
1604 	 * Perform the check only if we have accumulated more a second.
1605 	 */
1606 	if ((ticks + pending) >= hz) {
1607 		(void) rctl_test(rctlproc_legacy[RLIMIT_CPU], pp->p_rctls, pp,
1608 		    (pp->p_utime + pp->p_stime)/hz, RCA_UNSAFE_SIGINFO);
1609 	}
1610 
1611 	/*
1612 	 *   (b) task.max-cpu-time resource control
1613 	 *
1614 	 * If we have accumulated enough ticks, increment the task CPU
1615 	 * time usage and test for the resource limit. This minimizes the
1616 	 * number of calls to the rct_test(). The task CPU time mutex
1617 	 * is highly contentious as many processes can be sharing a task.
1618 	 */
1619 	if (pp->p_ttime >= clock_tick_proc_max) {
1620 		secs = task_cpu_time_incr(pp->p_task, pp->p_ttime);
1621 		pp->p_ttime = 0;
1622 		if (secs) {
1623 			(void) rctl_test(rc_task_cpu_time, pp->p_task->tk_rctls,
1624 			    pp, secs, RCA_UNSAFE_SIGINFO);
1625 		}
1626 	}
1627 
1628 	/*
1629 	 * Update memory usage for the currently running process.
1630 	 */
1631 	rss = rm_asrss(as);
1632 	PTOU(pp)->u_mem += rss;
1633 	if (rss > PTOU(pp)->u_mem_max)
1634 		PTOU(pp)->u_mem_max = rss;
1635 
1636 	/*
1637 	 * Notify the CPU the thread is running on.
1638 	 */
1639 	if (poke && t->t_cpu != CPU)
1640 		poke_cpu(t->t_cpu->cpu_id);
1641 }
1642 
1643 void
profil_tick(uintptr_t upc)1644 profil_tick(uintptr_t upc)
1645 {
1646 	int ticks;
1647 	proc_t *p = ttoproc(curthread);
1648 	klwp_t *lwp = ttolwp(curthread);
1649 	struct prof *pr = &p->p_prof;
1650 
1651 	do {
1652 		ticks = lwp->lwp_oweupc;
1653 	} while (atomic_cas_32(&lwp->lwp_oweupc, ticks, 0) != ticks);
1654 
1655 	mutex_enter(&p->p_pflock);
1656 	if (pr->pr_scale >= 2 && upc >= pr->pr_off) {
1657 		/*
1658 		 * Old-style profiling
1659 		 */
1660 		uint16_t *slot = pr->pr_base;
1661 		uint16_t old, new;
1662 		if (pr->pr_scale != 2) {
1663 			uintptr_t delta = upc - pr->pr_off;
1664 			uintptr_t byteoff = ((delta >> 16) * pr->pr_scale) +
1665 			    (((delta & 0xffff) * pr->pr_scale) >> 16);
1666 			if (byteoff >= (uintptr_t)pr->pr_size) {
1667 				mutex_exit(&p->p_pflock);
1668 				return;
1669 			}
1670 			slot += byteoff / sizeof (uint16_t);
1671 		}
1672 		if (fuword16(slot, &old) < 0 ||
1673 		    (new = old + ticks) > SHRT_MAX ||
1674 		    suword16(slot, new) < 0) {
1675 			pr->pr_scale = 0;
1676 		}
1677 	} else if (pr->pr_scale == 1) {
1678 		/*
1679 		 * PC Sampling
1680 		 */
1681 		model_t model = lwp_getdatamodel(lwp);
1682 		int result;
1683 #ifdef __lint
1684 		model = model;
1685 #endif
1686 		while (ticks-- > 0) {
1687 			if (pr->pr_samples == pr->pr_size) {
1688 				/* buffer full, turn off sampling */
1689 				pr->pr_scale = 0;
1690 				break;
1691 			}
1692 			switch (SIZEOF_PTR(model)) {
1693 			case sizeof (uint32_t):
1694 				result = suword32(pr->pr_base, (uint32_t)upc);
1695 				break;
1696 #ifdef _LP64
1697 			case sizeof (uint64_t):
1698 				result = suword64(pr->pr_base, (uint64_t)upc);
1699 				break;
1700 #endif
1701 			default:
1702 				cmn_err(CE_WARN, "profil_tick: unexpected "
1703 				    "data model");
1704 				result = -1;
1705 				break;
1706 			}
1707 			if (result != 0) {
1708 				pr->pr_scale = 0;
1709 				break;
1710 			}
1711 			pr->pr_base = (caddr_t)pr->pr_base + SIZEOF_PTR(model);
1712 			pr->pr_samples++;
1713 		}
1714 	}
1715 	mutex_exit(&p->p_pflock);
1716 }
1717 
1718 static void
delay_wakeup(void * arg)1719 delay_wakeup(void *arg)
1720 {
1721 	kthread_t	*t = arg;
1722 
1723 	mutex_enter(&t->t_delay_lock);
1724 	cv_signal(&t->t_delay_cv);
1725 	mutex_exit(&t->t_delay_lock);
1726 }
1727 
1728 /*
1729  * The delay(9F) man page indicates that it can only be called from user or
1730  * kernel context - detect and diagnose bad calls. The following macro will
1731  * produce a limited number of messages identifying bad callers.  This is done
1732  * in a macro so that caller() is meaningful. When a bad caller is identified,
1733  * switching to 'drv_usecwait(TICK_TO_USEC(ticks));' may be appropriate.
1734  */
1735 #define	DELAY_CONTEXT_CHECK()	{					\
1736 	uint32_t	m;						\
1737 	char		*f;						\
1738 	ulong_t		off;						\
1739 									\
1740 	m = delay_from_interrupt_msg;					\
1741 	if (delay_from_interrupt_diagnose && servicing_interrupt() &&	\
1742 	    !panicstr && !devinfo_freeze &&				\
1743 	    atomic_cas_32(&delay_from_interrupt_msg, m ? m : 1, m-1)) {	\
1744 		f = modgetsymname((uintptr_t)caller(), &off);		\
1745 		cmn_err(CE_WARN, "delay(9F) called from "		\
1746 		    "interrupt context: %s`%s",				\
1747 		    mod_containing_pc(caller()), f ? f : "...");	\
1748 	}								\
1749 }
1750 
1751 /*
1752  * delay_common: common delay code.
1753  */
1754 static void
delay_common(clock_t ticks)1755 delay_common(clock_t ticks)
1756 {
1757 	kthread_t	*t = curthread;
1758 	clock_t		deadline;
1759 	clock_t		timeleft;
1760 	callout_id_t	id;
1761 
1762 	/* If timeouts aren't running all we can do is spin. */
1763 	if (panicstr || devinfo_freeze) {
1764 		/* Convert delay(9F) call into drv_usecwait(9F) call. */
1765 		if (ticks > 0)
1766 			drv_usecwait(TICK_TO_USEC(ticks));
1767 		return;
1768 	}
1769 
1770 	deadline = ddi_get_lbolt() + ticks;
1771 	while ((timeleft = deadline - ddi_get_lbolt()) > 0) {
1772 		mutex_enter(&t->t_delay_lock);
1773 		id = timeout_default(delay_wakeup, t, timeleft);
1774 		cv_wait(&t->t_delay_cv, &t->t_delay_lock);
1775 		mutex_exit(&t->t_delay_lock);
1776 		(void) untimeout_default(id, 0);
1777 	}
1778 }
1779 
1780 /*
1781  * Delay specified number of clock ticks.
1782  */
1783 void
delay(clock_t ticks)1784 delay(clock_t ticks)
1785 {
1786 	DELAY_CONTEXT_CHECK();
1787 
1788 	delay_common(ticks);
1789 }
1790 
1791 /*
1792  * Delay a random number of clock ticks between 1 and ticks.
1793  */
1794 void
delay_random(clock_t ticks)1795 delay_random(clock_t ticks)
1796 {
1797 	int	r;
1798 
1799 	DELAY_CONTEXT_CHECK();
1800 
1801 	(void) random_get_pseudo_bytes((void *)&r, sizeof (r));
1802 	if (ticks == 0)
1803 		ticks = 1;
1804 	ticks = (r % ticks) + 1;
1805 	delay_common(ticks);
1806 }
1807 
1808 /*
1809  * Like delay, but interruptible by a signal.
1810  */
1811 int
delay_sig(clock_t ticks)1812 delay_sig(clock_t ticks)
1813 {
1814 	kthread_t	*t = curthread;
1815 	clock_t		deadline;
1816 	clock_t		rc;
1817 
1818 	/* If timeouts aren't running all we can do is spin. */
1819 	if (panicstr || devinfo_freeze) {
1820 		if (ticks > 0)
1821 			drv_usecwait(TICK_TO_USEC(ticks));
1822 		return (0);
1823 	}
1824 
1825 	deadline = ddi_get_lbolt() + ticks;
1826 	mutex_enter(&t->t_delay_lock);
1827 	do {
1828 		rc = cv_timedwait_sig(&t->t_delay_cv,
1829 		    &t->t_delay_lock, deadline);
1830 		/* loop until past deadline or signaled */
1831 	} while (rc > 0);
1832 	mutex_exit(&t->t_delay_lock);
1833 	if (rc == 0)
1834 		return (EINTR);
1835 	return (0);
1836 }
1837 
1838 
1839 #define	SECONDS_PER_DAY 86400
1840 
1841 /*
1842  * Initialize the system time based on the TOD chip.  approx is used as
1843  * an approximation of time (e.g. from the filesystem) in the event that
1844  * the TOD chip has been cleared or is unresponsive.  An approx of -1
1845  * means the filesystem doesn't keep time.
1846  */
1847 void
clkset(time_t approx)1848 clkset(time_t approx)
1849 {
1850 	timestruc_t ts;
1851 	int spl;
1852 	int set_clock = 0;
1853 
1854 	mutex_enter(&tod_lock);
1855 	ts = tod_get();
1856 
1857 	if (ts.tv_sec > 365 * SECONDS_PER_DAY) {
1858 		/*
1859 		 * If the TOD chip is reporting some time after 1971,
1860 		 * then it probably didn't lose power or become otherwise
1861 		 * cleared in the recent past;  check to assure that
1862 		 * the time coming from the filesystem isn't in the future
1863 		 * according to the TOD chip.
1864 		 */
1865 		if (approx != -1 && approx > ts.tv_sec) {
1866 			cmn_err(CE_WARN, "Last shutdown is later "
1867 			    "than time on time-of-day chip; check date.");
1868 		}
1869 	} else {
1870 		/*
1871 		 * If the TOD chip isn't giving correct time, set it to the
1872 		 * greater of i) approx and ii) 1987. That way if approx
1873 		 * is negative or is earlier than 1987, we set the clock
1874 		 * back to a time when Oliver North, ALF and Dire Straits
1875 		 * were all on the collective brain:  1987.
1876 		 */
1877 		timestruc_t tmp;
1878 		time_t diagnose_date = (1987 - 1970) * 365 * SECONDS_PER_DAY;
1879 		ts.tv_sec = (approx > diagnose_date ? approx : diagnose_date);
1880 		ts.tv_nsec = 0;
1881 
1882 		/*
1883 		 * Attempt to write the new time to the TOD chip.  Set spl high
1884 		 * to avoid getting preempted between the tod_set and tod_get.
1885 		 */
1886 		spl = splhi();
1887 		tod_set(ts);
1888 		tmp = tod_get();
1889 		splx(spl);
1890 
1891 		if (tmp.tv_sec != ts.tv_sec && tmp.tv_sec != ts.tv_sec + 1) {
1892 			tod_broken = 1;
1893 			dosynctodr = 0;
1894 			cmn_err(CE_WARN, "Time-of-day chip unresponsive.");
1895 		} else {
1896 			cmn_err(CE_WARN, "Time-of-day chip had "
1897 			    "incorrect date; check and reset.");
1898 		}
1899 		set_clock = 1;
1900 	}
1901 
1902 	if (!boot_time) {
1903 		boot_time = ts.tv_sec;
1904 		set_clock = 1;
1905 	}
1906 
1907 	if (set_clock)
1908 		set_hrestime(&ts);
1909 
1910 	mutex_exit(&tod_lock);
1911 }
1912 
1913 int	timechanged;	/* for testing if the system time has been reset */
1914 
1915 void
set_hrestime(timestruc_t * ts)1916 set_hrestime(timestruc_t *ts)
1917 {
1918 	int spl = hr_clock_lock();
1919 	hrestime = *ts;
1920 	membar_enter();	/* hrestime must be visible before timechanged++ */
1921 	timedelta = 0;
1922 	timechanged++;
1923 	hr_clock_unlock(spl);
1924 	callout_hrestime();
1925 }
1926 
1927 static uint_t deadman_seconds;
1928 static uint32_t deadman_panics;
1929 static int deadman_enabled = 0;
1930 static int deadman_panic_timers = 1;
1931 
1932 static void
deadman(void)1933 deadman(void)
1934 {
1935 	if (panicstr) {
1936 		/*
1937 		 * During panic, other CPUs besides the panic
1938 		 * master continue to handle cyclics and some other
1939 		 * interrupts.  The code below is intended to be
1940 		 * single threaded, so any CPU other than the master
1941 		 * must keep out.
1942 		 */
1943 		if (CPU->cpu_id != panic_cpu.cpu_id)
1944 			return;
1945 
1946 		if (!deadman_panic_timers)
1947 			return; /* allow all timers to be manually disabled */
1948 
1949 		/*
1950 		 * If we are generating a crash dump or syncing filesystems and
1951 		 * the corresponding timer is set, decrement it and re-enter
1952 		 * the panic code to abort it and advance to the next state.
1953 		 * The panic states and triggers are explained in panic.c.
1954 		 */
1955 		if (panic_dump) {
1956 			if (dump_timeleft && (--dump_timeleft == 0)) {
1957 				panic("panic dump timeout");
1958 				/*NOTREACHED*/
1959 			}
1960 		} else if (panic_sync) {
1961 			if (sync_timeleft && (--sync_timeleft == 0)) {
1962 				panic("panic sync timeout");
1963 				/*NOTREACHED*/
1964 			}
1965 		}
1966 
1967 		return;
1968 	}
1969 
1970 	if (deadman_counter != CPU->cpu_deadman_counter) {
1971 		CPU->cpu_deadman_counter = deadman_counter;
1972 		CPU->cpu_deadman_countdown = deadman_seconds;
1973 		return;
1974 	}
1975 
1976 	if (--CPU->cpu_deadman_countdown > 0)
1977 		return;
1978 
1979 	/*
1980 	 * Regardless of whether or not we actually bring the system down,
1981 	 * bump the deadman_panics variable.
1982 	 *
1983 	 * N.B. deadman_panics is incremented once for each CPU that
1984 	 * passes through here.  It's expected that all the CPUs will
1985 	 * detect this condition within one second of each other, so
1986 	 * when deadman_enabled is off, deadman_panics will
1987 	 * typically be a multiple of the total number of CPUs in
1988 	 * the system.
1989 	 */
1990 	atomic_inc_32(&deadman_panics);
1991 
1992 	if (!deadman_enabled) {
1993 		CPU->cpu_deadman_countdown = deadman_seconds;
1994 		return;
1995 	}
1996 
1997 	/*
1998 	 * If we're here, we want to bring the system down.
1999 	 */
2000 	panic("deadman: timed out after %d seconds of clock "
2001 	    "inactivity", deadman_seconds);
2002 	/*NOTREACHED*/
2003 }
2004 
2005 /*ARGSUSED*/
2006 static void
deadman_online(void * arg,cpu_t * cpu,cyc_handler_t * hdlr,cyc_time_t * when)2007 deadman_online(void *arg, cpu_t *cpu, cyc_handler_t *hdlr, cyc_time_t *when)
2008 {
2009 	cpu->cpu_deadman_counter = 0;
2010 	cpu->cpu_deadman_countdown = deadman_seconds;
2011 
2012 	hdlr->cyh_func = (cyc_func_t)deadman;
2013 	hdlr->cyh_level = CY_HIGH_LEVEL;
2014 	hdlr->cyh_arg = NULL;
2015 
2016 	/*
2017 	 * Stagger the CPUs so that they don't all run deadman() at
2018 	 * the same time.  Simplest reason to do this is to make it
2019 	 * more likely that only one CPU will panic in case of a
2020 	 * timeout.  This is (strictly speaking) an aesthetic, not a
2021 	 * technical consideration.
2022 	 */
2023 	when->cyt_when = cpu->cpu_id * (NANOSEC / NCPU);
2024 	when->cyt_interval = NANOSEC;
2025 }
2026 
2027 
2028 void
deadman_init(void)2029 deadman_init(void)
2030 {
2031 	cyc_omni_handler_t hdlr;
2032 
2033 	if (deadman_seconds == 0)
2034 		deadman_seconds = snoop_interval / MICROSEC;
2035 
2036 	if (snooping)
2037 		deadman_enabled = 1;
2038 
2039 	hdlr.cyo_online = deadman_online;
2040 	hdlr.cyo_offline = NULL;
2041 	hdlr.cyo_arg = NULL;
2042 
2043 	mutex_enter(&cpu_lock);
2044 	deadman_cyclic = cyclic_add_omni(&hdlr);
2045 	mutex_exit(&cpu_lock);
2046 }
2047 
2048 /*
2049  * tod_fault() is for updating tod validate mechanism state:
2050  * (1) TOD_NOFAULT: for resetting the state to 'normal'.
2051  *     currently used for debugging only
2052  * (2) The following four cases detected by tod validate mechanism:
2053  *       TOD_REVERSED: current tod value is less than previous value.
2054  *       TOD_STALLED: current tod value hasn't advanced.
2055  *       TOD_JUMPED: current tod value advanced too far from previous value.
2056  *       TOD_RATECHANGED: the ratio between average tod delta and
2057  *       average tick delta has changed.
2058  * (3) TOD_RDONLY: when the TOD clock is not writeable e.g. because it is
2059  *     a virtual TOD provided by a hypervisor.
2060  */
2061 enum tod_fault_type
tod_fault(enum tod_fault_type ftype,int off)2062 tod_fault(enum tod_fault_type ftype, int off)
2063 {
2064 	ASSERT(MUTEX_HELD(&tod_lock));
2065 
2066 	if (tod_faulted != ftype) {
2067 		switch (ftype) {
2068 		case TOD_NOFAULT:
2069 			plat_tod_fault(TOD_NOFAULT);
2070 			cmn_err(CE_NOTE, "Restarted tracking "
2071 			    "Time of Day clock.");
2072 			tod_faulted = ftype;
2073 			break;
2074 		case TOD_REVERSED:
2075 		case TOD_JUMPED:
2076 			if (tod_faulted == TOD_NOFAULT) {
2077 				plat_tod_fault(ftype);
2078 				cmn_err(CE_WARN, "Time of Day clock error: "
2079 				    "reason [%s by 0x%x]. -- "
2080 				    " Stopped tracking Time Of Day clock.",
2081 				    tod_fault_table[ftype], off);
2082 				tod_faulted = ftype;
2083 			}
2084 			break;
2085 		case TOD_STALLED:
2086 		case TOD_RATECHANGED:
2087 			if (tod_faulted == TOD_NOFAULT) {
2088 				plat_tod_fault(ftype);
2089 				cmn_err(CE_WARN, "Time of Day clock error: "
2090 				    "reason [%s]. -- "
2091 				    " Stopped tracking Time Of Day clock.",
2092 				    tod_fault_table[ftype]);
2093 				tod_faulted = ftype;
2094 			}
2095 			break;
2096 		case TOD_RDONLY:
2097 			if (tod_faulted == TOD_NOFAULT) {
2098 				plat_tod_fault(ftype);
2099 				cmn_err(CE_NOTE, "!Time of Day clock is "
2100 				    "Read-Only; set of Date/Time will not "
2101 				    "persist across reboot.");
2102 				tod_faulted = ftype;
2103 			}
2104 			break;
2105 		default:
2106 			break;
2107 		}
2108 	}
2109 	return (tod_faulted);
2110 }
2111 
2112 /*
2113  * Two functions that allow tod_status_flag to be manipulated by functions
2114  * external to this file.
2115  */
2116 
2117 void
tod_status_set(int tod_flag)2118 tod_status_set(int tod_flag)
2119 {
2120 	tod_status_flag |= tod_flag;
2121 }
2122 
2123 void
tod_status_clear(int tod_flag)2124 tod_status_clear(int tod_flag)
2125 {
2126 	tod_status_flag &= ~tod_flag;
2127 }
2128 
2129 /*
2130  * Record a timestamp and the value passed to tod_set().  The next call to
2131  * tod_validate() can use these values, prev_set_tick and prev_set_tod,
2132  * when checking the timestruc_t returned by tod_get().  Ordinarily,
2133  * tod_validate() will use prev_tick and prev_tod for this task but these
2134  * become obsolete, and will be re-assigned with the prev_set_* values,
2135  * in the case when the TOD is re-written.
2136  */
2137 void
tod_set_prev(timestruc_t ts)2138 tod_set_prev(timestruc_t ts)
2139 {
2140 	if ((tod_validate_enable == 0) || (tod_faulted != TOD_NOFAULT) ||
2141 	    tod_validate_deferred) {
2142 		return;
2143 	}
2144 	prev_set_tick = gethrtime();
2145 	/*
2146 	 * A negative value will be set to zero in utc_to_tod() so we fake
2147 	 * a zero here in such a case.  This would need to change if the
2148 	 * behavior of utc_to_tod() changes.
2149 	 */
2150 	prev_set_tod = ts.tv_sec < 0 ? 0 : ts.tv_sec;
2151 }
2152 
2153 /*
2154  * tod_validate() is used for checking values returned by tod_get().
2155  * Four error cases can be detected by this routine:
2156  *   TOD_REVERSED: current tod value is less than previous.
2157  *   TOD_STALLED: current tod value hasn't advanced.
2158  *   TOD_JUMPED: current tod value advanced too far from previous value.
2159  *   TOD_RATECHANGED: the ratio between average tod delta and
2160  *   average tick delta has changed.
2161  */
2162 time_t
tod_validate(time_t tod)2163 tod_validate(time_t tod)
2164 {
2165 	time_t diff_tod;
2166 	hrtime_t diff_tick;
2167 
2168 	long dtick;
2169 	int dtick_delta;
2170 
2171 	int off = 0;
2172 	enum tod_fault_type tod_bad = TOD_NOFAULT;
2173 
2174 	static int firsttime = 1;
2175 
2176 	static time_t prev_tod = 0;
2177 	static hrtime_t prev_tick = 0;
2178 	static long dtick_avg = TOD_REF_FREQ;
2179 
2180 	int cpr_resume_done = 0;
2181 	int dr_resume_done = 0;
2182 
2183 	hrtime_t tick = gethrtime();
2184 
2185 	ASSERT(MUTEX_HELD(&tod_lock));
2186 
2187 	/*
2188 	 * tod_validate_enable is patchable via /etc/system.
2189 	 * If TOD is already faulted, or if TOD validation is deferred,
2190 	 * there is nothing to do.
2191 	 */
2192 	if ((tod_validate_enable == 0) || (tod_faulted != TOD_NOFAULT) ||
2193 	    tod_validate_deferred) {
2194 		return (tod);
2195 	}
2196 
2197 	/*
2198 	 * If this is the first time through, we just need to save the tod
2199 	 * we were called with and hrtime so we can use them next time to
2200 	 * validate tod_get().
2201 	 */
2202 	if (firsttime) {
2203 		firsttime = 0;
2204 		prev_tod = tod;
2205 		prev_tick = tick;
2206 		return (tod);
2207 	}
2208 
2209 	/*
2210 	 * Handle any flags that have been turned on by tod_status_set().
2211 	 * In the case where a tod_set() is done and then a subsequent
2212 	 * tod_get() fails (ie, both TOD_SET_DONE and TOD_GET_FAILED are
2213 	 * true), we treat the TOD_GET_FAILED with precedence by switching
2214 	 * off the flag, returning tod and leaving TOD_SET_DONE asserted
2215 	 * until such time as tod_get() completes successfully.
2216 	 */
2217 	if (tod_status_flag & TOD_GET_FAILED) {
2218 		/*
2219 		 * tod_get() has encountered an issue, possibly transitory,
2220 		 * when reading TOD.  We'll just return the incoming tod
2221 		 * value (which is actually hrestime.tv_sec in this case)
2222 		 * and when we get a genuine tod, following a successful
2223 		 * tod_get(), we can validate using prev_tod and prev_tick.
2224 		 */
2225 		tod_status_flag &= ~TOD_GET_FAILED;
2226 		return (tod);
2227 	} else if (tod_status_flag & TOD_SET_DONE) {
2228 		/*
2229 		 * TOD has been modified.  Just before the TOD was written,
2230 		 * tod_set_prev() saved tod and hrtime; we can now use
2231 		 * those values, prev_set_tod and prev_set_tick, to validate
2232 		 * the incoming tod that's just been read.
2233 		 */
2234 		prev_tod = prev_set_tod;
2235 		prev_tick = prev_set_tick;
2236 		dtick_avg = TOD_REF_FREQ;
2237 		tod_status_flag &= ~TOD_SET_DONE;
2238 		/*
2239 		 * If a tod_set() preceded a cpr_suspend() without an
2240 		 * intervening tod_validate(), we need to ensure that a
2241 		 * TOD_JUMPED condition is ignored.
2242 		 * Note this isn't a concern in the case of DR as we've
2243 		 * just reassigned dtick_avg, above.
2244 		 */
2245 		if (tod_status_flag & TOD_CPR_RESUME_DONE) {
2246 			cpr_resume_done = 1;
2247 			tod_status_flag &= ~TOD_CPR_RESUME_DONE;
2248 		}
2249 	} else if (tod_status_flag & TOD_CPR_RESUME_DONE) {
2250 		/*
2251 		 * The system's coming back from a checkpoint resume.
2252 		 */
2253 		cpr_resume_done = 1;
2254 		tod_status_flag &= ~TOD_CPR_RESUME_DONE;
2255 		/*
2256 		 * We need to handle the possibility of a CPR suspend
2257 		 * operation having been initiated whilst a DR event was
2258 		 * in-flight.
2259 		 */
2260 		if (tod_status_flag & TOD_DR_RESUME_DONE) {
2261 			dr_resume_done = 1;
2262 			tod_status_flag &= ~TOD_DR_RESUME_DONE;
2263 		}
2264 	} else if (tod_status_flag & TOD_DR_RESUME_DONE) {
2265 		/*
2266 		 * A Dynamic Reconfiguration event has taken place.
2267 		 */
2268 		dr_resume_done = 1;
2269 		tod_status_flag &= ~TOD_DR_RESUME_DONE;
2270 	}
2271 
2272 	/* test hook */
2273 	switch (tod_unit_test) {
2274 	case 1: /* for testing jumping tod */
2275 		tod += tod_test_injector;
2276 		tod_unit_test = 0;
2277 		break;
2278 	case 2:	/* for testing stuck tod bit */
2279 		tod |= 1 << tod_test_injector;
2280 		tod_unit_test = 0;
2281 		break;
2282 	case 3:	/* for testing stalled tod */
2283 		tod = prev_tod;
2284 		tod_unit_test = 0;
2285 		break;
2286 	case 4:	/* reset tod fault status */
2287 		(void) tod_fault(TOD_NOFAULT, 0);
2288 		tod_unit_test = 0;
2289 		break;
2290 	default:
2291 		break;
2292 	}
2293 
2294 	diff_tod = tod - prev_tod;
2295 	diff_tick = tick - prev_tick;
2296 
2297 	ASSERT(diff_tick >= 0);
2298 
2299 	if (diff_tod < 0) {
2300 		/* ERROR - tod reversed */
2301 		tod_bad = TOD_REVERSED;
2302 		off = (int)(prev_tod - tod);
2303 	} else if (diff_tod == 0) {
2304 		/* tod did not advance */
2305 		if (diff_tick > TOD_STALL_THRESHOLD) {
2306 			/* ERROR - tod stalled */
2307 			tod_bad = TOD_STALLED;
2308 		} else {
2309 			/*
2310 			 * Make sure we don't update prev_tick
2311 			 * so that diff_tick is calculated since
2312 			 * the first diff_tod == 0
2313 			 */
2314 			return (tod);
2315 		}
2316 	} else {
2317 		/* calculate dtick */
2318 		dtick = diff_tick / diff_tod;
2319 
2320 		/* update dtick averages */
2321 		dtick_avg += ((dtick - dtick_avg) / TOD_FILTER_N);
2322 
2323 		/*
2324 		 * Calculate dtick_delta as
2325 		 * variation from reference freq in quartiles
2326 		 */
2327 		dtick_delta = (dtick_avg - TOD_REF_FREQ) /
2328 		    (TOD_REF_FREQ >> 2);
2329 
2330 		/*
2331 		 * Even with a perfectly functioning TOD device,
2332 		 * when the number of elapsed seconds is low the
2333 		 * algorithm can calculate a rate that is beyond
2334 		 * tolerance, causing an error.  The algorithm is
2335 		 * inaccurate when elapsed time is low (less than
2336 		 * 5 seconds).
2337 		 */
2338 		if (diff_tod > 4) {
2339 			if (dtick < TOD_JUMP_THRESHOLD) {
2340 				/*
2341 				 * If we've just done a CPR resume, we detect
2342 				 * a jump in the TOD but, actually, what's
2343 				 * happened is that the TOD has been increasing
2344 				 * whilst the system was suspended and the tick
2345 				 * count hasn't kept up.  We consider the first
2346 				 * occurrence of this after a resume as normal
2347 				 * and ignore it; otherwise, in a non-resume
2348 				 * case, we regard it as a TOD problem.
2349 				 */
2350 				if (!cpr_resume_done) {
2351 					/* ERROR - tod jumped */
2352 					tod_bad = TOD_JUMPED;
2353 					off = (int)diff_tod;
2354 				}
2355 			}
2356 			if (dtick_delta) {
2357 				/*
2358 				 * If we've just done a DR resume, dtick_avg
2359 				 * can go a bit askew so we reset it and carry
2360 				 * on; otherwise, the TOD is in error.
2361 				 */
2362 				if (dr_resume_done) {
2363 					dtick_avg = TOD_REF_FREQ;
2364 				} else {
2365 					/* ERROR - change in clock rate */
2366 					tod_bad = TOD_RATECHANGED;
2367 				}
2368 			}
2369 		}
2370 	}
2371 
2372 	if (tod_bad != TOD_NOFAULT) {
2373 		(void) tod_fault(tod_bad, off);
2374 
2375 		/*
2376 		 * Disable dosynctodr since we are going to fault
2377 		 * the TOD chip anyway here
2378 		 */
2379 		dosynctodr = 0;
2380 
2381 		/*
2382 		 * Set tod to the correct value from hrestime
2383 		 */
2384 		tod = hrestime.tv_sec;
2385 	}
2386 
2387 	prev_tod = tod;
2388 	prev_tick = tick;
2389 	return (tod);
2390 }
2391 
2392 static void
calcloadavg(int nrun,uint64_t * hp_ave)2393 calcloadavg(int nrun, uint64_t *hp_ave)
2394 {
2395 	static int64_t f[3] = { 135, 27, 9 };
2396 	uint_t i;
2397 	int64_t q, r;
2398 
2399 	/*
2400 	 * Compute load average over the last 1, 5, and 15 minutes
2401 	 * (60, 300, and 900 seconds).  The constants in f[3] are for
2402 	 * exponential decay:
2403 	 * (1 - exp(-1/60)) << 13 = 135,
2404 	 * (1 - exp(-1/300)) << 13 = 27,
2405 	 * (1 - exp(-1/900)) << 13 = 9.
2406 	 */
2407 
2408 	/*
2409 	 * a little hoop-jumping to avoid integer overflow
2410 	 */
2411 	for (i = 0; i < 3; i++) {
2412 		q = (hp_ave[i]  >> 16) << 7;
2413 		r = (hp_ave[i]  & 0xffff) << 7;
2414 		hp_ave[i] += ((nrun - q) * f[i] - ((r * f[i]) >> 16)) >> 4;
2415 	}
2416 }
2417 
2418 /*
2419  * lbolt_hybrid() is used by ddi_get_lbolt() and ddi_get_lbolt64() to
2420  * calculate the value of lbolt according to the current mode. In the event
2421  * driven mode (the default), lbolt is calculated by dividing the current hires
2422  * time by the number of nanoseconds per clock tick. In the cyclic driven mode
2423  * an internal variable is incremented at each firing of the lbolt cyclic
2424  * and returned by lbolt_cyclic_driven().
2425  *
2426  * The system will transition from event to cyclic driven mode when the number
2427  * of calls to lbolt_event_driven() exceeds the (per CPU) threshold within a
2428  * window of time. It does so by reprograming lbolt_cyclic from CY_INFINITY to
2429  * nsec_per_tick. The lbolt cyclic will remain ON while at least one CPU is
2430  * causing enough activity to cross the thresholds.
2431  */
2432 int64_t
lbolt_bootstrap(void)2433 lbolt_bootstrap(void)
2434 {
2435 	return (0);
2436 }
2437 
2438 /* ARGSUSED */
2439 uint_t
lbolt_ev_to_cyclic(caddr_t arg1,caddr_t arg2)2440 lbolt_ev_to_cyclic(caddr_t arg1, caddr_t arg2)
2441 {
2442 	hrtime_t ts, exp;
2443 	int ret;
2444 
2445 	ASSERT(lbolt_hybrid != lbolt_cyclic_driven);
2446 
2447 	kpreempt_disable();
2448 
2449 	ts = gethrtime();
2450 	lb_info->lbi_internal = (ts/nsec_per_tick);
2451 
2452 	/*
2453 	 * Align the next expiration to a clock tick boundary.
2454 	 */
2455 	exp = ts + nsec_per_tick - 1;
2456 	exp = (exp/nsec_per_tick) * nsec_per_tick;
2457 
2458 	ret = cyclic_reprogram(lb_info->id.lbi_cyclic_id, exp);
2459 	ASSERT(ret);
2460 
2461 	lbolt_hybrid = lbolt_cyclic_driven;
2462 	lb_info->lbi_cyc_deactivate = B_FALSE;
2463 	lb_info->lbi_cyc_deac_start = lb_info->lbi_internal;
2464 
2465 	kpreempt_enable();
2466 
2467 	ret = atomic_dec_32_nv(&lb_info->lbi_token);
2468 	ASSERT(ret == 0);
2469 
2470 	return (1);
2471 }
2472 
2473 int64_t
lbolt_event_driven(void)2474 lbolt_event_driven(void)
2475 {
2476 	hrtime_t ts;
2477 	int64_t lb;
2478 	int ret, cpu = CPU->cpu_seqid;
2479 
2480 	ts = gethrtime();
2481 	ASSERT(ts > 0);
2482 
2483 	ASSERT(nsec_per_tick > 0);
2484 	lb = (ts/nsec_per_tick);
2485 
2486 	/*
2487 	 * Switch to cyclic mode if the number of calls to this routine
2488 	 * has reached the threshold within the interval.
2489 	 */
2490 	if ((lb - lb_cpu[cpu].lbc_cnt_start) < lb_info->lbi_thresh_interval) {
2491 
2492 		if (--lb_cpu[cpu].lbc_counter == 0) {
2493 			/*
2494 			 * Reached the threshold within the interval, reset
2495 			 * the usage statistics.
2496 			 */
2497 			lb_cpu[cpu].lbc_counter = lb_info->lbi_thresh_calls;
2498 			lb_cpu[cpu].lbc_cnt_start = lb;
2499 
2500 			/*
2501 			 * Make sure only one thread reprograms the
2502 			 * lbolt cyclic and changes the mode.
2503 			 */
2504 			if (panicstr == NULL &&
2505 			    atomic_cas_32(&lb_info->lbi_token, 0, 1) == 0) {
2506 
2507 				if (lbolt_hybrid == lbolt_cyclic_driven) {
2508 					ret = atomic_dec_32_nv(
2509 					    &lb_info->lbi_token);
2510 					ASSERT(ret == 0);
2511 				} else {
2512 					lbolt_softint_post();
2513 				}
2514 			}
2515 		}
2516 	} else {
2517 		/*
2518 		 * Exceeded the interval, reset the usage statistics.
2519 		 */
2520 		lb_cpu[cpu].lbc_counter = lb_info->lbi_thresh_calls;
2521 		lb_cpu[cpu].lbc_cnt_start = lb;
2522 	}
2523 
2524 	ASSERT(lb >= lb_info->lbi_debug_time);
2525 
2526 	return (lb - lb_info->lbi_debug_time);
2527 }
2528 
2529 int64_t
lbolt_cyclic_driven(void)2530 lbolt_cyclic_driven(void)
2531 {
2532 	int64_t lb = lb_info->lbi_internal;
2533 	int cpu;
2534 
2535 	/*
2536 	 * If a CPU has already prevented the lbolt cyclic from deactivating
2537 	 * itself, don't bother tracking the usage. Otherwise check if we're
2538 	 * within the interval and how the per CPU counter is doing.
2539 	 */
2540 	if (lb_info->lbi_cyc_deactivate) {
2541 		cpu = CPU->cpu_seqid;
2542 		if ((lb - lb_cpu[cpu].lbc_cnt_start) <
2543 		    lb_info->lbi_thresh_interval) {
2544 
2545 			if (lb_cpu[cpu].lbc_counter == 0)
2546 				/*
2547 				 * Reached the threshold within the interval,
2548 				 * prevent the lbolt cyclic from turning itself
2549 				 * off.
2550 				 */
2551 				lb_info->lbi_cyc_deactivate = B_FALSE;
2552 			else
2553 				lb_cpu[cpu].lbc_counter--;
2554 		} else {
2555 			/*
2556 			 * Only reset the usage statistics when we have
2557 			 * exceeded the interval.
2558 			 */
2559 			lb_cpu[cpu].lbc_counter = lb_info->lbi_thresh_calls;
2560 			lb_cpu[cpu].lbc_cnt_start = lb;
2561 		}
2562 	}
2563 
2564 	ASSERT(lb >= lb_info->lbi_debug_time);
2565 
2566 	return (lb - lb_info->lbi_debug_time);
2567 }
2568 
2569 /*
2570  * The lbolt_cyclic() routine will fire at a nsec_per_tick interval to satisfy
2571  * performance needs of ddi_get_lbolt() and ddi_get_lbolt64() consumers.
2572  * It is inactive by default, and will be activated when switching from event
2573  * to cyclic driven lbolt. The cyclic will turn itself off unless signaled
2574  * by lbolt_cyclic_driven().
2575  */
2576 static void
lbolt_cyclic(void)2577 lbolt_cyclic(void)
2578 {
2579 	int ret;
2580 
2581 	lb_info->lbi_internal++;
2582 
2583 	if (!lbolt_cyc_only) {
2584 
2585 		if (lb_info->lbi_cyc_deactivate) {
2586 			/*
2587 			 * Switching from cyclic to event driven mode.
2588 			 */
2589 			if (panicstr == NULL &&
2590 			    atomic_cas_32(&lb_info->lbi_token, 0, 1) == 0) {
2591 
2592 				if (lbolt_hybrid == lbolt_event_driven) {
2593 					ret = atomic_dec_32_nv(
2594 					    &lb_info->lbi_token);
2595 					ASSERT(ret == 0);
2596 					return;
2597 				}
2598 
2599 				kpreempt_disable();
2600 
2601 				lbolt_hybrid = lbolt_event_driven;
2602 				ret = cyclic_reprogram(
2603 				    lb_info->id.lbi_cyclic_id,
2604 				    CY_INFINITY);
2605 				ASSERT(ret);
2606 
2607 				kpreempt_enable();
2608 
2609 				ret = atomic_dec_32_nv(&lb_info->lbi_token);
2610 				ASSERT(ret == 0);
2611 			}
2612 		}
2613 
2614 		/*
2615 		 * The lbolt cyclic should not try to deactivate itself before
2616 		 * the sampling period has elapsed.
2617 		 */
2618 		if (lb_info->lbi_internal - lb_info->lbi_cyc_deac_start >=
2619 		    lb_info->lbi_thresh_interval) {
2620 			lb_info->lbi_cyc_deactivate = B_TRUE;
2621 			lb_info->lbi_cyc_deac_start = lb_info->lbi_internal;
2622 		}
2623 	}
2624 }
2625 
2626 /*
2627  * Since the lbolt service was historically cyclic driven, it must be 'stopped'
2628  * when the system drops into the kernel debugger. lbolt_debug_entry() is
2629  * called by the KDI system claim callbacks to record a hires timestamp at
2630  * debug enter time. lbolt_debug_return() is called by the sistem release
2631  * callbacks to account for the time spent in the debugger. The value is then
2632  * accumulated in the lb_info structure and used by lbolt_event_driven() and
2633  * lbolt_cyclic_driven(), as well as the mdb_get_lbolt() routine.
2634  */
2635 void
lbolt_debug_entry(void)2636 lbolt_debug_entry(void)
2637 {
2638 	if (lbolt_hybrid != lbolt_bootstrap) {
2639 		ASSERT(lb_info != NULL);
2640 		lb_info->lbi_debug_ts = gethrtime();
2641 	}
2642 }
2643 
2644 /*
2645  * Calculate the time spent in the debugger and add it to the lbolt info
2646  * structure. We also update the internal lbolt value in case we were in
2647  * cyclic driven mode going in.
2648  */
2649 void
lbolt_debug_return(void)2650 lbolt_debug_return(void)
2651 {
2652 	hrtime_t ts;
2653 
2654 	if (lbolt_hybrid != lbolt_bootstrap) {
2655 		ASSERT(lb_info != NULL);
2656 		ASSERT(nsec_per_tick > 0);
2657 
2658 		ts = gethrtime();
2659 		lb_info->lbi_internal = (ts/nsec_per_tick);
2660 		lb_info->lbi_debug_time +=
2661 		    ((ts - lb_info->lbi_debug_ts)/nsec_per_tick);
2662 
2663 		lb_info->lbi_debug_ts = 0;
2664 	}
2665 }
2666