xref: /titanic_41/usr/src/uts/i86pc/os/timestamp.c (revision dd7afb26c5036958cddc0c2c1e499571664a6ed0)
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 
22 /*
23  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright 2012 Nexenta Systems, Inc. All rights reserved.
27  * Copyright 2016 Joyent, Inc.
28  */
29 
30 #include <sys/types.h>
31 #include <sys/param.h>
32 #include <sys/systm.h>
33 #include <sys/disp.h>
34 #include <sys/var.h>
35 #include <sys/cmn_err.h>
36 #include <sys/debug.h>
37 #include <sys/x86_archext.h>
38 #include <sys/archsystm.h>
39 #include <sys/cpuvar.h>
40 #include <sys/psm_defs.h>
41 #include <sys/clock.h>
42 #include <sys/atomic.h>
43 #include <sys/lockstat.h>
44 #include <sys/smp_impldefs.h>
45 #include <sys/dtrace.h>
46 #include <sys/time.h>
47 #include <sys/panic.h>
48 #include <sys/cpu.h>
49 #include <sys/comm_page.h>
50 
51 /*
52  * Using the Pentium's TSC register for gethrtime()
53  * ------------------------------------------------
54  *
55  * The Pentium family, like many chip architectures, has a high-resolution
56  * timestamp counter ("TSC") which increments once per CPU cycle.  The contents
57  * of the timestamp counter are read with the RDTSC instruction.
58  *
59  * As with its UltraSPARC equivalent (the %tick register), TSC's cycle count
60  * must be translated into nanoseconds in order to implement gethrtime().
61  * We avoid inducing floating point operations in this conversion by
62  * implementing the same nsec_scale algorithm as that found in the sun4u
63  * platform code.  The sun4u NATIVE_TIME_TO_NSEC_SCALE block comment contains
64  * a detailed description of the algorithm; the comment is not reproduced
65  * here.  This implementation differs only in its value for NSEC_SHIFT:
66  * we implement an NSEC_SHIFT of 5 (instead of sun4u's 4) to allow for
67  * 60 MHz Pentiums.
68  *
69  * While TSC and %tick are both cycle counting registers, TSC's functionality
70  * falls short in several critical ways:
71  *
72  *  (a)	TSCs on different CPUs are not guaranteed to be in sync.  While in
73  *	practice they often _are_ in sync, this isn't guaranteed by the
74  *	architecture.
75  *
76  *  (b)	The TSC cannot be reliably set to an arbitrary value.  The architecture
77  *	only supports writing the low 32-bits of TSC, making it impractical
78  *	to rewrite.
79  *
80  *  (c)	The architecture doesn't have the capacity to interrupt based on
81  *	arbitrary values of TSC; there is no TICK_CMPR equivalent.
82  *
83  * Together, (a) and (b) imply that software must track the skew between
84  * TSCs and account for it (it is assumed that while there may exist skew,
85  * there does not exist drift).  To determine the skew between CPUs, we
86  * have newly onlined CPUs call tsc_sync_slave(), while the CPU performing
87  * the online operation calls tsc_sync_master().
88  *
89  * In the absence of time-of-day clock adjustments, gethrtime() must stay in
90  * sync with gettimeofday().  This is problematic; given (c), the software
91  * cannot drive its time-of-day source from TSC, and yet they must somehow be
92  * kept in sync.  We implement this by having a routine, tsc_tick(), which
93  * is called once per second from the interrupt which drives time-of-day.
94  *
95  * Note that the hrtime base for gethrtime, tsc_hrtime_base, is modified
96  * atomically with nsec_scale under CLOCK_LOCK.  This assures that time
97  * monotonically increases.
98  */
99 
100 #define	NSEC_SHIFT 5
101 
102 static uint_t nsec_unscale;
103 
104 /*
105  * These two variables used to be grouped together inside of a structure that
106  * lived on a single cache line. A regression (bug ID 4623398) caused the
107  * compiler to emit code that "optimized" away the while-loops below. The
108  * result was that no synchronization between the onlining and onlined CPUs
109  * took place.
110  */
111 static volatile int tsc_ready;
112 static volatile int tsc_sync_go;
113 
114 /*
115  * Used as indices into the tsc_sync_snaps[] array.
116  */
117 #define	TSC_MASTER		0
118 #define	TSC_SLAVE		1
119 
120 /*
121  * Used in the tsc_master_sync()/tsc_slave_sync() rendezvous.
122  */
123 #define	TSC_SYNC_STOP		1
124 #define	TSC_SYNC_GO		2
125 #define	TSC_SYNC_DONE		3
126 #define	SYNC_ITERATIONS		10
127 
128 #define	TSC_CONVERT_AND_ADD(tsc, hrt, scale) {	 	\
129 	unsigned int *_l = (unsigned int *)&(tsc); 	\
130 	(hrt) += mul32(_l[1], scale) << NSEC_SHIFT; 	\
131 	(hrt) += mul32(_l[0], scale) >> (32 - NSEC_SHIFT); \
132 }
133 
134 #define	TSC_CONVERT(tsc, hrt, scale) { 			\
135 	unsigned int *_l = (unsigned int *)&(tsc); 	\
136 	(hrt) = mul32(_l[1], scale) << NSEC_SHIFT; 	\
137 	(hrt) += mul32(_l[0], scale) >> (32 - NSEC_SHIFT); \
138 }
139 
140 int tsc_master_slave_sync_needed = 1;
141 
142 typedef struct tsc_sync {
143 	volatile hrtime_t master_tsc, slave_tsc;
144 } tsc_sync_t;
145 static tsc_sync_t *tscp;
146 static hrtime_t largest_tsc_delta = 0;
147 static ulong_t shortest_write_time = ~0UL;
148 
149 static hrtime_t	tsc_last_jumped = 0;
150 static int	tsc_jumped = 0;
151 
152 static hrtime_t	shadow_tsc_hrtime_base;
153 static hrtime_t	shadow_tsc_last;
154 static uint_t	shadow_nsec_scale;
155 static uint32_t	shadow_hres_lock;
156 int get_tsc_ready();
157 
158 hrtime_t
tsc_gethrtime(void)159 tsc_gethrtime(void)
160 {
161 	uint32_t old_hres_lock;
162 	hrtime_t tsc, hrt;
163 
164 	do {
165 		old_hres_lock = hres_lock;
166 
167 		if ((tsc = tsc_read()) >= tsc_last) {
168 			/*
169 			 * It would seem to be obvious that this is true
170 			 * (that is, the past is less than the present),
171 			 * but it isn't true in the presence of suspend/resume
172 			 * cycles.  If we manage to call gethrtime()
173 			 * after a resume, but before the first call to
174 			 * tsc_tick(), we will see the jump.  In this case,
175 			 * we will simply use the value in TSC as the delta.
176 			 */
177 			tsc -= tsc_last;
178 		} else if (tsc >= tsc_last - 2*tsc_max_delta) {
179 			/*
180 			 * There is a chance that tsc_tick() has just run on
181 			 * another CPU, and we have drifted just enough so that
182 			 * we appear behind tsc_last.  In this case, force the
183 			 * delta to be zero.
184 			 */
185 			tsc = 0;
186 		}
187 
188 		hrt = tsc_hrtime_base;
189 
190 		TSC_CONVERT_AND_ADD(tsc, hrt, nsec_scale);
191 	} while ((old_hres_lock & ~1) != hres_lock);
192 
193 	return (hrt);
194 }
195 
196 hrtime_t
tsc_gethrtime_delta(void)197 tsc_gethrtime_delta(void)
198 {
199 	uint32_t old_hres_lock;
200 	hrtime_t tsc, hrt;
201 	ulong_t flags;
202 
203 	do {
204 		old_hres_lock = hres_lock;
205 
206 		/*
207 		 * We need to disable interrupts here to assure that we
208 		 * don't migrate between the call to tsc_read() and
209 		 * adding the CPU's TSC tick delta. Note that disabling
210 		 * and reenabling preemption is forbidden here because
211 		 * we may be in the middle of a fast trap. In the amd64
212 		 * kernel we cannot tolerate preemption during a fast
213 		 * trap. See _update_sregs().
214 		 */
215 
216 		flags = clear_int_flag();
217 		tsc = tsc_read() + tsc_sync_tick_delta[CPU->cpu_id];
218 		restore_int_flag(flags);
219 
220 		/* See comments in tsc_gethrtime() above */
221 
222 		if (tsc >= tsc_last) {
223 			tsc -= tsc_last;
224 		} else if (tsc >= tsc_last - 2 * tsc_max_delta) {
225 			tsc = 0;
226 		}
227 
228 		hrt = tsc_hrtime_base;
229 
230 		TSC_CONVERT_AND_ADD(tsc, hrt, nsec_scale);
231 	} while ((old_hres_lock & ~1) != hres_lock);
232 
233 	return (hrt);
234 }
235 
236 hrtime_t
tsc_gethrtime_tick_delta(void)237 tsc_gethrtime_tick_delta(void)
238 {
239 	hrtime_t hrt;
240 	ulong_t flags;
241 
242 	flags = clear_int_flag();
243 	hrt = tsc_sync_tick_delta[CPU->cpu_id];
244 	restore_int_flag(flags);
245 
246 	return (hrt);
247 }
248 
249 /*
250  * This is similar to the above, but it cannot actually spin on hres_lock.
251  * As a result, it caches all of the variables it needs; if the variables
252  * don't change, it's done.
253  */
254 hrtime_t
dtrace_gethrtime(void)255 dtrace_gethrtime(void)
256 {
257 	uint32_t old_hres_lock;
258 	hrtime_t tsc, hrt;
259 	ulong_t flags;
260 
261 	do {
262 		old_hres_lock = hres_lock;
263 
264 		/*
265 		 * Interrupts are disabled to ensure that the thread isn't
266 		 * migrated between the tsc_read() and adding the CPU's
267 		 * TSC tick delta.
268 		 */
269 		flags = clear_int_flag();
270 
271 		tsc = tsc_read();
272 
273 		if (gethrtimef == tsc_gethrtime_delta)
274 			tsc += tsc_sync_tick_delta[CPU->cpu_id];
275 
276 		restore_int_flag(flags);
277 
278 		/*
279 		 * See the comments in tsc_gethrtime(), above.
280 		 */
281 		if (tsc >= tsc_last)
282 			tsc -= tsc_last;
283 		else if (tsc >= tsc_last - 2*tsc_max_delta)
284 			tsc = 0;
285 
286 		hrt = tsc_hrtime_base;
287 
288 		TSC_CONVERT_AND_ADD(tsc, hrt, nsec_scale);
289 
290 		if ((old_hres_lock & ~1) == hres_lock)
291 			break;
292 
293 		/*
294 		 * If we're here, the clock lock is locked -- or it has been
295 		 * unlocked and locked since we looked.  This may be due to
296 		 * tsc_tick() running on another CPU -- or it may be because
297 		 * some code path has ended up in dtrace_probe() with
298 		 * CLOCK_LOCK held.  We'll try to determine that we're in
299 		 * the former case by taking another lap if the lock has
300 		 * changed since when we first looked at it.
301 		 */
302 		if (old_hres_lock != hres_lock)
303 			continue;
304 
305 		/*
306 		 * So the lock was and is locked.  We'll use the old data
307 		 * instead.
308 		 */
309 		old_hres_lock = shadow_hres_lock;
310 
311 		/*
312 		 * Again, disable interrupts to ensure that the thread
313 		 * isn't migrated between the tsc_read() and adding
314 		 * the CPU's TSC tick delta.
315 		 */
316 		flags = clear_int_flag();
317 
318 		tsc = tsc_read();
319 
320 		if (gethrtimef == tsc_gethrtime_delta)
321 			tsc += tsc_sync_tick_delta[CPU->cpu_id];
322 
323 		restore_int_flag(flags);
324 
325 		/*
326 		 * See the comments in tsc_gethrtime(), above.
327 		 */
328 		if (tsc >= shadow_tsc_last)
329 			tsc -= shadow_tsc_last;
330 		else if (tsc >= shadow_tsc_last - 2 * tsc_max_delta)
331 			tsc = 0;
332 
333 		hrt = shadow_tsc_hrtime_base;
334 
335 		TSC_CONVERT_AND_ADD(tsc, hrt, shadow_nsec_scale);
336 	} while ((old_hres_lock & ~1) != shadow_hres_lock);
337 
338 	return (hrt);
339 }
340 
341 hrtime_t
tsc_gethrtimeunscaled(void)342 tsc_gethrtimeunscaled(void)
343 {
344 	uint32_t old_hres_lock;
345 	hrtime_t tsc;
346 
347 	do {
348 		old_hres_lock = hres_lock;
349 
350 		/* See tsc_tick(). */
351 		tsc = tsc_read() + tsc_last_jumped;
352 	} while ((old_hres_lock & ~1) != hres_lock);
353 
354 	return (tsc);
355 }
356 
357 /*
358  * Convert a nanosecond based timestamp to tsc
359  */
360 uint64_t
tsc_unscalehrtime(hrtime_t nsec)361 tsc_unscalehrtime(hrtime_t nsec)
362 {
363 	hrtime_t tsc;
364 
365 	if (tsc_gethrtime_enable) {
366 		TSC_CONVERT(nsec, tsc, nsec_unscale);
367 		return (tsc);
368 	}
369 	return ((uint64_t)nsec);
370 }
371 
372 /* Convert a tsc timestamp to nanoseconds */
373 void
tsc_scalehrtime(hrtime_t * tsc)374 tsc_scalehrtime(hrtime_t *tsc)
375 {
376 	hrtime_t hrt;
377 	hrtime_t mytsc;
378 
379 	if (tsc == NULL)
380 		return;
381 	mytsc = *tsc;
382 
383 	TSC_CONVERT(mytsc, hrt, nsec_scale);
384 	*tsc  = hrt;
385 }
386 
387 hrtime_t
tsc_gethrtimeunscaled_delta(void)388 tsc_gethrtimeunscaled_delta(void)
389 {
390 	hrtime_t hrt;
391 	ulong_t flags;
392 
393 	/*
394 	 * Similarly to tsc_gethrtime_delta, we need to disable preemption
395 	 * to prevent migration between the call to tsc_gethrtimeunscaled
396 	 * and adding the CPU's hrtime delta. Note that disabling and
397 	 * reenabling preemption is forbidden here because we may be in the
398 	 * middle of a fast trap. In the amd64 kernel we cannot tolerate
399 	 * preemption during a fast trap. See _update_sregs().
400 	 */
401 
402 	flags = clear_int_flag();
403 	hrt = tsc_gethrtimeunscaled() + tsc_sync_tick_delta[CPU->cpu_id];
404 	restore_int_flag(flags);
405 
406 	return (hrt);
407 }
408 
409 /*
410  * Called by the master in the TSC sync operation (usually the boot CPU).
411  * If the slave is discovered to have a skew, gethrtimef will be changed to
412  * point to tsc_gethrtime_delta(). Calculating skews is precise only when
413  * the master and slave TSCs are read simultaneously; however, there is no
414  * algorithm that can read both CPUs in perfect simultaneity. The proposed
415  * algorithm is an approximate method based on the behaviour of cache
416  * management. The slave CPU continuously reads TSC and then reads a global
417  * variable which the master CPU updates. The moment the master's update reaches
418  * the slave's visibility (being forced by an mfence operation) we use the TSC
419  * reading taken on the slave. A corresponding TSC read will be taken on the
420  * master as soon as possible after finishing the mfence operation. But the
421  * delay between causing the slave to notice the invalid cache line and the
422  * competion of mfence is not repeatable. This error is heuristically assumed
423  * to be 1/4th of the total write time as being measured by the two TSC reads
424  * on the master sandwiching the mfence. Furthermore, due to the nature of
425  * bus arbitration, contention on memory bus, etc., the time taken for the write
426  * to reflect globally can vary a lot. So instead of taking a single reading,
427  * a set of readings are taken and the one with least write time is chosen
428  * to calculate the final skew.
429  *
430  * TSC sync is disabled in the context of virtualization because the CPUs
431  * assigned to the guest are virtual CPUs which means the real CPUs on which
432  * guest runs keep changing during life time of guest OS. So we would end up
433  * calculating TSC skews for a set of CPUs during boot whereas the guest
434  * might migrate to a different set of physical CPUs at a later point of
435  * time.
436  */
437 void
tsc_sync_master(processorid_t slave)438 tsc_sync_master(processorid_t slave)
439 {
440 	ulong_t flags, source, min_write_time = ~0UL;
441 	hrtime_t write_time, x, mtsc_after, tdelta;
442 	tsc_sync_t *tsc = tscp;
443 	int cnt;
444 	int hwtype;
445 
446 	hwtype = get_hwenv();
447 	if (!tsc_master_slave_sync_needed || (hwtype & HW_VIRTUAL) != 0)
448 		return;
449 
450 	flags = clear_int_flag();
451 	source = CPU->cpu_id;
452 
453 	for (cnt = 0; cnt < SYNC_ITERATIONS; cnt++) {
454 		while (tsc_sync_go != TSC_SYNC_GO)
455 			SMT_PAUSE();
456 
457 		tsc->master_tsc = tsc_read();
458 		membar_enter();
459 		mtsc_after = tsc_read();
460 		while (tsc_sync_go != TSC_SYNC_DONE)
461 			SMT_PAUSE();
462 		write_time =  mtsc_after - tsc->master_tsc;
463 		if (write_time <= min_write_time) {
464 			min_write_time = write_time;
465 			/*
466 			 * Apply heuristic adjustment only if the calculated
467 			 * delta is > 1/4th of the write time.
468 			 */
469 			x = tsc->slave_tsc - mtsc_after;
470 			if (x < 0)
471 				x = -x;
472 			if (x > (min_write_time/4))
473 				/*
474 				 * Subtract 1/4th of the measured write time
475 				 * from the master's TSC value, as an estimate
476 				 * of how late the mfence completion came
477 				 * after the slave noticed the cache line
478 				 * change.
479 				 */
480 				tdelta = tsc->slave_tsc -
481 				    (mtsc_after - (min_write_time/4));
482 			else
483 				tdelta = tsc->slave_tsc - mtsc_after;
484 			tsc_sync_tick_delta[slave] =
485 			    tsc_sync_tick_delta[source] - tdelta;
486 		}
487 
488 		tsc->master_tsc = tsc->slave_tsc = write_time = 0;
489 		membar_enter();
490 		tsc_sync_go = TSC_SYNC_STOP;
491 	}
492 	if (tdelta < 0)
493 		tdelta = -tdelta;
494 	if (tdelta > largest_tsc_delta)
495 		largest_tsc_delta = tdelta;
496 	if (min_write_time < shortest_write_time)
497 		shortest_write_time = min_write_time;
498 	/*
499 	 * Enable delta variants of tsc functions if the largest of all chosen
500 	 * deltas is > smallest of the write time.
501 	 */
502 	if (largest_tsc_delta > shortest_write_time) {
503 		gethrtimef = tsc_gethrtime_delta;
504 		gethrtimeunscaledf = tsc_gethrtimeunscaled_delta;
505 		tsc_ncpu = NCPU;
506 	}
507 	restore_int_flag(flags);
508 }
509 
510 /*
511  * Called by a CPU which has just been onlined.  It is expected that the CPU
512  * performing the online operation will call tsc_sync_master().
513  *
514  * TSC sync is disabled in the context of virtualization. See comments
515  * above tsc_sync_master.
516  */
517 void
tsc_sync_slave(void)518 tsc_sync_slave(void)
519 {
520 	ulong_t flags;
521 	hrtime_t s1;
522 	tsc_sync_t *tsc = tscp;
523 	int cnt;
524 	int hwtype;
525 
526 	hwtype = get_hwenv();
527 	if (!tsc_master_slave_sync_needed || (hwtype & HW_VIRTUAL) != 0)
528 		return;
529 
530 	flags = clear_int_flag();
531 
532 	for (cnt = 0; cnt < SYNC_ITERATIONS; cnt++) {
533 		/* Re-fill the cache line */
534 		s1 = tsc->master_tsc;
535 		membar_enter();
536 		tsc_sync_go = TSC_SYNC_GO;
537 		do {
538 			/*
539 			 * Do not put an SMT_PAUSE here. For instance,
540 			 * if the master and slave are really the same
541 			 * hyper-threaded CPU, then you want the master
542 			 * to yield to the slave as quickly as possible here,
543 			 * but not the other way.
544 			 */
545 			s1 = tsc_read();
546 		} while (tsc->master_tsc == 0);
547 		tsc->slave_tsc = s1;
548 		membar_enter();
549 		tsc_sync_go = TSC_SYNC_DONE;
550 
551 		while (tsc_sync_go != TSC_SYNC_STOP)
552 			SMT_PAUSE();
553 	}
554 
555 	restore_int_flag(flags);
556 }
557 
558 /*
559  * Called once per second on a CPU from the cyclic subsystem's
560  * CY_HIGH_LEVEL interrupt.  (No longer just cpu0-only)
561  */
562 void
tsc_tick(void)563 tsc_tick(void)
564 {
565 	hrtime_t now, delta;
566 	ushort_t spl;
567 
568 	/*
569 	 * Before we set the new variables, we set the shadow values.  This
570 	 * allows for lock free operation in dtrace_gethrtime().
571 	 */
572 	lock_set_spl((lock_t *)&shadow_hres_lock + HRES_LOCK_OFFSET,
573 	    ipltospl(CBE_HIGH_PIL), &spl);
574 
575 	shadow_tsc_hrtime_base = tsc_hrtime_base;
576 	shadow_tsc_last = tsc_last;
577 	shadow_nsec_scale = nsec_scale;
578 
579 	shadow_hres_lock++;
580 	splx(spl);
581 
582 	CLOCK_LOCK(&spl);
583 
584 	now = tsc_read();
585 
586 	if (gethrtimef == tsc_gethrtime_delta)
587 		now += tsc_sync_tick_delta[CPU->cpu_id];
588 
589 	if (now < tsc_last) {
590 		/*
591 		 * The TSC has just jumped into the past.  We assume that
592 		 * this is due to a suspend/resume cycle, and we're going
593 		 * to use the _current_ value of TSC as the delta.  This
594 		 * will keep tsc_hrtime_base correct.  We're also going to
595 		 * assume that rate of tsc does not change after a suspend
596 		 * resume (i.e nsec_scale remains the same).
597 		 */
598 		delta = now;
599 		tsc_last_jumped += tsc_last;
600 		tsc_jumped = 1;
601 	} else {
602 		/*
603 		 * Determine the number of TSC ticks since the last clock
604 		 * tick, and add that to the hrtime base.
605 		 */
606 		delta = now - tsc_last;
607 	}
608 
609 	TSC_CONVERT_AND_ADD(delta, tsc_hrtime_base, nsec_scale);
610 	tsc_last = now;
611 
612 	CLOCK_UNLOCK(spl);
613 }
614 
615 void
tsc_hrtimeinit(uint64_t cpu_freq_hz)616 tsc_hrtimeinit(uint64_t cpu_freq_hz)
617 {
618 	extern int gethrtime_hires;
619 	longlong_t tsc;
620 	ulong_t flags;
621 
622 	/*
623 	 * cpu_freq_hz is the measured cpu frequency in hertz
624 	 */
625 
626 	/*
627 	 * We can't accommodate CPUs slower than 31.25 MHz.
628 	 */
629 	ASSERT(cpu_freq_hz > NANOSEC / (1 << NSEC_SHIFT));
630 	nsec_scale =
631 	    (uint_t)(((uint64_t)NANOSEC << (32 - NSEC_SHIFT)) / cpu_freq_hz);
632 	nsec_unscale =
633 	    (uint_t)(((uint64_t)cpu_freq_hz << (32 - NSEC_SHIFT)) / NANOSEC);
634 
635 	flags = clear_int_flag();
636 	tsc = tsc_read();
637 	(void) tsc_gethrtime();
638 	tsc_max_delta = tsc_read() - tsc;
639 	restore_int_flag(flags);
640 	gethrtimef = tsc_gethrtime;
641 	gethrtimeunscaledf = tsc_gethrtimeunscaled;
642 	scalehrtimef = tsc_scalehrtime;
643 	unscalehrtimef = tsc_unscalehrtime;
644 	hrtime_tick = tsc_tick;
645 	gethrtime_hires = 1;
646 	/*
647 	 * Being part of the comm page, tsc_ncpu communicates the published
648 	 * length of the tsc_sync_tick_delta array.  This is kept zeroed to
649 	 * ignore the absent delta data while the TSCs are synced.
650 	 */
651 	tsc_ncpu = 0;
652 	/*
653 	 * Allocate memory for the structure used in the tsc sync logic.
654 	 * This structure should be aligned on a multiple of cache line size.
655 	 */
656 	tscp = kmem_zalloc(PAGESIZE, KM_SLEEP);
657 }
658 
659 int
get_tsc_ready()660 get_tsc_ready()
661 {
662 	return (tsc_ready);
663 }
664 
665 /*
666  * Adjust all the deltas by adding the passed value to the array.
667  * Then use the "delt" versions of the the gethrtime functions.
668  * Note that 'tdelta' _could_ be a negative number, which should
669  * reduce the values in the array (used, for example, if the Solaris
670  * instance was moved by a virtual manager to a machine with a higher
671  * value of tsc).
672  */
673 void
tsc_adjust_delta(hrtime_t tdelta)674 tsc_adjust_delta(hrtime_t tdelta)
675 {
676 	int		i;
677 
678 	for (i = 0; i < NCPU; i++) {
679 		tsc_sync_tick_delta[i] += tdelta;
680 	}
681 
682 	gethrtimef = tsc_gethrtime_delta;
683 	gethrtimeunscaledf = tsc_gethrtimeunscaled_delta;
684 	tsc_ncpu = NCPU;
685 }
686 
687 /*
688  * Functions to manage TSC and high-res time on suspend and resume.
689  */
690 
691 /*
692  * declarations needed for time adjustment
693  */
694 extern void	rtcsync(void);
695 extern tod_ops_t *tod_ops;
696 /* There must be a better way than exposing nsec_scale! */
697 extern uint_t	nsec_scale;
698 static uint64_t tsc_saved_tsc = 0; /* 1 in 2^64 chance this'll screw up! */
699 static timestruc_t tsc_saved_ts;
700 static int	tsc_needs_resume = 0;	/* We only want to do this once. */
701 int		tsc_delta_onsuspend = 0;
702 int		tsc_adjust_seconds = 1;
703 int		tsc_suspend_count = 0;
704 int		tsc_resume_in_cyclic = 0;
705 
706 /*
707  * Let timestamp.c know that we are suspending.  It needs to take
708  * snapshots of the current time, and do any pre-suspend work.
709  */
710 void
tsc_suspend(void)711 tsc_suspend(void)
712 {
713 /*
714  * What we need to do here, is to get the time we suspended, so that we
715  * know how much we should add to the resume.
716  * This routine is called by each CPU, so we need to handle reentry.
717  */
718 	if (tsc_gethrtime_enable) {
719 		/*
720 		 * We put the tsc_read() inside the lock as it
721 		 * as no locking constraints, and it puts the
722 		 * aquired value closer to the time stamp (in
723 		 * case we delay getting the lock).
724 		 */
725 		mutex_enter(&tod_lock);
726 		tsc_saved_tsc = tsc_read();
727 		tsc_saved_ts = TODOP_GET(tod_ops);
728 		mutex_exit(&tod_lock);
729 		/* We only want to do this once. */
730 		if (tsc_needs_resume == 0) {
731 			if (tsc_delta_onsuspend) {
732 				tsc_adjust_delta(tsc_saved_tsc);
733 			} else {
734 				tsc_adjust_delta(nsec_scale);
735 			}
736 			tsc_suspend_count++;
737 		}
738 	}
739 
740 	invalidate_cache();
741 	tsc_needs_resume = 1;
742 }
743 
744 /*
745  * Restore all timestamp state based on the snapshots taken at
746  * suspend time.
747  */
748 void
tsc_resume(void)749 tsc_resume(void)
750 {
751 	/*
752 	 * We only need to (and want to) do this once.  So let the first
753 	 * caller handle this (we are locked by the cpu lock), as it
754 	 * is preferential that we get the earliest sync.
755 	 */
756 	if (tsc_needs_resume) {
757 		/*
758 		 * If using the TSC, adjust the delta based on how long
759 		 * we were sleeping (or away).  We also adjust for
760 		 * migration and a grown TSC.
761 		 */
762 		if (tsc_saved_tsc != 0) {
763 			timestruc_t	ts;
764 			hrtime_t	now, sleep_tsc = 0;
765 			int		sleep_sec;
766 			extern void	tsc_tick(void);
767 			extern uint64_t cpu_freq_hz;
768 
769 			/* tsc_read() MUST be before TODOP_GET() */
770 			mutex_enter(&tod_lock);
771 			now = tsc_read();
772 			ts = TODOP_GET(tod_ops);
773 			mutex_exit(&tod_lock);
774 
775 			/* Compute seconds of sleep time */
776 			sleep_sec = ts.tv_sec - tsc_saved_ts.tv_sec;
777 
778 			/*
779 			 * If the saved sec is less that or equal to
780 			 * the current ts, then there is likely a
781 			 * problem with the clock.  Assume at least
782 			 * one second has passed, so that time goes forward.
783 			 */
784 			if (sleep_sec <= 0) {
785 				sleep_sec = 1;
786 			}
787 
788 			/* How many TSC's should have occured while sleeping */
789 			if (tsc_adjust_seconds)
790 				sleep_tsc = sleep_sec * cpu_freq_hz;
791 
792 			/*
793 			 * We also want to subtract from the "sleep_tsc"
794 			 * the current value of tsc_read(), so that our
795 			 * adjustment accounts for the amount of time we
796 			 * have been resumed _or_ an adjustment based on
797 			 * the fact that we didn't actually power off the
798 			 * CPU (migration is another issue, but _should_
799 			 * also comply with this calculation).  If the CPU
800 			 * never powered off, then:
801 			 *    'now == sleep_tsc + saved_tsc'
802 			 * and the delta will effectively be "0".
803 			 */
804 			sleep_tsc -= now;
805 			if (tsc_delta_onsuspend) {
806 				tsc_adjust_delta(sleep_tsc);
807 			} else {
808 				tsc_adjust_delta(tsc_saved_tsc + sleep_tsc);
809 			}
810 			tsc_saved_tsc = 0;
811 
812 			tsc_tick();
813 		}
814 		tsc_needs_resume = 0;
815 	}
816 
817 }
818