xref: /linux/arch/x86/kernel/tsc.c (revision ac6a0cf6716bb46813d0161024c66c2af66e53d1)
1 #include <linux/kernel.h>
2 #include <linux/sched.h>
3 #include <linux/init.h>
4 #include <linux/module.h>
5 #include <linux/timer.h>
6 #include <linux/acpi_pmtmr.h>
7 #include <linux/cpufreq.h>
8 #include <linux/dmi.h>
9 #include <linux/delay.h>
10 #include <linux/clocksource.h>
11 #include <linux/percpu.h>
12 #include <linux/timex.h>
13 
14 #include <asm/hpet.h>
15 #include <asm/timer.h>
16 #include <asm/vgtod.h>
17 #include <asm/time.h>
18 #include <asm/delay.h>
19 #include <asm/hypervisor.h>
20 
21 unsigned int __read_mostly cpu_khz;	/* TSC clocks / usec, not used here */
22 EXPORT_SYMBOL(cpu_khz);
23 
24 unsigned int __read_mostly tsc_khz;
25 EXPORT_SYMBOL(tsc_khz);
26 
27 /*
28  * TSC can be unstable due to cpufreq or due to unsynced TSCs
29  */
30 static int __read_mostly tsc_unstable;
31 
32 /* native_sched_clock() is called before tsc_init(), so
33    we must start with the TSC soft disabled to prevent
34    erroneous rdtsc usage on !cpu_has_tsc processors */
35 static int __read_mostly tsc_disabled = -1;
36 
37 static int tsc_clocksource_reliable;
38 /*
39  * Scheduler clock - returns current time in nanosec units.
40  */
41 u64 native_sched_clock(void)
42 {
43 	u64 this_offset;
44 
45 	/*
46 	 * Fall back to jiffies if there's no TSC available:
47 	 * ( But note that we still use it if the TSC is marked
48 	 *   unstable. We do this because unlike Time Of Day,
49 	 *   the scheduler clock tolerates small errors and it's
50 	 *   very important for it to be as fast as the platform
51 	 *   can achive it. )
52 	 */
53 	if (unlikely(tsc_disabled)) {
54 		/* No locking but a rare wrong value is not a big deal: */
55 		return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
56 	}
57 
58 	/* read the Time Stamp Counter: */
59 	rdtscll(this_offset);
60 
61 	/* return the value in ns */
62 	return __cycles_2_ns(this_offset);
63 }
64 
65 /* We need to define a real function for sched_clock, to override the
66    weak default version */
67 #ifdef CONFIG_PARAVIRT
68 unsigned long long sched_clock(void)
69 {
70 	return paravirt_sched_clock();
71 }
72 #else
73 unsigned long long
74 sched_clock(void) __attribute__((alias("native_sched_clock")));
75 #endif
76 
77 int check_tsc_unstable(void)
78 {
79 	return tsc_unstable;
80 }
81 EXPORT_SYMBOL_GPL(check_tsc_unstable);
82 
83 #ifdef CONFIG_X86_TSC
84 int __init notsc_setup(char *str)
85 {
86 	printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, "
87 			"cannot disable TSC completely.\n");
88 	tsc_disabled = 1;
89 	return 1;
90 }
91 #else
92 /*
93  * disable flag for tsc. Takes effect by clearing the TSC cpu flag
94  * in cpu/common.c
95  */
96 int __init notsc_setup(char *str)
97 {
98 	setup_clear_cpu_cap(X86_FEATURE_TSC);
99 	return 1;
100 }
101 #endif
102 
103 __setup("notsc", notsc_setup);
104 
105 static int __init tsc_setup(char *str)
106 {
107 	if (!strcmp(str, "reliable"))
108 		tsc_clocksource_reliable = 1;
109 	return 1;
110 }
111 
112 __setup("tsc=", tsc_setup);
113 
114 #define MAX_RETRIES     5
115 #define SMI_TRESHOLD    50000
116 
117 /*
118  * Read TSC and the reference counters. Take care of SMI disturbance
119  */
120 static u64 tsc_read_refs(u64 *p, int hpet)
121 {
122 	u64 t1, t2;
123 	int i;
124 
125 	for (i = 0; i < MAX_RETRIES; i++) {
126 		t1 = get_cycles();
127 		if (hpet)
128 			*p = hpet_readl(HPET_COUNTER) & 0xFFFFFFFF;
129 		else
130 			*p = acpi_pm_read_early();
131 		t2 = get_cycles();
132 		if ((t2 - t1) < SMI_TRESHOLD)
133 			return t2;
134 	}
135 	return ULLONG_MAX;
136 }
137 
138 /*
139  * Calculate the TSC frequency from HPET reference
140  */
141 static unsigned long calc_hpet_ref(u64 deltatsc, u64 hpet1, u64 hpet2)
142 {
143 	u64 tmp;
144 
145 	if (hpet2 < hpet1)
146 		hpet2 += 0x100000000ULL;
147 	hpet2 -= hpet1;
148 	tmp = ((u64)hpet2 * hpet_readl(HPET_PERIOD));
149 	do_div(tmp, 1000000);
150 	do_div(deltatsc, tmp);
151 
152 	return (unsigned long) deltatsc;
153 }
154 
155 /*
156  * Calculate the TSC frequency from PMTimer reference
157  */
158 static unsigned long calc_pmtimer_ref(u64 deltatsc, u64 pm1, u64 pm2)
159 {
160 	u64 tmp;
161 
162 	if (!pm1 && !pm2)
163 		return ULONG_MAX;
164 
165 	if (pm2 < pm1)
166 		pm2 += (u64)ACPI_PM_OVRRUN;
167 	pm2 -= pm1;
168 	tmp = pm2 * 1000000000LL;
169 	do_div(tmp, PMTMR_TICKS_PER_SEC);
170 	do_div(deltatsc, tmp);
171 
172 	return (unsigned long) deltatsc;
173 }
174 
175 #define CAL_MS		10
176 #define CAL_LATCH	(CLOCK_TICK_RATE / (1000 / CAL_MS))
177 #define CAL_PIT_LOOPS	1000
178 
179 #define CAL2_MS		50
180 #define CAL2_LATCH	(CLOCK_TICK_RATE / (1000 / CAL2_MS))
181 #define CAL2_PIT_LOOPS	5000
182 
183 
184 /*
185  * Try to calibrate the TSC against the Programmable
186  * Interrupt Timer and return the frequency of the TSC
187  * in kHz.
188  *
189  * Return ULONG_MAX on failure to calibrate.
190  */
191 static unsigned long pit_calibrate_tsc(u32 latch, unsigned long ms, int loopmin)
192 {
193 	u64 tsc, t1, t2, delta;
194 	unsigned long tscmin, tscmax;
195 	int pitcnt;
196 
197 	/* Set the Gate high, disable speaker */
198 	outb((inb(0x61) & ~0x02) | 0x01, 0x61);
199 
200 	/*
201 	 * Setup CTC channel 2* for mode 0, (interrupt on terminal
202 	 * count mode), binary count. Set the latch register to 50ms
203 	 * (LSB then MSB) to begin countdown.
204 	 */
205 	outb(0xb0, 0x43);
206 	outb(latch & 0xff, 0x42);
207 	outb(latch >> 8, 0x42);
208 
209 	tsc = t1 = t2 = get_cycles();
210 
211 	pitcnt = 0;
212 	tscmax = 0;
213 	tscmin = ULONG_MAX;
214 	while ((inb(0x61) & 0x20) == 0) {
215 		t2 = get_cycles();
216 		delta = t2 - tsc;
217 		tsc = t2;
218 		if ((unsigned long) delta < tscmin)
219 			tscmin = (unsigned int) delta;
220 		if ((unsigned long) delta > tscmax)
221 			tscmax = (unsigned int) delta;
222 		pitcnt++;
223 	}
224 
225 	/*
226 	 * Sanity checks:
227 	 *
228 	 * If we were not able to read the PIT more than loopmin
229 	 * times, then we have been hit by a massive SMI
230 	 *
231 	 * If the maximum is 10 times larger than the minimum,
232 	 * then we got hit by an SMI as well.
233 	 */
234 	if (pitcnt < loopmin || tscmax > 10 * tscmin)
235 		return ULONG_MAX;
236 
237 	/* Calculate the PIT value */
238 	delta = t2 - t1;
239 	do_div(delta, ms);
240 	return delta;
241 }
242 
243 /*
244  * This reads the current MSB of the PIT counter, and
245  * checks if we are running on sufficiently fast and
246  * non-virtualized hardware.
247  *
248  * Our expectations are:
249  *
250  *  - the PIT is running at roughly 1.19MHz
251  *
252  *  - each IO is going to take about 1us on real hardware,
253  *    but we allow it to be much faster (by a factor of 10) or
254  *    _slightly_ slower (ie we allow up to a 2us read+counter
255  *    update - anything else implies a unacceptably slow CPU
256  *    or PIT for the fast calibration to work.
257  *
258  *  - with 256 PIT ticks to read the value, we have 214us to
259  *    see the same MSB (and overhead like doing a single TSC
260  *    read per MSB value etc).
261  *
262  *  - We're doing 2 reads per loop (LSB, MSB), and we expect
263  *    them each to take about a microsecond on real hardware.
264  *    So we expect a count value of around 100. But we'll be
265  *    generous, and accept anything over 50.
266  *
267  *  - if the PIT is stuck, and we see *many* more reads, we
268  *    return early (and the next caller of pit_expect_msb()
269  *    then consider it a failure when they don't see the
270  *    next expected value).
271  *
272  * These expectations mean that we know that we have seen the
273  * transition from one expected value to another with a fairly
274  * high accuracy, and we didn't miss any events. We can thus
275  * use the TSC value at the transitions to calculate a pretty
276  * good value for the TSC frequencty.
277  */
278 static inline int pit_verify_msb(unsigned char val)
279 {
280 	/* Ignore LSB */
281 	inb(0x42);
282 	return inb(0x42) == val;
283 }
284 
285 static inline int pit_expect_msb(unsigned char val, u64 *tscp, unsigned long *deltap)
286 {
287 	int count;
288 	u64 tsc = 0;
289 
290 	for (count = 0; count < 50000; count++) {
291 		if (!pit_verify_msb(val))
292 			break;
293 		tsc = get_cycles();
294 	}
295 	*deltap = get_cycles() - tsc;
296 	*tscp = tsc;
297 
298 	/*
299 	 * We require _some_ success, but the quality control
300 	 * will be based on the error terms on the TSC values.
301 	 */
302 	return count > 5;
303 }
304 
305 /*
306  * How many MSB values do we want to see? We aim for
307  * a maximum error rate of 500ppm (in practice the
308  * real error is much smaller), but refuse to spend
309  * more than 25ms on it.
310  */
311 #define MAX_QUICK_PIT_MS 25
312 #define MAX_QUICK_PIT_ITERATIONS (MAX_QUICK_PIT_MS * PIT_TICK_RATE / 1000 / 256)
313 
314 static unsigned long quick_pit_calibrate(void)
315 {
316 	int i;
317 	u64 tsc, delta;
318 	unsigned long d1, d2;
319 
320 	/* Set the Gate high, disable speaker */
321 	outb((inb(0x61) & ~0x02) | 0x01, 0x61);
322 
323 	/*
324 	 * Counter 2, mode 0 (one-shot), binary count
325 	 *
326 	 * NOTE! Mode 2 decrements by two (and then the
327 	 * output is flipped each time, giving the same
328 	 * final output frequency as a decrement-by-one),
329 	 * so mode 0 is much better when looking at the
330 	 * individual counts.
331 	 */
332 	outb(0xb0, 0x43);
333 
334 	/* Start at 0xffff */
335 	outb(0xff, 0x42);
336 	outb(0xff, 0x42);
337 
338 	/*
339 	 * The PIT starts counting at the next edge, so we
340 	 * need to delay for a microsecond. The easiest way
341 	 * to do that is to just read back the 16-bit counter
342 	 * once from the PIT.
343 	 */
344 	pit_verify_msb(0);
345 
346 	if (pit_expect_msb(0xff, &tsc, &d1)) {
347 		for (i = 1; i <= MAX_QUICK_PIT_ITERATIONS; i++) {
348 			if (!pit_expect_msb(0xff-i, &delta, &d2))
349 				break;
350 
351 			/*
352 			 * Iterate until the error is less than 500 ppm
353 			 */
354 			delta -= tsc;
355 			if (d1+d2 >= delta >> 11)
356 				continue;
357 
358 			/*
359 			 * Check the PIT one more time to verify that
360 			 * all TSC reads were stable wrt the PIT.
361 			 *
362 			 * This also guarantees serialization of the
363 			 * last cycle read ('d2') in pit_expect_msb.
364 			 */
365 			if (!pit_verify_msb(0xfe - i))
366 				break;
367 			goto success;
368 		}
369 	}
370 	printk("Fast TSC calibration failed\n");
371 	return 0;
372 
373 success:
374 	/*
375 	 * Ok, if we get here, then we've seen the
376 	 * MSB of the PIT decrement 'i' times, and the
377 	 * error has shrunk to less than 500 ppm.
378 	 *
379 	 * As a result, we can depend on there not being
380 	 * any odd delays anywhere, and the TSC reads are
381 	 * reliable (within the error). We also adjust the
382 	 * delta to the middle of the error bars, just
383 	 * because it looks nicer.
384 	 *
385 	 * kHz = ticks / time-in-seconds / 1000;
386 	 * kHz = (t2 - t1) / (I * 256 / PIT_TICK_RATE) / 1000
387 	 * kHz = ((t2 - t1) * PIT_TICK_RATE) / (I * 256 * 1000)
388 	 */
389 	delta += (long)(d2 - d1)/2;
390 	delta *= PIT_TICK_RATE;
391 	do_div(delta, i*256*1000);
392 	printk("Fast TSC calibration using PIT\n");
393 	return delta;
394 }
395 
396 /**
397  * native_calibrate_tsc - calibrate the tsc on boot
398  */
399 unsigned long native_calibrate_tsc(void)
400 {
401 	u64 tsc1, tsc2, delta, ref1, ref2;
402 	unsigned long tsc_pit_min = ULONG_MAX, tsc_ref_min = ULONG_MAX;
403 	unsigned long flags, latch, ms, fast_calibrate, hv_tsc_khz;
404 	int hpet = is_hpet_enabled(), i, loopmin;
405 
406 	hv_tsc_khz = get_hypervisor_tsc_freq();
407 	if (hv_tsc_khz) {
408 		printk(KERN_INFO "TSC: Frequency read from the hypervisor\n");
409 		return hv_tsc_khz;
410 	}
411 
412 	local_irq_save(flags);
413 	fast_calibrate = quick_pit_calibrate();
414 	local_irq_restore(flags);
415 	if (fast_calibrate)
416 		return fast_calibrate;
417 
418 	/*
419 	 * Run 5 calibration loops to get the lowest frequency value
420 	 * (the best estimate). We use two different calibration modes
421 	 * here:
422 	 *
423 	 * 1) PIT loop. We set the PIT Channel 2 to oneshot mode and
424 	 * load a timeout of 50ms. We read the time right after we
425 	 * started the timer and wait until the PIT count down reaches
426 	 * zero. In each wait loop iteration we read the TSC and check
427 	 * the delta to the previous read. We keep track of the min
428 	 * and max values of that delta. The delta is mostly defined
429 	 * by the IO time of the PIT access, so we can detect when a
430 	 * SMI/SMM disturbance happend between the two reads. If the
431 	 * maximum time is significantly larger than the minimum time,
432 	 * then we discard the result and have another try.
433 	 *
434 	 * 2) Reference counter. If available we use the HPET or the
435 	 * PMTIMER as a reference to check the sanity of that value.
436 	 * We use separate TSC readouts and check inside of the
437 	 * reference read for a SMI/SMM disturbance. We dicard
438 	 * disturbed values here as well. We do that around the PIT
439 	 * calibration delay loop as we have to wait for a certain
440 	 * amount of time anyway.
441 	 */
442 
443 	/* Preset PIT loop values */
444 	latch = CAL_LATCH;
445 	ms = CAL_MS;
446 	loopmin = CAL_PIT_LOOPS;
447 
448 	for (i = 0; i < 3; i++) {
449 		unsigned long tsc_pit_khz;
450 
451 		/*
452 		 * Read the start value and the reference count of
453 		 * hpet/pmtimer when available. Then do the PIT
454 		 * calibration, which will take at least 50ms, and
455 		 * read the end value.
456 		 */
457 		local_irq_save(flags);
458 		tsc1 = tsc_read_refs(&ref1, hpet);
459 		tsc_pit_khz = pit_calibrate_tsc(latch, ms, loopmin);
460 		tsc2 = tsc_read_refs(&ref2, hpet);
461 		local_irq_restore(flags);
462 
463 		/* Pick the lowest PIT TSC calibration so far */
464 		tsc_pit_min = min(tsc_pit_min, tsc_pit_khz);
465 
466 		/* hpet or pmtimer available ? */
467 		if (!hpet && !ref1 && !ref2)
468 			continue;
469 
470 		/* Check, whether the sampling was disturbed by an SMI */
471 		if (tsc1 == ULLONG_MAX || tsc2 == ULLONG_MAX)
472 			continue;
473 
474 		tsc2 = (tsc2 - tsc1) * 1000000LL;
475 		if (hpet)
476 			tsc2 = calc_hpet_ref(tsc2, ref1, ref2);
477 		else
478 			tsc2 = calc_pmtimer_ref(tsc2, ref1, ref2);
479 
480 		tsc_ref_min = min(tsc_ref_min, (unsigned long) tsc2);
481 
482 		/* Check the reference deviation */
483 		delta = ((u64) tsc_pit_min) * 100;
484 		do_div(delta, tsc_ref_min);
485 
486 		/*
487 		 * If both calibration results are inside a 10% window
488 		 * then we can be sure, that the calibration
489 		 * succeeded. We break out of the loop right away. We
490 		 * use the reference value, as it is more precise.
491 		 */
492 		if (delta >= 90 && delta <= 110) {
493 			printk(KERN_INFO
494 			       "TSC: PIT calibration matches %s. %d loops\n",
495 			       hpet ? "HPET" : "PMTIMER", i + 1);
496 			return tsc_ref_min;
497 		}
498 
499 		/*
500 		 * Check whether PIT failed more than once. This
501 		 * happens in virtualized environments. We need to
502 		 * give the virtual PC a slightly longer timeframe for
503 		 * the HPET/PMTIMER to make the result precise.
504 		 */
505 		if (i == 1 && tsc_pit_min == ULONG_MAX) {
506 			latch = CAL2_LATCH;
507 			ms = CAL2_MS;
508 			loopmin = CAL2_PIT_LOOPS;
509 		}
510 	}
511 
512 	/*
513 	 * Now check the results.
514 	 */
515 	if (tsc_pit_min == ULONG_MAX) {
516 		/* PIT gave no useful value */
517 		printk(KERN_WARNING "TSC: Unable to calibrate against PIT\n");
518 
519 		/* We don't have an alternative source, disable TSC */
520 		if (!hpet && !ref1 && !ref2) {
521 			printk("TSC: No reference (HPET/PMTIMER) available\n");
522 			return 0;
523 		}
524 
525 		/* The alternative source failed as well, disable TSC */
526 		if (tsc_ref_min == ULONG_MAX) {
527 			printk(KERN_WARNING "TSC: HPET/PMTIMER calibration "
528 			       "failed.\n");
529 			return 0;
530 		}
531 
532 		/* Use the alternative source */
533 		printk(KERN_INFO "TSC: using %s reference calibration\n",
534 		       hpet ? "HPET" : "PMTIMER");
535 
536 		return tsc_ref_min;
537 	}
538 
539 	/* We don't have an alternative source, use the PIT calibration value */
540 	if (!hpet && !ref1 && !ref2) {
541 		printk(KERN_INFO "TSC: Using PIT calibration value\n");
542 		return tsc_pit_min;
543 	}
544 
545 	/* The alternative source failed, use the PIT calibration value */
546 	if (tsc_ref_min == ULONG_MAX) {
547 		printk(KERN_WARNING "TSC: HPET/PMTIMER calibration failed. "
548 		       "Using PIT calibration\n");
549 		return tsc_pit_min;
550 	}
551 
552 	/*
553 	 * The calibration values differ too much. In doubt, we use
554 	 * the PIT value as we know that there are PMTIMERs around
555 	 * running at double speed. At least we let the user know:
556 	 */
557 	printk(KERN_WARNING "TSC: PIT calibration deviates from %s: %lu %lu.\n",
558 	       hpet ? "HPET" : "PMTIMER", tsc_pit_min, tsc_ref_min);
559 	printk(KERN_INFO "TSC: Using PIT calibration value\n");
560 	return tsc_pit_min;
561 }
562 
563 int recalibrate_cpu_khz(void)
564 {
565 #ifndef CONFIG_SMP
566 	unsigned long cpu_khz_old = cpu_khz;
567 
568 	if (cpu_has_tsc) {
569 		tsc_khz = calibrate_tsc();
570 		cpu_khz = tsc_khz;
571 		cpu_data(0).loops_per_jiffy =
572 			cpufreq_scale(cpu_data(0).loops_per_jiffy,
573 					cpu_khz_old, cpu_khz);
574 		return 0;
575 	} else
576 		return -ENODEV;
577 #else
578 	return -ENODEV;
579 #endif
580 }
581 
582 EXPORT_SYMBOL(recalibrate_cpu_khz);
583 
584 
585 /* Accelerators for sched_clock()
586  * convert from cycles(64bits) => nanoseconds (64bits)
587  *  basic equation:
588  *              ns = cycles / (freq / ns_per_sec)
589  *              ns = cycles * (ns_per_sec / freq)
590  *              ns = cycles * (10^9 / (cpu_khz * 10^3))
591  *              ns = cycles * (10^6 / cpu_khz)
592  *
593  *      Then we use scaling math (suggested by george@mvista.com) to get:
594  *              ns = cycles * (10^6 * SC / cpu_khz) / SC
595  *              ns = cycles * cyc2ns_scale / SC
596  *
597  *      And since SC is a constant power of two, we can convert the div
598  *  into a shift.
599  *
600  *  We can use khz divisor instead of mhz to keep a better precision, since
601  *  cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
602  *  (mathieu.desnoyers@polymtl.ca)
603  *
604  *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
605  */
606 
607 DEFINE_PER_CPU(unsigned long, cyc2ns);
608 DEFINE_PER_CPU(unsigned long long, cyc2ns_offset);
609 
610 static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu)
611 {
612 	unsigned long long tsc_now, ns_now, *offset;
613 	unsigned long flags, *scale;
614 
615 	local_irq_save(flags);
616 	sched_clock_idle_sleep_event();
617 
618 	scale = &per_cpu(cyc2ns, cpu);
619 	offset = &per_cpu(cyc2ns_offset, cpu);
620 
621 	rdtscll(tsc_now);
622 	ns_now = __cycles_2_ns(tsc_now);
623 
624 	if (cpu_khz) {
625 		*scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz;
626 		*offset = ns_now - (tsc_now * *scale >> CYC2NS_SCALE_FACTOR);
627 	}
628 
629 	sched_clock_idle_wakeup_event(0);
630 	local_irq_restore(flags);
631 }
632 
633 #ifdef CONFIG_CPU_FREQ
634 
635 /* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
636  * changes.
637  *
638  * RED-PEN: On SMP we assume all CPUs run with the same frequency.  It's
639  * not that important because current Opteron setups do not support
640  * scaling on SMP anyroads.
641  *
642  * Should fix up last_tsc too. Currently gettimeofday in the
643  * first tick after the change will be slightly wrong.
644  */
645 
646 static unsigned int  ref_freq;
647 static unsigned long loops_per_jiffy_ref;
648 static unsigned long tsc_khz_ref;
649 
650 static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
651 				void *data)
652 {
653 	struct cpufreq_freqs *freq = data;
654 	unsigned long *lpj;
655 
656 	if (cpu_has(&cpu_data(freq->cpu), X86_FEATURE_CONSTANT_TSC))
657 		return 0;
658 
659 	lpj = &boot_cpu_data.loops_per_jiffy;
660 #ifdef CONFIG_SMP
661 	if (!(freq->flags & CPUFREQ_CONST_LOOPS))
662 		lpj = &cpu_data(freq->cpu).loops_per_jiffy;
663 #endif
664 
665 	if (!ref_freq) {
666 		ref_freq = freq->old;
667 		loops_per_jiffy_ref = *lpj;
668 		tsc_khz_ref = tsc_khz;
669 	}
670 	if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
671 			(val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
672 			(val == CPUFREQ_RESUMECHANGE)) {
673 		*lpj = 	cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
674 
675 		tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
676 		if (!(freq->flags & CPUFREQ_CONST_LOOPS))
677 			mark_tsc_unstable("cpufreq changes");
678 	}
679 
680 	set_cyc2ns_scale(tsc_khz, freq->cpu);
681 
682 	return 0;
683 }
684 
685 static struct notifier_block time_cpufreq_notifier_block = {
686 	.notifier_call  = time_cpufreq_notifier
687 };
688 
689 static int __init cpufreq_tsc(void)
690 {
691 	if (!cpu_has_tsc)
692 		return 0;
693 	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
694 		return 0;
695 	cpufreq_register_notifier(&time_cpufreq_notifier_block,
696 				CPUFREQ_TRANSITION_NOTIFIER);
697 	return 0;
698 }
699 
700 core_initcall(cpufreq_tsc);
701 
702 #endif /* CONFIG_CPU_FREQ */
703 
704 /* clocksource code */
705 
706 static struct clocksource clocksource_tsc;
707 
708 /*
709  * We compare the TSC to the cycle_last value in the clocksource
710  * structure to avoid a nasty time-warp. This can be observed in a
711  * very small window right after one CPU updated cycle_last under
712  * xtime/vsyscall_gtod lock and the other CPU reads a TSC value which
713  * is smaller than the cycle_last reference value due to a TSC which
714  * is slighty behind. This delta is nowhere else observable, but in
715  * that case it results in a forward time jump in the range of hours
716  * due to the unsigned delta calculation of the time keeping core
717  * code, which is necessary to support wrapping clocksources like pm
718  * timer.
719  */
720 static cycle_t read_tsc(struct clocksource *cs)
721 {
722 	cycle_t ret = (cycle_t)get_cycles();
723 
724 	return ret >= clocksource_tsc.cycle_last ?
725 		ret : clocksource_tsc.cycle_last;
726 }
727 
728 #ifdef CONFIG_X86_64
729 static cycle_t __vsyscall_fn vread_tsc(void)
730 {
731 	cycle_t ret;
732 
733 	/*
734 	 * Surround the RDTSC by barriers, to make sure it's not
735 	 * speculated to outside the seqlock critical section and
736 	 * does not cause time warps:
737 	 */
738 	rdtsc_barrier();
739 	ret = (cycle_t)vget_cycles();
740 	rdtsc_barrier();
741 
742 	return ret >= __vsyscall_gtod_data.clock.cycle_last ?
743 		ret : __vsyscall_gtod_data.clock.cycle_last;
744 }
745 #endif
746 
747 static struct clocksource clocksource_tsc = {
748 	.name                   = "tsc",
749 	.rating                 = 300,
750 	.read                   = read_tsc,
751 	.mask                   = CLOCKSOURCE_MASK(64),
752 	.shift                  = 22,
753 	.flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
754 				  CLOCK_SOURCE_MUST_VERIFY,
755 #ifdef CONFIG_X86_64
756 	.vread                  = vread_tsc,
757 #endif
758 };
759 
760 void mark_tsc_unstable(char *reason)
761 {
762 	if (!tsc_unstable) {
763 		tsc_unstable = 1;
764 		printk("Marking TSC unstable due to %s\n", reason);
765 		/* Change only the rating, when not registered */
766 		if (clocksource_tsc.mult)
767 			clocksource_change_rating(&clocksource_tsc, 0);
768 		else
769 			clocksource_tsc.rating = 0;
770 	}
771 }
772 
773 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
774 
775 static int __init dmi_mark_tsc_unstable(const struct dmi_system_id *d)
776 {
777 	printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
778 			d->ident);
779 	tsc_unstable = 1;
780 	return 0;
781 }
782 
783 /* List of systems that have known TSC problems */
784 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
785 	{
786 		.callback = dmi_mark_tsc_unstable,
787 		.ident = "IBM Thinkpad 380XD",
788 		.matches = {
789 			DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
790 			DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
791 		},
792 	},
793 	{}
794 };
795 
796 static void __init check_system_tsc_reliable(void)
797 {
798 #ifdef CONFIG_MGEODE_LX
799 	/* RTSC counts during suspend */
800 #define RTSC_SUSP 0x100
801 	unsigned long res_low, res_high;
802 
803 	rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
804 	/* Geode_LX - the OLPC CPU has a possibly a very reliable TSC */
805 	if (res_low & RTSC_SUSP)
806 		tsc_clocksource_reliable = 1;
807 #endif
808 	if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE))
809 		tsc_clocksource_reliable = 1;
810 }
811 
812 /*
813  * Make an educated guess if the TSC is trustworthy and synchronized
814  * over all CPUs.
815  */
816 __cpuinit int unsynchronized_tsc(void)
817 {
818 	if (!cpu_has_tsc || tsc_unstable)
819 		return 1;
820 
821 #ifdef CONFIG_SMP
822 	if (apic_is_clustered_box())
823 		return 1;
824 #endif
825 
826 	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
827 		return 0;
828 	/*
829 	 * Intel systems are normally all synchronized.
830 	 * Exceptions must mark TSC as unstable:
831 	 */
832 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
833 		/* assume multi socket systems are not synchronized: */
834 		if (num_possible_cpus() > 1)
835 			tsc_unstable = 1;
836 	}
837 
838 	return tsc_unstable;
839 }
840 
841 static void __init init_tsc_clocksource(void)
842 {
843 	clocksource_tsc.mult = clocksource_khz2mult(tsc_khz,
844 			clocksource_tsc.shift);
845 	if (tsc_clocksource_reliable)
846 		clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
847 	/* lower the rating if we already know its unstable: */
848 	if (check_tsc_unstable()) {
849 		clocksource_tsc.rating = 0;
850 		clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
851 	}
852 	clocksource_register(&clocksource_tsc);
853 }
854 
855 void __init tsc_init(void)
856 {
857 	u64 lpj;
858 	int cpu;
859 
860 	if (!cpu_has_tsc)
861 		return;
862 
863 	tsc_khz = calibrate_tsc();
864 	cpu_khz = tsc_khz;
865 
866 	if (!tsc_khz) {
867 		mark_tsc_unstable("could not calculate TSC khz");
868 		return;
869 	}
870 
871 #ifdef CONFIG_X86_64
872 	if (cpu_has(&boot_cpu_data, X86_FEATURE_CONSTANT_TSC) &&
873 			(boot_cpu_data.x86_vendor == X86_VENDOR_AMD))
874 		cpu_khz = calibrate_cpu();
875 #endif
876 
877 	printk("Detected %lu.%03lu MHz processor.\n",
878 			(unsigned long)cpu_khz / 1000,
879 			(unsigned long)cpu_khz % 1000);
880 
881 	/*
882 	 * Secondary CPUs do not run through tsc_init(), so set up
883 	 * all the scale factors for all CPUs, assuming the same
884 	 * speed as the bootup CPU. (cpufreq notifiers will fix this
885 	 * up if their speed diverges)
886 	 */
887 	for_each_possible_cpu(cpu)
888 		set_cyc2ns_scale(cpu_khz, cpu);
889 
890 	if (tsc_disabled > 0)
891 		return;
892 
893 	/* now allow native_sched_clock() to use rdtsc */
894 	tsc_disabled = 0;
895 
896 	lpj = ((u64)tsc_khz * 1000);
897 	do_div(lpj, HZ);
898 	lpj_fine = lpj;
899 
900 	use_tsc_delay();
901 	/* Check and install the TSC clocksource */
902 	dmi_check_system(bad_tsc_dmi_table);
903 
904 	if (unsynchronized_tsc())
905 		mark_tsc_unstable("TSCs unsynchronized");
906 
907 	check_system_tsc_reliable();
908 	init_tsc_clocksource();
909 }
910 
911