xref: /openbmc/linux/arch/x86/kernel/tsc.c (revision a09d2831)
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 #include <asm/nmi.h>
21 #include <asm/x86_init.h>
22 
23 unsigned int __read_mostly cpu_khz;	/* TSC clocks / usec, not used here */
24 EXPORT_SYMBOL(cpu_khz);
25 
26 unsigned int __read_mostly tsc_khz;
27 EXPORT_SYMBOL(tsc_khz);
28 
29 /*
30  * TSC can be unstable due to cpufreq or due to unsynced TSCs
31  */
32 static int __read_mostly tsc_unstable;
33 
34 /* native_sched_clock() is called before tsc_init(), so
35    we must start with the TSC soft disabled to prevent
36    erroneous rdtsc usage on !cpu_has_tsc processors */
37 static int __read_mostly tsc_disabled = -1;
38 
39 static int tsc_clocksource_reliable;
40 /*
41  * Scheduler clock - returns current time in nanosec units.
42  */
43 u64 native_sched_clock(void)
44 {
45 	u64 this_offset;
46 
47 	/*
48 	 * Fall back to jiffies if there's no TSC available:
49 	 * ( But note that we still use it if the TSC is marked
50 	 *   unstable. We do this because unlike Time Of Day,
51 	 *   the scheduler clock tolerates small errors and it's
52 	 *   very important for it to be as fast as the platform
53 	 *   can achive it. )
54 	 */
55 	if (unlikely(tsc_disabled)) {
56 		/* No locking but a rare wrong value is not a big deal: */
57 		return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
58 	}
59 
60 	/* read the Time Stamp Counter: */
61 	rdtscll(this_offset);
62 
63 	/* return the value in ns */
64 	return __cycles_2_ns(this_offset);
65 }
66 
67 /* We need to define a real function for sched_clock, to override the
68    weak default version */
69 #ifdef CONFIG_PARAVIRT
70 unsigned long long sched_clock(void)
71 {
72 	return paravirt_sched_clock();
73 }
74 #else
75 unsigned long long
76 sched_clock(void) __attribute__((alias("native_sched_clock")));
77 #endif
78 
79 int check_tsc_unstable(void)
80 {
81 	return tsc_unstable;
82 }
83 EXPORT_SYMBOL_GPL(check_tsc_unstable);
84 
85 #ifdef CONFIG_X86_TSC
86 int __init notsc_setup(char *str)
87 {
88 	printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, "
89 			"cannot disable TSC completely.\n");
90 	tsc_disabled = 1;
91 	return 1;
92 }
93 #else
94 /*
95  * disable flag for tsc. Takes effect by clearing the TSC cpu flag
96  * in cpu/common.c
97  */
98 int __init notsc_setup(char *str)
99 {
100 	setup_clear_cpu_cap(X86_FEATURE_TSC);
101 	return 1;
102 }
103 #endif
104 
105 __setup("notsc", notsc_setup);
106 
107 static int __init tsc_setup(char *str)
108 {
109 	if (!strcmp(str, "reliable"))
110 		tsc_clocksource_reliable = 1;
111 	return 1;
112 }
113 
114 __setup("tsc=", tsc_setup);
115 
116 #define MAX_RETRIES     5
117 #define SMI_TRESHOLD    50000
118 
119 /*
120  * Read TSC and the reference counters. Take care of SMI disturbance
121  */
122 static u64 tsc_read_refs(u64 *p, int hpet)
123 {
124 	u64 t1, t2;
125 	int i;
126 
127 	for (i = 0; i < MAX_RETRIES; i++) {
128 		t1 = get_cycles();
129 		if (hpet)
130 			*p = hpet_readl(HPET_COUNTER) & 0xFFFFFFFF;
131 		else
132 			*p = acpi_pm_read_early();
133 		t2 = get_cycles();
134 		if ((t2 - t1) < SMI_TRESHOLD)
135 			return t2;
136 	}
137 	return ULLONG_MAX;
138 }
139 
140 /*
141  * Calculate the TSC frequency from HPET reference
142  */
143 static unsigned long calc_hpet_ref(u64 deltatsc, u64 hpet1, u64 hpet2)
144 {
145 	u64 tmp;
146 
147 	if (hpet2 < hpet1)
148 		hpet2 += 0x100000000ULL;
149 	hpet2 -= hpet1;
150 	tmp = ((u64)hpet2 * hpet_readl(HPET_PERIOD));
151 	do_div(tmp, 1000000);
152 	do_div(deltatsc, tmp);
153 
154 	return (unsigned long) deltatsc;
155 }
156 
157 /*
158  * Calculate the TSC frequency from PMTimer reference
159  */
160 static unsigned long calc_pmtimer_ref(u64 deltatsc, u64 pm1, u64 pm2)
161 {
162 	u64 tmp;
163 
164 	if (!pm1 && !pm2)
165 		return ULONG_MAX;
166 
167 	if (pm2 < pm1)
168 		pm2 += (u64)ACPI_PM_OVRRUN;
169 	pm2 -= pm1;
170 	tmp = pm2 * 1000000000LL;
171 	do_div(tmp, PMTMR_TICKS_PER_SEC);
172 	do_div(deltatsc, tmp);
173 
174 	return (unsigned long) deltatsc;
175 }
176 
177 #define CAL_MS		10
178 #define CAL_LATCH	(CLOCK_TICK_RATE / (1000 / CAL_MS))
179 #define CAL_PIT_LOOPS	1000
180 
181 #define CAL2_MS		50
182 #define CAL2_LATCH	(CLOCK_TICK_RATE / (1000 / CAL2_MS))
183 #define CAL2_PIT_LOOPS	5000
184 
185 
186 /*
187  * Try to calibrate the TSC against the Programmable
188  * Interrupt Timer and return the frequency of the TSC
189  * in kHz.
190  *
191  * Return ULONG_MAX on failure to calibrate.
192  */
193 static unsigned long pit_calibrate_tsc(u32 latch, unsigned long ms, int loopmin)
194 {
195 	u64 tsc, t1, t2, delta;
196 	unsigned long tscmin, tscmax;
197 	int pitcnt;
198 
199 	/* Set the Gate high, disable speaker */
200 	outb((inb(0x61) & ~0x02) | 0x01, 0x61);
201 
202 	/*
203 	 * Setup CTC channel 2* for mode 0, (interrupt on terminal
204 	 * count mode), binary count. Set the latch register to 50ms
205 	 * (LSB then MSB) to begin countdown.
206 	 */
207 	outb(0xb0, 0x43);
208 	outb(latch & 0xff, 0x42);
209 	outb(latch >> 8, 0x42);
210 
211 	tsc = t1 = t2 = get_cycles();
212 
213 	pitcnt = 0;
214 	tscmax = 0;
215 	tscmin = ULONG_MAX;
216 	while ((inb(0x61) & 0x20) == 0) {
217 		t2 = get_cycles();
218 		delta = t2 - tsc;
219 		tsc = t2;
220 		if ((unsigned long) delta < tscmin)
221 			tscmin = (unsigned int) delta;
222 		if ((unsigned long) delta > tscmax)
223 			tscmax = (unsigned int) delta;
224 		pitcnt++;
225 	}
226 
227 	/*
228 	 * Sanity checks:
229 	 *
230 	 * If we were not able to read the PIT more than loopmin
231 	 * times, then we have been hit by a massive SMI
232 	 *
233 	 * If the maximum is 10 times larger than the minimum,
234 	 * then we got hit by an SMI as well.
235 	 */
236 	if (pitcnt < loopmin || tscmax > 10 * tscmin)
237 		return ULONG_MAX;
238 
239 	/* Calculate the PIT value */
240 	delta = t2 - t1;
241 	do_div(delta, ms);
242 	return delta;
243 }
244 
245 /*
246  * This reads the current MSB of the PIT counter, and
247  * checks if we are running on sufficiently fast and
248  * non-virtualized hardware.
249  *
250  * Our expectations are:
251  *
252  *  - the PIT is running at roughly 1.19MHz
253  *
254  *  - each IO is going to take about 1us on real hardware,
255  *    but we allow it to be much faster (by a factor of 10) or
256  *    _slightly_ slower (ie we allow up to a 2us read+counter
257  *    update - anything else implies a unacceptably slow CPU
258  *    or PIT for the fast calibration to work.
259  *
260  *  - with 256 PIT ticks to read the value, we have 214us to
261  *    see the same MSB (and overhead like doing a single TSC
262  *    read per MSB value etc).
263  *
264  *  - We're doing 2 reads per loop (LSB, MSB), and we expect
265  *    them each to take about a microsecond on real hardware.
266  *    So we expect a count value of around 100. But we'll be
267  *    generous, and accept anything over 50.
268  *
269  *  - if the PIT is stuck, and we see *many* more reads, we
270  *    return early (and the next caller of pit_expect_msb()
271  *    then consider it a failure when they don't see the
272  *    next expected value).
273  *
274  * These expectations mean that we know that we have seen the
275  * transition from one expected value to another with a fairly
276  * high accuracy, and we didn't miss any events. We can thus
277  * use the TSC value at the transitions to calculate a pretty
278  * good value for the TSC frequencty.
279  */
280 static inline int pit_verify_msb(unsigned char val)
281 {
282 	/* Ignore LSB */
283 	inb(0x42);
284 	return inb(0x42) == val;
285 }
286 
287 static inline int pit_expect_msb(unsigned char val, u64 *tscp, unsigned long *deltap)
288 {
289 	int count;
290 	u64 tsc = 0;
291 
292 	for (count = 0; count < 50000; count++) {
293 		if (!pit_verify_msb(val))
294 			break;
295 		tsc = get_cycles();
296 	}
297 	*deltap = get_cycles() - tsc;
298 	*tscp = tsc;
299 
300 	/*
301 	 * We require _some_ success, but the quality control
302 	 * will be based on the error terms on the TSC values.
303 	 */
304 	return count > 5;
305 }
306 
307 /*
308  * How many MSB values do we want to see? We aim for
309  * a maximum error rate of 500ppm (in practice the
310  * real error is much smaller), but refuse to spend
311  * more than 25ms on it.
312  */
313 #define MAX_QUICK_PIT_MS 25
314 #define MAX_QUICK_PIT_ITERATIONS (MAX_QUICK_PIT_MS * PIT_TICK_RATE / 1000 / 256)
315 
316 static unsigned long quick_pit_calibrate(void)
317 {
318 	int i;
319 	u64 tsc, delta;
320 	unsigned long d1, d2;
321 
322 	/* Set the Gate high, disable speaker */
323 	outb((inb(0x61) & ~0x02) | 0x01, 0x61);
324 
325 	/*
326 	 * Counter 2, mode 0 (one-shot), binary count
327 	 *
328 	 * NOTE! Mode 2 decrements by two (and then the
329 	 * output is flipped each time, giving the same
330 	 * final output frequency as a decrement-by-one),
331 	 * so mode 0 is much better when looking at the
332 	 * individual counts.
333 	 */
334 	outb(0xb0, 0x43);
335 
336 	/* Start at 0xffff */
337 	outb(0xff, 0x42);
338 	outb(0xff, 0x42);
339 
340 	/*
341 	 * The PIT starts counting at the next edge, so we
342 	 * need to delay for a microsecond. The easiest way
343 	 * to do that is to just read back the 16-bit counter
344 	 * once from the PIT.
345 	 */
346 	pit_verify_msb(0);
347 
348 	if (pit_expect_msb(0xff, &tsc, &d1)) {
349 		for (i = 1; i <= MAX_QUICK_PIT_ITERATIONS; i++) {
350 			if (!pit_expect_msb(0xff-i, &delta, &d2))
351 				break;
352 
353 			/*
354 			 * Iterate until the error is less than 500 ppm
355 			 */
356 			delta -= tsc;
357 			if (d1+d2 >= delta >> 11)
358 				continue;
359 
360 			/*
361 			 * Check the PIT one more time to verify that
362 			 * all TSC reads were stable wrt the PIT.
363 			 *
364 			 * This also guarantees serialization of the
365 			 * last cycle read ('d2') in pit_expect_msb.
366 			 */
367 			if (!pit_verify_msb(0xfe - i))
368 				break;
369 			goto success;
370 		}
371 	}
372 	printk("Fast TSC calibration failed\n");
373 	return 0;
374 
375 success:
376 	/*
377 	 * Ok, if we get here, then we've seen the
378 	 * MSB of the PIT decrement 'i' times, and the
379 	 * error has shrunk to less than 500 ppm.
380 	 *
381 	 * As a result, we can depend on there not being
382 	 * any odd delays anywhere, and the TSC reads are
383 	 * reliable (within the error). We also adjust the
384 	 * delta to the middle of the error bars, just
385 	 * because it looks nicer.
386 	 *
387 	 * kHz = ticks / time-in-seconds / 1000;
388 	 * kHz = (t2 - t1) / (I * 256 / PIT_TICK_RATE) / 1000
389 	 * kHz = ((t2 - t1) * PIT_TICK_RATE) / (I * 256 * 1000)
390 	 */
391 	delta += (long)(d2 - d1)/2;
392 	delta *= PIT_TICK_RATE;
393 	do_div(delta, i*256*1000);
394 	printk("Fast TSC calibration using PIT\n");
395 	return delta;
396 }
397 
398 /**
399  * native_calibrate_tsc - calibrate the tsc on boot
400  */
401 unsigned long native_calibrate_tsc(void)
402 {
403 	u64 tsc1, tsc2, delta, ref1, ref2;
404 	unsigned long tsc_pit_min = ULONG_MAX, tsc_ref_min = ULONG_MAX;
405 	unsigned long flags, latch, ms, fast_calibrate;
406 	int hpet = is_hpet_enabled(), i, loopmin;
407 
408 	local_irq_save(flags);
409 	fast_calibrate = quick_pit_calibrate();
410 	local_irq_restore(flags);
411 	if (fast_calibrate)
412 		return fast_calibrate;
413 
414 	/*
415 	 * Run 5 calibration loops to get the lowest frequency value
416 	 * (the best estimate). We use two different calibration modes
417 	 * here:
418 	 *
419 	 * 1) PIT loop. We set the PIT Channel 2 to oneshot mode and
420 	 * load a timeout of 50ms. We read the time right after we
421 	 * started the timer and wait until the PIT count down reaches
422 	 * zero. In each wait loop iteration we read the TSC and check
423 	 * the delta to the previous read. We keep track of the min
424 	 * and max values of that delta. The delta is mostly defined
425 	 * by the IO time of the PIT access, so we can detect when a
426 	 * SMI/SMM disturbance happend between the two reads. If the
427 	 * maximum time is significantly larger than the minimum time,
428 	 * then we discard the result and have another try.
429 	 *
430 	 * 2) Reference counter. If available we use the HPET or the
431 	 * PMTIMER as a reference to check the sanity of that value.
432 	 * We use separate TSC readouts and check inside of the
433 	 * reference read for a SMI/SMM disturbance. We dicard
434 	 * disturbed values here as well. We do that around the PIT
435 	 * calibration delay loop as we have to wait for a certain
436 	 * amount of time anyway.
437 	 */
438 
439 	/* Preset PIT loop values */
440 	latch = CAL_LATCH;
441 	ms = CAL_MS;
442 	loopmin = CAL_PIT_LOOPS;
443 
444 	for (i = 0; i < 3; i++) {
445 		unsigned long tsc_pit_khz;
446 
447 		/*
448 		 * Read the start value and the reference count of
449 		 * hpet/pmtimer when available. Then do the PIT
450 		 * calibration, which will take at least 50ms, and
451 		 * read the end value.
452 		 */
453 		local_irq_save(flags);
454 		tsc1 = tsc_read_refs(&ref1, hpet);
455 		tsc_pit_khz = pit_calibrate_tsc(latch, ms, loopmin);
456 		tsc2 = tsc_read_refs(&ref2, hpet);
457 		local_irq_restore(flags);
458 
459 		/* Pick the lowest PIT TSC calibration so far */
460 		tsc_pit_min = min(tsc_pit_min, tsc_pit_khz);
461 
462 		/* hpet or pmtimer available ? */
463 		if (!hpet && !ref1 && !ref2)
464 			continue;
465 
466 		/* Check, whether the sampling was disturbed by an SMI */
467 		if (tsc1 == ULLONG_MAX || tsc2 == ULLONG_MAX)
468 			continue;
469 
470 		tsc2 = (tsc2 - tsc1) * 1000000LL;
471 		if (hpet)
472 			tsc2 = calc_hpet_ref(tsc2, ref1, ref2);
473 		else
474 			tsc2 = calc_pmtimer_ref(tsc2, ref1, ref2);
475 
476 		tsc_ref_min = min(tsc_ref_min, (unsigned long) tsc2);
477 
478 		/* Check the reference deviation */
479 		delta = ((u64) tsc_pit_min) * 100;
480 		do_div(delta, tsc_ref_min);
481 
482 		/*
483 		 * If both calibration results are inside a 10% window
484 		 * then we can be sure, that the calibration
485 		 * succeeded. We break out of the loop right away. We
486 		 * use the reference value, as it is more precise.
487 		 */
488 		if (delta >= 90 && delta <= 110) {
489 			printk(KERN_INFO
490 			       "TSC: PIT calibration matches %s. %d loops\n",
491 			       hpet ? "HPET" : "PMTIMER", i + 1);
492 			return tsc_ref_min;
493 		}
494 
495 		/*
496 		 * Check whether PIT failed more than once. This
497 		 * happens in virtualized environments. We need to
498 		 * give the virtual PC a slightly longer timeframe for
499 		 * the HPET/PMTIMER to make the result precise.
500 		 */
501 		if (i == 1 && tsc_pit_min == ULONG_MAX) {
502 			latch = CAL2_LATCH;
503 			ms = CAL2_MS;
504 			loopmin = CAL2_PIT_LOOPS;
505 		}
506 	}
507 
508 	/*
509 	 * Now check the results.
510 	 */
511 	if (tsc_pit_min == ULONG_MAX) {
512 		/* PIT gave no useful value */
513 		printk(KERN_WARNING "TSC: Unable to calibrate against PIT\n");
514 
515 		/* We don't have an alternative source, disable TSC */
516 		if (!hpet && !ref1 && !ref2) {
517 			printk("TSC: No reference (HPET/PMTIMER) available\n");
518 			return 0;
519 		}
520 
521 		/* The alternative source failed as well, disable TSC */
522 		if (tsc_ref_min == ULONG_MAX) {
523 			printk(KERN_WARNING "TSC: HPET/PMTIMER calibration "
524 			       "failed.\n");
525 			return 0;
526 		}
527 
528 		/* Use the alternative source */
529 		printk(KERN_INFO "TSC: using %s reference calibration\n",
530 		       hpet ? "HPET" : "PMTIMER");
531 
532 		return tsc_ref_min;
533 	}
534 
535 	/* We don't have an alternative source, use the PIT calibration value */
536 	if (!hpet && !ref1 && !ref2) {
537 		printk(KERN_INFO "TSC: Using PIT calibration value\n");
538 		return tsc_pit_min;
539 	}
540 
541 	/* The alternative source failed, use the PIT calibration value */
542 	if (tsc_ref_min == ULONG_MAX) {
543 		printk(KERN_WARNING "TSC: HPET/PMTIMER calibration failed. "
544 		       "Using PIT calibration\n");
545 		return tsc_pit_min;
546 	}
547 
548 	/*
549 	 * The calibration values differ too much. In doubt, we use
550 	 * the PIT value as we know that there are PMTIMERs around
551 	 * running at double speed. At least we let the user know:
552 	 */
553 	printk(KERN_WARNING "TSC: PIT calibration deviates from %s: %lu %lu.\n",
554 	       hpet ? "HPET" : "PMTIMER", tsc_pit_min, tsc_ref_min);
555 	printk(KERN_INFO "TSC: Using PIT calibration value\n");
556 	return tsc_pit_min;
557 }
558 
559 int recalibrate_cpu_khz(void)
560 {
561 #ifndef CONFIG_SMP
562 	unsigned long cpu_khz_old = cpu_khz;
563 
564 	if (cpu_has_tsc) {
565 		tsc_khz = x86_platform.calibrate_tsc();
566 		cpu_khz = tsc_khz;
567 		cpu_data(0).loops_per_jiffy =
568 			cpufreq_scale(cpu_data(0).loops_per_jiffy,
569 					cpu_khz_old, cpu_khz);
570 		return 0;
571 	} else
572 		return -ENODEV;
573 #else
574 	return -ENODEV;
575 #endif
576 }
577 
578 EXPORT_SYMBOL(recalibrate_cpu_khz);
579 
580 
581 /* Accelerators for sched_clock()
582  * convert from cycles(64bits) => nanoseconds (64bits)
583  *  basic equation:
584  *              ns = cycles / (freq / ns_per_sec)
585  *              ns = cycles * (ns_per_sec / freq)
586  *              ns = cycles * (10^9 / (cpu_khz * 10^3))
587  *              ns = cycles * (10^6 / cpu_khz)
588  *
589  *      Then we use scaling math (suggested by george@mvista.com) to get:
590  *              ns = cycles * (10^6 * SC / cpu_khz) / SC
591  *              ns = cycles * cyc2ns_scale / SC
592  *
593  *      And since SC is a constant power of two, we can convert the div
594  *  into a shift.
595  *
596  *  We can use khz divisor instead of mhz to keep a better precision, since
597  *  cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
598  *  (mathieu.desnoyers@polymtl.ca)
599  *
600  *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
601  */
602 
603 DEFINE_PER_CPU(unsigned long, cyc2ns);
604 DEFINE_PER_CPU(unsigned long long, cyc2ns_offset);
605 
606 static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu)
607 {
608 	unsigned long long tsc_now, ns_now, *offset;
609 	unsigned long flags, *scale;
610 
611 	local_irq_save(flags);
612 	sched_clock_idle_sleep_event();
613 
614 	scale = &per_cpu(cyc2ns, cpu);
615 	offset = &per_cpu(cyc2ns_offset, cpu);
616 
617 	rdtscll(tsc_now);
618 	ns_now = __cycles_2_ns(tsc_now);
619 
620 	if (cpu_khz) {
621 		*scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz;
622 		*offset = ns_now - (tsc_now * *scale >> CYC2NS_SCALE_FACTOR);
623 	}
624 
625 	sched_clock_idle_wakeup_event(0);
626 	local_irq_restore(flags);
627 }
628 
629 #ifdef CONFIG_CPU_FREQ
630 
631 /* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
632  * changes.
633  *
634  * RED-PEN: On SMP we assume all CPUs run with the same frequency.  It's
635  * not that important because current Opteron setups do not support
636  * scaling on SMP anyroads.
637  *
638  * Should fix up last_tsc too. Currently gettimeofday in the
639  * first tick after the change will be slightly wrong.
640  */
641 
642 static unsigned int  ref_freq;
643 static unsigned long loops_per_jiffy_ref;
644 static unsigned long tsc_khz_ref;
645 
646 static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
647 				void *data)
648 {
649 	struct cpufreq_freqs *freq = data;
650 	unsigned long *lpj;
651 
652 	if (cpu_has(&cpu_data(freq->cpu), X86_FEATURE_CONSTANT_TSC))
653 		return 0;
654 
655 	lpj = &boot_cpu_data.loops_per_jiffy;
656 #ifdef CONFIG_SMP
657 	if (!(freq->flags & CPUFREQ_CONST_LOOPS))
658 		lpj = &cpu_data(freq->cpu).loops_per_jiffy;
659 #endif
660 
661 	if (!ref_freq) {
662 		ref_freq = freq->old;
663 		loops_per_jiffy_ref = *lpj;
664 		tsc_khz_ref = tsc_khz;
665 	}
666 	if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
667 			(val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
668 			(val == CPUFREQ_RESUMECHANGE)) {
669 		*lpj = cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
670 
671 		tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
672 		if (!(freq->flags & CPUFREQ_CONST_LOOPS))
673 			mark_tsc_unstable("cpufreq changes");
674 	}
675 
676 	set_cyc2ns_scale(tsc_khz, freq->cpu);
677 
678 	return 0;
679 }
680 
681 static struct notifier_block time_cpufreq_notifier_block = {
682 	.notifier_call  = time_cpufreq_notifier
683 };
684 
685 static int __init cpufreq_tsc(void)
686 {
687 	if (!cpu_has_tsc)
688 		return 0;
689 	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
690 		return 0;
691 	cpufreq_register_notifier(&time_cpufreq_notifier_block,
692 				CPUFREQ_TRANSITION_NOTIFIER);
693 	return 0;
694 }
695 
696 core_initcall(cpufreq_tsc);
697 
698 #endif /* CONFIG_CPU_FREQ */
699 
700 /* clocksource code */
701 
702 static struct clocksource clocksource_tsc;
703 
704 /*
705  * We compare the TSC to the cycle_last value in the clocksource
706  * structure to avoid a nasty time-warp. This can be observed in a
707  * very small window right after one CPU updated cycle_last under
708  * xtime/vsyscall_gtod lock and the other CPU reads a TSC value which
709  * is smaller than the cycle_last reference value due to a TSC which
710  * is slighty behind. This delta is nowhere else observable, but in
711  * that case it results in a forward time jump in the range of hours
712  * due to the unsigned delta calculation of the time keeping core
713  * code, which is necessary to support wrapping clocksources like pm
714  * timer.
715  */
716 static cycle_t read_tsc(struct clocksource *cs)
717 {
718 	cycle_t ret = (cycle_t)get_cycles();
719 
720 	return ret >= clocksource_tsc.cycle_last ?
721 		ret : clocksource_tsc.cycle_last;
722 }
723 
724 #ifdef CONFIG_X86_64
725 static cycle_t __vsyscall_fn vread_tsc(void)
726 {
727 	cycle_t ret;
728 
729 	/*
730 	 * Surround the RDTSC by barriers, to make sure it's not
731 	 * speculated to outside the seqlock critical section and
732 	 * does not cause time warps:
733 	 */
734 	rdtsc_barrier();
735 	ret = (cycle_t)vget_cycles();
736 	rdtsc_barrier();
737 
738 	return ret >= __vsyscall_gtod_data.clock.cycle_last ?
739 		ret : __vsyscall_gtod_data.clock.cycle_last;
740 }
741 #endif
742 
743 static void resume_tsc(void)
744 {
745 	clocksource_tsc.cycle_last = 0;
746 }
747 
748 static struct clocksource clocksource_tsc = {
749 	.name                   = "tsc",
750 	.rating                 = 300,
751 	.read                   = read_tsc,
752 	.resume			= resume_tsc,
753 	.mask                   = CLOCKSOURCE_MASK(64),
754 	.shift                  = 22,
755 	.flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
756 				  CLOCK_SOURCE_MUST_VERIFY,
757 #ifdef CONFIG_X86_64
758 	.vread                  = vread_tsc,
759 #endif
760 };
761 
762 void mark_tsc_unstable(char *reason)
763 {
764 	if (!tsc_unstable) {
765 		tsc_unstable = 1;
766 		sched_clock_stable = 0;
767 		printk(KERN_INFO "Marking TSC unstable due to %s\n", reason);
768 		/* Change only the rating, when not registered */
769 		if (clocksource_tsc.mult)
770 			clocksource_mark_unstable(&clocksource_tsc);
771 		else {
772 			clocksource_tsc.flags |= CLOCK_SOURCE_UNSTABLE;
773 			clocksource_tsc.rating = 0;
774 		}
775 	}
776 }
777 
778 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
779 
780 static int __init dmi_mark_tsc_unstable(const struct dmi_system_id *d)
781 {
782 	printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
783 			d->ident);
784 	tsc_unstable = 1;
785 	return 0;
786 }
787 
788 /* List of systems that have known TSC problems */
789 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
790 	{
791 		.callback = dmi_mark_tsc_unstable,
792 		.ident = "IBM Thinkpad 380XD",
793 		.matches = {
794 			DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
795 			DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
796 		},
797 	},
798 	{}
799 };
800 
801 static void __init check_system_tsc_reliable(void)
802 {
803 #ifdef CONFIG_MGEODE_LX
804 	/* RTSC counts during suspend */
805 #define RTSC_SUSP 0x100
806 	unsigned long res_low, res_high;
807 
808 	rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
809 	/* Geode_LX - the OLPC CPU has a possibly a very reliable TSC */
810 	if (res_low & RTSC_SUSP)
811 		tsc_clocksource_reliable = 1;
812 #endif
813 	if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE))
814 		tsc_clocksource_reliable = 1;
815 }
816 
817 /*
818  * Make an educated guess if the TSC is trustworthy and synchronized
819  * over all CPUs.
820  */
821 __cpuinit int unsynchronized_tsc(void)
822 {
823 	if (!cpu_has_tsc || tsc_unstable)
824 		return 1;
825 
826 #ifdef CONFIG_SMP
827 	if (apic_is_clustered_box())
828 		return 1;
829 #endif
830 
831 	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
832 		return 0;
833 	/*
834 	 * Intel systems are normally all synchronized.
835 	 * Exceptions must mark TSC as unstable:
836 	 */
837 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
838 		/* assume multi socket systems are not synchronized: */
839 		if (num_possible_cpus() > 1)
840 			tsc_unstable = 1;
841 	}
842 
843 	return tsc_unstable;
844 }
845 
846 static void __init init_tsc_clocksource(void)
847 {
848 	clocksource_tsc.mult = clocksource_khz2mult(tsc_khz,
849 			clocksource_tsc.shift);
850 	if (tsc_clocksource_reliable)
851 		clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
852 	/* lower the rating if we already know its unstable: */
853 	if (check_tsc_unstable()) {
854 		clocksource_tsc.rating = 0;
855 		clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
856 	}
857 	clocksource_register(&clocksource_tsc);
858 }
859 
860 #ifdef CONFIG_X86_64
861 /*
862  * calibrate_cpu is used on systems with fixed rate TSCs to determine
863  * processor frequency
864  */
865 #define TICK_COUNT 100000000
866 static unsigned long __init calibrate_cpu(void)
867 {
868 	int tsc_start, tsc_now;
869 	int i, no_ctr_free;
870 	unsigned long evntsel3 = 0, pmc3 = 0, pmc_now = 0;
871 	unsigned long flags;
872 
873 	for (i = 0; i < 4; i++)
874 		if (avail_to_resrv_perfctr_nmi_bit(i))
875 			break;
876 	no_ctr_free = (i == 4);
877 	if (no_ctr_free) {
878 		WARN(1, KERN_WARNING "Warning: AMD perfctrs busy ... "
879 		     "cpu_khz value may be incorrect.\n");
880 		i = 3;
881 		rdmsrl(MSR_K7_EVNTSEL3, evntsel3);
882 		wrmsrl(MSR_K7_EVNTSEL3, 0);
883 		rdmsrl(MSR_K7_PERFCTR3, pmc3);
884 	} else {
885 		reserve_perfctr_nmi(MSR_K7_PERFCTR0 + i);
886 		reserve_evntsel_nmi(MSR_K7_EVNTSEL0 + i);
887 	}
888 	local_irq_save(flags);
889 	/* start measuring cycles, incrementing from 0 */
890 	wrmsrl(MSR_K7_PERFCTR0 + i, 0);
891 	wrmsrl(MSR_K7_EVNTSEL0 + i, 1 << 22 | 3 << 16 | 0x76);
892 	rdtscl(tsc_start);
893 	do {
894 		rdmsrl(MSR_K7_PERFCTR0 + i, pmc_now);
895 		tsc_now = get_cycles();
896 	} while ((tsc_now - tsc_start) < TICK_COUNT);
897 
898 	local_irq_restore(flags);
899 	if (no_ctr_free) {
900 		wrmsrl(MSR_K7_EVNTSEL3, 0);
901 		wrmsrl(MSR_K7_PERFCTR3, pmc3);
902 		wrmsrl(MSR_K7_EVNTSEL3, evntsel3);
903 	} else {
904 		release_perfctr_nmi(MSR_K7_PERFCTR0 + i);
905 		release_evntsel_nmi(MSR_K7_EVNTSEL0 + i);
906 	}
907 
908 	return pmc_now * tsc_khz / (tsc_now - tsc_start);
909 }
910 #else
911 static inline unsigned long calibrate_cpu(void) { return cpu_khz; }
912 #endif
913 
914 void __init tsc_init(void)
915 {
916 	u64 lpj;
917 	int cpu;
918 
919 	x86_init.timers.tsc_pre_init();
920 
921 	if (!cpu_has_tsc)
922 		return;
923 
924 	tsc_khz = x86_platform.calibrate_tsc();
925 	cpu_khz = tsc_khz;
926 
927 	if (!tsc_khz) {
928 		mark_tsc_unstable("could not calculate TSC khz");
929 		return;
930 	}
931 
932 	if (cpu_has(&boot_cpu_data, X86_FEATURE_CONSTANT_TSC) &&
933 			(boot_cpu_data.x86_vendor == X86_VENDOR_AMD))
934 		cpu_khz = calibrate_cpu();
935 
936 	printk("Detected %lu.%03lu MHz processor.\n",
937 			(unsigned long)cpu_khz / 1000,
938 			(unsigned long)cpu_khz % 1000);
939 
940 	/*
941 	 * Secondary CPUs do not run through tsc_init(), so set up
942 	 * all the scale factors for all CPUs, assuming the same
943 	 * speed as the bootup CPU. (cpufreq notifiers will fix this
944 	 * up if their speed diverges)
945 	 */
946 	for_each_possible_cpu(cpu)
947 		set_cyc2ns_scale(cpu_khz, cpu);
948 
949 	if (tsc_disabled > 0)
950 		return;
951 
952 	/* now allow native_sched_clock() to use rdtsc */
953 	tsc_disabled = 0;
954 
955 	lpj = ((u64)tsc_khz * 1000);
956 	do_div(lpj, HZ);
957 	lpj_fine = lpj;
958 
959 	use_tsc_delay();
960 	/* Check and install the TSC clocksource */
961 	dmi_check_system(bad_tsc_dmi_table);
962 
963 	if (unsynchronized_tsc())
964 		mark_tsc_unstable("TSCs unsynchronized");
965 
966 	check_system_tsc_reliable();
967 	init_tsc_clocksource();
968 }
969 
970