xref: /linux/arch/x86/kernel/tsc.c (revision 5bdef865eb358b6f3760e25e591ae115e9eeddef)
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_expect_msb(unsigned char val, u64 *tscp, unsigned long *deltap)
279 {
280 	int count;
281 	u64 tsc = 0;
282 
283 	for (count = 0; count < 50000; count++) {
284 		/* Ignore LSB */
285 		inb(0x42);
286 		if (inb(0x42) != val)
287 			break;
288 		tsc = get_cycles();
289 	}
290 	*deltap = get_cycles() - tsc;
291 	*tscp = tsc;
292 
293 	/*
294 	 * We require _some_ success, but the quality control
295 	 * will be based on the error terms on the TSC values.
296 	 */
297 	return count > 5;
298 }
299 
300 /*
301  * How many MSB values do we want to see? We aim for
302  * a maximum error rate of 500ppm (in practice the
303  * real error is much smaller), but refuse to spend
304  * more than 25ms on it.
305  */
306 #define MAX_QUICK_PIT_MS 25
307 #define MAX_QUICK_PIT_ITERATIONS (MAX_QUICK_PIT_MS * PIT_TICK_RATE / 1000 / 256)
308 
309 static unsigned long quick_pit_calibrate(void)
310 {
311 	int i;
312 	u64 tsc, delta;
313 	unsigned long d1, d2;
314 
315 	/* Set the Gate high, disable speaker */
316 	outb((inb(0x61) & ~0x02) | 0x01, 0x61);
317 
318 	/*
319 	 * Counter 2, mode 0 (one-shot), binary count
320 	 *
321 	 * NOTE! Mode 2 decrements by two (and then the
322 	 * output is flipped each time, giving the same
323 	 * final output frequency as a decrement-by-one),
324 	 * so mode 0 is much better when looking at the
325 	 * individual counts.
326 	 */
327 	outb(0xb0, 0x43);
328 
329 	/* Start at 0xffff */
330 	outb(0xff, 0x42);
331 	outb(0xff, 0x42);
332 
333 	/*
334 	 * The PIT starts counting at the next edge, so we
335 	 * need to delay for a microsecond. The easiest way
336 	 * to do that is to just read back the 16-bit counter
337 	 * once from the PIT.
338 	 */
339 	inb(0x42);
340 	inb(0x42);
341 
342 	if (pit_expect_msb(0xff, &tsc, &d1)) {
343 		for (i = 1; i <= MAX_QUICK_PIT_ITERATIONS; i++) {
344 			if (!pit_expect_msb(0xff-i, &delta, &d2))
345 				break;
346 
347 			/*
348 			 * Iterate until the error is less than 500 ppm
349 			 */
350 			delta -= tsc;
351 			if (d1+d2 < delta >> 11)
352 				goto success;
353 		}
354 	}
355 	printk("Fast TSC calibration failed\n");
356 	return 0;
357 
358 success:
359 	/*
360 	 * Ok, if we get here, then we've seen the
361 	 * MSB of the PIT decrement 'i' times, and the
362 	 * error has shrunk to less than 500 ppm.
363 	 *
364 	 * As a result, we can depend on there not being
365 	 * any odd delays anywhere, and the TSC reads are
366 	 * reliable (within the error). We also adjust the
367 	 * delta to the middle of the error bars, just
368 	 * because it looks nicer.
369 	 *
370 	 * kHz = ticks / time-in-seconds / 1000;
371 	 * kHz = (t2 - t1) / (I * 256 / PIT_TICK_RATE) / 1000
372 	 * kHz = ((t2 - t1) * PIT_TICK_RATE) / (I * 256 * 1000)
373 	 */
374 	delta += (long)(d2 - d1)/2;
375 	delta *= PIT_TICK_RATE;
376 	do_div(delta, i*256*1000);
377 	printk("Fast TSC calibration using PIT\n");
378 	return delta;
379 }
380 
381 /**
382  * native_calibrate_tsc - calibrate the tsc on boot
383  */
384 unsigned long native_calibrate_tsc(void)
385 {
386 	u64 tsc1, tsc2, delta, ref1, ref2;
387 	unsigned long tsc_pit_min = ULONG_MAX, tsc_ref_min = ULONG_MAX;
388 	unsigned long flags, latch, ms, fast_calibrate, hv_tsc_khz;
389 	int hpet = is_hpet_enabled(), i, loopmin;
390 
391 	hv_tsc_khz = get_hypervisor_tsc_freq();
392 	if (hv_tsc_khz) {
393 		printk(KERN_INFO "TSC: Frequency read from the hypervisor\n");
394 		return hv_tsc_khz;
395 	}
396 
397 	local_irq_save(flags);
398 	fast_calibrate = quick_pit_calibrate();
399 	local_irq_restore(flags);
400 	if (fast_calibrate)
401 		return fast_calibrate;
402 
403 	/*
404 	 * Run 5 calibration loops to get the lowest frequency value
405 	 * (the best estimate). We use two different calibration modes
406 	 * here:
407 	 *
408 	 * 1) PIT loop. We set the PIT Channel 2 to oneshot mode and
409 	 * load a timeout of 50ms. We read the time right after we
410 	 * started the timer and wait until the PIT count down reaches
411 	 * zero. In each wait loop iteration we read the TSC and check
412 	 * the delta to the previous read. We keep track of the min
413 	 * and max values of that delta. The delta is mostly defined
414 	 * by the IO time of the PIT access, so we can detect when a
415 	 * SMI/SMM disturbance happend between the two reads. If the
416 	 * maximum time is significantly larger than the minimum time,
417 	 * then we discard the result and have another try.
418 	 *
419 	 * 2) Reference counter. If available we use the HPET or the
420 	 * PMTIMER as a reference to check the sanity of that value.
421 	 * We use separate TSC readouts and check inside of the
422 	 * reference read for a SMI/SMM disturbance. We dicard
423 	 * disturbed values here as well. We do that around the PIT
424 	 * calibration delay loop as we have to wait for a certain
425 	 * amount of time anyway.
426 	 */
427 
428 	/* Preset PIT loop values */
429 	latch = CAL_LATCH;
430 	ms = CAL_MS;
431 	loopmin = CAL_PIT_LOOPS;
432 
433 	for (i = 0; i < 3; i++) {
434 		unsigned long tsc_pit_khz;
435 
436 		/*
437 		 * Read the start value and the reference count of
438 		 * hpet/pmtimer when available. Then do the PIT
439 		 * calibration, which will take at least 50ms, and
440 		 * read the end value.
441 		 */
442 		local_irq_save(flags);
443 		tsc1 = tsc_read_refs(&ref1, hpet);
444 		tsc_pit_khz = pit_calibrate_tsc(latch, ms, loopmin);
445 		tsc2 = tsc_read_refs(&ref2, hpet);
446 		local_irq_restore(flags);
447 
448 		/* Pick the lowest PIT TSC calibration so far */
449 		tsc_pit_min = min(tsc_pit_min, tsc_pit_khz);
450 
451 		/* hpet or pmtimer available ? */
452 		if (!hpet && !ref1 && !ref2)
453 			continue;
454 
455 		/* Check, whether the sampling was disturbed by an SMI */
456 		if (tsc1 == ULLONG_MAX || tsc2 == ULLONG_MAX)
457 			continue;
458 
459 		tsc2 = (tsc2 - tsc1) * 1000000LL;
460 		if (hpet)
461 			tsc2 = calc_hpet_ref(tsc2, ref1, ref2);
462 		else
463 			tsc2 = calc_pmtimer_ref(tsc2, ref1, ref2);
464 
465 		tsc_ref_min = min(tsc_ref_min, (unsigned long) tsc2);
466 
467 		/* Check the reference deviation */
468 		delta = ((u64) tsc_pit_min) * 100;
469 		do_div(delta, tsc_ref_min);
470 
471 		/*
472 		 * If both calibration results are inside a 10% window
473 		 * then we can be sure, that the calibration
474 		 * succeeded. We break out of the loop right away. We
475 		 * use the reference value, as it is more precise.
476 		 */
477 		if (delta >= 90 && delta <= 110) {
478 			printk(KERN_INFO
479 			       "TSC: PIT calibration matches %s. %d loops\n",
480 			       hpet ? "HPET" : "PMTIMER", i + 1);
481 			return tsc_ref_min;
482 		}
483 
484 		/*
485 		 * Check whether PIT failed more than once. This
486 		 * happens in virtualized environments. We need to
487 		 * give the virtual PC a slightly longer timeframe for
488 		 * the HPET/PMTIMER to make the result precise.
489 		 */
490 		if (i == 1 && tsc_pit_min == ULONG_MAX) {
491 			latch = CAL2_LATCH;
492 			ms = CAL2_MS;
493 			loopmin = CAL2_PIT_LOOPS;
494 		}
495 	}
496 
497 	/*
498 	 * Now check the results.
499 	 */
500 	if (tsc_pit_min == ULONG_MAX) {
501 		/* PIT gave no useful value */
502 		printk(KERN_WARNING "TSC: Unable to calibrate against PIT\n");
503 
504 		/* We don't have an alternative source, disable TSC */
505 		if (!hpet && !ref1 && !ref2) {
506 			printk("TSC: No reference (HPET/PMTIMER) available\n");
507 			return 0;
508 		}
509 
510 		/* The alternative source failed as well, disable TSC */
511 		if (tsc_ref_min == ULONG_MAX) {
512 			printk(KERN_WARNING "TSC: HPET/PMTIMER calibration "
513 			       "failed.\n");
514 			return 0;
515 		}
516 
517 		/* Use the alternative source */
518 		printk(KERN_INFO "TSC: using %s reference calibration\n",
519 		       hpet ? "HPET" : "PMTIMER");
520 
521 		return tsc_ref_min;
522 	}
523 
524 	/* We don't have an alternative source, use the PIT calibration value */
525 	if (!hpet && !ref1 && !ref2) {
526 		printk(KERN_INFO "TSC: Using PIT calibration value\n");
527 		return tsc_pit_min;
528 	}
529 
530 	/* The alternative source failed, use the PIT calibration value */
531 	if (tsc_ref_min == ULONG_MAX) {
532 		printk(KERN_WARNING "TSC: HPET/PMTIMER calibration failed. "
533 		       "Using PIT calibration\n");
534 		return tsc_pit_min;
535 	}
536 
537 	/*
538 	 * The calibration values differ too much. In doubt, we use
539 	 * the PIT value as we know that there are PMTIMERs around
540 	 * running at double speed. At least we let the user know:
541 	 */
542 	printk(KERN_WARNING "TSC: PIT calibration deviates from %s: %lu %lu.\n",
543 	       hpet ? "HPET" : "PMTIMER", tsc_pit_min, tsc_ref_min);
544 	printk(KERN_INFO "TSC: Using PIT calibration value\n");
545 	return tsc_pit_min;
546 }
547 
548 int recalibrate_cpu_khz(void)
549 {
550 #ifndef CONFIG_SMP
551 	unsigned long cpu_khz_old = cpu_khz;
552 
553 	if (cpu_has_tsc) {
554 		tsc_khz = calibrate_tsc();
555 		cpu_khz = tsc_khz;
556 		cpu_data(0).loops_per_jiffy =
557 			cpufreq_scale(cpu_data(0).loops_per_jiffy,
558 					cpu_khz_old, cpu_khz);
559 		return 0;
560 	} else
561 		return -ENODEV;
562 #else
563 	return -ENODEV;
564 #endif
565 }
566 
567 EXPORT_SYMBOL(recalibrate_cpu_khz);
568 
569 
570 /* Accelerators for sched_clock()
571  * convert from cycles(64bits) => nanoseconds (64bits)
572  *  basic equation:
573  *              ns = cycles / (freq / ns_per_sec)
574  *              ns = cycles * (ns_per_sec / freq)
575  *              ns = cycles * (10^9 / (cpu_khz * 10^3))
576  *              ns = cycles * (10^6 / cpu_khz)
577  *
578  *      Then we use scaling math (suggested by george@mvista.com) to get:
579  *              ns = cycles * (10^6 * SC / cpu_khz) / SC
580  *              ns = cycles * cyc2ns_scale / SC
581  *
582  *      And since SC is a constant power of two, we can convert the div
583  *  into a shift.
584  *
585  *  We can use khz divisor instead of mhz to keep a better precision, since
586  *  cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
587  *  (mathieu.desnoyers@polymtl.ca)
588  *
589  *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
590  */
591 
592 DEFINE_PER_CPU(unsigned long, cyc2ns);
593 DEFINE_PER_CPU(unsigned long long, cyc2ns_offset);
594 
595 static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu)
596 {
597 	unsigned long long tsc_now, ns_now, *offset;
598 	unsigned long flags, *scale;
599 
600 	local_irq_save(flags);
601 	sched_clock_idle_sleep_event();
602 
603 	scale = &per_cpu(cyc2ns, cpu);
604 	offset = &per_cpu(cyc2ns_offset, cpu);
605 
606 	rdtscll(tsc_now);
607 	ns_now = __cycles_2_ns(tsc_now);
608 
609 	if (cpu_khz) {
610 		*scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz;
611 		*offset = ns_now - (tsc_now * *scale >> CYC2NS_SCALE_FACTOR);
612 	}
613 
614 	sched_clock_idle_wakeup_event(0);
615 	local_irq_restore(flags);
616 }
617 
618 #ifdef CONFIG_CPU_FREQ
619 
620 /* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
621  * changes.
622  *
623  * RED-PEN: On SMP we assume all CPUs run with the same frequency.  It's
624  * not that important because current Opteron setups do not support
625  * scaling on SMP anyroads.
626  *
627  * Should fix up last_tsc too. Currently gettimeofday in the
628  * first tick after the change will be slightly wrong.
629  */
630 
631 static unsigned int  ref_freq;
632 static unsigned long loops_per_jiffy_ref;
633 static unsigned long tsc_khz_ref;
634 
635 static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
636 				void *data)
637 {
638 	struct cpufreq_freqs *freq = data;
639 	unsigned long *lpj;
640 
641 	if (cpu_has(&cpu_data(freq->cpu), X86_FEATURE_CONSTANT_TSC))
642 		return 0;
643 
644 	lpj = &boot_cpu_data.loops_per_jiffy;
645 #ifdef CONFIG_SMP
646 	if (!(freq->flags & CPUFREQ_CONST_LOOPS))
647 		lpj = &cpu_data(freq->cpu).loops_per_jiffy;
648 #endif
649 
650 	if (!ref_freq) {
651 		ref_freq = freq->old;
652 		loops_per_jiffy_ref = *lpj;
653 		tsc_khz_ref = tsc_khz;
654 	}
655 	if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
656 			(val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
657 			(val == CPUFREQ_RESUMECHANGE)) {
658 		*lpj = 	cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
659 
660 		tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
661 		if (!(freq->flags & CPUFREQ_CONST_LOOPS))
662 			mark_tsc_unstable("cpufreq changes");
663 	}
664 
665 	set_cyc2ns_scale(tsc_khz, freq->cpu);
666 
667 	return 0;
668 }
669 
670 static struct notifier_block time_cpufreq_notifier_block = {
671 	.notifier_call  = time_cpufreq_notifier
672 };
673 
674 static int __init cpufreq_tsc(void)
675 {
676 	if (!cpu_has_tsc)
677 		return 0;
678 	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
679 		return 0;
680 	cpufreq_register_notifier(&time_cpufreq_notifier_block,
681 				CPUFREQ_TRANSITION_NOTIFIER);
682 	return 0;
683 }
684 
685 core_initcall(cpufreq_tsc);
686 
687 #endif /* CONFIG_CPU_FREQ */
688 
689 /* clocksource code */
690 
691 static struct clocksource clocksource_tsc;
692 
693 /*
694  * We compare the TSC to the cycle_last value in the clocksource
695  * structure to avoid a nasty time-warp. This can be observed in a
696  * very small window right after one CPU updated cycle_last under
697  * xtime/vsyscall_gtod lock and the other CPU reads a TSC value which
698  * is smaller than the cycle_last reference value due to a TSC which
699  * is slighty behind. This delta is nowhere else observable, but in
700  * that case it results in a forward time jump in the range of hours
701  * due to the unsigned delta calculation of the time keeping core
702  * code, which is necessary to support wrapping clocksources like pm
703  * timer.
704  */
705 static cycle_t read_tsc(struct clocksource *cs)
706 {
707 	cycle_t ret = (cycle_t)get_cycles();
708 
709 	return ret >= clocksource_tsc.cycle_last ?
710 		ret : clocksource_tsc.cycle_last;
711 }
712 
713 #ifdef CONFIG_X86_64
714 static cycle_t __vsyscall_fn vread_tsc(void)
715 {
716 	cycle_t ret;
717 
718 	/*
719 	 * Surround the RDTSC by barriers, to make sure it's not
720 	 * speculated to outside the seqlock critical section and
721 	 * does not cause time warps:
722 	 */
723 	rdtsc_barrier();
724 	ret = (cycle_t)vget_cycles();
725 	rdtsc_barrier();
726 
727 	return ret >= __vsyscall_gtod_data.clock.cycle_last ?
728 		ret : __vsyscall_gtod_data.clock.cycle_last;
729 }
730 #endif
731 
732 static struct clocksource clocksource_tsc = {
733 	.name                   = "tsc",
734 	.rating                 = 300,
735 	.read                   = read_tsc,
736 	.mask                   = CLOCKSOURCE_MASK(64),
737 	.shift                  = 22,
738 	.flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
739 				  CLOCK_SOURCE_MUST_VERIFY,
740 #ifdef CONFIG_X86_64
741 	.vread                  = vread_tsc,
742 #endif
743 };
744 
745 void mark_tsc_unstable(char *reason)
746 {
747 	if (!tsc_unstable) {
748 		tsc_unstable = 1;
749 		printk("Marking TSC unstable due to %s\n", reason);
750 		/* Change only the rating, when not registered */
751 		if (clocksource_tsc.mult)
752 			clocksource_change_rating(&clocksource_tsc, 0);
753 		else
754 			clocksource_tsc.rating = 0;
755 	}
756 }
757 
758 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
759 
760 static int __init dmi_mark_tsc_unstable(const struct dmi_system_id *d)
761 {
762 	printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
763 			d->ident);
764 	tsc_unstable = 1;
765 	return 0;
766 }
767 
768 /* List of systems that have known TSC problems */
769 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
770 	{
771 		.callback = dmi_mark_tsc_unstable,
772 		.ident = "IBM Thinkpad 380XD",
773 		.matches = {
774 			DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
775 			DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
776 		},
777 	},
778 	{}
779 };
780 
781 static void __init check_system_tsc_reliable(void)
782 {
783 #ifdef CONFIG_MGEODE_LX
784 	/* RTSC counts during suspend */
785 #define RTSC_SUSP 0x100
786 	unsigned long res_low, res_high;
787 
788 	rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
789 	/* Geode_LX - the OLPC CPU has a possibly a very reliable TSC */
790 	if (res_low & RTSC_SUSP)
791 		tsc_clocksource_reliable = 1;
792 #endif
793 	if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE))
794 		tsc_clocksource_reliable = 1;
795 }
796 
797 /*
798  * Make an educated guess if the TSC is trustworthy and synchronized
799  * over all CPUs.
800  */
801 __cpuinit int unsynchronized_tsc(void)
802 {
803 	if (!cpu_has_tsc || tsc_unstable)
804 		return 1;
805 
806 #ifdef CONFIG_SMP
807 	if (apic_is_clustered_box())
808 		return 1;
809 #endif
810 
811 	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
812 		return 0;
813 	/*
814 	 * Intel systems are normally all synchronized.
815 	 * Exceptions must mark TSC as unstable:
816 	 */
817 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
818 		/* assume multi socket systems are not synchronized: */
819 		if (num_possible_cpus() > 1)
820 			tsc_unstable = 1;
821 	}
822 
823 	return tsc_unstable;
824 }
825 
826 static void __init init_tsc_clocksource(void)
827 {
828 	clocksource_tsc.mult = clocksource_khz2mult(tsc_khz,
829 			clocksource_tsc.shift);
830 	if (tsc_clocksource_reliable)
831 		clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
832 	/* lower the rating if we already know its unstable: */
833 	if (check_tsc_unstable()) {
834 		clocksource_tsc.rating = 0;
835 		clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
836 	}
837 	clocksource_register(&clocksource_tsc);
838 }
839 
840 void __init tsc_init(void)
841 {
842 	u64 lpj;
843 	int cpu;
844 
845 	if (!cpu_has_tsc)
846 		return;
847 
848 	tsc_khz = calibrate_tsc();
849 	cpu_khz = tsc_khz;
850 
851 	if (!tsc_khz) {
852 		mark_tsc_unstable("could not calculate TSC khz");
853 		return;
854 	}
855 
856 #ifdef CONFIG_X86_64
857 	if (cpu_has(&boot_cpu_data, X86_FEATURE_CONSTANT_TSC) &&
858 			(boot_cpu_data.x86_vendor == X86_VENDOR_AMD))
859 		cpu_khz = calibrate_cpu();
860 #endif
861 
862 	printk("Detected %lu.%03lu MHz processor.\n",
863 			(unsigned long)cpu_khz / 1000,
864 			(unsigned long)cpu_khz % 1000);
865 
866 	/*
867 	 * Secondary CPUs do not run through tsc_init(), so set up
868 	 * all the scale factors for all CPUs, assuming the same
869 	 * speed as the bootup CPU. (cpufreq notifiers will fix this
870 	 * up if their speed diverges)
871 	 */
872 	for_each_possible_cpu(cpu)
873 		set_cyc2ns_scale(cpu_khz, cpu);
874 
875 	if (tsc_disabled > 0)
876 		return;
877 
878 	/* now allow native_sched_clock() to use rdtsc */
879 	tsc_disabled = 0;
880 
881 	lpj = ((u64)tsc_khz * 1000);
882 	do_div(lpj, HZ);
883 	lpj_fine = lpj;
884 
885 	use_tsc_delay();
886 	/* Check and install the TSC clocksource */
887 	dmi_check_system(bad_tsc_dmi_table);
888 
889 	if (unsynchronized_tsc())
890 		mark_tsc_unstable("TSCs unsynchronized");
891 
892 	check_system_tsc_reliable();
893 	init_tsc_clocksource();
894 }
895 
896