xref: /openbmc/linux/arch/x86/kernel/tsc.c (revision 82ced6fd)
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 
13 #include <asm/hpet.h>
14 #include <asm/timer.h>
15 #include <asm/vgtod.h>
16 #include <asm/time.h>
17 #include <asm/delay.h>
18 #include <asm/hypervisor.h>
19 
20 unsigned int __read_mostly cpu_khz;	/* TSC clocks / usec, not used here */
21 EXPORT_SYMBOL(cpu_khz);
22 
23 unsigned int __read_mostly tsc_khz;
24 EXPORT_SYMBOL(tsc_khz);
25 
26 /*
27  * TSC can be unstable due to cpufreq or due to unsynced TSCs
28  */
29 static int __read_mostly tsc_unstable;
30 
31 /* native_sched_clock() is called before tsc_init(), so
32    we must start with the TSC soft disabled to prevent
33    erroneous rdtsc usage on !cpu_has_tsc processors */
34 static int __read_mostly tsc_disabled = -1;
35 
36 static int tsc_clocksource_reliable;
37 /*
38  * Scheduler clock - returns current time in nanosec units.
39  */
40 u64 native_sched_clock(void)
41 {
42 	u64 this_offset;
43 
44 	/*
45 	 * Fall back to jiffies if there's no TSC available:
46 	 * ( But note that we still use it if the TSC is marked
47 	 *   unstable. We do this because unlike Time Of Day,
48 	 *   the scheduler clock tolerates small errors and it's
49 	 *   very important for it to be as fast as the platform
50 	 *   can achive it. )
51 	 */
52 	if (unlikely(tsc_disabled)) {
53 		/* No locking but a rare wrong value is not a big deal: */
54 		return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
55 	}
56 
57 	/* read the Time Stamp Counter: */
58 	rdtscll(this_offset);
59 
60 	/* return the value in ns */
61 	return __cycles_2_ns(this_offset);
62 }
63 
64 /* We need to define a real function for sched_clock, to override the
65    weak default version */
66 #ifdef CONFIG_PARAVIRT
67 unsigned long long sched_clock(void)
68 {
69 	return paravirt_sched_clock();
70 }
71 #else
72 unsigned long long
73 sched_clock(void) __attribute__((alias("native_sched_clock")));
74 #endif
75 
76 int check_tsc_unstable(void)
77 {
78 	return tsc_unstable;
79 }
80 EXPORT_SYMBOL_GPL(check_tsc_unstable);
81 
82 #ifdef CONFIG_X86_TSC
83 int __init notsc_setup(char *str)
84 {
85 	printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, "
86 			"cannot disable TSC completely.\n");
87 	tsc_disabled = 1;
88 	return 1;
89 }
90 #else
91 /*
92  * disable flag for tsc. Takes effect by clearing the TSC cpu flag
93  * in cpu/common.c
94  */
95 int __init notsc_setup(char *str)
96 {
97 	setup_clear_cpu_cap(X86_FEATURE_TSC);
98 	return 1;
99 }
100 #endif
101 
102 __setup("notsc", notsc_setup);
103 
104 static int __init tsc_setup(char *str)
105 {
106 	if (!strcmp(str, "reliable"))
107 		tsc_clocksource_reliable = 1;
108 	return 1;
109 }
110 
111 __setup("tsc=", tsc_setup);
112 
113 #define MAX_RETRIES     5
114 #define SMI_TRESHOLD    50000
115 
116 /*
117  * Read TSC and the reference counters. Take care of SMI disturbance
118  */
119 static u64 tsc_read_refs(u64 *p, int hpet)
120 {
121 	u64 t1, t2;
122 	int i;
123 
124 	for (i = 0; i < MAX_RETRIES; i++) {
125 		t1 = get_cycles();
126 		if (hpet)
127 			*p = hpet_readl(HPET_COUNTER) & 0xFFFFFFFF;
128 		else
129 			*p = acpi_pm_read_early();
130 		t2 = get_cycles();
131 		if ((t2 - t1) < SMI_TRESHOLD)
132 			return t2;
133 	}
134 	return ULLONG_MAX;
135 }
136 
137 /*
138  * Calculate the TSC frequency from HPET reference
139  */
140 static unsigned long calc_hpet_ref(u64 deltatsc, u64 hpet1, u64 hpet2)
141 {
142 	u64 tmp;
143 
144 	if (hpet2 < hpet1)
145 		hpet2 += 0x100000000ULL;
146 	hpet2 -= hpet1;
147 	tmp = ((u64)hpet2 * hpet_readl(HPET_PERIOD));
148 	do_div(tmp, 1000000);
149 	do_div(deltatsc, tmp);
150 
151 	return (unsigned long) deltatsc;
152 }
153 
154 /*
155  * Calculate the TSC frequency from PMTimer reference
156  */
157 static unsigned long calc_pmtimer_ref(u64 deltatsc, u64 pm1, u64 pm2)
158 {
159 	u64 tmp;
160 
161 	if (!pm1 && !pm2)
162 		return ULONG_MAX;
163 
164 	if (pm2 < pm1)
165 		pm2 += (u64)ACPI_PM_OVRRUN;
166 	pm2 -= pm1;
167 	tmp = pm2 * 1000000000LL;
168 	do_div(tmp, PMTMR_TICKS_PER_SEC);
169 	do_div(deltatsc, tmp);
170 
171 	return (unsigned long) deltatsc;
172 }
173 
174 #define CAL_MS		10
175 #define CAL_LATCH	(CLOCK_TICK_RATE / (1000 / CAL_MS))
176 #define CAL_PIT_LOOPS	1000
177 
178 #define CAL2_MS		50
179 #define CAL2_LATCH	(CLOCK_TICK_RATE / (1000 / CAL2_MS))
180 #define CAL2_PIT_LOOPS	5000
181 
182 
183 /*
184  * Try to calibrate the TSC against the Programmable
185  * Interrupt Timer and return the frequency of the TSC
186  * in kHz.
187  *
188  * Return ULONG_MAX on failure to calibrate.
189  */
190 static unsigned long pit_calibrate_tsc(u32 latch, unsigned long ms, int loopmin)
191 {
192 	u64 tsc, t1, t2, delta;
193 	unsigned long tscmin, tscmax;
194 	int pitcnt;
195 
196 	/* Set the Gate high, disable speaker */
197 	outb((inb(0x61) & ~0x02) | 0x01, 0x61);
198 
199 	/*
200 	 * Setup CTC channel 2* for mode 0, (interrupt on terminal
201 	 * count mode), binary count. Set the latch register to 50ms
202 	 * (LSB then MSB) to begin countdown.
203 	 */
204 	outb(0xb0, 0x43);
205 	outb(latch & 0xff, 0x42);
206 	outb(latch >> 8, 0x42);
207 
208 	tsc = t1 = t2 = get_cycles();
209 
210 	pitcnt = 0;
211 	tscmax = 0;
212 	tscmin = ULONG_MAX;
213 	while ((inb(0x61) & 0x20) == 0) {
214 		t2 = get_cycles();
215 		delta = t2 - tsc;
216 		tsc = t2;
217 		if ((unsigned long) delta < tscmin)
218 			tscmin = (unsigned int) delta;
219 		if ((unsigned long) delta > tscmax)
220 			tscmax = (unsigned int) delta;
221 		pitcnt++;
222 	}
223 
224 	/*
225 	 * Sanity checks:
226 	 *
227 	 * If we were not able to read the PIT more than loopmin
228 	 * times, then we have been hit by a massive SMI
229 	 *
230 	 * If the maximum is 10 times larger than the minimum,
231 	 * then we got hit by an SMI as well.
232 	 */
233 	if (pitcnt < loopmin || tscmax > 10 * tscmin)
234 		return ULONG_MAX;
235 
236 	/* Calculate the PIT value */
237 	delta = t2 - t1;
238 	do_div(delta, ms);
239 	return delta;
240 }
241 
242 /*
243  * This reads the current MSB of the PIT counter, and
244  * checks if we are running on sufficiently fast and
245  * non-virtualized hardware.
246  *
247  * Our expectations are:
248  *
249  *  - the PIT is running at roughly 1.19MHz
250  *
251  *  - each IO is going to take about 1us on real hardware,
252  *    but we allow it to be much faster (by a factor of 10) or
253  *    _slightly_ slower (ie we allow up to a 2us read+counter
254  *    update - anything else implies a unacceptably slow CPU
255  *    or PIT for the fast calibration to work.
256  *
257  *  - with 256 PIT ticks to read the value, we have 214us to
258  *    see the same MSB (and overhead like doing a single TSC
259  *    read per MSB value etc).
260  *
261  *  - We're doing 2 reads per loop (LSB, MSB), and we expect
262  *    them each to take about a microsecond on real hardware.
263  *    So we expect a count value of around 100. But we'll be
264  *    generous, and accept anything over 50.
265  *
266  *  - if the PIT is stuck, and we see *many* more reads, we
267  *    return early (and the next caller of pit_expect_msb()
268  *    then consider it a failure when they don't see the
269  *    next expected value).
270  *
271  * These expectations mean that we know that we have seen the
272  * transition from one expected value to another with a fairly
273  * high accuracy, and we didn't miss any events. We can thus
274  * use the TSC value at the transitions to calculate a pretty
275  * good value for the TSC frequencty.
276  */
277 static inline int pit_expect_msb(unsigned char val, u64 *tscp, unsigned long *deltap)
278 {
279 	int count;
280 	u64 tsc = 0;
281 
282 	for (count = 0; count < 50000; count++) {
283 		/* Ignore LSB */
284 		inb(0x42);
285 		if (inb(0x42) != val)
286 			break;
287 		tsc = get_cycles();
288 	}
289 	*deltap = get_cycles() - tsc;
290 	*tscp = tsc;
291 
292 	/*
293 	 * We require _some_ success, but the quality control
294 	 * will be based on the error terms on the TSC values.
295 	 */
296 	return count > 5;
297 }
298 
299 /*
300  * How many MSB values do we want to see? We aim for
301  * a maximum error rate of 500ppm (in practice the
302  * real error is much smaller), but refuse to spend
303  * more than 25ms on it.
304  */
305 #define MAX_QUICK_PIT_MS 25
306 #define MAX_QUICK_PIT_ITERATIONS (MAX_QUICK_PIT_MS * PIT_TICK_RATE / 1000 / 256)
307 
308 static unsigned long quick_pit_calibrate(void)
309 {
310 	int i;
311 	u64 tsc, delta;
312 	unsigned long d1, d2;
313 
314 	/* Set the Gate high, disable speaker */
315 	outb((inb(0x61) & ~0x02) | 0x01, 0x61);
316 
317 	/*
318 	 * Counter 2, mode 0 (one-shot), binary count
319 	 *
320 	 * NOTE! Mode 2 decrements by two (and then the
321 	 * output is flipped each time, giving the same
322 	 * final output frequency as a decrement-by-one),
323 	 * so mode 0 is much better when looking at the
324 	 * individual counts.
325 	 */
326 	outb(0xb0, 0x43);
327 
328 	/* Start at 0xffff */
329 	outb(0xff, 0x42);
330 	outb(0xff, 0x42);
331 
332 	/*
333 	 * The PIT starts counting at the next edge, so we
334 	 * need to delay for a microsecond. The easiest way
335 	 * to do that is to just read back the 16-bit counter
336 	 * once from the PIT.
337 	 */
338 	inb(0x42);
339 	inb(0x42);
340 
341 	if (pit_expect_msb(0xff, &tsc, &d1)) {
342 		for (i = 1; i <= MAX_QUICK_PIT_ITERATIONS; i++) {
343 			if (!pit_expect_msb(0xff-i, &delta, &d2))
344 				break;
345 
346 			/*
347 			 * Iterate until the error is less than 500 ppm
348 			 */
349 			delta -= tsc;
350 			if (d1+d2 < delta >> 11)
351 				goto success;
352 		}
353 	}
354 	printk("Fast TSC calibration failed\n");
355 	return 0;
356 
357 success:
358 	/*
359 	 * Ok, if we get here, then we've seen the
360 	 * MSB of the PIT decrement 'i' times, and the
361 	 * error has shrunk to less than 500 ppm.
362 	 *
363 	 * As a result, we can depend on there not being
364 	 * any odd delays anywhere, and the TSC reads are
365 	 * reliable (within the error). We also adjust the
366 	 * delta to the middle of the error bars, just
367 	 * because it looks nicer.
368 	 *
369 	 * kHz = ticks / time-in-seconds / 1000;
370 	 * kHz = (t2 - t1) / (I * 256 / PIT_TICK_RATE) / 1000
371 	 * kHz = ((t2 - t1) * PIT_TICK_RATE) / (I * 256 * 1000)
372 	 */
373 	delta += (long)(d2 - d1)/2;
374 	delta *= PIT_TICK_RATE;
375 	do_div(delta, i*256*1000);
376 	printk("Fast TSC calibration using PIT\n");
377 	return delta;
378 }
379 
380 /**
381  * native_calibrate_tsc - calibrate the tsc on boot
382  */
383 unsigned long native_calibrate_tsc(void)
384 {
385 	u64 tsc1, tsc2, delta, ref1, ref2;
386 	unsigned long tsc_pit_min = ULONG_MAX, tsc_ref_min = ULONG_MAX;
387 	unsigned long flags, latch, ms, fast_calibrate, tsc_khz;
388 	int hpet = is_hpet_enabled(), i, loopmin;
389 
390 	tsc_khz = get_hypervisor_tsc_freq();
391 	if (tsc_khz) {
392 		printk(KERN_INFO "TSC: Frequency read from the hypervisor\n");
393 		return tsc_khz;
394 	}
395 
396 	local_irq_save(flags);
397 	fast_calibrate = quick_pit_calibrate();
398 	local_irq_restore(flags);
399 	if (fast_calibrate)
400 		return fast_calibrate;
401 
402 	/*
403 	 * Run 5 calibration loops to get the lowest frequency value
404 	 * (the best estimate). We use two different calibration modes
405 	 * here:
406 	 *
407 	 * 1) PIT loop. We set the PIT Channel 2 to oneshot mode and
408 	 * load a timeout of 50ms. We read the time right after we
409 	 * started the timer and wait until the PIT count down reaches
410 	 * zero. In each wait loop iteration we read the TSC and check
411 	 * the delta to the previous read. We keep track of the min
412 	 * and max values of that delta. The delta is mostly defined
413 	 * by the IO time of the PIT access, so we can detect when a
414 	 * SMI/SMM disturbance happend between the two reads. If the
415 	 * maximum time is significantly larger than the minimum time,
416 	 * then we discard the result and have another try.
417 	 *
418 	 * 2) Reference counter. If available we use the HPET or the
419 	 * PMTIMER as a reference to check the sanity of that value.
420 	 * We use separate TSC readouts and check inside of the
421 	 * reference read for a SMI/SMM disturbance. We dicard
422 	 * disturbed values here as well. We do that around the PIT
423 	 * calibration delay loop as we have to wait for a certain
424 	 * amount of time anyway.
425 	 */
426 
427 	/* Preset PIT loop values */
428 	latch = CAL_LATCH;
429 	ms = CAL_MS;
430 	loopmin = CAL_PIT_LOOPS;
431 
432 	for (i = 0; i < 3; i++) {
433 		unsigned long tsc_pit_khz;
434 
435 		/*
436 		 * Read the start value and the reference count of
437 		 * hpet/pmtimer when available. Then do the PIT
438 		 * calibration, which will take at least 50ms, and
439 		 * read the end value.
440 		 */
441 		local_irq_save(flags);
442 		tsc1 = tsc_read_refs(&ref1, hpet);
443 		tsc_pit_khz = pit_calibrate_tsc(latch, ms, loopmin);
444 		tsc2 = tsc_read_refs(&ref2, hpet);
445 		local_irq_restore(flags);
446 
447 		/* Pick the lowest PIT TSC calibration so far */
448 		tsc_pit_min = min(tsc_pit_min, tsc_pit_khz);
449 
450 		/* hpet or pmtimer available ? */
451 		if (!hpet && !ref1 && !ref2)
452 			continue;
453 
454 		/* Check, whether the sampling was disturbed by an SMI */
455 		if (tsc1 == ULLONG_MAX || tsc2 == ULLONG_MAX)
456 			continue;
457 
458 		tsc2 = (tsc2 - tsc1) * 1000000LL;
459 		if (hpet)
460 			tsc2 = calc_hpet_ref(tsc2, ref1, ref2);
461 		else
462 			tsc2 = calc_pmtimer_ref(tsc2, ref1, ref2);
463 
464 		tsc_ref_min = min(tsc_ref_min, (unsigned long) tsc2);
465 
466 		/* Check the reference deviation */
467 		delta = ((u64) tsc_pit_min) * 100;
468 		do_div(delta, tsc_ref_min);
469 
470 		/*
471 		 * If both calibration results are inside a 10% window
472 		 * then we can be sure, that the calibration
473 		 * succeeded. We break out of the loop right away. We
474 		 * use the reference value, as it is more precise.
475 		 */
476 		if (delta >= 90 && delta <= 110) {
477 			printk(KERN_INFO
478 			       "TSC: PIT calibration matches %s. %d loops\n",
479 			       hpet ? "HPET" : "PMTIMER", i + 1);
480 			return tsc_ref_min;
481 		}
482 
483 		/*
484 		 * Check whether PIT failed more than once. This
485 		 * happens in virtualized environments. We need to
486 		 * give the virtual PC a slightly longer timeframe for
487 		 * the HPET/PMTIMER to make the result precise.
488 		 */
489 		if (i == 1 && tsc_pit_min == ULONG_MAX) {
490 			latch = CAL2_LATCH;
491 			ms = CAL2_MS;
492 			loopmin = CAL2_PIT_LOOPS;
493 		}
494 	}
495 
496 	/*
497 	 * Now check the results.
498 	 */
499 	if (tsc_pit_min == ULONG_MAX) {
500 		/* PIT gave no useful value */
501 		printk(KERN_WARNING "TSC: Unable to calibrate against PIT\n");
502 
503 		/* We don't have an alternative source, disable TSC */
504 		if (!hpet && !ref1 && !ref2) {
505 			printk("TSC: No reference (HPET/PMTIMER) available\n");
506 			return 0;
507 		}
508 
509 		/* The alternative source failed as well, disable TSC */
510 		if (tsc_ref_min == ULONG_MAX) {
511 			printk(KERN_WARNING "TSC: HPET/PMTIMER calibration "
512 			       "failed.\n");
513 			return 0;
514 		}
515 
516 		/* Use the alternative source */
517 		printk(KERN_INFO "TSC: using %s reference calibration\n",
518 		       hpet ? "HPET" : "PMTIMER");
519 
520 		return tsc_ref_min;
521 	}
522 
523 	/* We don't have an alternative source, use the PIT calibration value */
524 	if (!hpet && !ref1 && !ref2) {
525 		printk(KERN_INFO "TSC: Using PIT calibration value\n");
526 		return tsc_pit_min;
527 	}
528 
529 	/* The alternative source failed, use the PIT calibration value */
530 	if (tsc_ref_min == ULONG_MAX) {
531 		printk(KERN_WARNING "TSC: HPET/PMTIMER calibration failed. "
532 		       "Using PIT calibration\n");
533 		return tsc_pit_min;
534 	}
535 
536 	/*
537 	 * The calibration values differ too much. In doubt, we use
538 	 * the PIT value as we know that there are PMTIMERs around
539 	 * running at double speed. At least we let the user know:
540 	 */
541 	printk(KERN_WARNING "TSC: PIT calibration deviates from %s: %lu %lu.\n",
542 	       hpet ? "HPET" : "PMTIMER", tsc_pit_min, tsc_ref_min);
543 	printk(KERN_INFO "TSC: Using PIT calibration value\n");
544 	return tsc_pit_min;
545 }
546 
547 int recalibrate_cpu_khz(void)
548 {
549 #ifndef CONFIG_SMP
550 	unsigned long cpu_khz_old = cpu_khz;
551 
552 	if (cpu_has_tsc) {
553 		tsc_khz = calibrate_tsc();
554 		cpu_khz = tsc_khz;
555 		cpu_data(0).loops_per_jiffy =
556 			cpufreq_scale(cpu_data(0).loops_per_jiffy,
557 					cpu_khz_old, cpu_khz);
558 		return 0;
559 	} else
560 		return -ENODEV;
561 #else
562 	return -ENODEV;
563 #endif
564 }
565 
566 EXPORT_SYMBOL(recalibrate_cpu_khz);
567 
568 
569 /* Accelerators for sched_clock()
570  * convert from cycles(64bits) => nanoseconds (64bits)
571  *  basic equation:
572  *              ns = cycles / (freq / ns_per_sec)
573  *              ns = cycles * (ns_per_sec / freq)
574  *              ns = cycles * (10^9 / (cpu_khz * 10^3))
575  *              ns = cycles * (10^6 / cpu_khz)
576  *
577  *      Then we use scaling math (suggested by george@mvista.com) to get:
578  *              ns = cycles * (10^6 * SC / cpu_khz) / SC
579  *              ns = cycles * cyc2ns_scale / SC
580  *
581  *      And since SC is a constant power of two, we can convert the div
582  *  into a shift.
583  *
584  *  We can use khz divisor instead of mhz to keep a better precision, since
585  *  cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
586  *  (mathieu.desnoyers@polymtl.ca)
587  *
588  *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
589  */
590 
591 DEFINE_PER_CPU(unsigned long, cyc2ns);
592 
593 static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu)
594 {
595 	unsigned long long tsc_now, ns_now;
596 	unsigned long flags, *scale;
597 
598 	local_irq_save(flags);
599 	sched_clock_idle_sleep_event();
600 
601 	scale = &per_cpu(cyc2ns, cpu);
602 
603 	rdtscll(tsc_now);
604 	ns_now = __cycles_2_ns(tsc_now);
605 
606 	if (cpu_khz)
607 		*scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz;
608 
609 	sched_clock_idle_wakeup_event(0);
610 	local_irq_restore(flags);
611 }
612 
613 #ifdef CONFIG_CPU_FREQ
614 
615 /* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
616  * changes.
617  *
618  * RED-PEN: On SMP we assume all CPUs run with the same frequency.  It's
619  * not that important because current Opteron setups do not support
620  * scaling on SMP anyroads.
621  *
622  * Should fix up last_tsc too. Currently gettimeofday in the
623  * first tick after the change will be slightly wrong.
624  */
625 
626 static unsigned int  ref_freq;
627 static unsigned long loops_per_jiffy_ref;
628 static unsigned long tsc_khz_ref;
629 
630 static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
631 				void *data)
632 {
633 	struct cpufreq_freqs *freq = data;
634 	unsigned long *lpj, dummy;
635 
636 	if (cpu_has(&cpu_data(freq->cpu), X86_FEATURE_CONSTANT_TSC))
637 		return 0;
638 
639 	lpj = &dummy;
640 	if (!(freq->flags & CPUFREQ_CONST_LOOPS))
641 #ifdef CONFIG_SMP
642 		lpj = &cpu_data(freq->cpu).loops_per_jiffy;
643 #else
644 	lpj = &boot_cpu_data.loops_per_jiffy;
645 #endif
646 
647 	if (!ref_freq) {
648 		ref_freq = freq->old;
649 		loops_per_jiffy_ref = *lpj;
650 		tsc_khz_ref = tsc_khz;
651 	}
652 	if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
653 			(val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
654 			(val == CPUFREQ_RESUMECHANGE)) {
655 		*lpj = 	cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
656 
657 		tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
658 		if (!(freq->flags & CPUFREQ_CONST_LOOPS))
659 			mark_tsc_unstable("cpufreq changes");
660 	}
661 
662 	set_cyc2ns_scale(tsc_khz, freq->cpu);
663 
664 	return 0;
665 }
666 
667 static struct notifier_block time_cpufreq_notifier_block = {
668 	.notifier_call  = time_cpufreq_notifier
669 };
670 
671 static int __init cpufreq_tsc(void)
672 {
673 	if (!cpu_has_tsc)
674 		return 0;
675 	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
676 		return 0;
677 	cpufreq_register_notifier(&time_cpufreq_notifier_block,
678 				CPUFREQ_TRANSITION_NOTIFIER);
679 	return 0;
680 }
681 
682 core_initcall(cpufreq_tsc);
683 
684 #endif /* CONFIG_CPU_FREQ */
685 
686 /* clocksource code */
687 
688 static struct clocksource clocksource_tsc;
689 
690 /*
691  * We compare the TSC to the cycle_last value in the clocksource
692  * structure to avoid a nasty time-warp. This can be observed in a
693  * very small window right after one CPU updated cycle_last under
694  * xtime/vsyscall_gtod lock and the other CPU reads a TSC value which
695  * is smaller than the cycle_last reference value due to a TSC which
696  * is slighty behind. This delta is nowhere else observable, but in
697  * that case it results in a forward time jump in the range of hours
698  * due to the unsigned delta calculation of the time keeping core
699  * code, which is necessary to support wrapping clocksources like pm
700  * timer.
701  */
702 static cycle_t read_tsc(struct clocksource *cs)
703 {
704 	cycle_t ret = (cycle_t)get_cycles();
705 
706 	return ret >= clocksource_tsc.cycle_last ?
707 		ret : clocksource_tsc.cycle_last;
708 }
709 
710 #ifdef CONFIG_X86_64
711 static cycle_t __vsyscall_fn vread_tsc(void)
712 {
713 	cycle_t ret = (cycle_t)vget_cycles();
714 
715 	return ret >= __vsyscall_gtod_data.clock.cycle_last ?
716 		ret : __vsyscall_gtod_data.clock.cycle_last;
717 }
718 #endif
719 
720 static struct clocksource clocksource_tsc = {
721 	.name                   = "tsc",
722 	.rating                 = 300,
723 	.read                   = read_tsc,
724 	.mask                   = CLOCKSOURCE_MASK(64),
725 	.shift                  = 22,
726 	.flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
727 				  CLOCK_SOURCE_MUST_VERIFY,
728 #ifdef CONFIG_X86_64
729 	.vread                  = vread_tsc,
730 #endif
731 };
732 
733 void mark_tsc_unstable(char *reason)
734 {
735 	if (!tsc_unstable) {
736 		tsc_unstable = 1;
737 		printk("Marking TSC unstable due to %s\n", reason);
738 		/* Change only the rating, when not registered */
739 		if (clocksource_tsc.mult)
740 			clocksource_change_rating(&clocksource_tsc, 0);
741 		else
742 			clocksource_tsc.rating = 0;
743 	}
744 }
745 
746 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
747 
748 static int __init dmi_mark_tsc_unstable(const struct dmi_system_id *d)
749 {
750 	printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
751 			d->ident);
752 	tsc_unstable = 1;
753 	return 0;
754 }
755 
756 /* List of systems that have known TSC problems */
757 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
758 	{
759 		.callback = dmi_mark_tsc_unstable,
760 		.ident = "IBM Thinkpad 380XD",
761 		.matches = {
762 			DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
763 			DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
764 		},
765 	},
766 	{}
767 };
768 
769 static void __init check_system_tsc_reliable(void)
770 {
771 #ifdef CONFIG_MGEODE_LX
772 	/* RTSC counts during suspend */
773 #define RTSC_SUSP 0x100
774 	unsigned long res_low, res_high;
775 
776 	rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
777 	/* Geode_LX - the OLPC CPU has a possibly a very reliable TSC */
778 	if (res_low & RTSC_SUSP)
779 		tsc_clocksource_reliable = 1;
780 #endif
781 	if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE))
782 		tsc_clocksource_reliable = 1;
783 }
784 
785 /*
786  * Make an educated guess if the TSC is trustworthy and synchronized
787  * over all CPUs.
788  */
789 __cpuinit int unsynchronized_tsc(void)
790 {
791 	if (!cpu_has_tsc || tsc_unstable)
792 		return 1;
793 
794 #ifdef CONFIG_SMP
795 	if (apic_is_clustered_box())
796 		return 1;
797 #endif
798 
799 	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
800 		return 0;
801 	/*
802 	 * Intel systems are normally all synchronized.
803 	 * Exceptions must mark TSC as unstable:
804 	 */
805 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
806 		/* assume multi socket systems are not synchronized: */
807 		if (num_possible_cpus() > 1)
808 			tsc_unstable = 1;
809 	}
810 
811 	return tsc_unstable;
812 }
813 
814 static void __init init_tsc_clocksource(void)
815 {
816 	clocksource_tsc.mult = clocksource_khz2mult(tsc_khz,
817 			clocksource_tsc.shift);
818 	if (tsc_clocksource_reliable)
819 		clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
820 	/* lower the rating if we already know its unstable: */
821 	if (check_tsc_unstable()) {
822 		clocksource_tsc.rating = 0;
823 		clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
824 	}
825 	clocksource_register(&clocksource_tsc);
826 }
827 
828 void __init tsc_init(void)
829 {
830 	u64 lpj;
831 	int cpu;
832 
833 	if (!cpu_has_tsc)
834 		return;
835 
836 	tsc_khz = calibrate_tsc();
837 	cpu_khz = tsc_khz;
838 
839 	if (!tsc_khz) {
840 		mark_tsc_unstable("could not calculate TSC khz");
841 		return;
842 	}
843 
844 #ifdef CONFIG_X86_64
845 	if (cpu_has(&boot_cpu_data, X86_FEATURE_CONSTANT_TSC) &&
846 			(boot_cpu_data.x86_vendor == X86_VENDOR_AMD))
847 		cpu_khz = calibrate_cpu();
848 #endif
849 
850 	printk("Detected %lu.%03lu MHz processor.\n",
851 			(unsigned long)cpu_khz / 1000,
852 			(unsigned long)cpu_khz % 1000);
853 
854 	/*
855 	 * Secondary CPUs do not run through tsc_init(), so set up
856 	 * all the scale factors for all CPUs, assuming the same
857 	 * speed as the bootup CPU. (cpufreq notifiers will fix this
858 	 * up if their speed diverges)
859 	 */
860 	for_each_possible_cpu(cpu)
861 		set_cyc2ns_scale(cpu_khz, cpu);
862 
863 	if (tsc_disabled > 0)
864 		return;
865 
866 	/* now allow native_sched_clock() to use rdtsc */
867 	tsc_disabled = 0;
868 
869 	lpj = ((u64)tsc_khz * 1000);
870 	do_div(lpj, HZ);
871 	lpj_fine = lpj;
872 
873 	use_tsc_delay();
874 	/* Check and install the TSC clocksource */
875 	dmi_check_system(bad_tsc_dmi_table);
876 
877 	if (unsynchronized_tsc())
878 		mark_tsc_unstable("TSCs unsynchronized");
879 
880 	check_system_tsc_reliable();
881 	init_tsc_clocksource();
882 }
883 
884