xref: /linux/arch/x86/kernel/tsc.c (revision 81fa7a69c2174ed8de314b9c231ef30a8718e5e1)
1 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
2 
3 #include <linux/kernel.h>
4 #include <linux/sched.h>
5 #include <linux/sched/clock.h>
6 #include <linux/init.h>
7 #include <linux/export.h>
8 #include <linux/timer.h>
9 #include <linux/acpi_pmtmr.h>
10 #include <linux/cpufreq.h>
11 #include <linux/delay.h>
12 #include <linux/clocksource.h>
13 #include <linux/percpu.h>
14 #include <linux/timex.h>
15 #include <linux/static_key.h>
16 
17 #include <asm/hpet.h>
18 #include <asm/timer.h>
19 #include <asm/vgtod.h>
20 #include <asm/time.h>
21 #include <asm/delay.h>
22 #include <asm/hypervisor.h>
23 #include <asm/nmi.h>
24 #include <asm/x86_init.h>
25 #include <asm/geode.h>
26 #include <asm/apic.h>
27 #include <asm/intel-family.h>
28 #include <asm/i8259.h>
29 #include <asm/uv/uv.h>
30 
31 unsigned int __read_mostly cpu_khz;	/* TSC clocks / usec, not used here */
32 EXPORT_SYMBOL(cpu_khz);
33 
34 unsigned int __read_mostly tsc_khz;
35 EXPORT_SYMBOL(tsc_khz);
36 
37 #define KHZ	1000
38 
39 /*
40  * TSC can be unstable due to cpufreq or due to unsynced TSCs
41  */
42 static int __read_mostly tsc_unstable;
43 
44 static DEFINE_STATIC_KEY_FALSE(__use_tsc);
45 
46 int tsc_clocksource_reliable;
47 
48 static u32 art_to_tsc_numerator;
49 static u32 art_to_tsc_denominator;
50 static u64 art_to_tsc_offset;
51 struct clocksource *art_related_clocksource;
52 
53 struct cyc2ns {
54 	struct cyc2ns_data data[2];	/*  0 + 2*16 = 32 */
55 	seqcount_t	   seq;		/* 32 + 4    = 36 */
56 
57 }; /* fits one cacheline */
58 
59 static DEFINE_PER_CPU_ALIGNED(struct cyc2ns, cyc2ns);
60 
61 void __always_inline cyc2ns_read_begin(struct cyc2ns_data *data)
62 {
63 	int seq, idx;
64 
65 	preempt_disable_notrace();
66 
67 	do {
68 		seq = this_cpu_read(cyc2ns.seq.sequence);
69 		idx = seq & 1;
70 
71 		data->cyc2ns_offset = this_cpu_read(cyc2ns.data[idx].cyc2ns_offset);
72 		data->cyc2ns_mul    = this_cpu_read(cyc2ns.data[idx].cyc2ns_mul);
73 		data->cyc2ns_shift  = this_cpu_read(cyc2ns.data[idx].cyc2ns_shift);
74 
75 	} while (unlikely(seq != this_cpu_read(cyc2ns.seq.sequence)));
76 }
77 
78 void __always_inline cyc2ns_read_end(void)
79 {
80 	preempt_enable_notrace();
81 }
82 
83 /*
84  * Accelerators for sched_clock()
85  * convert from cycles(64bits) => nanoseconds (64bits)
86  *  basic equation:
87  *              ns = cycles / (freq / ns_per_sec)
88  *              ns = cycles * (ns_per_sec / freq)
89  *              ns = cycles * (10^9 / (cpu_khz * 10^3))
90  *              ns = cycles * (10^6 / cpu_khz)
91  *
92  *      Then we use scaling math (suggested by george@mvista.com) to get:
93  *              ns = cycles * (10^6 * SC / cpu_khz) / SC
94  *              ns = cycles * cyc2ns_scale / SC
95  *
96  *      And since SC is a constant power of two, we can convert the div
97  *  into a shift. The larger SC is, the more accurate the conversion, but
98  *  cyc2ns_scale needs to be a 32-bit value so that 32-bit multiplication
99  *  (64-bit result) can be used.
100  *
101  *  We can use khz divisor instead of mhz to keep a better precision.
102  *  (mathieu.desnoyers@polymtl.ca)
103  *
104  *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
105  */
106 
107 static __always_inline unsigned long long cycles_2_ns(unsigned long long cyc)
108 {
109 	struct cyc2ns_data data;
110 	unsigned long long ns;
111 
112 	cyc2ns_read_begin(&data);
113 
114 	ns = data.cyc2ns_offset;
115 	ns += mul_u64_u32_shr(cyc, data.cyc2ns_mul, data.cyc2ns_shift);
116 
117 	cyc2ns_read_end();
118 
119 	return ns;
120 }
121 
122 static void __set_cyc2ns_scale(unsigned long khz, int cpu, unsigned long long tsc_now)
123 {
124 	unsigned long long ns_now;
125 	struct cyc2ns_data data;
126 	struct cyc2ns *c2n;
127 
128 	ns_now = cycles_2_ns(tsc_now);
129 
130 	/*
131 	 * Compute a new multiplier as per the above comment and ensure our
132 	 * time function is continuous; see the comment near struct
133 	 * cyc2ns_data.
134 	 */
135 	clocks_calc_mult_shift(&data.cyc2ns_mul, &data.cyc2ns_shift, khz,
136 			       NSEC_PER_MSEC, 0);
137 
138 	/*
139 	 * cyc2ns_shift is exported via arch_perf_update_userpage() where it is
140 	 * not expected to be greater than 31 due to the original published
141 	 * conversion algorithm shifting a 32-bit value (now specifies a 64-bit
142 	 * value) - refer perf_event_mmap_page documentation in perf_event.h.
143 	 */
144 	if (data.cyc2ns_shift == 32) {
145 		data.cyc2ns_shift = 31;
146 		data.cyc2ns_mul >>= 1;
147 	}
148 
149 	data.cyc2ns_offset = ns_now -
150 		mul_u64_u32_shr(tsc_now, data.cyc2ns_mul, data.cyc2ns_shift);
151 
152 	c2n = per_cpu_ptr(&cyc2ns, cpu);
153 
154 	raw_write_seqcount_latch(&c2n->seq);
155 	c2n->data[0] = data;
156 	raw_write_seqcount_latch(&c2n->seq);
157 	c2n->data[1] = data;
158 }
159 
160 static void set_cyc2ns_scale(unsigned long khz, int cpu, unsigned long long tsc_now)
161 {
162 	unsigned long flags;
163 
164 	local_irq_save(flags);
165 	sched_clock_idle_sleep_event();
166 
167 	if (khz)
168 		__set_cyc2ns_scale(khz, cpu, tsc_now);
169 
170 	sched_clock_idle_wakeup_event();
171 	local_irq_restore(flags);
172 }
173 
174 /*
175  * Initialize cyc2ns for boot cpu
176  */
177 static void __init cyc2ns_init_boot_cpu(void)
178 {
179 	struct cyc2ns *c2n = this_cpu_ptr(&cyc2ns);
180 
181 	seqcount_init(&c2n->seq);
182 	__set_cyc2ns_scale(tsc_khz, smp_processor_id(), rdtsc());
183 }
184 
185 /*
186  * Secondary CPUs do not run through tsc_init(), so set up
187  * all the scale factors for all CPUs, assuming the same
188  * speed as the bootup CPU. (cpufreq notifiers will fix this
189  * up if their speed diverges)
190  */
191 static void __init cyc2ns_init_secondary_cpus(void)
192 {
193 	unsigned int cpu, this_cpu = smp_processor_id();
194 	struct cyc2ns *c2n = this_cpu_ptr(&cyc2ns);
195 	struct cyc2ns_data *data = c2n->data;
196 
197 	for_each_possible_cpu(cpu) {
198 		if (cpu != this_cpu) {
199 			seqcount_init(&c2n->seq);
200 			c2n = per_cpu_ptr(&cyc2ns, cpu);
201 			c2n->data[0] = data[0];
202 			c2n->data[1] = data[1];
203 		}
204 	}
205 }
206 
207 /*
208  * Scheduler clock - returns current time in nanosec units.
209  */
210 u64 native_sched_clock(void)
211 {
212 	if (static_branch_likely(&__use_tsc)) {
213 		u64 tsc_now = rdtsc();
214 
215 		/* return the value in ns */
216 		return cycles_2_ns(tsc_now);
217 	}
218 
219 	/*
220 	 * Fall back to jiffies if there's no TSC available:
221 	 * ( But note that we still use it if the TSC is marked
222 	 *   unstable. We do this because unlike Time Of Day,
223 	 *   the scheduler clock tolerates small errors and it's
224 	 *   very important for it to be as fast as the platform
225 	 *   can achieve it. )
226 	 */
227 
228 	/* No locking but a rare wrong value is not a big deal: */
229 	return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
230 }
231 
232 /*
233  * Generate a sched_clock if you already have a TSC value.
234  */
235 u64 native_sched_clock_from_tsc(u64 tsc)
236 {
237 	return cycles_2_ns(tsc);
238 }
239 
240 /* We need to define a real function for sched_clock, to override the
241    weak default version */
242 #ifdef CONFIG_PARAVIRT
243 unsigned long long sched_clock(void)
244 {
245 	return paravirt_sched_clock();
246 }
247 
248 bool using_native_sched_clock(void)
249 {
250 	return pv_time_ops.sched_clock == native_sched_clock;
251 }
252 #else
253 unsigned long long
254 sched_clock(void) __attribute__((alias("native_sched_clock")));
255 
256 bool using_native_sched_clock(void) { return true; }
257 #endif
258 
259 int check_tsc_unstable(void)
260 {
261 	return tsc_unstable;
262 }
263 EXPORT_SYMBOL_GPL(check_tsc_unstable);
264 
265 #ifdef CONFIG_X86_TSC
266 int __init notsc_setup(char *str)
267 {
268 	mark_tsc_unstable("boot parameter notsc");
269 	return 1;
270 }
271 #else
272 /*
273  * disable flag for tsc. Takes effect by clearing the TSC cpu flag
274  * in cpu/common.c
275  */
276 int __init notsc_setup(char *str)
277 {
278 	setup_clear_cpu_cap(X86_FEATURE_TSC);
279 	return 1;
280 }
281 #endif
282 
283 __setup("notsc", notsc_setup);
284 
285 static int no_sched_irq_time;
286 
287 static int __init tsc_setup(char *str)
288 {
289 	if (!strcmp(str, "reliable"))
290 		tsc_clocksource_reliable = 1;
291 	if (!strncmp(str, "noirqtime", 9))
292 		no_sched_irq_time = 1;
293 	if (!strcmp(str, "unstable"))
294 		mark_tsc_unstable("boot parameter");
295 	return 1;
296 }
297 
298 __setup("tsc=", tsc_setup);
299 
300 #define MAX_RETRIES     5
301 #define SMI_TRESHOLD    50000
302 
303 /*
304  * Read TSC and the reference counters. Take care of SMI disturbance
305  */
306 static u64 tsc_read_refs(u64 *p, int hpet)
307 {
308 	u64 t1, t2;
309 	int i;
310 
311 	for (i = 0; i < MAX_RETRIES; i++) {
312 		t1 = get_cycles();
313 		if (hpet)
314 			*p = hpet_readl(HPET_COUNTER) & 0xFFFFFFFF;
315 		else
316 			*p = acpi_pm_read_early();
317 		t2 = get_cycles();
318 		if ((t2 - t1) < SMI_TRESHOLD)
319 			return t2;
320 	}
321 	return ULLONG_MAX;
322 }
323 
324 /*
325  * Calculate the TSC frequency from HPET reference
326  */
327 static unsigned long calc_hpet_ref(u64 deltatsc, u64 hpet1, u64 hpet2)
328 {
329 	u64 tmp;
330 
331 	if (hpet2 < hpet1)
332 		hpet2 += 0x100000000ULL;
333 	hpet2 -= hpet1;
334 	tmp = ((u64)hpet2 * hpet_readl(HPET_PERIOD));
335 	do_div(tmp, 1000000);
336 	deltatsc = div64_u64(deltatsc, tmp);
337 
338 	return (unsigned long) deltatsc;
339 }
340 
341 /*
342  * Calculate the TSC frequency from PMTimer reference
343  */
344 static unsigned long calc_pmtimer_ref(u64 deltatsc, u64 pm1, u64 pm2)
345 {
346 	u64 tmp;
347 
348 	if (!pm1 && !pm2)
349 		return ULONG_MAX;
350 
351 	if (pm2 < pm1)
352 		pm2 += (u64)ACPI_PM_OVRRUN;
353 	pm2 -= pm1;
354 	tmp = pm2 * 1000000000LL;
355 	do_div(tmp, PMTMR_TICKS_PER_SEC);
356 	do_div(deltatsc, tmp);
357 
358 	return (unsigned long) deltatsc;
359 }
360 
361 #define CAL_MS		10
362 #define CAL_LATCH	(PIT_TICK_RATE / (1000 / CAL_MS))
363 #define CAL_PIT_LOOPS	1000
364 
365 #define CAL2_MS		50
366 #define CAL2_LATCH	(PIT_TICK_RATE / (1000 / CAL2_MS))
367 #define CAL2_PIT_LOOPS	5000
368 
369 
370 /*
371  * Try to calibrate the TSC against the Programmable
372  * Interrupt Timer and return the frequency of the TSC
373  * in kHz.
374  *
375  * Return ULONG_MAX on failure to calibrate.
376  */
377 static unsigned long pit_calibrate_tsc(u32 latch, unsigned long ms, int loopmin)
378 {
379 	u64 tsc, t1, t2, delta;
380 	unsigned long tscmin, tscmax;
381 	int pitcnt;
382 
383 	if (!has_legacy_pic()) {
384 		/*
385 		 * Relies on tsc_early_delay_calibrate() to have given us semi
386 		 * usable udelay(), wait for the same 50ms we would have with
387 		 * the PIT loop below.
388 		 */
389 		udelay(10 * USEC_PER_MSEC);
390 		udelay(10 * USEC_PER_MSEC);
391 		udelay(10 * USEC_PER_MSEC);
392 		udelay(10 * USEC_PER_MSEC);
393 		udelay(10 * USEC_PER_MSEC);
394 		return ULONG_MAX;
395 	}
396 
397 	/* Set the Gate high, disable speaker */
398 	outb((inb(0x61) & ~0x02) | 0x01, 0x61);
399 
400 	/*
401 	 * Setup CTC channel 2* for mode 0, (interrupt on terminal
402 	 * count mode), binary count. Set the latch register to 50ms
403 	 * (LSB then MSB) to begin countdown.
404 	 */
405 	outb(0xb0, 0x43);
406 	outb(latch & 0xff, 0x42);
407 	outb(latch >> 8, 0x42);
408 
409 	tsc = t1 = t2 = get_cycles();
410 
411 	pitcnt = 0;
412 	tscmax = 0;
413 	tscmin = ULONG_MAX;
414 	while ((inb(0x61) & 0x20) == 0) {
415 		t2 = get_cycles();
416 		delta = t2 - tsc;
417 		tsc = t2;
418 		if ((unsigned long) delta < tscmin)
419 			tscmin = (unsigned int) delta;
420 		if ((unsigned long) delta > tscmax)
421 			tscmax = (unsigned int) delta;
422 		pitcnt++;
423 	}
424 
425 	/*
426 	 * Sanity checks:
427 	 *
428 	 * If we were not able to read the PIT more than loopmin
429 	 * times, then we have been hit by a massive SMI
430 	 *
431 	 * If the maximum is 10 times larger than the minimum,
432 	 * then we got hit by an SMI as well.
433 	 */
434 	if (pitcnt < loopmin || tscmax > 10 * tscmin)
435 		return ULONG_MAX;
436 
437 	/* Calculate the PIT value */
438 	delta = t2 - t1;
439 	do_div(delta, ms);
440 	return delta;
441 }
442 
443 /*
444  * This reads the current MSB of the PIT counter, and
445  * checks if we are running on sufficiently fast and
446  * non-virtualized hardware.
447  *
448  * Our expectations are:
449  *
450  *  - the PIT is running at roughly 1.19MHz
451  *
452  *  - each IO is going to take about 1us on real hardware,
453  *    but we allow it to be much faster (by a factor of 10) or
454  *    _slightly_ slower (ie we allow up to a 2us read+counter
455  *    update - anything else implies a unacceptably slow CPU
456  *    or PIT for the fast calibration to work.
457  *
458  *  - with 256 PIT ticks to read the value, we have 214us to
459  *    see the same MSB (and overhead like doing a single TSC
460  *    read per MSB value etc).
461  *
462  *  - We're doing 2 reads per loop (LSB, MSB), and we expect
463  *    them each to take about a microsecond on real hardware.
464  *    So we expect a count value of around 100. But we'll be
465  *    generous, and accept anything over 50.
466  *
467  *  - if the PIT is stuck, and we see *many* more reads, we
468  *    return early (and the next caller of pit_expect_msb()
469  *    then consider it a failure when they don't see the
470  *    next expected value).
471  *
472  * These expectations mean that we know that we have seen the
473  * transition from one expected value to another with a fairly
474  * high accuracy, and we didn't miss any events. We can thus
475  * use the TSC value at the transitions to calculate a pretty
476  * good value for the TSC frequencty.
477  */
478 static inline int pit_verify_msb(unsigned char val)
479 {
480 	/* Ignore LSB */
481 	inb(0x42);
482 	return inb(0x42) == val;
483 }
484 
485 static inline int pit_expect_msb(unsigned char val, u64 *tscp, unsigned long *deltap)
486 {
487 	int count;
488 	u64 tsc = 0, prev_tsc = 0;
489 
490 	for (count = 0; count < 50000; count++) {
491 		if (!pit_verify_msb(val))
492 			break;
493 		prev_tsc = tsc;
494 		tsc = get_cycles();
495 	}
496 	*deltap = get_cycles() - prev_tsc;
497 	*tscp = tsc;
498 
499 	/*
500 	 * We require _some_ success, but the quality control
501 	 * will be based on the error terms on the TSC values.
502 	 */
503 	return count > 5;
504 }
505 
506 /*
507  * How many MSB values do we want to see? We aim for
508  * a maximum error rate of 500ppm (in practice the
509  * real error is much smaller), but refuse to spend
510  * more than 50ms on it.
511  */
512 #define MAX_QUICK_PIT_MS 50
513 #define MAX_QUICK_PIT_ITERATIONS (MAX_QUICK_PIT_MS * PIT_TICK_RATE / 1000 / 256)
514 
515 static unsigned long quick_pit_calibrate(void)
516 {
517 	int i;
518 	u64 tsc, delta;
519 	unsigned long d1, d2;
520 
521 	if (!has_legacy_pic())
522 		return 0;
523 
524 	/* Set the Gate high, disable speaker */
525 	outb((inb(0x61) & ~0x02) | 0x01, 0x61);
526 
527 	/*
528 	 * Counter 2, mode 0 (one-shot), binary count
529 	 *
530 	 * NOTE! Mode 2 decrements by two (and then the
531 	 * output is flipped each time, giving the same
532 	 * final output frequency as a decrement-by-one),
533 	 * so mode 0 is much better when looking at the
534 	 * individual counts.
535 	 */
536 	outb(0xb0, 0x43);
537 
538 	/* Start at 0xffff */
539 	outb(0xff, 0x42);
540 	outb(0xff, 0x42);
541 
542 	/*
543 	 * The PIT starts counting at the next edge, so we
544 	 * need to delay for a microsecond. The easiest way
545 	 * to do that is to just read back the 16-bit counter
546 	 * once from the PIT.
547 	 */
548 	pit_verify_msb(0);
549 
550 	if (pit_expect_msb(0xff, &tsc, &d1)) {
551 		for (i = 1; i <= MAX_QUICK_PIT_ITERATIONS; i++) {
552 			if (!pit_expect_msb(0xff-i, &delta, &d2))
553 				break;
554 
555 			delta -= tsc;
556 
557 			/*
558 			 * Extrapolate the error and fail fast if the error will
559 			 * never be below 500 ppm.
560 			 */
561 			if (i == 1 &&
562 			    d1 + d2 >= (delta * MAX_QUICK_PIT_ITERATIONS) >> 11)
563 				return 0;
564 
565 			/*
566 			 * Iterate until the error is less than 500 ppm
567 			 */
568 			if (d1+d2 >= delta >> 11)
569 				continue;
570 
571 			/*
572 			 * Check the PIT one more time to verify that
573 			 * all TSC reads were stable wrt the PIT.
574 			 *
575 			 * This also guarantees serialization of the
576 			 * last cycle read ('d2') in pit_expect_msb.
577 			 */
578 			if (!pit_verify_msb(0xfe - i))
579 				break;
580 			goto success;
581 		}
582 	}
583 	pr_info("Fast TSC calibration failed\n");
584 	return 0;
585 
586 success:
587 	/*
588 	 * Ok, if we get here, then we've seen the
589 	 * MSB of the PIT decrement 'i' times, and the
590 	 * error has shrunk to less than 500 ppm.
591 	 *
592 	 * As a result, we can depend on there not being
593 	 * any odd delays anywhere, and the TSC reads are
594 	 * reliable (within the error).
595 	 *
596 	 * kHz = ticks / time-in-seconds / 1000;
597 	 * kHz = (t2 - t1) / (I * 256 / PIT_TICK_RATE) / 1000
598 	 * kHz = ((t2 - t1) * PIT_TICK_RATE) / (I * 256 * 1000)
599 	 */
600 	delta *= PIT_TICK_RATE;
601 	do_div(delta, i*256*1000);
602 	pr_info("Fast TSC calibration using PIT\n");
603 	return delta;
604 }
605 
606 /**
607  * native_calibrate_tsc
608  * Determine TSC frequency via CPUID, else return 0.
609  */
610 unsigned long native_calibrate_tsc(void)
611 {
612 	unsigned int eax_denominator, ebx_numerator, ecx_hz, edx;
613 	unsigned int crystal_khz;
614 
615 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
616 		return 0;
617 
618 	if (boot_cpu_data.cpuid_level < 0x15)
619 		return 0;
620 
621 	eax_denominator = ebx_numerator = ecx_hz = edx = 0;
622 
623 	/* CPUID 15H TSC/Crystal ratio, plus optionally Crystal Hz */
624 	cpuid(0x15, &eax_denominator, &ebx_numerator, &ecx_hz, &edx);
625 
626 	if (ebx_numerator == 0 || eax_denominator == 0)
627 		return 0;
628 
629 	crystal_khz = ecx_hz / 1000;
630 
631 	if (crystal_khz == 0) {
632 		switch (boot_cpu_data.x86_model) {
633 		case INTEL_FAM6_SKYLAKE_MOBILE:
634 		case INTEL_FAM6_SKYLAKE_DESKTOP:
635 		case INTEL_FAM6_KABYLAKE_MOBILE:
636 		case INTEL_FAM6_KABYLAKE_DESKTOP:
637 			crystal_khz = 24000;	/* 24.0 MHz */
638 			break;
639 		case INTEL_FAM6_ATOM_DENVERTON:
640 			crystal_khz = 25000;	/* 25.0 MHz */
641 			break;
642 		case INTEL_FAM6_ATOM_GOLDMONT:
643 			crystal_khz = 19200;	/* 19.2 MHz */
644 			break;
645 		}
646 	}
647 
648 	if (crystal_khz == 0)
649 		return 0;
650 	/*
651 	 * TSC frequency determined by CPUID is a "hardware reported"
652 	 * frequency and is the most accurate one so far we have. This
653 	 * is considered a known frequency.
654 	 */
655 	setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ);
656 
657 	/*
658 	 * For Atom SoCs TSC is the only reliable clocksource.
659 	 * Mark TSC reliable so no watchdog on it.
660 	 */
661 	if (boot_cpu_data.x86_model == INTEL_FAM6_ATOM_GOLDMONT)
662 		setup_force_cpu_cap(X86_FEATURE_TSC_RELIABLE);
663 
664 	return crystal_khz * ebx_numerator / eax_denominator;
665 }
666 
667 static unsigned long cpu_khz_from_cpuid(void)
668 {
669 	unsigned int eax_base_mhz, ebx_max_mhz, ecx_bus_mhz, edx;
670 
671 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
672 		return 0;
673 
674 	if (boot_cpu_data.cpuid_level < 0x16)
675 		return 0;
676 
677 	eax_base_mhz = ebx_max_mhz = ecx_bus_mhz = edx = 0;
678 
679 	cpuid(0x16, &eax_base_mhz, &ebx_max_mhz, &ecx_bus_mhz, &edx);
680 
681 	return eax_base_mhz * 1000;
682 }
683 
684 /*
685  * calibrate cpu using pit, hpet, and ptimer methods. They are available
686  * later in boot after acpi is initialized.
687  */
688 static unsigned long pit_hpet_ptimer_calibrate_cpu(void)
689 {
690 	u64 tsc1, tsc2, delta, ref1, ref2;
691 	unsigned long tsc_pit_min = ULONG_MAX, tsc_ref_min = ULONG_MAX;
692 	unsigned long flags, latch, ms;
693 	int hpet = is_hpet_enabled(), i, loopmin;
694 
695 	/*
696 	 * Run 5 calibration loops to get the lowest frequency value
697 	 * (the best estimate). We use two different calibration modes
698 	 * here:
699 	 *
700 	 * 1) PIT loop. We set the PIT Channel 2 to oneshot mode and
701 	 * load a timeout of 50ms. We read the time right after we
702 	 * started the timer and wait until the PIT count down reaches
703 	 * zero. In each wait loop iteration we read the TSC and check
704 	 * the delta to the previous read. We keep track of the min
705 	 * and max values of that delta. The delta is mostly defined
706 	 * by the IO time of the PIT access, so we can detect when a
707 	 * SMI/SMM disturbance happened between the two reads. If the
708 	 * maximum time is significantly larger than the minimum time,
709 	 * then we discard the result and have another try.
710 	 *
711 	 * 2) Reference counter. If available we use the HPET or the
712 	 * PMTIMER as a reference to check the sanity of that value.
713 	 * We use separate TSC readouts and check inside of the
714 	 * reference read for a SMI/SMM disturbance. We dicard
715 	 * disturbed values here as well. We do that around the PIT
716 	 * calibration delay loop as we have to wait for a certain
717 	 * amount of time anyway.
718 	 */
719 
720 	/* Preset PIT loop values */
721 	latch = CAL_LATCH;
722 	ms = CAL_MS;
723 	loopmin = CAL_PIT_LOOPS;
724 
725 	for (i = 0; i < 3; i++) {
726 		unsigned long tsc_pit_khz;
727 
728 		/*
729 		 * Read the start value and the reference count of
730 		 * hpet/pmtimer when available. Then do the PIT
731 		 * calibration, which will take at least 50ms, and
732 		 * read the end value.
733 		 */
734 		local_irq_save(flags);
735 		tsc1 = tsc_read_refs(&ref1, hpet);
736 		tsc_pit_khz = pit_calibrate_tsc(latch, ms, loopmin);
737 		tsc2 = tsc_read_refs(&ref2, hpet);
738 		local_irq_restore(flags);
739 
740 		/* Pick the lowest PIT TSC calibration so far */
741 		tsc_pit_min = min(tsc_pit_min, tsc_pit_khz);
742 
743 		/* hpet or pmtimer available ? */
744 		if (ref1 == ref2)
745 			continue;
746 
747 		/* Check, whether the sampling was disturbed by an SMI */
748 		if (tsc1 == ULLONG_MAX || tsc2 == ULLONG_MAX)
749 			continue;
750 
751 		tsc2 = (tsc2 - tsc1) * 1000000LL;
752 		if (hpet)
753 			tsc2 = calc_hpet_ref(tsc2, ref1, ref2);
754 		else
755 			tsc2 = calc_pmtimer_ref(tsc2, ref1, ref2);
756 
757 		tsc_ref_min = min(tsc_ref_min, (unsigned long) tsc2);
758 
759 		/* Check the reference deviation */
760 		delta = ((u64) tsc_pit_min) * 100;
761 		do_div(delta, tsc_ref_min);
762 
763 		/*
764 		 * If both calibration results are inside a 10% window
765 		 * then we can be sure, that the calibration
766 		 * succeeded. We break out of the loop right away. We
767 		 * use the reference value, as it is more precise.
768 		 */
769 		if (delta >= 90 && delta <= 110) {
770 			pr_info("PIT calibration matches %s. %d loops\n",
771 				hpet ? "HPET" : "PMTIMER", i + 1);
772 			return tsc_ref_min;
773 		}
774 
775 		/*
776 		 * Check whether PIT failed more than once. This
777 		 * happens in virtualized environments. We need to
778 		 * give the virtual PC a slightly longer timeframe for
779 		 * the HPET/PMTIMER to make the result precise.
780 		 */
781 		if (i == 1 && tsc_pit_min == ULONG_MAX) {
782 			latch = CAL2_LATCH;
783 			ms = CAL2_MS;
784 			loopmin = CAL2_PIT_LOOPS;
785 		}
786 	}
787 
788 	/*
789 	 * Now check the results.
790 	 */
791 	if (tsc_pit_min == ULONG_MAX) {
792 		/* PIT gave no useful value */
793 		pr_warn("Unable to calibrate against PIT\n");
794 
795 		/* We don't have an alternative source, disable TSC */
796 		if (!hpet && !ref1 && !ref2) {
797 			pr_notice("No reference (HPET/PMTIMER) available\n");
798 			return 0;
799 		}
800 
801 		/* The alternative source failed as well, disable TSC */
802 		if (tsc_ref_min == ULONG_MAX) {
803 			pr_warn("HPET/PMTIMER calibration failed\n");
804 			return 0;
805 		}
806 
807 		/* Use the alternative source */
808 		pr_info("using %s reference calibration\n",
809 			hpet ? "HPET" : "PMTIMER");
810 
811 		return tsc_ref_min;
812 	}
813 
814 	/* We don't have an alternative source, use the PIT calibration value */
815 	if (!hpet && !ref1 && !ref2) {
816 		pr_info("Using PIT calibration value\n");
817 		return tsc_pit_min;
818 	}
819 
820 	/* The alternative source failed, use the PIT calibration value */
821 	if (tsc_ref_min == ULONG_MAX) {
822 		pr_warn("HPET/PMTIMER calibration failed. Using PIT calibration.\n");
823 		return tsc_pit_min;
824 	}
825 
826 	/*
827 	 * The calibration values differ too much. In doubt, we use
828 	 * the PIT value as we know that there are PMTIMERs around
829 	 * running at double speed. At least we let the user know:
830 	 */
831 	pr_warn("PIT calibration deviates from %s: %lu %lu\n",
832 		hpet ? "HPET" : "PMTIMER", tsc_pit_min, tsc_ref_min);
833 	pr_info("Using PIT calibration value\n");
834 	return tsc_pit_min;
835 }
836 
837 /**
838  * native_calibrate_cpu_early - can calibrate the cpu early in boot
839  */
840 unsigned long native_calibrate_cpu_early(void)
841 {
842 	unsigned long flags, fast_calibrate = cpu_khz_from_cpuid();
843 
844 	if (!fast_calibrate)
845 		fast_calibrate = cpu_khz_from_msr();
846 	if (!fast_calibrate) {
847 		local_irq_save(flags);
848 		fast_calibrate = quick_pit_calibrate();
849 		local_irq_restore(flags);
850 	}
851 	return fast_calibrate;
852 }
853 
854 
855 /**
856  * native_calibrate_cpu - calibrate the cpu
857  */
858 static unsigned long native_calibrate_cpu(void)
859 {
860 	unsigned long tsc_freq = native_calibrate_cpu_early();
861 
862 	if (!tsc_freq)
863 		tsc_freq = pit_hpet_ptimer_calibrate_cpu();
864 
865 	return tsc_freq;
866 }
867 
868 void recalibrate_cpu_khz(void)
869 {
870 #ifndef CONFIG_SMP
871 	unsigned long cpu_khz_old = cpu_khz;
872 
873 	if (!boot_cpu_has(X86_FEATURE_TSC))
874 		return;
875 
876 	cpu_khz = x86_platform.calibrate_cpu();
877 	tsc_khz = x86_platform.calibrate_tsc();
878 	if (tsc_khz == 0)
879 		tsc_khz = cpu_khz;
880 	else if (abs(cpu_khz - tsc_khz) * 10 > tsc_khz)
881 		cpu_khz = tsc_khz;
882 	cpu_data(0).loops_per_jiffy = cpufreq_scale(cpu_data(0).loops_per_jiffy,
883 						    cpu_khz_old, cpu_khz);
884 #endif
885 }
886 
887 EXPORT_SYMBOL(recalibrate_cpu_khz);
888 
889 
890 static unsigned long long cyc2ns_suspend;
891 
892 void tsc_save_sched_clock_state(void)
893 {
894 	if (!sched_clock_stable())
895 		return;
896 
897 	cyc2ns_suspend = sched_clock();
898 }
899 
900 /*
901  * Even on processors with invariant TSC, TSC gets reset in some the
902  * ACPI system sleep states. And in some systems BIOS seem to reinit TSC to
903  * arbitrary value (still sync'd across cpu's) during resume from such sleep
904  * states. To cope up with this, recompute the cyc2ns_offset for each cpu so
905  * that sched_clock() continues from the point where it was left off during
906  * suspend.
907  */
908 void tsc_restore_sched_clock_state(void)
909 {
910 	unsigned long long offset;
911 	unsigned long flags;
912 	int cpu;
913 
914 	if (!sched_clock_stable())
915 		return;
916 
917 	local_irq_save(flags);
918 
919 	/*
920 	 * We're coming out of suspend, there's no concurrency yet; don't
921 	 * bother being nice about the RCU stuff, just write to both
922 	 * data fields.
923 	 */
924 
925 	this_cpu_write(cyc2ns.data[0].cyc2ns_offset, 0);
926 	this_cpu_write(cyc2ns.data[1].cyc2ns_offset, 0);
927 
928 	offset = cyc2ns_suspend - sched_clock();
929 
930 	for_each_possible_cpu(cpu) {
931 		per_cpu(cyc2ns.data[0].cyc2ns_offset, cpu) = offset;
932 		per_cpu(cyc2ns.data[1].cyc2ns_offset, cpu) = offset;
933 	}
934 
935 	local_irq_restore(flags);
936 }
937 
938 #ifdef CONFIG_CPU_FREQ
939 /* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
940  * changes.
941  *
942  * RED-PEN: On SMP we assume all CPUs run with the same frequency.  It's
943  * not that important because current Opteron setups do not support
944  * scaling on SMP anyroads.
945  *
946  * Should fix up last_tsc too. Currently gettimeofday in the
947  * first tick after the change will be slightly wrong.
948  */
949 
950 static unsigned int  ref_freq;
951 static unsigned long loops_per_jiffy_ref;
952 static unsigned long tsc_khz_ref;
953 
954 static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
955 				void *data)
956 {
957 	struct cpufreq_freqs *freq = data;
958 	unsigned long *lpj;
959 
960 	lpj = &boot_cpu_data.loops_per_jiffy;
961 #ifdef CONFIG_SMP
962 	if (!(freq->flags & CPUFREQ_CONST_LOOPS))
963 		lpj = &cpu_data(freq->cpu).loops_per_jiffy;
964 #endif
965 
966 	if (!ref_freq) {
967 		ref_freq = freq->old;
968 		loops_per_jiffy_ref = *lpj;
969 		tsc_khz_ref = tsc_khz;
970 	}
971 	if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
972 			(val == CPUFREQ_POSTCHANGE && freq->old > freq->new)) {
973 		*lpj = cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
974 
975 		tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
976 		if (!(freq->flags & CPUFREQ_CONST_LOOPS))
977 			mark_tsc_unstable("cpufreq changes");
978 
979 		set_cyc2ns_scale(tsc_khz, freq->cpu, rdtsc());
980 	}
981 
982 	return 0;
983 }
984 
985 static struct notifier_block time_cpufreq_notifier_block = {
986 	.notifier_call  = time_cpufreq_notifier
987 };
988 
989 static int __init cpufreq_register_tsc_scaling(void)
990 {
991 	if (!boot_cpu_has(X86_FEATURE_TSC))
992 		return 0;
993 	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
994 		return 0;
995 	cpufreq_register_notifier(&time_cpufreq_notifier_block,
996 				CPUFREQ_TRANSITION_NOTIFIER);
997 	return 0;
998 }
999 
1000 core_initcall(cpufreq_register_tsc_scaling);
1001 
1002 #endif /* CONFIG_CPU_FREQ */
1003 
1004 #define ART_CPUID_LEAF (0x15)
1005 #define ART_MIN_DENOMINATOR (1)
1006 
1007 
1008 /*
1009  * If ART is present detect the numerator:denominator to convert to TSC
1010  */
1011 static void __init detect_art(void)
1012 {
1013 	unsigned int unused[2];
1014 
1015 	if (boot_cpu_data.cpuid_level < ART_CPUID_LEAF)
1016 		return;
1017 
1018 	/*
1019 	 * Don't enable ART in a VM, non-stop TSC and TSC_ADJUST required,
1020 	 * and the TSC counter resets must not occur asynchronously.
1021 	 */
1022 	if (boot_cpu_has(X86_FEATURE_HYPERVISOR) ||
1023 	    !boot_cpu_has(X86_FEATURE_NONSTOP_TSC) ||
1024 	    !boot_cpu_has(X86_FEATURE_TSC_ADJUST) ||
1025 	    tsc_async_resets)
1026 		return;
1027 
1028 	cpuid(ART_CPUID_LEAF, &art_to_tsc_denominator,
1029 	      &art_to_tsc_numerator, unused, unused+1);
1030 
1031 	if (art_to_tsc_denominator < ART_MIN_DENOMINATOR)
1032 		return;
1033 
1034 	rdmsrl(MSR_IA32_TSC_ADJUST, art_to_tsc_offset);
1035 
1036 	/* Make this sticky over multiple CPU init calls */
1037 	setup_force_cpu_cap(X86_FEATURE_ART);
1038 }
1039 
1040 
1041 /* clocksource code */
1042 
1043 static void tsc_resume(struct clocksource *cs)
1044 {
1045 	tsc_verify_tsc_adjust(true);
1046 }
1047 
1048 /*
1049  * We used to compare the TSC to the cycle_last value in the clocksource
1050  * structure to avoid a nasty time-warp. This can be observed in a
1051  * very small window right after one CPU updated cycle_last under
1052  * xtime/vsyscall_gtod lock and the other CPU reads a TSC value which
1053  * is smaller than the cycle_last reference value due to a TSC which
1054  * is slighty behind. This delta is nowhere else observable, but in
1055  * that case it results in a forward time jump in the range of hours
1056  * due to the unsigned delta calculation of the time keeping core
1057  * code, which is necessary to support wrapping clocksources like pm
1058  * timer.
1059  *
1060  * This sanity check is now done in the core timekeeping code.
1061  * checking the result of read_tsc() - cycle_last for being negative.
1062  * That works because CLOCKSOURCE_MASK(64) does not mask out any bit.
1063  */
1064 static u64 read_tsc(struct clocksource *cs)
1065 {
1066 	return (u64)rdtsc_ordered();
1067 }
1068 
1069 static void tsc_cs_mark_unstable(struct clocksource *cs)
1070 {
1071 	if (tsc_unstable)
1072 		return;
1073 
1074 	tsc_unstable = 1;
1075 	if (using_native_sched_clock())
1076 		clear_sched_clock_stable();
1077 	disable_sched_clock_irqtime();
1078 	pr_info("Marking TSC unstable due to clocksource watchdog\n");
1079 }
1080 
1081 static void tsc_cs_tick_stable(struct clocksource *cs)
1082 {
1083 	if (tsc_unstable)
1084 		return;
1085 
1086 	if (using_native_sched_clock())
1087 		sched_clock_tick_stable();
1088 }
1089 
1090 /*
1091  * .mask MUST be CLOCKSOURCE_MASK(64). See comment above read_tsc()
1092  */
1093 static struct clocksource clocksource_tsc_early = {
1094 	.name                   = "tsc-early",
1095 	.rating                 = 299,
1096 	.read                   = read_tsc,
1097 	.mask                   = CLOCKSOURCE_MASK(64),
1098 	.flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
1099 				  CLOCK_SOURCE_MUST_VERIFY,
1100 	.archdata               = { .vclock_mode = VCLOCK_TSC },
1101 	.resume			= tsc_resume,
1102 	.mark_unstable		= tsc_cs_mark_unstable,
1103 	.tick_stable		= tsc_cs_tick_stable,
1104 	.list			= LIST_HEAD_INIT(clocksource_tsc_early.list),
1105 };
1106 
1107 /*
1108  * Must mark VALID_FOR_HRES early such that when we unregister tsc_early
1109  * this one will immediately take over. We will only register if TSC has
1110  * been found good.
1111  */
1112 static struct clocksource clocksource_tsc = {
1113 	.name                   = "tsc",
1114 	.rating                 = 300,
1115 	.read                   = read_tsc,
1116 	.mask                   = CLOCKSOURCE_MASK(64),
1117 	.flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
1118 				  CLOCK_SOURCE_VALID_FOR_HRES |
1119 				  CLOCK_SOURCE_MUST_VERIFY,
1120 	.archdata               = { .vclock_mode = VCLOCK_TSC },
1121 	.resume			= tsc_resume,
1122 	.mark_unstable		= tsc_cs_mark_unstable,
1123 	.tick_stable		= tsc_cs_tick_stable,
1124 	.list			= LIST_HEAD_INIT(clocksource_tsc.list),
1125 };
1126 
1127 void mark_tsc_unstable(char *reason)
1128 {
1129 	if (tsc_unstable)
1130 		return;
1131 
1132 	tsc_unstable = 1;
1133 	if (using_native_sched_clock())
1134 		clear_sched_clock_stable();
1135 	disable_sched_clock_irqtime();
1136 	pr_info("Marking TSC unstable due to %s\n", reason);
1137 
1138 	clocksource_mark_unstable(&clocksource_tsc_early);
1139 	clocksource_mark_unstable(&clocksource_tsc);
1140 }
1141 
1142 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
1143 
1144 static void __init check_system_tsc_reliable(void)
1145 {
1146 #if defined(CONFIG_MGEODEGX1) || defined(CONFIG_MGEODE_LX) || defined(CONFIG_X86_GENERIC)
1147 	if (is_geode_lx()) {
1148 		/* RTSC counts during suspend */
1149 #define RTSC_SUSP 0x100
1150 		unsigned long res_low, res_high;
1151 
1152 		rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
1153 		/* Geode_LX - the OLPC CPU has a very reliable TSC */
1154 		if (res_low & RTSC_SUSP)
1155 			tsc_clocksource_reliable = 1;
1156 	}
1157 #endif
1158 	if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE))
1159 		tsc_clocksource_reliable = 1;
1160 }
1161 
1162 /*
1163  * Make an educated guess if the TSC is trustworthy and synchronized
1164  * over all CPUs.
1165  */
1166 int unsynchronized_tsc(void)
1167 {
1168 	if (!boot_cpu_has(X86_FEATURE_TSC) || tsc_unstable)
1169 		return 1;
1170 
1171 #ifdef CONFIG_SMP
1172 	if (apic_is_clustered_box())
1173 		return 1;
1174 #endif
1175 
1176 	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
1177 		return 0;
1178 
1179 	if (tsc_clocksource_reliable)
1180 		return 0;
1181 	/*
1182 	 * Intel systems are normally all synchronized.
1183 	 * Exceptions must mark TSC as unstable:
1184 	 */
1185 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
1186 		/* assume multi socket systems are not synchronized: */
1187 		if (num_possible_cpus() > 1)
1188 			return 1;
1189 	}
1190 
1191 	return 0;
1192 }
1193 
1194 /*
1195  * Convert ART to TSC given numerator/denominator found in detect_art()
1196  */
1197 struct system_counterval_t convert_art_to_tsc(u64 art)
1198 {
1199 	u64 tmp, res, rem;
1200 
1201 	rem = do_div(art, art_to_tsc_denominator);
1202 
1203 	res = art * art_to_tsc_numerator;
1204 	tmp = rem * art_to_tsc_numerator;
1205 
1206 	do_div(tmp, art_to_tsc_denominator);
1207 	res += tmp + art_to_tsc_offset;
1208 
1209 	return (struct system_counterval_t) {.cs = art_related_clocksource,
1210 			.cycles = res};
1211 }
1212 EXPORT_SYMBOL(convert_art_to_tsc);
1213 
1214 /**
1215  * convert_art_ns_to_tsc() - Convert ART in nanoseconds to TSC.
1216  * @art_ns: ART (Always Running Timer) in unit of nanoseconds
1217  *
1218  * PTM requires all timestamps to be in units of nanoseconds. When user
1219  * software requests a cross-timestamp, this function converts system timestamp
1220  * to TSC.
1221  *
1222  * This is valid when CPU feature flag X86_FEATURE_TSC_KNOWN_FREQ is set
1223  * indicating the tsc_khz is derived from CPUID[15H]. Drivers should check
1224  * that this flag is set before conversion to TSC is attempted.
1225  *
1226  * Return:
1227  * struct system_counterval_t - system counter value with the pointer to the
1228  *	corresponding clocksource
1229  *	@cycles:	System counter value
1230  *	@cs:		Clocksource corresponding to system counter value. Used
1231  *			by timekeeping code to verify comparibility of two cycle
1232  *			values.
1233  */
1234 
1235 struct system_counterval_t convert_art_ns_to_tsc(u64 art_ns)
1236 {
1237 	u64 tmp, res, rem;
1238 
1239 	rem = do_div(art_ns, USEC_PER_SEC);
1240 
1241 	res = art_ns * tsc_khz;
1242 	tmp = rem * tsc_khz;
1243 
1244 	do_div(tmp, USEC_PER_SEC);
1245 	res += tmp;
1246 
1247 	return (struct system_counterval_t) { .cs = art_related_clocksource,
1248 					      .cycles = res};
1249 }
1250 EXPORT_SYMBOL(convert_art_ns_to_tsc);
1251 
1252 
1253 static void tsc_refine_calibration_work(struct work_struct *work);
1254 static DECLARE_DELAYED_WORK(tsc_irqwork, tsc_refine_calibration_work);
1255 /**
1256  * tsc_refine_calibration_work - Further refine tsc freq calibration
1257  * @work - ignored.
1258  *
1259  * This functions uses delayed work over a period of a
1260  * second to further refine the TSC freq value. Since this is
1261  * timer based, instead of loop based, we don't block the boot
1262  * process while this longer calibration is done.
1263  *
1264  * If there are any calibration anomalies (too many SMIs, etc),
1265  * or the refined calibration is off by 1% of the fast early
1266  * calibration, we throw out the new calibration and use the
1267  * early calibration.
1268  */
1269 static void tsc_refine_calibration_work(struct work_struct *work)
1270 {
1271 	static u64 tsc_start = -1, ref_start;
1272 	static int hpet;
1273 	u64 tsc_stop, ref_stop, delta;
1274 	unsigned long freq;
1275 	int cpu;
1276 
1277 	/* Don't bother refining TSC on unstable systems */
1278 	if (tsc_unstable)
1279 		goto unreg;
1280 
1281 	/*
1282 	 * Since the work is started early in boot, we may be
1283 	 * delayed the first time we expire. So set the workqueue
1284 	 * again once we know timers are working.
1285 	 */
1286 	if (tsc_start == -1) {
1287 		/*
1288 		 * Only set hpet once, to avoid mixing hardware
1289 		 * if the hpet becomes enabled later.
1290 		 */
1291 		hpet = is_hpet_enabled();
1292 		schedule_delayed_work(&tsc_irqwork, HZ);
1293 		tsc_start = tsc_read_refs(&ref_start, hpet);
1294 		return;
1295 	}
1296 
1297 	tsc_stop = tsc_read_refs(&ref_stop, hpet);
1298 
1299 	/* hpet or pmtimer available ? */
1300 	if (ref_start == ref_stop)
1301 		goto out;
1302 
1303 	/* Check, whether the sampling was disturbed by an SMI */
1304 	if (tsc_start == ULLONG_MAX || tsc_stop == ULLONG_MAX)
1305 		goto out;
1306 
1307 	delta = tsc_stop - tsc_start;
1308 	delta *= 1000000LL;
1309 	if (hpet)
1310 		freq = calc_hpet_ref(delta, ref_start, ref_stop);
1311 	else
1312 		freq = calc_pmtimer_ref(delta, ref_start, ref_stop);
1313 
1314 	/* Make sure we're within 1% */
1315 	if (abs(tsc_khz - freq) > tsc_khz/100)
1316 		goto out;
1317 
1318 	tsc_khz = freq;
1319 	pr_info("Refined TSC clocksource calibration: %lu.%03lu MHz\n",
1320 		(unsigned long)tsc_khz / 1000,
1321 		(unsigned long)tsc_khz % 1000);
1322 
1323 	/* Inform the TSC deadline clockevent devices about the recalibration */
1324 	lapic_update_tsc_freq();
1325 
1326 	/* Update the sched_clock() rate to match the clocksource one */
1327 	for_each_possible_cpu(cpu)
1328 		set_cyc2ns_scale(tsc_khz, cpu, tsc_stop);
1329 
1330 out:
1331 	if (tsc_unstable)
1332 		goto unreg;
1333 
1334 	if (boot_cpu_has(X86_FEATURE_ART))
1335 		art_related_clocksource = &clocksource_tsc;
1336 	clocksource_register_khz(&clocksource_tsc, tsc_khz);
1337 unreg:
1338 	clocksource_unregister(&clocksource_tsc_early);
1339 }
1340 
1341 
1342 static int __init init_tsc_clocksource(void)
1343 {
1344 	if (!boot_cpu_has(X86_FEATURE_TSC) || !tsc_khz)
1345 		return 0;
1346 
1347 	if (tsc_unstable)
1348 		goto unreg;
1349 
1350 	if (tsc_clocksource_reliable)
1351 		clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
1352 
1353 	if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC_S3))
1354 		clocksource_tsc.flags |= CLOCK_SOURCE_SUSPEND_NONSTOP;
1355 
1356 	/*
1357 	 * When TSC frequency is known (retrieved via MSR or CPUID), we skip
1358 	 * the refined calibration and directly register it as a clocksource.
1359 	 */
1360 	if (boot_cpu_has(X86_FEATURE_TSC_KNOWN_FREQ)) {
1361 		if (boot_cpu_has(X86_FEATURE_ART))
1362 			art_related_clocksource = &clocksource_tsc;
1363 		clocksource_register_khz(&clocksource_tsc, tsc_khz);
1364 unreg:
1365 		clocksource_unregister(&clocksource_tsc_early);
1366 		return 0;
1367 	}
1368 
1369 	schedule_delayed_work(&tsc_irqwork, 0);
1370 	return 0;
1371 }
1372 /*
1373  * We use device_initcall here, to ensure we run after the hpet
1374  * is fully initialized, which may occur at fs_initcall time.
1375  */
1376 device_initcall(init_tsc_clocksource);
1377 
1378 static bool __init determine_cpu_tsc_frequencies(bool early)
1379 {
1380 	/* Make sure that cpu and tsc are not already calibrated */
1381 	WARN_ON(cpu_khz || tsc_khz);
1382 
1383 	if (early) {
1384 		cpu_khz = x86_platform.calibrate_cpu();
1385 		tsc_khz = x86_platform.calibrate_tsc();
1386 	} else {
1387 		/* We should not be here with non-native cpu calibration */
1388 		WARN_ON(x86_platform.calibrate_cpu != native_calibrate_cpu);
1389 		cpu_khz = pit_hpet_ptimer_calibrate_cpu();
1390 	}
1391 
1392 	/*
1393 	 * Trust non-zero tsc_khz as authoritative,
1394 	 * and use it to sanity check cpu_khz,
1395 	 * which will be off if system timer is off.
1396 	 */
1397 	if (tsc_khz == 0)
1398 		tsc_khz = cpu_khz;
1399 	else if (abs(cpu_khz - tsc_khz) * 10 > tsc_khz)
1400 		cpu_khz = tsc_khz;
1401 
1402 	if (tsc_khz == 0)
1403 		return false;
1404 
1405 	pr_info("Detected %lu.%03lu MHz processor\n",
1406 		(unsigned long)cpu_khz / KHZ,
1407 		(unsigned long)cpu_khz % KHZ);
1408 
1409 	if (cpu_khz != tsc_khz) {
1410 		pr_info("Detected %lu.%03lu MHz TSC",
1411 			(unsigned long)tsc_khz / KHZ,
1412 			(unsigned long)tsc_khz % KHZ);
1413 	}
1414 	return true;
1415 }
1416 
1417 static unsigned long __init get_loops_per_jiffy(void)
1418 {
1419 	u64 lpj = (u64)tsc_khz * KHZ;
1420 
1421 	do_div(lpj, HZ);
1422 	return lpj;
1423 }
1424 
1425 static void __init tsc_enable_sched_clock(void)
1426 {
1427 	/* Sanitize TSC ADJUST before cyc2ns gets initialized */
1428 	tsc_store_and_check_tsc_adjust(true);
1429 	cyc2ns_init_boot_cpu();
1430 	static_branch_enable(&__use_tsc);
1431 }
1432 
1433 void __init tsc_early_init(void)
1434 {
1435 	if (!boot_cpu_has(X86_FEATURE_TSC))
1436 		return;
1437 	/* Don't change UV TSC multi-chassis synchronization */
1438 	if (is_early_uv_system())
1439 		return;
1440 	if (!determine_cpu_tsc_frequencies(true))
1441 		return;
1442 	loops_per_jiffy = get_loops_per_jiffy();
1443 
1444 	tsc_enable_sched_clock();
1445 }
1446 
1447 void __init tsc_init(void)
1448 {
1449 	/*
1450 	 * native_calibrate_cpu_early can only calibrate using methods that are
1451 	 * available early in boot.
1452 	 */
1453 	if (x86_platform.calibrate_cpu == native_calibrate_cpu_early)
1454 		x86_platform.calibrate_cpu = native_calibrate_cpu;
1455 
1456 	if (!boot_cpu_has(X86_FEATURE_TSC)) {
1457 		setup_clear_cpu_cap(X86_FEATURE_TSC_DEADLINE_TIMER);
1458 		return;
1459 	}
1460 
1461 	if (!tsc_khz) {
1462 		/* We failed to determine frequencies earlier, try again */
1463 		if (!determine_cpu_tsc_frequencies(false)) {
1464 			mark_tsc_unstable("could not calculate TSC khz");
1465 			setup_clear_cpu_cap(X86_FEATURE_TSC_DEADLINE_TIMER);
1466 			return;
1467 		}
1468 		tsc_enable_sched_clock();
1469 	}
1470 
1471 	cyc2ns_init_secondary_cpus();
1472 
1473 	if (!no_sched_irq_time)
1474 		enable_sched_clock_irqtime();
1475 
1476 	lpj_fine = get_loops_per_jiffy();
1477 	use_tsc_delay();
1478 
1479 	check_system_tsc_reliable();
1480 
1481 	if (unsynchronized_tsc()) {
1482 		mark_tsc_unstable("TSCs unsynchronized");
1483 		return;
1484 	}
1485 
1486 	clocksource_register_khz(&clocksource_tsc_early, tsc_khz);
1487 	detect_art();
1488 }
1489 
1490 #ifdef CONFIG_SMP
1491 /*
1492  * If we have a constant TSC and are using the TSC for the delay loop,
1493  * we can skip clock calibration if another cpu in the same socket has already
1494  * been calibrated. This assumes that CONSTANT_TSC applies to all
1495  * cpus in the socket - this should be a safe assumption.
1496  */
1497 unsigned long calibrate_delay_is_known(void)
1498 {
1499 	int sibling, cpu = smp_processor_id();
1500 	int constant_tsc = cpu_has(&cpu_data(cpu), X86_FEATURE_CONSTANT_TSC);
1501 	const struct cpumask *mask = topology_core_cpumask(cpu);
1502 
1503 	if (!constant_tsc || !mask)
1504 		return 0;
1505 
1506 	sibling = cpumask_any_but(mask, cpu);
1507 	if (sibling < nr_cpu_ids)
1508 		return cpu_data(sibling).loops_per_jiffy;
1509 	return 0;
1510 }
1511 #endif
1512