1 // SPDX-License-Identifier: GPL-2.0-only
2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
3 
4 #include <linux/kernel.h>
5 #include <linux/sched.h>
6 #include <linux/sched/clock.h>
7 #include <linux/init.h>
8 #include <linux/export.h>
9 #include <linux/timer.h>
10 #include <linux/acpi_pmtmr.h>
11 #include <linux/cpufreq.h>
12 #include <linux/delay.h>
13 #include <linux/clocksource.h>
14 #include <linux/percpu.h>
15 #include <linux/timex.h>
16 #include <linux/static_key.h>
17 #include <linux/static_call.h>
18 
19 #include <asm/hpet.h>
20 #include <asm/timer.h>
21 #include <asm/vgtod.h>
22 #include <asm/time.h>
23 #include <asm/delay.h>
24 #include <asm/hypervisor.h>
25 #include <asm/nmi.h>
26 #include <asm/x86_init.h>
27 #include <asm/geode.h>
28 #include <asm/apic.h>
29 #include <asm/intel-family.h>
30 #include <asm/i8259.h>
31 #include <asm/uv/uv.h>
32 
33 unsigned int __read_mostly cpu_khz;	/* TSC clocks / usec, not used here */
34 EXPORT_SYMBOL(cpu_khz);
35 
36 unsigned int __read_mostly tsc_khz;
37 EXPORT_SYMBOL(tsc_khz);
38 
39 #define KHZ	1000
40 
41 /*
42  * TSC can be unstable due to cpufreq or due to unsynced TSCs
43  */
44 static int __read_mostly tsc_unstable;
45 static unsigned int __initdata tsc_early_khz;
46 
47 static DEFINE_STATIC_KEY_FALSE(__use_tsc);
48 
49 int tsc_clocksource_reliable;
50 
51 static int __read_mostly tsc_force_recalibrate;
52 
53 static u32 art_to_tsc_numerator;
54 static u32 art_to_tsc_denominator;
55 static u64 art_to_tsc_offset;
56 static struct clocksource *art_related_clocksource;
57 
58 struct cyc2ns {
59 	struct cyc2ns_data data[2];	/*  0 + 2*16 = 32 */
60 	seqcount_latch_t   seq;		/* 32 + 4    = 36 */
61 
62 }; /* fits one cacheline */
63 
64 static DEFINE_PER_CPU_ALIGNED(struct cyc2ns, cyc2ns);
65 
tsc_early_khz_setup(char * buf)66 static int __init tsc_early_khz_setup(char *buf)
67 {
68 	return kstrtouint(buf, 0, &tsc_early_khz);
69 }
70 early_param("tsc_early_khz", tsc_early_khz_setup);
71 
__cyc2ns_read(struct cyc2ns_data * data)72 __always_inline void __cyc2ns_read(struct cyc2ns_data *data)
73 {
74 	int seq, idx;
75 
76 	do {
77 		seq = this_cpu_read(cyc2ns.seq.seqcount.sequence);
78 		idx = seq & 1;
79 
80 		data->cyc2ns_offset = this_cpu_read(cyc2ns.data[idx].cyc2ns_offset);
81 		data->cyc2ns_mul    = this_cpu_read(cyc2ns.data[idx].cyc2ns_mul);
82 		data->cyc2ns_shift  = this_cpu_read(cyc2ns.data[idx].cyc2ns_shift);
83 
84 	} while (unlikely(seq != this_cpu_read(cyc2ns.seq.seqcount.sequence)));
85 }
86 
cyc2ns_read_begin(struct cyc2ns_data * data)87 __always_inline void cyc2ns_read_begin(struct cyc2ns_data *data)
88 {
89 	preempt_disable_notrace();
90 	__cyc2ns_read(data);
91 }
92 
cyc2ns_read_end(void)93 __always_inline void cyc2ns_read_end(void)
94 {
95 	preempt_enable_notrace();
96 }
97 
98 /*
99  * Accelerators for sched_clock()
100  * convert from cycles(64bits) => nanoseconds (64bits)
101  *  basic equation:
102  *              ns = cycles / (freq / ns_per_sec)
103  *              ns = cycles * (ns_per_sec / freq)
104  *              ns = cycles * (10^9 / (cpu_khz * 10^3))
105  *              ns = cycles * (10^6 / cpu_khz)
106  *
107  *      Then we use scaling math (suggested by george@mvista.com) to get:
108  *              ns = cycles * (10^6 * SC / cpu_khz) / SC
109  *              ns = cycles * cyc2ns_scale / SC
110  *
111  *      And since SC is a constant power of two, we can convert the div
112  *  into a shift. The larger SC is, the more accurate the conversion, but
113  *  cyc2ns_scale needs to be a 32-bit value so that 32-bit multiplication
114  *  (64-bit result) can be used.
115  *
116  *  We can use khz divisor instead of mhz to keep a better precision.
117  *  (mathieu.desnoyers@polymtl.ca)
118  *
119  *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
120  */
121 
__cycles_2_ns(unsigned long long cyc)122 static __always_inline unsigned long long __cycles_2_ns(unsigned long long cyc)
123 {
124 	struct cyc2ns_data data;
125 	unsigned long long ns;
126 
127 	__cyc2ns_read(&data);
128 
129 	ns = data.cyc2ns_offset;
130 	ns += mul_u64_u32_shr(cyc, data.cyc2ns_mul, data.cyc2ns_shift);
131 
132 	return ns;
133 }
134 
cycles_2_ns(unsigned long long cyc)135 static __always_inline unsigned long long cycles_2_ns(unsigned long long cyc)
136 {
137 	unsigned long long ns;
138 	preempt_disable_notrace();
139 	ns = __cycles_2_ns(cyc);
140 	preempt_enable_notrace();
141 	return ns;
142 }
143 
__set_cyc2ns_scale(unsigned long khz,int cpu,unsigned long long tsc_now)144 static void __set_cyc2ns_scale(unsigned long khz, int cpu, unsigned long long tsc_now)
145 {
146 	unsigned long long ns_now;
147 	struct cyc2ns_data data;
148 	struct cyc2ns *c2n;
149 
150 	ns_now = cycles_2_ns(tsc_now);
151 
152 	/*
153 	 * Compute a new multiplier as per the above comment and ensure our
154 	 * time function is continuous; see the comment near struct
155 	 * cyc2ns_data.
156 	 */
157 	clocks_calc_mult_shift(&data.cyc2ns_mul, &data.cyc2ns_shift, khz,
158 			       NSEC_PER_MSEC, 0);
159 
160 	/*
161 	 * cyc2ns_shift is exported via arch_perf_update_userpage() where it is
162 	 * not expected to be greater than 31 due to the original published
163 	 * conversion algorithm shifting a 32-bit value (now specifies a 64-bit
164 	 * value) - refer perf_event_mmap_page documentation in perf_event.h.
165 	 */
166 	if (data.cyc2ns_shift == 32) {
167 		data.cyc2ns_shift = 31;
168 		data.cyc2ns_mul >>= 1;
169 	}
170 
171 	data.cyc2ns_offset = ns_now -
172 		mul_u64_u32_shr(tsc_now, data.cyc2ns_mul, data.cyc2ns_shift);
173 
174 	c2n = per_cpu_ptr(&cyc2ns, cpu);
175 
176 	raw_write_seqcount_latch(&c2n->seq);
177 	c2n->data[0] = data;
178 	raw_write_seqcount_latch(&c2n->seq);
179 	c2n->data[1] = data;
180 }
181 
set_cyc2ns_scale(unsigned long khz,int cpu,unsigned long long tsc_now)182 static void set_cyc2ns_scale(unsigned long khz, int cpu, unsigned long long tsc_now)
183 {
184 	unsigned long flags;
185 
186 	local_irq_save(flags);
187 	sched_clock_idle_sleep_event();
188 
189 	if (khz)
190 		__set_cyc2ns_scale(khz, cpu, tsc_now);
191 
192 	sched_clock_idle_wakeup_event();
193 	local_irq_restore(flags);
194 }
195 
196 /*
197  * Initialize cyc2ns for boot cpu
198  */
cyc2ns_init_boot_cpu(void)199 static void __init cyc2ns_init_boot_cpu(void)
200 {
201 	struct cyc2ns *c2n = this_cpu_ptr(&cyc2ns);
202 
203 	seqcount_latch_init(&c2n->seq);
204 	__set_cyc2ns_scale(tsc_khz, smp_processor_id(), rdtsc());
205 }
206 
207 /*
208  * Secondary CPUs do not run through tsc_init(), so set up
209  * all the scale factors for all CPUs, assuming the same
210  * speed as the bootup CPU.
211  */
cyc2ns_init_secondary_cpus(void)212 static void __init cyc2ns_init_secondary_cpus(void)
213 {
214 	unsigned int cpu, this_cpu = smp_processor_id();
215 	struct cyc2ns *c2n = this_cpu_ptr(&cyc2ns);
216 	struct cyc2ns_data *data = c2n->data;
217 
218 	for_each_possible_cpu(cpu) {
219 		if (cpu != this_cpu) {
220 			seqcount_latch_init(&c2n->seq);
221 			c2n = per_cpu_ptr(&cyc2ns, cpu);
222 			c2n->data[0] = data[0];
223 			c2n->data[1] = data[1];
224 		}
225 	}
226 }
227 
228 /*
229  * Scheduler clock - returns current time in nanosec units.
230  */
native_sched_clock(void)231 noinstr u64 native_sched_clock(void)
232 {
233 	if (static_branch_likely(&__use_tsc)) {
234 		u64 tsc_now = rdtsc();
235 
236 		/* return the value in ns */
237 		return __cycles_2_ns(tsc_now);
238 	}
239 
240 	/*
241 	 * Fall back to jiffies if there's no TSC available:
242 	 * ( But note that we still use it if the TSC is marked
243 	 *   unstable. We do this because unlike Time Of Day,
244 	 *   the scheduler clock tolerates small errors and it's
245 	 *   very important for it to be as fast as the platform
246 	 *   can achieve it. )
247 	 */
248 
249 	/* No locking but a rare wrong value is not a big deal: */
250 	return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
251 }
252 
253 /*
254  * Generate a sched_clock if you already have a TSC value.
255  */
native_sched_clock_from_tsc(u64 tsc)256 u64 native_sched_clock_from_tsc(u64 tsc)
257 {
258 	return cycles_2_ns(tsc);
259 }
260 
261 /* We need to define a real function for sched_clock, to override the
262    weak default version */
263 #ifdef CONFIG_PARAVIRT
sched_clock_noinstr(void)264 noinstr u64 sched_clock_noinstr(void)
265 {
266 	return paravirt_sched_clock();
267 }
268 
using_native_sched_clock(void)269 bool using_native_sched_clock(void)
270 {
271 	return static_call_query(pv_sched_clock) == native_sched_clock;
272 }
273 #else
274 u64 sched_clock_noinstr(void) __attribute__((alias("native_sched_clock")));
275 
using_native_sched_clock(void)276 bool using_native_sched_clock(void) { return true; }
277 #endif
278 
sched_clock(void)279 notrace u64 sched_clock(void)
280 {
281 	u64 now;
282 	preempt_disable_notrace();
283 	now = sched_clock_noinstr();
284 	preempt_enable_notrace();
285 	return now;
286 }
287 
check_tsc_unstable(void)288 int check_tsc_unstable(void)
289 {
290 	return tsc_unstable;
291 }
292 EXPORT_SYMBOL_GPL(check_tsc_unstable);
293 
294 #ifdef CONFIG_X86_TSC
notsc_setup(char * str)295 int __init notsc_setup(char *str)
296 {
297 	mark_tsc_unstable("boot parameter notsc");
298 	return 1;
299 }
300 #else
301 /*
302  * disable flag for tsc. Takes effect by clearing the TSC cpu flag
303  * in cpu/common.c
304  */
notsc_setup(char * str)305 int __init notsc_setup(char *str)
306 {
307 	setup_clear_cpu_cap(X86_FEATURE_TSC);
308 	return 1;
309 }
310 #endif
311 
312 __setup("notsc", notsc_setup);
313 
314 static int no_sched_irq_time;
315 static int no_tsc_watchdog;
316 static int tsc_as_watchdog;
317 
tsc_setup(char * str)318 static int __init tsc_setup(char *str)
319 {
320 	if (!strcmp(str, "reliable"))
321 		tsc_clocksource_reliable = 1;
322 	if (!strncmp(str, "noirqtime", 9))
323 		no_sched_irq_time = 1;
324 	if (!strcmp(str, "unstable"))
325 		mark_tsc_unstable("boot parameter");
326 	if (!strcmp(str, "nowatchdog")) {
327 		no_tsc_watchdog = 1;
328 		if (tsc_as_watchdog)
329 			pr_alert("%s: Overriding earlier tsc=watchdog with tsc=nowatchdog\n",
330 				 __func__);
331 		tsc_as_watchdog = 0;
332 	}
333 	if (!strcmp(str, "recalibrate"))
334 		tsc_force_recalibrate = 1;
335 	if (!strcmp(str, "watchdog")) {
336 		if (no_tsc_watchdog)
337 			pr_alert("%s: tsc=watchdog overridden by earlier tsc=nowatchdog\n",
338 				 __func__);
339 		else
340 			tsc_as_watchdog = 1;
341 	}
342 	return 1;
343 }
344 
345 __setup("tsc=", tsc_setup);
346 
347 #define MAX_RETRIES		5
348 #define TSC_DEFAULT_THRESHOLD	0x20000
349 
350 /*
351  * Read TSC and the reference counters. Take care of any disturbances
352  */
tsc_read_refs(u64 * p,int hpet)353 static u64 tsc_read_refs(u64 *p, int hpet)
354 {
355 	u64 t1, t2;
356 	u64 thresh = tsc_khz ? tsc_khz >> 5 : TSC_DEFAULT_THRESHOLD;
357 	int i;
358 
359 	for (i = 0; i < MAX_RETRIES; i++) {
360 		t1 = get_cycles();
361 		if (hpet)
362 			*p = hpet_readl(HPET_COUNTER) & 0xFFFFFFFF;
363 		else
364 			*p = acpi_pm_read_early();
365 		t2 = get_cycles();
366 		if ((t2 - t1) < thresh)
367 			return t2;
368 	}
369 	return ULLONG_MAX;
370 }
371 
372 /*
373  * Calculate the TSC frequency from HPET reference
374  */
calc_hpet_ref(u64 deltatsc,u64 hpet1,u64 hpet2)375 static unsigned long calc_hpet_ref(u64 deltatsc, u64 hpet1, u64 hpet2)
376 {
377 	u64 tmp;
378 
379 	if (hpet2 < hpet1)
380 		hpet2 += 0x100000000ULL;
381 	hpet2 -= hpet1;
382 	tmp = ((u64)hpet2 * hpet_readl(HPET_PERIOD));
383 	do_div(tmp, 1000000);
384 	deltatsc = div64_u64(deltatsc, tmp);
385 
386 	return (unsigned long) deltatsc;
387 }
388 
389 /*
390  * Calculate the TSC frequency from PMTimer reference
391  */
calc_pmtimer_ref(u64 deltatsc,u64 pm1,u64 pm2)392 static unsigned long calc_pmtimer_ref(u64 deltatsc, u64 pm1, u64 pm2)
393 {
394 	u64 tmp;
395 
396 	if (!pm1 && !pm2)
397 		return ULONG_MAX;
398 
399 	if (pm2 < pm1)
400 		pm2 += (u64)ACPI_PM_OVRRUN;
401 	pm2 -= pm1;
402 	tmp = pm2 * 1000000000LL;
403 	do_div(tmp, PMTMR_TICKS_PER_SEC);
404 	do_div(deltatsc, tmp);
405 
406 	return (unsigned long) deltatsc;
407 }
408 
409 #define CAL_MS		10
410 #define CAL_LATCH	(PIT_TICK_RATE / (1000 / CAL_MS))
411 #define CAL_PIT_LOOPS	1000
412 
413 #define CAL2_MS		50
414 #define CAL2_LATCH	(PIT_TICK_RATE / (1000 / CAL2_MS))
415 #define CAL2_PIT_LOOPS	5000
416 
417 
418 /*
419  * Try to calibrate the TSC against the Programmable
420  * Interrupt Timer and return the frequency of the TSC
421  * in kHz.
422  *
423  * Return ULONG_MAX on failure to calibrate.
424  */
pit_calibrate_tsc(u32 latch,unsigned long ms,int loopmin)425 static unsigned long pit_calibrate_tsc(u32 latch, unsigned long ms, int loopmin)
426 {
427 	u64 tsc, t1, t2, delta;
428 	unsigned long tscmin, tscmax;
429 	int pitcnt;
430 
431 	if (!has_legacy_pic()) {
432 		/*
433 		 * Relies on tsc_early_delay_calibrate() to have given us semi
434 		 * usable udelay(), wait for the same 50ms we would have with
435 		 * the PIT loop below.
436 		 */
437 		udelay(10 * USEC_PER_MSEC);
438 		udelay(10 * USEC_PER_MSEC);
439 		udelay(10 * USEC_PER_MSEC);
440 		udelay(10 * USEC_PER_MSEC);
441 		udelay(10 * USEC_PER_MSEC);
442 		return ULONG_MAX;
443 	}
444 
445 	/* Set the Gate high, disable speaker */
446 	outb((inb(0x61) & ~0x02) | 0x01, 0x61);
447 
448 	/*
449 	 * Setup CTC channel 2* for mode 0, (interrupt on terminal
450 	 * count mode), binary count. Set the latch register to 50ms
451 	 * (LSB then MSB) to begin countdown.
452 	 */
453 	outb(0xb0, 0x43);
454 	outb(latch & 0xff, 0x42);
455 	outb(latch >> 8, 0x42);
456 
457 	tsc = t1 = t2 = get_cycles();
458 
459 	pitcnt = 0;
460 	tscmax = 0;
461 	tscmin = ULONG_MAX;
462 	while ((inb(0x61) & 0x20) == 0) {
463 		t2 = get_cycles();
464 		delta = t2 - tsc;
465 		tsc = t2;
466 		if ((unsigned long) delta < tscmin)
467 			tscmin = (unsigned int) delta;
468 		if ((unsigned long) delta > tscmax)
469 			tscmax = (unsigned int) delta;
470 		pitcnt++;
471 	}
472 
473 	/*
474 	 * Sanity checks:
475 	 *
476 	 * If we were not able to read the PIT more than loopmin
477 	 * times, then we have been hit by a massive SMI
478 	 *
479 	 * If the maximum is 10 times larger than the minimum,
480 	 * then we got hit by an SMI as well.
481 	 */
482 	if (pitcnt < loopmin || tscmax > 10 * tscmin)
483 		return ULONG_MAX;
484 
485 	/* Calculate the PIT value */
486 	delta = t2 - t1;
487 	do_div(delta, ms);
488 	return delta;
489 }
490 
491 /*
492  * This reads the current MSB of the PIT counter, and
493  * checks if we are running on sufficiently fast and
494  * non-virtualized hardware.
495  *
496  * Our expectations are:
497  *
498  *  - the PIT is running at roughly 1.19MHz
499  *
500  *  - each IO is going to take about 1us on real hardware,
501  *    but we allow it to be much faster (by a factor of 10) or
502  *    _slightly_ slower (ie we allow up to a 2us read+counter
503  *    update - anything else implies a unacceptably slow CPU
504  *    or PIT for the fast calibration to work.
505  *
506  *  - with 256 PIT ticks to read the value, we have 214us to
507  *    see the same MSB (and overhead like doing a single TSC
508  *    read per MSB value etc).
509  *
510  *  - We're doing 2 reads per loop (LSB, MSB), and we expect
511  *    them each to take about a microsecond on real hardware.
512  *    So we expect a count value of around 100. But we'll be
513  *    generous, and accept anything over 50.
514  *
515  *  - if the PIT is stuck, and we see *many* more reads, we
516  *    return early (and the next caller of pit_expect_msb()
517  *    then consider it a failure when they don't see the
518  *    next expected value).
519  *
520  * These expectations mean that we know that we have seen the
521  * transition from one expected value to another with a fairly
522  * high accuracy, and we didn't miss any events. We can thus
523  * use the TSC value at the transitions to calculate a pretty
524  * good value for the TSC frequency.
525  */
pit_verify_msb(unsigned char val)526 static inline int pit_verify_msb(unsigned char val)
527 {
528 	/* Ignore LSB */
529 	inb(0x42);
530 	return inb(0x42) == val;
531 }
532 
pit_expect_msb(unsigned char val,u64 * tscp,unsigned long * deltap)533 static inline int pit_expect_msb(unsigned char val, u64 *tscp, unsigned long *deltap)
534 {
535 	int count;
536 	u64 tsc = 0, prev_tsc = 0;
537 
538 	for (count = 0; count < 50000; count++) {
539 		if (!pit_verify_msb(val))
540 			break;
541 		prev_tsc = tsc;
542 		tsc = get_cycles();
543 	}
544 	*deltap = get_cycles() - prev_tsc;
545 	*tscp = tsc;
546 
547 	/*
548 	 * We require _some_ success, but the quality control
549 	 * will be based on the error terms on the TSC values.
550 	 */
551 	return count > 5;
552 }
553 
554 /*
555  * How many MSB values do we want to see? We aim for
556  * a maximum error rate of 500ppm (in practice the
557  * real error is much smaller), but refuse to spend
558  * more than 50ms on it.
559  */
560 #define MAX_QUICK_PIT_MS 50
561 #define MAX_QUICK_PIT_ITERATIONS (MAX_QUICK_PIT_MS * PIT_TICK_RATE / 1000 / 256)
562 
quick_pit_calibrate(void)563 static unsigned long quick_pit_calibrate(void)
564 {
565 	int i;
566 	u64 tsc, delta;
567 	unsigned long d1, d2;
568 
569 	if (!has_legacy_pic())
570 		return 0;
571 
572 	/* Set the Gate high, disable speaker */
573 	outb((inb(0x61) & ~0x02) | 0x01, 0x61);
574 
575 	/*
576 	 * Counter 2, mode 0 (one-shot), binary count
577 	 *
578 	 * NOTE! Mode 2 decrements by two (and then the
579 	 * output is flipped each time, giving the same
580 	 * final output frequency as a decrement-by-one),
581 	 * so mode 0 is much better when looking at the
582 	 * individual counts.
583 	 */
584 	outb(0xb0, 0x43);
585 
586 	/* Start at 0xffff */
587 	outb(0xff, 0x42);
588 	outb(0xff, 0x42);
589 
590 	/*
591 	 * The PIT starts counting at the next edge, so we
592 	 * need to delay for a microsecond. The easiest way
593 	 * to do that is to just read back the 16-bit counter
594 	 * once from the PIT.
595 	 */
596 	pit_verify_msb(0);
597 
598 	if (pit_expect_msb(0xff, &tsc, &d1)) {
599 		for (i = 1; i <= MAX_QUICK_PIT_ITERATIONS; i++) {
600 			if (!pit_expect_msb(0xff-i, &delta, &d2))
601 				break;
602 
603 			delta -= tsc;
604 
605 			/*
606 			 * Extrapolate the error and fail fast if the error will
607 			 * never be below 500 ppm.
608 			 */
609 			if (i == 1 &&
610 			    d1 + d2 >= (delta * MAX_QUICK_PIT_ITERATIONS) >> 11)
611 				return 0;
612 
613 			/*
614 			 * Iterate until the error is less than 500 ppm
615 			 */
616 			if (d1+d2 >= delta >> 11)
617 				continue;
618 
619 			/*
620 			 * Check the PIT one more time to verify that
621 			 * all TSC reads were stable wrt the PIT.
622 			 *
623 			 * This also guarantees serialization of the
624 			 * last cycle read ('d2') in pit_expect_msb.
625 			 */
626 			if (!pit_verify_msb(0xfe - i))
627 				break;
628 			goto success;
629 		}
630 	}
631 	pr_info("Fast TSC calibration failed\n");
632 	return 0;
633 
634 success:
635 	/*
636 	 * Ok, if we get here, then we've seen the
637 	 * MSB of the PIT decrement 'i' times, and the
638 	 * error has shrunk to less than 500 ppm.
639 	 *
640 	 * As a result, we can depend on there not being
641 	 * any odd delays anywhere, and the TSC reads are
642 	 * reliable (within the error).
643 	 *
644 	 * kHz = ticks / time-in-seconds / 1000;
645 	 * kHz = (t2 - t1) / (I * 256 / PIT_TICK_RATE) / 1000
646 	 * kHz = ((t2 - t1) * PIT_TICK_RATE) / (I * 256 * 1000)
647 	 */
648 	delta *= PIT_TICK_RATE;
649 	do_div(delta, i*256*1000);
650 	pr_info("Fast TSC calibration using PIT\n");
651 	return delta;
652 }
653 
654 /**
655  * native_calibrate_tsc
656  * Determine TSC frequency via CPUID, else return 0.
657  */
native_calibrate_tsc(void)658 unsigned long native_calibrate_tsc(void)
659 {
660 	unsigned int eax_denominator, ebx_numerator, ecx_hz, edx;
661 	unsigned int crystal_khz;
662 
663 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
664 		return 0;
665 
666 	if (boot_cpu_data.cpuid_level < 0x15)
667 		return 0;
668 
669 	eax_denominator = ebx_numerator = ecx_hz = edx = 0;
670 
671 	/* CPUID 15H TSC/Crystal ratio, plus optionally Crystal Hz */
672 	cpuid(0x15, &eax_denominator, &ebx_numerator, &ecx_hz, &edx);
673 
674 	if (ebx_numerator == 0 || eax_denominator == 0)
675 		return 0;
676 
677 	crystal_khz = ecx_hz / 1000;
678 
679 	/*
680 	 * Denverton SoCs don't report crystal clock, and also don't support
681 	 * CPUID.0x16 for the calculation below, so hardcode the 25MHz crystal
682 	 * clock.
683 	 */
684 	if (crystal_khz == 0 &&
685 			boot_cpu_data.x86_model == INTEL_FAM6_ATOM_GOLDMONT_D)
686 		crystal_khz = 25000;
687 
688 	/*
689 	 * TSC frequency reported directly by CPUID is a "hardware reported"
690 	 * frequency and is the most accurate one so far we have. This
691 	 * is considered a known frequency.
692 	 */
693 	if (crystal_khz != 0)
694 		setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ);
695 
696 	/*
697 	 * Some Intel SoCs like Skylake and Kabylake don't report the crystal
698 	 * clock, but we can easily calculate it to a high degree of accuracy
699 	 * by considering the crystal ratio and the CPU speed.
700 	 */
701 	if (crystal_khz == 0 && boot_cpu_data.cpuid_level >= 0x16) {
702 		unsigned int eax_base_mhz, ebx, ecx, edx;
703 
704 		cpuid(0x16, &eax_base_mhz, &ebx, &ecx, &edx);
705 		crystal_khz = eax_base_mhz * 1000 *
706 			eax_denominator / ebx_numerator;
707 	}
708 
709 	if (crystal_khz == 0)
710 		return 0;
711 
712 	/*
713 	 * For Atom SoCs TSC is the only reliable clocksource.
714 	 * Mark TSC reliable so no watchdog on it.
715 	 */
716 	if (boot_cpu_data.x86_model == INTEL_FAM6_ATOM_GOLDMONT)
717 		setup_force_cpu_cap(X86_FEATURE_TSC_RELIABLE);
718 
719 #ifdef CONFIG_X86_LOCAL_APIC
720 	/*
721 	 * The local APIC appears to be fed by the core crystal clock
722 	 * (which sounds entirely sensible). We can set the global
723 	 * lapic_timer_period here to avoid having to calibrate the APIC
724 	 * timer later.
725 	 */
726 	lapic_timer_period = crystal_khz * 1000 / HZ;
727 #endif
728 
729 	return crystal_khz * ebx_numerator / eax_denominator;
730 }
731 
cpu_khz_from_cpuid(void)732 static unsigned long cpu_khz_from_cpuid(void)
733 {
734 	unsigned int eax_base_mhz, ebx_max_mhz, ecx_bus_mhz, edx;
735 
736 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
737 		return 0;
738 
739 	if (boot_cpu_data.cpuid_level < 0x16)
740 		return 0;
741 
742 	eax_base_mhz = ebx_max_mhz = ecx_bus_mhz = edx = 0;
743 
744 	cpuid(0x16, &eax_base_mhz, &ebx_max_mhz, &ecx_bus_mhz, &edx);
745 
746 	return eax_base_mhz * 1000;
747 }
748 
749 /*
750  * calibrate cpu using pit, hpet, and ptimer methods. They are available
751  * later in boot after acpi is initialized.
752  */
pit_hpet_ptimer_calibrate_cpu(void)753 static unsigned long pit_hpet_ptimer_calibrate_cpu(void)
754 {
755 	u64 tsc1, tsc2, delta, ref1, ref2;
756 	unsigned long tsc_pit_min = ULONG_MAX, tsc_ref_min = ULONG_MAX;
757 	unsigned long flags, latch, ms;
758 	int hpet = is_hpet_enabled(), i, loopmin;
759 
760 	/*
761 	 * Run 5 calibration loops to get the lowest frequency value
762 	 * (the best estimate). We use two different calibration modes
763 	 * here:
764 	 *
765 	 * 1) PIT loop. We set the PIT Channel 2 to oneshot mode and
766 	 * load a timeout of 50ms. We read the time right after we
767 	 * started the timer and wait until the PIT count down reaches
768 	 * zero. In each wait loop iteration we read the TSC and check
769 	 * the delta to the previous read. We keep track of the min
770 	 * and max values of that delta. The delta is mostly defined
771 	 * by the IO time of the PIT access, so we can detect when
772 	 * any disturbance happened between the two reads. If the
773 	 * maximum time is significantly larger than the minimum time,
774 	 * then we discard the result and have another try.
775 	 *
776 	 * 2) Reference counter. If available we use the HPET or the
777 	 * PMTIMER as a reference to check the sanity of that value.
778 	 * We use separate TSC readouts and check inside of the
779 	 * reference read for any possible disturbance. We discard
780 	 * disturbed values here as well. We do that around the PIT
781 	 * calibration delay loop as we have to wait for a certain
782 	 * amount of time anyway.
783 	 */
784 
785 	/* Preset PIT loop values */
786 	latch = CAL_LATCH;
787 	ms = CAL_MS;
788 	loopmin = CAL_PIT_LOOPS;
789 
790 	for (i = 0; i < 3; i++) {
791 		unsigned long tsc_pit_khz;
792 
793 		/*
794 		 * Read the start value and the reference count of
795 		 * hpet/pmtimer when available. Then do the PIT
796 		 * calibration, which will take at least 50ms, and
797 		 * read the end value.
798 		 */
799 		local_irq_save(flags);
800 		tsc1 = tsc_read_refs(&ref1, hpet);
801 		tsc_pit_khz = pit_calibrate_tsc(latch, ms, loopmin);
802 		tsc2 = tsc_read_refs(&ref2, hpet);
803 		local_irq_restore(flags);
804 
805 		/* Pick the lowest PIT TSC calibration so far */
806 		tsc_pit_min = min(tsc_pit_min, tsc_pit_khz);
807 
808 		/* hpet or pmtimer available ? */
809 		if (ref1 == ref2)
810 			continue;
811 
812 		/* Check, whether the sampling was disturbed */
813 		if (tsc1 == ULLONG_MAX || tsc2 == ULLONG_MAX)
814 			continue;
815 
816 		tsc2 = (tsc2 - tsc1) * 1000000LL;
817 		if (hpet)
818 			tsc2 = calc_hpet_ref(tsc2, ref1, ref2);
819 		else
820 			tsc2 = calc_pmtimer_ref(tsc2, ref1, ref2);
821 
822 		tsc_ref_min = min(tsc_ref_min, (unsigned long) tsc2);
823 
824 		/* Check the reference deviation */
825 		delta = ((u64) tsc_pit_min) * 100;
826 		do_div(delta, tsc_ref_min);
827 
828 		/*
829 		 * If both calibration results are inside a 10% window
830 		 * then we can be sure, that the calibration
831 		 * succeeded. We break out of the loop right away. We
832 		 * use the reference value, as it is more precise.
833 		 */
834 		if (delta >= 90 && delta <= 110) {
835 			pr_info("PIT calibration matches %s. %d loops\n",
836 				hpet ? "HPET" : "PMTIMER", i + 1);
837 			return tsc_ref_min;
838 		}
839 
840 		/*
841 		 * Check whether PIT failed more than once. This
842 		 * happens in virtualized environments. We need to
843 		 * give the virtual PC a slightly longer timeframe for
844 		 * the HPET/PMTIMER to make the result precise.
845 		 */
846 		if (i == 1 && tsc_pit_min == ULONG_MAX) {
847 			latch = CAL2_LATCH;
848 			ms = CAL2_MS;
849 			loopmin = CAL2_PIT_LOOPS;
850 		}
851 	}
852 
853 	/*
854 	 * Now check the results.
855 	 */
856 	if (tsc_pit_min == ULONG_MAX) {
857 		/* PIT gave no useful value */
858 		pr_warn("Unable to calibrate against PIT\n");
859 
860 		/* We don't have an alternative source, disable TSC */
861 		if (!hpet && !ref1 && !ref2) {
862 			pr_notice("No reference (HPET/PMTIMER) available\n");
863 			return 0;
864 		}
865 
866 		/* The alternative source failed as well, disable TSC */
867 		if (tsc_ref_min == ULONG_MAX) {
868 			pr_warn("HPET/PMTIMER calibration failed\n");
869 			return 0;
870 		}
871 
872 		/* Use the alternative source */
873 		pr_info("using %s reference calibration\n",
874 			hpet ? "HPET" : "PMTIMER");
875 
876 		return tsc_ref_min;
877 	}
878 
879 	/* We don't have an alternative source, use the PIT calibration value */
880 	if (!hpet && !ref1 && !ref2) {
881 		pr_info("Using PIT calibration value\n");
882 		return tsc_pit_min;
883 	}
884 
885 	/* The alternative source failed, use the PIT calibration value */
886 	if (tsc_ref_min == ULONG_MAX) {
887 		pr_warn("HPET/PMTIMER calibration failed. Using PIT calibration.\n");
888 		return tsc_pit_min;
889 	}
890 
891 	/*
892 	 * The calibration values differ too much. In doubt, we use
893 	 * the PIT value as we know that there are PMTIMERs around
894 	 * running at double speed. At least we let the user know:
895 	 */
896 	pr_warn("PIT calibration deviates from %s: %lu %lu\n",
897 		hpet ? "HPET" : "PMTIMER", tsc_pit_min, tsc_ref_min);
898 	pr_info("Using PIT calibration value\n");
899 	return tsc_pit_min;
900 }
901 
902 /**
903  * native_calibrate_cpu_early - can calibrate the cpu early in boot
904  */
native_calibrate_cpu_early(void)905 unsigned long native_calibrate_cpu_early(void)
906 {
907 	unsigned long flags, fast_calibrate = cpu_khz_from_cpuid();
908 
909 	if (!fast_calibrate)
910 		fast_calibrate = cpu_khz_from_msr();
911 	if (!fast_calibrate) {
912 		local_irq_save(flags);
913 		fast_calibrate = quick_pit_calibrate();
914 		local_irq_restore(flags);
915 	}
916 	return fast_calibrate;
917 }
918 
919 
920 /**
921  * native_calibrate_cpu - calibrate the cpu
922  */
native_calibrate_cpu(void)923 static unsigned long native_calibrate_cpu(void)
924 {
925 	unsigned long tsc_freq = native_calibrate_cpu_early();
926 
927 	if (!tsc_freq)
928 		tsc_freq = pit_hpet_ptimer_calibrate_cpu();
929 
930 	return tsc_freq;
931 }
932 
recalibrate_cpu_khz(void)933 void recalibrate_cpu_khz(void)
934 {
935 #ifndef CONFIG_SMP
936 	unsigned long cpu_khz_old = cpu_khz;
937 
938 	if (!boot_cpu_has(X86_FEATURE_TSC))
939 		return;
940 
941 	cpu_khz = x86_platform.calibrate_cpu();
942 	tsc_khz = x86_platform.calibrate_tsc();
943 	if (tsc_khz == 0)
944 		tsc_khz = cpu_khz;
945 	else if (abs(cpu_khz - tsc_khz) * 10 > tsc_khz)
946 		cpu_khz = tsc_khz;
947 	cpu_data(0).loops_per_jiffy = cpufreq_scale(cpu_data(0).loops_per_jiffy,
948 						    cpu_khz_old, cpu_khz);
949 #endif
950 }
951 EXPORT_SYMBOL_GPL(recalibrate_cpu_khz);
952 
953 
954 static unsigned long long cyc2ns_suspend;
955 
tsc_save_sched_clock_state(void)956 void tsc_save_sched_clock_state(void)
957 {
958 	if (!sched_clock_stable())
959 		return;
960 
961 	cyc2ns_suspend = sched_clock();
962 }
963 
964 /*
965  * Even on processors with invariant TSC, TSC gets reset in some the
966  * ACPI system sleep states. And in some systems BIOS seem to reinit TSC to
967  * arbitrary value (still sync'd across cpu's) during resume from such sleep
968  * states. To cope up with this, recompute the cyc2ns_offset for each cpu so
969  * that sched_clock() continues from the point where it was left off during
970  * suspend.
971  */
tsc_restore_sched_clock_state(void)972 void tsc_restore_sched_clock_state(void)
973 {
974 	unsigned long long offset;
975 	unsigned long flags;
976 	int cpu;
977 
978 	if (!sched_clock_stable())
979 		return;
980 
981 	local_irq_save(flags);
982 
983 	/*
984 	 * We're coming out of suspend, there's no concurrency yet; don't
985 	 * bother being nice about the RCU stuff, just write to both
986 	 * data fields.
987 	 */
988 
989 	this_cpu_write(cyc2ns.data[0].cyc2ns_offset, 0);
990 	this_cpu_write(cyc2ns.data[1].cyc2ns_offset, 0);
991 
992 	offset = cyc2ns_suspend - sched_clock();
993 
994 	for_each_possible_cpu(cpu) {
995 		per_cpu(cyc2ns.data[0].cyc2ns_offset, cpu) = offset;
996 		per_cpu(cyc2ns.data[1].cyc2ns_offset, cpu) = offset;
997 	}
998 
999 	local_irq_restore(flags);
1000 }
1001 
1002 #ifdef CONFIG_CPU_FREQ
1003 /*
1004  * Frequency scaling support. Adjust the TSC based timer when the CPU frequency
1005  * changes.
1006  *
1007  * NOTE: On SMP the situation is not fixable in general, so simply mark the TSC
1008  * as unstable and give up in those cases.
1009  *
1010  * Should fix up last_tsc too. Currently gettimeofday in the
1011  * first tick after the change will be slightly wrong.
1012  */
1013 
1014 static unsigned int  ref_freq;
1015 static unsigned long loops_per_jiffy_ref;
1016 static unsigned long tsc_khz_ref;
1017 
time_cpufreq_notifier(struct notifier_block * nb,unsigned long val,void * data)1018 static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
1019 				void *data)
1020 {
1021 	struct cpufreq_freqs *freq = data;
1022 
1023 	if (num_online_cpus() > 1) {
1024 		mark_tsc_unstable("cpufreq changes on SMP");
1025 		return 0;
1026 	}
1027 
1028 	if (!ref_freq) {
1029 		ref_freq = freq->old;
1030 		loops_per_jiffy_ref = boot_cpu_data.loops_per_jiffy;
1031 		tsc_khz_ref = tsc_khz;
1032 	}
1033 
1034 	if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
1035 	    (val == CPUFREQ_POSTCHANGE && freq->old > freq->new)) {
1036 		boot_cpu_data.loops_per_jiffy =
1037 			cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
1038 
1039 		tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
1040 		if (!(freq->flags & CPUFREQ_CONST_LOOPS))
1041 			mark_tsc_unstable("cpufreq changes");
1042 
1043 		set_cyc2ns_scale(tsc_khz, freq->policy->cpu, rdtsc());
1044 	}
1045 
1046 	return 0;
1047 }
1048 
1049 static struct notifier_block time_cpufreq_notifier_block = {
1050 	.notifier_call  = time_cpufreq_notifier
1051 };
1052 
cpufreq_register_tsc_scaling(void)1053 static int __init cpufreq_register_tsc_scaling(void)
1054 {
1055 	if (!boot_cpu_has(X86_FEATURE_TSC))
1056 		return 0;
1057 	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
1058 		return 0;
1059 	cpufreq_register_notifier(&time_cpufreq_notifier_block,
1060 				CPUFREQ_TRANSITION_NOTIFIER);
1061 	return 0;
1062 }
1063 
1064 core_initcall(cpufreq_register_tsc_scaling);
1065 
1066 #endif /* CONFIG_CPU_FREQ */
1067 
1068 #define ART_CPUID_LEAF (0x15)
1069 #define ART_MIN_DENOMINATOR (1)
1070 
1071 
1072 /*
1073  * If ART is present detect the numerator:denominator to convert to TSC
1074  */
detect_art(void)1075 static void __init detect_art(void)
1076 {
1077 	unsigned int unused[2];
1078 
1079 	if (boot_cpu_data.cpuid_level < ART_CPUID_LEAF)
1080 		return;
1081 
1082 	/*
1083 	 * Don't enable ART in a VM, non-stop TSC and TSC_ADJUST required,
1084 	 * and the TSC counter resets must not occur asynchronously.
1085 	 */
1086 	if (boot_cpu_has(X86_FEATURE_HYPERVISOR) ||
1087 	    !boot_cpu_has(X86_FEATURE_NONSTOP_TSC) ||
1088 	    !boot_cpu_has(X86_FEATURE_TSC_ADJUST) ||
1089 	    tsc_async_resets)
1090 		return;
1091 
1092 	cpuid(ART_CPUID_LEAF, &art_to_tsc_denominator,
1093 	      &art_to_tsc_numerator, unused, unused+1);
1094 
1095 	if (art_to_tsc_denominator < ART_MIN_DENOMINATOR)
1096 		return;
1097 
1098 	rdmsrl(MSR_IA32_TSC_ADJUST, art_to_tsc_offset);
1099 
1100 	/* Make this sticky over multiple CPU init calls */
1101 	setup_force_cpu_cap(X86_FEATURE_ART);
1102 }
1103 
1104 
1105 /* clocksource code */
1106 
tsc_resume(struct clocksource * cs)1107 static void tsc_resume(struct clocksource *cs)
1108 {
1109 	tsc_verify_tsc_adjust(true);
1110 }
1111 
1112 /*
1113  * We used to compare the TSC to the cycle_last value in the clocksource
1114  * structure to avoid a nasty time-warp. This can be observed in a
1115  * very small window right after one CPU updated cycle_last under
1116  * xtime/vsyscall_gtod lock and the other CPU reads a TSC value which
1117  * is smaller than the cycle_last reference value due to a TSC which
1118  * is slightly behind. This delta is nowhere else observable, but in
1119  * that case it results in a forward time jump in the range of hours
1120  * due to the unsigned delta calculation of the time keeping core
1121  * code, which is necessary to support wrapping clocksources like pm
1122  * timer.
1123  *
1124  * This sanity check is now done in the core timekeeping code.
1125  * checking the result of read_tsc() - cycle_last for being negative.
1126  * That works because CLOCKSOURCE_MASK(64) does not mask out any bit.
1127  */
read_tsc(struct clocksource * cs)1128 static u64 read_tsc(struct clocksource *cs)
1129 {
1130 	return (u64)rdtsc_ordered();
1131 }
1132 
tsc_cs_mark_unstable(struct clocksource * cs)1133 static void tsc_cs_mark_unstable(struct clocksource *cs)
1134 {
1135 	if (tsc_unstable)
1136 		return;
1137 
1138 	tsc_unstable = 1;
1139 	if (using_native_sched_clock())
1140 		clear_sched_clock_stable();
1141 	disable_sched_clock_irqtime();
1142 	pr_info("Marking TSC unstable due to clocksource watchdog\n");
1143 }
1144 
tsc_cs_tick_stable(struct clocksource * cs)1145 static void tsc_cs_tick_stable(struct clocksource *cs)
1146 {
1147 	if (tsc_unstable)
1148 		return;
1149 
1150 	if (using_native_sched_clock())
1151 		sched_clock_tick_stable();
1152 }
1153 
tsc_cs_enable(struct clocksource * cs)1154 static int tsc_cs_enable(struct clocksource *cs)
1155 {
1156 	vclocks_set_used(VDSO_CLOCKMODE_TSC);
1157 	return 0;
1158 }
1159 
1160 /*
1161  * .mask MUST be CLOCKSOURCE_MASK(64). See comment above read_tsc()
1162  */
1163 static struct clocksource clocksource_tsc_early = {
1164 	.name			= "tsc-early",
1165 	.rating			= 299,
1166 	.uncertainty_margin	= 32 * NSEC_PER_MSEC,
1167 	.read			= read_tsc,
1168 	.mask			= CLOCKSOURCE_MASK(64),
1169 	.flags			= CLOCK_SOURCE_IS_CONTINUOUS |
1170 				  CLOCK_SOURCE_MUST_VERIFY,
1171 	.vdso_clock_mode	= VDSO_CLOCKMODE_TSC,
1172 	.enable			= tsc_cs_enable,
1173 	.resume			= tsc_resume,
1174 	.mark_unstable		= tsc_cs_mark_unstable,
1175 	.tick_stable		= tsc_cs_tick_stable,
1176 	.list			= LIST_HEAD_INIT(clocksource_tsc_early.list),
1177 };
1178 
1179 /*
1180  * Must mark VALID_FOR_HRES early such that when we unregister tsc_early
1181  * this one will immediately take over. We will only register if TSC has
1182  * been found good.
1183  */
1184 static struct clocksource clocksource_tsc = {
1185 	.name			= "tsc",
1186 	.rating			= 300,
1187 	.read			= read_tsc,
1188 	.mask			= CLOCKSOURCE_MASK(64),
1189 	.flags			= CLOCK_SOURCE_IS_CONTINUOUS |
1190 				  CLOCK_SOURCE_VALID_FOR_HRES |
1191 				  CLOCK_SOURCE_MUST_VERIFY |
1192 				  CLOCK_SOURCE_VERIFY_PERCPU,
1193 	.vdso_clock_mode	= VDSO_CLOCKMODE_TSC,
1194 	.enable			= tsc_cs_enable,
1195 	.resume			= tsc_resume,
1196 	.mark_unstable		= tsc_cs_mark_unstable,
1197 	.tick_stable		= tsc_cs_tick_stable,
1198 	.list			= LIST_HEAD_INIT(clocksource_tsc.list),
1199 };
1200 
mark_tsc_unstable(char * reason)1201 void mark_tsc_unstable(char *reason)
1202 {
1203 	if (tsc_unstable)
1204 		return;
1205 
1206 	tsc_unstable = 1;
1207 	if (using_native_sched_clock())
1208 		clear_sched_clock_stable();
1209 	disable_sched_clock_irqtime();
1210 	pr_info("Marking TSC unstable due to %s\n", reason);
1211 
1212 	clocksource_mark_unstable(&clocksource_tsc_early);
1213 	clocksource_mark_unstable(&clocksource_tsc);
1214 }
1215 
1216 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
1217 
tsc_disable_clocksource_watchdog(void)1218 static void __init tsc_disable_clocksource_watchdog(void)
1219 {
1220 	clocksource_tsc_early.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
1221 	clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
1222 }
1223 
tsc_clocksource_watchdog_disabled(void)1224 bool tsc_clocksource_watchdog_disabled(void)
1225 {
1226 	return !(clocksource_tsc.flags & CLOCK_SOURCE_MUST_VERIFY) &&
1227 	       tsc_as_watchdog && !no_tsc_watchdog;
1228 }
1229 
check_system_tsc_reliable(void)1230 static void __init check_system_tsc_reliable(void)
1231 {
1232 #if defined(CONFIG_MGEODEGX1) || defined(CONFIG_MGEODE_LX) || defined(CONFIG_X86_GENERIC)
1233 	if (is_geode_lx()) {
1234 		/* RTSC counts during suspend */
1235 #define RTSC_SUSP 0x100
1236 		unsigned long res_low, res_high;
1237 
1238 		rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
1239 		/* Geode_LX - the OLPC CPU has a very reliable TSC */
1240 		if (res_low & RTSC_SUSP)
1241 			tsc_clocksource_reliable = 1;
1242 	}
1243 #endif
1244 	if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE))
1245 		tsc_clocksource_reliable = 1;
1246 
1247 	/*
1248 	 * Disable the clocksource watchdog when the system has:
1249 	 *  - TSC running at constant frequency
1250 	 *  - TSC which does not stop in C-States
1251 	 *  - the TSC_ADJUST register which allows to detect even minimal
1252 	 *    modifications
1253 	 *  - not more than two sockets. As the number of sockets cannot be
1254 	 *    evaluated at the early boot stage where this has to be
1255 	 *    invoked, check the number of online memory nodes as a
1256 	 *    fallback solution which is an reasonable estimate.
1257 	 */
1258 	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC) &&
1259 	    boot_cpu_has(X86_FEATURE_NONSTOP_TSC) &&
1260 	    boot_cpu_has(X86_FEATURE_TSC_ADJUST) &&
1261 	    nr_online_nodes <= 4)
1262 		tsc_disable_clocksource_watchdog();
1263 }
1264 
1265 /*
1266  * Make an educated guess if the TSC is trustworthy and synchronized
1267  * over all CPUs.
1268  */
unsynchronized_tsc(void)1269 int unsynchronized_tsc(void)
1270 {
1271 	if (!boot_cpu_has(X86_FEATURE_TSC) || tsc_unstable)
1272 		return 1;
1273 
1274 #ifdef CONFIG_SMP
1275 	if (apic_is_clustered_box())
1276 		return 1;
1277 #endif
1278 
1279 	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
1280 		return 0;
1281 
1282 	if (tsc_clocksource_reliable)
1283 		return 0;
1284 	/*
1285 	 * Intel systems are normally all synchronized.
1286 	 * Exceptions must mark TSC as unstable:
1287 	 */
1288 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
1289 		/* assume multi socket systems are not synchronized: */
1290 		if (num_possible_cpus() > 1)
1291 			return 1;
1292 	}
1293 
1294 	return 0;
1295 }
1296 
1297 /*
1298  * Convert ART to TSC given numerator/denominator found in detect_art()
1299  */
convert_art_to_tsc(u64 art)1300 struct system_counterval_t convert_art_to_tsc(u64 art)
1301 {
1302 	u64 tmp, res, rem;
1303 
1304 	rem = do_div(art, art_to_tsc_denominator);
1305 
1306 	res = art * art_to_tsc_numerator;
1307 	tmp = rem * art_to_tsc_numerator;
1308 
1309 	do_div(tmp, art_to_tsc_denominator);
1310 	res += tmp + art_to_tsc_offset;
1311 
1312 	return (struct system_counterval_t) {.cs = art_related_clocksource,
1313 			.cycles = res};
1314 }
1315 EXPORT_SYMBOL(convert_art_to_tsc);
1316 
1317 /**
1318  * convert_art_ns_to_tsc() - Convert ART in nanoseconds to TSC.
1319  * @art_ns: ART (Always Running Timer) in unit of nanoseconds
1320  *
1321  * PTM requires all timestamps to be in units of nanoseconds. When user
1322  * software requests a cross-timestamp, this function converts system timestamp
1323  * to TSC.
1324  *
1325  * This is valid when CPU feature flag X86_FEATURE_TSC_KNOWN_FREQ is set
1326  * indicating the tsc_khz is derived from CPUID[15H]. Drivers should check
1327  * that this flag is set before conversion to TSC is attempted.
1328  *
1329  * Return:
1330  * struct system_counterval_t - system counter value with the pointer to the
1331  *	corresponding clocksource
1332  *	@cycles:	System counter value
1333  *	@cs:		Clocksource corresponding to system counter value. Used
1334  *			by timekeeping code to verify comparability of two cycle
1335  *			values.
1336  */
1337 
convert_art_ns_to_tsc(u64 art_ns)1338 struct system_counterval_t convert_art_ns_to_tsc(u64 art_ns)
1339 {
1340 	u64 tmp, res, rem;
1341 
1342 	rem = do_div(art_ns, USEC_PER_SEC);
1343 
1344 	res = art_ns * tsc_khz;
1345 	tmp = rem * tsc_khz;
1346 
1347 	do_div(tmp, USEC_PER_SEC);
1348 	res += tmp;
1349 
1350 	return (struct system_counterval_t) { .cs = art_related_clocksource,
1351 					      .cycles = res};
1352 }
1353 EXPORT_SYMBOL(convert_art_ns_to_tsc);
1354 
1355 
1356 static void tsc_refine_calibration_work(struct work_struct *work);
1357 static DECLARE_DELAYED_WORK(tsc_irqwork, tsc_refine_calibration_work);
1358 /**
1359  * tsc_refine_calibration_work - Further refine tsc freq calibration
1360  * @work - ignored.
1361  *
1362  * This functions uses delayed work over a period of a
1363  * second to further refine the TSC freq value. Since this is
1364  * timer based, instead of loop based, we don't block the boot
1365  * process while this longer calibration is done.
1366  *
1367  * If there are any calibration anomalies (too many SMIs, etc),
1368  * or the refined calibration is off by 1% of the fast early
1369  * calibration, we throw out the new calibration and use the
1370  * early calibration.
1371  */
tsc_refine_calibration_work(struct work_struct * work)1372 static void tsc_refine_calibration_work(struct work_struct *work)
1373 {
1374 	static u64 tsc_start = ULLONG_MAX, ref_start;
1375 	static int hpet;
1376 	u64 tsc_stop, ref_stop, delta;
1377 	unsigned long freq;
1378 	int cpu;
1379 
1380 	/* Don't bother refining TSC on unstable systems */
1381 	if (tsc_unstable)
1382 		goto unreg;
1383 
1384 	/*
1385 	 * Since the work is started early in boot, we may be
1386 	 * delayed the first time we expire. So set the workqueue
1387 	 * again once we know timers are working.
1388 	 */
1389 	if (tsc_start == ULLONG_MAX) {
1390 restart:
1391 		/*
1392 		 * Only set hpet once, to avoid mixing hardware
1393 		 * if the hpet becomes enabled later.
1394 		 */
1395 		hpet = is_hpet_enabled();
1396 		tsc_start = tsc_read_refs(&ref_start, hpet);
1397 		schedule_delayed_work(&tsc_irqwork, HZ);
1398 		return;
1399 	}
1400 
1401 	tsc_stop = tsc_read_refs(&ref_stop, hpet);
1402 
1403 	/* hpet or pmtimer available ? */
1404 	if (ref_start == ref_stop)
1405 		goto out;
1406 
1407 	/* Check, whether the sampling was disturbed */
1408 	if (tsc_stop == ULLONG_MAX)
1409 		goto restart;
1410 
1411 	delta = tsc_stop - tsc_start;
1412 	delta *= 1000000LL;
1413 	if (hpet)
1414 		freq = calc_hpet_ref(delta, ref_start, ref_stop);
1415 	else
1416 		freq = calc_pmtimer_ref(delta, ref_start, ref_stop);
1417 
1418 	/* Will hit this only if tsc_force_recalibrate has been set */
1419 	if (boot_cpu_has(X86_FEATURE_TSC_KNOWN_FREQ)) {
1420 
1421 		/* Warn if the deviation exceeds 500 ppm */
1422 		if (abs(tsc_khz - freq) > (tsc_khz >> 11)) {
1423 			pr_warn("Warning: TSC freq calibrated by CPUID/MSR differs from what is calibrated by HW timer, please check with vendor!!\n");
1424 			pr_info("Previous calibrated TSC freq:\t %lu.%03lu MHz\n",
1425 				(unsigned long)tsc_khz / 1000,
1426 				(unsigned long)tsc_khz % 1000);
1427 		}
1428 
1429 		pr_info("TSC freq recalibrated by [%s]:\t %lu.%03lu MHz\n",
1430 			hpet ? "HPET" : "PM_TIMER",
1431 			(unsigned long)freq / 1000,
1432 			(unsigned long)freq % 1000);
1433 
1434 		return;
1435 	}
1436 
1437 	/* Make sure we're within 1% */
1438 	if (abs(tsc_khz - freq) > tsc_khz/100)
1439 		goto out;
1440 
1441 	tsc_khz = freq;
1442 	pr_info("Refined TSC clocksource calibration: %lu.%03lu MHz\n",
1443 		(unsigned long)tsc_khz / 1000,
1444 		(unsigned long)tsc_khz % 1000);
1445 
1446 	/* Inform the TSC deadline clockevent devices about the recalibration */
1447 	lapic_update_tsc_freq();
1448 
1449 	/* Update the sched_clock() rate to match the clocksource one */
1450 	for_each_possible_cpu(cpu)
1451 		set_cyc2ns_scale(tsc_khz, cpu, tsc_stop);
1452 
1453 out:
1454 	if (tsc_unstable)
1455 		goto unreg;
1456 
1457 	if (boot_cpu_has(X86_FEATURE_ART))
1458 		art_related_clocksource = &clocksource_tsc;
1459 	clocksource_register_khz(&clocksource_tsc, tsc_khz);
1460 unreg:
1461 	clocksource_unregister(&clocksource_tsc_early);
1462 }
1463 
1464 
init_tsc_clocksource(void)1465 static int __init init_tsc_clocksource(void)
1466 {
1467 	if (!boot_cpu_has(X86_FEATURE_TSC) || !tsc_khz)
1468 		return 0;
1469 
1470 	if (tsc_unstable) {
1471 		clocksource_unregister(&clocksource_tsc_early);
1472 		return 0;
1473 	}
1474 
1475 	if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC_S3))
1476 		clocksource_tsc.flags |= CLOCK_SOURCE_SUSPEND_NONSTOP;
1477 
1478 	/*
1479 	 * When TSC frequency is known (retrieved via MSR or CPUID), we skip
1480 	 * the refined calibration and directly register it as a clocksource.
1481 	 */
1482 	if (boot_cpu_has(X86_FEATURE_TSC_KNOWN_FREQ)) {
1483 		if (boot_cpu_has(X86_FEATURE_ART))
1484 			art_related_clocksource = &clocksource_tsc;
1485 		clocksource_register_khz(&clocksource_tsc, tsc_khz);
1486 		clocksource_unregister(&clocksource_tsc_early);
1487 
1488 		if (!tsc_force_recalibrate)
1489 			return 0;
1490 	}
1491 
1492 	schedule_delayed_work(&tsc_irqwork, 0);
1493 	return 0;
1494 }
1495 /*
1496  * We use device_initcall here, to ensure we run after the hpet
1497  * is fully initialized, which may occur at fs_initcall time.
1498  */
1499 device_initcall(init_tsc_clocksource);
1500 
determine_cpu_tsc_frequencies(bool early)1501 static bool __init determine_cpu_tsc_frequencies(bool early)
1502 {
1503 	/* Make sure that cpu and tsc are not already calibrated */
1504 	WARN_ON(cpu_khz || tsc_khz);
1505 
1506 	if (early) {
1507 		cpu_khz = x86_platform.calibrate_cpu();
1508 		if (tsc_early_khz)
1509 			tsc_khz = tsc_early_khz;
1510 		else
1511 			tsc_khz = x86_platform.calibrate_tsc();
1512 	} else {
1513 		/* We should not be here with non-native cpu calibration */
1514 		WARN_ON(x86_platform.calibrate_cpu != native_calibrate_cpu);
1515 		cpu_khz = pit_hpet_ptimer_calibrate_cpu();
1516 	}
1517 
1518 	/*
1519 	 * Trust non-zero tsc_khz as authoritative,
1520 	 * and use it to sanity check cpu_khz,
1521 	 * which will be off if system timer is off.
1522 	 */
1523 	if (tsc_khz == 0)
1524 		tsc_khz = cpu_khz;
1525 	else if (abs(cpu_khz - tsc_khz) * 10 > tsc_khz)
1526 		cpu_khz = tsc_khz;
1527 
1528 	if (tsc_khz == 0)
1529 		return false;
1530 
1531 	pr_info("Detected %lu.%03lu MHz processor\n",
1532 		(unsigned long)cpu_khz / KHZ,
1533 		(unsigned long)cpu_khz % KHZ);
1534 
1535 	if (cpu_khz != tsc_khz) {
1536 		pr_info("Detected %lu.%03lu MHz TSC",
1537 			(unsigned long)tsc_khz / KHZ,
1538 			(unsigned long)tsc_khz % KHZ);
1539 	}
1540 	return true;
1541 }
1542 
get_loops_per_jiffy(void)1543 static unsigned long __init get_loops_per_jiffy(void)
1544 {
1545 	u64 lpj = (u64)tsc_khz * KHZ;
1546 
1547 	do_div(lpj, HZ);
1548 	return lpj;
1549 }
1550 
tsc_enable_sched_clock(void)1551 static void __init tsc_enable_sched_clock(void)
1552 {
1553 	loops_per_jiffy = get_loops_per_jiffy();
1554 	use_tsc_delay();
1555 
1556 	/* Sanitize TSC ADJUST before cyc2ns gets initialized */
1557 	tsc_store_and_check_tsc_adjust(true);
1558 	cyc2ns_init_boot_cpu();
1559 	static_branch_enable(&__use_tsc);
1560 }
1561 
tsc_early_init(void)1562 void __init tsc_early_init(void)
1563 {
1564 	if (!boot_cpu_has(X86_FEATURE_TSC))
1565 		return;
1566 	/* Don't change UV TSC multi-chassis synchronization */
1567 	if (is_early_uv_system())
1568 		return;
1569 	if (!determine_cpu_tsc_frequencies(true))
1570 		return;
1571 	tsc_enable_sched_clock();
1572 }
1573 
tsc_init(void)1574 void __init tsc_init(void)
1575 {
1576 	if (!cpu_feature_enabled(X86_FEATURE_TSC)) {
1577 		setup_clear_cpu_cap(X86_FEATURE_TSC_DEADLINE_TIMER);
1578 		return;
1579 	}
1580 
1581 	/*
1582 	 * native_calibrate_cpu_early can only calibrate using methods that are
1583 	 * available early in boot.
1584 	 */
1585 	if (x86_platform.calibrate_cpu == native_calibrate_cpu_early)
1586 		x86_platform.calibrate_cpu = native_calibrate_cpu;
1587 
1588 	if (!tsc_khz) {
1589 		/* We failed to determine frequencies earlier, try again */
1590 		if (!determine_cpu_tsc_frequencies(false)) {
1591 			mark_tsc_unstable("could not calculate TSC khz");
1592 			setup_clear_cpu_cap(X86_FEATURE_TSC_DEADLINE_TIMER);
1593 			return;
1594 		}
1595 		tsc_enable_sched_clock();
1596 	}
1597 
1598 	cyc2ns_init_secondary_cpus();
1599 
1600 	if (!no_sched_irq_time)
1601 		enable_sched_clock_irqtime();
1602 
1603 	lpj_fine = get_loops_per_jiffy();
1604 
1605 	check_system_tsc_reliable();
1606 
1607 	if (unsynchronized_tsc()) {
1608 		mark_tsc_unstable("TSCs unsynchronized");
1609 		return;
1610 	}
1611 
1612 	if (tsc_clocksource_reliable || no_tsc_watchdog)
1613 		tsc_disable_clocksource_watchdog();
1614 
1615 	clocksource_register_khz(&clocksource_tsc_early, tsc_khz);
1616 	detect_art();
1617 }
1618 
1619 #ifdef CONFIG_SMP
1620 /*
1621  * Check whether existing calibration data can be reused.
1622  */
calibrate_delay_is_known(void)1623 unsigned long calibrate_delay_is_known(void)
1624 {
1625 	int sibling, cpu = smp_processor_id();
1626 	int constant_tsc = cpu_has(&cpu_data(cpu), X86_FEATURE_CONSTANT_TSC);
1627 	const struct cpumask *mask = topology_core_cpumask(cpu);
1628 
1629 	/*
1630 	 * If TSC has constant frequency and TSC is synchronized across
1631 	 * sockets then reuse CPU0 calibration.
1632 	 */
1633 	if (constant_tsc && !tsc_unstable)
1634 		return cpu_data(0).loops_per_jiffy;
1635 
1636 	/*
1637 	 * If TSC has constant frequency and TSC is not synchronized across
1638 	 * sockets and this is not the first CPU in the socket, then reuse
1639 	 * the calibration value of an already online CPU on that socket.
1640 	 *
1641 	 * This assumes that CONSTANT_TSC is consistent for all CPUs in a
1642 	 * socket.
1643 	 */
1644 	if (!constant_tsc || !mask)
1645 		return 0;
1646 
1647 	sibling = cpumask_any_but(mask, cpu);
1648 	if (sibling < nr_cpu_ids)
1649 		return cpu_data(sibling).loops_per_jiffy;
1650 	return 0;
1651 }
1652 #endif
1653