Linux Audio

Check our new training course

Loading...
v5.4
   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
  18#include <asm/hpet.h>
  19#include <asm/timer.h>
  20#include <asm/vgtod.h>
  21#include <asm/time.h>
  22#include <asm/delay.h>
  23#include <asm/hypervisor.h>
  24#include <asm/nmi.h>
  25#include <asm/x86_init.h>
  26#include <asm/geode.h>
  27#include <asm/apic.h>
  28#include <asm/intel-family.h>
  29#include <asm/i8259.h>
  30#include <asm/uv/uv.h>
  31
  32unsigned int __read_mostly cpu_khz;	/* TSC clocks / usec, not used here */
  33EXPORT_SYMBOL(cpu_khz);
  34
  35unsigned int __read_mostly tsc_khz;
  36EXPORT_SYMBOL(tsc_khz);
  37
  38#define KHZ	1000
  39
  40/*
  41 * TSC can be unstable due to cpufreq or due to unsynced TSCs
  42 */
  43static int __read_mostly tsc_unstable;
  44
 
 
 
 
 
  45static DEFINE_STATIC_KEY_FALSE(__use_tsc);
  46
  47int tsc_clocksource_reliable;
  48
  49static u32 art_to_tsc_numerator;
  50static u32 art_to_tsc_denominator;
  51static u64 art_to_tsc_offset;
  52struct clocksource *art_related_clocksource;
  53
  54struct cyc2ns {
  55	struct cyc2ns_data data[2];	/*  0 + 2*16 = 32 */
  56	seqcount_t	   seq;		/* 32 + 4    = 36 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  57
  58}; /* fits one cacheline */
 
 
 
 
  59
  60static DEFINE_PER_CPU_ALIGNED(struct cyc2ns, cyc2ns);
  61
  62__always_inline void cyc2ns_read_begin(struct cyc2ns_data *data)
  63{
  64	int seq, idx;
  65
  66	preempt_disable_notrace();
  67
  68	do {
  69		seq = this_cpu_read(cyc2ns.seq.sequence);
  70		idx = seq & 1;
 
 
 
 
 
  71
  72		data->cyc2ns_offset = this_cpu_read(cyc2ns.data[idx].cyc2ns_offset);
  73		data->cyc2ns_mul    = this_cpu_read(cyc2ns.data[idx].cyc2ns_mul);
  74		data->cyc2ns_shift  = this_cpu_read(cyc2ns.data[idx].cyc2ns_shift);
  75
  76	} while (unlikely(seq != this_cpu_read(cyc2ns.seq.sequence)));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  77}
  78
  79__always_inline void cyc2ns_read_end(void)
 
 
 
 
 
 
  80{
  81	preempt_enable_notrace();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  82}
  83
  84/*
  85 * Accelerators for sched_clock()
  86 * convert from cycles(64bits) => nanoseconds (64bits)
  87 *  basic equation:
  88 *              ns = cycles / (freq / ns_per_sec)
  89 *              ns = cycles * (ns_per_sec / freq)
  90 *              ns = cycles * (10^9 / (cpu_khz * 10^3))
  91 *              ns = cycles * (10^6 / cpu_khz)
  92 *
  93 *      Then we use scaling math (suggested by george@mvista.com) to get:
  94 *              ns = cycles * (10^6 * SC / cpu_khz) / SC
  95 *              ns = cycles * cyc2ns_scale / SC
  96 *
  97 *      And since SC is a constant power of two, we can convert the div
  98 *  into a shift. The larger SC is, the more accurate the conversion, but
  99 *  cyc2ns_scale needs to be a 32-bit value so that 32-bit multiplication
 100 *  (64-bit result) can be used.
 101 *
 102 *  We can use khz divisor instead of mhz to keep a better precision.
 103 *  (mathieu.desnoyers@polymtl.ca)
 104 *
 105 *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
 106 */
 107
 108static __always_inline unsigned long long cycles_2_ns(unsigned long long cyc)
 
 
 
 
 
 
 
 
 109{
 110	struct cyc2ns_data data;
 
 
 
 
 
 
 
 
 
 
 
 111	unsigned long long ns;
 112
 113	cyc2ns_read_begin(&data);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 114
 115	ns = data.cyc2ns_offset;
 116	ns += mul_u64_u32_shr(cyc, data.cyc2ns_mul, data.cyc2ns_shift);
 117
 118	cyc2ns_read_end();
 
 
 
 
 
 119
 120	return ns;
 121}
 122
 123static void __set_cyc2ns_scale(unsigned long khz, int cpu, unsigned long long tsc_now)
 124{
 125	unsigned long long ns_now;
 126	struct cyc2ns_data data;
 127	struct cyc2ns *c2n;
 128
 
 
 
 
 
 
 
 
 
 129	ns_now = cycles_2_ns(tsc_now);
 130
 131	/*
 132	 * Compute a new multiplier as per the above comment and ensure our
 133	 * time function is continuous; see the comment near struct
 134	 * cyc2ns_data.
 135	 */
 136	clocks_calc_mult_shift(&data.cyc2ns_mul, &data.cyc2ns_shift, khz,
 137			       NSEC_PER_MSEC, 0);
 138
 139	/*
 140	 * cyc2ns_shift is exported via arch_perf_update_userpage() where it is
 141	 * not expected to be greater than 31 due to the original published
 142	 * conversion algorithm shifting a 32-bit value (now specifies a 64-bit
 143	 * value) - refer perf_event_mmap_page documentation in perf_event.h.
 144	 */
 145	if (data.cyc2ns_shift == 32) {
 146		data.cyc2ns_shift = 31;
 147		data.cyc2ns_mul >>= 1;
 148	}
 149
 150	data.cyc2ns_offset = ns_now -
 151		mul_u64_u32_shr(tsc_now, data.cyc2ns_mul, data.cyc2ns_shift);
 152
 153	c2n = per_cpu_ptr(&cyc2ns, cpu);
 154
 155	raw_write_seqcount_latch(&c2n->seq);
 156	c2n->data[0] = data;
 157	raw_write_seqcount_latch(&c2n->seq);
 158	c2n->data[1] = data;
 159}
 160
 161static void set_cyc2ns_scale(unsigned long khz, int cpu, unsigned long long tsc_now)
 162{
 163	unsigned long flags;
 164
 165	local_irq_save(flags);
 166	sched_clock_idle_sleep_event();
 167
 168	if (khz)
 169		__set_cyc2ns_scale(khz, cpu, tsc_now);
 170
 171	sched_clock_idle_wakeup_event();
 172	local_irq_restore(flags);
 173}
 174
 175/*
 176 * Initialize cyc2ns for boot cpu
 177 */
 178static void __init cyc2ns_init_boot_cpu(void)
 179{
 180	struct cyc2ns *c2n = this_cpu_ptr(&cyc2ns);
 181
 182	seqcount_init(&c2n->seq);
 183	__set_cyc2ns_scale(tsc_khz, smp_processor_id(), rdtsc());
 184}
 185
 186/*
 187 * Secondary CPUs do not run through tsc_init(), so set up
 188 * all the scale factors for all CPUs, assuming the same
 189 * speed as the bootup CPU.
 190 */
 191static 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 */
 210u64 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 */
 235u64 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
 243unsigned long long sched_clock(void)
 244{
 245	return paravirt_sched_clock();
 246}
 247
 248bool using_native_sched_clock(void)
 249{
 250	return pv_ops.time.sched_clock == native_sched_clock;
 251}
 252#else
 253unsigned long long
 254sched_clock(void) __attribute__((alias("native_sched_clock")));
 255
 256bool using_native_sched_clock(void) { return true; }
 257#endif
 258
 259int check_tsc_unstable(void)
 260{
 261	return tsc_unstable;
 262}
 263EXPORT_SYMBOL_GPL(check_tsc_unstable);
 264
 265#ifdef CONFIG_X86_TSC
 266int __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 */
 276int __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
 285static int no_sched_irq_time;
 286static int no_tsc_watchdog;
 287
 288static int __init tsc_setup(char *str)
 289{
 290	if (!strcmp(str, "reliable"))
 291		tsc_clocksource_reliable = 1;
 292	if (!strncmp(str, "noirqtime", 9))
 293		no_sched_irq_time = 1;
 294	if (!strcmp(str, "unstable"))
 295		mark_tsc_unstable("boot parameter");
 296	if (!strcmp(str, "nowatchdog"))
 297		no_tsc_watchdog = 1;
 298	return 1;
 299}
 300
 301__setup("tsc=", tsc_setup);
 302
 303#define MAX_RETRIES		5
 304#define TSC_DEFAULT_THRESHOLD	0x20000
 305
 306/*
 307 * Read TSC and the reference counters. Take care of any disturbances
 308 */
 309static u64 tsc_read_refs(u64 *p, int hpet)
 310{
 311	u64 t1, t2;
 312	u64 thresh = tsc_khz ? tsc_khz >> 5 : TSC_DEFAULT_THRESHOLD;
 313	int i;
 314
 315	for (i = 0; i < MAX_RETRIES; i++) {
 316		t1 = get_cycles();
 317		if (hpet)
 318			*p = hpet_readl(HPET_COUNTER) & 0xFFFFFFFF;
 319		else
 320			*p = acpi_pm_read_early();
 321		t2 = get_cycles();
 322		if ((t2 - t1) < thresh)
 323			return t2;
 324	}
 325	return ULLONG_MAX;
 326}
 327
 328/*
 329 * Calculate the TSC frequency from HPET reference
 330 */
 331static unsigned long calc_hpet_ref(u64 deltatsc, u64 hpet1, u64 hpet2)
 332{
 333	u64 tmp;
 334
 335	if (hpet2 < hpet1)
 336		hpet2 += 0x100000000ULL;
 337	hpet2 -= hpet1;
 338	tmp = ((u64)hpet2 * hpet_readl(HPET_PERIOD));
 339	do_div(tmp, 1000000);
 340	deltatsc = div64_u64(deltatsc, tmp);
 341
 342	return (unsigned long) deltatsc;
 343}
 344
 345/*
 346 * Calculate the TSC frequency from PMTimer reference
 347 */
 348static unsigned long calc_pmtimer_ref(u64 deltatsc, u64 pm1, u64 pm2)
 349{
 350	u64 tmp;
 351
 352	if (!pm1 && !pm2)
 353		return ULONG_MAX;
 354
 355	if (pm2 < pm1)
 356		pm2 += (u64)ACPI_PM_OVRRUN;
 357	pm2 -= pm1;
 358	tmp = pm2 * 1000000000LL;
 359	do_div(tmp, PMTMR_TICKS_PER_SEC);
 360	do_div(deltatsc, tmp);
 361
 362	return (unsigned long) deltatsc;
 363}
 364
 365#define CAL_MS		10
 366#define CAL_LATCH	(PIT_TICK_RATE / (1000 / CAL_MS))
 367#define CAL_PIT_LOOPS	1000
 368
 369#define CAL2_MS		50
 370#define CAL2_LATCH	(PIT_TICK_RATE / (1000 / CAL2_MS))
 371#define CAL2_PIT_LOOPS	5000
 372
 373
 374/*
 375 * Try to calibrate the TSC against the Programmable
 376 * Interrupt Timer and return the frequency of the TSC
 377 * in kHz.
 378 *
 379 * Return ULONG_MAX on failure to calibrate.
 380 */
 381static unsigned long pit_calibrate_tsc(u32 latch, unsigned long ms, int loopmin)
 382{
 383	u64 tsc, t1, t2, delta;
 384	unsigned long tscmin, tscmax;
 385	int pitcnt;
 386
 387	if (!has_legacy_pic()) {
 388		/*
 389		 * Relies on tsc_early_delay_calibrate() to have given us semi
 390		 * usable udelay(), wait for the same 50ms we would have with
 391		 * the PIT loop below.
 392		 */
 393		udelay(10 * USEC_PER_MSEC);
 394		udelay(10 * USEC_PER_MSEC);
 395		udelay(10 * USEC_PER_MSEC);
 396		udelay(10 * USEC_PER_MSEC);
 397		udelay(10 * USEC_PER_MSEC);
 398		return ULONG_MAX;
 399	}
 400
 401	/* Set the Gate high, disable speaker */
 402	outb((inb(0x61) & ~0x02) | 0x01, 0x61);
 403
 404	/*
 405	 * Setup CTC channel 2* for mode 0, (interrupt on terminal
 406	 * count mode), binary count. Set the latch register to 50ms
 407	 * (LSB then MSB) to begin countdown.
 408	 */
 409	outb(0xb0, 0x43);
 410	outb(latch & 0xff, 0x42);
 411	outb(latch >> 8, 0x42);
 412
 413	tsc = t1 = t2 = get_cycles();
 414
 415	pitcnt = 0;
 416	tscmax = 0;
 417	tscmin = ULONG_MAX;
 418	while ((inb(0x61) & 0x20) == 0) {
 419		t2 = get_cycles();
 420		delta = t2 - tsc;
 421		tsc = t2;
 422		if ((unsigned long) delta < tscmin)
 423			tscmin = (unsigned int) delta;
 424		if ((unsigned long) delta > tscmax)
 425			tscmax = (unsigned int) delta;
 426		pitcnt++;
 427	}
 428
 429	/*
 430	 * Sanity checks:
 431	 *
 432	 * If we were not able to read the PIT more than loopmin
 433	 * times, then we have been hit by a massive SMI
 434	 *
 435	 * If the maximum is 10 times larger than the minimum,
 436	 * then we got hit by an SMI as well.
 437	 */
 438	if (pitcnt < loopmin || tscmax > 10 * tscmin)
 439		return ULONG_MAX;
 440
 441	/* Calculate the PIT value */
 442	delta = t2 - t1;
 443	do_div(delta, ms);
 444	return delta;
 445}
 446
 447/*
 448 * This reads the current MSB of the PIT counter, and
 449 * checks if we are running on sufficiently fast and
 450 * non-virtualized hardware.
 451 *
 452 * Our expectations are:
 453 *
 454 *  - the PIT is running at roughly 1.19MHz
 455 *
 456 *  - each IO is going to take about 1us on real hardware,
 457 *    but we allow it to be much faster (by a factor of 10) or
 458 *    _slightly_ slower (ie we allow up to a 2us read+counter
 459 *    update - anything else implies a unacceptably slow CPU
 460 *    or PIT for the fast calibration to work.
 461 *
 462 *  - with 256 PIT ticks to read the value, we have 214us to
 463 *    see the same MSB (and overhead like doing a single TSC
 464 *    read per MSB value etc).
 465 *
 466 *  - We're doing 2 reads per loop (LSB, MSB), and we expect
 467 *    them each to take about a microsecond on real hardware.
 468 *    So we expect a count value of around 100. But we'll be
 469 *    generous, and accept anything over 50.
 470 *
 471 *  - if the PIT is stuck, and we see *many* more reads, we
 472 *    return early (and the next caller of pit_expect_msb()
 473 *    then consider it a failure when they don't see the
 474 *    next expected value).
 475 *
 476 * These expectations mean that we know that we have seen the
 477 * transition from one expected value to another with a fairly
 478 * high accuracy, and we didn't miss any events. We can thus
 479 * use the TSC value at the transitions to calculate a pretty
 480 * good value for the TSC frequencty.
 481 */
 482static inline int pit_verify_msb(unsigned char val)
 483{
 484	/* Ignore LSB */
 485	inb(0x42);
 486	return inb(0x42) == val;
 487}
 488
 489static inline int pit_expect_msb(unsigned char val, u64 *tscp, unsigned long *deltap)
 490{
 491	int count;
 492	u64 tsc = 0, prev_tsc = 0;
 493
 494	for (count = 0; count < 50000; count++) {
 495		if (!pit_verify_msb(val))
 496			break;
 497		prev_tsc = tsc;
 498		tsc = get_cycles();
 499	}
 500	*deltap = get_cycles() - prev_tsc;
 501	*tscp = tsc;
 502
 503	/*
 504	 * We require _some_ success, but the quality control
 505	 * will be based on the error terms on the TSC values.
 506	 */
 507	return count > 5;
 508}
 509
 510/*
 511 * How many MSB values do we want to see? We aim for
 512 * a maximum error rate of 500ppm (in practice the
 513 * real error is much smaller), but refuse to spend
 514 * more than 50ms on it.
 515 */
 516#define MAX_QUICK_PIT_MS 50
 517#define MAX_QUICK_PIT_ITERATIONS (MAX_QUICK_PIT_MS * PIT_TICK_RATE / 1000 / 256)
 518
 519static unsigned long quick_pit_calibrate(void)
 520{
 521	int i;
 522	u64 tsc, delta;
 523	unsigned long d1, d2;
 524
 525	if (!has_legacy_pic())
 526		return 0;
 527
 528	/* Set the Gate high, disable speaker */
 529	outb((inb(0x61) & ~0x02) | 0x01, 0x61);
 530
 531	/*
 532	 * Counter 2, mode 0 (one-shot), binary count
 533	 *
 534	 * NOTE! Mode 2 decrements by two (and then the
 535	 * output is flipped each time, giving the same
 536	 * final output frequency as a decrement-by-one),
 537	 * so mode 0 is much better when looking at the
 538	 * individual counts.
 539	 */
 540	outb(0xb0, 0x43);
 541
 542	/* Start at 0xffff */
 543	outb(0xff, 0x42);
 544	outb(0xff, 0x42);
 545
 546	/*
 547	 * The PIT starts counting at the next edge, so we
 548	 * need to delay for a microsecond. The easiest way
 549	 * to do that is to just read back the 16-bit counter
 550	 * once from the PIT.
 551	 */
 552	pit_verify_msb(0);
 553
 554	if (pit_expect_msb(0xff, &tsc, &d1)) {
 555		for (i = 1; i <= MAX_QUICK_PIT_ITERATIONS; i++) {
 556			if (!pit_expect_msb(0xff-i, &delta, &d2))
 557				break;
 558
 559			delta -= tsc;
 560
 561			/*
 562			 * Extrapolate the error and fail fast if the error will
 563			 * never be below 500 ppm.
 564			 */
 565			if (i == 1 &&
 566			    d1 + d2 >= (delta * MAX_QUICK_PIT_ITERATIONS) >> 11)
 567				return 0;
 568
 569			/*
 570			 * Iterate until the error is less than 500 ppm
 571			 */
 572			if (d1+d2 >= delta >> 11)
 573				continue;
 574
 575			/*
 576			 * Check the PIT one more time to verify that
 577			 * all TSC reads were stable wrt the PIT.
 578			 *
 579			 * This also guarantees serialization of the
 580			 * last cycle read ('d2') in pit_expect_msb.
 581			 */
 582			if (!pit_verify_msb(0xfe - i))
 583				break;
 584			goto success;
 585		}
 586	}
 587	pr_info("Fast TSC calibration failed\n");
 588	return 0;
 589
 590success:
 591	/*
 592	 * Ok, if we get here, then we've seen the
 593	 * MSB of the PIT decrement 'i' times, and the
 594	 * error has shrunk to less than 500 ppm.
 595	 *
 596	 * As a result, we can depend on there not being
 597	 * any odd delays anywhere, and the TSC reads are
 598	 * reliable (within the error).
 599	 *
 600	 * kHz = ticks / time-in-seconds / 1000;
 601	 * kHz = (t2 - t1) / (I * 256 / PIT_TICK_RATE) / 1000
 602	 * kHz = ((t2 - t1) * PIT_TICK_RATE) / (I * 256 * 1000)
 603	 */
 604	delta *= PIT_TICK_RATE;
 605	do_div(delta, i*256*1000);
 606	pr_info("Fast TSC calibration using PIT\n");
 607	return delta;
 608}
 609
 610/**
 611 * native_calibrate_tsc
 612 * Determine TSC frequency via CPUID, else return 0.
 613 */
 614unsigned long native_calibrate_tsc(void)
 615{
 616	unsigned int eax_denominator, ebx_numerator, ecx_hz, edx;
 617	unsigned int crystal_khz;
 618
 619	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
 620		return 0;
 621
 622	if (boot_cpu_data.cpuid_level < 0x15)
 623		return 0;
 624
 625	eax_denominator = ebx_numerator = ecx_hz = edx = 0;
 626
 627	/* CPUID 15H TSC/Crystal ratio, plus optionally Crystal Hz */
 628	cpuid(0x15, &eax_denominator, &ebx_numerator, &ecx_hz, &edx);
 629
 630	if (ebx_numerator == 0 || eax_denominator == 0)
 631		return 0;
 632
 633	crystal_khz = ecx_hz / 1000;
 634
 635	/*
 636	 * Denverton SoCs don't report crystal clock, and also don't support
 637	 * CPUID.0x16 for the calculation below, so hardcode the 25MHz crystal
 638	 * clock.
 639	 */
 640	if (crystal_khz == 0 &&
 641			boot_cpu_data.x86_model == INTEL_FAM6_ATOM_GOLDMONT_D)
 642		crystal_khz = 25000;
 
 
 
 
 
 
 
 
 
 643
 644	/*
 645	 * TSC frequency reported directly by CPUID is a "hardware reported"
 646	 * frequency and is the most accurate one so far we have. This
 647	 * is considered a known frequency.
 648	 */
 649	if (crystal_khz != 0)
 650		setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ);
 651
 652	/*
 653	 * Some Intel SoCs like Skylake and Kabylake don't report the crystal
 654	 * clock, but we can easily calculate it to a high degree of accuracy
 655	 * by considering the crystal ratio and the CPU speed.
 656	 */
 657	if (crystal_khz == 0 && boot_cpu_data.cpuid_level >= 0x16) {
 658		unsigned int eax_base_mhz, ebx, ecx, edx;
 659
 660		cpuid(0x16, &eax_base_mhz, &ebx, &ecx, &edx);
 661		crystal_khz = eax_base_mhz * 1000 *
 662			eax_denominator / ebx_numerator;
 663	}
 664
 665	if (crystal_khz == 0)
 666		return 0;
 667
 668	/*
 669	 * For Atom SoCs TSC is the only reliable clocksource.
 670	 * Mark TSC reliable so no watchdog on it.
 671	 */
 672	if (boot_cpu_data.x86_model == INTEL_FAM6_ATOM_GOLDMONT)
 673		setup_force_cpu_cap(X86_FEATURE_TSC_RELIABLE);
 674
 675#ifdef CONFIG_X86_LOCAL_APIC
 676	/*
 677	 * The local APIC appears to be fed by the core crystal clock
 678	 * (which sounds entirely sensible). We can set the global
 679	 * lapic_timer_period here to avoid having to calibrate the APIC
 680	 * timer later.
 681	 */
 682	lapic_timer_period = crystal_khz * 1000 / HZ;
 683#endif
 684
 685	return crystal_khz * ebx_numerator / eax_denominator;
 686}
 687
 688static unsigned long cpu_khz_from_cpuid(void)
 689{
 690	unsigned int eax_base_mhz, ebx_max_mhz, ecx_bus_mhz, edx;
 691
 692	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
 693		return 0;
 694
 695	if (boot_cpu_data.cpuid_level < 0x16)
 696		return 0;
 697
 698	eax_base_mhz = ebx_max_mhz = ecx_bus_mhz = edx = 0;
 699
 700	cpuid(0x16, &eax_base_mhz, &ebx_max_mhz, &ecx_bus_mhz, &edx);
 701
 702	return eax_base_mhz * 1000;
 703}
 704
 705/*
 706 * calibrate cpu using pit, hpet, and ptimer methods. They are available
 707 * later in boot after acpi is initialized.
 708 */
 709static unsigned long pit_hpet_ptimer_calibrate_cpu(void)
 710{
 711	u64 tsc1, tsc2, delta, ref1, ref2;
 712	unsigned long tsc_pit_min = ULONG_MAX, tsc_ref_min = ULONG_MAX;
 713	unsigned long flags, latch, ms;
 714	int hpet = is_hpet_enabled(), i, loopmin;
 715
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 716	/*
 717	 * Run 5 calibration loops to get the lowest frequency value
 718	 * (the best estimate). We use two different calibration modes
 719	 * here:
 720	 *
 721	 * 1) PIT loop. We set the PIT Channel 2 to oneshot mode and
 722	 * load a timeout of 50ms. We read the time right after we
 723	 * started the timer and wait until the PIT count down reaches
 724	 * zero. In each wait loop iteration we read the TSC and check
 725	 * the delta to the previous read. We keep track of the min
 726	 * and max values of that delta. The delta is mostly defined
 727	 * by the IO time of the PIT access, so we can detect when
 728	 * any disturbance happened between the two reads. If the
 729	 * maximum time is significantly larger than the minimum time,
 730	 * then we discard the result and have another try.
 731	 *
 732	 * 2) Reference counter. If available we use the HPET or the
 733	 * PMTIMER as a reference to check the sanity of that value.
 734	 * We use separate TSC readouts and check inside of the
 735	 * reference read for any possible disturbance. We dicard
 736	 * disturbed values here as well. We do that around the PIT
 737	 * calibration delay loop as we have to wait for a certain
 738	 * amount of time anyway.
 739	 */
 740
 741	/* Preset PIT loop values */
 742	latch = CAL_LATCH;
 743	ms = CAL_MS;
 744	loopmin = CAL_PIT_LOOPS;
 745
 746	for (i = 0; i < 3; i++) {
 747		unsigned long tsc_pit_khz;
 748
 749		/*
 750		 * Read the start value and the reference count of
 751		 * hpet/pmtimer when available. Then do the PIT
 752		 * calibration, which will take at least 50ms, and
 753		 * read the end value.
 754		 */
 755		local_irq_save(flags);
 756		tsc1 = tsc_read_refs(&ref1, hpet);
 757		tsc_pit_khz = pit_calibrate_tsc(latch, ms, loopmin);
 758		tsc2 = tsc_read_refs(&ref2, hpet);
 759		local_irq_restore(flags);
 760
 761		/* Pick the lowest PIT TSC calibration so far */
 762		tsc_pit_min = min(tsc_pit_min, tsc_pit_khz);
 763
 764		/* hpet or pmtimer available ? */
 765		if (ref1 == ref2)
 766			continue;
 767
 768		/* Check, whether the sampling was disturbed */
 769		if (tsc1 == ULLONG_MAX || tsc2 == ULLONG_MAX)
 770			continue;
 771
 772		tsc2 = (tsc2 - tsc1) * 1000000LL;
 773		if (hpet)
 774			tsc2 = calc_hpet_ref(tsc2, ref1, ref2);
 775		else
 776			tsc2 = calc_pmtimer_ref(tsc2, ref1, ref2);
 777
 778		tsc_ref_min = min(tsc_ref_min, (unsigned long) tsc2);
 779
 780		/* Check the reference deviation */
 781		delta = ((u64) tsc_pit_min) * 100;
 782		do_div(delta, tsc_ref_min);
 783
 784		/*
 785		 * If both calibration results are inside a 10% window
 786		 * then we can be sure, that the calibration
 787		 * succeeded. We break out of the loop right away. We
 788		 * use the reference value, as it is more precise.
 789		 */
 790		if (delta >= 90 && delta <= 110) {
 791			pr_info("PIT calibration matches %s. %d loops\n",
 792				hpet ? "HPET" : "PMTIMER", i + 1);
 793			return tsc_ref_min;
 794		}
 795
 796		/*
 797		 * Check whether PIT failed more than once. This
 798		 * happens in virtualized environments. We need to
 799		 * give the virtual PC a slightly longer timeframe for
 800		 * the HPET/PMTIMER to make the result precise.
 801		 */
 802		if (i == 1 && tsc_pit_min == ULONG_MAX) {
 803			latch = CAL2_LATCH;
 804			ms = CAL2_MS;
 805			loopmin = CAL2_PIT_LOOPS;
 806		}
 807	}
 808
 809	/*
 810	 * Now check the results.
 811	 */
 812	if (tsc_pit_min == ULONG_MAX) {
 813		/* PIT gave no useful value */
 814		pr_warn("Unable to calibrate against PIT\n");
 815
 816		/* We don't have an alternative source, disable TSC */
 817		if (!hpet && !ref1 && !ref2) {
 818			pr_notice("No reference (HPET/PMTIMER) available\n");
 819			return 0;
 820		}
 821
 822		/* The alternative source failed as well, disable TSC */
 823		if (tsc_ref_min == ULONG_MAX) {
 824			pr_warn("HPET/PMTIMER calibration failed\n");
 825			return 0;
 826		}
 827
 828		/* Use the alternative source */
 829		pr_info("using %s reference calibration\n",
 830			hpet ? "HPET" : "PMTIMER");
 831
 832		return tsc_ref_min;
 833	}
 834
 835	/* We don't have an alternative source, use the PIT calibration value */
 836	if (!hpet && !ref1 && !ref2) {
 837		pr_info("Using PIT calibration value\n");
 838		return tsc_pit_min;
 839	}
 840
 841	/* The alternative source failed, use the PIT calibration value */
 842	if (tsc_ref_min == ULONG_MAX) {
 843		pr_warn("HPET/PMTIMER calibration failed. Using PIT calibration.\n");
 844		return tsc_pit_min;
 845	}
 846
 847	/*
 848	 * The calibration values differ too much. In doubt, we use
 849	 * the PIT value as we know that there are PMTIMERs around
 850	 * running at double speed. At least we let the user know:
 851	 */
 852	pr_warn("PIT calibration deviates from %s: %lu %lu\n",
 853		hpet ? "HPET" : "PMTIMER", tsc_pit_min, tsc_ref_min);
 854	pr_info("Using PIT calibration value\n");
 855	return tsc_pit_min;
 856}
 857
 858/**
 859 * native_calibrate_cpu_early - can calibrate the cpu early in boot
 860 */
 861unsigned long native_calibrate_cpu_early(void)
 862{
 863	unsigned long flags, fast_calibrate = cpu_khz_from_cpuid();
 864
 865	if (!fast_calibrate)
 866		fast_calibrate = cpu_khz_from_msr();
 867	if (!fast_calibrate) {
 868		local_irq_save(flags);
 869		fast_calibrate = quick_pit_calibrate();
 870		local_irq_restore(flags);
 871	}
 872	return fast_calibrate;
 873}
 874
 875
 876/**
 877 * native_calibrate_cpu - calibrate the cpu
 878 */
 879static unsigned long native_calibrate_cpu(void)
 880{
 881	unsigned long tsc_freq = native_calibrate_cpu_early();
 882
 883	if (!tsc_freq)
 884		tsc_freq = pit_hpet_ptimer_calibrate_cpu();
 885
 886	return tsc_freq;
 887}
 888
 889void recalibrate_cpu_khz(void)
 890{
 891#ifndef CONFIG_SMP
 892	unsigned long cpu_khz_old = cpu_khz;
 893
 894	if (!boot_cpu_has(X86_FEATURE_TSC))
 895		return;
 896
 897	cpu_khz = x86_platform.calibrate_cpu();
 898	tsc_khz = x86_platform.calibrate_tsc();
 899	if (tsc_khz == 0)
 900		tsc_khz = cpu_khz;
 901	else if (abs(cpu_khz - tsc_khz) * 10 > tsc_khz)
 902		cpu_khz = tsc_khz;
 903	cpu_data(0).loops_per_jiffy = cpufreq_scale(cpu_data(0).loops_per_jiffy,
 904						    cpu_khz_old, cpu_khz);
 
 
 
 
 905#endif
 906}
 907
 908EXPORT_SYMBOL(recalibrate_cpu_khz);
 909
 910
 911static unsigned long long cyc2ns_suspend;
 912
 913void tsc_save_sched_clock_state(void)
 914{
 915	if (!sched_clock_stable())
 916		return;
 917
 918	cyc2ns_suspend = sched_clock();
 919}
 920
 921/*
 922 * Even on processors with invariant TSC, TSC gets reset in some the
 923 * ACPI system sleep states. And in some systems BIOS seem to reinit TSC to
 924 * arbitrary value (still sync'd across cpu's) during resume from such sleep
 925 * states. To cope up with this, recompute the cyc2ns_offset for each cpu so
 926 * that sched_clock() continues from the point where it was left off during
 927 * suspend.
 928 */
 929void tsc_restore_sched_clock_state(void)
 930{
 931	unsigned long long offset;
 932	unsigned long flags;
 933	int cpu;
 934
 935	if (!sched_clock_stable())
 936		return;
 937
 938	local_irq_save(flags);
 939
 940	/*
 941	 * We're coming out of suspend, there's no concurrency yet; don't
 942	 * bother being nice about the RCU stuff, just write to both
 943	 * data fields.
 944	 */
 945
 946	this_cpu_write(cyc2ns.data[0].cyc2ns_offset, 0);
 947	this_cpu_write(cyc2ns.data[1].cyc2ns_offset, 0);
 948
 949	offset = cyc2ns_suspend - sched_clock();
 950
 951	for_each_possible_cpu(cpu) {
 952		per_cpu(cyc2ns.data[0].cyc2ns_offset, cpu) = offset;
 953		per_cpu(cyc2ns.data[1].cyc2ns_offset, cpu) = offset;
 954	}
 955
 956	local_irq_restore(flags);
 957}
 958
 959#ifdef CONFIG_CPU_FREQ
 960/*
 961 * Frequency scaling support. Adjust the TSC based timer when the CPU frequency
 962 * changes.
 963 *
 964 * NOTE: On SMP the situation is not fixable in general, so simply mark the TSC
 965 * as unstable and give up in those cases.
 
 966 *
 967 * Should fix up last_tsc too. Currently gettimeofday in the
 968 * first tick after the change will be slightly wrong.
 969 */
 970
 971static unsigned int  ref_freq;
 972static unsigned long loops_per_jiffy_ref;
 973static unsigned long tsc_khz_ref;
 974
 975static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
 976				void *data)
 977{
 978	struct cpufreq_freqs *freq = data;
 
 979
 980	if (num_online_cpus() > 1) {
 981		mark_tsc_unstable("cpufreq changes on SMP");
 982		return 0;
 983	}
 
 984
 985	if (!ref_freq) {
 986		ref_freq = freq->old;
 987		loops_per_jiffy_ref = boot_cpu_data.loops_per_jiffy;
 988		tsc_khz_ref = tsc_khz;
 989	}
 990
 991	if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
 992	    (val == CPUFREQ_POSTCHANGE && freq->old > freq->new)) {
 993		boot_cpu_data.loops_per_jiffy =
 994			cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
 995
 996		tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
 997		if (!(freq->flags & CPUFREQ_CONST_LOOPS))
 998			mark_tsc_unstable("cpufreq changes");
 999
1000		set_cyc2ns_scale(tsc_khz, freq->policy->cpu, rdtsc());
1001	}
1002
1003	return 0;
1004}
1005
1006static struct notifier_block time_cpufreq_notifier_block = {
1007	.notifier_call  = time_cpufreq_notifier
1008};
1009
1010static int __init cpufreq_register_tsc_scaling(void)
1011{
1012	if (!boot_cpu_has(X86_FEATURE_TSC))
1013		return 0;
1014	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
1015		return 0;
1016	cpufreq_register_notifier(&time_cpufreq_notifier_block,
1017				CPUFREQ_TRANSITION_NOTIFIER);
1018	return 0;
1019}
1020
1021core_initcall(cpufreq_register_tsc_scaling);
1022
1023#endif /* CONFIG_CPU_FREQ */
1024
1025#define ART_CPUID_LEAF (0x15)
1026#define ART_MIN_DENOMINATOR (1)
1027
1028
1029/*
1030 * If ART is present detect the numerator:denominator to convert to TSC
1031 */
1032static void __init detect_art(void)
1033{
1034	unsigned int unused[2];
1035
1036	if (boot_cpu_data.cpuid_level < ART_CPUID_LEAF)
1037		return;
1038
1039	/*
1040	 * Don't enable ART in a VM, non-stop TSC and TSC_ADJUST required,
1041	 * and the TSC counter resets must not occur asynchronously.
1042	 */
1043	if (boot_cpu_has(X86_FEATURE_HYPERVISOR) ||
1044	    !boot_cpu_has(X86_FEATURE_NONSTOP_TSC) ||
1045	    !boot_cpu_has(X86_FEATURE_TSC_ADJUST) ||
1046	    tsc_async_resets)
1047		return;
1048
1049	cpuid(ART_CPUID_LEAF, &art_to_tsc_denominator,
1050	      &art_to_tsc_numerator, unused, unused+1);
1051
1052	if (art_to_tsc_denominator < ART_MIN_DENOMINATOR)
1053		return;
1054
1055	rdmsrl(MSR_IA32_TSC_ADJUST, art_to_tsc_offset);
1056
1057	/* Make this sticky over multiple CPU init calls */
1058	setup_force_cpu_cap(X86_FEATURE_ART);
1059}
1060
1061
1062/* clocksource code */
1063
 
 
1064static void tsc_resume(struct clocksource *cs)
1065{
1066	tsc_verify_tsc_adjust(true);
1067}
1068
1069/*
1070 * We used to compare the TSC to the cycle_last value in the clocksource
1071 * structure to avoid a nasty time-warp. This can be observed in a
1072 * very small window right after one CPU updated cycle_last under
1073 * xtime/vsyscall_gtod lock and the other CPU reads a TSC value which
1074 * is smaller than the cycle_last reference value due to a TSC which
1075 * is slighty behind. This delta is nowhere else observable, but in
1076 * that case it results in a forward time jump in the range of hours
1077 * due to the unsigned delta calculation of the time keeping core
1078 * code, which is necessary to support wrapping clocksources like pm
1079 * timer.
1080 *
1081 * This sanity check is now done in the core timekeeping code.
1082 * checking the result of read_tsc() - cycle_last for being negative.
1083 * That works because CLOCKSOURCE_MASK(64) does not mask out any bit.
1084 */
1085static u64 read_tsc(struct clocksource *cs)
1086{
1087	return (u64)rdtsc_ordered();
1088}
1089
1090static void tsc_cs_mark_unstable(struct clocksource *cs)
1091{
1092	if (tsc_unstable)
1093		return;
1094
1095	tsc_unstable = 1;
1096	if (using_native_sched_clock())
1097		clear_sched_clock_stable();
1098	disable_sched_clock_irqtime();
1099	pr_info("Marking TSC unstable due to clocksource watchdog\n");
1100}
1101
1102static void tsc_cs_tick_stable(struct clocksource *cs)
1103{
1104	if (tsc_unstable)
1105		return;
1106
1107	if (using_native_sched_clock())
1108		sched_clock_tick_stable();
1109}
1110
1111/*
1112 * .mask MUST be CLOCKSOURCE_MASK(64). See comment above read_tsc()
1113 */
1114static struct clocksource clocksource_tsc_early = {
1115	.name                   = "tsc-early",
1116	.rating                 = 299,
1117	.read                   = read_tsc,
1118	.mask                   = CLOCKSOURCE_MASK(64),
1119	.flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
1120				  CLOCK_SOURCE_MUST_VERIFY,
1121	.archdata               = { .vclock_mode = VCLOCK_TSC },
1122	.resume			= tsc_resume,
1123	.mark_unstable		= tsc_cs_mark_unstable,
1124	.tick_stable		= tsc_cs_tick_stable,
1125	.list			= LIST_HEAD_INIT(clocksource_tsc_early.list),
1126};
1127
1128/*
1129 * Must mark VALID_FOR_HRES early such that when we unregister tsc_early
1130 * this one will immediately take over. We will only register if TSC has
1131 * been found good.
1132 */
1133static struct clocksource clocksource_tsc = {
1134	.name                   = "tsc",
1135	.rating                 = 300,
1136	.read                   = read_tsc,
1137	.mask                   = CLOCKSOURCE_MASK(64),
1138	.flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
1139				  CLOCK_SOURCE_VALID_FOR_HRES |
1140				  CLOCK_SOURCE_MUST_VERIFY,
1141	.archdata               = { .vclock_mode = VCLOCK_TSC },
1142	.resume			= tsc_resume,
1143	.mark_unstable		= tsc_cs_mark_unstable,
1144	.tick_stable		= tsc_cs_tick_stable,
1145	.list			= LIST_HEAD_INIT(clocksource_tsc.list),
1146};
1147
1148void mark_tsc_unstable(char *reason)
1149{
1150	if (tsc_unstable)
1151		return;
1152
1153	tsc_unstable = 1;
1154	if (using_native_sched_clock())
1155		clear_sched_clock_stable();
1156	disable_sched_clock_irqtime();
1157	pr_info("Marking TSC unstable due to %s\n", reason);
1158
1159	clocksource_mark_unstable(&clocksource_tsc_early);
1160	clocksource_mark_unstable(&clocksource_tsc);
 
 
 
 
 
1161}
1162
1163EXPORT_SYMBOL_GPL(mark_tsc_unstable);
1164
1165static void __init check_system_tsc_reliable(void)
1166{
1167#if defined(CONFIG_MGEODEGX1) || defined(CONFIG_MGEODE_LX) || defined(CONFIG_X86_GENERIC)
1168	if (is_geode_lx()) {
1169		/* RTSC counts during suspend */
1170#define RTSC_SUSP 0x100
1171		unsigned long res_low, res_high;
1172
1173		rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
1174		/* Geode_LX - the OLPC CPU has a very reliable TSC */
1175		if (res_low & RTSC_SUSP)
1176			tsc_clocksource_reliable = 1;
1177	}
1178#endif
1179	if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE))
1180		tsc_clocksource_reliable = 1;
1181}
1182
1183/*
1184 * Make an educated guess if the TSC is trustworthy and synchronized
1185 * over all CPUs.
1186 */
1187int unsynchronized_tsc(void)
1188{
1189	if (!boot_cpu_has(X86_FEATURE_TSC) || tsc_unstable)
1190		return 1;
1191
1192#ifdef CONFIG_SMP
1193	if (apic_is_clustered_box())
1194		return 1;
1195#endif
1196
1197	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
1198		return 0;
1199
1200	if (tsc_clocksource_reliable)
1201		return 0;
1202	/*
1203	 * Intel systems are normally all synchronized.
1204	 * Exceptions must mark TSC as unstable:
1205	 */
1206	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
1207		/* assume multi socket systems are not synchronized: */
1208		if (num_possible_cpus() > 1)
1209			return 1;
1210	}
1211
1212	return 0;
1213}
1214
1215/*
1216 * Convert ART to TSC given numerator/denominator found in detect_art()
1217 */
1218struct system_counterval_t convert_art_to_tsc(u64 art)
1219{
1220	u64 tmp, res, rem;
1221
1222	rem = do_div(art, art_to_tsc_denominator);
1223
1224	res = art * art_to_tsc_numerator;
1225	tmp = rem * art_to_tsc_numerator;
1226
1227	do_div(tmp, art_to_tsc_denominator);
1228	res += tmp + art_to_tsc_offset;
1229
1230	return (struct system_counterval_t) {.cs = art_related_clocksource,
1231			.cycles = res};
1232}
1233EXPORT_SYMBOL(convert_art_to_tsc);
1234
1235/**
1236 * convert_art_ns_to_tsc() - Convert ART in nanoseconds to TSC.
1237 * @art_ns: ART (Always Running Timer) in unit of nanoseconds
1238 *
1239 * PTM requires all timestamps to be in units of nanoseconds. When user
1240 * software requests a cross-timestamp, this function converts system timestamp
1241 * to TSC.
1242 *
1243 * This is valid when CPU feature flag X86_FEATURE_TSC_KNOWN_FREQ is set
1244 * indicating the tsc_khz is derived from CPUID[15H]. Drivers should check
1245 * that this flag is set before conversion to TSC is attempted.
1246 *
1247 * Return:
1248 * struct system_counterval_t - system counter value with the pointer to the
1249 *	corresponding clocksource
1250 *	@cycles:	System counter value
1251 *	@cs:		Clocksource corresponding to system counter value. Used
1252 *			by timekeeping code to verify comparibility of two cycle
1253 *			values.
1254 */
1255
1256struct system_counterval_t convert_art_ns_to_tsc(u64 art_ns)
1257{
1258	u64 tmp, res, rem;
1259
1260	rem = do_div(art_ns, USEC_PER_SEC);
1261
1262	res = art_ns * tsc_khz;
1263	tmp = rem * tsc_khz;
1264
1265	do_div(tmp, USEC_PER_SEC);
1266	res += tmp;
1267
1268	return (struct system_counterval_t) { .cs = art_related_clocksource,
1269					      .cycles = res};
1270}
1271EXPORT_SYMBOL(convert_art_ns_to_tsc);
1272
1273
1274static void tsc_refine_calibration_work(struct work_struct *work);
1275static DECLARE_DELAYED_WORK(tsc_irqwork, tsc_refine_calibration_work);
1276/**
1277 * tsc_refine_calibration_work - Further refine tsc freq calibration
1278 * @work - ignored.
1279 *
1280 * This functions uses delayed work over a period of a
1281 * second to further refine the TSC freq value. Since this is
1282 * timer based, instead of loop based, we don't block the boot
1283 * process while this longer calibration is done.
1284 *
1285 * If there are any calibration anomalies (too many SMIs, etc),
1286 * or the refined calibration is off by 1% of the fast early
1287 * calibration, we throw out the new calibration and use the
1288 * early calibration.
1289 */
1290static void tsc_refine_calibration_work(struct work_struct *work)
1291{
1292	static u64 tsc_start = ULLONG_MAX, ref_start;
1293	static int hpet;
1294	u64 tsc_stop, ref_stop, delta;
1295	unsigned long freq;
1296	int cpu;
1297
1298	/* Don't bother refining TSC on unstable systems */
1299	if (tsc_unstable)
1300		goto unreg;
1301
1302	/*
1303	 * Since the work is started early in boot, we may be
1304	 * delayed the first time we expire. So set the workqueue
1305	 * again once we know timers are working.
1306	 */
1307	if (tsc_start == ULLONG_MAX) {
1308restart:
1309		/*
1310		 * Only set hpet once, to avoid mixing hardware
1311		 * if the hpet becomes enabled later.
1312		 */
1313		hpet = is_hpet_enabled();
1314		tsc_start = tsc_read_refs(&ref_start, hpet);
1315		schedule_delayed_work(&tsc_irqwork, HZ);
 
1316		return;
1317	}
1318
1319	tsc_stop = tsc_read_refs(&ref_stop, hpet);
1320
1321	/* hpet or pmtimer available ? */
1322	if (ref_start == ref_stop)
1323		goto out;
1324
1325	/* Check, whether the sampling was disturbed */
1326	if (tsc_stop == ULLONG_MAX)
1327		goto restart;
1328
1329	delta = tsc_stop - tsc_start;
1330	delta *= 1000000LL;
1331	if (hpet)
1332		freq = calc_hpet_ref(delta, ref_start, ref_stop);
1333	else
1334		freq = calc_pmtimer_ref(delta, ref_start, ref_stop);
1335
1336	/* Make sure we're within 1% */
1337	if (abs(tsc_khz - freq) > tsc_khz/100)
1338		goto out;
1339
1340	tsc_khz = freq;
1341	pr_info("Refined TSC clocksource calibration: %lu.%03lu MHz\n",
1342		(unsigned long)tsc_khz / 1000,
1343		(unsigned long)tsc_khz % 1000);
1344
1345	/* Inform the TSC deadline clockevent devices about the recalibration */
1346	lapic_update_tsc_freq();
1347
1348	/* Update the sched_clock() rate to match the clocksource one */
1349	for_each_possible_cpu(cpu)
1350		set_cyc2ns_scale(tsc_khz, cpu, tsc_stop);
1351
1352out:
1353	if (tsc_unstable)
1354		goto unreg;
1355
1356	if (boot_cpu_has(X86_FEATURE_ART))
1357		art_related_clocksource = &clocksource_tsc;
1358	clocksource_register_khz(&clocksource_tsc, tsc_khz);
1359unreg:
1360	clocksource_unregister(&clocksource_tsc_early);
1361}
1362
1363
1364static int __init init_tsc_clocksource(void)
1365{
1366	if (!boot_cpu_has(X86_FEATURE_TSC) || !tsc_khz)
1367		return 0;
1368
1369	if (tsc_unstable)
1370		goto unreg;
1371
1372	if (tsc_clocksource_reliable || no_tsc_watchdog)
1373		clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
 
 
 
 
 
1374
1375	if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC_S3))
1376		clocksource_tsc.flags |= CLOCK_SOURCE_SUSPEND_NONSTOP;
1377
1378	/*
1379	 * When TSC frequency is known (retrieved via MSR or CPUID), we skip
1380	 * the refined calibration and directly register it as a clocksource.
1381	 */
1382	if (boot_cpu_has(X86_FEATURE_TSC_KNOWN_FREQ)) {
1383		if (boot_cpu_has(X86_FEATURE_ART))
1384			art_related_clocksource = &clocksource_tsc;
1385		clocksource_register_khz(&clocksource_tsc, tsc_khz);
1386unreg:
1387		clocksource_unregister(&clocksource_tsc_early);
1388		return 0;
1389	}
1390
1391	schedule_delayed_work(&tsc_irqwork, 0);
1392	return 0;
1393}
1394/*
1395 * We use device_initcall here, to ensure we run after the hpet
1396 * is fully initialized, which may occur at fs_initcall time.
1397 */
1398device_initcall(init_tsc_clocksource);
1399
1400static bool __init determine_cpu_tsc_frequencies(bool early)
1401{
1402	/* Make sure that cpu and tsc are not already calibrated */
1403	WARN_ON(cpu_khz || tsc_khz);
1404
1405	if (early) {
1406		cpu_khz = x86_platform.calibrate_cpu();
1407		tsc_khz = x86_platform.calibrate_tsc();
1408	} else {
1409		/* We should not be here with non-native cpu calibration */
1410		WARN_ON(x86_platform.calibrate_cpu != native_calibrate_cpu);
1411		cpu_khz = pit_hpet_ptimer_calibrate_cpu();
1412	}
1413
 
 
 
1414	/*
1415	 * Trust non-zero tsc_khz as authoritative,
1416	 * and use it to sanity check cpu_khz,
1417	 * which will be off if system timer is off.
1418	 */
1419	if (tsc_khz == 0)
1420		tsc_khz = cpu_khz;
1421	else if (abs(cpu_khz - tsc_khz) * 10 > tsc_khz)
1422		cpu_khz = tsc_khz;
1423
1424	if (tsc_khz == 0)
1425		return false;
1426
1427	pr_info("Detected %lu.%03lu MHz processor\n",
1428		(unsigned long)cpu_khz / KHZ,
1429		(unsigned long)cpu_khz % KHZ);
1430
1431	if (cpu_khz != tsc_khz) {
1432		pr_info("Detected %lu.%03lu MHz TSC",
1433			(unsigned long)tsc_khz / KHZ,
1434			(unsigned long)tsc_khz % KHZ);
1435	}
1436	return true;
1437}
1438
1439static unsigned long __init get_loops_per_jiffy(void)
1440{
1441	u64 lpj = (u64)tsc_khz * KHZ;
1442
1443	do_div(lpj, HZ);
1444	return lpj;
1445}
1446
1447static void __init tsc_enable_sched_clock(void)
1448{
1449	/* Sanitize TSC ADJUST before cyc2ns gets initialized */
1450	tsc_store_and_check_tsc_adjust(true);
1451	cyc2ns_init_boot_cpu();
1452	static_branch_enable(&__use_tsc);
1453}
1454
1455void __init tsc_early_init(void)
1456{
1457	if (!boot_cpu_has(X86_FEATURE_TSC))
1458		return;
1459	/* Don't change UV TSC multi-chassis synchronization */
1460	if (is_early_uv_system())
1461		return;
1462	if (!determine_cpu_tsc_frequencies(true))
1463		return;
1464	loops_per_jiffy = get_loops_per_jiffy();
1465
1466	tsc_enable_sched_clock();
1467}
1468
1469void __init tsc_init(void)
1470{
1471	/*
1472	 * native_calibrate_cpu_early can only calibrate using methods that are
1473	 * available early in boot.
 
 
1474	 */
1475	if (x86_platform.calibrate_cpu == native_calibrate_cpu_early)
1476		x86_platform.calibrate_cpu = native_calibrate_cpu;
 
 
1477
1478	if (!boot_cpu_has(X86_FEATURE_TSC)) {
1479		setup_clear_cpu_cap(X86_FEATURE_TSC_DEADLINE_TIMER);
1480		return;
1481	}
1482
1483	if (!tsc_khz) {
1484		/* We failed to determine frequencies earlier, try again */
1485		if (!determine_cpu_tsc_frequencies(false)) {
1486			mark_tsc_unstable("could not calculate TSC khz");
1487			setup_clear_cpu_cap(X86_FEATURE_TSC_DEADLINE_TIMER);
1488			return;
1489		}
1490		tsc_enable_sched_clock();
1491	}
1492
1493	cyc2ns_init_secondary_cpus();
 
1494
1495	if (!no_sched_irq_time)
1496		enable_sched_clock_irqtime();
1497
1498	lpj_fine = get_loops_per_jiffy();
1499	use_tsc_delay();
 
1500
1501	check_system_tsc_reliable();
1502
1503	if (unsynchronized_tsc()) {
1504		mark_tsc_unstable("TSCs unsynchronized");
1505		return;
1506	}
1507
1508	if (tsc_clocksource_reliable || no_tsc_watchdog)
1509		clocksource_tsc_early.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
1510
1511	clocksource_register_khz(&clocksource_tsc_early, tsc_khz);
1512	detect_art();
1513}
1514
1515#ifdef CONFIG_SMP
1516/*
1517 * If we have a constant TSC and are using the TSC for the delay loop,
1518 * we can skip clock calibration if another cpu in the same socket has already
1519 * been calibrated. This assumes that CONSTANT_TSC applies to all
1520 * cpus in the socket - this should be a safe assumption.
1521 */
1522unsigned long calibrate_delay_is_known(void)
1523{
1524	int sibling, cpu = smp_processor_id();
1525	int constant_tsc = cpu_has(&cpu_data(cpu), X86_FEATURE_CONSTANT_TSC);
1526	const struct cpumask *mask = topology_core_cpumask(cpu);
 
 
1527
1528	if (!constant_tsc || !mask)
1529		return 0;
1530
1531	sibling = cpumask_any_but(mask, cpu);
1532	if (sibling < nr_cpu_ids)
1533		return cpu_data(sibling).loops_per_jiffy;
1534	return 0;
1535}
1536#endif
v4.10.11
 
   1#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   2
   3#include <linux/kernel.h>
   4#include <linux/sched.h>
 
   5#include <linux/init.h>
   6#include <linux/export.h>
   7#include <linux/timer.h>
   8#include <linux/acpi_pmtmr.h>
   9#include <linux/cpufreq.h>
  10#include <linux/delay.h>
  11#include <linux/clocksource.h>
  12#include <linux/percpu.h>
  13#include <linux/timex.h>
  14#include <linux/static_key.h>
  15
  16#include <asm/hpet.h>
  17#include <asm/timer.h>
  18#include <asm/vgtod.h>
  19#include <asm/time.h>
  20#include <asm/delay.h>
  21#include <asm/hypervisor.h>
  22#include <asm/nmi.h>
  23#include <asm/x86_init.h>
  24#include <asm/geode.h>
  25#include <asm/apic.h>
  26#include <asm/intel-family.h>
 
 
  27
  28unsigned int __read_mostly cpu_khz;	/* TSC clocks / usec, not used here */
  29EXPORT_SYMBOL(cpu_khz);
  30
  31unsigned int __read_mostly tsc_khz;
  32EXPORT_SYMBOL(tsc_khz);
  33
 
 
  34/*
  35 * TSC can be unstable due to cpufreq or due to unsynced TSCs
  36 */
  37static int __read_mostly tsc_unstable;
  38
  39/* native_sched_clock() is called before tsc_init(), so
  40   we must start with the TSC soft disabled to prevent
  41   erroneous rdtsc usage on !boot_cpu_has(X86_FEATURE_TSC) processors */
  42static int __read_mostly tsc_disabled = -1;
  43
  44static DEFINE_STATIC_KEY_FALSE(__use_tsc);
  45
  46int tsc_clocksource_reliable;
  47
  48static u32 art_to_tsc_numerator;
  49static u32 art_to_tsc_denominator;
  50static u64 art_to_tsc_offset;
  51struct clocksource *art_related_clocksource;
  52
  53/*
  54 * Use a ring-buffer like data structure, where a writer advances the head by
  55 * writing a new data entry and a reader advances the tail when it observes a
  56 * new entry.
  57 *
  58 * Writers are made to wait on readers until there's space to write a new
  59 * entry.
  60 *
  61 * This means that we can always use an {offset, mul} pair to compute a ns
  62 * value that is 'roughly' in the right direction, even if we're writing a new
  63 * {offset, mul} pair during the clock read.
  64 *
  65 * The down-side is that we can no longer guarantee strict monotonicity anymore
  66 * (assuming the TSC was that to begin with), because while we compute the
  67 * intersection point of the two clock slopes and make sure the time is
  68 * continuous at the point of switching; we can no longer guarantee a reader is
  69 * strictly before or after the switch point.
  70 *
  71 * It does mean a reader no longer needs to disable IRQs in order to avoid
  72 * CPU-Freq updates messing with his times, and similarly an NMI reader will
  73 * no longer run the risk of hitting half-written state.
  74 */
  75
  76struct cyc2ns {
  77	struct cyc2ns_data data[2];	/*  0 + 2*24 = 48 */
  78	struct cyc2ns_data *head;	/* 48 + 8    = 56 */
  79	struct cyc2ns_data *tail;	/* 56 + 8    = 64 */
  80}; /* exactly fits one cacheline */
  81
  82static DEFINE_PER_CPU_ALIGNED(struct cyc2ns, cyc2ns);
  83
  84struct cyc2ns_data *cyc2ns_read_begin(void)
  85{
  86	struct cyc2ns_data *head;
  87
  88	preempt_disable();
  89
  90	head = this_cpu_read(cyc2ns.head);
  91	/*
  92	 * Ensure we observe the entry when we observe the pointer to it.
  93	 * matches the wmb from cyc2ns_write_end().
  94	 */
  95	smp_read_barrier_depends();
  96	head->__count++;
  97	barrier();
  98
  99	return head;
 100}
 
 101
 102void cyc2ns_read_end(struct cyc2ns_data *head)
 103{
 104	barrier();
 105	/*
 106	 * If we're the outer most nested read; update the tail pointer
 107	 * when we're done. This notifies possible pending writers
 108	 * that we've observed the head pointer and that the other
 109	 * entry is now free.
 110	 */
 111	if (!--head->__count) {
 112		/*
 113		 * x86-TSO does not reorder writes with older reads;
 114		 * therefore once this write becomes visible to another
 115		 * cpu, we must be finished reading the cyc2ns_data.
 116		 *
 117		 * matches with cyc2ns_write_begin().
 118		 */
 119		this_cpu_write(cyc2ns.tail, head);
 120	}
 121	preempt_enable();
 122}
 123
 124/*
 125 * Begin writing a new @data entry for @cpu.
 126 *
 127 * Assumes some sort of write side lock; currently 'provided' by the assumption
 128 * that cpufreq will call its notifiers sequentially.
 129 */
 130static struct cyc2ns_data *cyc2ns_write_begin(int cpu)
 131{
 132	struct cyc2ns *c2n = &per_cpu(cyc2ns, cpu);
 133	struct cyc2ns_data *data = c2n->data;
 134
 135	if (data == c2n->head)
 136		data++;
 137
 138	/* XXX send an IPI to @cpu in order to guarantee a read? */
 139
 140	/*
 141	 * When we observe the tail write from cyc2ns_read_end(),
 142	 * the cpu must be done with that entry and its safe
 143	 * to start writing to it.
 144	 */
 145	while (c2n->tail == data)
 146		cpu_relax();
 147
 148	return data;
 149}
 150
 151static void cyc2ns_write_end(int cpu, struct cyc2ns_data *data)
 152{
 153	struct cyc2ns *c2n = &per_cpu(cyc2ns, cpu);
 154
 155	/*
 156	 * Ensure the @data writes are visible before we publish the
 157	 * entry. Matches the data-depencency in cyc2ns_read_begin().
 158	 */
 159	smp_wmb();
 160
 161	ACCESS_ONCE(c2n->head) = data;
 162}
 163
 164/*
 165 * Accelerators for sched_clock()
 166 * convert from cycles(64bits) => nanoseconds (64bits)
 167 *  basic equation:
 168 *              ns = cycles / (freq / ns_per_sec)
 169 *              ns = cycles * (ns_per_sec / freq)
 170 *              ns = cycles * (10^9 / (cpu_khz * 10^3))
 171 *              ns = cycles * (10^6 / cpu_khz)
 172 *
 173 *      Then we use scaling math (suggested by george@mvista.com) to get:
 174 *              ns = cycles * (10^6 * SC / cpu_khz) / SC
 175 *              ns = cycles * cyc2ns_scale / SC
 176 *
 177 *      And since SC is a constant power of two, we can convert the div
 178 *  into a shift. The larger SC is, the more accurate the conversion, but
 179 *  cyc2ns_scale needs to be a 32-bit value so that 32-bit multiplication
 180 *  (64-bit result) can be used.
 181 *
 182 *  We can use khz divisor instead of mhz to keep a better precision.
 183 *  (mathieu.desnoyers@polymtl.ca)
 184 *
 185 *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
 186 */
 187
 188static void cyc2ns_data_init(struct cyc2ns_data *data)
 189{
 190	data->cyc2ns_mul = 0;
 191	data->cyc2ns_shift = 0;
 192	data->cyc2ns_offset = 0;
 193	data->__count = 0;
 194}
 195
 196static void cyc2ns_init(int cpu)
 197{
 198	struct cyc2ns *c2n = &per_cpu(cyc2ns, cpu);
 199
 200	cyc2ns_data_init(&c2n->data[0]);
 201	cyc2ns_data_init(&c2n->data[1]);
 202
 203	c2n->head = c2n->data;
 204	c2n->tail = c2n->data;
 205}
 206
 207static inline unsigned long long cycles_2_ns(unsigned long long cyc)
 208{
 209	struct cyc2ns_data *data, *tail;
 210	unsigned long long ns;
 211
 212	/*
 213	 * See cyc2ns_read_*() for details; replicated in order to avoid
 214	 * an extra few instructions that came with the abstraction.
 215	 * Notable, it allows us to only do the __count and tail update
 216	 * dance when its actually needed.
 217	 */
 218
 219	preempt_disable_notrace();
 220	data = this_cpu_read(cyc2ns.head);
 221	tail = this_cpu_read(cyc2ns.tail);
 222
 223	if (likely(data == tail)) {
 224		ns = data->cyc2ns_offset;
 225		ns += mul_u64_u32_shr(cyc, data->cyc2ns_mul, data->cyc2ns_shift);
 226	} else {
 227		data->__count++;
 228
 229		barrier();
 230
 231		ns = data->cyc2ns_offset;
 232		ns += mul_u64_u32_shr(cyc, data->cyc2ns_mul, data->cyc2ns_shift);
 233
 234		barrier();
 235
 236		if (!--data->__count)
 237			this_cpu_write(cyc2ns.tail, data);
 238	}
 239	preempt_enable_notrace();
 240
 241	return ns;
 242}
 243
 244static void set_cyc2ns_scale(unsigned long khz, int cpu)
 245{
 246	unsigned long long tsc_now, ns_now;
 247	struct cyc2ns_data *data;
 248	unsigned long flags;
 249
 250	local_irq_save(flags);
 251	sched_clock_idle_sleep_event();
 252
 253	if (!khz)
 254		goto done;
 255
 256	data = cyc2ns_write_begin(cpu);
 257
 258	tsc_now = rdtsc();
 259	ns_now = cycles_2_ns(tsc_now);
 260
 261	/*
 262	 * Compute a new multiplier as per the above comment and ensure our
 263	 * time function is continuous; see the comment near struct
 264	 * cyc2ns_data.
 265	 */
 266	clocks_calc_mult_shift(&data->cyc2ns_mul, &data->cyc2ns_shift, khz,
 267			       NSEC_PER_MSEC, 0);
 268
 269	/*
 270	 * cyc2ns_shift is exported via arch_perf_update_userpage() where it is
 271	 * not expected to be greater than 31 due to the original published
 272	 * conversion algorithm shifting a 32-bit value (now specifies a 64-bit
 273	 * value) - refer perf_event_mmap_page documentation in perf_event.h.
 274	 */
 275	if (data->cyc2ns_shift == 32) {
 276		data->cyc2ns_shift = 31;
 277		data->cyc2ns_mul >>= 1;
 278	}
 279
 280	data->cyc2ns_offset = ns_now -
 281		mul_u64_u32_shr(tsc_now, data->cyc2ns_mul, data->cyc2ns_shift);
 
 
 282
 283	cyc2ns_write_end(cpu, data);
 
 
 
 
 284
 285done:
 286	sched_clock_idle_wakeup_event(0);
 
 
 
 
 
 
 
 
 
 287	local_irq_restore(flags);
 288}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 289/*
 290 * Scheduler clock - returns current time in nanosec units.
 291 */
 292u64 native_sched_clock(void)
 293{
 294	if (static_branch_likely(&__use_tsc)) {
 295		u64 tsc_now = rdtsc();
 296
 297		/* return the value in ns */
 298		return cycles_2_ns(tsc_now);
 299	}
 300
 301	/*
 302	 * Fall back to jiffies if there's no TSC available:
 303	 * ( But note that we still use it if the TSC is marked
 304	 *   unstable. We do this because unlike Time Of Day,
 305	 *   the scheduler clock tolerates small errors and it's
 306	 *   very important for it to be as fast as the platform
 307	 *   can achieve it. )
 308	 */
 309
 310	/* No locking but a rare wrong value is not a big deal: */
 311	return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
 312}
 313
 314/*
 315 * Generate a sched_clock if you already have a TSC value.
 316 */
 317u64 native_sched_clock_from_tsc(u64 tsc)
 318{
 319	return cycles_2_ns(tsc);
 320}
 321
 322/* We need to define a real function for sched_clock, to override the
 323   weak default version */
 324#ifdef CONFIG_PARAVIRT
 325unsigned long long sched_clock(void)
 326{
 327	return paravirt_sched_clock();
 328}
 
 
 
 
 
 329#else
 330unsigned long long
 331sched_clock(void) __attribute__((alias("native_sched_clock")));
 
 
 332#endif
 333
 334int check_tsc_unstable(void)
 335{
 336	return tsc_unstable;
 337}
 338EXPORT_SYMBOL_GPL(check_tsc_unstable);
 339
 340#ifdef CONFIG_X86_TSC
 341int __init notsc_setup(char *str)
 342{
 343	pr_warn("Kernel compiled with CONFIG_X86_TSC, cannot disable TSC completely\n");
 344	tsc_disabled = 1;
 345	return 1;
 346}
 347#else
 348/*
 349 * disable flag for tsc. Takes effect by clearing the TSC cpu flag
 350 * in cpu/common.c
 351 */
 352int __init notsc_setup(char *str)
 353{
 354	setup_clear_cpu_cap(X86_FEATURE_TSC);
 355	return 1;
 356}
 357#endif
 358
 359__setup("notsc", notsc_setup);
 360
 361static int no_sched_irq_time;
 
 362
 363static int __init tsc_setup(char *str)
 364{
 365	if (!strcmp(str, "reliable"))
 366		tsc_clocksource_reliable = 1;
 367	if (!strncmp(str, "noirqtime", 9))
 368		no_sched_irq_time = 1;
 
 
 
 
 369	return 1;
 370}
 371
 372__setup("tsc=", tsc_setup);
 373
 374#define MAX_RETRIES     5
 375#define SMI_TRESHOLD    50000
 376
 377/*
 378 * Read TSC and the reference counters. Take care of SMI disturbance
 379 */
 380static u64 tsc_read_refs(u64 *p, int hpet)
 381{
 382	u64 t1, t2;
 
 383	int i;
 384
 385	for (i = 0; i < MAX_RETRIES; i++) {
 386		t1 = get_cycles();
 387		if (hpet)
 388			*p = hpet_readl(HPET_COUNTER) & 0xFFFFFFFF;
 389		else
 390			*p = acpi_pm_read_early();
 391		t2 = get_cycles();
 392		if ((t2 - t1) < SMI_TRESHOLD)
 393			return t2;
 394	}
 395	return ULLONG_MAX;
 396}
 397
 398/*
 399 * Calculate the TSC frequency from HPET reference
 400 */
 401static unsigned long calc_hpet_ref(u64 deltatsc, u64 hpet1, u64 hpet2)
 402{
 403	u64 tmp;
 404
 405	if (hpet2 < hpet1)
 406		hpet2 += 0x100000000ULL;
 407	hpet2 -= hpet1;
 408	tmp = ((u64)hpet2 * hpet_readl(HPET_PERIOD));
 409	do_div(tmp, 1000000);
 410	do_div(deltatsc, tmp);
 411
 412	return (unsigned long) deltatsc;
 413}
 414
 415/*
 416 * Calculate the TSC frequency from PMTimer reference
 417 */
 418static unsigned long calc_pmtimer_ref(u64 deltatsc, u64 pm1, u64 pm2)
 419{
 420	u64 tmp;
 421
 422	if (!pm1 && !pm2)
 423		return ULONG_MAX;
 424
 425	if (pm2 < pm1)
 426		pm2 += (u64)ACPI_PM_OVRRUN;
 427	pm2 -= pm1;
 428	tmp = pm2 * 1000000000LL;
 429	do_div(tmp, PMTMR_TICKS_PER_SEC);
 430	do_div(deltatsc, tmp);
 431
 432	return (unsigned long) deltatsc;
 433}
 434
 435#define CAL_MS		10
 436#define CAL_LATCH	(PIT_TICK_RATE / (1000 / CAL_MS))
 437#define CAL_PIT_LOOPS	1000
 438
 439#define CAL2_MS		50
 440#define CAL2_LATCH	(PIT_TICK_RATE / (1000 / CAL2_MS))
 441#define CAL2_PIT_LOOPS	5000
 442
 443
 444/*
 445 * Try to calibrate the TSC against the Programmable
 446 * Interrupt Timer and return the frequency of the TSC
 447 * in kHz.
 448 *
 449 * Return ULONG_MAX on failure to calibrate.
 450 */
 451static unsigned long pit_calibrate_tsc(u32 latch, unsigned long ms, int loopmin)
 452{
 453	u64 tsc, t1, t2, delta;
 454	unsigned long tscmin, tscmax;
 455	int pitcnt;
 456
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 457	/* Set the Gate high, disable speaker */
 458	outb((inb(0x61) & ~0x02) | 0x01, 0x61);
 459
 460	/*
 461	 * Setup CTC channel 2* for mode 0, (interrupt on terminal
 462	 * count mode), binary count. Set the latch register to 50ms
 463	 * (LSB then MSB) to begin countdown.
 464	 */
 465	outb(0xb0, 0x43);
 466	outb(latch & 0xff, 0x42);
 467	outb(latch >> 8, 0x42);
 468
 469	tsc = t1 = t2 = get_cycles();
 470
 471	pitcnt = 0;
 472	tscmax = 0;
 473	tscmin = ULONG_MAX;
 474	while ((inb(0x61) & 0x20) == 0) {
 475		t2 = get_cycles();
 476		delta = t2 - tsc;
 477		tsc = t2;
 478		if ((unsigned long) delta < tscmin)
 479			tscmin = (unsigned int) delta;
 480		if ((unsigned long) delta > tscmax)
 481			tscmax = (unsigned int) delta;
 482		pitcnt++;
 483	}
 484
 485	/*
 486	 * Sanity checks:
 487	 *
 488	 * If we were not able to read the PIT more than loopmin
 489	 * times, then we have been hit by a massive SMI
 490	 *
 491	 * If the maximum is 10 times larger than the minimum,
 492	 * then we got hit by an SMI as well.
 493	 */
 494	if (pitcnt < loopmin || tscmax > 10 * tscmin)
 495		return ULONG_MAX;
 496
 497	/* Calculate the PIT value */
 498	delta = t2 - t1;
 499	do_div(delta, ms);
 500	return delta;
 501}
 502
 503/*
 504 * This reads the current MSB of the PIT counter, and
 505 * checks if we are running on sufficiently fast and
 506 * non-virtualized hardware.
 507 *
 508 * Our expectations are:
 509 *
 510 *  - the PIT is running at roughly 1.19MHz
 511 *
 512 *  - each IO is going to take about 1us on real hardware,
 513 *    but we allow it to be much faster (by a factor of 10) or
 514 *    _slightly_ slower (ie we allow up to a 2us read+counter
 515 *    update - anything else implies a unacceptably slow CPU
 516 *    or PIT for the fast calibration to work.
 517 *
 518 *  - with 256 PIT ticks to read the value, we have 214us to
 519 *    see the same MSB (and overhead like doing a single TSC
 520 *    read per MSB value etc).
 521 *
 522 *  - We're doing 2 reads per loop (LSB, MSB), and we expect
 523 *    them each to take about a microsecond on real hardware.
 524 *    So we expect a count value of around 100. But we'll be
 525 *    generous, and accept anything over 50.
 526 *
 527 *  - if the PIT is stuck, and we see *many* more reads, we
 528 *    return early (and the next caller of pit_expect_msb()
 529 *    then consider it a failure when they don't see the
 530 *    next expected value).
 531 *
 532 * These expectations mean that we know that we have seen the
 533 * transition from one expected value to another with a fairly
 534 * high accuracy, and we didn't miss any events. We can thus
 535 * use the TSC value at the transitions to calculate a pretty
 536 * good value for the TSC frequencty.
 537 */
 538static inline int pit_verify_msb(unsigned char val)
 539{
 540	/* Ignore LSB */
 541	inb(0x42);
 542	return inb(0x42) == val;
 543}
 544
 545static inline int pit_expect_msb(unsigned char val, u64 *tscp, unsigned long *deltap)
 546{
 547	int count;
 548	u64 tsc = 0, prev_tsc = 0;
 549
 550	for (count = 0; count < 50000; count++) {
 551		if (!pit_verify_msb(val))
 552			break;
 553		prev_tsc = tsc;
 554		tsc = get_cycles();
 555	}
 556	*deltap = get_cycles() - prev_tsc;
 557	*tscp = tsc;
 558
 559	/*
 560	 * We require _some_ success, but the quality control
 561	 * will be based on the error terms on the TSC values.
 562	 */
 563	return count > 5;
 564}
 565
 566/*
 567 * How many MSB values do we want to see? We aim for
 568 * a maximum error rate of 500ppm (in practice the
 569 * real error is much smaller), but refuse to spend
 570 * more than 50ms on it.
 571 */
 572#define MAX_QUICK_PIT_MS 50
 573#define MAX_QUICK_PIT_ITERATIONS (MAX_QUICK_PIT_MS * PIT_TICK_RATE / 1000 / 256)
 574
 575static unsigned long quick_pit_calibrate(void)
 576{
 577	int i;
 578	u64 tsc, delta;
 579	unsigned long d1, d2;
 580
 
 
 
 581	/* Set the Gate high, disable speaker */
 582	outb((inb(0x61) & ~0x02) | 0x01, 0x61);
 583
 584	/*
 585	 * Counter 2, mode 0 (one-shot), binary count
 586	 *
 587	 * NOTE! Mode 2 decrements by two (and then the
 588	 * output is flipped each time, giving the same
 589	 * final output frequency as a decrement-by-one),
 590	 * so mode 0 is much better when looking at the
 591	 * individual counts.
 592	 */
 593	outb(0xb0, 0x43);
 594
 595	/* Start at 0xffff */
 596	outb(0xff, 0x42);
 597	outb(0xff, 0x42);
 598
 599	/*
 600	 * The PIT starts counting at the next edge, so we
 601	 * need to delay for a microsecond. The easiest way
 602	 * to do that is to just read back the 16-bit counter
 603	 * once from the PIT.
 604	 */
 605	pit_verify_msb(0);
 606
 607	if (pit_expect_msb(0xff, &tsc, &d1)) {
 608		for (i = 1; i <= MAX_QUICK_PIT_ITERATIONS; i++) {
 609			if (!pit_expect_msb(0xff-i, &delta, &d2))
 610				break;
 611
 612			delta -= tsc;
 613
 614			/*
 615			 * Extrapolate the error and fail fast if the error will
 616			 * never be below 500 ppm.
 617			 */
 618			if (i == 1 &&
 619			    d1 + d2 >= (delta * MAX_QUICK_PIT_ITERATIONS) >> 11)
 620				return 0;
 621
 622			/*
 623			 * Iterate until the error is less than 500 ppm
 624			 */
 625			if (d1+d2 >= delta >> 11)
 626				continue;
 627
 628			/*
 629			 * Check the PIT one more time to verify that
 630			 * all TSC reads were stable wrt the PIT.
 631			 *
 632			 * This also guarantees serialization of the
 633			 * last cycle read ('d2') in pit_expect_msb.
 634			 */
 635			if (!pit_verify_msb(0xfe - i))
 636				break;
 637			goto success;
 638		}
 639	}
 640	pr_info("Fast TSC calibration failed\n");
 641	return 0;
 642
 643success:
 644	/*
 645	 * Ok, if we get here, then we've seen the
 646	 * MSB of the PIT decrement 'i' times, and the
 647	 * error has shrunk to less than 500 ppm.
 648	 *
 649	 * As a result, we can depend on there not being
 650	 * any odd delays anywhere, and the TSC reads are
 651	 * reliable (within the error).
 652	 *
 653	 * kHz = ticks / time-in-seconds / 1000;
 654	 * kHz = (t2 - t1) / (I * 256 / PIT_TICK_RATE) / 1000
 655	 * kHz = ((t2 - t1) * PIT_TICK_RATE) / (I * 256 * 1000)
 656	 */
 657	delta *= PIT_TICK_RATE;
 658	do_div(delta, i*256*1000);
 659	pr_info("Fast TSC calibration using PIT\n");
 660	return delta;
 661}
 662
 663/**
 664 * native_calibrate_tsc
 665 * Determine TSC frequency via CPUID, else return 0.
 666 */
 667unsigned long native_calibrate_tsc(void)
 668{
 669	unsigned int eax_denominator, ebx_numerator, ecx_hz, edx;
 670	unsigned int crystal_khz;
 671
 672	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
 673		return 0;
 674
 675	if (boot_cpu_data.cpuid_level < 0x15)
 676		return 0;
 677
 678	eax_denominator = ebx_numerator = ecx_hz = edx = 0;
 679
 680	/* CPUID 15H TSC/Crystal ratio, plus optionally Crystal Hz */
 681	cpuid(0x15, &eax_denominator, &ebx_numerator, &ecx_hz, &edx);
 682
 683	if (ebx_numerator == 0 || eax_denominator == 0)
 684		return 0;
 685
 686	crystal_khz = ecx_hz / 1000;
 687
 688	if (crystal_khz == 0) {
 689		switch (boot_cpu_data.x86_model) {
 690		case INTEL_FAM6_SKYLAKE_MOBILE:
 691		case INTEL_FAM6_SKYLAKE_DESKTOP:
 692		case INTEL_FAM6_KABYLAKE_MOBILE:
 693		case INTEL_FAM6_KABYLAKE_DESKTOP:
 694			crystal_khz = 24000;	/* 24.0 MHz */
 695			break;
 696		case INTEL_FAM6_SKYLAKE_X:
 697		case INTEL_FAM6_ATOM_DENVERTON:
 698			crystal_khz = 25000;	/* 25.0 MHz */
 699			break;
 700		case INTEL_FAM6_ATOM_GOLDMONT:
 701			crystal_khz = 19200;	/* 19.2 MHz */
 702			break;
 703		}
 704	}
 705
 706	/*
 707	 * TSC frequency determined by CPUID is a "hardware reported"
 708	 * frequency and is the most accurate one so far we have. This
 709	 * is considered a known frequency.
 710	 */
 711	setup_force_cpu_cap(X86_FEATURE_TSC_KNOWN_FREQ);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 712
 713	/*
 714	 * For Atom SoCs TSC is the only reliable clocksource.
 715	 * Mark TSC reliable so no watchdog on it.
 716	 */
 717	if (boot_cpu_data.x86_model == INTEL_FAM6_ATOM_GOLDMONT)
 718		setup_force_cpu_cap(X86_FEATURE_TSC_RELIABLE);
 719
 
 
 
 
 
 
 
 
 
 
 720	return crystal_khz * ebx_numerator / eax_denominator;
 721}
 722
 723static unsigned long cpu_khz_from_cpuid(void)
 724{
 725	unsigned int eax_base_mhz, ebx_max_mhz, ecx_bus_mhz, edx;
 726
 727	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
 728		return 0;
 729
 730	if (boot_cpu_data.cpuid_level < 0x16)
 731		return 0;
 732
 733	eax_base_mhz = ebx_max_mhz = ecx_bus_mhz = edx = 0;
 734
 735	cpuid(0x16, &eax_base_mhz, &ebx_max_mhz, &ecx_bus_mhz, &edx);
 736
 737	return eax_base_mhz * 1000;
 738}
 739
 740/**
 741 * native_calibrate_cpu - calibrate the cpu on boot
 
 742 */
 743unsigned long native_calibrate_cpu(void)
 744{
 745	u64 tsc1, tsc2, delta, ref1, ref2;
 746	unsigned long tsc_pit_min = ULONG_MAX, tsc_ref_min = ULONG_MAX;
 747	unsigned long flags, latch, ms, fast_calibrate;
 748	int hpet = is_hpet_enabled(), i, loopmin;
 749
 750	fast_calibrate = cpu_khz_from_cpuid();
 751	if (fast_calibrate)
 752		return fast_calibrate;
 753
 754	fast_calibrate = cpu_khz_from_msr();
 755	if (fast_calibrate)
 756		return fast_calibrate;
 757
 758	local_irq_save(flags);
 759	fast_calibrate = quick_pit_calibrate();
 760	local_irq_restore(flags);
 761	if (fast_calibrate)
 762		return fast_calibrate;
 763
 764	/*
 765	 * Run 5 calibration loops to get the lowest frequency value
 766	 * (the best estimate). We use two different calibration modes
 767	 * here:
 768	 *
 769	 * 1) PIT loop. We set the PIT Channel 2 to oneshot mode and
 770	 * load a timeout of 50ms. We read the time right after we
 771	 * started the timer and wait until the PIT count down reaches
 772	 * zero. In each wait loop iteration we read the TSC and check
 773	 * the delta to the previous read. We keep track of the min
 774	 * and max values of that delta. The delta is mostly defined
 775	 * by the IO time of the PIT access, so we can detect when a
 776	 * SMI/SMM disturbance happened between the two reads. If the
 777	 * maximum time is significantly larger than the minimum time,
 778	 * then we discard the result and have another try.
 779	 *
 780	 * 2) Reference counter. If available we use the HPET or the
 781	 * PMTIMER as a reference to check the sanity of that value.
 782	 * We use separate TSC readouts and check inside of the
 783	 * reference read for a SMI/SMM disturbance. We dicard
 784	 * disturbed values here as well. We do that around the PIT
 785	 * calibration delay loop as we have to wait for a certain
 786	 * amount of time anyway.
 787	 */
 788
 789	/* Preset PIT loop values */
 790	latch = CAL_LATCH;
 791	ms = CAL_MS;
 792	loopmin = CAL_PIT_LOOPS;
 793
 794	for (i = 0; i < 3; i++) {
 795		unsigned long tsc_pit_khz;
 796
 797		/*
 798		 * Read the start value and the reference count of
 799		 * hpet/pmtimer when available. Then do the PIT
 800		 * calibration, which will take at least 50ms, and
 801		 * read the end value.
 802		 */
 803		local_irq_save(flags);
 804		tsc1 = tsc_read_refs(&ref1, hpet);
 805		tsc_pit_khz = pit_calibrate_tsc(latch, ms, loopmin);
 806		tsc2 = tsc_read_refs(&ref2, hpet);
 807		local_irq_restore(flags);
 808
 809		/* Pick the lowest PIT TSC calibration so far */
 810		tsc_pit_min = min(tsc_pit_min, tsc_pit_khz);
 811
 812		/* hpet or pmtimer available ? */
 813		if (ref1 == ref2)
 814			continue;
 815
 816		/* Check, whether the sampling was disturbed by an SMI */
 817		if (tsc1 == ULLONG_MAX || tsc2 == ULLONG_MAX)
 818			continue;
 819
 820		tsc2 = (tsc2 - tsc1) * 1000000LL;
 821		if (hpet)
 822			tsc2 = calc_hpet_ref(tsc2, ref1, ref2);
 823		else
 824			tsc2 = calc_pmtimer_ref(tsc2, ref1, ref2);
 825
 826		tsc_ref_min = min(tsc_ref_min, (unsigned long) tsc2);
 827
 828		/* Check the reference deviation */
 829		delta = ((u64) tsc_pit_min) * 100;
 830		do_div(delta, tsc_ref_min);
 831
 832		/*
 833		 * If both calibration results are inside a 10% window
 834		 * then we can be sure, that the calibration
 835		 * succeeded. We break out of the loop right away. We
 836		 * use the reference value, as it is more precise.
 837		 */
 838		if (delta >= 90 && delta <= 110) {
 839			pr_info("PIT calibration matches %s. %d loops\n",
 840				hpet ? "HPET" : "PMTIMER", i + 1);
 841			return tsc_ref_min;
 842		}
 843
 844		/*
 845		 * Check whether PIT failed more than once. This
 846		 * happens in virtualized environments. We need to
 847		 * give the virtual PC a slightly longer timeframe for
 848		 * the HPET/PMTIMER to make the result precise.
 849		 */
 850		if (i == 1 && tsc_pit_min == ULONG_MAX) {
 851			latch = CAL2_LATCH;
 852			ms = CAL2_MS;
 853			loopmin = CAL2_PIT_LOOPS;
 854		}
 855	}
 856
 857	/*
 858	 * Now check the results.
 859	 */
 860	if (tsc_pit_min == ULONG_MAX) {
 861		/* PIT gave no useful value */
 862		pr_warn("Unable to calibrate against PIT\n");
 863
 864		/* We don't have an alternative source, disable TSC */
 865		if (!hpet && !ref1 && !ref2) {
 866			pr_notice("No reference (HPET/PMTIMER) available\n");
 867			return 0;
 868		}
 869
 870		/* The alternative source failed as well, disable TSC */
 871		if (tsc_ref_min == ULONG_MAX) {
 872			pr_warn("HPET/PMTIMER calibration failed\n");
 873			return 0;
 874		}
 875
 876		/* Use the alternative source */
 877		pr_info("using %s reference calibration\n",
 878			hpet ? "HPET" : "PMTIMER");
 879
 880		return tsc_ref_min;
 881	}
 882
 883	/* We don't have an alternative source, use the PIT calibration value */
 884	if (!hpet && !ref1 && !ref2) {
 885		pr_info("Using PIT calibration value\n");
 886		return tsc_pit_min;
 887	}
 888
 889	/* The alternative source failed, use the PIT calibration value */
 890	if (tsc_ref_min == ULONG_MAX) {
 891		pr_warn("HPET/PMTIMER calibration failed. Using PIT calibration.\n");
 892		return tsc_pit_min;
 893	}
 894
 895	/*
 896	 * The calibration values differ too much. In doubt, we use
 897	 * the PIT value as we know that there are PMTIMERs around
 898	 * running at double speed. At least we let the user know:
 899	 */
 900	pr_warn("PIT calibration deviates from %s: %lu %lu\n",
 901		hpet ? "HPET" : "PMTIMER", tsc_pit_min, tsc_ref_min);
 902	pr_info("Using PIT calibration value\n");
 903	return tsc_pit_min;
 904}
 905
 906int recalibrate_cpu_khz(void)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 907{
 908#ifndef CONFIG_SMP
 909	unsigned long cpu_khz_old = cpu_khz;
 910
 911	if (!boot_cpu_has(X86_FEATURE_TSC))
 912		return -ENODEV;
 913
 914	cpu_khz = x86_platform.calibrate_cpu();
 915	tsc_khz = x86_platform.calibrate_tsc();
 916	if (tsc_khz == 0)
 917		tsc_khz = cpu_khz;
 918	else if (abs(cpu_khz - tsc_khz) * 10 > tsc_khz)
 919		cpu_khz = tsc_khz;
 920	cpu_data(0).loops_per_jiffy = cpufreq_scale(cpu_data(0).loops_per_jiffy,
 921						    cpu_khz_old, cpu_khz);
 922
 923	return 0;
 924#else
 925	return -ENODEV;
 926#endif
 927}
 928
 929EXPORT_SYMBOL(recalibrate_cpu_khz);
 930
 931
 932static unsigned long long cyc2ns_suspend;
 933
 934void tsc_save_sched_clock_state(void)
 935{
 936	if (!sched_clock_stable())
 937		return;
 938
 939	cyc2ns_suspend = sched_clock();
 940}
 941
 942/*
 943 * Even on processors with invariant TSC, TSC gets reset in some the
 944 * ACPI system sleep states. And in some systems BIOS seem to reinit TSC to
 945 * arbitrary value (still sync'd across cpu's) during resume from such sleep
 946 * states. To cope up with this, recompute the cyc2ns_offset for each cpu so
 947 * that sched_clock() continues from the point where it was left off during
 948 * suspend.
 949 */
 950void tsc_restore_sched_clock_state(void)
 951{
 952	unsigned long long offset;
 953	unsigned long flags;
 954	int cpu;
 955
 956	if (!sched_clock_stable())
 957		return;
 958
 959	local_irq_save(flags);
 960
 961	/*
 962	 * We're coming out of suspend, there's no concurrency yet; don't
 963	 * bother being nice about the RCU stuff, just write to both
 964	 * data fields.
 965	 */
 966
 967	this_cpu_write(cyc2ns.data[0].cyc2ns_offset, 0);
 968	this_cpu_write(cyc2ns.data[1].cyc2ns_offset, 0);
 969
 970	offset = cyc2ns_suspend - sched_clock();
 971
 972	for_each_possible_cpu(cpu) {
 973		per_cpu(cyc2ns.data[0].cyc2ns_offset, cpu) = offset;
 974		per_cpu(cyc2ns.data[1].cyc2ns_offset, cpu) = offset;
 975	}
 976
 977	local_irq_restore(flags);
 978}
 979
 980#ifdef CONFIG_CPU_FREQ
 981
 982/* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
 983 * changes.
 984 *
 985 * RED-PEN: On SMP we assume all CPUs run with the same frequency.  It's
 986 * not that important because current Opteron setups do not support
 987 * scaling on SMP anyroads.
 988 *
 989 * Should fix up last_tsc too. Currently gettimeofday in the
 990 * first tick after the change will be slightly wrong.
 991 */
 992
 993static unsigned int  ref_freq;
 994static unsigned long loops_per_jiffy_ref;
 995static unsigned long tsc_khz_ref;
 996
 997static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
 998				void *data)
 999{
1000	struct cpufreq_freqs *freq = data;
1001	unsigned long *lpj;
1002
1003	lpj = &boot_cpu_data.loops_per_jiffy;
1004#ifdef CONFIG_SMP
1005	if (!(freq->flags & CPUFREQ_CONST_LOOPS))
1006		lpj = &cpu_data(freq->cpu).loops_per_jiffy;
1007#endif
1008
1009	if (!ref_freq) {
1010		ref_freq = freq->old;
1011		loops_per_jiffy_ref = *lpj;
1012		tsc_khz_ref = tsc_khz;
1013	}
 
1014	if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
1015			(val == CPUFREQ_POSTCHANGE && freq->old > freq->new)) {
1016		*lpj = cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
 
1017
1018		tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
1019		if (!(freq->flags & CPUFREQ_CONST_LOOPS))
1020			mark_tsc_unstable("cpufreq changes");
1021
1022		set_cyc2ns_scale(tsc_khz, freq->cpu);
1023	}
1024
1025	return 0;
1026}
1027
1028static struct notifier_block time_cpufreq_notifier_block = {
1029	.notifier_call  = time_cpufreq_notifier
1030};
1031
1032static int __init cpufreq_register_tsc_scaling(void)
1033{
1034	if (!boot_cpu_has(X86_FEATURE_TSC))
1035		return 0;
1036	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
1037		return 0;
1038	cpufreq_register_notifier(&time_cpufreq_notifier_block,
1039				CPUFREQ_TRANSITION_NOTIFIER);
1040	return 0;
1041}
1042
1043core_initcall(cpufreq_register_tsc_scaling);
1044
1045#endif /* CONFIG_CPU_FREQ */
1046
1047#define ART_CPUID_LEAF (0x15)
1048#define ART_MIN_DENOMINATOR (1)
1049
1050
1051/*
1052 * If ART is present detect the numerator:denominator to convert to TSC
1053 */
1054static void detect_art(void)
1055{
1056	unsigned int unused[2];
1057
1058	if (boot_cpu_data.cpuid_level < ART_CPUID_LEAF)
1059		return;
1060
1061	/* Don't enable ART in a VM, non-stop TSC and TSC_ADJUST required */
 
 
 
1062	if (boot_cpu_has(X86_FEATURE_HYPERVISOR) ||
1063	    !boot_cpu_has(X86_FEATURE_NONSTOP_TSC) ||
1064	    !boot_cpu_has(X86_FEATURE_TSC_ADJUST))
 
1065		return;
1066
1067	cpuid(ART_CPUID_LEAF, &art_to_tsc_denominator,
1068	      &art_to_tsc_numerator, unused, unused+1);
1069
1070	if (art_to_tsc_denominator < ART_MIN_DENOMINATOR)
1071		return;
1072
1073	rdmsrl(MSR_IA32_TSC_ADJUST, art_to_tsc_offset);
1074
1075	/* Make this sticky over multiple CPU init calls */
1076	setup_force_cpu_cap(X86_FEATURE_ART);
1077}
1078
1079
1080/* clocksource code */
1081
1082static struct clocksource clocksource_tsc;
1083
1084static void tsc_resume(struct clocksource *cs)
1085{
1086	tsc_verify_tsc_adjust(true);
1087}
1088
1089/*
1090 * We used to compare the TSC to the cycle_last value in the clocksource
1091 * structure to avoid a nasty time-warp. This can be observed in a
1092 * very small window right after one CPU updated cycle_last under
1093 * xtime/vsyscall_gtod lock and the other CPU reads a TSC value which
1094 * is smaller than the cycle_last reference value due to a TSC which
1095 * is slighty behind. This delta is nowhere else observable, but in
1096 * that case it results in a forward time jump in the range of hours
1097 * due to the unsigned delta calculation of the time keeping core
1098 * code, which is necessary to support wrapping clocksources like pm
1099 * timer.
1100 *
1101 * This sanity check is now done in the core timekeeping code.
1102 * checking the result of read_tsc() - cycle_last for being negative.
1103 * That works because CLOCKSOURCE_MASK(64) does not mask out any bit.
1104 */
1105static u64 read_tsc(struct clocksource *cs)
1106{
1107	return (u64)rdtsc_ordered();
1108}
1109
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1110/*
1111 * .mask MUST be CLOCKSOURCE_MASK(64). See comment above read_tsc()
1112 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1113static struct clocksource clocksource_tsc = {
1114	.name                   = "tsc",
1115	.rating                 = 300,
1116	.read                   = read_tsc,
1117	.mask                   = CLOCKSOURCE_MASK(64),
1118	.flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
 
1119				  CLOCK_SOURCE_MUST_VERIFY,
1120	.archdata               = { .vclock_mode = VCLOCK_TSC },
1121	.resume			= tsc_resume,
 
 
 
1122};
1123
1124void mark_tsc_unstable(char *reason)
1125{
1126	if (!tsc_unstable) {
1127		tsc_unstable = 1;
 
 
 
1128		clear_sched_clock_stable();
1129		disable_sched_clock_irqtime();
1130		pr_info("Marking TSC unstable due to %s\n", reason);
1131		/* Change only the rating, when not registered */
1132		if (clocksource_tsc.mult)
1133			clocksource_mark_unstable(&clocksource_tsc);
1134		else {
1135			clocksource_tsc.flags |= CLOCK_SOURCE_UNSTABLE;
1136			clocksource_tsc.rating = 0;
1137		}
1138	}
1139}
1140
1141EXPORT_SYMBOL_GPL(mark_tsc_unstable);
1142
1143static void __init check_system_tsc_reliable(void)
1144{
1145#if defined(CONFIG_MGEODEGX1) || defined(CONFIG_MGEODE_LX) || defined(CONFIG_X86_GENERIC)
1146	if (is_geode_lx()) {
1147		/* RTSC counts during suspend */
1148#define RTSC_SUSP 0x100
1149		unsigned long res_low, res_high;
1150
1151		rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
1152		/* Geode_LX - the OLPC CPU has a very reliable TSC */
1153		if (res_low & RTSC_SUSP)
1154			tsc_clocksource_reliable = 1;
1155	}
1156#endif
1157	if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE))
1158		tsc_clocksource_reliable = 1;
1159}
1160
1161/*
1162 * Make an educated guess if the TSC is trustworthy and synchronized
1163 * over all CPUs.
1164 */
1165int unsynchronized_tsc(void)
1166{
1167	if (!boot_cpu_has(X86_FEATURE_TSC) || tsc_unstable)
1168		return 1;
1169
1170#ifdef CONFIG_SMP
1171	if (apic_is_clustered_box())
1172		return 1;
1173#endif
1174
1175	if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
1176		return 0;
1177
1178	if (tsc_clocksource_reliable)
1179		return 0;
1180	/*
1181	 * Intel systems are normally all synchronized.
1182	 * Exceptions must mark TSC as unstable:
1183	 */
1184	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
1185		/* assume multi socket systems are not synchronized: */
1186		if (num_possible_cpus() > 1)
1187			return 1;
1188	}
1189
1190	return 0;
1191}
1192
1193/*
1194 * Convert ART to TSC given numerator/denominator found in detect_art()
1195 */
1196struct system_counterval_t convert_art_to_tsc(u64 art)
1197{
1198	u64 tmp, res, rem;
1199
1200	rem = do_div(art, art_to_tsc_denominator);
1201
1202	res = art * art_to_tsc_numerator;
1203	tmp = rem * art_to_tsc_numerator;
1204
1205	do_div(tmp, art_to_tsc_denominator);
1206	res += tmp + art_to_tsc_offset;
1207
1208	return (struct system_counterval_t) {.cs = art_related_clocksource,
1209			.cycles = res};
1210}
1211EXPORT_SYMBOL(convert_art_to_tsc);
1212
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1213static void tsc_refine_calibration_work(struct work_struct *work);
1214static DECLARE_DELAYED_WORK(tsc_irqwork, tsc_refine_calibration_work);
1215/**
1216 * tsc_refine_calibration_work - Further refine tsc freq calibration
1217 * @work - ignored.
1218 *
1219 * This functions uses delayed work over a period of a
1220 * second to further refine the TSC freq value. Since this is
1221 * timer based, instead of loop based, we don't block the boot
1222 * process while this longer calibration is done.
1223 *
1224 * If there are any calibration anomalies (too many SMIs, etc),
1225 * or the refined calibration is off by 1% of the fast early
1226 * calibration, we throw out the new calibration and use the
1227 * early calibration.
1228 */
1229static void tsc_refine_calibration_work(struct work_struct *work)
1230{
1231	static u64 tsc_start = -1, ref_start;
1232	static int hpet;
1233	u64 tsc_stop, ref_stop, delta;
1234	unsigned long freq;
 
1235
1236	/* Don't bother refining TSC on unstable systems */
1237	if (check_tsc_unstable())
1238		goto out;
1239
1240	/*
1241	 * Since the work is started early in boot, we may be
1242	 * delayed the first time we expire. So set the workqueue
1243	 * again once we know timers are working.
1244	 */
1245	if (tsc_start == -1) {
 
1246		/*
1247		 * Only set hpet once, to avoid mixing hardware
1248		 * if the hpet becomes enabled later.
1249		 */
1250		hpet = is_hpet_enabled();
 
1251		schedule_delayed_work(&tsc_irqwork, HZ);
1252		tsc_start = tsc_read_refs(&ref_start, hpet);
1253		return;
1254	}
1255
1256	tsc_stop = tsc_read_refs(&ref_stop, hpet);
1257
1258	/* hpet or pmtimer available ? */
1259	if (ref_start == ref_stop)
1260		goto out;
1261
1262	/* Check, whether the sampling was disturbed by an SMI */
1263	if (tsc_start == ULLONG_MAX || tsc_stop == ULLONG_MAX)
1264		goto out;
1265
1266	delta = tsc_stop - tsc_start;
1267	delta *= 1000000LL;
1268	if (hpet)
1269		freq = calc_hpet_ref(delta, ref_start, ref_stop);
1270	else
1271		freq = calc_pmtimer_ref(delta, ref_start, ref_stop);
1272
1273	/* Make sure we're within 1% */
1274	if (abs(tsc_khz - freq) > tsc_khz/100)
1275		goto out;
1276
1277	tsc_khz = freq;
1278	pr_info("Refined TSC clocksource calibration: %lu.%03lu MHz\n",
1279		(unsigned long)tsc_khz / 1000,
1280		(unsigned long)tsc_khz % 1000);
1281
1282	/* Inform the TSC deadline clockevent devices about the recalibration */
1283	lapic_update_tsc_freq();
1284
 
 
 
 
1285out:
 
 
 
1286	if (boot_cpu_has(X86_FEATURE_ART))
1287		art_related_clocksource = &clocksource_tsc;
1288	clocksource_register_khz(&clocksource_tsc, tsc_khz);
 
 
1289}
1290
1291
1292static int __init init_tsc_clocksource(void)
1293{
1294	if (!boot_cpu_has(X86_FEATURE_TSC) || tsc_disabled > 0 || !tsc_khz)
1295		return 0;
1296
1297	if (tsc_clocksource_reliable)
 
 
 
1298		clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
1299	/* lower the rating if we already know its unstable: */
1300	if (check_tsc_unstable()) {
1301		clocksource_tsc.rating = 0;
1302		clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
1303	}
1304
1305	if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC_S3))
1306		clocksource_tsc.flags |= CLOCK_SOURCE_SUSPEND_NONSTOP;
1307
1308	/*
1309	 * When TSC frequency is known (retrieved via MSR or CPUID), we skip
1310	 * the refined calibration and directly register it as a clocksource.
1311	 */
1312	if (boot_cpu_has(X86_FEATURE_TSC_KNOWN_FREQ)) {
1313		if (boot_cpu_has(X86_FEATURE_ART))
1314			art_related_clocksource = &clocksource_tsc;
1315		clocksource_register_khz(&clocksource_tsc, tsc_khz);
 
 
1316		return 0;
1317	}
1318
1319	schedule_delayed_work(&tsc_irqwork, 0);
1320	return 0;
1321}
1322/*
1323 * We use device_initcall here, to ensure we run after the hpet
1324 * is fully initialized, which may occur at fs_initcall time.
1325 */
1326device_initcall(init_tsc_clocksource);
1327
1328void __init tsc_init(void)
1329{
1330	u64 lpj;
1331	int cpu;
1332
1333	if (!boot_cpu_has(X86_FEATURE_TSC)) {
1334		setup_clear_cpu_cap(X86_FEATURE_TSC_DEADLINE_TIMER);
1335		return;
 
 
 
 
1336	}
1337
1338	cpu_khz = x86_platform.calibrate_cpu();
1339	tsc_khz = x86_platform.calibrate_tsc();
1340
1341	/*
1342	 * Trust non-zero tsc_khz as authorative,
1343	 * and use it to sanity check cpu_khz,
1344	 * which will be off if system timer is off.
1345	 */
1346	if (tsc_khz == 0)
1347		tsc_khz = cpu_khz;
1348	else if (abs(cpu_khz - tsc_khz) * 10 > tsc_khz)
1349		cpu_khz = tsc_khz;
1350
1351	if (!tsc_khz) {
1352		mark_tsc_unstable("could not calculate TSC khz");
1353		setup_clear_cpu_cap(X86_FEATURE_TSC_DEADLINE_TIMER);
1354		return;
 
 
 
 
 
 
 
1355	}
 
 
1356
1357	pr_info("Detected %lu.%03lu MHz processor\n",
1358		(unsigned long)cpu_khz / 1000,
1359		(unsigned long)cpu_khz % 1000);
 
 
 
 
1360
 
 
1361	/* Sanitize TSC ADJUST before cyc2ns gets initialized */
1362	tsc_store_and_check_tsc_adjust(true);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1363
 
 
1364	/*
1365	 * Secondary CPUs do not run through tsc_init(), so set up
1366	 * all the scale factors for all CPUs, assuming the same
1367	 * speed as the bootup CPU. (cpufreq notifiers will fix this
1368	 * up if their speed diverges)
1369	 */
1370	for_each_possible_cpu(cpu) {
1371		cyc2ns_init(cpu);
1372		set_cyc2ns_scale(tsc_khz, cpu);
1373	}
1374
1375	if (tsc_disabled > 0)
 
1376		return;
 
1377
1378	/* now allow native_sched_clock() to use rdtsc */
 
 
 
 
 
 
 
 
1379
1380	tsc_disabled = 0;
1381	static_branch_enable(&__use_tsc);
1382
1383	if (!no_sched_irq_time)
1384		enable_sched_clock_irqtime();
1385
1386	lpj = ((u64)tsc_khz * 1000);
1387	do_div(lpj, HZ);
1388	lpj_fine = lpj;
1389
1390	use_tsc_delay();
1391
1392	if (unsynchronized_tsc())
1393		mark_tsc_unstable("TSCs unsynchronized");
 
 
1394
1395	check_system_tsc_reliable();
 
1396
 
1397	detect_art();
1398}
1399
1400#ifdef CONFIG_SMP
1401/*
1402 * If we have a constant TSC and are using the TSC for the delay loop,
1403 * we can skip clock calibration if another cpu in the same socket has already
1404 * been calibrated. This assumes that CONSTANT_TSC applies to all
1405 * cpus in the socket - this should be a safe assumption.
1406 */
1407unsigned long calibrate_delay_is_known(void)
1408{
1409	int sibling, cpu = smp_processor_id();
1410	struct cpumask *mask = topology_core_cpumask(cpu);
1411
1412	if (!tsc_disabled && !cpu_has(&cpu_data(cpu), X86_FEATURE_CONSTANT_TSC))
1413		return 0;
1414
1415	if (!mask)
1416		return 0;
1417
1418	sibling = cpumask_any_but(mask, cpu);
1419	if (sibling < nr_cpu_ids)
1420		return cpu_data(sibling).loops_per_jiffy;
1421	return 0;
1422}
1423#endif