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