linux/arch/x86/kernel/tsc.c
<<
>>
Prefs
   1#include <linux/kernel.h>
   2#include <linux/sched.h>
   3#include <linux/init.h>
   4#include <linux/module.h>
   5#include <linux/timer.h>
   6#include <linux/acpi_pmtmr.h>
   7#include <linux/cpufreq.h>
   8#include <linux/dmi.h>
   9#include <linux/delay.h>
  10#include <linux/clocksource.h>
  11#include <linux/percpu.h>
  12
  13#include <asm/hpet.h>
  14#include <asm/timer.h>
  15#include <asm/vgtod.h>
  16#include <asm/time.h>
  17#include <asm/delay.h>
  18
  19unsigned int cpu_khz;           /* TSC clocks / usec, not used here */
  20EXPORT_SYMBOL(cpu_khz);
  21unsigned int tsc_khz;
  22EXPORT_SYMBOL(tsc_khz);
  23
  24/*
  25 * TSC can be unstable due to cpufreq or due to unsynced TSCs
  26 */
  27static int tsc_unstable;
  28
  29/* native_sched_clock() is called before tsc_init(), so
  30   we must start with the TSC soft disabled to prevent
  31   erroneous rdtsc usage on !cpu_has_tsc processors */
  32static int tsc_disabled = -1;
  33
  34/*
  35 * Scheduler clock - returns current time in nanosec units.
  36 */
  37u64 native_sched_clock(void)
  38{
  39        u64 this_offset;
  40
  41        /*
  42         * Fall back to jiffies if there's no TSC available:
  43         * ( But note that we still use it if the TSC is marked
  44         *   unstable. We do this because unlike Time Of Day,
  45         *   the scheduler clock tolerates small errors and it's
  46         *   very important for it to be as fast as the platform
  47         *   can achive it. )
  48         */
  49        if (unlikely(tsc_disabled)) {
  50                /* No locking but a rare wrong value is not a big deal: */
  51                return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
  52        }
  53
  54        /* read the Time Stamp Counter: */
  55        rdtscll(this_offset);
  56
  57        /* return the value in ns */
  58        return __cycles_2_ns(this_offset);
  59}
  60
  61/* We need to define a real function for sched_clock, to override the
  62   weak default version */
  63#ifdef CONFIG_PARAVIRT
  64unsigned long long sched_clock(void)
  65{
  66        return paravirt_sched_clock();
  67}
  68#else
  69unsigned long long
  70sched_clock(void) __attribute__((alias("native_sched_clock")));
  71#endif
  72
  73int check_tsc_unstable(void)
  74{
  75        return tsc_unstable;
  76}
  77EXPORT_SYMBOL_GPL(check_tsc_unstable);
  78
  79#ifdef CONFIG_X86_TSC
  80int __init notsc_setup(char *str)
  81{
  82        printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, "
  83                        "cannot disable TSC completely.\n");
  84        tsc_disabled = 1;
  85        return 1;
  86}
  87#else
  88/*
  89 * disable flag for tsc. Takes effect by clearing the TSC cpu flag
  90 * in cpu/common.c
  91 */
  92int __init notsc_setup(char *str)
  93{
  94        setup_clear_cpu_cap(X86_FEATURE_TSC);
  95        return 1;
  96}
  97#endif
  98
  99__setup("notsc", notsc_setup);
 100
 101#define MAX_RETRIES     5
 102#define SMI_TRESHOLD    50000
 103
 104/*
 105 * Read TSC and the reference counters. Take care of SMI disturbance
 106 */
 107static u64 tsc_read_refs(u64 *p, int hpet)
 108{
 109        u64 t1, t2;
 110        int i;
 111
 112        for (i = 0; i < MAX_RETRIES; i++) {
 113                t1 = get_cycles();
 114                if (hpet)
 115                        *p = hpet_readl(HPET_COUNTER) & 0xFFFFFFFF;
 116                else
 117                        *p = acpi_pm_read_early();
 118                t2 = get_cycles();
 119                if ((t2 - t1) < SMI_TRESHOLD)
 120                        return t2;
 121        }
 122        return ULLONG_MAX;
 123}
 124
 125/*
 126 * Calculate the TSC frequency from HPET reference
 127 */
 128static unsigned long calc_hpet_ref(u64 deltatsc, u64 hpet1, u64 hpet2)
 129{
 130        u64 tmp;
 131
 132        if (hpet2 < hpet1)
 133                hpet2 += 0x100000000ULL;
 134        hpet2 -= hpet1;
 135        tmp = ((u64)hpet2 * hpet_readl(HPET_PERIOD));
 136        do_div(tmp, 1000000);
 137        do_div(deltatsc, tmp);
 138
 139        return (unsigned long) deltatsc;
 140}
 141
 142/*
 143 * Calculate the TSC frequency from PMTimer reference
 144 */
 145static unsigned long calc_pmtimer_ref(u64 deltatsc, u64 pm1, u64 pm2)
 146{
 147        u64 tmp;
 148
 149        if (!pm1 && !pm2)
 150                return ULONG_MAX;
 151
 152        if (pm2 < pm1)
 153                pm2 += (u64)ACPI_PM_OVRRUN;
 154        pm2 -= pm1;
 155        tmp = pm2 * 1000000000LL;
 156        do_div(tmp, PMTMR_TICKS_PER_SEC);
 157        do_div(deltatsc, tmp);
 158
 159        return (unsigned long) deltatsc;
 160}
 161
 162#define CAL_MS          10
 163#define CAL_LATCH       (CLOCK_TICK_RATE / (1000 / CAL_MS))
 164#define CAL_PIT_LOOPS   1000
 165
 166#define CAL2_MS         50
 167#define CAL2_LATCH      (CLOCK_TICK_RATE / (1000 / CAL2_MS))
 168#define CAL2_PIT_LOOPS  5000
 169
 170
 171/*
 172 * Try to calibrate the TSC against the Programmable
 173 * Interrupt Timer and return the frequency of the TSC
 174 * in kHz.
 175 *
 176 * Return ULONG_MAX on failure to calibrate.
 177 */
 178static unsigned long pit_calibrate_tsc(u32 latch, unsigned long ms, int loopmin)
 179{
 180        u64 tsc, t1, t2, delta;
 181        unsigned long tscmin, tscmax;
 182        int pitcnt;
 183
 184        /* Set the Gate high, disable speaker */
 185        outb((inb(0x61) & ~0x02) | 0x01, 0x61);
 186
 187        /*
 188         * Setup CTC channel 2* for mode 0, (interrupt on terminal
 189         * count mode), binary count. Set the latch register to 50ms
 190         * (LSB then MSB) to begin countdown.
 191         */
 192        outb(0xb0, 0x43);
 193        outb(latch & 0xff, 0x42);
 194        outb(latch >> 8, 0x42);
 195
 196        tsc = t1 = t2 = get_cycles();
 197
 198        pitcnt = 0;
 199        tscmax = 0;
 200        tscmin = ULONG_MAX;
 201        while ((inb(0x61) & 0x20) == 0) {
 202                t2 = get_cycles();
 203                delta = t2 - tsc;
 204                tsc = t2;
 205                if ((unsigned long) delta < tscmin)
 206                        tscmin = (unsigned int) delta;
 207                if ((unsigned long) delta > tscmax)
 208                        tscmax = (unsigned int) delta;
 209                pitcnt++;
 210        }
 211
 212        /*
 213         * Sanity checks:
 214         *
 215         * If we were not able to read the PIT more than loopmin
 216         * times, then we have been hit by a massive SMI
 217         *
 218         * If the maximum is 10 times larger than the minimum,
 219         * then we got hit by an SMI as well.
 220         */
 221        if (pitcnt < loopmin || tscmax > 10 * tscmin)
 222                return ULONG_MAX;
 223
 224        /* Calculate the PIT value */
 225        delta = t2 - t1;
 226        do_div(delta, ms);
 227        return delta;
 228}
 229
 230/*
 231 * This reads the current MSB of the PIT counter, and
 232 * checks if we are running on sufficiently fast and
 233 * non-virtualized hardware.
 234 *
 235 * Our expectations are:
 236 *
 237 *  - the PIT is running at roughly 1.19MHz
 238 *
 239 *  - each IO is going to take about 1us on real hardware,
 240 *    but we allow it to be much faster (by a factor of 10) or
 241 *    _slightly_ slower (ie we allow up to a 2us read+counter
 242 *    update - anything else implies a unacceptably slow CPU
 243 *    or PIT for the fast calibration to work.
 244 *
 245 *  - with 256 PIT ticks to read the value, we have 214us to
 246 *    see the same MSB (and overhead like doing a single TSC
 247 *    read per MSB value etc).
 248 *
 249 *  - We're doing 2 reads per loop (LSB, MSB), and we expect
 250 *    them each to take about a microsecond on real hardware.
 251 *    So we expect a count value of around 100. But we'll be
 252 *    generous, and accept anything over 50.
 253 *
 254 *  - if the PIT is stuck, and we see *many* more reads, we
 255 *    return early (and the next caller of pit_expect_msb()
 256 *    then consider it a failure when they don't see the
 257 *    next expected value).
 258 *
 259 * These expectations mean that we know that we have seen the
 260 * transition from one expected value to another with a fairly
 261 * high accuracy, and we didn't miss any events. We can thus
 262 * use the TSC value at the transitions to calculate a pretty
 263 * good value for the TSC frequencty.
 264 */
 265static inline int pit_expect_msb(unsigned char val)
 266{
 267        int count = 0;
 268
 269        for (count = 0; count < 50000; count++) {
 270                /* Ignore LSB */
 271                inb(0x42);
 272                if (inb(0x42) != val)
 273                        break;
 274        }
 275        return count > 50;
 276}
 277
 278/*
 279 * How many MSB values do we want to see? We aim for a
 280 * 15ms calibration, which assuming a 2us counter read
 281 * error should give us roughly 150 ppm precision for
 282 * the calibration.
 283 */
 284#define QUICK_PIT_MS 15
 285#define QUICK_PIT_ITERATIONS (QUICK_PIT_MS * PIT_TICK_RATE / 1000 / 256)
 286
 287static unsigned long quick_pit_calibrate(void)
 288{
 289        /* Set the Gate high, disable speaker */
 290        outb((inb(0x61) & ~0x02) | 0x01, 0x61);
 291
 292        /*
 293         * Counter 2, mode 0 (one-shot), binary count
 294         *
 295         * NOTE! Mode 2 decrements by two (and then the
 296         * output is flipped each time, giving the same
 297         * final output frequency as a decrement-by-one),
 298         * so mode 0 is much better when looking at the
 299         * individual counts.
 300         */
 301        outb(0xb0, 0x43);
 302
 303        /* Start at 0xffff */
 304        outb(0xff, 0x42);
 305        outb(0xff, 0x42);
 306
 307        if (pit_expect_msb(0xff)) {
 308                int i;
 309                u64 t1, t2, delta;
 310                unsigned char expect = 0xfe;
 311
 312                t1 = get_cycles();
 313                for (i = 0; i < QUICK_PIT_ITERATIONS; i++, expect--) {
 314                        if (!pit_expect_msb(expect))
 315                                goto failed;
 316                }
 317                t2 = get_cycles();
 318
 319                /*
 320                 * Make sure we can rely on the second TSC timestamp:
 321                 */
 322                if (!pit_expect_msb(expect))
 323                        goto failed;
 324
 325                /*
 326                 * Ok, if we get here, then we've seen the
 327                 * MSB of the PIT decrement QUICK_PIT_ITERATIONS
 328                 * times, and each MSB had many hits, so we never
 329                 * had any sudden jumps.
 330                 *
 331                 * As a result, we can depend on there not being
 332                 * any odd delays anywhere, and the TSC reads are
 333                 * reliable.
 334                 *
 335                 * kHz = ticks / time-in-seconds / 1000;
 336                 * kHz = (t2 - t1) / (QPI * 256 / PIT_TICK_RATE) / 1000
 337                 * kHz = ((t2 - t1) * PIT_TICK_RATE) / (QPI * 256 * 1000)
 338                 */
 339                delta = (t2 - t1)*PIT_TICK_RATE;
 340                do_div(delta, QUICK_PIT_ITERATIONS*256*1000);
 341                printk("Fast TSC calibration using PIT\n");
 342                return delta;
 343        }
 344failed:
 345        return 0;
 346}
 347
 348/**
 349 * native_calibrate_tsc - calibrate the tsc on boot
 350 */
 351unsigned long native_calibrate_tsc(void)
 352{
 353        u64 tsc1, tsc2, delta, ref1, ref2;
 354        unsigned long tsc_pit_min = ULONG_MAX, tsc_ref_min = ULONG_MAX;
 355        unsigned long flags, latch, ms, fast_calibrate;
 356        int hpet = is_hpet_enabled(), i, loopmin;
 357
 358        local_irq_save(flags);
 359        fast_calibrate = quick_pit_calibrate();
 360        local_irq_restore(flags);
 361        if (fast_calibrate)
 362                return fast_calibrate;
 363
 364        /*
 365         * Run 5 calibration loops to get the lowest frequency value
 366         * (the best estimate). We use two different calibration modes
 367         * here:
 368         *
 369         * 1) PIT loop. We set the PIT Channel 2 to oneshot mode and
 370         * load a timeout of 50ms. We read the time right after we
 371         * started the timer and wait until the PIT count down reaches
 372         * zero. In each wait loop iteration we read the TSC and check
 373         * the delta to the previous read. We keep track of the min
 374         * and max values of that delta. The delta is mostly defined
 375         * by the IO time of the PIT access, so we can detect when a
 376         * SMI/SMM disturbance happend between the two reads. If the
 377         * maximum time is significantly larger than the minimum time,
 378         * then we discard the result and have another try.
 379         *
 380         * 2) Reference counter. If available we use the HPET or the
 381         * PMTIMER as a reference to check the sanity of that value.
 382         * We use separate TSC readouts and check inside of the
 383         * reference read for a SMI/SMM disturbance. We dicard
 384         * disturbed values here as well. We do that around the PIT
 385         * calibration delay loop as we have to wait for a certain
 386         * amount of time anyway.
 387         */
 388
 389        /* Preset PIT loop values */
 390        latch = CAL_LATCH;
 391        ms = CAL_MS;
 392        loopmin = CAL_PIT_LOOPS;
 393
 394        for (i = 0; i < 3; i++) {
 395                unsigned long tsc_pit_khz;
 396
 397                /*
 398                 * Read the start value and the reference count of
 399                 * hpet/pmtimer when available. Then do the PIT
 400                 * calibration, which will take at least 50ms, and
 401                 * read the end value.
 402                 */
 403                local_irq_save(flags);
 404                tsc1 = tsc_read_refs(&ref1, hpet);
 405                tsc_pit_khz = pit_calibrate_tsc(latch, ms, loopmin);
 406                tsc2 = tsc_read_refs(&ref2, hpet);
 407                local_irq_restore(flags);
 408
 409                /* Pick the lowest PIT TSC calibration so far */
 410                tsc_pit_min = min(tsc_pit_min, tsc_pit_khz);
 411
 412                /* hpet or pmtimer available ? */
 413                if (!hpet && !ref1 && !ref2)
 414                        continue;
 415
 416                /* Check, whether the sampling was disturbed by an SMI */
 417                if (tsc1 == ULLONG_MAX || tsc2 == ULLONG_MAX)
 418                        continue;
 419
 420                tsc2 = (tsc2 - tsc1) * 1000000LL;
 421                if (hpet)
 422                        tsc2 = calc_hpet_ref(tsc2, ref1, ref2);
 423                else
 424                        tsc2 = calc_pmtimer_ref(tsc2, ref1, ref2);
 425
 426                tsc_ref_min = min(tsc_ref_min, (unsigned long) tsc2);
 427
 428                /* Check the reference deviation */
 429                delta = ((u64) tsc_pit_min) * 100;
 430                do_div(delta, tsc_ref_min);
 431
 432                /*
 433                 * If both calibration results are inside a 10% window
 434                 * then we can be sure, that the calibration
 435                 * succeeded. We break out of the loop right away. We
 436                 * use the reference value, as it is more precise.
 437                 */
 438                if (delta >= 90 && delta <= 110) {
 439                        printk(KERN_INFO
 440                               "TSC: PIT calibration matches %s. %d loops\n",
 441                               hpet ? "HPET" : "PMTIMER", i + 1);
 442                        return tsc_ref_min;
 443                }
 444
 445                /*
 446                 * Check whether PIT failed more than once. This
 447                 * happens in virtualized environments. We need to
 448                 * give the virtual PC a slightly longer timeframe for
 449                 * the HPET/PMTIMER to make the result precise.
 450                 */
 451                if (i == 1 && tsc_pit_min == ULONG_MAX) {
 452                        latch = CAL2_LATCH;
 453                        ms = CAL2_MS;
 454                        loopmin = CAL2_PIT_LOOPS;
 455                }
 456        }
 457
 458        /*
 459         * Now check the results.
 460         */
 461        if (tsc_pit_min == ULONG_MAX) {
 462                /* PIT gave no useful value */
 463                printk(KERN_WARNING "TSC: Unable to calibrate against PIT\n");
 464
 465                /* We don't have an alternative source, disable TSC */
 466                if (!hpet && !ref1 && !ref2) {
 467                        printk("TSC: No reference (HPET/PMTIMER) available\n");
 468                        return 0;
 469                }
 470
 471                /* The alternative source failed as well, disable TSC */
 472                if (tsc_ref_min == ULONG_MAX) {
 473                        printk(KERN_WARNING "TSC: HPET/PMTIMER calibration "
 474                               "failed.\n");
 475                        return 0;
 476                }
 477
 478                /* Use the alternative source */
 479                printk(KERN_INFO "TSC: using %s reference calibration\n",
 480                       hpet ? "HPET" : "PMTIMER");
 481
 482                return tsc_ref_min;
 483        }
 484
 485        /* We don't have an alternative source, use the PIT calibration value */
 486        if (!hpet && !ref1 && !ref2) {
 487                printk(KERN_INFO "TSC: Using PIT calibration value\n");
 488                return tsc_pit_min;
 489        }
 490
 491        /* The alternative source failed, use the PIT calibration value */
 492        if (tsc_ref_min == ULONG_MAX) {
 493                printk(KERN_WARNING "TSC: HPET/PMTIMER calibration failed. "
 494                       "Using PIT calibration\n");
 495                return tsc_pit_min;
 496        }
 497
 498        /*
 499         * The calibration values differ too much. In doubt, we use
 500         * the PIT value as we know that there are PMTIMERs around
 501         * running at double speed. At least we let the user know:
 502         */
 503        printk(KERN_WARNING "TSC: PIT calibration deviates from %s: %lu %lu.\n",
 504               hpet ? "HPET" : "PMTIMER", tsc_pit_min, tsc_ref_min);
 505        printk(KERN_INFO "TSC: Using PIT calibration value\n");
 506        return tsc_pit_min;
 507}
 508
 509#ifdef CONFIG_X86_32
 510/* Only called from the Powernow K7 cpu freq driver */
 511int recalibrate_cpu_khz(void)
 512{
 513#ifndef CONFIG_SMP
 514        unsigned long cpu_khz_old = cpu_khz;
 515
 516        if (cpu_has_tsc) {
 517                tsc_khz = calibrate_tsc();
 518                cpu_khz = tsc_khz;
 519                cpu_data(0).loops_per_jiffy =
 520                        cpufreq_scale(cpu_data(0).loops_per_jiffy,
 521                                        cpu_khz_old, cpu_khz);
 522                return 0;
 523        } else
 524                return -ENODEV;
 525#else
 526        return -ENODEV;
 527#endif
 528}
 529
 530EXPORT_SYMBOL(recalibrate_cpu_khz);
 531
 532#endif /* CONFIG_X86_32 */
 533
 534/* Accelerators for sched_clock()
 535 * convert from cycles(64bits) => nanoseconds (64bits)
 536 *  basic equation:
 537 *              ns = cycles / (freq / ns_per_sec)
 538 *              ns = cycles * (ns_per_sec / freq)
 539 *              ns = cycles * (10^9 / (cpu_khz * 10^3))
 540 *              ns = cycles * (10^6 / cpu_khz)
 541 *
 542 *      Then we use scaling math (suggested by george@mvista.com) to get:
 543 *              ns = cycles * (10^6 * SC / cpu_khz) / SC
 544 *              ns = cycles * cyc2ns_scale / SC
 545 *
 546 *      And since SC is a constant power of two, we can convert the div
 547 *  into a shift.
 548 *
 549 *  We can use khz divisor instead of mhz to keep a better precision, since
 550 *  cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
 551 *  (mathieu.desnoyers@polymtl.ca)
 552 *
 553 *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
 554 */
 555
 556DEFINE_PER_CPU(unsigned long, cyc2ns);
 557
 558static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu)
 559{
 560        unsigned long long tsc_now, ns_now;
 561        unsigned long flags, *scale;
 562
 563        local_irq_save(flags);
 564        sched_clock_idle_sleep_event();
 565
 566        scale = &per_cpu(cyc2ns, cpu);
 567
 568        rdtscll(tsc_now);
 569        ns_now = __cycles_2_ns(tsc_now);
 570
 571        if (cpu_khz)
 572                *scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz;
 573
 574        sched_clock_idle_wakeup_event(0);
 575        local_irq_restore(flags);
 576}
 577
 578#ifdef CONFIG_CPU_FREQ
 579
 580/* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
 581 * changes.
 582 *
 583 * RED-PEN: On SMP we assume all CPUs run with the same frequency.  It's
 584 * not that important because current Opteron setups do not support
 585 * scaling on SMP anyroads.
 586 *
 587 * Should fix up last_tsc too. Currently gettimeofday in the
 588 * first tick after the change will be slightly wrong.
 589 */
 590
 591static unsigned int  ref_freq;
 592static unsigned long loops_per_jiffy_ref;
 593static unsigned long tsc_khz_ref;
 594
 595static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
 596                                void *data)
 597{
 598        struct cpufreq_freqs *freq = data;
 599        unsigned long *lpj, dummy;
 600
 601        if (cpu_has(&cpu_data(freq->cpu), X86_FEATURE_CONSTANT_TSC))
 602                return 0;
 603
 604        lpj = &dummy;
 605        if (!(freq->flags & CPUFREQ_CONST_LOOPS))
 606#ifdef CONFIG_SMP
 607                lpj = &cpu_data(freq->cpu).loops_per_jiffy;
 608#else
 609        lpj = &boot_cpu_data.loops_per_jiffy;
 610#endif
 611
 612        if (!ref_freq) {
 613                ref_freq = freq->old;
 614                loops_per_jiffy_ref = *lpj;
 615                tsc_khz_ref = tsc_khz;
 616        }
 617        if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
 618                        (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
 619                        (val == CPUFREQ_RESUMECHANGE)) {
 620                *lpj =  cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
 621
 622                tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
 623                if (!(freq->flags & CPUFREQ_CONST_LOOPS))
 624                        mark_tsc_unstable("cpufreq changes");
 625        }
 626
 627        set_cyc2ns_scale(tsc_khz, freq->cpu);
 628
 629        return 0;
 630}
 631
 632static struct notifier_block time_cpufreq_notifier_block = {
 633        .notifier_call  = time_cpufreq_notifier
 634};
 635
 636static int __init cpufreq_tsc(void)
 637{
 638        if (!cpu_has_tsc)
 639                return 0;
 640        if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
 641                return 0;
 642        cpufreq_register_notifier(&time_cpufreq_notifier_block,
 643                                CPUFREQ_TRANSITION_NOTIFIER);
 644        return 0;
 645}
 646
 647core_initcall(cpufreq_tsc);
 648
 649#endif /* CONFIG_CPU_FREQ */
 650
 651/* clocksource code */
 652
 653static struct clocksource clocksource_tsc;
 654
 655/*
 656 * We compare the TSC to the cycle_last value in the clocksource
 657 * structure to avoid a nasty time-warp. This can be observed in a
 658 * very small window right after one CPU updated cycle_last under
 659 * xtime/vsyscall_gtod lock and the other CPU reads a TSC value which
 660 * is smaller than the cycle_last reference value due to a TSC which
 661 * is slighty behind. This delta is nowhere else observable, but in
 662 * that case it results in a forward time jump in the range of hours
 663 * due to the unsigned delta calculation of the time keeping core
 664 * code, which is necessary to support wrapping clocksources like pm
 665 * timer.
 666 */
 667static cycle_t read_tsc(void)
 668{
 669        cycle_t ret = (cycle_t)get_cycles();
 670
 671        return ret >= clocksource_tsc.cycle_last ?
 672                ret : clocksource_tsc.cycle_last;
 673}
 674
 675#ifdef CONFIG_X86_64
 676static cycle_t __vsyscall_fn vread_tsc(void)
 677{
 678        cycle_t ret = (cycle_t)vget_cycles();
 679
 680        return ret >= __vsyscall_gtod_data.clock.cycle_last ?
 681                ret : __vsyscall_gtod_data.clock.cycle_last;
 682}
 683#endif
 684
 685static struct clocksource clocksource_tsc = {
 686        .name                   = "tsc",
 687        .rating                 = 300,
 688        .read                   = read_tsc,
 689        .mask                   = CLOCKSOURCE_MASK(64),
 690        .shift                  = 22,
 691        .flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
 692                                  CLOCK_SOURCE_MUST_VERIFY,
 693#ifdef CONFIG_X86_64
 694        .vread                  = vread_tsc,
 695#endif
 696};
 697
 698void mark_tsc_unstable(char *reason)
 699{
 700        if (!tsc_unstable) {
 701                tsc_unstable = 1;
 702                printk("Marking TSC unstable due to %s\n", reason);
 703                /* Change only the rating, when not registered */
 704                if (clocksource_tsc.mult)
 705                        clocksource_change_rating(&clocksource_tsc, 0);
 706                else
 707                        clocksource_tsc.rating = 0;
 708        }
 709}
 710
 711EXPORT_SYMBOL_GPL(mark_tsc_unstable);
 712
 713static int __init dmi_mark_tsc_unstable(const struct dmi_system_id *d)
 714{
 715        printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
 716                        d->ident);
 717        tsc_unstable = 1;
 718        return 0;
 719}
 720
 721/* List of systems that have known TSC problems */
 722static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
 723        {
 724                .callback = dmi_mark_tsc_unstable,
 725                .ident = "IBM Thinkpad 380XD",
 726                .matches = {
 727                        DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
 728                        DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
 729                },
 730        },
 731        {}
 732};
 733
 734/*
 735 * Geode_LX - the OLPC CPU has a possibly a very reliable TSC
 736 */
 737#ifdef CONFIG_MGEODE_LX
 738/* RTSC counts during suspend */
 739#define RTSC_SUSP 0x100
 740
 741static void __init check_geode_tsc_reliable(void)
 742{
 743        unsigned long res_low, res_high;
 744
 745        rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
 746        if (res_low & RTSC_SUSP)
 747                clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
 748}
 749#else
 750static inline void check_geode_tsc_reliable(void) { }
 751#endif
 752
 753/*
 754 * Make an educated guess if the TSC is trustworthy and synchronized
 755 * over all CPUs.
 756 */
 757__cpuinit int unsynchronized_tsc(void)
 758{
 759        if (!cpu_has_tsc || tsc_unstable)
 760                return 1;
 761
 762#ifdef CONFIG_X86_SMP
 763        if (apic_is_clustered_box())
 764                return 1;
 765#endif
 766
 767        if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
 768                return 0;
 769        /*
 770         * Intel systems are normally all synchronized.
 771         * Exceptions must mark TSC as unstable:
 772         */
 773        if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
 774                /* assume multi socket systems are not synchronized: */
 775                if (num_possible_cpus() > 1)
 776                        tsc_unstable = 1;
 777        }
 778
 779        return tsc_unstable;
 780}
 781
 782static void __init init_tsc_clocksource(void)
 783{
 784        clocksource_tsc.mult = clocksource_khz2mult(tsc_khz,
 785                        clocksource_tsc.shift);
 786        /* lower the rating if we already know its unstable: */
 787        if (check_tsc_unstable()) {
 788                clocksource_tsc.rating = 0;
 789                clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
 790        }
 791        clocksource_register(&clocksource_tsc);
 792}
 793
 794void __init tsc_init(void)
 795{
 796        u64 lpj;
 797        int cpu;
 798
 799        if (!cpu_has_tsc)
 800                return;
 801
 802        tsc_khz = calibrate_tsc();
 803        cpu_khz = tsc_khz;
 804
 805        if (!tsc_khz) {
 806                mark_tsc_unstable("could not calculate TSC khz");
 807                return;
 808        }
 809
 810#ifdef CONFIG_X86_64
 811        if (cpu_has(&boot_cpu_data, X86_FEATURE_CONSTANT_TSC) &&
 812                        (boot_cpu_data.x86_vendor == X86_VENDOR_AMD))
 813                cpu_khz = calibrate_cpu();
 814#endif
 815
 816        printk("Detected %lu.%03lu MHz processor.\n",
 817                        (unsigned long)cpu_khz / 1000,
 818                        (unsigned long)cpu_khz % 1000);
 819
 820        /*
 821         * Secondary CPUs do not run through tsc_init(), so set up
 822         * all the scale factors for all CPUs, assuming the same
 823         * speed as the bootup CPU. (cpufreq notifiers will fix this
 824         * up if their speed diverges)
 825         */
 826        for_each_possible_cpu(cpu)
 827                set_cyc2ns_scale(cpu_khz, cpu);
 828
 829        if (tsc_disabled > 0)
 830                return;
 831
 832        /* now allow native_sched_clock() to use rdtsc */
 833        tsc_disabled = 0;
 834
 835        lpj = ((u64)tsc_khz * 1000);
 836        do_div(lpj, HZ);
 837        lpj_fine = lpj;
 838
 839        use_tsc_delay();
 840        /* Check and install the TSC clocksource */
 841        dmi_check_system(bad_tsc_dmi_table);
 842
 843        if (unsynchronized_tsc())
 844                mark_tsc_unstable("TSCs unsynchronized");
 845
 846        check_geode_tsc_reliable();
 847        init_tsc_clocksource();
 848}
 849
 850
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.