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#include <asm/hypervisor.h>
  19
  20unsigned int cpu_khz;           /* TSC clocks / usec, not used here */
  21EXPORT_SYMBOL(cpu_khz);
  22unsigned int tsc_khz;
  23EXPORT_SYMBOL(tsc_khz);
  24
  25/*
  26 * TSC can be unstable due to cpufreq or due to unsynced TSCs
  27 */
  28static int tsc_unstable;
  29
  30/* native_sched_clock() is called before tsc_init(), so
  31   we must start with the TSC soft disabled to prevent
  32   erroneous rdtsc usage on !cpu_has_tsc processors */
  33static int tsc_disabled = -1;
  34
  35static int tsc_clocksource_reliable;
  36/*
  37 * Scheduler clock - returns current time in nanosec units.
  38 */
  39u64 native_sched_clock(void)
  40{
  41        u64 this_offset;
  42
  43        /*
  44         * Fall back to jiffies if there's no TSC available:
  45         * ( But note that we still use it if the TSC is marked
  46         *   unstable. We do this because unlike Time Of Day,
  47         *   the scheduler clock tolerates small errors and it's
  48         *   very important for it to be as fast as the platform
  49         *   can achive it. )
  50         */
  51        if (unlikely(tsc_disabled)) {
  52                /* No locking but a rare wrong value is not a big deal: */
  53                return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
  54        }
  55
  56        /* read the Time Stamp Counter: */
  57        rdtscll(this_offset);
  58
  59        /* return the value in ns */
  60        return __cycles_2_ns(this_offset);
  61}
  62
  63/* We need to define a real function for sched_clock, to override the
  64   weak default version */
  65#ifdef CONFIG_PARAVIRT
  66unsigned long long sched_clock(void)
  67{
  68        return paravirt_sched_clock();
  69}
  70#else
  71unsigned long long
  72sched_clock(void) __attribute__((alias("native_sched_clock")));
  73#endif
  74
  75int check_tsc_unstable(void)
  76{
  77        return tsc_unstable;
  78}
  79EXPORT_SYMBOL_GPL(check_tsc_unstable);
  80
  81#ifdef CONFIG_X86_TSC
  82int __init notsc_setup(char *str)
  83{
  84        printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, "
  85                        "cannot disable TSC completely.\n");
  86        tsc_disabled = 1;
  87        return 1;
  88}
  89#else
  90/*
  91 * disable flag for tsc. Takes effect by clearing the TSC cpu flag
  92 * in cpu/common.c
  93 */
  94int __init notsc_setup(char *str)
  95{
  96        setup_clear_cpu_cap(X86_FEATURE_TSC);
  97        return 1;
  98}
  99#endif
 100
 101__setup("notsc", notsc_setup);
 102
 103static int __init tsc_setup(char *str)
 104{
 105        if (!strcmp(str, "reliable"))
 106                tsc_clocksource_reliable = 1;
 107        return 1;
 108}
 109
 110__setup("tsc=", tsc_setup);
 111
 112#define MAX_RETRIES     5
 113#define SMI_TRESHOLD    50000
 114
 115/*
 116 * Read TSC and the reference counters. Take care of SMI disturbance
 117 */
 118static u64 tsc_read_refs(u64 *p, int hpet)
 119{
 120        u64 t1, t2;
 121        int i;
 122
 123        for (i = 0; i < MAX_RETRIES; i++) {
 124                t1 = get_cycles();
 125                if (hpet)
 126                        *p = hpet_readl(HPET_COUNTER) & 0xFFFFFFFF;
 127                else
 128                        *p = acpi_pm_read_early();
 129                t2 = get_cycles();
 130                if ((t2 - t1) < SMI_TRESHOLD)
 131                        return t2;
 132        }
 133        return ULLONG_MAX;
 134}
 135
 136/*
 137 * Calculate the TSC frequency from HPET reference
 138 */
 139static unsigned long calc_hpet_ref(u64 deltatsc, u64 hpet1, u64 hpet2)
 140{
 141        u64 tmp;
 142
 143        if (hpet2 < hpet1)
 144                hpet2 += 0x100000000ULL;
 145        hpet2 -= hpet1;
 146        tmp = ((u64)hpet2 * hpet_readl(HPET_PERIOD));
 147        do_div(tmp, 1000000);
 148        do_div(deltatsc, tmp);
 149
 150        return (unsigned long) deltatsc;
 151}
 152
 153/*
 154 * Calculate the TSC frequency from PMTimer reference
 155 */
 156static unsigned long calc_pmtimer_ref(u64 deltatsc, u64 pm1, u64 pm2)
 157{
 158        u64 tmp;
 159
 160        if (!pm1 && !pm2)
 161                return ULONG_MAX;
 162
 163        if (pm2 < pm1)
 164                pm2 += (u64)ACPI_PM_OVRRUN;
 165        pm2 -= pm1;
 166        tmp = pm2 * 1000000000LL;
 167        do_div(tmp, PMTMR_TICKS_PER_SEC);
 168        do_div(deltatsc, tmp);
 169
 170        return (unsigned long) deltatsc;
 171}
 172
 173#define CAL_MS          10
 174#define CAL_LATCH       (CLOCK_TICK_RATE / (1000 / CAL_MS))
 175#define CAL_PIT_LOOPS   1000
 176
 177#define CAL2_MS         50
 178#define CAL2_LATCH      (CLOCK_TICK_RATE / (1000 / CAL2_MS))
 179#define CAL2_PIT_LOOPS  5000
 180
 181
 182/*
 183 * Try to calibrate the TSC against the Programmable
 184 * Interrupt Timer and return the frequency of the TSC
 185 * in kHz.
 186 *
 187 * Return ULONG_MAX on failure to calibrate.
 188 */
 189static unsigned long pit_calibrate_tsc(u32 latch, unsigned long ms, int loopmin)
 190{
 191        u64 tsc, t1, t2, delta;
 192        unsigned long tscmin, tscmax;
 193        int pitcnt;
 194
 195        /* Set the Gate high, disable speaker */
 196        outb((inb(0x61) & ~0x02) | 0x01, 0x61);
 197
 198        /*
 199         * Setup CTC channel 2* for mode 0, (interrupt on terminal
 200         * count mode), binary count. Set the latch register to 50ms
 201         * (LSB then MSB) to begin countdown.
 202         */
 203        outb(0xb0, 0x43);
 204        outb(latch & 0xff, 0x42);
 205        outb(latch >> 8, 0x42);
 206
 207        tsc = t1 = t2 = get_cycles();
 208
 209        pitcnt = 0;
 210        tscmax = 0;
 211        tscmin = ULONG_MAX;
 212        while ((inb(0x61) & 0x20) == 0) {
 213                t2 = get_cycles();
 214                delta = t2 - tsc;
 215                tsc = t2;
 216                if ((unsigned long) delta < tscmin)
 217                        tscmin = (unsigned int) delta;
 218                if ((unsigned long) delta > tscmax)
 219                        tscmax = (unsigned int) delta;
 220                pitcnt++;
 221        }
 222
 223        /*
 224         * Sanity checks:
 225         *
 226         * If we were not able to read the PIT more than loopmin
 227         * times, then we have been hit by a massive SMI
 228         *
 229         * If the maximum is 10 times larger than the minimum,
 230         * then we got hit by an SMI as well.
 231         */
 232        if (pitcnt < loopmin || tscmax > 10 * tscmin)
 233                return ULONG_MAX;
 234
 235        /* Calculate the PIT value */
 236        delta = t2 - t1;
 237        do_div(delta, ms);
 238        return delta;
 239}
 240
 241/*
 242 * This reads the current MSB of the PIT counter, and
 243 * checks if we are running on sufficiently fast and
 244 * non-virtualized hardware.
 245 *
 246 * Our expectations are:
 247 *
 248 *  - the PIT is running at roughly 1.19MHz
 249 *
 250 *  - each IO is going to take about 1us on real hardware,
 251 *    but we allow it to be much faster (by a factor of 10) or
 252 *    _slightly_ slower (ie we allow up to a 2us read+counter
 253 *    update - anything else implies a unacceptably slow CPU
 254 *    or PIT for the fast calibration to work.
 255 *
 256 *  - with 256 PIT ticks to read the value, we have 214us to
 257 *    see the same MSB (and overhead like doing a single TSC
 258 *    read per MSB value etc).
 259 *
 260 *  - We're doing 2 reads per loop (LSB, MSB), and we expect
 261 *    them each to take about a microsecond on real hardware.
 262 *    So we expect a count value of around 100. But we'll be
 263 *    generous, and accept anything over 50.
 264 *
 265 *  - if the PIT is stuck, and we see *many* more reads, we
 266 *    return early (and the next caller of pit_expect_msb()
 267 *    then consider it a failure when they don't see the
 268 *    next expected value).
 269 *
 270 * These expectations mean that we know that we have seen the
 271 * transition from one expected value to another with a fairly
 272 * high accuracy, and we didn't miss any events. We can thus
 273 * use the TSC value at the transitions to calculate a pretty
 274 * good value for the TSC frequencty.
 275 */
 276static inline int pit_expect_msb(unsigned char val, u64 *tscp, unsigned long *deltap)
 277{
 278        int count;
 279        u64 tsc = 0;
 280
 281        for (count = 0; count < 50000; count++) {
 282                /* Ignore LSB */
 283                inb(0x42);
 284                if (inb(0x42) != val)
 285                        break;
 286                tsc = get_cycles();
 287        }
 288        *deltap = get_cycles() - tsc;
 289        *tscp = tsc;
 290
 291        /*
 292         * We require _some_ success, but the quality control
 293         * will be based on the error terms on the TSC values.
 294         */
 295        return count > 5;
 296}
 297
 298/*
 299 * How many MSB values do we want to see? We aim for
 300 * a maximum error rate of 500ppm (in practice the
 301 * real error is much smaller), but refuse to spend
 302 * more than 25ms on it.
 303 */
 304#define MAX_QUICK_PIT_MS 25
 305#define MAX_QUICK_PIT_ITERATIONS (MAX_QUICK_PIT_MS * PIT_TICK_RATE / 1000 / 256)
 306
 307static unsigned long quick_pit_calibrate(void)
 308{
 309        int i;
 310        u64 tsc, delta;
 311        unsigned long d1, d2;
 312
 313        /* Set the Gate high, disable speaker */
 314        outb((inb(0x61) & ~0x02) | 0x01, 0x61);
 315
 316        /*
 317         * Counter 2, mode 0 (one-shot), binary count
 318         *
 319         * NOTE! Mode 2 decrements by two (and then the
 320         * output is flipped each time, giving the same
 321         * final output frequency as a decrement-by-one),
 322         * so mode 0 is much better when looking at the
 323         * individual counts.
 324         */
 325        outb(0xb0, 0x43);
 326
 327        /* Start at 0xffff */
 328        outb(0xff, 0x42);
 329        outb(0xff, 0x42);
 330
 331        /*
 332         * The PIT starts counting at the next edge, so we
 333         * need to delay for a microsecond. The easiest way
 334         * to do that is to just read back the 16-bit counter
 335         * once from the PIT.
 336         */
 337        inb(0x42);
 338        inb(0x42);
 339
 340        if (pit_expect_msb(0xff, &tsc, &d1)) {
 341                for (i = 1; i <= MAX_QUICK_PIT_ITERATIONS; i++) {
 342                        if (!pit_expect_msb(0xff-i, &delta, &d2))
 343                                break;
 344
 345                        /*
 346                         * Iterate until the error is less than 500 ppm
 347                         */
 348                        delta -= tsc;
 349                        if (d1+d2 < delta >> 11)
 350                                goto success;
 351                }
 352        }
 353        printk("Fast TSC calibration failed\n");
 354        return 0;
 355
 356success:
 357        /*
 358         * Ok, if we get here, then we've seen the
 359         * MSB of the PIT decrement 'i' times, and the
 360         * error has shrunk to less than 500 ppm.
 361         *
 362         * As a result, we can depend on there not being
 363         * any odd delays anywhere, and the TSC reads are
 364         * reliable (within the error). We also adjust the
 365         * delta to the middle of the error bars, just
 366         * because it looks nicer.
 367         *
 368         * kHz = ticks / time-in-seconds / 1000;
 369         * kHz = (t2 - t1) / (I * 256 / PIT_TICK_RATE) / 1000
 370         * kHz = ((t2 - t1) * PIT_TICK_RATE) / (I * 256 * 1000)
 371         */
 372        delta += (long)(d2 - d1)/2;
 373        delta *= PIT_TICK_RATE;
 374        do_div(delta, i*256*1000);
 375        printk("Fast TSC calibration using PIT\n");
 376        return delta;
 377}
 378
 379/**
 380 * native_calibrate_tsc - calibrate the tsc on boot
 381 */
 382unsigned long native_calibrate_tsc(void)
 383{
 384        u64 tsc1, tsc2, delta, ref1, ref2;
 385        unsigned long tsc_pit_min = ULONG_MAX, tsc_ref_min = ULONG_MAX;
 386        unsigned long flags, latch, ms, fast_calibrate, tsc_khz;
 387        int hpet = is_hpet_enabled(), i, loopmin;
 388
 389        tsc_khz = get_hypervisor_tsc_freq();
 390        if (tsc_khz) {
 391                printk(KERN_INFO "TSC: Frequency read from the hypervisor\n");
 392                return tsc_khz;
 393        }
 394
 395        local_irq_save(flags);
 396        fast_calibrate = quick_pit_calibrate();
 397        local_irq_restore(flags);
 398        if (fast_calibrate)
 399                return fast_calibrate;
 400
 401        /*
 402         * Run 5 calibration loops to get the lowest frequency value
 403         * (the best estimate). We use two different calibration modes
 404         * here:
 405         *
 406         * 1) PIT loop. We set the PIT Channel 2 to oneshot mode and
 407         * load a timeout of 50ms. We read the time right after we
 408         * started the timer and wait until the PIT count down reaches
 409         * zero. In each wait loop iteration we read the TSC and check
 410         * the delta to the previous read. We keep track of the min
 411         * and max values of that delta. The delta is mostly defined
 412         * by the IO time of the PIT access, so we can detect when a
 413         * SMI/SMM disturbance happend between the two reads. If the
 414         * maximum time is significantly larger than the minimum time,
 415         * then we discard the result and have another try.
 416         *
 417         * 2) Reference counter. If available we use the HPET or the
 418         * PMTIMER as a reference to check the sanity of that value.
 419         * We use separate TSC readouts and check inside of the
 420         * reference read for a SMI/SMM disturbance. We dicard
 421         * disturbed values here as well. We do that around the PIT
 422         * calibration delay loop as we have to wait for a certain
 423         * amount of time anyway.
 424         */
 425
 426        /* Preset PIT loop values */
 427        latch = CAL_LATCH;
 428        ms = CAL_MS;
 429        loopmin = CAL_PIT_LOOPS;
 430
 431        for (i = 0; i < 3; i++) {
 432                unsigned long tsc_pit_khz;
 433
 434                /*
 435                 * Read the start value and the reference count of
 436                 * hpet/pmtimer when available. Then do the PIT
 437                 * calibration, which will take at least 50ms, and
 438                 * read the end value.
 439                 */
 440                local_irq_save(flags);
 441                tsc1 = tsc_read_refs(&ref1, hpet);
 442                tsc_pit_khz = pit_calibrate_tsc(latch, ms, loopmin);
 443                tsc2 = tsc_read_refs(&ref2, hpet);
 444                local_irq_restore(flags);
 445
 446                /* Pick the lowest PIT TSC calibration so far */
 447                tsc_pit_min = min(tsc_pit_min, tsc_pit_khz);
 448
 449                /* hpet or pmtimer available ? */
 450                if (!hpet && !ref1 && !ref2)
 451                        continue;
 452
 453                /* Check, whether the sampling was disturbed by an SMI */
 454                if (tsc1 == ULLONG_MAX || tsc2 == ULLONG_MAX)
 455                        continue;
 456
 457                tsc2 = (tsc2 - tsc1) * 1000000LL;
 458                if (hpet)
 459                        tsc2 = calc_hpet_ref(tsc2, ref1, ref2);
 460                else
 461                        tsc2 = calc_pmtimer_ref(tsc2, ref1, ref2);
 462
 463                tsc_ref_min = min(tsc_ref_min, (unsigned long) tsc2);
 464
 465                /* Check the reference deviation */
 466                delta = ((u64) tsc_pit_min) * 100;
 467                do_div(delta, tsc_ref_min);
 468
 469                /*
 470                 * If both calibration results are inside a 10% window
 471                 * then we can be sure, that the calibration
 472                 * succeeded. We break out of the loop right away. We
 473                 * use the reference value, as it is more precise.
 474                 */
 475                if (delta >= 90 && delta <= 110) {
 476                        printk(KERN_INFO
 477                               "TSC: PIT calibration matches %s. %d loops\n",
 478                               hpet ? "HPET" : "PMTIMER", i + 1);
 479                        return tsc_ref_min;
 480                }
 481
 482                /*
 483                 * Check whether PIT failed more than once. This
 484                 * happens in virtualized environments. We need to
 485                 * give the virtual PC a slightly longer timeframe for
 486                 * the HPET/PMTIMER to make the result precise.
 487                 */
 488                if (i == 1 && tsc_pit_min == ULONG_MAX) {
 489                        latch = CAL2_LATCH;
 490                        ms = CAL2_MS;
 491                        loopmin = CAL2_PIT_LOOPS;
 492                }
 493        }
 494
 495        /*
 496         * Now check the results.
 497         */
 498        if (tsc_pit_min == ULONG_MAX) {
 499                /* PIT gave no useful value */
 500                printk(KERN_WARNING "TSC: Unable to calibrate against PIT\n");
 501
 502                /* We don't have an alternative source, disable TSC */
 503                if (!hpet && !ref1 && !ref2) {
 504                        printk("TSC: No reference (HPET/PMTIMER) available\n");
 505                        return 0;
 506                }
 507
 508                /* The alternative source failed as well, disable TSC */
 509                if (tsc_ref_min == ULONG_MAX) {
 510                        printk(KERN_WARNING "TSC: HPET/PMTIMER calibration "
 511                               "failed.\n");
 512                        return 0;
 513                }
 514
 515                /* Use the alternative source */
 516                printk(KERN_INFO "TSC: using %s reference calibration\n",
 517                       hpet ? "HPET" : "PMTIMER");
 518
 519                return tsc_ref_min;
 520        }
 521
 522        /* We don't have an alternative source, use the PIT calibration value */
 523        if (!hpet && !ref1 && !ref2) {
 524                printk(KERN_INFO "TSC: Using PIT calibration value\n");
 525                return tsc_pit_min;
 526        }
 527
 528        /* The alternative source failed, use the PIT calibration value */
 529        if (tsc_ref_min == ULONG_MAX) {
 530                printk(KERN_WARNING "TSC: HPET/PMTIMER calibration failed. "
 531                       "Using PIT calibration\n");
 532                return tsc_pit_min;
 533        }
 534
 535        /*
 536         * The calibration values differ too much. In doubt, we use
 537         * the PIT value as we know that there are PMTIMERs around
 538         * running at double speed. At least we let the user know:
 539         */
 540        printk(KERN_WARNING "TSC: PIT calibration deviates from %s: %lu %lu.\n",
 541               hpet ? "HPET" : "PMTIMER", tsc_pit_min, tsc_ref_min);
 542        printk(KERN_INFO "TSC: Using PIT calibration value\n");
 543        return tsc_pit_min;
 544}
 545
 546#ifdef CONFIG_X86_32
 547/* Only called from the Powernow K7 cpu freq driver */
 548int recalibrate_cpu_khz(void)
 549{
 550#ifndef CONFIG_SMP
 551        unsigned long cpu_khz_old = cpu_khz;
 552
 553        if (cpu_has_tsc) {
 554                tsc_khz = calibrate_tsc();
 555                cpu_khz = tsc_khz;
 556                cpu_data(0).loops_per_jiffy =
 557                        cpufreq_scale(cpu_data(0).loops_per_jiffy,
 558                                        cpu_khz_old, cpu_khz);
 559                return 0;
 560        } else
 561                return -ENODEV;
 562#else
 563        return -ENODEV;
 564#endif
 565}
 566
 567EXPORT_SYMBOL(recalibrate_cpu_khz);
 568
 569#endif /* CONFIG_X86_32 */
 570
 571/* Accelerators for sched_clock()
 572 * convert from cycles(64bits) => nanoseconds (64bits)
 573 *  basic equation:
 574 *              ns = cycles / (freq / ns_per_sec)
 575 *              ns = cycles * (ns_per_sec / freq)
 576 *              ns = cycles * (10^9 / (cpu_khz * 10^3))
 577 *              ns = cycles * (10^6 / cpu_khz)
 578 *
 579 *      Then we use scaling math (suggested by george@mvista.com) to get:
 580 *              ns = cycles * (10^6 * SC / cpu_khz) / SC
 581 *              ns = cycles * cyc2ns_scale / SC
 582 *
 583 *      And since SC is a constant power of two, we can convert the div
 584 *  into a shift.
 585 *
 586 *  We can use khz divisor instead of mhz to keep a better precision, since
 587 *  cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
 588 *  (mathieu.desnoyers@polymtl.ca)
 589 *
 590 *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
 591 */
 592
 593DEFINE_PER_CPU(unsigned long, cyc2ns);
 594
 595static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu)
 596{
 597        unsigned long long tsc_now, ns_now;
 598        unsigned long flags, *scale;
 599
 600        local_irq_save(flags);
 601        sched_clock_idle_sleep_event();
 602
 603        scale = &per_cpu(cyc2ns, cpu);
 604
 605        rdtscll(tsc_now);
 606        ns_now = __cycles_2_ns(tsc_now);
 607
 608        if (cpu_khz)
 609                *scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz;
 610
 611        sched_clock_idle_wakeup_event(0);
 612        local_irq_restore(flags);
 613}
 614
 615#ifdef CONFIG_CPU_FREQ
 616
 617/* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
 618 * changes.
 619 *
 620 * RED-PEN: On SMP we assume all CPUs run with the same frequency.  It's
 621 * not that important because current Opteron setups do not support
 622 * scaling on SMP anyroads.
 623 *
 624 * Should fix up last_tsc too. Currently gettimeofday in the
 625 * first tick after the change will be slightly wrong.
 626 */
 627
 628static unsigned int  ref_freq;
 629static unsigned long loops_per_jiffy_ref;
 630static unsigned long tsc_khz_ref;
 631
 632static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
 633                                void *data)
 634{
 635        struct cpufreq_freqs *freq = data;
 636        unsigned long *lpj, dummy;
 637
 638        if (cpu_has(&cpu_data(freq->cpu), X86_FEATURE_CONSTANT_TSC))
 639                return 0;
 640
 641        lpj = &dummy;
 642        if (!(freq->flags & CPUFREQ_CONST_LOOPS))
 643#ifdef CONFIG_SMP
 644                lpj = &cpu_data(freq->cpu).loops_per_jiffy;
 645#else
 646        lpj = &boot_cpu_data.loops_per_jiffy;
 647#endif
 648
 649        if (!ref_freq) {
 650                ref_freq = freq->old;
 651                loops_per_jiffy_ref = *lpj;
 652                tsc_khz_ref = tsc_khz;
 653        }
 654        if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
 655                        (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
 656                        (val == CPUFREQ_RESUMECHANGE)) {
 657                *lpj =  cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
 658
 659                tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
 660                if (!(freq->flags & CPUFREQ_CONST_LOOPS))
 661                        mark_tsc_unstable("cpufreq changes");
 662        }
 663
 664        set_cyc2ns_scale(tsc_khz, freq->cpu);
 665
 666        return 0;
 667}
 668
 669static struct notifier_block time_cpufreq_notifier_block = {
 670        .notifier_call  = time_cpufreq_notifier
 671};
 672
 673static int __init cpufreq_tsc(void)
 674{
 675        if (!cpu_has_tsc)
 676                return 0;
 677        if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
 678                return 0;
 679        cpufreq_register_notifier(&time_cpufreq_notifier_block,
 680                                CPUFREQ_TRANSITION_NOTIFIER);
 681        return 0;
 682}
 683
 684core_initcall(cpufreq_tsc);
 685
 686#endif /* CONFIG_CPU_FREQ */
 687
 688/* clocksource code */
 689
 690static struct clocksource clocksource_tsc;
 691
 692/*
 693 * We compare the TSC to the cycle_last value in the clocksource
 694 * structure to avoid a nasty time-warp. This can be observed in a
 695 * very small window right after one CPU updated cycle_last under
 696 * xtime/vsyscall_gtod lock and the other CPU reads a TSC value which
 697 * is smaller than the cycle_last reference value due to a TSC which
 698 * is slighty behind. This delta is nowhere else observable, but in
 699 * that case it results in a forward time jump in the range of hours
 700 * due to the unsigned delta calculation of the time keeping core
 701 * code, which is necessary to support wrapping clocksources like pm
 702 * timer.
 703 */
 704static cycle_t read_tsc(void)
 705{
 706        cycle_t ret = (cycle_t)get_cycles();
 707
 708        return ret >= clocksource_tsc.cycle_last ?
 709                ret : clocksource_tsc.cycle_last;
 710}
 711
 712#ifdef CONFIG_X86_64
 713static cycle_t __vsyscall_fn vread_tsc(void)
 714{
 715        cycle_t ret = (cycle_t)vget_cycles();
 716
 717        return ret >= __vsyscall_gtod_data.clock.cycle_last ?
 718                ret : __vsyscall_gtod_data.clock.cycle_last;
 719}
 720#endif
 721
 722static struct clocksource clocksource_tsc = {
 723        .name                   = "tsc",
 724        .rating                 = 300,
 725        .read                   = read_tsc,
 726        .mask                   = CLOCKSOURCE_MASK(64),
 727        .shift                  = 22,
 728        .flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
 729                                  CLOCK_SOURCE_MUST_VERIFY,
 730#ifdef CONFIG_X86_64
 731        .vread                  = vread_tsc,
 732#endif
 733};
 734
 735void mark_tsc_unstable(char *reason)
 736{
 737        if (!tsc_unstable) {
 738                tsc_unstable = 1;
 739                printk("Marking TSC unstable due to %s\n", reason);
 740                /* Change only the rating, when not registered */
 741                if (clocksource_tsc.mult)
 742                        clocksource_change_rating(&clocksource_tsc, 0);
 743                else
 744                        clocksource_tsc.rating = 0;
 745        }
 746}
 747
 748EXPORT_SYMBOL_GPL(mark_tsc_unstable);
 749
 750static int __init dmi_mark_tsc_unstable(const struct dmi_system_id *d)
 751{
 752        printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
 753                        d->ident);
 754        tsc_unstable = 1;
 755        return 0;
 756}
 757
 758/* List of systems that have known TSC problems */
 759static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
 760        {
 761                .callback = dmi_mark_tsc_unstable,
 762                .ident = "IBM Thinkpad 380XD",
 763                .matches = {
 764                        DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
 765                        DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
 766                },
 767        },
 768        {}
 769};
 770
 771static void __init check_system_tsc_reliable(void)
 772{
 773#ifdef CONFIG_MGEODE_LX
 774        /* RTSC counts during suspend */
 775#define RTSC_SUSP 0x100
 776        unsigned long res_low, res_high;
 777
 778        rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
 779        /* Geode_LX - the OLPC CPU has a possibly a very reliable TSC */
 780        if (res_low & RTSC_SUSP)
 781                tsc_clocksource_reliable = 1;
 782#endif
 783        if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE))
 784                tsc_clocksource_reliable = 1;
 785}
 786
 787/*
 788 * Make an educated guess if the TSC is trustworthy and synchronized
 789 * over all CPUs.
 790 */
 791__cpuinit int unsynchronized_tsc(void)
 792{
 793        if (!cpu_has_tsc || tsc_unstable)
 794                return 1;
 795
 796#ifdef CONFIG_X86_SMP
 797        if (apic_is_clustered_box())
 798                return 1;
 799#endif
 800
 801        if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
 802                return 0;
 803        /*
 804         * Intel systems are normally all synchronized.
 805         * Exceptions must mark TSC as unstable:
 806         */
 807        if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
 808                /* assume multi socket systems are not synchronized: */
 809                if (num_possible_cpus() > 1)
 810                        tsc_unstable = 1;
 811        }
 812
 813        return tsc_unstable;
 814}
 815
 816static void __init init_tsc_clocksource(void)
 817{
 818        clocksource_tsc.mult = clocksource_khz2mult(tsc_khz,
 819                        clocksource_tsc.shift);
 820        if (tsc_clocksource_reliable)
 821                clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
 822        /* lower the rating if we already know its unstable: */
 823        if (check_tsc_unstable()) {
 824                clocksource_tsc.rating = 0;
 825                clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
 826        }
 827        clocksource_register(&clocksource_tsc);
 828}
 829
 830void __init tsc_init(void)
 831{
 832        u64 lpj;
 833        int cpu;
 834
 835        if (!cpu_has_tsc)
 836                return;
 837
 838        tsc_khz = calibrate_tsc();
 839        cpu_khz = tsc_khz;
 840
 841        if (!tsc_khz) {
 842                mark_tsc_unstable("could not calculate TSC khz");
 843                return;
 844        }
 845
 846#ifdef CONFIG_X86_64
 847        if (cpu_has(&boot_cpu_data, X86_FEATURE_CONSTANT_TSC) &&
 848                        (boot_cpu_data.x86_vendor == X86_VENDOR_AMD))
 849                cpu_khz = calibrate_cpu();
 850#endif
 851
 852        printk("Detected %lu.%03lu MHz processor.\n",
 853                        (unsigned long)cpu_khz / 1000,
 854                        (unsigned long)cpu_khz % 1000);
 855
 856        /*
 857         * Secondary CPUs do not run through tsc_init(), so set up
 858         * all the scale factors for all CPUs, assuming the same
 859         * speed as the bootup CPU. (cpufreq notifiers will fix this
 860         * up if their speed diverges)
 861         */
 862        for_each_possible_cpu(cpu)
 863                set_cyc2ns_scale(cpu_khz, cpu);
 864
 865        if (tsc_disabled > 0)
 866                return;
 867
 868        /* now allow native_sched_clock() to use rdtsc */
 869        tsc_disabled = 0;
 870
 871        lpj = ((u64)tsc_khz * 1000);
 872        do_div(lpj, HZ);
 873        lpj_fine = lpj;
 874
 875        use_tsc_delay();
 876        /* Check and install the TSC clocksource */
 877        dmi_check_system(bad_tsc_dmi_table);
 878
 879        if (unsynchronized_tsc())
 880                mark_tsc_unstable("TSCs unsynchronized");
 881
 882        check_system_tsc_reliable();
 883        init_tsc_clocksource();
 884}
 885
 886
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.