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 20 unsigned int cpu_khz; /* TSC clocks / usec, not used here */ 21 EXPORT_SYMBOL(cpu_khz); 22 unsigned int tsc_khz; 23 EXPORT_SYMBOL(tsc_khz); 24 25 /* 26 * TSC can be unstable due to cpufreq or due to unsynced TSCs 27 */ 28 static 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 */ 33 static int tsc_disabled = -1; 34 35 static int tsc_clocksource_reliable; 36 /* 37 * Scheduler clock - returns current time in nanosec units. 38 */ 39 u64 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 66 unsigned long long sched_clock(void) 67 { 68 return paravirt_sched_clock(); 69 } 70 #else 71 unsigned long long 72 sched_clock(void) __attribute__((alias("native_sched_clock"))); 73 #endif 74 75 int check_tsc_unstable(void) 76 { 77 return tsc_unstable; 78 } 79 EXPORT_SYMBOL_GPL(check_tsc_unstable); 80 81 #ifdef CONFIG_X86_TSC 82 int __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 */ 94 int __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); 102 103 static 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 */ 118 static 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 */ 139 static 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 */ 156 static 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 */ 189 static 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 */ 276 static inline int pit_expect_msb(unsigned char val) 277 { 278 int count = 0; 279 280 for (count = 0; count < 50000; count++) { 281 /* Ignore LSB */ 282 inb(0x42); 283 if (inb(0x42) != val) 284 break; 285 } 286 return count > 50; 287 } 288 289 /* 290 * How many MSB values do we want to see? We aim for a 291 * 15ms calibration, which assuming a 2us counter read 292 * error should give us roughly 150 ppm precision for 293 * the calibration. 294 */ 295 #define QUICK_PIT_MS 15 296 #define QUICK_PIT_ITERATIONS (QUICK_PIT_MS * PIT_TICK_RATE / 1000 / 256) 297 298 static unsigned long quick_pit_calibrate(void) 299 { 300 /* Set the Gate high, disable speaker */ 301 outb((inb(0x61) & ~0x02) | 0x01, 0x61); 302 303 /* 304 * Counter 2, mode 0 (one-shot), binary count 305 * 306 * NOTE! Mode 2 decrements by two (and then the 307 * output is flipped each time, giving the same 308 * final output frequency as a decrement-by-one), 309 * so mode 0 is much better when looking at the 310 * individual counts. 311 */ 312 outb(0xb0, 0x43); 313 314 /* Start at 0xffff */ 315 outb(0xff, 0x42); 316 outb(0xff, 0x42); 317 318 if (pit_expect_msb(0xff)) { 319 int i; 320 u64 t1, t2, delta; 321 unsigned char expect = 0xfe; 322 323 t1 = get_cycles(); 324 for (i = 0; i < QUICK_PIT_ITERATIONS; i++, expect--) { 325 if (!pit_expect_msb(expect)) 326 goto failed; 327 } 328 t2 = get_cycles(); 329 330 /* 331 * Make sure we can rely on the second TSC timestamp: 332 */ 333 if (!pit_expect_msb(expect)) 334 goto failed; 335 336 /* 337 * Ok, if we get here, then we've seen the 338 * MSB of the PIT decrement QUICK_PIT_ITERATIONS 339 * times, and each MSB had many hits, so we never 340 * had any sudden jumps. 341 * 342 * As a result, we can depend on there not being 343 * any odd delays anywhere, and the TSC reads are 344 * reliable. 345 * 346 * kHz = ticks / time-in-seconds / 1000; 347 * kHz = (t2 - t1) / (QPI * 256 / PIT_TICK_RATE) / 1000 348 * kHz = ((t2 - t1) * PIT_TICK_RATE) / (QPI * 256 * 1000) 349 */ 350 delta = (t2 - t1)*PIT_TICK_RATE; 351 do_div(delta, QUICK_PIT_ITERATIONS*256*1000); 352 printk("Fast TSC calibration using PIT\n"); 353 return delta; 354 } 355 failed: 356 return 0; 357 } 358 359 /** 360 * native_calibrate_tsc - calibrate the tsc on boot 361 */ 362 unsigned long native_calibrate_tsc(void) 363 { 364 u64 tsc1, tsc2, delta, ref1, ref2; 365 unsigned long tsc_pit_min = ULONG_MAX, tsc_ref_min = ULONG_MAX; 366 unsigned long flags, latch, ms, fast_calibrate, tsc_khz; 367 int hpet = is_hpet_enabled(), i, loopmin; 368 369 tsc_khz = get_hypervisor_tsc_freq(); 370 if (tsc_khz) { 371 printk(KERN_INFO "TSC: Frequency read from the hypervisor\n"); 372 return tsc_khz; 373 } 374 375 local_irq_save(flags); 376 fast_calibrate = quick_pit_calibrate(); 377 local_irq_restore(flags); 378 if (fast_calibrate) 379 return fast_calibrate; 380 381 /* 382 * Run 5 calibration loops to get the lowest frequency value 383 * (the best estimate). We use two different calibration modes 384 * here: 385 * 386 * 1) PIT loop. We set the PIT Channel 2 to oneshot mode and 387 * load a timeout of 50ms. We read the time right after we 388 * started the timer and wait until the PIT count down reaches 389 * zero. In each wait loop iteration we read the TSC and check 390 * the delta to the previous read. We keep track of the min 391 * and max values of that delta. The delta is mostly defined 392 * by the IO time of the PIT access, so we can detect when a 393 * SMI/SMM disturbance happend between the two reads. If the 394 * maximum time is significantly larger than the minimum time, 395 * then we discard the result and have another try. 396 * 397 * 2) Reference counter. If available we use the HPET or the 398 * PMTIMER as a reference to check the sanity of that value. 399 * We use separate TSC readouts and check inside of the 400 * reference read for a SMI/SMM disturbance. We dicard 401 * disturbed values here as well. We do that around the PIT 402 * calibration delay loop as we have to wait for a certain 403 * amount of time anyway. 404 */ 405 406 /* Preset PIT loop values */ 407 latch = CAL_LATCH; 408 ms = CAL_MS; 409 loopmin = CAL_PIT_LOOPS; 410 411 for (i = 0; i < 3; i++) { 412 unsigned long tsc_pit_khz; 413 414 /* 415 * Read the start value and the reference count of 416 * hpet/pmtimer when available. Then do the PIT 417 * calibration, which will take at least 50ms, and 418 * read the end value. 419 */ 420 local_irq_save(flags); 421 tsc1 = tsc_read_refs(&ref1, hpet); 422 tsc_pit_khz = pit_calibrate_tsc(latch, ms, loopmin); 423 tsc2 = tsc_read_refs(&ref2, hpet); 424 local_irq_restore(flags); 425 426 /* Pick the lowest PIT TSC calibration so far */ 427 tsc_pit_min = min(tsc_pit_min, tsc_pit_khz); 428 429 /* hpet or pmtimer available ? */ 430 if (!hpet && !ref1 && !ref2) 431 continue; 432 433 /* Check, whether the sampling was disturbed by an SMI */ 434 if (tsc1 == ULLONG_MAX || tsc2 == ULLONG_MAX) 435 continue; 436 437 tsc2 = (tsc2 - tsc1) * 1000000LL; 438 if (hpet) 439 tsc2 = calc_hpet_ref(tsc2, ref1, ref2); 440 else 441 tsc2 = calc_pmtimer_ref(tsc2, ref1, ref2); 442 443 tsc_ref_min = min(tsc_ref_min, (unsigned long) tsc2); 444 445 /* Check the reference deviation */ 446 delta = ((u64) tsc_pit_min) * 100; 447 do_div(delta, tsc_ref_min); 448 449 /* 450 * If both calibration results are inside a 10% window 451 * then we can be sure, that the calibration 452 * succeeded. We break out of the loop right away. We 453 * use the reference value, as it is more precise. 454 */ 455 if (delta >= 90 && delta <= 110) { 456 printk(KERN_INFO 457 "TSC: PIT calibration matches %s. %d loops\n", 458 hpet ? "HPET" : "PMTIMER", i + 1); 459 return tsc_ref_min; 460 } 461 462 /* 463 * Check whether PIT failed more than once. This 464 * happens in virtualized environments. We need to 465 * give the virtual PC a slightly longer timeframe for 466 * the HPET/PMTIMER to make the result precise. 467 */ 468 if (i == 1 && tsc_pit_min == ULONG_MAX) { 469 latch = CAL2_LATCH; 470 ms = CAL2_MS; 471 loopmin = CAL2_PIT_LOOPS; 472 } 473 } 474 475 /* 476 * Now check the results. 477 */ 478 if (tsc_pit_min == ULONG_MAX) { 479 /* PIT gave no useful value */ 480 printk(KERN_WARNING "TSC: Unable to calibrate against PIT\n"); 481 482 /* We don't have an alternative source, disable TSC */ 483 if (!hpet && !ref1 && !ref2) { 484 printk("TSC: No reference (HPET/PMTIMER) available\n"); 485 return 0; 486 } 487 488 /* The alternative source failed as well, disable TSC */ 489 if (tsc_ref_min == ULONG_MAX) { 490 printk(KERN_WARNING "TSC: HPET/PMTIMER calibration " 491 "failed.\n"); 492 return 0; 493 } 494 495 /* Use the alternative source */ 496 printk(KERN_INFO "TSC: using %s reference calibration\n", 497 hpet ? "HPET" : "PMTIMER"); 498 499 return tsc_ref_min; 500 } 501 502 /* We don't have an alternative source, use the PIT calibration value */ 503 if (!hpet && !ref1 && !ref2) { 504 printk(KERN_INFO "TSC: Using PIT calibration value\n"); 505 return tsc_pit_min; 506 } 507 508 /* The alternative source failed, use the PIT calibration value */ 509 if (tsc_ref_min == ULONG_MAX) { 510 printk(KERN_WARNING "TSC: HPET/PMTIMER calibration failed. " 511 "Using PIT calibration\n"); 512 return tsc_pit_min; 513 } 514 515 /* 516 * The calibration values differ too much. In doubt, we use 517 * the PIT value as we know that there are PMTIMERs around 518 * running at double speed. At least we let the user know: 519 */ 520 printk(KERN_WARNING "TSC: PIT calibration deviates from %s: %lu %lu.\n", 521 hpet ? "HPET" : "PMTIMER", tsc_pit_min, tsc_ref_min); 522 printk(KERN_INFO "TSC: Using PIT calibration value\n"); 523 return tsc_pit_min; 524 } 525 526 #ifdef CONFIG_X86_32 527 /* Only called from the Powernow K7 cpu freq driver */ 528 int recalibrate_cpu_khz(void) 529 { 530 #ifndef CONFIG_SMP 531 unsigned long cpu_khz_old = cpu_khz; 532 533 if (cpu_has_tsc) { 534 tsc_khz = calibrate_tsc(); 535 cpu_khz = tsc_khz; 536 cpu_data(0).loops_per_jiffy = 537 cpufreq_scale(cpu_data(0).loops_per_jiffy, 538 cpu_khz_old, cpu_khz); 539 return 0; 540 } else 541 return -ENODEV; 542 #else 543 return -ENODEV; 544 #endif 545 } 546 547 EXPORT_SYMBOL(recalibrate_cpu_khz); 548 549 #endif /* CONFIG_X86_32 */ 550 551 /* Accelerators for sched_clock() 552 * convert from cycles(64bits) => nanoseconds (64bits) 553 * basic equation: 554 * ns = cycles / (freq / ns_per_sec) 555 * ns = cycles * (ns_per_sec / freq) 556 * ns = cycles * (10^9 / (cpu_khz * 10^3)) 557 * ns = cycles * (10^6 / cpu_khz) 558 * 559 * Then we use scaling math (suggested by george@mvista.com) to get: 560 * ns = cycles * (10^6 * SC / cpu_khz) / SC 561 * ns = cycles * cyc2ns_scale / SC 562 * 563 * And since SC is a constant power of two, we can convert the div 564 * into a shift. 565 * 566 * We can use khz divisor instead of mhz to keep a better precision, since 567 * cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits. 568 * (mathieu.desnoyers@polymtl.ca) 569 * 570 * -johnstul@us.ibm.com "math is hard, lets go shopping!" 571 */ 572 573 DEFINE_PER_CPU(unsigned long, cyc2ns); 574 575 static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu) 576 { 577 unsigned long long tsc_now, ns_now; 578 unsigned long flags, *scale; 579 580 local_irq_save(flags); 581 sched_clock_idle_sleep_event(); 582 583 scale = &per_cpu(cyc2ns, cpu); 584 585 rdtscll(tsc_now); 586 ns_now = __cycles_2_ns(tsc_now); 587 588 if (cpu_khz) 589 *scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz; 590 591 sched_clock_idle_wakeup_event(0); 592 local_irq_restore(flags); 593 } 594 595 #ifdef CONFIG_CPU_FREQ 596 597 /* Frequency scaling support. Adjust the TSC based timer when the cpu frequency 598 * changes. 599 * 600 * RED-PEN: On SMP we assume all CPUs run with the same frequency. It's 601 * not that important because current Opteron setups do not support 602 * scaling on SMP anyroads. 603 * 604 * Should fix up last_tsc too. Currently gettimeofday in the 605 * first tick after the change will be slightly wrong. 606 */ 607 608 static unsigned int ref_freq; 609 static unsigned long loops_per_jiffy_ref; 610 static unsigned long tsc_khz_ref; 611 612 static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, 613 void *data) 614 { 615 struct cpufreq_freqs *freq = data; 616 unsigned long *lpj, dummy; 617 618 if (cpu_has(&cpu_data(freq->cpu), X86_FEATURE_CONSTANT_TSC)) 619 return 0; 620 621 lpj = &dummy; 622 if (!(freq->flags & CPUFREQ_CONST_LOOPS)) 623 #ifdef CONFIG_SMP 624 lpj = &cpu_data(freq->cpu).loops_per_jiffy; 625 #else 626 lpj = &boot_cpu_data.loops_per_jiffy; 627 #endif 628 629 if (!ref_freq) { 630 ref_freq = freq->old; 631 loops_per_jiffy_ref = *lpj; 632 tsc_khz_ref = tsc_khz; 633 } 634 if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) || 635 (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) || 636 (val == CPUFREQ_RESUMECHANGE)) { 637 *lpj = cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new); 638 639 tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new); 640 if (!(freq->flags & CPUFREQ_CONST_LOOPS)) 641 mark_tsc_unstable("cpufreq changes"); 642 } 643 644 set_cyc2ns_scale(tsc_khz, freq->cpu); 645 646 return 0; 647 } 648 649 static struct notifier_block time_cpufreq_notifier_block = { 650 .notifier_call = time_cpufreq_notifier 651 }; 652 653 static int __init cpufreq_tsc(void) 654 { 655 if (!cpu_has_tsc) 656 return 0; 657 if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC)) 658 return 0; 659 cpufreq_register_notifier(&time_cpufreq_notifier_block, 660 CPUFREQ_TRANSITION_NOTIFIER); 661 return 0; 662 } 663 664 core_initcall(cpufreq_tsc); 665 666 #endif /* CONFIG_CPU_FREQ */ 667 668 /* clocksource code */ 669 670 static struct clocksource clocksource_tsc; 671 672 /* 673 * We compare the TSC to the cycle_last value in the clocksource 674 * structure to avoid a nasty time-warp. This can be observed in a 675 * very small window right after one CPU updated cycle_last under 676 * xtime/vsyscall_gtod lock and the other CPU reads a TSC value which 677 * is smaller than the cycle_last reference value due to a TSC which 678 * is slighty behind. This delta is nowhere else observable, but in 679 * that case it results in a forward time jump in the range of hours 680 * due to the unsigned delta calculation of the time keeping core 681 * code, which is necessary to support wrapping clocksources like pm 682 * timer. 683 */ 684 static cycle_t read_tsc(void) 685 { 686 cycle_t ret = (cycle_t)get_cycles(); 687 688 return ret >= clocksource_tsc.cycle_last ? 689 ret : clocksource_tsc.cycle_last; 690 } 691 692 #ifdef CONFIG_X86_64 693 static cycle_t __vsyscall_fn vread_tsc(void) 694 { 695 cycle_t ret = (cycle_t)vget_cycles(); 696 697 return ret >= __vsyscall_gtod_data.clock.cycle_last ? 698 ret : __vsyscall_gtod_data.clock.cycle_last; 699 } 700 #endif 701 702 static struct clocksource clocksource_tsc = { 703 .name = "tsc", 704 .rating = 300, 705 .read = read_tsc, 706 .mask = CLOCKSOURCE_MASK(64), 707 .shift = 22, 708 .flags = CLOCK_SOURCE_IS_CONTINUOUS | 709 CLOCK_SOURCE_MUST_VERIFY, 710 #ifdef CONFIG_X86_64 711 .vread = vread_tsc, 712 #endif 713 }; 714 715 void mark_tsc_unstable(char *reason) 716 { 717 if (!tsc_unstable) { 718 tsc_unstable = 1; 719 printk("Marking TSC unstable due to %s\n", reason); 720 /* Change only the rating, when not registered */ 721 if (clocksource_tsc.mult) 722 clocksource_change_rating(&clocksource_tsc, 0); 723 else 724 clocksource_tsc.rating = 0; 725 } 726 } 727 728 EXPORT_SYMBOL_GPL(mark_tsc_unstable); 729 730 static int __init dmi_mark_tsc_unstable(const struct dmi_system_id *d) 731 { 732 printk(KERN_NOTICE "%s detected: marking TSC unstable.\n", 733 d->ident); 734 tsc_unstable = 1; 735 return 0; 736 } 737 738 /* List of systems that have known TSC problems */ 739 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = { 740 { 741 .callback = dmi_mark_tsc_unstable, 742 .ident = "IBM Thinkpad 380XD", 743 .matches = { 744 DMI_MATCH(DMI_BOARD_VENDOR, "IBM"), 745 DMI_MATCH(DMI_BOARD_NAME, "2635FA0"), 746 }, 747 }, 748 {} 749 }; 750 751 static void __init check_system_tsc_reliable(void) 752 { 753 #ifdef CONFIG_MGEODE_LX 754 /* RTSC counts during suspend */ 755 #define RTSC_SUSP 0x100 756 unsigned long res_low, res_high; 757 758 rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high); 759 /* Geode_LX - the OLPC CPU has a possibly a very reliable TSC */ 760 if (res_low & RTSC_SUSP) 761 tsc_clocksource_reliable = 1; 762 #endif 763 if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE)) 764 tsc_clocksource_reliable = 1; 765 } 766 767 /* 768 * Make an educated guess if the TSC is trustworthy and synchronized 769 * over all CPUs. 770 */ 771 __cpuinit int unsynchronized_tsc(void) 772 { 773 if (!cpu_has_tsc || tsc_unstable) 774 return 1; 775 776 #ifdef CONFIG_X86_SMP 777 if (apic_is_clustered_box()) 778 return 1; 779 #endif 780 781 if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC)) 782 return 0; 783 /* 784 * Intel systems are normally all synchronized. 785 * Exceptions must mark TSC as unstable: 786 */ 787 if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) { 788 /* assume multi socket systems are not synchronized: */ 789 if (num_possible_cpus() > 1) 790 tsc_unstable = 1; 791 } 792 793 return tsc_unstable; 794 } 795 796 static void __init init_tsc_clocksource(void) 797 { 798 clocksource_tsc.mult = clocksource_khz2mult(tsc_khz, 799 clocksource_tsc.shift); 800 if (tsc_clocksource_reliable) 801 clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY; 802 /* lower the rating if we already know its unstable: */ 803 if (check_tsc_unstable()) { 804 clocksource_tsc.rating = 0; 805 clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS; 806 } 807 clocksource_register(&clocksource_tsc); 808 } 809 810 void __init tsc_init(void) 811 { 812 u64 lpj; 813 int cpu; 814 815 if (!cpu_has_tsc) 816 return; 817 818 tsc_khz = calibrate_tsc(); 819 cpu_khz = tsc_khz; 820 821 if (!tsc_khz) { 822 mark_tsc_unstable("could not calculate TSC khz"); 823 return; 824 } 825 826 #ifdef CONFIG_X86_64 827 if (cpu_has(&boot_cpu_data, X86_FEATURE_CONSTANT_TSC) && 828 (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)) 829 cpu_khz = calibrate_cpu(); 830 #endif 831 832 printk("Detected %lu.%03lu MHz processor.\n", 833 (unsigned long)cpu_khz / 1000, 834 (unsigned long)cpu_khz % 1000); 835 836 /* 837 * Secondary CPUs do not run through tsc_init(), so set up 838 * all the scale factors for all CPUs, assuming the same 839 * speed as the bootup CPU. (cpufreq notifiers will fix this 840 * up if their speed diverges) 841 */ 842 for_each_possible_cpu(cpu) 843 set_cyc2ns_scale(cpu_khz, cpu); 844 845 if (tsc_disabled > 0) 846 return; 847 848 /* now allow native_sched_clock() to use rdtsc */ 849 tsc_disabled = 0; 850 851 lpj = ((u64)tsc_khz * 1000); 852 do_div(lpj, HZ); 853 lpj_fine = lpj; 854 855 use_tsc_delay(); 856 /* Check and install the TSC clocksource */ 857 dmi_check_system(bad_tsc_dmi_table); 858 859 if (unsynchronized_tsc()) 860 mark_tsc_unstable("TSCs unsynchronized"); 861 862 check_system_tsc_reliable(); 863 init_tsc_clocksource(); 864 } 865 866