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>
9 #include <linux/delay.h>
10 #include <linux/clocksource.h>
11 #include <linux/percpu.h>
14 #include <asm/timer.h>
15 #include <asm/vgtod.h>
17 #include <asm/delay.h>
19 unsigned int cpu_khz; /* TSC clocks / usec, not used here */
20 EXPORT_SYMBOL(cpu_khz);
22 EXPORT_SYMBOL(tsc_khz);
25 * TSC can be unstable due to cpufreq or due to unsynced TSCs
27 static int tsc_unstable;
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 */
32 static int tsc_disabled = -1;
35 * Scheduler clock - returns current time in nanosec units.
37 u64 native_sched_clock(void)
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
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);
54 /* read the Time Stamp Counter: */
57 /* return the value in ns */
58 return cycles_2_ns(this_offset);
61 /* We need to define a real function for sched_clock, to override the
62 weak default version */
63 #ifdef CONFIG_PARAVIRT
64 unsigned long long sched_clock(void)
66 return paravirt_sched_clock();
70 sched_clock(void) __attribute__((alias("native_sched_clock")));
73 int check_tsc_unstable(void)
77 EXPORT_SYMBOL_GPL(check_tsc_unstable);
80 int __init notsc_setup(char *str)
82 printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, "
83 "cannot disable TSC completely.\n");
89 * disable flag for tsc. Takes effect by clearing the TSC cpu flag
92 int __init notsc_setup(char *str)
94 setup_clear_cpu_cap(X86_FEATURE_TSC);
99 __setup("notsc", notsc_setup);
101 #define MAX_RETRIES 5
102 #define SMI_TRESHOLD 50000
105 * Read TSC and the reference counters. Take care of SMI disturbance
107 static u64 tsc_read_refs(u64 *p, int hpet)
112 for (i = 0; i < MAX_RETRIES; i++) {
115 *p = hpet_readl(HPET_COUNTER) & 0xFFFFFFFF;
117 *p = acpi_pm_read_early();
119 if ((t2 - t1) < SMI_TRESHOLD)
126 * Calculate the TSC frequency from HPET reference
128 static unsigned long calc_hpet_ref(u64 deltatsc, u64 hpet1, u64 hpet2)
133 hpet2 += 0x100000000ULL;
135 tmp = ((u64)hpet2 * hpet_readl(HPET_PERIOD));
136 do_div(tmp, 1000000);
137 do_div(deltatsc, tmp);
139 return (unsigned long) deltatsc;
143 * Calculate the TSC frequency from PMTimer reference
145 static unsigned long calc_pmtimer_ref(u64 deltatsc, u64 pm1, u64 pm2)
153 pm2 += (u64)ACPI_PM_OVRRUN;
155 tmp = pm2 * 1000000000LL;
156 do_div(tmp, PMTMR_TICKS_PER_SEC);
157 do_div(deltatsc, tmp);
159 return (unsigned long) deltatsc;
163 #define CAL_LATCH (CLOCK_TICK_RATE / (1000 / CAL_MS))
164 #define CAL_PIT_LOOPS 1000
167 #define CAL2_LATCH (CLOCK_TICK_RATE / (1000 / CAL2_MS))
168 #define CAL2_PIT_LOOPS 5000
172 * Try to calibrate the TSC against the Programmable
173 * Interrupt Timer and return the frequency of the TSC
176 * Return ULONG_MAX on failure to calibrate.
178 static unsigned long pit_calibrate_tsc(u32 latch, unsigned long ms, int loopmin)
180 u64 tsc, t1, t2, delta;
181 unsigned long tscmin, tscmax;
184 /* Set the Gate high, disable speaker */
185 outb((inb(0x61) & ~0x02) | 0x01, 0x61);
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.
193 outb(latch & 0xff, 0x42);
194 outb(latch >> 8, 0x42);
196 tsc = t1 = t2 = get_cycles();
201 while ((inb(0x61) & 0x20) == 0) {
205 if ((unsigned long) delta < tscmin)
206 tscmin = (unsigned int) delta;
207 if ((unsigned long) delta > tscmax)
208 tscmax = (unsigned int) delta;
215 * If we were not able to read the PIT more than loopmin
216 * times, then we have been hit by a massive SMI
218 * If the maximum is 10 times larger than the minimum,
219 * then we got hit by an SMI as well.
221 if (pitcnt < loopmin || tscmax > 10 * tscmin)
224 /* Calculate the PIT value */
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.
235 * Our expectations are:
237 * - the PIT is running at roughly 1.19MHz
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.
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).
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.
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).
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.
265 static inline int pit_expect_msb(unsigned char val)
269 for (count = 0; count < 50000; count++) {
272 if (inb(0x42) != val)
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
284 #define QUICK_PIT_MS 15
285 #define QUICK_PIT_ITERATIONS (QUICK_PIT_MS * PIT_TICK_RATE / 1000 / 256)
287 static unsigned long quick_pit_calibrate(void)
289 /* Set the Gate high, disable speaker */
290 outb((inb(0x61) & ~0x02) | 0x01, 0x61);
293 * Counter 2, mode 0 (one-shot), binary count
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
303 /* Start at 0xffff */
307 if (pit_expect_msb(0xff)) {
310 unsigned char expect = 0xfe;
313 for (i = 0; i < QUICK_PIT_ITERATIONS; i++, expect--) {
314 if (!pit_expect_msb(expect))
320 * Make sure we can rely on the second TSC timestamp:
322 if (!pit_expect_msb(--expect))
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.
331 * As a result, we can depend on there not being
332 * any odd delays anywhere, and the TSC reads are
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)
339 delta = (t2 - t1)*PIT_TICK_RATE;
340 do_div(delta, QUICK_PIT_ITERATIONS*256*1000);
341 printk("Fast TSC calibration using PIT\n");
349 * native_calibrate_tsc - calibrate the tsc on boot
351 unsigned long native_calibrate_tsc(void)
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;
358 local_irq_save(flags);
359 fast_calibrate = quick_pit_calibrate();
360 local_irq_restore(flags);
362 return fast_calibrate;
365 * Run 5 calibration loops to get the lowest frequency value
366 * (the best estimate). We use two different calibration modes
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.
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.
389 /* Preset PIT loop values */
392 loopmin = CAL_PIT_LOOPS;
394 for (i = 0; i < 3; i++) {
395 unsigned long tsc_pit_khz;
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.
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);
409 /* Pick the lowest PIT TSC calibration so far */
410 tsc_pit_min = min(tsc_pit_min, tsc_pit_khz);
412 /* hpet or pmtimer available ? */
413 if (!hpet && !ref1 && !ref2)
416 /* Check, whether the sampling was disturbed by an SMI */
417 if (tsc1 == ULLONG_MAX || tsc2 == ULLONG_MAX)
420 tsc2 = (tsc2 - tsc1) * 1000000LL;
422 tsc2 = calc_hpet_ref(tsc2, ref1, ref2);
424 tsc2 = calc_pmtimer_ref(tsc2, ref1, ref2);
426 tsc_ref_min = min(tsc_ref_min, (unsigned long) tsc2);
428 /* Check the reference deviation */
429 delta = ((u64) tsc_pit_min) * 100;
430 do_div(delta, tsc_ref_min);
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.
438 if (delta >= 90 && delta <= 110) {
440 "TSC: PIT calibration matches %s. %d loops\n",
441 hpet ? "HPET" : "PMTIMER", i + 1);
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.
451 if (i == 1 && tsc_pit_min == ULONG_MAX) {
454 loopmin = CAL2_PIT_LOOPS;
459 * Now check the results.
461 if (tsc_pit_min == ULONG_MAX) {
462 /* PIT gave no useful value */
463 printk(KERN_WARNING "TSC: PIT calibration failed due to "
464 "SMI disturbance.\n");
466 /* We don't have an alternative source, disable TSC */
467 if (!hpet && !ref1 && !ref2) {
468 printk("TSC: No reference (HPET/PMTIMER) available\n");
472 /* The alternative source failed as well, disable TSC */
473 if (tsc_ref_min == ULONG_MAX) {
474 printk(KERN_WARNING "TSC: HPET/PMTIMER calibration "
479 /* Use the alternative source */
480 printk(KERN_INFO "TSC: using %s reference calibration\n",
481 hpet ? "HPET" : "PMTIMER");
486 /* We don't have an alternative source, use the PIT calibration value */
487 if (!hpet && !ref1 && !ref2) {
488 printk(KERN_INFO "TSC: Using PIT calibration value\n");
492 /* The alternative source failed, use the PIT calibration value */
493 if (tsc_ref_min == ULONG_MAX) {
494 printk(KERN_WARNING "TSC: HPET/PMTIMER calibration failed. "
495 "Using PIT calibration\n");
500 * The calibration values differ too much. In doubt, we use
501 * the PIT value as we know that there are PMTIMERs around
502 * running at double speed. At least we let the user know:
504 printk(KERN_WARNING "TSC: PIT calibration deviates from %s: %lu %lu.\n",
505 hpet ? "HPET" : "PMTIMER", tsc_pit_min, tsc_ref_min);
506 printk(KERN_INFO "TSC: Using PIT calibration value\n");
511 /* Only called from the Powernow K7 cpu freq driver */
512 int recalibrate_cpu_khz(void)
515 unsigned long cpu_khz_old = cpu_khz;
518 tsc_khz = calibrate_tsc();
520 cpu_data(0).loops_per_jiffy =
521 cpufreq_scale(cpu_data(0).loops_per_jiffy,
522 cpu_khz_old, cpu_khz);
531 EXPORT_SYMBOL(recalibrate_cpu_khz);
533 #endif /* CONFIG_X86_32 */
535 /* Accelerators for sched_clock()
536 * convert from cycles(64bits) => nanoseconds (64bits)
538 * ns = cycles / (freq / ns_per_sec)
539 * ns = cycles * (ns_per_sec / freq)
540 * ns = cycles * (10^9 / (cpu_khz * 10^3))
541 * ns = cycles * (10^6 / cpu_khz)
543 * Then we use scaling math (suggested by george@mvista.com) to get:
544 * ns = cycles * (10^6 * SC / cpu_khz) / SC
545 * ns = cycles * cyc2ns_scale / SC
547 * And since SC is a constant power of two, we can convert the div
550 * We can use khz divisor instead of mhz to keep a better precision, since
551 * cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
552 * (mathieu.desnoyers@polymtl.ca)
554 * -johnstul@us.ibm.com "math is hard, lets go shopping!"
557 DEFINE_PER_CPU(unsigned long, cyc2ns);
559 static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu)
561 unsigned long long tsc_now, ns_now;
562 unsigned long flags, *scale;
564 local_irq_save(flags);
565 sched_clock_idle_sleep_event();
567 scale = &per_cpu(cyc2ns, cpu);
570 ns_now = __cycles_2_ns(tsc_now);
573 *scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz;
575 sched_clock_idle_wakeup_event(0);
576 local_irq_restore(flags);
579 #ifdef CONFIG_CPU_FREQ
581 /* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
584 * RED-PEN: On SMP we assume all CPUs run with the same frequency. It's
585 * not that important because current Opteron setups do not support
586 * scaling on SMP anyroads.
588 * Should fix up last_tsc too. Currently gettimeofday in the
589 * first tick after the change will be slightly wrong.
592 static unsigned int ref_freq;
593 static unsigned long loops_per_jiffy_ref;
594 static unsigned long tsc_khz_ref;
596 static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
599 struct cpufreq_freqs *freq = data;
600 unsigned long *lpj, dummy;
602 if (cpu_has(&cpu_data(freq->cpu), X86_FEATURE_CONSTANT_TSC))
606 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
608 lpj = &cpu_data(freq->cpu).loops_per_jiffy;
610 lpj = &boot_cpu_data.loops_per_jiffy;
614 ref_freq = freq->old;
615 loops_per_jiffy_ref = *lpj;
616 tsc_khz_ref = tsc_khz;
618 if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) ||
619 (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
620 (val == CPUFREQ_RESUMECHANGE)) {
621 *lpj = cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
623 tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
624 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
625 mark_tsc_unstable("cpufreq changes");
628 set_cyc2ns_scale(tsc_khz, freq->cpu);
633 static struct notifier_block time_cpufreq_notifier_block = {
634 .notifier_call = time_cpufreq_notifier
637 static int __init cpufreq_tsc(void)
641 if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
643 cpufreq_register_notifier(&time_cpufreq_notifier_block,
644 CPUFREQ_TRANSITION_NOTIFIER);
648 core_initcall(cpufreq_tsc);
650 #endif /* CONFIG_CPU_FREQ */
652 /* clocksource code */
654 static struct clocksource clocksource_tsc;
657 * We compare the TSC to the cycle_last value in the clocksource
658 * structure to avoid a nasty time-warp. This can be observed in a
659 * very small window right after one CPU updated cycle_last under
660 * xtime/vsyscall_gtod lock and the other CPU reads a TSC value which
661 * is smaller than the cycle_last reference value due to a TSC which
662 * is slighty behind. This delta is nowhere else observable, but in
663 * that case it results in a forward time jump in the range of hours
664 * due to the unsigned delta calculation of the time keeping core
665 * code, which is necessary to support wrapping clocksources like pm
668 static cycle_t read_tsc(void)
670 cycle_t ret = (cycle_t)get_cycles();
672 return ret >= clocksource_tsc.cycle_last ?
673 ret : clocksource_tsc.cycle_last;
677 static cycle_t __vsyscall_fn vread_tsc(void)
679 cycle_t ret = (cycle_t)vget_cycles();
681 return ret >= __vsyscall_gtod_data.clock.cycle_last ?
682 ret : __vsyscall_gtod_data.clock.cycle_last;
686 static struct clocksource clocksource_tsc = {
690 .mask = CLOCKSOURCE_MASK(64),
692 .flags = CLOCK_SOURCE_IS_CONTINUOUS |
693 CLOCK_SOURCE_MUST_VERIFY,
699 void mark_tsc_unstable(char *reason)
703 printk("Marking TSC unstable due to %s\n", reason);
704 /* Change only the rating, when not registered */
705 if (clocksource_tsc.mult)
706 clocksource_change_rating(&clocksource_tsc, 0);
708 clocksource_tsc.rating = 0;
712 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
714 static int __init dmi_mark_tsc_unstable(const struct dmi_system_id *d)
716 printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
722 /* List of systems that have known TSC problems */
723 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
725 .callback = dmi_mark_tsc_unstable,
726 .ident = "IBM Thinkpad 380XD",
728 DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
729 DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
736 * Geode_LX - the OLPC CPU has a possibly a very reliable TSC
738 #ifdef CONFIG_MGEODE_LX
739 /* RTSC counts during suspend */
740 #define RTSC_SUSP 0x100
742 static void __init check_geode_tsc_reliable(void)
744 unsigned long res_low, res_high;
746 rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
747 if (res_low & RTSC_SUSP)
748 clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
751 static inline void check_geode_tsc_reliable(void) { }
755 * Make an educated guess if the TSC is trustworthy and synchronized
758 __cpuinit int unsynchronized_tsc(void)
760 if (!cpu_has_tsc || tsc_unstable)
764 if (apic_is_clustered_box())
768 if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
771 * Intel systems are normally all synchronized.
772 * Exceptions must mark TSC as unstable:
774 if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
775 /* assume multi socket systems are not synchronized: */
776 if (num_possible_cpus() > 1)
783 static void __init init_tsc_clocksource(void)
785 clocksource_tsc.mult = clocksource_khz2mult(tsc_khz,
786 clocksource_tsc.shift);
787 /* lower the rating if we already know its unstable: */
788 if (check_tsc_unstable()) {
789 clocksource_tsc.rating = 0;
790 clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
792 clocksource_register(&clocksource_tsc);
795 void __init tsc_init(void)
803 tsc_khz = calibrate_tsc();
807 mark_tsc_unstable("could not calculate TSC khz");
812 if (cpu_has(&boot_cpu_data, X86_FEATURE_CONSTANT_TSC) &&
813 (boot_cpu_data.x86_vendor == X86_VENDOR_AMD))
814 cpu_khz = calibrate_cpu();
817 lpj = ((u64)tsc_khz * 1000);
821 printk("Detected %lu.%03lu MHz processor.\n",
822 (unsigned long)cpu_khz / 1000,
823 (unsigned long)cpu_khz % 1000);
826 * Secondary CPUs do not run through tsc_init(), so set up
827 * all the scale factors for all CPUs, assuming the same
828 * speed as the bootup CPU. (cpufreq notifiers will fix this
829 * up if their speed diverges)
831 for_each_possible_cpu(cpu)
832 set_cyc2ns_scale(cpu_khz, cpu);
834 if (tsc_disabled > 0)
837 /* now allow native_sched_clock() to use rdtsc */
841 /* Check and install the TSC clocksource */
842 dmi_check_system(bad_tsc_dmi_table);
844 if (unsynchronized_tsc())
845 mark_tsc_unstable("TSCs unsynchronized");
847 check_geode_tsc_reliable();
848 init_tsc_clocksource();