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 *pm, u64 *hpet)
112 for (i = 0; i < MAX_RETRIES; i++) {
115 *hpet = hpet_readl(HPET_COUNTER) & 0xFFFFFFFF;
117 *pm = acpi_pm_read_early();
119 if ((t2 - t1) < SMI_TRESHOLD)
126 * native_calibrate_tsc - calibrate the tsc on boot
128 unsigned long native_calibrate_tsc(void)
131 u64 tsc1, tsc2, tr1, tr2, delta, pm1, pm2, hpet1, hpet2;
132 int hpet = is_hpet_enabled();
133 unsigned int tsc_khz_val = 0;
135 local_irq_save(flags);
137 tsc1 = tsc_read_refs(&pm1, hpet ? &hpet1 : NULL);
139 outb((inb(0x61) & ~0x02) | 0x01, 0x61);
142 outb((CLOCK_TICK_RATE / (1000 / 50)) & 0xff, 0x42);
143 outb((CLOCK_TICK_RATE / (1000 / 50)) >> 8, 0x42);
145 while ((inb(0x61) & 0x20) == 0);
148 tsc2 = tsc_read_refs(&pm2, hpet ? &hpet2 : NULL);
150 local_irq_restore(flags);
153 * Preset the result with the raw and inaccurate PIT
160 /* hpet or pmtimer available ? */
161 if (!hpet && !pm1 && !pm2) {
162 printk(KERN_INFO "TSC calibrated against PIT\n");
166 /* Check, whether the sampling was disturbed by an SMI */
167 if (tsc1 == ULLONG_MAX || tsc2 == ULLONG_MAX) {
168 printk(KERN_WARNING "TSC calibration disturbed by SMI, "
169 "using PIT calibration result\n");
173 tsc2 = (tsc2 - tsc1) * 1000000LL;
176 printk(KERN_INFO "TSC calibrated against HPET\n");
178 hpet2 += 0x100000000ULL;
180 tsc1 = ((u64)hpet2 * hpet_readl(HPET_PERIOD));
181 do_div(tsc1, 1000000);
183 printk(KERN_INFO "TSC calibrated against PM_TIMER\n");
185 pm2 += (u64)ACPI_PM_OVRRUN;
187 tsc1 = pm2 * 1000000000LL;
188 do_div(tsc1, PMTMR_TICKS_PER_SEC);
200 /* Only called from the Powernow K7 cpu freq driver */
201 int recalibrate_cpu_khz(void)
204 unsigned long cpu_khz_old = cpu_khz;
207 tsc_khz = calibrate_tsc();
209 cpu_data(0).loops_per_jiffy =
210 cpufreq_scale(cpu_data(0).loops_per_jiffy,
211 cpu_khz_old, cpu_khz);
220 EXPORT_SYMBOL(recalibrate_cpu_khz);
222 #endif /* CONFIG_X86_32 */
224 /* Accelerators for sched_clock()
225 * convert from cycles(64bits) => nanoseconds (64bits)
227 * ns = cycles / (freq / ns_per_sec)
228 * ns = cycles * (ns_per_sec / freq)
229 * ns = cycles * (10^9 / (cpu_khz * 10^3))
230 * ns = cycles * (10^6 / cpu_khz)
232 * Then we use scaling math (suggested by george@mvista.com) to get:
233 * ns = cycles * (10^6 * SC / cpu_khz) / SC
234 * ns = cycles * cyc2ns_scale / SC
236 * And since SC is a constant power of two, we can convert the div
239 * We can use khz divisor instead of mhz to keep a better precision, since
240 * cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
241 * (mathieu.desnoyers@polymtl.ca)
243 * -johnstul@us.ibm.com "math is hard, lets go shopping!"
246 DEFINE_PER_CPU(unsigned long, cyc2ns);
248 static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu)
250 unsigned long long tsc_now, ns_now;
251 unsigned long flags, *scale;
253 local_irq_save(flags);
254 sched_clock_idle_sleep_event();
256 scale = &per_cpu(cyc2ns, cpu);
259 ns_now = __cycles_2_ns(tsc_now);
262 *scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz;
264 sched_clock_idle_wakeup_event(0);
265 local_irq_restore(flags);
268 #ifdef CONFIG_CPU_FREQ
270 /* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
273 * RED-PEN: On SMP we assume all CPUs run with the same frequency. It's
274 * not that important because current Opteron setups do not support
275 * scaling on SMP anyroads.
277 * Should fix up last_tsc too. Currently gettimeofday in the
278 * first tick after the change will be slightly wrong.
281 static unsigned int ref_freq;
282 static unsigned long loops_per_jiffy_ref;
283 static unsigned long tsc_khz_ref;
285 static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
288 struct cpufreq_freqs *freq = data;
289 unsigned long *lpj, dummy;
291 if (cpu_has(&cpu_data(freq->cpu), X86_FEATURE_CONSTANT_TSC))
295 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
297 lpj = &cpu_data(freq->cpu).loops_per_jiffy;
299 lpj = &boot_cpu_data.loops_per_jiffy;
303 ref_freq = freq->old;
304 loops_per_jiffy_ref = *lpj;
305 tsc_khz_ref = tsc_khz;
307 if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) ||
308 (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
309 (val == CPUFREQ_RESUMECHANGE)) {
310 *lpj = cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
312 tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
313 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
314 mark_tsc_unstable("cpufreq changes");
317 set_cyc2ns_scale(tsc_khz, freq->cpu);
322 static struct notifier_block time_cpufreq_notifier_block = {
323 .notifier_call = time_cpufreq_notifier
326 static int __init cpufreq_tsc(void)
330 if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
332 cpufreq_register_notifier(&time_cpufreq_notifier_block,
333 CPUFREQ_TRANSITION_NOTIFIER);
337 core_initcall(cpufreq_tsc);
339 #endif /* CONFIG_CPU_FREQ */
341 /* clocksource code */
343 static struct clocksource clocksource_tsc;
346 * We compare the TSC to the cycle_last value in the clocksource
347 * structure to avoid a nasty time-warp. This can be observed in a
348 * very small window right after one CPU updated cycle_last under
349 * xtime/vsyscall_gtod lock and the other CPU reads a TSC value which
350 * is smaller than the cycle_last reference value due to a TSC which
351 * is slighty behind. This delta is nowhere else observable, but in
352 * that case it results in a forward time jump in the range of hours
353 * due to the unsigned delta calculation of the time keeping core
354 * code, which is necessary to support wrapping clocksources like pm
357 static cycle_t read_tsc(void)
359 cycle_t ret = (cycle_t)get_cycles();
361 return ret >= clocksource_tsc.cycle_last ?
362 ret : clocksource_tsc.cycle_last;
366 static cycle_t __vsyscall_fn vread_tsc(void)
368 cycle_t ret = (cycle_t)vget_cycles();
370 return ret >= __vsyscall_gtod_data.clock.cycle_last ?
371 ret : __vsyscall_gtod_data.clock.cycle_last;
375 static struct clocksource clocksource_tsc = {
379 .mask = CLOCKSOURCE_MASK(64),
381 .flags = CLOCK_SOURCE_IS_CONTINUOUS |
382 CLOCK_SOURCE_MUST_VERIFY,
388 void mark_tsc_unstable(char *reason)
392 printk("Marking TSC unstable due to %s\n", reason);
393 /* Change only the rating, when not registered */
394 if (clocksource_tsc.mult)
395 clocksource_change_rating(&clocksource_tsc, 0);
397 clocksource_tsc.rating = 0;
401 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
403 static int __init dmi_mark_tsc_unstable(const struct dmi_system_id *d)
405 printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
411 /* List of systems that have known TSC problems */
412 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
414 .callback = dmi_mark_tsc_unstable,
415 .ident = "IBM Thinkpad 380XD",
417 DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
418 DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
425 * Geode_LX - the OLPC CPU has a possibly a very reliable TSC
427 #ifdef CONFIG_MGEODE_LX
428 /* RTSC counts during suspend */
429 #define RTSC_SUSP 0x100
431 static void __init check_geode_tsc_reliable(void)
433 unsigned long res_low, res_high;
435 rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
436 if (res_low & RTSC_SUSP)
437 clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
440 static inline void check_geode_tsc_reliable(void) { }
444 * Make an educated guess if the TSC is trustworthy and synchronized
447 __cpuinit int unsynchronized_tsc(void)
449 if (!cpu_has_tsc || tsc_unstable)
453 if (apic_is_clustered_box())
457 if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
460 * Intel systems are normally all synchronized.
461 * Exceptions must mark TSC as unstable:
463 if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
464 /* assume multi socket systems are not synchronized: */
465 if (num_possible_cpus() > 1)
472 static void __init init_tsc_clocksource(void)
474 clocksource_tsc.mult = clocksource_khz2mult(tsc_khz,
475 clocksource_tsc.shift);
476 /* lower the rating if we already know its unstable: */
477 if (check_tsc_unstable()) {
478 clocksource_tsc.rating = 0;
479 clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
481 clocksource_register(&clocksource_tsc);
484 void __init tsc_init(void)
492 tsc_khz = calibrate_tsc();
496 mark_tsc_unstable("could not calculate TSC khz");
501 if (cpu_has(&boot_cpu_data, X86_FEATURE_CONSTANT_TSC) &&
502 (boot_cpu_data.x86_vendor == X86_VENDOR_AMD))
503 cpu_khz = calibrate_cpu();
506 lpj = ((u64)tsc_khz * 1000);
510 printk("Detected %lu.%03lu MHz processor.\n",
511 (unsigned long)cpu_khz / 1000,
512 (unsigned long)cpu_khz % 1000);
515 * Secondary CPUs do not run through tsc_init(), so set up
516 * all the scale factors for all CPUs, assuming the same
517 * speed as the bootup CPU. (cpufreq notifiers will fix this
518 * up if their speed diverges)
520 for_each_possible_cpu(cpu)
521 set_cyc2ns_scale(cpu_khz, cpu);
523 if (tsc_disabled > 0)
526 /* now allow native_sched_clock() to use rdtsc */
530 /* Check and install the TSC clocksource */
531 dmi_check_system(bad_tsc_dmi_table);
533 if (unsynchronized_tsc())
534 mark_tsc_unstable("TSCs unsynchronized");
536 check_geode_tsc_reliable();
537 init_tsc_clocksource();