1 #include <linux/kernel.h>
2 #include <linux/sched.h>
3 #include <linux/interrupt.h>
4 #include <linux/init.h>
5 #include <linux/clocksource.h>
6 #include <linux/time.h>
7 #include <linux/acpi.h>
8 #include <linux/cpufreq.h>
9 #include <linux/acpi_pmtmr.h>
12 #include <asm/timex.h>
13 #include <asm/timer.h>
15 static int notsc __initdata = 0;
17 unsigned int cpu_khz; /* TSC clocks / usec, not used here */
18 EXPORT_SYMBOL(cpu_khz);
20 EXPORT_SYMBOL(tsc_khz);
22 /* Accelerators for sched_clock()
23 * convert from cycles(64bits) => nanoseconds (64bits)
25 * ns = cycles / (freq / ns_per_sec)
26 * ns = cycles * (ns_per_sec / freq)
27 * ns = cycles * (10^9 / (cpu_khz * 10^3))
28 * ns = cycles * (10^6 / cpu_khz)
30 * Then we use scaling math (suggested by george@mvista.com) to get:
31 * ns = cycles * (10^6 * SC / cpu_khz) / SC
32 * ns = cycles * cyc2ns_scale / SC
34 * And since SC is a constant power of two, we can convert the div
37 * We can use khz divisor instead of mhz to keep a better precision, since
38 * cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
39 * (mathieu.desnoyers@polymtl.ca)
41 * -johnstul@us.ibm.com "math is hard, lets go shopping!"
43 DEFINE_PER_CPU(unsigned long, cyc2ns);
45 static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu)
47 unsigned long flags, prev_scale, *scale;
48 unsigned long long tsc_now, ns_now;
50 local_irq_save(flags);
51 sched_clock_idle_sleep_event();
53 scale = &per_cpu(cyc2ns, cpu);
56 ns_now = __cycles_2_ns(tsc_now);
60 *scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz;
62 sched_clock_idle_wakeup_event(0);
63 local_irq_restore(flags);
66 unsigned long long sched_clock(void)
70 /* Could do CPU core sync here. Opteron can execute rdtsc speculatively,
71 * which means it is not completely exact and may not be monotonous
72 * between CPUs. But the errors should be too small to matter for
73 * scheduling purposes.
77 return cycles_2_ns(a);
80 static int tsc_unstable;
82 inline int check_tsc_unstable(void)
86 #ifdef CONFIG_CPU_FREQ
88 /* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
91 * RED-PEN: On SMP we assume all CPUs run with the same frequency. It's
92 * not that important because current Opteron setups do not support
93 * scaling on SMP anyroads.
95 * Should fix up last_tsc too. Currently gettimeofday in the
96 * first tick after the change will be slightly wrong.
99 static unsigned int ref_freq;
100 static unsigned long loops_per_jiffy_ref;
101 static unsigned long tsc_khz_ref;
103 static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
106 struct cpufreq_freqs *freq = data;
107 unsigned long *lpj, dummy;
109 if (cpu_has(&cpu_data(freq->cpu), X86_FEATURE_CONSTANT_TSC))
113 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
115 lpj = &cpu_data(freq->cpu).loops_per_jiffy;
117 lpj = &boot_cpu_data.loops_per_jiffy;
121 ref_freq = freq->old;
122 loops_per_jiffy_ref = *lpj;
123 tsc_khz_ref = tsc_khz;
125 if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) ||
126 (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
127 (val == CPUFREQ_RESUMECHANGE)) {
129 cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
131 tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
132 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
133 mark_tsc_unstable("cpufreq changes");
137 set_cyc2ns_scale(tsc_khz_ref, smp_processor_id());
143 static struct notifier_block time_cpufreq_notifier_block = {
144 .notifier_call = time_cpufreq_notifier
147 static int __init cpufreq_tsc(void)
149 cpufreq_register_notifier(&time_cpufreq_notifier_block,
150 CPUFREQ_TRANSITION_NOTIFIER);
154 core_initcall(cpufreq_tsc);
158 #define MAX_RETRIES 5
159 #define SMI_TRESHOLD 50000
162 * Read TSC and the reference counters. Take care of SMI disturbance
164 static unsigned long __init tsc_read_refs(unsigned long *pm,
167 unsigned long t1, t2;
170 for (i = 0; i < MAX_RETRIES; i++) {
171 t1 = get_cycles_sync();
173 *hpet = hpet_readl(HPET_COUNTER) & 0xFFFFFFFF;
175 *pm = acpi_pm_read_early();
176 t2 = get_cycles_sync();
177 if ((t2 - t1) < SMI_TRESHOLD)
184 * tsc_calibrate - calibrate the tsc on boot
186 void __init tsc_calibrate(void)
188 unsigned long flags, tsc1, tsc2, tr1, tr2, pm1, pm2, hpet1, hpet2;
189 int hpet = is_hpet_enabled(), cpu;
191 local_irq_save(flags);
193 tsc1 = tsc_read_refs(&pm1, hpet ? &hpet1 : NULL);
195 outb((inb(0x61) & ~0x02) | 0x01, 0x61);
198 outb((CLOCK_TICK_RATE / (1000 / 50)) & 0xff, 0x42);
199 outb((CLOCK_TICK_RATE / (1000 / 50)) >> 8, 0x42);
200 tr1 = get_cycles_sync();
201 while ((inb(0x61) & 0x20) == 0);
202 tr2 = get_cycles_sync();
204 tsc2 = tsc_read_refs(&pm2, hpet ? &hpet2 : NULL);
206 local_irq_restore(flags);
209 * Preset the result with the raw and inaccurate PIT
212 tsc_khz = (tr2 - tr1) / 50;
214 /* hpet or pmtimer available ? */
215 if (!hpet && !pm1 && !pm2) {
216 printk(KERN_INFO "TSC calibrated against PIT\n");
220 /* Check, whether the sampling was disturbed by an SMI */
221 if (tsc1 == ULONG_MAX || tsc2 == ULONG_MAX) {
222 printk(KERN_WARNING "TSC calibration disturbed by SMI, "
223 "using PIT calibration result\n");
227 tsc2 = (tsc2 - tsc1) * 1000000L;
230 printk(KERN_INFO "TSC calibrated against HPET\n");
232 hpet2 += 0x100000000;
234 tsc1 = (hpet2 * hpet_readl(HPET_PERIOD)) / 1000000;
236 printk(KERN_INFO "TSC calibrated against PM_TIMER\n");
238 pm2 += ACPI_PM_OVRRUN;
240 tsc1 = (pm2 * 1000000000) / PMTMR_TICKS_PER_SEC;
243 tsc_khz = tsc2 / tsc1;
245 for_each_possible_cpu(cpu)
246 set_cyc2ns_scale(tsc_khz, cpu);
250 * Make an educated guess if the TSC is trustworthy and synchronized
253 __cpuinit int unsynchronized_tsc(void)
259 if (apic_is_clustered_box())
262 /* Most intel systems have synchronized TSCs except for
263 multi node systems */
264 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) {
266 /* But TSC doesn't tick in C3 so don't use it there */
267 if (acpi_gbl_FADT.header.length > 0 &&
268 acpi_gbl_FADT.C3latency < 1000)
274 /* Assume multi socket systems are not synchronized */
275 return num_present_cpus() > 1;
278 int __init notsc_setup(char *s)
284 __setup("notsc", notsc_setup);
287 /* clock source code: */
288 static cycle_t read_tsc(void)
290 cycle_t ret = (cycle_t)get_cycles_sync();
294 static cycle_t __vsyscall_fn vread_tsc(void)
296 cycle_t ret = (cycle_t)get_cycles_sync();
300 static struct clocksource clocksource_tsc = {
304 .mask = CLOCKSOURCE_MASK(64),
306 .flags = CLOCK_SOURCE_IS_CONTINUOUS |
307 CLOCK_SOURCE_MUST_VERIFY,
311 void mark_tsc_unstable(char *reason)
315 printk("Marking TSC unstable due to %s\n", reason);
316 /* Change only the rating, when not registered */
317 if (clocksource_tsc.mult)
318 clocksource_change_rating(&clocksource_tsc, 0);
320 clocksource_tsc.rating = 0;
323 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
325 void __init init_tsc_clocksource(void)
328 clocksource_tsc.mult = clocksource_khz2mult(tsc_khz,
329 clocksource_tsc.shift);
330 if (check_tsc_unstable())
331 clocksource_tsc.rating = 0;
333 clocksource_register(&clocksource_tsc);