arch/x86/kernel/tsc_64.c

   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>
  10
  11 #include <asm/hpet.h>
  12 #include <asm/timex.h>
  13 #include <asm/timer.h>
  14
  15 static int notsc __initdata = 0;
  16
  17 unsigned int cpu_khz;           /* TSC clocks / usec, not used here */
  18 EXPORT_SYMBOL(cpu_khz);
  19 unsigned int tsc_khz;
  20 EXPORT_SYMBOL(tsc_khz);
  21
  22 /* Accelerators for sched_clock()
  23  * convert from cycles(64bits) => nanoseconds (64bits)
  24  *  basic equation:
  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)
  29  *
  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
  33  *
  34  *      And since SC is a constant power of two, we can convert the div
  35  *  into a shift.
  36  *
  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)
  40  *
  41  *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
  42  */
  43 DEFINE_PER_CPU(unsigned long, cyc2ns);
  44
  45 static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu)
  46 {
  47         unsigned long long tsc_now, ns_now;
  48         unsigned long flags, *scale;
  49
  50         local_irq_save(flags);
  51         sched_clock_idle_sleep_event();
  52
  53         scale = &per_cpu(cyc2ns, cpu);
  54
  55         rdtscll(tsc_now);
  56         ns_now = __cycles_2_ns(tsc_now);
  57
  58         if (cpu_khz)
  59                 *scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz;
  60
  61         sched_clock_idle_wakeup_event(0);
  62         local_irq_restore(flags);
  63 }
  64
  65 unsigned long long native_sched_clock(void)
  66 {
  67         unsigned long a = 0;
  68
  69         /* Could do CPU core sync here. Opteron can execute rdtsc speculatively,
  70          * which means it is not completely exact and may not be monotonous
  71          * between CPUs. But the errors should be too small to matter for
  72          * scheduling purposes.
  73          */
  74
  75         rdtscll(a);
  76         return cycles_2_ns(a);
  77 }
  78
  79 /* We need to define a real function for sched_clock, to override the
  80    weak default version */
  81 #ifdef CONFIG_PARAVIRT
  82 unsigned long long sched_clock(void)
  83 {
  84         return paravirt_sched_clock();
  85 }
  86 #else
  87 unsigned long long
  88 sched_clock(void) __attribute__((alias("native_sched_clock")));
  89 #endif
  90
  91
  92 static int tsc_unstable;
  93
  94 int check_tsc_unstable(void)
  95 {
  96         return tsc_unstable;
  97 }
  98 EXPORT_SYMBOL_GPL(check_tsc_unstable);
  99
 100 #ifdef CONFIG_CPU_FREQ
 101
 102 /* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
 103  * changes.
 104  *
 105  * RED-PEN: On SMP we assume all CPUs run with the same frequency.  It's
 106  * not that important because current Opteron setups do not support
 107  * scaling on SMP anyroads.
 108  *
 109  * Should fix up last_tsc too. Currently gettimeofday in the
 110  * first tick after the change will be slightly wrong.
 111  */
 112
 113 static unsigned int  ref_freq;
 114 static unsigned long loops_per_jiffy_ref;
 115 static unsigned long tsc_khz_ref;
 116
 117 static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
 118                                  void *data)
 119 {
 120         struct cpufreq_freqs *freq = data;
 121         unsigned long *lpj, dummy;
 122
 123         if (cpu_has(&cpu_data(freq->cpu), X86_FEATURE_CONSTANT_TSC))
 124                 return 0;
 125
 126         lpj = &dummy;
 127         if (!(freq->flags & CPUFREQ_CONST_LOOPS))
 128 #ifdef CONFIG_SMP
 129                 lpj = &cpu_data(freq->cpu).loops_per_jiffy;
 130 #else
 131                 lpj = &boot_cpu_data.loops_per_jiffy;
 132 #endif
 133
 134         if (!ref_freq) {
 135                 ref_freq = freq->old;
 136                 loops_per_jiffy_ref = *lpj;
 137                 tsc_khz_ref = tsc_khz;
 138         }
 139         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
 140                 (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
 141                 (val == CPUFREQ_RESUMECHANGE)) {
 142                 *lpj =
 143                 cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
 144
 145                 tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
 146                 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
 147                         mark_tsc_unstable("cpufreq changes");
 148         }
 149
 150         set_cyc2ns_scale(tsc_khz_ref, freq->cpu);
 151
 152         return 0;
 153 }
 154
 155 static struct notifier_block time_cpufreq_notifier_block = {
 156         .notifier_call  = time_cpufreq_notifier
 157 };
 158
 159 static int __init cpufreq_tsc(void)
 160 {
 161         cpufreq_register_notifier(&time_cpufreq_notifier_block,
 162                                   CPUFREQ_TRANSITION_NOTIFIER);
 163         return 0;
 164 }
 165
 166 core_initcall(cpufreq_tsc);
 167
 168 #endif
 169
 170 #define MAX_RETRIES     5
 171 #define SMI_TRESHOLD    50000
 172
 173 /*
 174  * Read TSC and the reference counters. Take care of SMI disturbance
 175  */
 176 static unsigned long __init tsc_read_refs(unsigned long *pm,
 177                                           unsigned long *hpet)
 178 {
 179         unsigned long t1, t2;
 180         int i;
 181
 182         for (i = 0; i < MAX_RETRIES; i++) {
 183                 t1 = get_cycles();
 184                 if (hpet)
 185                         *hpet = hpet_readl(HPET_COUNTER) & 0xFFFFFFFF;
 186                 else
 187                         *pm = acpi_pm_read_early();
 188                 t2 = get_cycles();
 189                 if ((t2 - t1) < SMI_TRESHOLD)
 190                         return t2;
 191         }
 192         return ULONG_MAX;
 193 }
 194
 195 /**
 196  * tsc_calibrate - calibrate the tsc on boot
 197  */
 198 void __init tsc_calibrate(void)
 199 {
 200         unsigned long flags, tsc1, tsc2, tr1, tr2, pm1, pm2, hpet1, hpet2;
 201         int hpet = is_hpet_enabled(), cpu;
 202
 203         local_irq_save(flags);
 204
 205         tsc1 = tsc_read_refs(&pm1, hpet ? &hpet1 : NULL);
 206
 207         outb((inb(0x61) & ~0x02) | 0x01, 0x61);
 208
 209         outb(0xb0, 0x43);
 210         outb((CLOCK_TICK_RATE / (1000 / 50)) & 0xff, 0x42);
 211         outb((CLOCK_TICK_RATE / (1000 / 50)) >> 8, 0x42);
 212         tr1 = get_cycles();
 213         while ((inb(0x61) & 0x20) == 0);
 214         tr2 = get_cycles();
 215
 216         tsc2 = tsc_read_refs(&pm2, hpet ? &hpet2 : NULL);
 217
 218         local_irq_restore(flags);
 219
 220         /*
 221          * Preset the result with the raw and inaccurate PIT
 222          * calibration value
 223          */
 224         tsc_khz = (tr2 - tr1) / 50;
 225
 226         /* hpet or pmtimer available ? */
 227         if (!hpet && !pm1 && !pm2) {
 228                 printk(KERN_INFO "TSC calibrated against PIT\n");
 229                 return;
 230         }
 231
 232         /* Check, whether the sampling was disturbed by an SMI */
 233         if (tsc1 == ULONG_MAX || tsc2 == ULONG_MAX) {
 234                 printk(KERN_WARNING "TSC calibration disturbed by SMI, "
 235                        "using PIT calibration result\n");
 236                 return;
 237         }
 238
 239         tsc2 = (tsc2 - tsc1) * 1000000L;
 240
 241         if (hpet) {
 242                 printk(KERN_INFO "TSC calibrated against HPET\n");
 243                 if (hpet2 < hpet1)
 244                         hpet2 += 0x100000000;
 245                 hpet2 -= hpet1;
 246                 tsc1 = (hpet2 * hpet_readl(HPET_PERIOD)) / 1000000;
 247         } else {
 248                 printk(KERN_INFO "TSC calibrated against PM_TIMER\n");
 249                 if (pm2 < pm1)
 250                         pm2 += ACPI_PM_OVRRUN;
 251                 pm2 -= pm1;
 252                 tsc1 = (pm2 * 1000000000) / PMTMR_TICKS_PER_SEC;
 253         }
 254
 255         tsc_khz = tsc2 / tsc1;
 256
 257         for_each_possible_cpu(cpu)
 258                 set_cyc2ns_scale(tsc_khz, cpu);
 259 }
 260
 261 /*
 262  * Make an educated guess if the TSC is trustworthy and synchronized
 263  * over all CPUs.
 264  */
 265 __cpuinit int unsynchronized_tsc(void)
 266 {
 267         if (tsc_unstable)
 268                 return 1;
 269
 270 #ifdef CONFIG_SMP
 271         if (apic_is_clustered_box())
 272                 return 1;
 273 #endif
 274
 275         if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
 276                 return 0;
 277
 278         /* Assume multi socket systems are not synchronized */
 279         return num_present_cpus() > 1;
 280 }
 281
 282 int __init notsc_setup(char *s)
 283 {
 284         notsc = 1;
 285         return 1;
 286 }
 287
 288 __setup("notsc", notsc_setup);
 289
 290
 291 /* clock source code: */
 292 static cycle_t read_tsc(void)
 293 {
 294         cycle_t ret = (cycle_t)get_cycles();
 295         return ret;
 296 }
 297
 298 static cycle_t __vsyscall_fn vread_tsc(void)
 299 {
 300         cycle_t ret = (cycle_t)vget_cycles();
 301         return ret;
 302 }
 303
 304 static struct clocksource clocksource_tsc = {
 305         .name                   = "tsc",
 306         .rating                 = 300,
 307         .read                   = read_tsc,
 308         .mask                   = CLOCKSOURCE_MASK(64),
 309         .shift                  = 22,
 310         .flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
 311                                   CLOCK_SOURCE_MUST_VERIFY,
 312         .vread                  = vread_tsc,
 313 };
 314
 315 void mark_tsc_unstable(char *reason)
 316 {
 317         if (!tsc_unstable) {
 318                 tsc_unstable = 1;
 319                 printk("Marking TSC unstable due to %s\n", reason);
 320                 /* Change only the rating, when not registered */
 321                 if (clocksource_tsc.mult)
 322                         clocksource_change_rating(&clocksource_tsc, 0);
 323                 else
 324                         clocksource_tsc.rating = 0;
 325         }
 326 }
 327 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
 328
 329 void __init init_tsc_clocksource(void)
 330 {
 331         if (!notsc) {
 332                 clocksource_tsc.mult = clocksource_khz2mult(tsc_khz,
 333                                                         clocksource_tsc.shift);
 334                 if (check_tsc_unstable())
 335                         clocksource_tsc.rating = 0;
 336
 337                 clocksource_register(&clocksource_tsc);
 338         }
 339 }