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 flags, prev_scale, *scale;
  48         unsigned long long tsc_now, ns_now;
  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         prev_scale = *scale;
  59         if (cpu_khz)
  60                 *scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz;
  61
  62         sched_clock_idle_wakeup_event(0);
  63         local_irq_restore(flags);
  64 }
  65
  66 unsigned long long native_sched_clock(void)
  67 {
  68         unsigned long a = 0;
  69
  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.
  74          */
  75
  76         rdtscll(a);
  77         return cycles_2_ns(a);
  78 }
  79
  80 /* We need to define a real function for sched_clock, to override the
  81    weak default version */
  82 #ifdef CONFIG_PARAVIRT
  83 unsigned long long sched_clock(void)
  84 {
  85         return paravirt_sched_clock();
  86 }
  87 #else
  88 unsigned long long
  89 sched_clock(void) __attribute__((alias("native_sched_clock")));
  90 #endif
  91
  92
  93 static int tsc_unstable;
  94
  95 int check_tsc_unstable(void)
  96 {
  97         return tsc_unstable;
  98 }
  99 EXPORT_SYMBOL_GPL(check_tsc_unstable);
 100
 101 #ifdef CONFIG_CPU_FREQ
 102
 103 /* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
 104  * changes.
 105  *
 106  * RED-PEN: On SMP we assume all CPUs run with the same frequency.  It's
 107  * not that important because current Opteron setups do not support
 108  * scaling on SMP anyroads.
 109  *
 110  * Should fix up last_tsc too. Currently gettimeofday in the
 111  * first tick after the change will be slightly wrong.
 112  */
 113
 114 static unsigned int  ref_freq;
 115 static unsigned long loops_per_jiffy_ref;
 116 static unsigned long tsc_khz_ref;
 117
 118 static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
 119                                  void *data)
 120 {
 121         struct cpufreq_freqs *freq = data;
 122         unsigned long *lpj, dummy;
 123
 124         if (cpu_has(&cpu_data(freq->cpu), X86_FEATURE_CONSTANT_TSC))
 125                 return 0;
 126
 127         lpj = &dummy;
 128         if (!(freq->flags & CPUFREQ_CONST_LOOPS))
 129 #ifdef CONFIG_SMP
 130                 lpj = &cpu_data(freq->cpu).loops_per_jiffy;
 131 #else
 132                 lpj = &boot_cpu_data.loops_per_jiffy;
 133 #endif
 134
 135         if (!ref_freq) {
 136                 ref_freq = freq->old;
 137                 loops_per_jiffy_ref = *lpj;
 138                 tsc_khz_ref = tsc_khz;
 139         }
 140         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
 141                 (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
 142                 (val == CPUFREQ_RESUMECHANGE)) {
 143                 *lpj =
 144                 cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
 145
 146                 tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
 147                 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
 148                         mark_tsc_unstable("cpufreq changes");
 149         }
 150
 151         preempt_disable();
 152         set_cyc2ns_scale(tsc_khz_ref, smp_processor_id());
 153         preempt_enable();
 154
 155         return 0;
 156 }
 157
 158 static struct notifier_block time_cpufreq_notifier_block = {
 159         .notifier_call  = time_cpufreq_notifier
 160 };
 161
 162 static int __init cpufreq_tsc(void)
 163 {
 164         cpufreq_register_notifier(&time_cpufreq_notifier_block,
 165                                   CPUFREQ_TRANSITION_NOTIFIER);
 166         return 0;
 167 }
 168
 169 core_initcall(cpufreq_tsc);
 170
 171 #endif
 172
 173 #define MAX_RETRIES     5
 174 #define SMI_TRESHOLD    50000
 175
 176 /*
 177  * Read TSC and the reference counters. Take care of SMI disturbance
 178  */
 179 static unsigned long __init tsc_read_refs(unsigned long *pm,
 180                                           unsigned long *hpet)
 181 {
 182         unsigned long t1, t2;
 183         int i;
 184
 185         for (i = 0; i < MAX_RETRIES; i++) {
 186                 t1 = get_cycles();
 187                 if (hpet)
 188                         *hpet = hpet_readl(HPET_COUNTER) & 0xFFFFFFFF;
 189                 else
 190                         *pm = acpi_pm_read_early();
 191                 t2 = get_cycles();
 192                 if ((t2 - t1) < SMI_TRESHOLD)
 193                         return t2;
 194         }
 195         return ULONG_MAX;
 196 }
 197
 198 /**
 199  * tsc_calibrate - calibrate the tsc on boot
 200  */
 201 void __init tsc_calibrate(void)
 202 {
 203         unsigned long flags, tsc1, tsc2, tr1, tr2, pm1, pm2, hpet1, hpet2;
 204         int hpet = is_hpet_enabled(), cpu;
 205
 206         local_irq_save(flags);
 207
 208         tsc1 = tsc_read_refs(&pm1, hpet ? &hpet1 : NULL);
 209
 210         outb((inb(0x61) & ~0x02) | 0x01, 0x61);
 211
 212         outb(0xb0, 0x43);
 213         outb((CLOCK_TICK_RATE / (1000 / 50)) & 0xff, 0x42);
 214         outb((CLOCK_TICK_RATE / (1000 / 50)) >> 8, 0x42);
 215         tr1 = get_cycles();
 216         while ((inb(0x61) & 0x20) == 0);
 217         tr2 = get_cycles();
 218
 219         tsc2 = tsc_read_refs(&pm2, hpet ? &hpet2 : NULL);
 220
 221         local_irq_restore(flags);
 222
 223         /*
 224          * Preset the result with the raw and inaccurate PIT
 225          * calibration value
 226          */
 227         tsc_khz = (tr2 - tr1) / 50;
 228
 229         /* hpet or pmtimer available ? */
 230         if (!hpet && !pm1 && !pm2) {
 231                 printk(KERN_INFO "TSC calibrated against PIT\n");
 232                 return;
 233         }
 234
 235         /* Check, whether the sampling was disturbed by an SMI */
 236         if (tsc1 == ULONG_MAX || tsc2 == ULONG_MAX) {
 237                 printk(KERN_WARNING "TSC calibration disturbed by SMI, "
 238                        "using PIT calibration result\n");
 239                 return;
 240         }
 241
 242         tsc2 = (tsc2 - tsc1) * 1000000L;
 243
 244         if (hpet) {
 245                 printk(KERN_INFO "TSC calibrated against HPET\n");
 246                 if (hpet2 < hpet1)
 247                         hpet2 += 0x100000000;
 248                 hpet2 -= hpet1;
 249                 tsc1 = (hpet2 * hpet_readl(HPET_PERIOD)) / 1000000;
 250         } else {
 251                 printk(KERN_INFO "TSC calibrated against PM_TIMER\n");
 252                 if (pm2 < pm1)
 253                         pm2 += ACPI_PM_OVRRUN;
 254                 pm2 -= pm1;
 255                 tsc1 = (pm2 * 1000000000) / PMTMR_TICKS_PER_SEC;
 256         }
 257
 258         tsc_khz = tsc2 / tsc1;
 259
 260         for_each_possible_cpu(cpu)
 261                 set_cyc2ns_scale(tsc_khz, cpu);
 262 }
 263
 264 /*
 265  * Make an educated guess if the TSC is trustworthy and synchronized
 266  * over all CPUs.
 267  */
 268 __cpuinit int unsynchronized_tsc(void)
 269 {
 270         if (tsc_unstable)
 271                 return 1;
 272
 273 #ifdef CONFIG_SMP
 274         if (apic_is_clustered_box())
 275                 return 1;
 276 #endif
 277
 278         if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
 279                 return 0;
 280
 281         /* Assume multi socket systems are not synchronized */
 282         return num_present_cpus() > 1;
 283 }
 284
 285 int __init notsc_setup(char *s)
 286 {
 287         notsc = 1;
 288         return 1;
 289 }
 290
 291 __setup("notsc", notsc_setup);
 292
 293
 294 /* clock source code: */
 295 static cycle_t read_tsc(void)
 296 {
 297         cycle_t ret = (cycle_t)get_cycles();
 298         return ret;
 299 }
 300
 301 static cycle_t __vsyscall_fn vread_tsc(void)
 302 {
 303         cycle_t ret = (cycle_t)vget_cycles();
 304         return ret;
 305 }
 306
 307 static struct clocksource clocksource_tsc = {
 308         .name                   = "tsc",
 309         .rating                 = 300,
 310         .read                   = read_tsc,
 311         .mask                   = CLOCKSOURCE_MASK(64),
 312         .shift                  = 22,
 313         .flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
 314                                   CLOCK_SOURCE_MUST_VERIFY,
 315         .vread                  = vread_tsc,
 316 };
 317
 318 void mark_tsc_unstable(char *reason)
 319 {
 320         if (!tsc_unstable) {
 321                 tsc_unstable = 1;
 322                 printk("Marking TSC unstable due to %s\n", reason);
 323                 /* Change only the rating, when not registered */
 324                 if (clocksource_tsc.mult)
 325                         clocksource_change_rating(&clocksource_tsc, 0);
 326                 else
 327                         clocksource_tsc.rating = 0;
 328         }
 329 }
 330 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
 331
 332 void __init init_tsc_clocksource(void)
 333 {
 334         if (!notsc) {
 335                 clocksource_tsc.mult = clocksource_khz2mult(tsc_khz,
 336                                                         clocksource_tsc.shift);
 337                 if (check_tsc_unstable())
 338                         clocksource_tsc.rating = 0;
 339
 340                 clocksource_register(&clocksource_tsc);
 341         }
 342 }