arch/i386/kernel/tsc.c

   1 /*
   2  * This code largely moved from arch/i386/kernel/timer/timer_tsc.c
   3  * which was originally moved from arch/i386/kernel/time.c.
   4  * See comments there for proper credits.
   5  */
   6
   7 #include <linux/workqueue.h>
   8 #include <linux/cpufreq.h>
   9 #include <linux/jiffies.h>
  10 #include <linux/init.h>
  11
  12 #include <asm/tsc.h>
  13 #include <asm/io.h>
  14
  15 #include "mach_timer.h"
  16
  17 /*
  18  * On some systems the TSC frequency does not
  19  * change with the cpu frequency. So we need
  20  * an extra value to store the TSC freq
  21  */
  22 unsigned int tsc_khz;
  23
  24 int tsc_disable __cpuinitdata = 0;
  25
  26 #ifdef CONFIG_X86_TSC
  27 static int __init tsc_setup(char *str)
  28 {
  29         printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, "
  30                                 "cannot disable TSC.\n");
  31         return 1;
  32 }
  33 #else
  34 /*
  35  * disable flag for tsc. Takes effect by clearing the TSC cpu flag
  36  * in cpu/common.c
  37  */
  38 static int __init tsc_setup(char *str)
  39 {
  40         tsc_disable = 1;
  41
  42         return 1;
  43 }
  44 #endif
  45
  46 __setup("notsc", tsc_setup);
  47
  48
  49 /*
  50  * code to mark and check if the TSC is unstable
  51  * due to cpufreq or due to unsynced TSCs
  52  */
  53 static int tsc_unstable;
  54
  55 static inline int check_tsc_unstable(void)
  56 {
  57         return tsc_unstable;
  58 }
  59
  60 void mark_tsc_unstable(void)
  61 {
  62         tsc_unstable = 1;
  63 }
  64 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
  65
  66 /* Accellerators for sched_clock()
  67  * convert from cycles(64bits) => nanoseconds (64bits)
  68  *  basic equation:
  69  *              ns = cycles / (freq / ns_per_sec)
  70  *              ns = cycles * (ns_per_sec / freq)
  71  *              ns = cycles * (10^9 / (cpu_khz * 10^3))
  72  *              ns = cycles * (10^6 / cpu_khz)
  73  *
  74  *      Then we use scaling math (suggested by george@mvista.com) to get:
  75  *              ns = cycles * (10^6 * SC / cpu_khz) / SC
  76  *              ns = cycles * cyc2ns_scale / SC
  77  *
  78  *      And since SC is a constant power of two, we can convert the div
  79  *  into a shift.
  80  *
  81  *  We can use khz divisor instead of mhz to keep a better percision, since
  82  *  cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
  83  *  (mathieu.desnoyers@polymtl.ca)
  84  *
  85  *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
  86  */
  87 static unsigned long cyc2ns_scale __read_mostly;
  88
  89 #define CYC2NS_SCALE_FACTOR 10 /* 2^10, carefully chosen */
  90
  91 static inline void set_cyc2ns_scale(unsigned long cpu_khz)
  92 {
  93         cyc2ns_scale = (1000000 << CYC2NS_SCALE_FACTOR)/cpu_khz;
  94 }
  95
  96 static inline unsigned long long cycles_2_ns(unsigned long long cyc)
  97 {
  98         return (cyc * cyc2ns_scale) >> CYC2NS_SCALE_FACTOR;
  99 }
 100
 101 /*
 102  * Scheduler clock - returns current time in nanosec units.
 103  */
 104 unsigned long long sched_clock(void)
 105 {
 106         unsigned long long this_offset;
 107
 108         /*
 109          * in the NUMA case we dont use the TSC as they are not
 110          * synchronized across all CPUs.
 111          */
 112 #ifndef CONFIG_NUMA
 113         if (!cpu_khz || check_tsc_unstable())
 114 #endif
 115                 /* no locking but a rare wrong value is not a big deal */
 116                 return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
 117
 118         /* read the Time Stamp Counter: */
 119         rdtscll(this_offset);
 120
 121         /* return the value in ns */
 122         return cycles_2_ns(this_offset);
 123 }
 124
 125 static unsigned long calculate_cpu_khz(void)
 126 {
 127         unsigned long long start, end;
 128         unsigned long count;
 129         u64 delta64;
 130         int i;
 131         unsigned long flags;
 132
 133         local_irq_save(flags);
 134
 135         /* run 3 times to ensure the cache is warm */
 136         for (i = 0; i < 3; i++) {
 137                 mach_prepare_counter();
 138                 rdtscll(start);
 139                 mach_countup(&count);
 140                 rdtscll(end);
 141         }
 142         /*
 143          * Error: ECTCNEVERSET
 144          * The CTC wasn't reliable: we got a hit on the very first read,
 145          * or the CPU was so fast/slow that the quotient wouldn't fit in
 146          * 32 bits..
 147          */
 148         if (count <= 1)
 149                 goto err;
 150
 151         delta64 = end - start;
 152
 153         /* cpu freq too fast: */
 154         if (delta64 > (1ULL<<32))
 155                 goto err;
 156
 157         /* cpu freq too slow: */
 158         if (delta64 <= CALIBRATE_TIME_MSEC)
 159                 goto err;
 160
 161         delta64 += CALIBRATE_TIME_MSEC/2; /* round for do_div */
 162         do_div(delta64,CALIBRATE_TIME_MSEC);
 163
 164         local_irq_restore(flags);
 165         return (unsigned long)delta64;
 166 err:
 167         local_irq_restore(flags);
 168         return 0;
 169 }
 170
 171 int recalibrate_cpu_khz(void)
 172 {
 173 #ifndef CONFIG_SMP
 174         unsigned long cpu_khz_old = cpu_khz;
 175
 176         if (cpu_has_tsc) {
 177                 cpu_khz = calculate_cpu_khz();
 178                 tsc_khz = cpu_khz;
 179                 cpu_data[0].loops_per_jiffy =
 180                         cpufreq_scale(cpu_data[0].loops_per_jiffy,
 181                                         cpu_khz_old, cpu_khz);
 182                 return 0;
 183         } else
 184                 return -ENODEV;
 185 #else
 186         return -ENODEV;
 187 #endif
 188 }
 189
 190 EXPORT_SYMBOL(recalibrate_cpu_khz);
 191
 192 void tsc_init(void)
 193 {
 194         if (!cpu_has_tsc || tsc_disable)
 195                 return;
 196
 197         cpu_khz = calculate_cpu_khz();
 198         tsc_khz = cpu_khz;
 199
 200         if (!cpu_khz)
 201                 return;
 202
 203         printk("Detected %lu.%03lu MHz processor.\n",
 204                                 (unsigned long)cpu_khz / 1000,
 205                                 (unsigned long)cpu_khz % 1000);
 206
 207         set_cyc2ns_scale(cpu_khz);
 208 }
 209
 210 #ifdef CONFIG_CPU_FREQ
 211
 212 static unsigned int cpufreq_delayed_issched = 0;
 213 static unsigned int cpufreq_init = 0;
 214 static struct work_struct cpufreq_delayed_get_work;
 215
 216 static void handle_cpufreq_delayed_get(void *v)
 217 {
 218         unsigned int cpu;
 219
 220         for_each_online_cpu(cpu)
 221                 cpufreq_get(cpu);
 222
 223         cpufreq_delayed_issched = 0;
 224 }
 225
 226 /*
 227  * if we notice cpufreq oddness, schedule a call to cpufreq_get() as it tries
 228  * to verify the CPU frequency the timing core thinks the CPU is running
 229  * at is still correct.
 230  */
 231 static inline void cpufreq_delayed_get(void)
 232 {
 233         if (cpufreq_init && !cpufreq_delayed_issched) {
 234                 cpufreq_delayed_issched = 1;
 235                 printk(KERN_DEBUG "Checking if CPU frequency changed.\n");
 236                 schedule_work(&cpufreq_delayed_get_work);
 237         }
 238 }
 239
 240 /*
 241  * if the CPU frequency is scaled, TSC-based delays will need a different
 242  * loops_per_jiffy value to function properly.
 243  */
 244 static unsigned int ref_freq = 0;
 245 static unsigned long loops_per_jiffy_ref = 0;
 246 static unsigned long cpu_khz_ref = 0;
 247
 248 static int
 249 time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, void *data)
 250 {
 251         struct cpufreq_freqs *freq = data;
 252
 253         if (val != CPUFREQ_RESUMECHANGE && val != CPUFREQ_SUSPENDCHANGE)
 254                 write_seqlock_irq(&xtime_lock);
 255
 256         if (!ref_freq) {
 257                 if (!freq->old){
 258                         ref_freq = freq->new;
 259                         goto end;
 260                 }
 261                 ref_freq = freq->old;
 262                 loops_per_jiffy_ref = cpu_data[freq->cpu].loops_per_jiffy;
 263                 cpu_khz_ref = cpu_khz;
 264         }
 265
 266         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
 267             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
 268             (val == CPUFREQ_RESUMECHANGE)) {
 269                 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
 270                         cpu_data[freq->cpu].loops_per_jiffy =
 271                                 cpufreq_scale(loops_per_jiffy_ref,
 272                                                 ref_freq, freq->new);
 273
 274                 if (cpu_khz) {
 275
 276                         if (num_online_cpus() == 1)
 277                                 cpu_khz = cpufreq_scale(cpu_khz_ref,
 278                                                 ref_freq, freq->new);
 279                         if (!(freq->flags & CPUFREQ_CONST_LOOPS)) {
 280                                 tsc_khz = cpu_khz;
 281                                 set_cyc2ns_scale(cpu_khz);
 282                                 /*
 283                                  * TSC based sched_clock turns
 284                                  * to junk w/ cpufreq
 285                                  */
 286                                 mark_tsc_unstable();
 287                         }
 288                 }
 289         }
 290 end:
 291         if (val != CPUFREQ_RESUMECHANGE && val != CPUFREQ_SUSPENDCHANGE)
 292                 write_sequnlock_irq(&xtime_lock);
 293
 294         return 0;
 295 }
 296
 297 static struct notifier_block time_cpufreq_notifier_block = {
 298         .notifier_call  = time_cpufreq_notifier
 299 };
 300
 301 static int __init cpufreq_tsc(void)
 302 {
 303         int ret;
 304
 305         INIT_WORK(&cpufreq_delayed_get_work, handle_cpufreq_delayed_get, NULL);
 306         ret = cpufreq_register_notifier(&time_cpufreq_notifier_block,
 307                                         CPUFREQ_TRANSITION_NOTIFIER);
 308         if (!ret)
 309                 cpufreq_init = 1;
 310
 311         return ret;
 312 }
 313
 314 core_initcall(cpufreq_tsc);
 315
 316 #endif