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/sched.h>
   8 #include <linux/clocksource.h>
   9 #include <linux/workqueue.h>
  10 #include <linux/cpufreq.h>
  11 #include <linux/jiffies.h>
  12 #include <linux/init.h>
  13 #include <linux/dmi.h>
  14
  15 #include <asm/delay.h>
  16 #include <asm/tsc.h>
  17 #include <asm/io.h>
  18 #include <asm/timer.h>
  19
  20 #include "mach_timer.h"
  21
  22 static int tsc_enabled;
  23
  24 /*
  25  * On some systems the TSC frequency does not
  26  * change with the cpu frequency. So we need
  27  * an extra value to store the TSC freq
  28  */
  29 unsigned int tsc_khz;
  30
  31 int tsc_disable;
  32
  33 #ifdef CONFIG_X86_TSC
  34 static int __init tsc_setup(char *str)
  35 {
  36         printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, "
  37                                 "cannot disable TSC.\n");
  38         return 1;
  39 }
  40 #else
  41 /*
  42  * disable flag for tsc. Takes effect by clearing the TSC cpu flag
  43  * in cpu/common.c
  44  */
  45 static int __init tsc_setup(char *str)
  46 {
  47         tsc_disable = 1;
  48
  49         return 1;
  50 }
  51 #endif
  52
  53 __setup("notsc", tsc_setup);
  54
  55 /*
  56  * code to mark and check if the TSC is unstable
  57  * due to cpufreq or due to unsynced TSCs
  58  */
  59 static int tsc_unstable;
  60
  61 static inline int check_tsc_unstable(void)
  62 {
  63         return tsc_unstable;
  64 }
  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          * Fall back to jiffies if there's no TSC available:
 110          * ( But note that we still use it if the TSC is marked
 111          *   unstable. We do this because unlike Time Of Day,
 112          *   the scheduler clock tolerates small errors and it's
 113          *   very important for it to be as fast as the platform
 114          *   can achive it. )
 115          */
 116         if (unlikely(!tsc_enabled && !tsc_unstable))
 117                 /* No locking but a rare wrong value is not a big deal: */
 118                 return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
 119
 120         /* read the Time Stamp Counter: */
 121         get_scheduled_cycles(this_offset);
 122
 123         /* return the value in ns */
 124         return cycles_2_ns(this_offset);
 125 }
 126
 127 unsigned long native_calculate_cpu_khz(void)
 128 {
 129         unsigned long long start, end;
 130         unsigned long count;
 131         u64 delta64;
 132         int i;
 133         unsigned long flags;
 134
 135         local_irq_save(flags);
 136
 137         /* run 3 times to ensure the cache is warm */
 138         for (i = 0; i < 3; i++) {
 139                 mach_prepare_counter();
 140                 rdtscll(start);
 141                 mach_countup(&count);
 142                 rdtscll(end);
 143         }
 144         /*
 145          * Error: ECTCNEVERSET
 146          * The CTC wasn't reliable: we got a hit on the very first read,
 147          * or the CPU was so fast/slow that the quotient wouldn't fit in
 148          * 32 bits..
 149          */
 150         if (count <= 1)
 151                 goto err;
 152
 153         delta64 = end - start;
 154
 155         /* cpu freq too fast: */
 156         if (delta64 > (1ULL<<32))
 157                 goto err;
 158
 159         /* cpu freq too slow: */
 160         if (delta64 <= CALIBRATE_TIME_MSEC)
 161                 goto err;
 162
 163         delta64 += CALIBRATE_TIME_MSEC/2; /* round for do_div */
 164         do_div(delta64,CALIBRATE_TIME_MSEC);
 165
 166         local_irq_restore(flags);
 167         return (unsigned long)delta64;
 168 err:
 169         local_irq_restore(flags);
 170         return 0;
 171 }
 172
 173 int recalibrate_cpu_khz(void)
 174 {
 175 #ifndef CONFIG_SMP
 176         unsigned long cpu_khz_old = cpu_khz;
 177
 178         if (cpu_has_tsc) {
 179                 cpu_khz = calculate_cpu_khz();
 180                 tsc_khz = cpu_khz;
 181                 cpu_data[0].loops_per_jiffy =
 182                         cpufreq_scale(cpu_data[0].loops_per_jiffy,
 183                                         cpu_khz_old, cpu_khz);
 184                 return 0;
 185         } else
 186                 return -ENODEV;
 187 #else
 188         return -ENODEV;
 189 #endif
 190 }
 191
 192 EXPORT_SYMBOL(recalibrate_cpu_khz);
 193
 194 #ifdef CONFIG_CPU_FREQ
 195
 196 /*
 197  * if the CPU frequency is scaled, TSC-based delays will need a different
 198  * loops_per_jiffy value to function properly.
 199  */
 200 static unsigned int ref_freq = 0;
 201 static unsigned long loops_per_jiffy_ref = 0;
 202 static unsigned long cpu_khz_ref = 0;
 203
 204 static int
 205 time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, void *data)
 206 {
 207         struct cpufreq_freqs *freq = data;
 208
 209         if (!ref_freq) {
 210                 if (!freq->old){
 211                         ref_freq = freq->new;
 212                         return 0;
 213                 }
 214                 ref_freq = freq->old;
 215                 loops_per_jiffy_ref = cpu_data[freq->cpu].loops_per_jiffy;
 216                 cpu_khz_ref = cpu_khz;
 217         }
 218
 219         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
 220             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
 221             (val == CPUFREQ_RESUMECHANGE)) {
 222                 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
 223                         cpu_data[freq->cpu].loops_per_jiffy =
 224                                 cpufreq_scale(loops_per_jiffy_ref,
 225                                                 ref_freq, freq->new);
 226
 227                 if (cpu_khz) {
 228
 229                         if (num_online_cpus() == 1)
 230                                 cpu_khz = cpufreq_scale(cpu_khz_ref,
 231                                                 ref_freq, freq->new);
 232                         if (!(freq->flags & CPUFREQ_CONST_LOOPS)) {
 233                                 tsc_khz = cpu_khz;
 234                                 set_cyc2ns_scale(cpu_khz);
 235                                 /*
 236                                  * TSC based sched_clock turns
 237                                  * to junk w/ cpufreq
 238                                  */
 239                                 mark_tsc_unstable("cpufreq changes");
 240                         }
 241                 }
 242         }
 243
 244         return 0;
 245 }
 246
 247 static struct notifier_block time_cpufreq_notifier_block = {
 248         .notifier_call  = time_cpufreq_notifier
 249 };
 250
 251 static int __init cpufreq_tsc(void)
 252 {
 253         return cpufreq_register_notifier(&time_cpufreq_notifier_block,
 254                                          CPUFREQ_TRANSITION_NOTIFIER);
 255 }
 256 core_initcall(cpufreq_tsc);
 257
 258 #endif
 259
 260 /* clock source code */
 261
 262 static unsigned long current_tsc_khz = 0;
 263
 264 static cycle_t read_tsc(void)
 265 {
 266         cycle_t ret;
 267
 268         rdtscll(ret);
 269
 270         return ret;
 271 }
 272
 273 static struct clocksource clocksource_tsc = {
 274         .name                   = "tsc",
 275         .rating                 = 300,
 276         .read                   = read_tsc,
 277         .mask                   = CLOCKSOURCE_MASK(64),
 278         .mult                   = 0, /* to be set */
 279         .shift                  = 22,
 280         .flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
 281                                   CLOCK_SOURCE_MUST_VERIFY,
 282 };
 283
 284 void mark_tsc_unstable(char *reason)
 285 {
 286         sched_clock_unstable_event();
 287         if (!tsc_unstable) {
 288                 tsc_unstable = 1;
 289                 tsc_enabled = 0;
 290                 printk("Marking TSC unstable due to: %s.\n", reason);
 291                 /* Can be called before registration */
 292                 if (clocksource_tsc.mult)
 293                         clocksource_change_rating(&clocksource_tsc, 0);
 294                 else
 295                         clocksource_tsc.rating = 0;
 296         }
 297 }
 298 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
 299
 300 static int __init dmi_mark_tsc_unstable(struct dmi_system_id *d)
 301 {
 302         printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
 303                        d->ident);
 304         tsc_unstable = 1;
 305         return 0;
 306 }
 307
 308 /* List of systems that have known TSC problems */
 309 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
 310         {
 311          .callback = dmi_mark_tsc_unstable,
 312          .ident = "IBM Thinkpad 380XD",
 313          .matches = {
 314                      DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
 315                      DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
 316                      },
 317          },
 318          {}
 319 };
 320
 321 /*
 322  * Make an educated guess if the TSC is trustworthy and synchronized
 323  * over all CPUs.
 324  */
 325 __cpuinit int unsynchronized_tsc(void)
 326 {
 327         if (!cpu_has_tsc || tsc_unstable)
 328                 return 1;
 329         /*
 330          * Intel systems are normally all synchronized.
 331          * Exceptions must mark TSC as unstable:
 332          */
 333         if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
 334                 /* assume multi socket systems are not synchronized: */
 335                 if (num_possible_cpus() > 1)
 336                         tsc_unstable = 1;
 337         }
 338         return tsc_unstable;
 339 }
 340
 341 /*
 342  * Geode_LX - the OLPC CPU has a possibly a very reliable TSC
 343  */
 344 #ifdef CONFIG_MGEODE_LX
 345 /* RTSC counts during suspend */
 346 #define RTSC_SUSP 0x100
 347
 348 static void __init check_geode_tsc_reliable(void)
 349 {
 350         unsigned long val;
 351
 352         rdmsrl(MSR_GEODE_BUSCONT_CONF0, val);
 353         if ((val & RTSC_SUSP))
 354                 clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
 355 }
 356 #else
 357 static inline void check_geode_tsc_reliable(void) { }
 358 #endif
 359
 360
 361 void __init tsc_init(void)
 362 {
 363         if (!cpu_has_tsc || tsc_disable)
 364                 goto out_no_tsc;
 365
 366         cpu_khz = calculate_cpu_khz();
 367         tsc_khz = cpu_khz;
 368
 369         if (!cpu_khz)
 370                 goto out_no_tsc;
 371
 372         printk("Detected %lu.%03lu MHz processor.\n",
 373                                 (unsigned long)cpu_khz / 1000,
 374                                 (unsigned long)cpu_khz % 1000);
 375
 376         set_cyc2ns_scale(cpu_khz);
 377         use_tsc_delay();
 378
 379         /* Check and install the TSC clocksource */
 380         dmi_check_system(bad_tsc_dmi_table);
 381
 382         unsynchronized_tsc();
 383         check_geode_tsc_reliable();
 384         current_tsc_khz = tsc_khz;
 385         clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz,
 386                                                         clocksource_tsc.shift);
 387         /* lower the rating if we already know its unstable: */
 388         if (check_tsc_unstable()) {
 389                 clocksource_tsc.rating = 0;
 390                 clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
 391         } else
 392                 tsc_enabled = 1;
 393
 394         clocksource_register(&clocksource_tsc);
 395
 396         return;
 397
 398 out_no_tsc:
 399         /*
 400          * Set the tsc_disable flag if there's no TSC support, this
 401          * makes it a fast flag for the kernel to see whether it
 402          * should be using the TSC.
 403          */
 404         tsc_disable = 1;
 405 }