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 EXPORT_SYMBOL_GPL(tsc_khz);
  31
  32 int tsc_disable;
  33
  34 #ifdef CONFIG_X86_TSC
  35 static int __init tsc_setup(char *str)
  36 {
  37         printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, "
  38                                 "cannot disable TSC.\n");
  39         return 1;
  40 }
  41 #else
  42 /*
  43  * disable flag for tsc. Takes effect by clearing the TSC cpu flag
  44  * in cpu/common.c
  45  */
  46 static int __init tsc_setup(char *str)
  47 {
  48         tsc_disable = 1;
  49
  50         return 1;
  51 }
  52 #endif
  53
  54 __setup("notsc", tsc_setup);
  55
  56 /*
  57  * code to mark and check if the TSC is unstable
  58  * due to cpufreq or due to unsynced TSCs
  59  */
  60 static int tsc_unstable;
  61
  62 int check_tsc_unstable(void)
  63 {
  64         return tsc_unstable;
  65 }
  66 EXPORT_SYMBOL_GPL(check_tsc_unstable);
  67
  68 /* Accellerators for sched_clock()
  69  * convert from cycles(64bits) => nanoseconds (64bits)
  70  *  basic equation:
  71  *              ns = cycles / (freq / ns_per_sec)
  72  *              ns = cycles * (ns_per_sec / freq)
  73  *              ns = cycles * (10^9 / (cpu_khz * 10^3))
  74  *              ns = cycles * (10^6 / cpu_khz)
  75  *
  76  *      Then we use scaling math (suggested by george@mvista.com) to get:
  77  *              ns = cycles * (10^6 * SC / cpu_khz) / SC
  78  *              ns = cycles * cyc2ns_scale / SC
  79  *
  80  *      And since SC is a constant power of two, we can convert the div
  81  *  into a shift.
  82  *
  83  *  We can use khz divisor instead of mhz to keep a better percision, since
  84  *  cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
  85  *  (mathieu.desnoyers@polymtl.ca)
  86  *
  87  *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
  88  */
  89 unsigned long cyc2ns_scale __read_mostly;
  90
  91 #define CYC2NS_SCALE_FACTOR 10 /* 2^10, carefully chosen */
  92
  93 static inline void set_cyc2ns_scale(unsigned long cpu_khz)
  94 {
  95         cyc2ns_scale = (1000000 << CYC2NS_SCALE_FACTOR)/cpu_khz;
  96 }
  97
  98 /*
  99  * Scheduler clock - returns current time in nanosec units.
 100  */
 101 unsigned long long native_sched_clock(void)
 102 {
 103         unsigned long long this_offset;
 104
 105         /*
 106          * Fall back to jiffies if there's no TSC available:
 107          * ( But note that we still use it if the TSC is marked
 108          *   unstable. We do this because unlike Time Of Day,
 109          *   the scheduler clock tolerates small errors and it's
 110          *   very important for it to be as fast as the platform
 111          *   can achive it. )
 112          */
 113         if (unlikely(!tsc_enabled && !tsc_unstable))
 114                 /* No locking but a rare wrong value is not a big deal: */
 115                 return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
 116
 117         /* read the Time Stamp Counter: */
 118         rdtscll(this_offset);
 119
 120         /* return the value in ns */
 121         return cycles_2_ns(this_offset);
 122 }
 123
 124 /* We need to define a real function for sched_clock, to override the
 125    weak default version */
 126 #ifdef CONFIG_PARAVIRT
 127 unsigned long long sched_clock(void)
 128 {
 129         return paravirt_sched_clock();
 130 }
 131 #else
 132 unsigned long long sched_clock(void)
 133         __attribute__((alias("native_sched_clock")));
 134 #endif
 135
 136 unsigned long native_calculate_cpu_khz(void)
 137 {
 138         unsigned long long start, end;
 139         unsigned long count;
 140         u64 delta64;
 141         int i;
 142         unsigned long flags;
 143
 144         local_irq_save(flags);
 145
 146         /* run 3 times to ensure the cache is warm */
 147         for (i = 0; i < 3; i++) {
 148                 mach_prepare_counter();
 149                 rdtscll(start);
 150                 mach_countup(&count);
 151                 rdtscll(end);
 152         }
 153         /*
 154          * Error: ECTCNEVERSET
 155          * The CTC wasn't reliable: we got a hit on the very first read,
 156          * or the CPU was so fast/slow that the quotient wouldn't fit in
 157          * 32 bits..
 158          */
 159         if (count <= 1)
 160                 goto err;
 161
 162         delta64 = end - start;
 163
 164         /* cpu freq too fast: */
 165         if (delta64 > (1ULL<<32))
 166                 goto err;
 167
 168         /* cpu freq too slow: */
 169         if (delta64 <= CALIBRATE_TIME_MSEC)
 170                 goto err;
 171
 172         delta64 += CALIBRATE_TIME_MSEC/2; /* round for do_div */
 173         do_div(delta64,CALIBRATE_TIME_MSEC);
 174
 175         local_irq_restore(flags);
 176         return (unsigned long)delta64;
 177 err:
 178         local_irq_restore(flags);
 179         return 0;
 180 }
 181
 182 int recalibrate_cpu_khz(void)
 183 {
 184 #ifndef CONFIG_SMP
 185         unsigned long cpu_khz_old = cpu_khz;
 186
 187         if (cpu_has_tsc) {
 188                 cpu_khz = calculate_cpu_khz();
 189                 tsc_khz = cpu_khz;
 190                 cpu_data[0].loops_per_jiffy =
 191                         cpufreq_scale(cpu_data[0].loops_per_jiffy,
 192                                         cpu_khz_old, cpu_khz);
 193                 return 0;
 194         } else
 195                 return -ENODEV;
 196 #else
 197         return -ENODEV;
 198 #endif
 199 }
 200
 201 EXPORT_SYMBOL(recalibrate_cpu_khz);
 202
 203 #ifdef CONFIG_CPU_FREQ
 204
 205 /*
 206  * if the CPU frequency is scaled, TSC-based delays will need a different
 207  * loops_per_jiffy value to function properly.
 208  */
 209 static unsigned int ref_freq = 0;
 210 static unsigned long loops_per_jiffy_ref = 0;
 211 static unsigned long cpu_khz_ref = 0;
 212
 213 static int
 214 time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, void *data)
 215 {
 216         struct cpufreq_freqs *freq = data;
 217
 218         if (!ref_freq) {
 219                 if (!freq->old){
 220                         ref_freq = freq->new;
 221                         return 0;
 222                 }
 223                 ref_freq = freq->old;
 224                 loops_per_jiffy_ref = cpu_data[freq->cpu].loops_per_jiffy;
 225                 cpu_khz_ref = cpu_khz;
 226         }
 227
 228         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
 229             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
 230             (val == CPUFREQ_RESUMECHANGE)) {
 231                 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
 232                         cpu_data[freq->cpu].loops_per_jiffy =
 233                                 cpufreq_scale(loops_per_jiffy_ref,
 234                                                 ref_freq, freq->new);
 235
 236                 if (cpu_khz) {
 237
 238                         if (num_online_cpus() == 1)
 239                                 cpu_khz = cpufreq_scale(cpu_khz_ref,
 240                                                 ref_freq, freq->new);
 241                         if (!(freq->flags & CPUFREQ_CONST_LOOPS)) {
 242                                 tsc_khz = cpu_khz;
 243                                 set_cyc2ns_scale(cpu_khz);
 244                                 /*
 245                                  * TSC based sched_clock turns
 246                                  * to junk w/ cpufreq
 247                                  */
 248                                 mark_tsc_unstable("cpufreq changes");
 249                         }
 250                 }
 251         }
 252
 253         return 0;
 254 }
 255
 256 static struct notifier_block time_cpufreq_notifier_block = {
 257         .notifier_call  = time_cpufreq_notifier
 258 };
 259
 260 static int __init cpufreq_tsc(void)
 261 {
 262         return cpufreq_register_notifier(&time_cpufreq_notifier_block,
 263                                          CPUFREQ_TRANSITION_NOTIFIER);
 264 }
 265 core_initcall(cpufreq_tsc);
 266
 267 #endif
 268
 269 /* clock source code */
 270
 271 static unsigned long current_tsc_khz = 0;
 272
 273 static cycle_t read_tsc(void)
 274 {
 275         cycle_t ret;
 276
 277         rdtscll(ret);
 278
 279         return ret;
 280 }
 281
 282 static struct clocksource clocksource_tsc = {
 283         .name                   = "tsc",
 284         .rating                 = 300,
 285         .read                   = read_tsc,
 286         .mask                   = CLOCKSOURCE_MASK(64),
 287         .mult                   = 0, /* to be set */
 288         .shift                  = 22,
 289         .flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
 290                                   CLOCK_SOURCE_MUST_VERIFY,
 291 };
 292
 293 void mark_tsc_unstable(char *reason)
 294 {
 295         if (!tsc_unstable) {
 296                 tsc_unstable = 1;
 297                 tsc_enabled = 0;
 298                 printk("Marking TSC unstable due to: %s.\n", reason);
 299                 /* Can be called before registration */
 300                 if (clocksource_tsc.mult)
 301                         clocksource_change_rating(&clocksource_tsc, 0);
 302                 else
 303                         clocksource_tsc.rating = 0;
 304         }
 305 }
 306 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
 307
 308 static int __init dmi_mark_tsc_unstable(struct dmi_system_id *d)
 309 {
 310         printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
 311                        d->ident);
 312         tsc_unstable = 1;
 313         return 0;
 314 }
 315
 316 /* List of systems that have known TSC problems */
 317 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
 318         {
 319          .callback = dmi_mark_tsc_unstable,
 320          .ident = "IBM Thinkpad 380XD",
 321          .matches = {
 322                      DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
 323                      DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
 324                      },
 325          },
 326          {}
 327 };
 328
 329 /*
 330  * Make an educated guess if the TSC is trustworthy and synchronized
 331  * over all CPUs.
 332  */
 333 __cpuinit int unsynchronized_tsc(void)
 334 {
 335         if (!cpu_has_tsc || tsc_unstable)
 336                 return 1;
 337         /*
 338          * Intel systems are normally all synchronized.
 339          * Exceptions must mark TSC as unstable:
 340          */
 341         if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
 342                 /* assume multi socket systems are not synchronized: */
 343                 if (num_possible_cpus() > 1)
 344                         tsc_unstable = 1;
 345         }
 346         return tsc_unstable;
 347 }
 348
 349 /*
 350  * Geode_LX - the OLPC CPU has a possibly a very reliable TSC
 351  */
 352 #ifdef CONFIG_MGEODE_LX
 353 /* RTSC counts during suspend */
 354 #define RTSC_SUSP 0x100
 355
 356 static void __init check_geode_tsc_reliable(void)
 357 {
 358         unsigned long val;
 359
 360         rdmsrl(MSR_GEODE_BUSCONT_CONF0, val);
 361         if ((val & RTSC_SUSP))
 362                 clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
 363 }
 364 #else
 365 static inline void check_geode_tsc_reliable(void) { }
 366 #endif
 367
 368
 369 void __init tsc_init(void)
 370 {
 371         if (!cpu_has_tsc || tsc_disable)
 372                 goto out_no_tsc;
 373
 374         cpu_khz = calculate_cpu_khz();
 375         tsc_khz = cpu_khz;
 376
 377         if (!cpu_khz)
 378                 goto out_no_tsc;
 379
 380         printk("Detected %lu.%03lu MHz processor.\n",
 381                                 (unsigned long)cpu_khz / 1000,
 382                                 (unsigned long)cpu_khz % 1000);
 383
 384         set_cyc2ns_scale(cpu_khz);
 385         use_tsc_delay();
 386
 387         /* Check and install the TSC clocksource */
 388         dmi_check_system(bad_tsc_dmi_table);
 389
 390         unsynchronized_tsc();
 391         check_geode_tsc_reliable();
 392         current_tsc_khz = tsc_khz;
 393         clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz,
 394                                                         clocksource_tsc.shift);
 395         /* lower the rating if we already know its unstable: */
 396         if (check_tsc_unstable()) {
 397                 clocksource_tsc.rating = 0;
 398                 clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
 399         } else
 400                 tsc_enabled = 1;
 401
 402         clocksource_register(&clocksource_tsc);
 403
 404         return;
 405
 406 out_no_tsc:
 407         /*
 408          * Set the tsc_disable flag if there's no TSC support, this
 409          * makes it a fast flag for the kernel to see whether it
 410          * should be using the TSC.
 411          */
 412         tsc_disable = 1;
 413 }