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/clocksource.h>
   8 #include <linux/workqueue.h>
   9 #include <linux/cpufreq.h>
  10 #include <linux/jiffies.h>
  11 #include <linux/init.h>
  12 #include <linux/dmi.h>
  13
  14 #include <asm/delay.h>
  15 #include <asm/tsc.h>
  16 #include <asm/io.h>
  17
  18 #include "mach_timer.h"
  19
  20 /*
  21  * On some systems the TSC frequency does not
  22  * change with the cpu frequency. So we need
  23  * an extra value to store the TSC freq
  24  */
  25 unsigned int tsc_khz;
  26
  27 int tsc_disable;
  28
  29 #ifdef CONFIG_X86_TSC
  30 static int __init tsc_setup(char *str)
  31 {
  32         printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, "
  33                                 "cannot disable TSC.\n");
  34         return 1;
  35 }
  36 #else
  37 /*
  38  * disable flag for tsc. Takes effect by clearing the TSC cpu flag
  39  * in cpu/common.c
  40  */
  41 static int __init tsc_setup(char *str)
  42 {
  43         tsc_disable = 1;
  44
  45         return 1;
  46 }
  47 #endif
  48
  49 __setup("notsc", tsc_setup);
  50
  51 /*
  52  * code to mark and check if the TSC is unstable
  53  * due to cpufreq or due to unsynced TSCs
  54  */
  55 static int tsc_unstable;
  56
  57 static inline int check_tsc_unstable(void)
  58 {
  59         return tsc_unstable;
  60 }
  61
  62 void mark_tsc_unstable(void)
  63 {
  64         tsc_unstable = 1;
  65 }
  66 EXPORT_SYMBOL_GPL(mark_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 static 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 static inline unsigned long long cycles_2_ns(unsigned long long cyc)
  99 {
 100         return (cyc * cyc2ns_scale) >> CYC2NS_SCALE_FACTOR;
 101 }
 102
 103 #ifndef CONFIG_XEN
 104 /*
 105  * Scheduler clock - returns current time in nanosec units.
 106  */
 107 unsigned long long sched_clock(void)
 108 {
 109         unsigned long long this_offset;
 110
 111         /*
 112          * in the NUMA case we dont use the TSC as they are not
 113          * synchronized across all CPUs.
 114          */
 115 #ifndef CONFIG_NUMA
 116         if (!cpu_khz || check_tsc_unstable())
 117 #endif
 118                 /* no locking but a rare wrong value is not a big deal */
 119                 return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
 120
 121         /* read the Time Stamp Counter: */
 122         rdtscll(this_offset);
 123
 124         /* return the value in ns */
 125         return cycles_2_ns(this_offset);
 126 }
 127 #endif
 128
 129 static unsigned long calculate_cpu_khz(void)
 130 {
 131         unsigned long long start, end;
 132         unsigned long count;
 133         u64 delta64;
 134         int i;
 135         unsigned long flags;
 136
 137         local_irq_save(flags);
 138
 139         /* run 3 times to ensure the cache is warm */
 140         for (i = 0; i < 3; i++) {
 141                 mach_prepare_counter();
 142                 rdtscll(start);
 143                 mach_countup(&count);
 144                 rdtscll(end);
 145         }
 146         /*
 147          * Error: ECTCNEVERSET
 148          * The CTC wasn't reliable: we got a hit on the very first read,
 149          * or the CPU was so fast/slow that the quotient wouldn't fit in
 150          * 32 bits..
 151          */
 152         if (count <= 1)
 153                 goto err;
 154
 155         delta64 = end - start;
 156
 157         /* cpu freq too fast: */
 158         if (delta64 > (1ULL<<32))
 159                 goto err;
 160
 161         /* cpu freq too slow: */
 162         if (delta64 <= CALIBRATE_TIME_MSEC)
 163                 goto err;
 164
 165         delta64 += CALIBRATE_TIME_MSEC/2; /* round for do_div */
 166         do_div(delta64,CALIBRATE_TIME_MSEC);
 167
 168         local_irq_restore(flags);
 169         return (unsigned long)delta64;
 170 err:
 171         local_irq_restore(flags);
 172         return 0;
 173 }
 174
 175 int recalibrate_cpu_khz(void)
 176 {
 177 #ifndef CONFIG_SMP
 178         unsigned long cpu_khz_old = cpu_khz;
 179
 180         if (cpu_has_tsc) {
 181                 cpu_khz = calculate_cpu_khz();
 182                 tsc_khz = cpu_khz;
 183                 cpu_data[0].loops_per_jiffy =
 184                         cpufreq_scale(cpu_data[0].loops_per_jiffy,
 185                                         cpu_khz_old, cpu_khz);
 186                 return 0;
 187         } else
 188                 return -ENODEV;
 189 #else
 190         return -ENODEV;
 191 #endif
 192 }
 193
 194 EXPORT_SYMBOL(recalibrate_cpu_khz);
 195
 196 void __init tsc_init(void)
 197 {
 198         if (!cpu_has_tsc || tsc_disable)
 199                 return;
 200
 201         cpu_khz = calculate_cpu_khz();
 202         tsc_khz = cpu_khz;
 203
 204         if (!cpu_khz)
 205                 return;
 206
 207         printk("Detected %lu.%03lu MHz processor.\n",
 208                                 (unsigned long)cpu_khz / 1000,
 209                                 (unsigned long)cpu_khz % 1000);
 210
 211         set_cyc2ns_scale(cpu_khz);
 212         use_tsc_delay();
 213 }
 214
 215 #ifdef CONFIG_CPU_FREQ
 216
 217 static unsigned int cpufreq_delayed_issched = 0;
 218 static unsigned int cpufreq_init = 0;
 219 static struct work_struct cpufreq_delayed_get_work;
 220
 221 static void handle_cpufreq_delayed_get(struct work_struct *work)
 222 {
 223         unsigned int cpu;
 224
 225         for_each_online_cpu(cpu)
 226                 cpufreq_get(cpu);
 227
 228         cpufreq_delayed_issched = 0;
 229 }
 230
 231 /*
 232  * if we notice cpufreq oddness, schedule a call to cpufreq_get() as it tries
 233  * to verify the CPU frequency the timing core thinks the CPU is running
 234  * at is still correct.
 235  */
 236 static inline void cpufreq_delayed_get(void)
 237 {
 238         if (cpufreq_init && !cpufreq_delayed_issched) {
 239                 cpufreq_delayed_issched = 1;
 240                 printk(KERN_DEBUG "Checking if CPU frequency changed.\n");
 241                 schedule_work(&cpufreq_delayed_get_work);
 242         }
 243 }
 244
 245 /*
 246  * if the CPU frequency is scaled, TSC-based delays will need a different
 247  * loops_per_jiffy value to function properly.
 248  */
 249 static unsigned int ref_freq = 0;
 250 static unsigned long loops_per_jiffy_ref = 0;
 251 static unsigned long cpu_khz_ref = 0;
 252
 253 static int
 254 time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, void *data)
 255 {
 256         struct cpufreq_freqs *freq = data;
 257
 258         if (val != CPUFREQ_RESUMECHANGE && val != CPUFREQ_SUSPENDCHANGE)
 259                 write_seqlock_irq(&xtime_lock);
 260
 261         if (!ref_freq) {
 262                 if (!freq->old){
 263                         ref_freq = freq->new;
 264                         goto end;
 265                 }
 266                 ref_freq = freq->old;
 267                 loops_per_jiffy_ref = cpu_data[freq->cpu].loops_per_jiffy;
 268                 cpu_khz_ref = cpu_khz;
 269         }
 270
 271         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
 272             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
 273             (val == CPUFREQ_RESUMECHANGE)) {
 274                 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
 275                         cpu_data[freq->cpu].loops_per_jiffy =
 276                                 cpufreq_scale(loops_per_jiffy_ref,
 277                                                 ref_freq, freq->new);
 278
 279                 if (cpu_khz) {
 280
 281                         if (num_online_cpus() == 1)
 282                                 cpu_khz = cpufreq_scale(cpu_khz_ref,
 283                                                 ref_freq, freq->new);
 284                         if (!(freq->flags & CPUFREQ_CONST_LOOPS)) {
 285                                 tsc_khz = cpu_khz;
 286                                 set_cyc2ns_scale(cpu_khz);
 287                                 /*
 288                                  * TSC based sched_clock turns
 289                                  * to junk w/ cpufreq
 290                                  */
 291                                 mark_tsc_unstable();
 292                         }
 293                 }
 294         }
 295 end:
 296         if (val != CPUFREQ_RESUMECHANGE && val != CPUFREQ_SUSPENDCHANGE)
 297                 write_sequnlock_irq(&xtime_lock);
 298
 299         return 0;
 300 }
 301
 302 static struct notifier_block time_cpufreq_notifier_block = {
 303         .notifier_call  = time_cpufreq_notifier
 304 };
 305
 306 static int __init cpufreq_tsc(void)
 307 {
 308         int ret;
 309
 310         INIT_WORK(&cpufreq_delayed_get_work, handle_cpufreq_delayed_get);
 311         ret = cpufreq_register_notifier(&time_cpufreq_notifier_block,
 312                                         CPUFREQ_TRANSITION_NOTIFIER);
 313         if (!ret)
 314                 cpufreq_init = 1;
 315
 316         return ret;
 317 }
 318
 319 core_initcall(cpufreq_tsc);
 320
 321 #endif
 322
 323 /* clock source code */
 324
 325 static unsigned long current_tsc_khz = 0;
 326 static int tsc_update_callback(void);
 327
 328 static cycle_t read_tsc(void)
 329 {
 330         cycle_t ret;
 331
 332         rdtscll(ret);
 333
 334         return ret;
 335 }
 336
 337 static struct clocksource clocksource_tsc = {
 338         .name                   = "tsc",
 339         .rating                 = 300,
 340         .read                   = read_tsc,
 341         .mask                   = CLOCKSOURCE_MASK(64),
 342         .mult                   = 0, /* to be set */
 343         .shift                  = 22,
 344         .update_callback        = tsc_update_callback,
 345         .is_continuous          = 1,
 346 };
 347
 348 static int tsc_update_callback(void)
 349 {
 350         int change = 0;
 351
 352         /* check to see if we should switch to the safe clocksource: */
 353         if (clocksource_tsc.rating != 0 && check_tsc_unstable()) {
 354                 clocksource_tsc.rating = 0;
 355                 clocksource_reselect();
 356                 change = 1;
 357         }
 358
 359         /* only update if tsc_khz has changed: */
 360         if (current_tsc_khz != tsc_khz) {
 361                 current_tsc_khz = tsc_khz;
 362                 clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz,
 363                                                         clocksource_tsc.shift);
 364                 change = 1;
 365         }
 366
 367         return change;
 368 }
 369
 370 static int __init dmi_mark_tsc_unstable(struct dmi_system_id *d)
 371 {
 372         printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
 373                        d->ident);
 374         mark_tsc_unstable();
 375         return 0;
 376 }
 377
 378 /* List of systems that have known TSC problems */
 379 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
 380         {
 381          .callback = dmi_mark_tsc_unstable,
 382          .ident = "IBM Thinkpad 380XD",
 383          .matches = {
 384                      DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
 385                      DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
 386                      },
 387          },
 388          {}
 389 };
 390
 391 #define TSC_FREQ_CHECK_INTERVAL (10*MSEC_PER_SEC) /* 10sec in MS */
 392 static struct timer_list verify_tsc_freq_timer;
 393
 394 /* XXX - Probably should add locking */
 395 static void verify_tsc_freq(unsigned long unused)
 396 {
 397         static u64 last_tsc;
 398         static unsigned long last_jiffies;
 399
 400         u64 now_tsc, interval_tsc;
 401         unsigned long now_jiffies, interval_jiffies;
 402
 403
 404         if (check_tsc_unstable())
 405                 return;
 406
 407         rdtscll(now_tsc);
 408         now_jiffies = jiffies;
 409
 410         if (!last_jiffies) {
 411                 goto out;
 412         }
 413
 414         interval_jiffies = now_jiffies - last_jiffies;
 415         interval_tsc = now_tsc - last_tsc;
 416         interval_tsc *= HZ;
 417         do_div(interval_tsc, cpu_khz*1000);
 418
 419         if (interval_tsc < (interval_jiffies * 3 / 4)) {
 420                 printk("TSC appears to be running slowly. "
 421                         "Marking it as unstable\n");
 422                 mark_tsc_unstable();
 423                 return;
 424         }
 425
 426 out:
 427         last_tsc = now_tsc;
 428         last_jiffies = now_jiffies;
 429         /* set us up to go off on the next interval: */
 430         mod_timer(&verify_tsc_freq_timer,
 431                 jiffies + msecs_to_jiffies(TSC_FREQ_CHECK_INTERVAL));
 432 }
 433
 434 /*
 435  * Make an educated guess if the TSC is trustworthy and synchronized
 436  * over all CPUs.
 437  */
 438 static __init int unsynchronized_tsc(void)
 439 {
 440         /*
 441          * Intel systems are normally all synchronized.
 442          * Exceptions must mark TSC as unstable:
 443          */
 444         if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
 445                 return 0;
 446
 447         /* assume multi socket systems are not synchronized: */
 448         return num_possible_cpus() > 1;
 449 }
 450
 451 static int __init init_tsc_clocksource(void)
 452 {
 453
 454         if (cpu_has_tsc && tsc_khz && !tsc_disable) {
 455                 /* check blacklist */
 456                 dmi_check_system(bad_tsc_dmi_table);
 457
 458                 if (unsynchronized_tsc()) /* mark unstable if unsynced */
 459                         mark_tsc_unstable();
 460                 current_tsc_khz = tsc_khz;
 461                 clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz,
 462                                                         clocksource_tsc.shift);
 463                 /* lower the rating if we already know its unstable: */
 464                 if (check_tsc_unstable())
 465                         clocksource_tsc.rating = 0;
 466
 467                 init_timer(&verify_tsc_freq_timer);
 468                 verify_tsc_freq_timer.function = verify_tsc_freq;
 469                 verify_tsc_freq_timer.expires =
 470                         jiffies + msecs_to_jiffies(TSC_FREQ_CHECK_INTERVAL);
 471                 add_timer(&verify_tsc_freq_timer);
 472
 473                 return clocksource_register(&clocksource_tsc);
 474         }
 475
 476         return 0;
 477 }
 478
 479 module_init(init_tsc_clocksource);