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 unsigned long long (*custom_sched_clock)(void);
  27
  28 int tsc_disable;
  29
  30 #ifdef CONFIG_X86_TSC
  31 static int __init tsc_setup(char *str)
  32 {
  33         printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, "
  34                                 "cannot disable TSC.\n");
  35         return 1;
  36 }
  37 #else
  38 /*
  39  * disable flag for tsc. Takes effect by clearing the TSC cpu flag
  40  * in cpu/common.c
  41  */
  42 static int __init tsc_setup(char *str)
  43 {
  44         tsc_disable = 1;
  45
  46         return 1;
  47 }
  48 #endif
  49
  50 __setup("notsc", tsc_setup);
  51
  52 /*
  53  * code to mark and check if the TSC is unstable
  54  * due to cpufreq or due to unsynced TSCs
  55  */
  56 static int tsc_unstable;
  57
  58 static inline int check_tsc_unstable(void)
  59 {
  60         return tsc_unstable;
  61 }
  62
  63 void mark_tsc_unstable(void)
  64 {
  65         tsc_unstable = 1;
  66 }
  67 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
  68
  69 /* Accellerators for sched_clock()
  70  * convert from cycles(64bits) => nanoseconds (64bits)
  71  *  basic equation:
  72  *              ns = cycles / (freq / ns_per_sec)
  73  *              ns = cycles * (ns_per_sec / freq)
  74  *              ns = cycles * (10^9 / (cpu_khz * 10^3))
  75  *              ns = cycles * (10^6 / cpu_khz)
  76  *
  77  *      Then we use scaling math (suggested by george@mvista.com) to get:
  78  *              ns = cycles * (10^6 * SC / cpu_khz) / SC
  79  *              ns = cycles * cyc2ns_scale / SC
  80  *
  81  *      And since SC is a constant power of two, we can convert the div
  82  *  into a shift.
  83  *
  84  *  We can use khz divisor instead of mhz to keep a better percision, since
  85  *  cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
  86  *  (mathieu.desnoyers@polymtl.ca)
  87  *
  88  *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
  89  */
  90 static unsigned long cyc2ns_scale __read_mostly;
  91
  92 #define CYC2NS_SCALE_FACTOR 10 /* 2^10, carefully chosen */
  93
  94 static inline void set_cyc2ns_scale(unsigned long cpu_khz)
  95 {
  96         cyc2ns_scale = (1000000 << CYC2NS_SCALE_FACTOR)/cpu_khz;
  97 }
  98
  99 static inline unsigned long long cycles_2_ns(unsigned long long cyc)
 100 {
 101         return (cyc * cyc2ns_scale) >> CYC2NS_SCALE_FACTOR;
 102 }
 103
 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         if (unlikely(custom_sched_clock))
 112                 return (*custom_sched_clock)();
 113
 114         /*
 115          * in the NUMA case we dont use the TSC as they are not
 116          * synchronized across all CPUs.
 117          */
 118 #ifndef CONFIG_NUMA
 119         if (!cpu_khz || check_tsc_unstable())
 120 #endif
 121                 /* no locking but a rare wrong value is not a big deal */
 122                 return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
 123
 124         /* read the Time Stamp Counter: */
 125         rdtscll(this_offset);
 126
 127         /* return the value in ns */
 128         return cycles_2_ns(this_offset);
 129 }
 130
 131 static unsigned long calculate_cpu_khz(void)
 132 {
 133         unsigned long long start, end;
 134         unsigned long count;
 135         u64 delta64;
 136         int i;
 137         unsigned long flags;
 138
 139         local_irq_save(flags);
 140
 141         /* run 3 times to ensure the cache is warm */
 142         for (i = 0; i < 3; i++) {
 143                 mach_prepare_counter();
 144                 rdtscll(start);
 145                 mach_countup(&count);
 146                 rdtscll(end);
 147         }
 148         /*
 149          * Error: ECTCNEVERSET
 150          * The CTC wasn't reliable: we got a hit on the very first read,
 151          * or the CPU was so fast/slow that the quotient wouldn't fit in
 152          * 32 bits..
 153          */
 154         if (count <= 1)
 155                 goto err;
 156
 157         delta64 = end - start;
 158
 159         /* cpu freq too fast: */
 160         if (delta64 > (1ULL<<32))
 161                 goto err;
 162
 163         /* cpu freq too slow: */
 164         if (delta64 <= CALIBRATE_TIME_MSEC)
 165                 goto err;
 166
 167         delta64 += CALIBRATE_TIME_MSEC/2; /* round for do_div */
 168         do_div(delta64,CALIBRATE_TIME_MSEC);
 169
 170         local_irq_restore(flags);
 171         return (unsigned long)delta64;
 172 err:
 173         local_irq_restore(flags);
 174         return 0;
 175 }
 176
 177 int recalibrate_cpu_khz(void)
 178 {
 179 #ifndef CONFIG_SMP
 180         unsigned long cpu_khz_old = cpu_khz;
 181
 182         if (cpu_has_tsc) {
 183                 cpu_khz = calculate_cpu_khz();
 184                 tsc_khz = cpu_khz;
 185                 cpu_data[0].loops_per_jiffy =
 186                         cpufreq_scale(cpu_data[0].loops_per_jiffy,
 187                                         cpu_khz_old, cpu_khz);
 188                 return 0;
 189         } else
 190                 return -ENODEV;
 191 #else
 192         return -ENODEV;
 193 #endif
 194 }
 195
 196 EXPORT_SYMBOL(recalibrate_cpu_khz);
 197
 198 void __init tsc_init(void)
 199 {
 200         if (!cpu_has_tsc || tsc_disable)
 201                 return;
 202
 203         cpu_khz = calculate_cpu_khz();
 204         tsc_khz = cpu_khz;
 205
 206         if (!cpu_khz)
 207                 return;
 208
 209         printk("Detected %lu.%03lu MHz processor.\n",
 210                                 (unsigned long)cpu_khz / 1000,
 211                                 (unsigned long)cpu_khz % 1000);
 212
 213         set_cyc2ns_scale(cpu_khz);
 214         use_tsc_delay();
 215 }
 216
 217 #ifdef CONFIG_CPU_FREQ
 218
 219 static unsigned int cpufreq_delayed_issched = 0;
 220 static unsigned int cpufreq_init = 0;
 221 static struct work_struct cpufreq_delayed_get_work;
 222
 223 static void handle_cpufreq_delayed_get(struct work_struct *work)
 224 {
 225         unsigned int cpu;
 226
 227         for_each_online_cpu(cpu)
 228                 cpufreq_get(cpu);
 229
 230         cpufreq_delayed_issched = 0;
 231 }
 232
 233 /*
 234  * if we notice cpufreq oddness, schedule a call to cpufreq_get() as it tries
 235  * to verify the CPU frequency the timing core thinks the CPU is running
 236  * at is still correct.
 237  */
 238 static inline void cpufreq_delayed_get(void)
 239 {
 240         if (cpufreq_init && !cpufreq_delayed_issched) {
 241                 cpufreq_delayed_issched = 1;
 242                 printk(KERN_DEBUG "Checking if CPU frequency changed.\n");
 243                 schedule_work(&cpufreq_delayed_get_work);
 244         }
 245 }
 246
 247 /*
 248  * if the CPU frequency is scaled, TSC-based delays will need a different
 249  * loops_per_jiffy value to function properly.
 250  */
 251 static unsigned int ref_freq = 0;
 252 static unsigned long loops_per_jiffy_ref = 0;
 253 static unsigned long cpu_khz_ref = 0;
 254
 255 static int
 256 time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, void *data)
 257 {
 258         struct cpufreq_freqs *freq = data;
 259
 260         if (val != CPUFREQ_RESUMECHANGE && val != CPUFREQ_SUSPENDCHANGE)
 261                 write_seqlock_irq(&xtime_lock);
 262
 263         if (!ref_freq) {
 264                 if (!freq->old){
 265                         ref_freq = freq->new;
 266                         goto end;
 267                 }
 268                 ref_freq = freq->old;
 269                 loops_per_jiffy_ref = cpu_data[freq->cpu].loops_per_jiffy;
 270                 cpu_khz_ref = cpu_khz;
 271         }
 272
 273         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
 274             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
 275             (val == CPUFREQ_RESUMECHANGE)) {
 276                 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
 277                         cpu_data[freq->cpu].loops_per_jiffy =
 278                                 cpufreq_scale(loops_per_jiffy_ref,
 279                                                 ref_freq, freq->new);
 280
 281                 if (cpu_khz) {
 282
 283                         if (num_online_cpus() == 1)
 284                                 cpu_khz = cpufreq_scale(cpu_khz_ref,
 285                                                 ref_freq, freq->new);
 286                         if (!(freq->flags & CPUFREQ_CONST_LOOPS)) {
 287                                 tsc_khz = cpu_khz;
 288                                 set_cyc2ns_scale(cpu_khz);
 289                                 /*
 290                                  * TSC based sched_clock turns
 291                                  * to junk w/ cpufreq
 292                                  */
 293                                 mark_tsc_unstable();
 294                         }
 295                 }
 296         }
 297 end:
 298         if (val != CPUFREQ_RESUMECHANGE && val != CPUFREQ_SUSPENDCHANGE)
 299                 write_sequnlock_irq(&xtime_lock);
 300
 301         return 0;
 302 }
 303
 304 static struct notifier_block time_cpufreq_notifier_block = {
 305         .notifier_call  = time_cpufreq_notifier
 306 };
 307
 308 static int __init cpufreq_tsc(void)
 309 {
 310         int ret;
 311
 312         INIT_WORK(&cpufreq_delayed_get_work, handle_cpufreq_delayed_get);
 313         ret = cpufreq_register_notifier(&time_cpufreq_notifier_block,
 314                                         CPUFREQ_TRANSITION_NOTIFIER);
 315         if (!ret)
 316                 cpufreq_init = 1;
 317
 318         return ret;
 319 }
 320
 321 core_initcall(cpufreq_tsc);
 322
 323 #endif
 324
 325 /* clock source code */
 326
 327 static unsigned long current_tsc_khz = 0;
 328 static int tsc_update_callback(void);
 329
 330 static cycle_t read_tsc(void)
 331 {
 332         cycle_t ret;
 333
 334         rdtscll(ret);
 335
 336         return ret;
 337 }
 338
 339 static struct clocksource clocksource_tsc = {
 340         .name                   = "tsc",
 341         .rating                 = 300,
 342         .read                   = read_tsc,
 343         .mask                   = CLOCKSOURCE_MASK(64),
 344         .mult                   = 0, /* to be set */
 345         .shift                  = 22,
 346         .update_callback        = tsc_update_callback,
 347         .is_continuous          = 1,
 348 };
 349
 350 static int tsc_update_callback(void)
 351 {
 352         int change = 0;
 353
 354         /* check to see if we should switch to the safe clocksource: */
 355         if (clocksource_tsc.rating != 0 && check_tsc_unstable()) {
 356                 clocksource_tsc.rating = 0;
 357                 clocksource_reselect();
 358                 change = 1;
 359         }
 360
 361         /* only update if tsc_khz has changed: */
 362         if (current_tsc_khz != tsc_khz) {
 363                 current_tsc_khz = tsc_khz;
 364                 clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz,
 365                                                         clocksource_tsc.shift);
 366                 change = 1;
 367         }
 368
 369         return change;
 370 }
 371
 372 static int __init dmi_mark_tsc_unstable(struct dmi_system_id *d)
 373 {
 374         printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
 375                        d->ident);
 376         mark_tsc_unstable();
 377         return 0;
 378 }
 379
 380 /* List of systems that have known TSC problems */
 381 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
 382         {
 383          .callback = dmi_mark_tsc_unstable,
 384          .ident = "IBM Thinkpad 380XD",
 385          .matches = {
 386                      DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
 387                      DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
 388                      },
 389          },
 390          {}
 391 };
 392
 393 #define TSC_FREQ_CHECK_INTERVAL (10*MSEC_PER_SEC) /* 10sec in MS */
 394 static struct timer_list verify_tsc_freq_timer;
 395
 396 /* XXX - Probably should add locking */
 397 static void verify_tsc_freq(unsigned long unused)
 398 {
 399         static u64 last_tsc;
 400         static unsigned long last_jiffies;
 401
 402         u64 now_tsc, interval_tsc;
 403         unsigned long now_jiffies, interval_jiffies;
 404
 405
 406         if (check_tsc_unstable())
 407                 return;
 408
 409         rdtscll(now_tsc);
 410         now_jiffies = jiffies;
 411
 412         if (!last_jiffies) {
 413                 goto out;
 414         }
 415
 416         interval_jiffies = now_jiffies - last_jiffies;
 417         interval_tsc = now_tsc - last_tsc;
 418         interval_tsc *= HZ;
 419         do_div(interval_tsc, cpu_khz*1000);
 420
 421         if (interval_tsc < (interval_jiffies * 3 / 4)) {
 422                 printk("TSC appears to be running slowly. "
 423                         "Marking it as unstable\n");
 424                 mark_tsc_unstable();
 425                 return;
 426         }
 427
 428 out:
 429         last_tsc = now_tsc;
 430         last_jiffies = now_jiffies;
 431         /* set us up to go off on the next interval: */
 432         mod_timer(&verify_tsc_freq_timer,
 433                 jiffies + msecs_to_jiffies(TSC_FREQ_CHECK_INTERVAL));
 434 }
 435
 436 /*
 437  * Make an educated guess if the TSC is trustworthy and synchronized
 438  * over all CPUs.
 439  */
 440 static __init int unsynchronized_tsc(void)
 441 {
 442         /*
 443          * Intel systems are normally all synchronized.
 444          * Exceptions must mark TSC as unstable:
 445          */
 446         if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
 447                 return 0;
 448
 449         /* assume multi socket systems are not synchronized: */
 450         return num_possible_cpus() > 1;
 451 }
 452
 453 static int __init init_tsc_clocksource(void)
 454 {
 455
 456         if (cpu_has_tsc && tsc_khz && !tsc_disable) {
 457                 /* check blacklist */
 458                 dmi_check_system(bad_tsc_dmi_table);
 459
 460                 if (unsynchronized_tsc()) /* mark unstable if unsynced */
 461                         mark_tsc_unstable();
 462                 current_tsc_khz = tsc_khz;
 463                 clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz,
 464                                                         clocksource_tsc.shift);
 465                 /* lower the rating if we already know its unstable: */
 466                 if (check_tsc_unstable())
 467                         clocksource_tsc.rating = 0;
 468
 469                 init_timer(&verify_tsc_freq_timer);
 470                 verify_tsc_freq_timer.function = verify_tsc_freq;
 471                 verify_tsc_freq_timer.expires =
 472                         jiffies + msecs_to_jiffies(TSC_FREQ_CHECK_INTERVAL);
 473                 add_timer(&verify_tsc_freq_timer);
 474
 475                 return clocksource_register(&clocksource_tsc);
 476         }
 477
 478         return 0;
 479 }
 480
 481 module_init(init_tsc_clocksource);