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