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          * Fall back to jiffies if there's no TSC available:
 116          */
 117         if (unlikely(tsc_disable))
 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
 128 static unsigned long calculate_cpu_khz(void)
 129 {
 130         unsigned long long start, end;
 131         unsigned long count;
 132         u64 delta64;
 133         int i;
 134         unsigned long flags;
 135
 136         local_irq_save(flags);
 137
 138         /* run 3 times to ensure the cache is warm */
 139         for (i = 0; i < 3; i++) {
 140                 mach_prepare_counter();
 141                 rdtscll(start);
 142                 mach_countup(&count);
 143                 rdtscll(end);
 144         }
 145         /*
 146          * Error: ECTCNEVERSET
 147          * The CTC wasn't reliable: we got a hit on the very first read,
 148          * or the CPU was so fast/slow that the quotient wouldn't fit in
 149          * 32 bits..
 150          */
 151         if (count <= 1)
 152                 goto err;
 153
 154         delta64 = end - start;
 155
 156         /* cpu freq too fast: */
 157         if (delta64 > (1ULL<<32))
 158                 goto err;
 159
 160         /* cpu freq too slow: */
 161         if (delta64 <= CALIBRATE_TIME_MSEC)
 162                 goto err;
 163
 164         delta64 += CALIBRATE_TIME_MSEC/2; /* round for do_div */
 165         do_div(delta64,CALIBRATE_TIME_MSEC);
 166
 167         local_irq_restore(flags);
 168         return (unsigned long)delta64;
 169 err:
 170         local_irq_restore(flags);
 171         return 0;
 172 }
 173
 174 int recalibrate_cpu_khz(void)
 175 {
 176 #ifndef CONFIG_SMP
 177         unsigned long cpu_khz_old = cpu_khz;
 178
 179         if (cpu_has_tsc) {
 180                 cpu_khz = calculate_cpu_khz();
 181                 tsc_khz = cpu_khz;
 182                 cpu_data[0].loops_per_jiffy =
 183                         cpufreq_scale(cpu_data[0].loops_per_jiffy,
 184                                         cpu_khz_old, cpu_khz);
 185                 return 0;
 186         } else
 187                 return -ENODEV;
 188 #else
 189         return -ENODEV;
 190 #endif
 191 }
 192
 193 EXPORT_SYMBOL(recalibrate_cpu_khz);
 194
 195 void __init tsc_init(void)
 196 {
 197         if (!cpu_has_tsc || tsc_disable)
 198                 goto out_no_tsc;
 199
 200         cpu_khz = calculate_cpu_khz();
 201         tsc_khz = cpu_khz;
 202
 203         if (!cpu_khz)
 204                 goto out_no_tsc;
 205
 206         printk("Detected %lu.%03lu MHz processor.\n",
 207                                 (unsigned long)cpu_khz / 1000,
 208                                 (unsigned long)cpu_khz % 1000);
 209
 210         set_cyc2ns_scale(cpu_khz);
 211         use_tsc_delay();
 212         return;
 213
 214 out_no_tsc:
 215         /*
 216          * Set the tsc_disable flag if there's no TSC support, this
 217          * makes it a fast flag for the kernel to see whether it
 218          * should be using the TSC.
 219          */
 220         tsc_disable = 1;
 221 }
 222
 223 #ifdef CONFIG_CPU_FREQ
 224
 225 static unsigned int cpufreq_delayed_issched = 0;
 226 static unsigned int cpufreq_init = 0;
 227 static struct work_struct cpufreq_delayed_get_work;
 228
 229 static void handle_cpufreq_delayed_get(struct work_struct *work)
 230 {
 231         unsigned int cpu;
 232
 233         for_each_online_cpu(cpu)
 234                 cpufreq_get(cpu);
 235
 236         cpufreq_delayed_issched = 0;
 237 }
 238
 239 /*
 240  * if we notice cpufreq oddness, schedule a call to cpufreq_get() as it tries
 241  * to verify the CPU frequency the timing core thinks the CPU is running
 242  * at is still correct.
 243  */
 244 static inline void cpufreq_delayed_get(void)
 245 {
 246         if (cpufreq_init && !cpufreq_delayed_issched) {
 247                 cpufreq_delayed_issched = 1;
 248                 printk(KERN_DEBUG "Checking if CPU frequency changed.\n");
 249                 schedule_work(&cpufreq_delayed_get_work);
 250         }
 251 }
 252
 253 /*
 254  * if the CPU frequency is scaled, TSC-based delays will need a different
 255  * loops_per_jiffy value to function properly.
 256  */
 257 static unsigned int ref_freq = 0;
 258 static unsigned long loops_per_jiffy_ref = 0;
 259 static unsigned long cpu_khz_ref = 0;
 260
 261 static int
 262 time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, void *data)
 263 {
 264         struct cpufreq_freqs *freq = data;
 265
 266         if (val != CPUFREQ_RESUMECHANGE && val != CPUFREQ_SUSPENDCHANGE)
 267                 write_seqlock_irq(&xtime_lock);
 268
 269         if (!ref_freq) {
 270                 if (!freq->old){
 271                         ref_freq = freq->new;
 272                         goto end;
 273                 }
 274                 ref_freq = freq->old;
 275                 loops_per_jiffy_ref = cpu_data[freq->cpu].loops_per_jiffy;
 276                 cpu_khz_ref = cpu_khz;
 277         }
 278
 279         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
 280             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
 281             (val == CPUFREQ_RESUMECHANGE)) {
 282                 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
 283                         cpu_data[freq->cpu].loops_per_jiffy =
 284                                 cpufreq_scale(loops_per_jiffy_ref,
 285                                                 ref_freq, freq->new);
 286
 287                 if (cpu_khz) {
 288
 289                         if (num_online_cpus() == 1)
 290                                 cpu_khz = cpufreq_scale(cpu_khz_ref,
 291                                                 ref_freq, freq->new);
 292                         if (!(freq->flags & CPUFREQ_CONST_LOOPS)) {
 293                                 tsc_khz = cpu_khz;
 294                                 set_cyc2ns_scale(cpu_khz);
 295                                 /*
 296                                  * TSC based sched_clock turns
 297                                  * to junk w/ cpufreq
 298                                  */
 299                                 mark_tsc_unstable();
 300                         }
 301                 }
 302         }
 303 end:
 304         if (val != CPUFREQ_RESUMECHANGE && val != CPUFREQ_SUSPENDCHANGE)
 305                 write_sequnlock_irq(&xtime_lock);
 306
 307         return 0;
 308 }
 309
 310 static struct notifier_block time_cpufreq_notifier_block = {
 311         .notifier_call  = time_cpufreq_notifier
 312 };
 313
 314 static int __init cpufreq_tsc(void)
 315 {
 316         int ret;
 317
 318         INIT_WORK(&cpufreq_delayed_get_work, handle_cpufreq_delayed_get);
 319         ret = cpufreq_register_notifier(&time_cpufreq_notifier_block,
 320                                         CPUFREQ_TRANSITION_NOTIFIER);
 321         if (!ret)
 322                 cpufreq_init = 1;
 323
 324         return ret;
 325 }
 326
 327 core_initcall(cpufreq_tsc);
 328
 329 #endif
 330
 331 /* clock source code */
 332
 333 static unsigned long current_tsc_khz = 0;
 334 static int tsc_update_callback(void);
 335
 336 static cycle_t read_tsc(void)
 337 {
 338         cycle_t ret;
 339
 340         rdtscll(ret);
 341
 342         return ret;
 343 }
 344
 345 static struct clocksource clocksource_tsc = {
 346         .name                   = "tsc",
 347         .rating                 = 300,
 348         .read                   = read_tsc,
 349         .mask                   = CLOCKSOURCE_MASK(64),
 350         .mult                   = 0, /* to be set */
 351         .shift                  = 22,
 352         .update_callback        = tsc_update_callback,
 353         .is_continuous          = 1,
 354 };
 355
 356 static int tsc_update_callback(void)
 357 {
 358         int change = 0;
 359
 360         /* check to see if we should switch to the safe clocksource: */
 361         if (clocksource_tsc.rating != 0 && check_tsc_unstable()) {
 362                 clocksource_tsc.rating = 0;
 363                 clocksource_reselect();
 364                 change = 1;
 365         }
 366
 367         /* only update if tsc_khz has changed: */
 368         if (current_tsc_khz != tsc_khz) {
 369                 current_tsc_khz = tsc_khz;
 370                 clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz,
 371                                                         clocksource_tsc.shift);
 372                 change = 1;
 373         }
 374
 375         return change;
 376 }
 377
 378 static int __init dmi_mark_tsc_unstable(struct dmi_system_id *d)
 379 {
 380         printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
 381                        d->ident);
 382         mark_tsc_unstable();
 383         return 0;
 384 }
 385
 386 /* List of systems that have known TSC problems */
 387 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
 388         {
 389          .callback = dmi_mark_tsc_unstable,
 390          .ident = "IBM Thinkpad 380XD",
 391          .matches = {
 392                      DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
 393                      DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
 394                      },
 395          },
 396          {}
 397 };
 398
 399 #define TSC_FREQ_CHECK_INTERVAL (10*MSEC_PER_SEC) /* 10sec in MS */
 400 static struct timer_list verify_tsc_freq_timer;
 401
 402 /* XXX - Probably should add locking */
 403 static void verify_tsc_freq(unsigned long unused)
 404 {
 405         static u64 last_tsc;
 406         static unsigned long last_jiffies;
 407
 408         u64 now_tsc, interval_tsc;
 409         unsigned long now_jiffies, interval_jiffies;
 410
 411
 412         if (check_tsc_unstable())
 413                 return;
 414
 415         rdtscll(now_tsc);
 416         now_jiffies = jiffies;
 417
 418         if (!last_jiffies) {
 419                 goto out;
 420         }
 421
 422         interval_jiffies = now_jiffies - last_jiffies;
 423         interval_tsc = now_tsc - last_tsc;
 424         interval_tsc *= HZ;
 425         do_div(interval_tsc, cpu_khz*1000);
 426
 427         if (interval_tsc < (interval_jiffies * 3 / 4)) {
 428                 printk("TSC appears to be running slowly. "
 429                         "Marking it as unstable\n");
 430                 mark_tsc_unstable();
 431                 return;
 432         }
 433
 434 out:
 435         last_tsc = now_tsc;
 436         last_jiffies = now_jiffies;
 437         /* set us up to go off on the next interval: */
 438         mod_timer(&verify_tsc_freq_timer,
 439                 jiffies + msecs_to_jiffies(TSC_FREQ_CHECK_INTERVAL));
 440 }
 441
 442 /*
 443  * Make an educated guess if the TSC is trustworthy and synchronized
 444  * over all CPUs.
 445  */
 446 static __init int unsynchronized_tsc(void)
 447 {
 448         /*
 449          * Intel systems are normally all synchronized.
 450          * Exceptions must mark TSC as unstable:
 451          */
 452         if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
 453                 return 0;
 454
 455         /* assume multi socket systems are not synchronized: */
 456         return num_possible_cpus() > 1;
 457 }
 458
 459 static int __init init_tsc_clocksource(void)
 460 {
 461
 462         if (cpu_has_tsc && tsc_khz && !tsc_disable) {
 463                 /* check blacklist */
 464                 dmi_check_system(bad_tsc_dmi_table);
 465
 466                 if (unsynchronized_tsc()) /* mark unstable if unsynced */
 467                         mark_tsc_unstable();
 468                 current_tsc_khz = tsc_khz;
 469                 clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz,
 470                                                         clocksource_tsc.shift);
 471                 /* lower the rating if we already know its unstable: */
 472                 if (check_tsc_unstable())
 473                         clocksource_tsc.rating = 0;
 474
 475                 init_timer(&verify_tsc_freq_timer);
 476                 verify_tsc_freq_timer.function = verify_tsc_freq;
 477                 verify_tsc_freq_timer.expires =
 478                         jiffies + msecs_to_jiffies(TSC_FREQ_CHECK_INTERVAL);
 479                 add_timer(&verify_tsc_freq_timer);
 480
 481                 return clocksource_register(&clocksource_tsc);
 482         }
 483
 484         return 0;
 485 }
 486
 487 module_init(init_tsc_clocksource);