arch/x86/kernel/tsc_32.c

   1 #include <linux/sched.h>
   2 #include <linux/clocksource.h>
   3 #include <linux/workqueue.h>
   4 #include <linux/cpufreq.h>
   5 #include <linux/jiffies.h>
   6 #include <linux/init.h>
   7 #include <linux/dmi.h>
   8 #include <linux/percpu.h>
   9
  10 #include <asm/delay.h>
  11 #include <asm/tsc.h>
  12 #include <asm/io.h>
  13 #include <asm/timer.h>
  14
  15 #include "mach_timer.h"
  16
  17 static int tsc_enabled;
  18
  19 /*
  20  * On some systems the TSC frequency does not
  21  * change with the cpu frequency. So we need
  22  * an extra value to store the TSC freq
  23  */
  24 unsigned int tsc_khz;
  25 EXPORT_SYMBOL_GPL(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 int check_tsc_unstable(void)
  58 {
  59         return tsc_unstable;
  60 }
  61 EXPORT_SYMBOL_GPL(check_tsc_unstable);
  62
  63 /* Accelerators for sched_clock()
  64  * convert from cycles(64bits) => nanoseconds (64bits)
  65  *  basic equation:
  66  *              ns = cycles / (freq / ns_per_sec)
  67  *              ns = cycles * (ns_per_sec / freq)
  68  *              ns = cycles * (10^9 / (cpu_khz * 10^3))
  69  *              ns = cycles * (10^6 / cpu_khz)
  70  *
  71  *      Then we use scaling math (suggested by george@mvista.com) to get:
  72  *              ns = cycles * (10^6 * SC / cpu_khz) / SC
  73  *              ns = cycles * cyc2ns_scale / SC
  74  *
  75  *      And since SC is a constant power of two, we can convert the div
  76  *  into a shift.
  77  *
  78  *  We can use khz divisor instead of mhz to keep a better precision, since
  79  *  cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
  80  *  (mathieu.desnoyers@polymtl.ca)
  81  *
  82  *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
  83  */
  84
  85 DEFINE_PER_CPU(unsigned long, cyc2ns);
  86
  87 static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu)
  88 {
  89         unsigned long flags, prev_scale, *scale;
  90         unsigned long long tsc_now, ns_now;
  91
  92         local_irq_save(flags);
  93         sched_clock_idle_sleep_event();
  94
  95         scale = &per_cpu(cyc2ns, cpu);
  96
  97         rdtscll(tsc_now);
  98         ns_now = __cycles_2_ns(tsc_now);
  99
 100         prev_scale = *scale;
 101         if (cpu_khz)
 102                 *scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz;
 103
 104         /*
 105          * Start smoothly with the new frequency:
 106          */
 107         sched_clock_idle_wakeup_event(0);
 108         local_irq_restore(flags);
 109 }
 110
 111 /*
 112  * Scheduler clock - returns current time in nanosec units.
 113  */
 114 unsigned long long native_sched_clock(void)
 115 {
 116         unsigned long long this_offset;
 117
 118         /*
 119          * Fall back to jiffies if there's no TSC available:
 120          * ( But note that we still use it if the TSC is marked
 121          *   unstable. We do this because unlike Time Of Day,
 122          *   the scheduler clock tolerates small errors and it's
 123          *   very important for it to be as fast as the platform
 124          *   can achive it. )
 125          */
 126         if (unlikely(!tsc_enabled && !tsc_unstable))
 127                 /* No locking but a rare wrong value is not a big deal: */
 128                 return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
 129
 130         /* read the Time Stamp Counter: */
 131         rdtscll(this_offset);
 132
 133         /* return the value in ns */
 134         return cycles_2_ns(this_offset);
 135 }
 136
 137 /* We need to define a real function for sched_clock, to override the
 138    weak default version */
 139 #ifdef CONFIG_PARAVIRT
 140 unsigned long long sched_clock(void)
 141 {
 142         return paravirt_sched_clock();
 143 }
 144 #else
 145 unsigned long long sched_clock(void)
 146         __attribute__((alias("native_sched_clock")));
 147 #endif
 148
 149 unsigned long native_calculate_cpu_khz(void)
 150 {
 151         unsigned long long start, end;
 152         unsigned long count;
 153         u64 delta64 = (u64)ULLONG_MAX;
 154         int i;
 155         unsigned long flags;
 156
 157         local_irq_save(flags);
 158
 159         /* run 3 times to ensure the cache is warm and to get an accurate reading */
 160         for (i = 0; i < 3; i++) {
 161                 mach_prepare_counter();
 162                 rdtscll(start);
 163                 mach_countup(&count);
 164                 rdtscll(end);
 165
 166                 /*
 167                  * Error: ECTCNEVERSET
 168                  * The CTC wasn't reliable: we got a hit on the very first read,
 169                  * or the CPU was so fast/slow that the quotient wouldn't fit in
 170                  * 32 bits..
 171                  */
 172                 if (count <= 1)
 173                         continue;
 174
 175                 /* cpu freq too slow: */
 176                 if ((end - start) <= CALIBRATE_TIME_MSEC)
 177                         continue;
 178
 179                 /*
 180                  * We want the minimum time of all runs in case one of them
 181                  * is inaccurate due to SMI or other delay
 182                  */
 183                 delta64 = min(delta64, (end - start));
 184         }
 185
 186         /* cpu freq too fast (or every run was bad): */
 187         if (delta64 > (1ULL<<32))
 188                 goto err;
 189
 190         delta64 += CALIBRATE_TIME_MSEC/2; /* round for do_div */
 191         do_div(delta64,CALIBRATE_TIME_MSEC);
 192
 193         local_irq_restore(flags);
 194         return (unsigned long)delta64;
 195 err:
 196         local_irq_restore(flags);
 197         return 0;
 198 }
 199
 200 int recalibrate_cpu_khz(void)
 201 {
 202 #ifndef CONFIG_SMP
 203         unsigned long cpu_khz_old = cpu_khz;
 204
 205         if (cpu_has_tsc) {
 206                 cpu_khz = calculate_cpu_khz();
 207                 tsc_khz = cpu_khz;
 208                 cpu_data(0).loops_per_jiffy =
 209                         cpufreq_scale(cpu_data(0).loops_per_jiffy,
 210                                         cpu_khz_old, cpu_khz);
 211                 return 0;
 212         } else
 213                 return -ENODEV;
 214 #else
 215         return -ENODEV;
 216 #endif
 217 }
 218
 219 EXPORT_SYMBOL(recalibrate_cpu_khz);
 220
 221 #ifdef CONFIG_CPU_FREQ
 222
 223 /*
 224  * if the CPU frequency is scaled, TSC-based delays will need a different
 225  * loops_per_jiffy value to function properly.
 226  */
 227 static unsigned int ref_freq = 0;
 228 static unsigned long loops_per_jiffy_ref = 0;
 229 static unsigned long cpu_khz_ref = 0;
 230
 231 static int
 232 time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, void *data)
 233 {
 234         struct cpufreq_freqs *freq = data;
 235
 236         if (!ref_freq) {
 237                 if (!freq->old){
 238                         ref_freq = freq->new;
 239                         return 0;
 240                 }
 241                 ref_freq = freq->old;
 242                 loops_per_jiffy_ref = cpu_data(freq->cpu).loops_per_jiffy;
 243                 cpu_khz_ref = cpu_khz;
 244         }
 245
 246         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
 247             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
 248             (val == CPUFREQ_RESUMECHANGE)) {
 249                 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
 250                         cpu_data(freq->cpu).loops_per_jiffy =
 251                                 cpufreq_scale(loops_per_jiffy_ref,
 252                                                 ref_freq, freq->new);
 253
 254                 if (cpu_khz) {
 255
 256                         if (num_online_cpus() == 1)
 257                                 cpu_khz = cpufreq_scale(cpu_khz_ref,
 258                                                 ref_freq, freq->new);
 259                         if (!(freq->flags & CPUFREQ_CONST_LOOPS)) {
 260                                 tsc_khz = cpu_khz;
 261                                 preempt_disable();
 262                                 set_cyc2ns_scale(cpu_khz, smp_processor_id());
 263                                 preempt_enable();
 264                                 /*
 265                                  * TSC based sched_clock turns
 266                                  * to junk w/ cpufreq
 267                                  */
 268                                 mark_tsc_unstable("cpufreq changes");
 269                         }
 270                 }
 271         }
 272
 273         return 0;
 274 }
 275
 276 static struct notifier_block time_cpufreq_notifier_block = {
 277         .notifier_call  = time_cpufreq_notifier
 278 };
 279
 280 static int __init cpufreq_tsc(void)
 281 {
 282         return cpufreq_register_notifier(&time_cpufreq_notifier_block,
 283                                          CPUFREQ_TRANSITION_NOTIFIER);
 284 }
 285 core_initcall(cpufreq_tsc);
 286
 287 #endif
 288
 289 /* clock source code */
 290
 291 static unsigned long current_tsc_khz = 0;
 292
 293 static cycle_t read_tsc(void)
 294 {
 295         cycle_t ret;
 296
 297         rdtscll(ret);
 298
 299         return ret;
 300 }
 301
 302 static struct clocksource clocksource_tsc = {
 303         .name                   = "tsc",
 304         .rating                 = 300,
 305         .read                   = read_tsc,
 306         .mask                   = CLOCKSOURCE_MASK(64),
 307         .mult                   = 0, /* to be set */
 308         .shift                  = 22,
 309         .flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
 310                                   CLOCK_SOURCE_MUST_VERIFY,
 311 };
 312
 313 void mark_tsc_unstable(char *reason)
 314 {
 315         if (!tsc_unstable) {
 316                 tsc_unstable = 1;
 317                 tsc_enabled = 0;
 318                 printk("Marking TSC unstable due to: %s.\n", reason);
 319                 /* Can be called before registration */
 320                 if (clocksource_tsc.mult)
 321                         clocksource_change_rating(&clocksource_tsc, 0);
 322                 else
 323                         clocksource_tsc.rating = 0;
 324         }
 325 }
 326 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
 327
 328 static int __init dmi_mark_tsc_unstable(const struct dmi_system_id *d)
 329 {
 330         printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
 331                        d->ident);
 332         tsc_unstable = 1;
 333         return 0;
 334 }
 335
 336 /* List of systems that have known TSC problems */
 337 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
 338         {
 339          .callback = dmi_mark_tsc_unstable,
 340          .ident = "IBM Thinkpad 380XD",
 341          .matches = {
 342                      DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
 343                      DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
 344                      },
 345          },
 346          {}
 347 };
 348
 349 /*
 350  * Make an educated guess if the TSC is trustworthy and synchronized
 351  * over all CPUs.
 352  */
 353 __cpuinit int unsynchronized_tsc(void)
 354 {
 355         if (!cpu_has_tsc || tsc_unstable)
 356                 return 1;
 357         /*
 358          * Intel systems are normally all synchronized.
 359          * Exceptions must mark TSC as unstable:
 360          */
 361         if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
 362                 /* assume multi socket systems are not synchronized: */
 363                 if (num_possible_cpus() > 1)
 364                         tsc_unstable = 1;
 365         }
 366         return tsc_unstable;
 367 }
 368
 369 /*
 370  * Geode_LX - the OLPC CPU has a possibly a very reliable TSC
 371  */
 372 #ifdef CONFIG_MGEODE_LX
 373 /* RTSC counts during suspend */
 374 #define RTSC_SUSP 0x100
 375
 376 static void __init check_geode_tsc_reliable(void)
 377 {
 378         unsigned long res_low, res_high;
 379
 380         rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
 381         if (res_low & RTSC_SUSP)
 382                 clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
 383 }
 384 #else
 385 static inline void check_geode_tsc_reliable(void) { }
 386 #endif
 387
 388
 389 void __init tsc_init(void)
 390 {
 391         int cpu;
 392
 393         if (!cpu_has_tsc || tsc_disable)
 394                 goto out_no_tsc;
 395
 396         cpu_khz = calculate_cpu_khz();
 397         tsc_khz = cpu_khz;
 398
 399         if (!cpu_khz)
 400                 goto out_no_tsc;
 401
 402         printk("Detected %lu.%03lu MHz processor.\n",
 403                                 (unsigned long)cpu_khz / 1000,
 404                                 (unsigned long)cpu_khz % 1000);
 405
 406         /*
 407          * Secondary CPUs do not run through tsc_init(), so set up
 408          * all the scale factors for all CPUs, assuming the same
 409          * speed as the bootup CPU. (cpufreq notifiers will fix this
 410          * up if their speed diverges)
 411          */
 412         for_each_possible_cpu(cpu)
 413                 set_cyc2ns_scale(cpu_khz, cpu);
 414
 415         use_tsc_delay();
 416
 417         /* Check and install the TSC clocksource */
 418         dmi_check_system(bad_tsc_dmi_table);
 419
 420         unsynchronized_tsc();
 421         check_geode_tsc_reliable();
 422         current_tsc_khz = tsc_khz;
 423         clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz,
 424                                                         clocksource_tsc.shift);
 425         /* lower the rating if we already know its unstable: */
 426         if (check_tsc_unstable()) {
 427                 clocksource_tsc.rating = 0;
 428                 clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
 429         } else
 430                 tsc_enabled = 1;
 431
 432         clocksource_register(&clocksource_tsc);
 433
 434         return;
 435
 436 out_no_tsc:
 437         /*
 438          * Set the tsc_disable flag if there's no TSC support, this
 439          * makes it a fast flag for the kernel to see whether it
 440          * should be using the TSC.
 441          */
 442         tsc_disable = 1;
 443 }