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 #include <asm/timer.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 unsigned long long (*custom_sched_clock)(void);
  28
  29 int tsc_disable;
  30
  31 #ifdef CONFIG_X86_TSC
  32 static int __init tsc_setup(char *str)
  33 {
  34         printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, "
  35                                 "cannot disable TSC.\n");
  36         return 1;
  37 }
  38 #else
  39 /*
  40  * disable flag for tsc. Takes effect by clearing the TSC cpu flag
  41  * in cpu/common.c
  42  */
  43 static int __init tsc_setup(char *str)
  44 {
  45         tsc_disable = 1;
  46
  47         return 1;
  48 }
  49 #endif
  50
  51 __setup("notsc", tsc_setup);
  52
  53 /*
  54  * code to mark and check if the TSC is unstable
  55  * due to cpufreq or due to unsynced TSCs
  56  */
  57 static int tsc_unstable;
  58
  59 static inline int check_tsc_unstable(void)
  60 {
  61         return tsc_unstable;
  62 }
  63
  64 /* Accellerators for sched_clock()
  65  * convert from cycles(64bits) => nanoseconds (64bits)
  66  *  basic equation:
  67  *              ns = cycles / (freq / ns_per_sec)
  68  *              ns = cycles * (ns_per_sec / freq)
  69  *              ns = cycles * (10^9 / (cpu_khz * 10^3))
  70  *              ns = cycles * (10^6 / cpu_khz)
  71  *
  72  *      Then we use scaling math (suggested by george@mvista.com) to get:
  73  *              ns = cycles * (10^6 * SC / cpu_khz) / SC
  74  *              ns = cycles * cyc2ns_scale / SC
  75  *
  76  *      And since SC is a constant power of two, we can convert the div
  77  *  into a shift.
  78  *
  79  *  We can use khz divisor instead of mhz to keep a better percision, since
  80  *  cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
  81  *  (mathieu.desnoyers@polymtl.ca)
  82  *
  83  *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
  84  */
  85 static unsigned long cyc2ns_scale __read_mostly;
  86
  87 #define CYC2NS_SCALE_FACTOR 10 /* 2^10, carefully chosen */
  88
  89 static inline void set_cyc2ns_scale(unsigned long cpu_khz)
  90 {
  91         cyc2ns_scale = (1000000 << CYC2NS_SCALE_FACTOR)/cpu_khz;
  92 }
  93
  94 static inline unsigned long long cycles_2_ns(unsigned long long cyc)
  95 {
  96         return (cyc * cyc2ns_scale) >> CYC2NS_SCALE_FACTOR;
  97 }
  98
  99 /*
 100  * Scheduler clock - returns current time in nanosec units.
 101  */
 102 unsigned long long sched_clock(void)
 103 {
 104         unsigned long long this_offset;
 105
 106         /*
 107          * Fall back to jiffies if there's no TSC available:
 108          */
 109         if (unlikely(tsc_disable))
 110                 /* No locking but a rare wrong value is not a big deal: */
 111                 return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
 112
 113         /* read the Time Stamp Counter: */
 114         get_scheduled_cycles(this_offset);
 115
 116         /* return the value in ns */
 117         return cycles_2_ns(this_offset);
 118 }
 119
 120 static unsigned long calculate_cpu_khz(void)
 121 {
 122         unsigned long long start, end;
 123         unsigned long count;
 124         u64 delta64;
 125         int i;
 126         unsigned long flags;
 127
 128         local_irq_save(flags);
 129
 130         /* run 3 times to ensure the cache is warm */
 131         for (i = 0; i < 3; i++) {
 132                 mach_prepare_counter();
 133                 rdtscll(start);
 134                 mach_countup(&count);
 135                 rdtscll(end);
 136         }
 137         /*
 138          * Error: ECTCNEVERSET
 139          * The CTC wasn't reliable: we got a hit on the very first read,
 140          * or the CPU was so fast/slow that the quotient wouldn't fit in
 141          * 32 bits..
 142          */
 143         if (count <= 1)
 144                 goto err;
 145
 146         delta64 = end - start;
 147
 148         /* cpu freq too fast: */
 149         if (delta64 > (1ULL<<32))
 150                 goto err;
 151
 152         /* cpu freq too slow: */
 153         if (delta64 <= CALIBRATE_TIME_MSEC)
 154                 goto err;
 155
 156         delta64 += CALIBRATE_TIME_MSEC/2; /* round for do_div */
 157         do_div(delta64,CALIBRATE_TIME_MSEC);
 158
 159         local_irq_restore(flags);
 160         return (unsigned long)delta64;
 161 err:
 162         local_irq_restore(flags);
 163         return 0;
 164 }
 165
 166 int recalibrate_cpu_khz(void)
 167 {
 168 #ifndef CONFIG_SMP
 169         unsigned long cpu_khz_old = cpu_khz;
 170
 171         if (cpu_has_tsc) {
 172                 cpu_khz = calculate_cpu_khz();
 173                 tsc_khz = cpu_khz;
 174                 cpu_data[0].loops_per_jiffy =
 175                         cpufreq_scale(cpu_data[0].loops_per_jiffy,
 176                                         cpu_khz_old, cpu_khz);
 177                 return 0;
 178         } else
 179                 return -ENODEV;
 180 #else
 181         return -ENODEV;
 182 #endif
 183 }
 184
 185 EXPORT_SYMBOL(recalibrate_cpu_khz);
 186
 187 void __init tsc_init(void)
 188 {
 189         if (!cpu_has_tsc || tsc_disable)
 190                 goto out_no_tsc;
 191
 192         cpu_khz = calculate_cpu_khz();
 193         tsc_khz = cpu_khz;
 194
 195         if (!cpu_khz)
 196                 goto out_no_tsc;
 197
 198         printk("Detected %lu.%03lu MHz processor.\n",
 199                                 (unsigned long)cpu_khz / 1000,
 200                                 (unsigned long)cpu_khz % 1000);
 201
 202         set_cyc2ns_scale(cpu_khz);
 203         use_tsc_delay();
 204         return;
 205
 206 out_no_tsc:
 207         /*
 208          * Set the tsc_disable flag if there's no TSC support, this
 209          * makes it a fast flag for the kernel to see whether it
 210          * should be using the TSC.
 211          */
 212         tsc_disable = 1;
 213 }
 214
 215 #ifdef CONFIG_CPU_FREQ
 216
 217 /*
 218  * if the CPU frequency is scaled, TSC-based delays will need a different
 219  * loops_per_jiffy value to function properly.
 220  */
 221 static unsigned int ref_freq = 0;
 222 static unsigned long loops_per_jiffy_ref = 0;
 223 static unsigned long cpu_khz_ref = 0;
 224
 225 static int
 226 time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, void *data)
 227 {
 228         struct cpufreq_freqs *freq = data;
 229
 230         if (val != CPUFREQ_RESUMECHANGE && val != CPUFREQ_SUSPENDCHANGE)
 231                 write_seqlock_irq(&xtime_lock);
 232
 233         if (!ref_freq) {
 234                 if (!freq->old){
 235                         ref_freq = freq->new;
 236                         goto end;
 237                 }
 238                 ref_freq = freq->old;
 239                 loops_per_jiffy_ref = cpu_data[freq->cpu].loops_per_jiffy;
 240                 cpu_khz_ref = cpu_khz;
 241         }
 242
 243         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
 244             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
 245             (val == CPUFREQ_RESUMECHANGE)) {
 246                 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
 247                         cpu_data[freq->cpu].loops_per_jiffy =
 248                                 cpufreq_scale(loops_per_jiffy_ref,
 249                                                 ref_freq, freq->new);
 250
 251                 if (cpu_khz) {
 252
 253                         if (num_online_cpus() == 1)
 254                                 cpu_khz = cpufreq_scale(cpu_khz_ref,
 255                                                 ref_freq, freq->new);
 256                         if (!(freq->flags & CPUFREQ_CONST_LOOPS)) {
 257                                 tsc_khz = cpu_khz;
 258                                 set_cyc2ns_scale(cpu_khz);
 259                                 /*
 260                                  * TSC based sched_clock turns
 261                                  * to junk w/ cpufreq
 262                                  */
 263                                 mark_tsc_unstable();
 264                         }
 265                 }
 266         }
 267 end:
 268         if (val != CPUFREQ_RESUMECHANGE && val != CPUFREQ_SUSPENDCHANGE)
 269                 write_sequnlock_irq(&xtime_lock);
 270
 271         return 0;
 272 }
 273
 274 static struct notifier_block time_cpufreq_notifier_block = {
 275         .notifier_call  = time_cpufreq_notifier
 276 };
 277
 278 static int __init cpufreq_tsc(void)
 279 {
 280         return cpufreq_register_notifier(&time_cpufreq_notifier_block,
 281                                          CPUFREQ_TRANSITION_NOTIFIER);
 282 }
 283 core_initcall(cpufreq_tsc);
 284
 285 #endif
 286
 287 /* clock source code */
 288
 289 static unsigned long current_tsc_khz = 0;
 290
 291 static cycle_t read_tsc(void)
 292 {
 293         cycle_t ret;
 294
 295         rdtscll(ret);
 296
 297         return ret;
 298 }
 299
 300 static struct clocksource clocksource_tsc = {
 301         .name                   = "tsc",
 302         .rating                 = 300,
 303         .read                   = read_tsc,
 304         .mask                   = CLOCKSOURCE_MASK(64),
 305         .mult                   = 0, /* to be set */
 306         .shift                  = 22,
 307         .flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
 308                                   CLOCK_SOURCE_MUST_VERIFY,
 309 };
 310
 311 void mark_tsc_unstable(void)
 312 {
 313         if (!tsc_unstable) {
 314                 tsc_unstable = 1;
 315                 /* Can be called before registration */
 316                 if (clocksource_tsc.mult)
 317                         clocksource_change_rating(&clocksource_tsc, 0);
 318                 else
 319                         clocksource_tsc.rating = 0;
 320         }
 321 }
 322 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
 323
 324 static int __init dmi_mark_tsc_unstable(struct dmi_system_id *d)
 325 {
 326         printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
 327                        d->ident);
 328         tsc_unstable = 1;
 329         return 0;
 330 }
 331
 332 /* List of systems that have known TSC problems */
 333 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
 334         {
 335          .callback = dmi_mark_tsc_unstable,
 336          .ident = "IBM Thinkpad 380XD",
 337          .matches = {
 338                      DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
 339                      DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
 340                      },
 341          },
 342          {}
 343 };
 344
 345 /*
 346  * Make an educated guess if the TSC is trustworthy and synchronized
 347  * over all CPUs.
 348  */
 349 __cpuinit int unsynchronized_tsc(void)
 350 {
 351         if (!cpu_has_tsc || tsc_unstable)
 352                 return 1;
 353         /*
 354          * Intel systems are normally all synchronized.
 355          * Exceptions must mark TSC as unstable:
 356          */
 357         if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
 358                 /* assume multi socket systems are not synchronized: */
 359                 if (num_possible_cpus() > 1)
 360                         tsc_unstable = 1;
 361         }
 362         return tsc_unstable;
 363 }
 364
 365 /*
 366  * Geode_LX - the OLPC CPU has a possibly a very reliable TSC
 367  */
 368 #ifdef CONFIG_MGEODE_LX
 369 /* RTSC counts during suspend */
 370 #define RTSC_SUSP 0x100
 371
 372 static void __init check_geode_tsc_reliable(void)
 373 {
 374         unsigned long val;
 375
 376         rdmsrl(MSR_GEODE_BUSCONT_CONF0, val);
 377         if ((val & RTSC_SUSP))
 378                 clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
 379 }
 380 #else
 381 static inline void check_geode_tsc_reliable(void) { }
 382 #endif
 383
 384 static int __init init_tsc_clocksource(void)
 385 {
 386
 387         if (cpu_has_tsc && tsc_khz && !tsc_disable) {
 388                 /* check blacklist */
 389                 dmi_check_system(bad_tsc_dmi_table);
 390
 391                 unsynchronized_tsc();
 392                 check_geode_tsc_reliable();
 393                 current_tsc_khz = tsc_khz;
 394                 clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz,
 395                                                         clocksource_tsc.shift);
 396                 /* lower the rating if we already know its unstable: */
 397                 if (check_tsc_unstable()) {
 398                         clocksource_tsc.rating = 0;
 399                         clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
 400                 }
 401
 402                 return clocksource_register(&clocksource_tsc);
 403         }
 404
 405         return 0;
 406 }
 407
 408 module_init(init_tsc_clocksource);