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/sched.h>
   8 #include <linux/clocksource.h>
   9 #include <linux/workqueue.h>
  10 #include <linux/cpufreq.h>
  11 #include <linux/jiffies.h>
  12 #include <linux/init.h>
  13 #include <linux/dmi.h>
  14
  15 #include <asm/delay.h>
  16 #include <asm/tsc.h>
  17 #include <asm/io.h>
  18 #include <asm/timer.h>
  19
  20 #include "mach_timer.h"
  21
  22 static int tsc_enabled;
  23
  24 /*
  25  * On some systems the TSC frequency does not
  26  * change with the cpu frequency. So we need
  27  * an extra value to store the TSC freq
  28  */
  29 unsigned int tsc_khz;
  30
  31 int tsc_disable;
  32
  33 #ifdef CONFIG_X86_TSC
  34 static int __init tsc_setup(char *str)
  35 {
  36         printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, "
  37                                 "cannot disable TSC.\n");
  38         return 1;
  39 }
  40 #else
  41 /*
  42  * disable flag for tsc. Takes effect by clearing the TSC cpu flag
  43  * in cpu/common.c
  44  */
  45 static int __init tsc_setup(char *str)
  46 {
  47         tsc_disable = 1;
  48
  49         return 1;
  50 }
  51 #endif
  52
  53 __setup("notsc", tsc_setup);
  54
  55 /*
  56  * code to mark and check if the TSC is unstable
  57  * due to cpufreq or due to unsynced TSCs
  58  */
  59 static int tsc_unstable;
  60
  61 static inline int check_tsc_unstable(void)
  62 {
  63         return tsc_unstable;
  64 }
  65
  66 /* Accellerators for sched_clock()
  67  * convert from cycles(64bits) => nanoseconds (64bits)
  68  *  basic equation:
  69  *              ns = cycles / (freq / ns_per_sec)
  70  *              ns = cycles * (ns_per_sec / freq)
  71  *              ns = cycles * (10^9 / (cpu_khz * 10^3))
  72  *              ns = cycles * (10^6 / cpu_khz)
  73  *
  74  *      Then we use scaling math (suggested by george@mvista.com) to get:
  75  *              ns = cycles * (10^6 * SC / cpu_khz) / SC
  76  *              ns = cycles * cyc2ns_scale / SC
  77  *
  78  *      And since SC is a constant power of two, we can convert the div
  79  *  into a shift.
  80  *
  81  *  We can use khz divisor instead of mhz to keep a better percision, since
  82  *  cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
  83  *  (mathieu.desnoyers@polymtl.ca)
  84  *
  85  *                      -johnstul@us.ibm.com "math is hard, lets go shopping!"
  86  */
  87 unsigned long cyc2ns_scale __read_mostly;
  88
  89 #define CYC2NS_SCALE_FACTOR 10 /* 2^10, carefully chosen */
  90
  91 static inline void set_cyc2ns_scale(unsigned long cpu_khz)
  92 {
  93         cyc2ns_scale = (1000000 << CYC2NS_SCALE_FACTOR)/cpu_khz;
  94 }
  95
  96 /*
  97  * Scheduler clock - returns current time in nanosec units.
  98  */
  99 unsigned long long native_sched_clock(void)
 100 {
 101         unsigned long long this_offset;
 102
 103         /*
 104          * Fall back to jiffies if there's no TSC available:
 105          * ( But note that we still use it if the TSC is marked
 106          *   unstable. We do this because unlike Time Of Day,
 107          *   the scheduler clock tolerates small errors and it's
 108          *   very important for it to be as fast as the platform
 109          *   can achive it. )
 110          */
 111         if (unlikely(!tsc_enabled && !tsc_unstable))
 112                 /* No locking but a rare wrong value is not a big deal: */
 113                 return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
 114
 115         /* read the Time Stamp Counter: */
 116         rdtscll(this_offset);
 117
 118         /* return the value in ns */
 119         return cycles_2_ns(this_offset);
 120 }
 121
 122 /* We need to define a real function for sched_clock, to override the
 123    weak default version */
 124 #ifdef CONFIG_PARAVIRT
 125 unsigned long long sched_clock(void)
 126 {
 127         return paravirt_sched_clock();
 128 }
 129 #else
 130 unsigned long long sched_clock(void)
 131         __attribute__((alias("native_sched_clock")));
 132 #endif
 133
 134 unsigned long native_calculate_cpu_khz(void)
 135 {
 136         unsigned long long start, end;
 137         unsigned long count;
 138         u64 delta64;
 139         int i;
 140         unsigned long flags;
 141
 142         local_irq_save(flags);
 143
 144         /* run 3 times to ensure the cache is warm */
 145         for (i = 0; i < 3; i++) {
 146                 mach_prepare_counter();
 147                 rdtscll(start);
 148                 mach_countup(&count);
 149                 rdtscll(end);
 150         }
 151         /*
 152          * Error: ECTCNEVERSET
 153          * The CTC wasn't reliable: we got a hit on the very first read,
 154          * or the CPU was so fast/slow that the quotient wouldn't fit in
 155          * 32 bits..
 156          */
 157         if (count <= 1)
 158                 goto err;
 159
 160         delta64 = end - start;
 161
 162         /* cpu freq too fast: */
 163         if (delta64 > (1ULL<<32))
 164                 goto err;
 165
 166         /* cpu freq too slow: */
 167         if (delta64 <= CALIBRATE_TIME_MSEC)
 168                 goto err;
 169
 170         delta64 += CALIBRATE_TIME_MSEC/2; /* round for do_div */
 171         do_div(delta64,CALIBRATE_TIME_MSEC);
 172
 173         local_irq_restore(flags);
 174         return (unsigned long)delta64;
 175 err:
 176         local_irq_restore(flags);
 177         return 0;
 178 }
 179
 180 int recalibrate_cpu_khz(void)
 181 {
 182 #ifndef CONFIG_SMP
 183         unsigned long cpu_khz_old = cpu_khz;
 184
 185         if (cpu_has_tsc) {
 186                 cpu_khz = calculate_cpu_khz();
 187                 tsc_khz = cpu_khz;
 188                 cpu_data[0].loops_per_jiffy =
 189                         cpufreq_scale(cpu_data[0].loops_per_jiffy,
 190                                         cpu_khz_old, cpu_khz);
 191                 return 0;
 192         } else
 193                 return -ENODEV;
 194 #else
 195         return -ENODEV;
 196 #endif
 197 }
 198
 199 EXPORT_SYMBOL(recalibrate_cpu_khz);
 200
 201 #ifdef CONFIG_CPU_FREQ
 202
 203 /*
 204  * if the CPU frequency is scaled, TSC-based delays will need a different
 205  * loops_per_jiffy value to function properly.
 206  */
 207 static unsigned int ref_freq = 0;
 208 static unsigned long loops_per_jiffy_ref = 0;
 209 static unsigned long cpu_khz_ref = 0;
 210
 211 static int
 212 time_cpufreq_notifier(struct notifier_block *nb, unsigned long val, void *data)
 213 {
 214         struct cpufreq_freqs *freq = data;
 215
 216         if (!ref_freq) {
 217                 if (!freq->old){
 218                         ref_freq = freq->new;
 219                         return 0;
 220                 }
 221                 ref_freq = freq->old;
 222                 loops_per_jiffy_ref = cpu_data[freq->cpu].loops_per_jiffy;
 223                 cpu_khz_ref = cpu_khz;
 224         }
 225
 226         if ((val == CPUFREQ_PRECHANGE  && freq->old < freq->new) ||
 227             (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
 228             (val == CPUFREQ_RESUMECHANGE)) {
 229                 if (!(freq->flags & CPUFREQ_CONST_LOOPS))
 230                         cpu_data[freq->cpu].loops_per_jiffy =
 231                                 cpufreq_scale(loops_per_jiffy_ref,
 232                                                 ref_freq, freq->new);
 233
 234                 if (cpu_khz) {
 235
 236                         if (num_online_cpus() == 1)
 237                                 cpu_khz = cpufreq_scale(cpu_khz_ref,
 238                                                 ref_freq, freq->new);
 239                         if (!(freq->flags & CPUFREQ_CONST_LOOPS)) {
 240                                 tsc_khz = cpu_khz;
 241                                 set_cyc2ns_scale(cpu_khz);
 242                                 /*
 243                                  * TSC based sched_clock turns
 244                                  * to junk w/ cpufreq
 245                                  */
 246                                 mark_tsc_unstable("cpufreq changes");
 247                         }
 248                 }
 249         }
 250
 251         return 0;
 252 }
 253
 254 static struct notifier_block time_cpufreq_notifier_block = {
 255         .notifier_call  = time_cpufreq_notifier
 256 };
 257
 258 static int __init cpufreq_tsc(void)
 259 {
 260         return cpufreq_register_notifier(&time_cpufreq_notifier_block,
 261                                          CPUFREQ_TRANSITION_NOTIFIER);
 262 }
 263 core_initcall(cpufreq_tsc);
 264
 265 #endif
 266
 267 /* clock source code */
 268
 269 static unsigned long current_tsc_khz = 0;
 270
 271 static cycle_t read_tsc(void)
 272 {
 273         cycle_t ret;
 274
 275         rdtscll(ret);
 276
 277         return ret;
 278 }
 279
 280 static struct clocksource clocksource_tsc = {
 281         .name                   = "tsc",
 282         .rating                 = 300,
 283         .read                   = read_tsc,
 284         .mask                   = CLOCKSOURCE_MASK(64),
 285         .mult                   = 0, /* to be set */
 286         .shift                  = 22,
 287         .flags                  = CLOCK_SOURCE_IS_CONTINUOUS |
 288                                   CLOCK_SOURCE_MUST_VERIFY,
 289 };
 290
 291 void mark_tsc_unstable(char *reason)
 292 {
 293         sched_clock_unstable_event();
 294         if (!tsc_unstable) {
 295                 tsc_unstable = 1;
 296                 tsc_enabled = 0;
 297                 printk("Marking TSC unstable due to: %s.\n", reason);
 298                 /* Can be called before registration */
 299                 if (clocksource_tsc.mult)
 300                         clocksource_change_rating(&clocksource_tsc, 0);
 301                 else
 302                         clocksource_tsc.rating = 0;
 303         }
 304 }
 305 EXPORT_SYMBOL_GPL(mark_tsc_unstable);
 306
 307 static int __init dmi_mark_tsc_unstable(struct dmi_system_id *d)
 308 {
 309         printk(KERN_NOTICE "%s detected: marking TSC unstable.\n",
 310                        d->ident);
 311         tsc_unstable = 1;
 312         return 0;
 313 }
 314
 315 /* List of systems that have known TSC problems */
 316 static struct dmi_system_id __initdata bad_tsc_dmi_table[] = {
 317         {
 318          .callback = dmi_mark_tsc_unstable,
 319          .ident = "IBM Thinkpad 380XD",
 320          .matches = {
 321                      DMI_MATCH(DMI_BOARD_VENDOR, "IBM"),
 322                      DMI_MATCH(DMI_BOARD_NAME, "2635FA0"),
 323                      },
 324          },
 325          {}
 326 };
 327
 328 /*
 329  * Make an educated guess if the TSC is trustworthy and synchronized
 330  * over all CPUs.
 331  */
 332 __cpuinit int unsynchronized_tsc(void)
 333 {
 334         if (!cpu_has_tsc || tsc_unstable)
 335                 return 1;
 336         /*
 337          * Intel systems are normally all synchronized.
 338          * Exceptions must mark TSC as unstable:
 339          */
 340         if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
 341                 /* assume multi socket systems are not synchronized: */
 342                 if (num_possible_cpus() > 1)
 343                         tsc_unstable = 1;
 344         }
 345         return tsc_unstable;
 346 }
 347
 348 /*
 349  * Geode_LX - the OLPC CPU has a possibly a very reliable TSC
 350  */
 351 #ifdef CONFIG_MGEODE_LX
 352 /* RTSC counts during suspend */
 353 #define RTSC_SUSP 0x100
 354
 355 static void __init check_geode_tsc_reliable(void)
 356 {
 357         unsigned long val;
 358
 359         rdmsrl(MSR_GEODE_BUSCONT_CONF0, val);
 360         if ((val & RTSC_SUSP))
 361                 clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
 362 }
 363 #else
 364 static inline void check_geode_tsc_reliable(void) { }
 365 #endif
 366
 367
 368 void __init tsc_init(void)
 369 {
 370         if (!cpu_has_tsc || tsc_disable)
 371                 goto out_no_tsc;
 372
 373         cpu_khz = calculate_cpu_khz();
 374         tsc_khz = cpu_khz;
 375
 376         if (!cpu_khz)
 377                 goto out_no_tsc;
 378
 379         printk("Detected %lu.%03lu MHz processor.\n",
 380                                 (unsigned long)cpu_khz / 1000,
 381                                 (unsigned long)cpu_khz % 1000);
 382
 383         set_cyc2ns_scale(cpu_khz);
 384         use_tsc_delay();
 385
 386         /* Check and install the TSC clocksource */
 387         dmi_check_system(bad_tsc_dmi_table);
 388
 389         unsynchronized_tsc();
 390         check_geode_tsc_reliable();
 391         current_tsc_khz = tsc_khz;
 392         clocksource_tsc.mult = clocksource_khz2mult(current_tsc_khz,
 393                                                         clocksource_tsc.shift);
 394         /* lower the rating if we already know its unstable: */
 395         if (check_tsc_unstable()) {
 396                 clocksource_tsc.rating = 0;
 397                 clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
 398         } else
 399                 tsc_enabled = 1;
 400
 401         clocksource_register(&clocksource_tsc);
 402
 403         return;
 404
 405 out_no_tsc:
 406         /*
 407          * Set the tsc_disable flag if there's no TSC support, this
 408          * makes it a fast flag for the kernel to see whether it
 409          * should be using the TSC.
 410          */
 411         tsc_disable = 1;
 412 }