err.no Git - linux-2.6/blob - kernel/time/clocksource.c

   1 /*
   2  * linux/kernel/time/clocksource.c
   3  *
   4  * This file contains the functions which manage clocksource drivers.
   5  *
   6  * Copyright (C) 2004, 2005 IBM, John Stultz (johnstul@us.ibm.com)
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License as published by
  10  * the Free Software Foundation; either version 2 of the License, or
  11  * (at your option) any later version.
  12  *
  13  * This program is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * along with this program; if not, write to the Free Software
  20  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  21  *
  22  * TODO WishList:
  23  *   o Allow clocksource drivers to be unregistered
  24  *   o get rid of clocksource_jiffies extern
  25  */
  26
  27 #include <linux/clocksource.h>
  28 #include <linux/sysdev.h>
  29 #include <linux/init.h>
  30 #include <linux/module.h>
  31 #include <linux/sched.h> /* for spin_unlock_irq() using preempt_count() m68k */
  32 #include <linux/tick.h>
  33
  34 /* XXX - Would like a better way for initializing curr_clocksource */
  35 extern struct clocksource clocksource_jiffies;
  36
  37 /*[Clocksource internal variables]---------
  38  * curr_clocksource:
  39  *      currently selected clocksource. Initialized to clocksource_jiffies.
  40  * next_clocksource:
  41  *      pending next selected clocksource.
  42  * clocksource_list:
  43  *      linked list with the registered clocksources
  44  * clocksource_lock:
  45  *      protects manipulations to curr_clocksource and next_clocksource
  46  *      and the clocksource_list
  47  * override_name:
  48  *      Name of the user-specified clocksource.
  49  */
  50 static struct clocksource *curr_clocksource = &clocksource_jiffies;
  51 static struct clocksource *next_clocksource;
  52 static struct clocksource *clocksource_override;
  53 static LIST_HEAD(clocksource_list);
  54 static DEFINE_SPINLOCK(clocksource_lock);
  55 static char override_name[32];
  56 static int finished_booting;
  57
  58 /* clocksource_done_booting - Called near the end of core bootup
  59  *
  60  * Hack to avoid lots of clocksource churn at boot time.
  61  * We use fs_initcall because we want this to start before
  62  * device_initcall but after subsys_initcall.
  63  */
  64 static int __init clocksource_done_booting(void)
  65 {
  66         finished_booting = 1;
  67         return 0;
  68 }
  69 fs_initcall(clocksource_done_booting);
  70
  71 #ifdef CONFIG_CLOCKSOURCE_WATCHDOG
  72 static LIST_HEAD(watchdog_list);
  73 static struct clocksource *watchdog;
  74 static struct timer_list watchdog_timer;
  75 static DEFINE_SPINLOCK(watchdog_lock);
  76 static cycle_t watchdog_last;
  77 static unsigned long watchdog_resumed;
  78
  79 /*
  80  * Interval: 0.5sec Threshold: 0.0625s
  81  */
  82 #define WATCHDOG_INTERVAL (HZ >> 1)
  83 #define WATCHDOG_THRESHOLD (NSEC_PER_SEC >> 4)
  84
  85 static void clocksource_ratewd(struct clocksource *cs, int64_t delta)
  86 {
  87         if (delta > -WATCHDOG_THRESHOLD && delta < WATCHDOG_THRESHOLD)
  88                 return;
  89
  90         printk(KERN_WARNING "Clocksource %s unstable (delta = %Ld ns)\n",
  91                cs->name, delta);
  92         cs->flags &= ~(CLOCK_SOURCE_VALID_FOR_HRES | CLOCK_SOURCE_WATCHDOG);
  93         clocksource_change_rating(cs, 0);
  94         list_del(&cs->wd_list);
  95 }
  96
  97 static void clocksource_watchdog(unsigned long data)
  98 {
  99         struct clocksource *cs, *tmp;
 100         cycle_t csnow, wdnow;
 101         int64_t wd_nsec, cs_nsec;
 102         int resumed;
 103
 104         spin_lock(&watchdog_lock);
 105
 106         resumed = test_and_clear_bit(0, &watchdog_resumed);
 107
 108         wdnow = watchdog->read();
 109         wd_nsec = cyc2ns(watchdog, (wdnow - watchdog_last) & watchdog->mask);
 110         watchdog_last = wdnow;
 111
 112         list_for_each_entry_safe(cs, tmp, &watchdog_list, wd_list) {
 113                 csnow = cs->read();
 114
 115                 if (unlikely(resumed)) {
 116                         cs->wd_last = csnow;
 117                         continue;
 118                 }
 119
 120                 /* Initialized ? */
 121                 if (!(cs->flags & CLOCK_SOURCE_WATCHDOG)) {
 122                         if ((cs->flags & CLOCK_SOURCE_IS_CONTINUOUS) &&
 123                             (watchdog->flags & CLOCK_SOURCE_IS_CONTINUOUS)) {
 124                                 cs->flags |= CLOCK_SOURCE_VALID_FOR_HRES;
 125                                 /*
 126                                  * We just marked the clocksource as
 127                                  * highres-capable, notify the rest of the
 128                                  * system as well so that we transition
 129                                  * into high-res mode:
 130                                  */
 131                                 tick_clock_notify();
 132                         }
 133                         cs->flags |= CLOCK_SOURCE_WATCHDOG;
 134                         cs->wd_last = csnow;
 135                 } else {
 136                         cs_nsec = cyc2ns(cs, (csnow - cs->wd_last) & cs->mask);
 137                         cs->wd_last = csnow;
 138                         /* Check the delta. Might remove from the list ! */
 139                         clocksource_ratewd(cs, cs_nsec - wd_nsec);
 140                 }
 141         }
 142
 143         if (!list_empty(&watchdog_list)) {
 144                 /* Cycle through CPUs to check if the CPUs stay synchronized to
 145                  * each other. */
 146                 int next_cpu = next_cpu(raw_smp_processor_id(), cpu_online_map);
 147                 if (next_cpu >= NR_CPUS)
 148                         next_cpu = first_cpu(cpu_online_map);
 149                 watchdog_timer.expires += WATCHDOG_INTERVAL;
 150                 add_timer_on(&watchdog_timer, next_cpu);
 151         }
 152         spin_unlock(&watchdog_lock);
 153 }
 154 static void clocksource_resume_watchdog(void)
 155 {
 156         set_bit(0, &watchdog_resumed);
 157 }
 158
 159 static void clocksource_check_watchdog(struct clocksource *cs)
 160 {
 161         struct clocksource *cse;
 162         unsigned long flags;
 163
 164         spin_lock_irqsave(&watchdog_lock, flags);
 165         if (cs->flags & CLOCK_SOURCE_MUST_VERIFY) {
 166                 int started = !list_empty(&watchdog_list);
 167
 168                 list_add(&cs->wd_list, &watchdog_list);
 169                 if (!started && watchdog) {
 170                         watchdog_last = watchdog->read();
 171                         watchdog_timer.expires = jiffies + WATCHDOG_INTERVAL;
 172                         add_timer_on(&watchdog_timer, first_cpu(cpu_online_map));
 173                 }
 174         } else {
 175                 if (cs->flags & CLOCK_SOURCE_IS_CONTINUOUS)
 176                         cs->flags |= CLOCK_SOURCE_VALID_FOR_HRES;
 177
 178                 if (!watchdog || cs->rating > watchdog->rating) {
 179                         if (watchdog)
 180                                 del_timer(&watchdog_timer);
 181                         watchdog = cs;
 182                         init_timer_deferrable(&watchdog_timer);
 183                         watchdog_timer.function = clocksource_watchdog;
 184
 185                         /* Reset watchdog cycles */
 186                         list_for_each_entry(cse, &watchdog_list, wd_list)
 187                                 cse->flags &= ~CLOCK_SOURCE_WATCHDOG;
 188                         /* Start if list is not empty */
 189                         if (!list_empty(&watchdog_list)) {
 190                                 watchdog_last = watchdog->read();
 191                                 watchdog_timer.expires =
 192                                         jiffies + WATCHDOG_INTERVAL;
 193                                 add_timer_on(&watchdog_timer,
 194                                                 first_cpu(cpu_online_map));
 195                         }
 196                 }
 197         }
 198         spin_unlock_irqrestore(&watchdog_lock, flags);
 199 }
 200 #else
 201 static void clocksource_check_watchdog(struct clocksource *cs)
 202 {
 203         if (cs->flags & CLOCK_SOURCE_IS_CONTINUOUS)
 204                 cs->flags |= CLOCK_SOURCE_VALID_FOR_HRES;
 205 }
 206
 207 static inline void clocksource_resume_watchdog(void) { }
 208 #endif
 209
 210 /**
 211  * clocksource_resume - resume the clocksource(s)
 212  */
 213 void clocksource_resume(void)
 214 {
 215         struct clocksource *cs;
 216         unsigned long flags;
 217
 218         spin_lock_irqsave(&clocksource_lock, flags);
 219
 220         list_for_each_entry(cs, &clocksource_list, list) {
 221                 if (cs->resume)
 222                         cs->resume();
 223         }
 224
 225         clocksource_resume_watchdog();
 226
 227         spin_unlock_irqrestore(&clocksource_lock, flags);
 228 }
 229
 230 /**
 231  * clocksource_get_next - Returns the selected clocksource
 232  *
 233  */
 234 struct clocksource *clocksource_get_next(void)
 235 {
 236         unsigned long flags;
 237
 238         spin_lock_irqsave(&clocksource_lock, flags);
 239         if (next_clocksource && finished_booting) {
 240                 curr_clocksource = next_clocksource;
 241                 next_clocksource = NULL;
 242         }
 243         spin_unlock_irqrestore(&clocksource_lock, flags);
 244
 245         return curr_clocksource;
 246 }
 247
 248 /**
 249  * select_clocksource - Selects the best registered clocksource.
 250  *
 251  * Private function. Must hold clocksource_lock when called.
 252  *
 253  * Select the clocksource with the best rating, or the clocksource,
 254  * which is selected by userspace override.
 255  */
 256 static struct clocksource *select_clocksource(void)
 257 {
 258         struct clocksource *next;
 259
 260         if (list_empty(&clocksource_list))
 261                 return NULL;
 262
 263         if (clocksource_override)
 264                 next = clocksource_override;
 265         else
 266                 next = list_entry(clocksource_list.next, struct clocksource,
 267                                   list);
 268
 269         if (next == curr_clocksource)
 270                 return NULL;
 271
 272         return next;
 273 }
 274
 275 /*
 276  * Enqueue the clocksource sorted by rating
 277  */
 278 static int clocksource_enqueue(struct clocksource *c)
 279 {
 280         struct list_head *tmp, *entry = &clocksource_list;
 281
 282         list_for_each(tmp, &clocksource_list) {
 283                 struct clocksource *cs;
 284
 285                 cs = list_entry(tmp, struct clocksource, list);
 286                 if (cs == c)
 287                         return -EBUSY;
 288                 /* Keep track of the place, where to insert */
 289                 if (cs->rating >= c->rating)
 290                         entry = tmp;
 291         }
 292         list_add(&c->list, entry);
 293
 294         if (strlen(c->name) == strlen(override_name) &&
 295             !strcmp(c->name, override_name))
 296                 clocksource_override = c;
 297
 298         return 0;
 299 }
 300
 301 /**
 302  * clocksource_register - Used to install new clocksources
 303  * @t:          clocksource to be registered
 304  *
 305  * Returns -EBUSY if registration fails, zero otherwise.
 306  */
 307 int clocksource_register(struct clocksource *c)
 308 {
 309         unsigned long flags;
 310         int ret;
 311
 312         spin_lock_irqsave(&clocksource_lock, flags);
 313         ret = clocksource_enqueue(c);
 314         if (!ret)
 315                 next_clocksource = select_clocksource();
 316         spin_unlock_irqrestore(&clocksource_lock, flags);
 317         if (!ret)
 318                 clocksource_check_watchdog(c);
 319         return ret;
 320 }
 321 EXPORT_SYMBOL(clocksource_register);
 322
 323 /**
 324  * clocksource_change_rating - Change the rating of a registered clocksource
 325  *
 326  */
 327 void clocksource_change_rating(struct clocksource *cs, int rating)
 328 {
 329         unsigned long flags;
 330
 331         spin_lock_irqsave(&clocksource_lock, flags);
 332         list_del(&cs->list);
 333         cs->rating = rating;
 334         clocksource_enqueue(cs);
 335         next_clocksource = select_clocksource();
 336         spin_unlock_irqrestore(&clocksource_lock, flags);
 337 }
 338
 339 /**
 340  * clocksource_unregister - remove a registered clocksource
 341  */
 342 void clocksource_unregister(struct clocksource *cs)
 343 {
 344         unsigned long flags;
 345
 346         spin_lock_irqsave(&clocksource_lock, flags);
 347         list_del(&cs->list);
 348         if (clocksource_override == cs)
 349                 clocksource_override = NULL;
 350         next_clocksource = select_clocksource();
 351         spin_unlock_irqrestore(&clocksource_lock, flags);
 352 }
 353
 354 #ifdef CONFIG_SYSFS
 355 /**
 356  * sysfs_show_current_clocksources - sysfs interface for current clocksource
 357  * @dev:        unused
 358  * @buf:        char buffer to be filled with clocksource list
 359  *
 360  * Provides sysfs interface for listing current clocksource.
 361  */
 362 static ssize_t
 363 sysfs_show_current_clocksources(struct sys_device *dev, char *buf)
 364 {
 365         ssize_t count = 0;
 366
 367         spin_lock_irq(&clocksource_lock);
 368         count = snprintf(buf, PAGE_SIZE, "%s\n", curr_clocksource->name);
 369         spin_unlock_irq(&clocksource_lock);
 370
 371         return count;
 372 }
 373
 374 /**
 375  * sysfs_override_clocksource - interface for manually overriding clocksource
 376  * @dev:        unused
 377  * @buf:        name of override clocksource
 378  * @count:      length of buffer
 379  *
 380  * Takes input from sysfs interface for manually overriding the default
 381  * clocksource selction.
 382  */
 383 static ssize_t sysfs_override_clocksource(struct sys_device *dev,
 384                                           const char *buf, size_t count)
 385 {
 386         struct clocksource *ovr = NULL;
 387         size_t ret = count;
 388         int len;
 389
 390         /* strings from sysfs write are not 0 terminated! */
 391         if (count >= sizeof(override_name))
 392                 return -EINVAL;
 393
 394         /* strip of \n: */
 395         if (buf[count-1] == '\n')
 396                 count--;
 397
 398         spin_lock_irq(&clocksource_lock);
 399
 400         if (count > 0)
 401                 memcpy(override_name, buf, count);
 402         override_name[count] = 0;
 403
 404         len = strlen(override_name);
 405         if (len) {
 406                 struct clocksource *cs;
 407
 408                 ovr = clocksource_override;
 409                 /* try to select it: */
 410                 list_for_each_entry(cs, &clocksource_list, list) {
 411                         if (strlen(cs->name) == len &&
 412                             !strcmp(cs->name, override_name))
 413                                 ovr = cs;
 414                 }
 415         }
 416
 417         /* Reselect, when the override name has changed */
 418         if (ovr != clocksource_override) {
 419                 clocksource_override = ovr;
 420                 next_clocksource = select_clocksource();
 421         }
 422
 423         spin_unlock_irq(&clocksource_lock);
 424
 425         return ret;
 426 }
 427
 428 /**
 429  * sysfs_show_available_clocksources - sysfs interface for listing clocksource
 430  * @dev:        unused
 431  * @buf:        char buffer to be filled with clocksource list
 432  *
 433  * Provides sysfs interface for listing registered clocksources
 434  */
 435 static ssize_t
 436 sysfs_show_available_clocksources(struct sys_device *dev, char *buf)
 437 {
 438         struct clocksource *src;
 439         ssize_t count = 0;
 440
 441         spin_lock_irq(&clocksource_lock);
 442         list_for_each_entry(src, &clocksource_list, list) {
 443                 count += snprintf(buf + count,
 444                                   max((ssize_t)PAGE_SIZE - count, (ssize_t)0),
 445                                   "%s ", src->name);
 446         }
 447         spin_unlock_irq(&clocksource_lock);
 448
 449         count += snprintf(buf + count,
 450                           max((ssize_t)PAGE_SIZE - count, (ssize_t)0), "\n");
 451
 452         return count;
 453 }
 454
 455 /*
 456  * Sysfs setup bits:
 457  */
 458 static SYSDEV_ATTR(current_clocksource, 0600, sysfs_show_current_clocksources,
 459                    sysfs_override_clocksource);
 460
 461 static SYSDEV_ATTR(available_clocksource, 0600,
 462                    sysfs_show_available_clocksources, NULL);
 463
 464 static struct sysdev_class clocksource_sysclass = {
 465         .name = "clocksource",
 466 };
 467
 468 static struct sys_device device_clocksource = {
 469         .id     = 0,
 470         .cls    = &clocksource_sysclass,
 471 };
 472
 473 static int __init init_clocksource_sysfs(void)
 474 {
 475         int error = sysdev_class_register(&clocksource_sysclass);
 476
 477         if (!error)
 478                 error = sysdev_register(&device_clocksource);
 479         if (!error)
 480                 error = sysdev_create_file(
 481                                 &device_clocksource,
 482                                 &attr_current_clocksource);
 483         if (!error)
 484                 error = sysdev_create_file(
 485                                 &device_clocksource,
 486                                 &attr_available_clocksource);
 487         return error;
 488 }
 489
 490 device_initcall(init_clocksource_sysfs);
 491 #endif /* CONFIG_SYSFS */
 492
 493 /**
 494  * boot_override_clocksource - boot clock override
 495  * @str:        override name
 496  *
 497  * Takes a clocksource= boot argument and uses it
 498  * as the clocksource override name.
 499  */
 500 static int __init boot_override_clocksource(char* str)
 501 {
 502         unsigned long flags;
 503         spin_lock_irqsave(&clocksource_lock, flags);
 504         if (str)
 505                 strlcpy(override_name, str, sizeof(override_name));
 506         spin_unlock_irqrestore(&clocksource_lock, flags);
 507         return 1;
 508 }
 509
 510 __setup("clocksource=", boot_override_clocksource);
 511
 512 /**
 513  * boot_override_clock - Compatibility layer for deprecated boot option
 514  * @str:        override name
 515  *
 516  * DEPRECATED! Takes a clock= boot argument and uses it
 517  * as the clocksource override name
 518  */
 519 static int __init boot_override_clock(char* str)
 520 {
 521         if (!strcmp(str, "pmtmr")) {
 522                 printk("Warning: clock=pmtmr is deprecated. "
 523                         "Use clocksource=acpi_pm.\n");
 524                 return boot_override_clocksource("acpi_pm");
 525         }
 526         printk("Warning! clock= boot option is deprecated. "
 527                 "Use clocksource=xyz\n");
 528         return boot_override_clocksource(str);
 529 }
 530
 531 __setup("clock=", boot_override_clock);