Merge git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi-rc-fixes-2.6
[linux-2.6] / 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                 __mod_timer(&watchdog_timer,
145                             watchdog_timer.expires + WATCHDOG_INTERVAL);
146         }
147         spin_unlock(&watchdog_lock);
148 }
149 static void clocksource_resume_watchdog(void)
150 {
151         set_bit(0, &watchdog_resumed);
152 }
153
154 static void clocksource_check_watchdog(struct clocksource *cs)
155 {
156         struct clocksource *cse;
157         unsigned long flags;
158
159         spin_lock_irqsave(&watchdog_lock, flags);
160         if (cs->flags & CLOCK_SOURCE_MUST_VERIFY) {
161                 int started = !list_empty(&watchdog_list);
162
163                 list_add(&cs->wd_list, &watchdog_list);
164                 if (!started && watchdog) {
165                         watchdog_last = watchdog->read();
166                         watchdog_timer.expires = jiffies + WATCHDOG_INTERVAL;
167                         add_timer(&watchdog_timer);
168                 }
169         } else {
170                 if (cs->flags & CLOCK_SOURCE_IS_CONTINUOUS)
171                         cs->flags |= CLOCK_SOURCE_VALID_FOR_HRES;
172
173                 if (!watchdog || cs->rating > watchdog->rating) {
174                         if (watchdog)
175                                 del_timer(&watchdog_timer);
176                         watchdog = cs;
177                         init_timer_deferrable(&watchdog_timer);
178                         watchdog_timer.function = clocksource_watchdog;
179
180                         /* Reset watchdog cycles */
181                         list_for_each_entry(cse, &watchdog_list, wd_list)
182                                 cse->flags &= ~CLOCK_SOURCE_WATCHDOG;
183                         /* Start if list is not empty */
184                         if (!list_empty(&watchdog_list)) {
185                                 watchdog_last = watchdog->read();
186                                 watchdog_timer.expires =
187                                         jiffies + WATCHDOG_INTERVAL;
188                                 add_timer(&watchdog_timer);
189                         }
190                 }
191         }
192         spin_unlock_irqrestore(&watchdog_lock, flags);
193 }
194 #else
195 static void clocksource_check_watchdog(struct clocksource *cs)
196 {
197         if (cs->flags & CLOCK_SOURCE_IS_CONTINUOUS)
198                 cs->flags |= CLOCK_SOURCE_VALID_FOR_HRES;
199 }
200
201 static inline void clocksource_resume_watchdog(void) { }
202 #endif
203
204 /**
205  * clocksource_resume - resume the clocksource(s)
206  */
207 void clocksource_resume(void)
208 {
209         struct clocksource *cs;
210         unsigned long flags;
211
212         spin_lock_irqsave(&clocksource_lock, flags);
213
214         list_for_each_entry(cs, &clocksource_list, list) {
215                 if (cs->resume)
216                         cs->resume();
217         }
218
219         clocksource_resume_watchdog();
220
221         spin_unlock_irqrestore(&clocksource_lock, flags);
222 }
223
224 /**
225  * clocksource_get_next - Returns the selected clocksource
226  *
227  */
228 struct clocksource *clocksource_get_next(void)
229 {
230         unsigned long flags;
231
232         spin_lock_irqsave(&clocksource_lock, flags);
233         if (next_clocksource && finished_booting) {
234                 curr_clocksource = next_clocksource;
235                 next_clocksource = NULL;
236         }
237         spin_unlock_irqrestore(&clocksource_lock, flags);
238
239         return curr_clocksource;
240 }
241
242 /**
243  * select_clocksource - Selects the best registered clocksource.
244  *
245  * Private function. Must hold clocksource_lock when called.
246  *
247  * Select the clocksource with the best rating, or the clocksource,
248  * which is selected by userspace override.
249  */
250 static struct clocksource *select_clocksource(void)
251 {
252         struct clocksource *next;
253
254         if (list_empty(&clocksource_list))
255                 return NULL;
256
257         if (clocksource_override)
258                 next = clocksource_override;
259         else
260                 next = list_entry(clocksource_list.next, struct clocksource,
261                                   list);
262
263         if (next == curr_clocksource)
264                 return NULL;
265
266         return next;
267 }
268
269 /*
270  * Enqueue the clocksource sorted by rating
271  */
272 static int clocksource_enqueue(struct clocksource *c)
273 {
274         struct list_head *tmp, *entry = &clocksource_list;
275
276         list_for_each(tmp, &clocksource_list) {
277                 struct clocksource *cs;
278
279                 cs = list_entry(tmp, struct clocksource, list);
280                 if (cs == c)
281                         return -EBUSY;
282                 /* Keep track of the place, where to insert */
283                 if (cs->rating >= c->rating)
284                         entry = tmp;
285         }
286         list_add(&c->list, entry);
287
288         if (strlen(c->name) == strlen(override_name) &&
289             !strcmp(c->name, override_name))
290                 clocksource_override = c;
291
292         return 0;
293 }
294
295 /**
296  * clocksource_register - Used to install new clocksources
297  * @t:          clocksource to be registered
298  *
299  * Returns -EBUSY if registration fails, zero otherwise.
300  */
301 int clocksource_register(struct clocksource *c)
302 {
303         unsigned long flags;
304         int ret;
305
306         spin_lock_irqsave(&clocksource_lock, flags);
307         ret = clocksource_enqueue(c);
308         if (!ret)
309                 next_clocksource = select_clocksource();
310         spin_unlock_irqrestore(&clocksource_lock, flags);
311         if (!ret)
312                 clocksource_check_watchdog(c);
313         return ret;
314 }
315 EXPORT_SYMBOL(clocksource_register);
316
317 /**
318  * clocksource_change_rating - Change the rating of a registered clocksource
319  *
320  */
321 void clocksource_change_rating(struct clocksource *cs, int rating)
322 {
323         unsigned long flags;
324
325         spin_lock_irqsave(&clocksource_lock, flags);
326         list_del(&cs->list);
327         cs->rating = rating;
328         clocksource_enqueue(cs);
329         next_clocksource = select_clocksource();
330         spin_unlock_irqrestore(&clocksource_lock, flags);
331 }
332
333 /**
334  * clocksource_unregister - remove a registered clocksource
335  */
336 void clocksource_unregister(struct clocksource *cs)
337 {
338         unsigned long flags;
339
340         spin_lock_irqsave(&clocksource_lock, flags);
341         list_del(&cs->list);
342         if (clocksource_override == cs)
343                 clocksource_override = NULL;
344         next_clocksource = select_clocksource();
345         spin_unlock_irqrestore(&clocksource_lock, flags);
346 }
347
348 #ifdef CONFIG_SYSFS
349 /**
350  * sysfs_show_current_clocksources - sysfs interface for current clocksource
351  * @dev:        unused
352  * @buf:        char buffer to be filled with clocksource list
353  *
354  * Provides sysfs interface for listing current clocksource.
355  */
356 static ssize_t
357 sysfs_show_current_clocksources(struct sys_device *dev, char *buf)
358 {
359         ssize_t count = 0;
360
361         spin_lock_irq(&clocksource_lock);
362         count = snprintf(buf, PAGE_SIZE, "%s\n", curr_clocksource->name);
363         spin_unlock_irq(&clocksource_lock);
364
365         return count;
366 }
367
368 /**
369  * sysfs_override_clocksource - interface for manually overriding clocksource
370  * @dev:        unused
371  * @buf:        name of override clocksource
372  * @count:      length of buffer
373  *
374  * Takes input from sysfs interface for manually overriding the default
375  * clocksource selction.
376  */
377 static ssize_t sysfs_override_clocksource(struct sys_device *dev,
378                                           const char *buf, size_t count)
379 {
380         struct clocksource *ovr = NULL;
381         size_t ret = count;
382         int len;
383
384         /* strings from sysfs write are not 0 terminated! */
385         if (count >= sizeof(override_name))
386                 return -EINVAL;
387
388         /* strip of \n: */
389         if (buf[count-1] == '\n')
390                 count--;
391
392         spin_lock_irq(&clocksource_lock);
393
394         if (count > 0)
395                 memcpy(override_name, buf, count);
396         override_name[count] = 0;
397
398         len = strlen(override_name);
399         if (len) {
400                 struct clocksource *cs;
401
402                 ovr = clocksource_override;
403                 /* try to select it: */
404                 list_for_each_entry(cs, &clocksource_list, list) {
405                         if (strlen(cs->name) == len &&
406                             !strcmp(cs->name, override_name))
407                                 ovr = cs;
408                 }
409         }
410
411         /* Reselect, when the override name has changed */
412         if (ovr != clocksource_override) {
413                 clocksource_override = ovr;
414                 next_clocksource = select_clocksource();
415         }
416
417         spin_unlock_irq(&clocksource_lock);
418
419         return ret;
420 }
421
422 /**
423  * sysfs_show_available_clocksources - sysfs interface for listing clocksource
424  * @dev:        unused
425  * @buf:        char buffer to be filled with clocksource list
426  *
427  * Provides sysfs interface for listing registered clocksources
428  */
429 static ssize_t
430 sysfs_show_available_clocksources(struct sys_device *dev, char *buf)
431 {
432         struct clocksource *src;
433         ssize_t count = 0;
434
435         spin_lock_irq(&clocksource_lock);
436         list_for_each_entry(src, &clocksource_list, list) {
437                 count += snprintf(buf + count,
438                                   max((ssize_t)PAGE_SIZE - count, (ssize_t)0),
439                                   "%s ", src->name);
440         }
441         spin_unlock_irq(&clocksource_lock);
442
443         count += snprintf(buf + count,
444                           max((ssize_t)PAGE_SIZE - count, (ssize_t)0), "\n");
445
446         return count;
447 }
448
449 /*
450  * Sysfs setup bits:
451  */
452 static SYSDEV_ATTR(current_clocksource, 0600, sysfs_show_current_clocksources,
453                    sysfs_override_clocksource);
454
455 static SYSDEV_ATTR(available_clocksource, 0600,
456                    sysfs_show_available_clocksources, NULL);
457
458 static struct sysdev_class clocksource_sysclass = {
459         .name = "clocksource",
460 };
461
462 static struct sys_device device_clocksource = {
463         .id     = 0,
464         .cls    = &clocksource_sysclass,
465 };
466
467 static int __init init_clocksource_sysfs(void)
468 {
469         int error = sysdev_class_register(&clocksource_sysclass);
470
471         if (!error)
472                 error = sysdev_register(&device_clocksource);
473         if (!error)
474                 error = sysdev_create_file(
475                                 &device_clocksource,
476                                 &attr_current_clocksource);
477         if (!error)
478                 error = sysdev_create_file(
479                                 &device_clocksource,
480                                 &attr_available_clocksource);
481         return error;
482 }
483
484 device_initcall(init_clocksource_sysfs);
485 #endif /* CONFIG_SYSFS */
486
487 /**
488  * boot_override_clocksource - boot clock override
489  * @str:        override name
490  *
491  * Takes a clocksource= boot argument and uses it
492  * as the clocksource override name.
493  */
494 static int __init boot_override_clocksource(char* str)
495 {
496         unsigned long flags;
497         spin_lock_irqsave(&clocksource_lock, flags);
498         if (str)
499                 strlcpy(override_name, str, sizeof(override_name));
500         spin_unlock_irqrestore(&clocksource_lock, flags);
501         return 1;
502 }
503
504 __setup("clocksource=", boot_override_clocksource);
505
506 /**
507  * boot_override_clock - Compatibility layer for deprecated boot option
508  * @str:        override name
509  *
510  * DEPRECATED! Takes a clock= boot argument and uses it
511  * as the clocksource override name
512  */
513 static int __init boot_override_clock(char* str)
514 {
515         if (!strcmp(str, "pmtmr")) {
516                 printk("Warning: clock=pmtmr is deprecated. "
517                         "Use clocksource=acpi_pm.\n");
518                 return boot_override_clocksource("acpi_pm");
519         }
520         printk("Warning! clock= boot option is deprecated. "
521                 "Use clocksource=xyz\n");
522         return boot_override_clocksource(str);
523 }
524
525 __setup("clock=", boot_override_clock);