clock events: allow replacement of broadcast timer
[safe/jmp/linux-2.6] / kernel / time / tick-broadcast.c
1 /*
2  * linux/kernel/time/tick-broadcast.c
3  *
4  * This file contains functions which emulate a local clock-event
5  * device via a broadcast event source.
6  *
7  * Copyright(C) 2005-2006, Thomas Gleixner <tglx@linutronix.de>
8  * Copyright(C) 2005-2007, Red Hat, Inc., Ingo Molnar
9  * Copyright(C) 2006-2007, Timesys Corp., Thomas Gleixner
10  *
11  * This code is licenced under the GPL version 2. For details see
12  * kernel-base/COPYING.
13  */
14 #include <linux/cpu.h>
15 #include <linux/err.h>
16 #include <linux/hrtimer.h>
17 #include <linux/irq.h>
18 #include <linux/percpu.h>
19 #include <linux/profile.h>
20 #include <linux/sched.h>
21 #include <linux/tick.h>
22
23 #include "tick-internal.h"
24
25 /*
26  * Broadcast support for broken x86 hardware, where the local apic
27  * timer stops in C3 state.
28  */
29
30 struct tick_device tick_broadcast_device;
31 static cpumask_t tick_broadcast_mask;
32 static DEFINE_SPINLOCK(tick_broadcast_lock);
33
34 #ifdef CONFIG_TICK_ONESHOT
35 static void tick_broadcast_clear_oneshot(int cpu);
36 #else
37 static inline void tick_broadcast_clear_oneshot(int cpu) { }
38 #endif
39
40 /*
41  * Debugging: see timer_list.c
42  */
43 struct tick_device *tick_get_broadcast_device(void)
44 {
45         return &tick_broadcast_device;
46 }
47
48 cpumask_t *tick_get_broadcast_mask(void)
49 {
50         return &tick_broadcast_mask;
51 }
52
53 /*
54  * Start the device in periodic mode
55  */
56 static void tick_broadcast_start_periodic(struct clock_event_device *bc)
57 {
58         if (bc)
59                 tick_setup_periodic(bc, 1);
60 }
61
62 /*
63  * Check, if the device can be utilized as broadcast device:
64  */
65 int tick_check_broadcast_device(struct clock_event_device *dev)
66 {
67         if ((tick_broadcast_device.evtdev &&
68              tick_broadcast_device.evtdev->rating >= dev->rating) ||
69              (dev->features & CLOCK_EVT_FEAT_C3STOP))
70                 return 0;
71
72         clockevents_exchange_device(NULL, dev);
73         tick_broadcast_device.evtdev = dev;
74         if (!cpus_empty(tick_broadcast_mask))
75                 tick_broadcast_start_periodic(dev);
76         return 1;
77 }
78
79 /*
80  * Check, if the device is the broadcast device
81  */
82 int tick_is_broadcast_device(struct clock_event_device *dev)
83 {
84         return (dev && tick_broadcast_device.evtdev == dev);
85 }
86
87 /*
88  * Check, if the device is disfunctional and a place holder, which
89  * needs to be handled by the broadcast device.
90  */
91 int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
92 {
93         unsigned long flags;
94         int ret = 0;
95
96         spin_lock_irqsave(&tick_broadcast_lock, flags);
97
98         /*
99          * Devices might be registered with both periodic and oneshot
100          * mode disabled. This signals, that the device needs to be
101          * operated from the broadcast device and is a placeholder for
102          * the cpu local device.
103          */
104         if (!tick_device_is_functional(dev)) {
105                 dev->event_handler = tick_handle_periodic;
106                 cpu_set(cpu, tick_broadcast_mask);
107                 tick_broadcast_start_periodic(tick_broadcast_device.evtdev);
108                 ret = 1;
109         } else {
110                 /*
111                  * When the new device is not affected by the stop
112                  * feature and the cpu is marked in the broadcast mask
113                  * then clear the broadcast bit.
114                  */
115                 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP)) {
116                         int cpu = smp_processor_id();
117
118                         cpu_clear(cpu, tick_broadcast_mask);
119                         tick_broadcast_clear_oneshot(cpu);
120                 }
121         }
122         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
123         return ret;
124 }
125
126 /*
127  * Broadcast the event to the cpus, which are set in the mask
128  */
129 int tick_do_broadcast(cpumask_t mask)
130 {
131         int ret = 0, cpu = smp_processor_id();
132         struct tick_device *td;
133
134         /*
135          * Check, if the current cpu is in the mask
136          */
137         if (cpu_isset(cpu, mask)) {
138                 cpu_clear(cpu, mask);
139                 td = &per_cpu(tick_cpu_device, cpu);
140                 td->evtdev->event_handler(td->evtdev);
141                 ret = 1;
142         }
143
144         if (!cpus_empty(mask)) {
145                 /*
146                  * It might be necessary to actually check whether the devices
147                  * have different broadcast functions. For now, just use the
148                  * one of the first device. This works as long as we have this
149                  * misfeature only on x86 (lapic)
150                  */
151                 cpu = first_cpu(mask);
152                 td = &per_cpu(tick_cpu_device, cpu);
153                 td->evtdev->broadcast(mask);
154                 ret = 1;
155         }
156         return ret;
157 }
158
159 /*
160  * Periodic broadcast:
161  * - invoke the broadcast handlers
162  */
163 static void tick_do_periodic_broadcast(void)
164 {
165         cpumask_t mask;
166
167         spin_lock(&tick_broadcast_lock);
168
169         cpus_and(mask, cpu_online_map, tick_broadcast_mask);
170         tick_do_broadcast(mask);
171
172         spin_unlock(&tick_broadcast_lock);
173 }
174
175 /*
176  * Event handler for periodic broadcast ticks
177  */
178 static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
179 {
180         tick_do_periodic_broadcast();
181
182         /*
183          * The device is in periodic mode. No reprogramming necessary:
184          */
185         if (dev->mode == CLOCK_EVT_MODE_PERIODIC)
186                 return;
187
188         /*
189          * Setup the next period for devices, which do not have
190          * periodic mode:
191          */
192         for (;;) {
193                 ktime_t next = ktime_add(dev->next_event, tick_period);
194
195                 if (!clockevents_program_event(dev, next, ktime_get()))
196                         return;
197                 tick_do_periodic_broadcast();
198         }
199 }
200
201 /*
202  * Powerstate information: The system enters/leaves a state, where
203  * affected devices might stop
204  */
205 static void tick_do_broadcast_on_off(void *why)
206 {
207         struct clock_event_device *bc, *dev;
208         struct tick_device *td;
209         unsigned long flags, *reason = why;
210         int cpu;
211
212         spin_lock_irqsave(&tick_broadcast_lock, flags);
213
214         cpu = smp_processor_id();
215         td = &per_cpu(tick_cpu_device, cpu);
216         dev = td->evtdev;
217         bc = tick_broadcast_device.evtdev;
218
219         /*
220          * Is the device in broadcast mode forever or is it not
221          * affected by the powerstate ?
222          */
223         if (!dev || !tick_device_is_functional(dev) ||
224             !(dev->features & CLOCK_EVT_FEAT_C3STOP))
225                 goto out;
226
227         if (*reason == CLOCK_EVT_NOTIFY_BROADCAST_ON) {
228                 if (!cpu_isset(cpu, tick_broadcast_mask)) {
229                         cpu_set(cpu, tick_broadcast_mask);
230                         if (td->mode == TICKDEV_MODE_PERIODIC)
231                                 clockevents_set_mode(dev,
232                                                      CLOCK_EVT_MODE_SHUTDOWN);
233                 }
234         } else {
235                 if (cpu_isset(cpu, tick_broadcast_mask)) {
236                         cpu_clear(cpu, tick_broadcast_mask);
237                         if (td->mode == TICKDEV_MODE_PERIODIC)
238                                 tick_setup_periodic(dev, 0);
239                 }
240         }
241
242         if (cpus_empty(tick_broadcast_mask))
243                 clockevents_set_mode(bc, CLOCK_EVT_MODE_SHUTDOWN);
244         else {
245                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
246                         tick_broadcast_start_periodic(bc);
247                 else
248                         tick_broadcast_setup_oneshot(bc);
249         }
250 out:
251         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
252 }
253
254 /*
255  * Powerstate information: The system enters/leaves a state, where
256  * affected devices might stop.
257  */
258 void tick_broadcast_on_off(unsigned long reason, int *oncpu)
259 {
260         int cpu = get_cpu();
261
262         if (!cpu_isset(*oncpu, cpu_online_map)) {
263                 printk(KERN_ERR "tick-braodcast: ignoring broadcast for "
264                        "offline CPU #%d\n", *oncpu);
265         } else {
266
267                 if (cpu == *oncpu)
268                         tick_do_broadcast_on_off(&reason);
269                 else
270                         smp_call_function_single(*oncpu,
271                                                  tick_do_broadcast_on_off,
272                                                  &reason, 1, 1);
273         }
274         put_cpu();
275 }
276
277 /*
278  * Set the periodic handler depending on broadcast on/off
279  */
280 void tick_set_periodic_handler(struct clock_event_device *dev, int broadcast)
281 {
282         if (!broadcast)
283                 dev->event_handler = tick_handle_periodic;
284         else
285                 dev->event_handler = tick_handle_periodic_broadcast;
286 }
287
288 /*
289  * Remove a CPU from broadcasting
290  */
291 void tick_shutdown_broadcast(unsigned int *cpup)
292 {
293         struct clock_event_device *bc;
294         unsigned long flags;
295         unsigned int cpu = *cpup;
296
297         spin_lock_irqsave(&tick_broadcast_lock, flags);
298
299         bc = tick_broadcast_device.evtdev;
300         cpu_clear(cpu, tick_broadcast_mask);
301
302         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
303                 if (bc && cpus_empty(tick_broadcast_mask))
304                         clockevents_set_mode(bc, CLOCK_EVT_MODE_SHUTDOWN);
305         }
306
307         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
308 }
309
310 void tick_suspend_broadcast(void)
311 {
312         struct clock_event_device *bc;
313         unsigned long flags;
314
315         spin_lock_irqsave(&tick_broadcast_lock, flags);
316
317         bc = tick_broadcast_device.evtdev;
318         if (bc)
319                 clockevents_set_mode(bc, CLOCK_EVT_MODE_SHUTDOWN);
320
321         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
322 }
323
324 int tick_resume_broadcast(void)
325 {
326         struct clock_event_device *bc;
327         unsigned long flags;
328         int broadcast = 0;
329
330         spin_lock_irqsave(&tick_broadcast_lock, flags);
331
332         bc = tick_broadcast_device.evtdev;
333
334         if (bc) {
335                 clockevents_set_mode(bc, CLOCK_EVT_MODE_RESUME);
336
337                 switch (tick_broadcast_device.mode) {
338                 case TICKDEV_MODE_PERIODIC:
339                         if(!cpus_empty(tick_broadcast_mask))
340                                 tick_broadcast_start_periodic(bc);
341                         broadcast = cpu_isset(smp_processor_id(),
342                                               tick_broadcast_mask);
343                         break;
344                 case TICKDEV_MODE_ONESHOT:
345                         broadcast = tick_resume_broadcast_oneshot(bc);
346                         break;
347                 }
348         }
349         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
350
351         return broadcast;
352 }
353
354
355 #ifdef CONFIG_TICK_ONESHOT
356
357 static cpumask_t tick_broadcast_oneshot_mask;
358
359 /*
360  * Debugging: see timer_list.c
361  */
362 cpumask_t *tick_get_broadcast_oneshot_mask(void)
363 {
364         return &tick_broadcast_oneshot_mask;
365 }
366
367 static int tick_broadcast_set_event(ktime_t expires, int force)
368 {
369         struct clock_event_device *bc = tick_broadcast_device.evtdev;
370         ktime_t now = ktime_get();
371         int res;
372
373         for(;;) {
374                 res = clockevents_program_event(bc, expires, now);
375                 if (!res || !force)
376                         return res;
377                 now = ktime_get();
378                 expires = ktime_add(now, ktime_set(0, bc->min_delta_ns));
379         }
380 }
381
382 int tick_resume_broadcast_oneshot(struct clock_event_device *bc)
383 {
384         clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
385         return 0;
386 }
387
388 /*
389  * Reprogram the broadcast device:
390  *
391  * Called with tick_broadcast_lock held and interrupts disabled.
392  */
393 static int tick_broadcast_reprogram(void)
394 {
395         ktime_t expires = { .tv64 = KTIME_MAX };
396         struct tick_device *td;
397         int cpu;
398
399         /*
400          * Find the event which expires next:
401          */
402         for (cpu = first_cpu(tick_broadcast_oneshot_mask); cpu != NR_CPUS;
403              cpu = next_cpu(cpu, tick_broadcast_oneshot_mask)) {
404                 td = &per_cpu(tick_cpu_device, cpu);
405                 if (td->evtdev->next_event.tv64 < expires.tv64)
406                         expires = td->evtdev->next_event;
407         }
408
409         if (expires.tv64 == KTIME_MAX)
410                 return 0;
411
412         return tick_broadcast_set_event(expires, 0);
413 }
414
415 /*
416  * Handle oneshot mode broadcasting
417  */
418 static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
419 {
420         struct tick_device *td;
421         cpumask_t mask;
422         ktime_t now;
423         int cpu;
424
425         spin_lock(&tick_broadcast_lock);
426 again:
427         dev->next_event.tv64 = KTIME_MAX;
428         mask = CPU_MASK_NONE;
429         now = ktime_get();
430         /* Find all expired events */
431         for (cpu = first_cpu(tick_broadcast_oneshot_mask); cpu != NR_CPUS;
432              cpu = next_cpu(cpu, tick_broadcast_oneshot_mask)) {
433                 td = &per_cpu(tick_cpu_device, cpu);
434                 if (td->evtdev->next_event.tv64 <= now.tv64)
435                         cpu_set(cpu, mask);
436         }
437
438         /*
439          * Wakeup the cpus which have an expired event. The broadcast
440          * device is reprogrammed in the return from idle code.
441          */
442         if (!tick_do_broadcast(mask)) {
443                 /*
444                  * The global event did not expire any CPU local
445                  * events. This happens in dyntick mode, as the
446                  * maximum PIT delta is quite small.
447                  */
448                 if (tick_broadcast_reprogram())
449                         goto again;
450         }
451         spin_unlock(&tick_broadcast_lock);
452 }
453
454 /*
455  * Powerstate information: The system enters/leaves a state, where
456  * affected devices might stop
457  */
458 void tick_broadcast_oneshot_control(unsigned long reason)
459 {
460         struct clock_event_device *bc, *dev;
461         struct tick_device *td;
462         unsigned long flags;
463         int cpu;
464
465         spin_lock_irqsave(&tick_broadcast_lock, flags);
466
467         /*
468          * Periodic mode does not care about the enter/exit of power
469          * states
470          */
471         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
472                 goto out;
473
474         bc = tick_broadcast_device.evtdev;
475         cpu = smp_processor_id();
476         td = &per_cpu(tick_cpu_device, cpu);
477         dev = td->evtdev;
478
479         if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
480                 goto out;
481
482         if (reason == CLOCK_EVT_NOTIFY_BROADCAST_ENTER) {
483                 if (!cpu_isset(cpu, tick_broadcast_oneshot_mask)) {
484                         cpu_set(cpu, tick_broadcast_oneshot_mask);
485                         clockevents_set_mode(dev, CLOCK_EVT_MODE_SHUTDOWN);
486                         if (dev->next_event.tv64 < bc->next_event.tv64)
487                                 tick_broadcast_set_event(dev->next_event, 1);
488                 }
489         } else {
490                 if (cpu_isset(cpu, tick_broadcast_oneshot_mask)) {
491                         cpu_clear(cpu, tick_broadcast_oneshot_mask);
492                         clockevents_set_mode(dev, CLOCK_EVT_MODE_ONESHOT);
493                         if (dev->next_event.tv64 != KTIME_MAX)
494                                 tick_program_event(dev->next_event, 1);
495                 }
496         }
497
498 out:
499         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
500 }
501
502 /*
503  * Reset the one shot broadcast for a cpu
504  *
505  * Called with tick_broadcast_lock held
506  */
507 static void tick_broadcast_clear_oneshot(int cpu)
508 {
509         cpu_clear(cpu, tick_broadcast_oneshot_mask);
510 }
511
512 /**
513  * tick_broadcast_setup_highres - setup the broadcast device for highres
514  */
515 void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
516 {
517         bc->event_handler = tick_handle_oneshot_broadcast;
518         clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
519         bc->next_event.tv64 = KTIME_MAX;
520 }
521
522 /*
523  * Select oneshot operating mode for the broadcast device
524  */
525 void tick_broadcast_switch_to_oneshot(void)
526 {
527         struct clock_event_device *bc;
528         unsigned long flags;
529
530         spin_lock_irqsave(&tick_broadcast_lock, flags);
531
532         tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
533         bc = tick_broadcast_device.evtdev;
534         if (bc)
535                 tick_broadcast_setup_oneshot(bc);
536         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
537 }
538
539
540 /*
541  * Remove a dead CPU from broadcasting
542  */
543 void tick_shutdown_broadcast_oneshot(unsigned int *cpup)
544 {
545         unsigned long flags;
546         unsigned int cpu = *cpup;
547
548         spin_lock_irqsave(&tick_broadcast_lock, flags);
549
550         /*
551          * Clear the broadcast mask flag for the dead cpu, but do not
552          * stop the broadcast device!
553          */
554         cpu_clear(cpu, tick_broadcast_oneshot_mask);
555
556         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
557 }
558
559 #endif