x86, clockevents: add C1E aware idle function
[linux-2.6.git] / 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/interrupt.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 static int tick_broadcast_force;
34
35 #ifdef CONFIG_TICK_ONESHOT
36 static void tick_broadcast_clear_oneshot(int cpu);
37 #else
38 static inline void tick_broadcast_clear_oneshot(int cpu) { }
39 #endif
40
41 /*
42  * Debugging: see timer_list.c
43  */
44 struct tick_device *tick_get_broadcast_device(void)
45 {
46         return &tick_broadcast_device;
47 }
48
49 cpumask_t *tick_get_broadcast_mask(void)
50 {
51         return &tick_broadcast_mask;
52 }
53
54 /*
55  * Start the device in periodic mode
56  */
57 static void tick_broadcast_start_periodic(struct clock_event_device *bc)
58 {
59         if (bc)
60                 tick_setup_periodic(bc, 1);
61 }
62
63 /*
64  * Check, if the device can be utilized as broadcast device:
65  */
66 int tick_check_broadcast_device(struct clock_event_device *dev)
67 {
68         if ((tick_broadcast_device.evtdev &&
69              tick_broadcast_device.evtdev->rating >= dev->rating) ||
70              (dev->features & CLOCK_EVT_FEAT_C3STOP))
71                 return 0;
72
73         clockevents_exchange_device(NULL, dev);
74         tick_broadcast_device.evtdev = dev;
75         if (!cpus_empty(tick_broadcast_mask))
76                 tick_broadcast_start_periodic(dev);
77         return 1;
78 }
79
80 /*
81  * Check, if the device is the broadcast device
82  */
83 int tick_is_broadcast_device(struct clock_event_device *dev)
84 {
85         return (dev && tick_broadcast_device.evtdev == dev);
86 }
87
88 /*
89  * Check, if the device is disfunctional and a place holder, which
90  * needs to be handled by the broadcast device.
91  */
92 int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
93 {
94         unsigned long flags;
95         int ret = 0;
96
97         spin_lock_irqsave(&tick_broadcast_lock, flags);
98
99         /*
100          * Devices might be registered with both periodic and oneshot
101          * mode disabled. This signals, that the device needs to be
102          * operated from the broadcast device and is a placeholder for
103          * the cpu local device.
104          */
105         if (!tick_device_is_functional(dev)) {
106                 dev->event_handler = tick_handle_periodic;
107                 cpu_set(cpu, tick_broadcast_mask);
108                 tick_broadcast_start_periodic(tick_broadcast_device.evtdev);
109                 ret = 1;
110         } else {
111                 /*
112                  * When the new device is not affected by the stop
113                  * feature and the cpu is marked in the broadcast mask
114                  * then clear the broadcast bit.
115                  */
116                 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP)) {
117                         int cpu = smp_processor_id();
118
119                         cpu_clear(cpu, tick_broadcast_mask);
120                         tick_broadcast_clear_oneshot(cpu);
121                 }
122         }
123         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
124         return ret;
125 }
126
127 /*
128  * Broadcast the event to the cpus, which are set in the mask
129  */
130 static void tick_do_broadcast(cpumask_t mask)
131 {
132         int cpu = smp_processor_id();
133         struct tick_device *td;
134
135         /*
136          * Check, if the current cpu is in the mask
137          */
138         if (cpu_isset(cpu, mask)) {
139                 cpu_clear(cpu, mask);
140                 td = &per_cpu(tick_cpu_device, cpu);
141                 td->evtdev->event_handler(td->evtdev);
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         }
155 }
156
157 /*
158  * Periodic broadcast:
159  * - invoke the broadcast handlers
160  */
161 static void tick_do_periodic_broadcast(void)
162 {
163         cpumask_t mask;
164
165         spin_lock(&tick_broadcast_lock);
166
167         cpus_and(mask, cpu_online_map, tick_broadcast_mask);
168         tick_do_broadcast(mask);
169
170         spin_unlock(&tick_broadcast_lock);
171 }
172
173 /*
174  * Event handler for periodic broadcast ticks
175  */
176 static void tick_handle_periodic_broadcast(struct clock_event_device *dev)
177 {
178         tick_do_periodic_broadcast();
179
180         /*
181          * The device is in periodic mode. No reprogramming necessary:
182          */
183         if (dev->mode == CLOCK_EVT_MODE_PERIODIC)
184                 return;
185
186         /*
187          * Setup the next period for devices, which do not have
188          * periodic mode:
189          */
190         for (;;) {
191                 ktime_t next = ktime_add(dev->next_event, tick_period);
192
193                 if (!clockevents_program_event(dev, next, ktime_get()))
194                         return;
195                 tick_do_periodic_broadcast();
196         }
197 }
198
199 /*
200  * Powerstate information: The system enters/leaves a state, where
201  * affected devices might stop
202  */
203 static void tick_do_broadcast_on_off(void *why)
204 {
205         struct clock_event_device *bc, *dev;
206         struct tick_device *td;
207         unsigned long flags, *reason = why;
208         int cpu;
209
210         spin_lock_irqsave(&tick_broadcast_lock, flags);
211
212         cpu = smp_processor_id();
213         td = &per_cpu(tick_cpu_device, cpu);
214         dev = td->evtdev;
215         bc = tick_broadcast_device.evtdev;
216
217         /*
218          * Is the device not affected by the powerstate ?
219          */
220         if (!dev || !(dev->features & CLOCK_EVT_FEAT_C3STOP))
221                 goto out;
222
223         if (!tick_device_is_functional(dev))
224                 goto out;
225
226         switch (*reason) {
227         case CLOCK_EVT_NOTIFY_BROADCAST_ON:
228         case CLOCK_EVT_NOTIFY_BROADCAST_FORCE:
229                 if (!cpu_isset(cpu, tick_broadcast_mask)) {
230                         cpu_set(cpu, tick_broadcast_mask);
231                         if (td->mode == TICKDEV_MODE_PERIODIC)
232                                 clockevents_set_mode(dev,
233                                                      CLOCK_EVT_MODE_SHUTDOWN);
234                 }
235                 if (*reason == CLOCK_EVT_NOTIFY_BROADCAST_FORCE)
236                         tick_broadcast_force = 1;
237                 break;
238         case CLOCK_EVT_NOTIFY_BROADCAST_OFF:
239                 if (!tick_broadcast_force &&
240                     cpu_isset(cpu, tick_broadcast_mask)) {
241                         cpu_clear(cpu, tick_broadcast_mask);
242                         if (td->mode == TICKDEV_MODE_PERIODIC)
243                                 tick_setup_periodic(dev, 0);
244                 }
245                 break;
246         }
247
248         if (cpus_empty(tick_broadcast_mask))
249                 clockevents_set_mode(bc, CLOCK_EVT_MODE_SHUTDOWN);
250         else {
251                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
252                         tick_broadcast_start_periodic(bc);
253                 else
254                         tick_broadcast_setup_oneshot(bc);
255         }
256 out:
257         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
258 }
259
260 /*
261  * Powerstate information: The system enters/leaves a state, where
262  * affected devices might stop.
263  */
264 void tick_broadcast_on_off(unsigned long reason, int *oncpu)
265 {
266         if (!cpu_isset(*oncpu, cpu_online_map))
267                 printk(KERN_ERR "tick-broadcast: ignoring broadcast for "
268                        "offline CPU #%d\n", *oncpu);
269         else
270                 smp_call_function_single(*oncpu, tick_do_broadcast_on_off,
271                                          &reason, 1, 1);
272 }
273
274 /*
275  * Set the periodic handler depending on broadcast on/off
276  */
277 void tick_set_periodic_handler(struct clock_event_device *dev, int broadcast)
278 {
279         if (!broadcast)
280                 dev->event_handler = tick_handle_periodic;
281         else
282                 dev->event_handler = tick_handle_periodic_broadcast;
283 }
284
285 /*
286  * Remove a CPU from broadcasting
287  */
288 void tick_shutdown_broadcast(unsigned int *cpup)
289 {
290         struct clock_event_device *bc;
291         unsigned long flags;
292         unsigned int cpu = *cpup;
293
294         spin_lock_irqsave(&tick_broadcast_lock, flags);
295
296         bc = tick_broadcast_device.evtdev;
297         cpu_clear(cpu, tick_broadcast_mask);
298
299         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
300                 if (bc && cpus_empty(tick_broadcast_mask))
301                         clockevents_set_mode(bc, CLOCK_EVT_MODE_SHUTDOWN);
302         }
303
304         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
305 }
306
307 void tick_suspend_broadcast(void)
308 {
309         struct clock_event_device *bc;
310         unsigned long flags;
311
312         spin_lock_irqsave(&tick_broadcast_lock, flags);
313
314         bc = tick_broadcast_device.evtdev;
315         if (bc)
316                 clockevents_set_mode(bc, CLOCK_EVT_MODE_SHUTDOWN);
317
318         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
319 }
320
321 int tick_resume_broadcast(void)
322 {
323         struct clock_event_device *bc;
324         unsigned long flags;
325         int broadcast = 0;
326
327         spin_lock_irqsave(&tick_broadcast_lock, flags);
328
329         bc = tick_broadcast_device.evtdev;
330
331         if (bc) {
332                 clockevents_set_mode(bc, CLOCK_EVT_MODE_RESUME);
333
334                 switch (tick_broadcast_device.mode) {
335                 case TICKDEV_MODE_PERIODIC:
336                         if(!cpus_empty(tick_broadcast_mask))
337                                 tick_broadcast_start_periodic(bc);
338                         broadcast = cpu_isset(smp_processor_id(),
339                                               tick_broadcast_mask);
340                         break;
341                 case TICKDEV_MODE_ONESHOT:
342                         broadcast = tick_resume_broadcast_oneshot(bc);
343                         break;
344                 }
345         }
346         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
347
348         return broadcast;
349 }
350
351
352 #ifdef CONFIG_TICK_ONESHOT
353
354 static cpumask_t tick_broadcast_oneshot_mask;
355
356 /*
357  * Debugging: see timer_list.c
358  */
359 cpumask_t *tick_get_broadcast_oneshot_mask(void)
360 {
361         return &tick_broadcast_oneshot_mask;
362 }
363
364 static int tick_broadcast_set_event(ktime_t expires, int force)
365 {
366         struct clock_event_device *bc = tick_broadcast_device.evtdev;
367         ktime_t now = ktime_get();
368         int res;
369
370         for(;;) {
371                 res = clockevents_program_event(bc, expires, now);
372                 if (!res || !force)
373                         return res;
374                 now = ktime_get();
375                 expires = ktime_add(now, ktime_set(0, bc->min_delta_ns));
376         }
377 }
378
379 int tick_resume_broadcast_oneshot(struct clock_event_device *bc)
380 {
381         clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
382         return 0;
383 }
384
385 /*
386  * Handle oneshot mode broadcasting
387  */
388 static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
389 {
390         struct tick_device *td;
391         cpumask_t mask;
392         ktime_t now, next_event;
393         int cpu;
394
395         spin_lock(&tick_broadcast_lock);
396 again:
397         dev->next_event.tv64 = KTIME_MAX;
398         next_event.tv64 = KTIME_MAX;
399         mask = CPU_MASK_NONE;
400         now = ktime_get();
401         /* Find all expired events */
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 <= now.tv64)
406                         cpu_set(cpu, mask);
407                 else if (td->evtdev->next_event.tv64 < next_event.tv64)
408                         next_event.tv64 = td->evtdev->next_event.tv64;
409         }
410
411         /*
412          * Wakeup the cpus which have an expired event.
413          */
414         tick_do_broadcast(mask);
415
416         /*
417          * Two reasons for reprogram:
418          *
419          * - The global event did not expire any CPU local
420          * events. This happens in dyntick mode, as the maximum PIT
421          * delta is quite small.
422          *
423          * - There are pending events on sleeping CPUs which were not
424          * in the event mask
425          */
426         if (next_event.tv64 != KTIME_MAX) {
427                 /*
428                  * Rearm the broadcast device. If event expired,
429                  * repeat the above
430                  */
431                 if (tick_broadcast_set_event(next_event, 0))
432                         goto again;
433         }
434         spin_unlock(&tick_broadcast_lock);
435 }
436
437 /*
438  * Powerstate information: The system enters/leaves a state, where
439  * affected devices might stop
440  */
441 void tick_broadcast_oneshot_control(unsigned long reason)
442 {
443         struct clock_event_device *bc, *dev;
444         struct tick_device *td;
445         unsigned long flags;
446         int cpu;
447
448         spin_lock_irqsave(&tick_broadcast_lock, flags);
449
450         /*
451          * Periodic mode does not care about the enter/exit of power
452          * states
453          */
454         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
455                 goto out;
456
457         bc = tick_broadcast_device.evtdev;
458         cpu = smp_processor_id();
459         td = &per_cpu(tick_cpu_device, cpu);
460         dev = td->evtdev;
461
462         if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
463                 goto out;
464
465         if (reason == CLOCK_EVT_NOTIFY_BROADCAST_ENTER) {
466                 if (!cpu_isset(cpu, tick_broadcast_oneshot_mask)) {
467                         cpu_set(cpu, tick_broadcast_oneshot_mask);
468                         clockevents_set_mode(dev, CLOCK_EVT_MODE_SHUTDOWN);
469                         if (dev->next_event.tv64 < bc->next_event.tv64)
470                                 tick_broadcast_set_event(dev->next_event, 1);
471                 }
472         } else {
473                 if (cpu_isset(cpu, tick_broadcast_oneshot_mask)) {
474                         cpu_clear(cpu, tick_broadcast_oneshot_mask);
475                         clockevents_set_mode(dev, CLOCK_EVT_MODE_ONESHOT);
476                         if (dev->next_event.tv64 != KTIME_MAX)
477                                 tick_program_event(dev->next_event, 1);
478                 }
479         }
480
481 out:
482         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
483 }
484
485 /*
486  * Reset the one shot broadcast for a cpu
487  *
488  * Called with tick_broadcast_lock held
489  */
490 static void tick_broadcast_clear_oneshot(int cpu)
491 {
492         cpu_clear(cpu, tick_broadcast_oneshot_mask);
493 }
494
495 /**
496  * tick_broadcast_setup_oneshot - setup the broadcast device
497  */
498 void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
499 {
500         bc->event_handler = tick_handle_oneshot_broadcast;
501         clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
502         bc->next_event.tv64 = KTIME_MAX;
503 }
504
505 /*
506  * Select oneshot operating mode for the broadcast device
507  */
508 void tick_broadcast_switch_to_oneshot(void)
509 {
510         struct clock_event_device *bc;
511         unsigned long flags;
512
513         spin_lock_irqsave(&tick_broadcast_lock, flags);
514
515         tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
516         bc = tick_broadcast_device.evtdev;
517         if (bc)
518                 tick_broadcast_setup_oneshot(bc);
519         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
520 }
521
522
523 /*
524  * Remove a dead CPU from broadcasting
525  */
526 void tick_shutdown_broadcast_oneshot(unsigned int *cpup)
527 {
528         unsigned long flags;
529         unsigned int cpu = *cpup;
530
531         spin_lock_irqsave(&tick_broadcast_lock, flags);
532
533         /*
534          * Clear the broadcast mask flag for the dead cpu, but do not
535          * stop the broadcast device!
536          */
537         cpu_clear(cpu, tick_broadcast_oneshot_mask);
538
539         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
540 }
541
542 #endif