NOHZ: split tick_nohz_restart_sched_tick()
[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/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         ktime_t next;
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. We read dev->next_event first and add to it
191          * when the event alrady expired. clockevents_program_event()
192          * sets dev->next_event only when the event is really
193          * programmed to the device.
194          */
195         for (next = dev->next_event; ;) {
196                 next = ktime_add(next, tick_period);
197
198                 if (!clockevents_program_event(dev, next, ktime_get()))
199                         return;
200                 tick_do_periodic_broadcast();
201         }
202 }
203
204 /*
205  * Powerstate information: The system enters/leaves a state, where
206  * affected devices might stop
207  */
208 static void tick_do_broadcast_on_off(void *why)
209 {
210         struct clock_event_device *bc, *dev;
211         struct tick_device *td;
212         unsigned long flags, *reason = why;
213         int cpu, bc_stopped;
214
215         spin_lock_irqsave(&tick_broadcast_lock, flags);
216
217         cpu = smp_processor_id();
218         td = &per_cpu(tick_cpu_device, cpu);
219         dev = td->evtdev;
220         bc = tick_broadcast_device.evtdev;
221
222         /*
223          * Is the device not affected by the powerstate ?
224          */
225         if (!dev || !(dev->features & CLOCK_EVT_FEAT_C3STOP))
226                 goto out;
227
228         if (!tick_device_is_functional(dev))
229                 goto out;
230
231         bc_stopped = cpus_empty(tick_broadcast_mask);
232
233         switch (*reason) {
234         case CLOCK_EVT_NOTIFY_BROADCAST_ON:
235         case CLOCK_EVT_NOTIFY_BROADCAST_FORCE:
236                 if (!cpu_isset(cpu, tick_broadcast_mask)) {
237                         cpu_set(cpu, tick_broadcast_mask);
238                         if (tick_broadcast_device.mode ==
239                             TICKDEV_MODE_PERIODIC)
240                                 clockevents_shutdown(dev);
241                 }
242                 if (*reason == CLOCK_EVT_NOTIFY_BROADCAST_FORCE)
243                         tick_broadcast_force = 1;
244                 break;
245         case CLOCK_EVT_NOTIFY_BROADCAST_OFF:
246                 if (!tick_broadcast_force &&
247                     cpu_isset(cpu, tick_broadcast_mask)) {
248                         cpu_clear(cpu, tick_broadcast_mask);
249                         if (tick_broadcast_device.mode ==
250                             TICKDEV_MODE_PERIODIC)
251                                 tick_setup_periodic(dev, 0);
252                 }
253                 break;
254         }
255
256         if (cpus_empty(tick_broadcast_mask)) {
257                 if (!bc_stopped)
258                         clockevents_shutdown(bc);
259         } else if (bc_stopped) {
260                 if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
261                         tick_broadcast_start_periodic(bc);
262                 else
263                         tick_broadcast_setup_oneshot(bc);
264         }
265 out:
266         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
267 }
268
269 /*
270  * Powerstate information: The system enters/leaves a state, where
271  * affected devices might stop.
272  */
273 void tick_broadcast_on_off(unsigned long reason, int *oncpu)
274 {
275         if (!cpu_isset(*oncpu, cpu_online_map))
276                 printk(KERN_ERR "tick-broadcast: ignoring broadcast for "
277                        "offline CPU #%d\n", *oncpu);
278         else
279                 smp_call_function_single(*oncpu, tick_do_broadcast_on_off,
280                                          &reason, 1);
281 }
282
283 /*
284  * Set the periodic handler depending on broadcast on/off
285  */
286 void tick_set_periodic_handler(struct clock_event_device *dev, int broadcast)
287 {
288         if (!broadcast)
289                 dev->event_handler = tick_handle_periodic;
290         else
291                 dev->event_handler = tick_handle_periodic_broadcast;
292 }
293
294 /*
295  * Remove a CPU from broadcasting
296  */
297 void tick_shutdown_broadcast(unsigned int *cpup)
298 {
299         struct clock_event_device *bc;
300         unsigned long flags;
301         unsigned int cpu = *cpup;
302
303         spin_lock_irqsave(&tick_broadcast_lock, flags);
304
305         bc = tick_broadcast_device.evtdev;
306         cpu_clear(cpu, tick_broadcast_mask);
307
308         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
309                 if (bc && cpus_empty(tick_broadcast_mask))
310                         clockevents_shutdown(bc);
311         }
312
313         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
314 }
315
316 void tick_suspend_broadcast(void)
317 {
318         struct clock_event_device *bc;
319         unsigned long flags;
320
321         spin_lock_irqsave(&tick_broadcast_lock, flags);
322
323         bc = tick_broadcast_device.evtdev;
324         if (bc)
325                 clockevents_shutdown(bc);
326
327         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
328 }
329
330 int tick_resume_broadcast(void)
331 {
332         struct clock_event_device *bc;
333         unsigned long flags;
334         int broadcast = 0;
335
336         spin_lock_irqsave(&tick_broadcast_lock, flags);
337
338         bc = tick_broadcast_device.evtdev;
339
340         if (bc) {
341                 clockevents_set_mode(bc, CLOCK_EVT_MODE_RESUME);
342
343                 switch (tick_broadcast_device.mode) {
344                 case TICKDEV_MODE_PERIODIC:
345                         if(!cpus_empty(tick_broadcast_mask))
346                                 tick_broadcast_start_periodic(bc);
347                         broadcast = cpu_isset(smp_processor_id(),
348                                               tick_broadcast_mask);
349                         break;
350                 case TICKDEV_MODE_ONESHOT:
351                         broadcast = tick_resume_broadcast_oneshot(bc);
352                         break;
353                 }
354         }
355         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
356
357         return broadcast;
358 }
359
360
361 #ifdef CONFIG_TICK_ONESHOT
362
363 static cpumask_t tick_broadcast_oneshot_mask;
364
365 /*
366  * Debugging: see timer_list.c
367  */
368 cpumask_t *tick_get_broadcast_oneshot_mask(void)
369 {
370         return &tick_broadcast_oneshot_mask;
371 }
372
373 static int tick_broadcast_set_event(ktime_t expires, int force)
374 {
375         struct clock_event_device *bc = tick_broadcast_device.evtdev;
376
377         return tick_dev_program_event(bc, expires, force);
378 }
379
380 int tick_resume_broadcast_oneshot(struct clock_event_device *bc)
381 {
382         clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
383         return 0;
384 }
385
386 /*
387  * Handle oneshot mode broadcasting
388  */
389 static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
390 {
391         struct tick_device *td;
392         cpumask_t mask;
393         ktime_t now, next_event;
394         int cpu;
395
396         spin_lock(&tick_broadcast_lock);
397 again:
398         dev->next_event.tv64 = KTIME_MAX;
399         next_event.tv64 = KTIME_MAX;
400         mask = CPU_MASK_NONE;
401         now = ktime_get();
402         /* Find all expired events */
403         for_each_cpu_mask_nr(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 static void tick_broadcast_init_next_event(cpumask_t *mask, ktime_t expires)
496 {
497         struct tick_device *td;
498         int cpu;
499
500         for_each_cpu_mask_nr(cpu, *mask) {
501                 td = &per_cpu(tick_cpu_device, cpu);
502                 if (td->evtdev)
503                         td->evtdev->next_event = expires;
504         }
505 }
506
507 /**
508  * tick_broadcast_setup_oneshot - setup the broadcast device
509  */
510 void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
511 {
512         /* Set it up only once ! */
513         if (bc->event_handler != tick_handle_oneshot_broadcast) {
514                 int was_periodic = bc->mode == CLOCK_EVT_MODE_PERIODIC;
515                 int cpu = smp_processor_id();
516                 cpumask_t mask;
517
518                 bc->event_handler = tick_handle_oneshot_broadcast;
519                 clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
520
521                 /* Take the do_timer update */
522                 tick_do_timer_cpu = cpu;
523
524                 /*
525                  * We must be careful here. There might be other CPUs
526                  * waiting for periodic broadcast. We need to set the
527                  * oneshot_mask bits for those and program the
528                  * broadcast device to fire.
529                  */
530                 mask = tick_broadcast_mask;
531                 cpu_clear(cpu, mask);
532                 cpus_or(tick_broadcast_oneshot_mask,
533                         tick_broadcast_oneshot_mask, mask);
534
535                 if (was_periodic && !cpus_empty(mask)) {
536                         tick_broadcast_init_next_event(&mask, tick_next_period);
537                         tick_broadcast_set_event(tick_next_period, 1);
538                 } else
539                         bc->next_event.tv64 = KTIME_MAX;
540         }
541 }
542
543 /*
544  * Select oneshot operating mode for the broadcast device
545  */
546 void tick_broadcast_switch_to_oneshot(void)
547 {
548         struct clock_event_device *bc;
549         unsigned long flags;
550
551         spin_lock_irqsave(&tick_broadcast_lock, flags);
552
553         tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
554         bc = tick_broadcast_device.evtdev;
555         if (bc)
556                 tick_broadcast_setup_oneshot(bc);
557         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
558 }
559
560
561 /*
562  * Remove a dead CPU from broadcasting
563  */
564 void tick_shutdown_broadcast_oneshot(unsigned int *cpup)
565 {
566         unsigned long flags;
567         unsigned int cpu = *cpup;
568
569         spin_lock_irqsave(&tick_broadcast_lock, flags);
570
571         /*
572          * Clear the broadcast mask flag for the dead cpu, but do not
573          * stop the broadcast device!
574          */
575         cpu_clear(cpu, tick_broadcast_oneshot_mask);
576
577         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
578 }
579
580 /*
581  * Check, whether the broadcast device is in one shot mode
582  */
583 int tick_broadcast_oneshot_active(void)
584 {
585         return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
586 }
587
588 #endif