clockevent: Prevent dead lock on clockevents_lock
[linux-3.10.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 static struct tick_device tick_broadcast_device;
31 /* FIXME: Use cpumask_var_t. */
32 static DECLARE_BITMAP(tick_broadcast_mask, NR_CPUS);
33 static DECLARE_BITMAP(tmpmask, NR_CPUS);
34 static DEFINE_SPINLOCK(tick_broadcast_lock);
35 static int tick_broadcast_force;
36
37 #ifdef CONFIG_TICK_ONESHOT
38 static void tick_broadcast_clear_oneshot(int cpu);
39 #else
40 static inline void tick_broadcast_clear_oneshot(int cpu) { }
41 #endif
42
43 /*
44  * Debugging: see timer_list.c
45  */
46 struct tick_device *tick_get_broadcast_device(void)
47 {
48         return &tick_broadcast_device;
49 }
50
51 struct cpumask *tick_get_broadcast_mask(void)
52 {
53         return to_cpumask(tick_broadcast_mask);
54 }
55
56 /*
57  * Start the device in periodic mode
58  */
59 static void tick_broadcast_start_periodic(struct clock_event_device *bc)
60 {
61         if (bc)
62                 tick_setup_periodic(bc, 1);
63 }
64
65 /*
66  * Check, if the device can be utilized as broadcast device:
67  */
68 int tick_check_broadcast_device(struct clock_event_device *dev)
69 {
70         if ((tick_broadcast_device.evtdev &&
71              tick_broadcast_device.evtdev->rating >= dev->rating) ||
72              (dev->features & CLOCK_EVT_FEAT_C3STOP))
73                 return 0;
74
75         clockevents_exchange_device(NULL, dev);
76         tick_broadcast_device.evtdev = dev;
77         if (!cpumask_empty(tick_get_broadcast_mask()))
78                 tick_broadcast_start_periodic(dev);
79         return 1;
80 }
81
82 /*
83  * Check, if the device is the broadcast device
84  */
85 int tick_is_broadcast_device(struct clock_event_device *dev)
86 {
87         return (dev && tick_broadcast_device.evtdev == dev);
88 }
89
90 /*
91  * Check, if the device is disfunctional and a place holder, which
92  * needs to be handled by the broadcast device.
93  */
94 int tick_device_uses_broadcast(struct clock_event_device *dev, int cpu)
95 {
96         unsigned long flags;
97         int ret = 0;
98
99         spin_lock_irqsave(&tick_broadcast_lock, flags);
100
101         /*
102          * Devices might be registered with both periodic and oneshot
103          * mode disabled. This signals, that the device needs to be
104          * operated from the broadcast device and is a placeholder for
105          * the cpu local device.
106          */
107         if (!tick_device_is_functional(dev)) {
108                 dev->event_handler = tick_handle_periodic;
109                 cpumask_set_cpu(cpu, tick_get_broadcast_mask());
110                 tick_broadcast_start_periodic(tick_broadcast_device.evtdev);
111                 ret = 1;
112         } else {
113                 /*
114                  * When the new device is not affected by the stop
115                  * feature and the cpu is marked in the broadcast mask
116                  * then clear the broadcast bit.
117                  */
118                 if (!(dev->features & CLOCK_EVT_FEAT_C3STOP)) {
119                         int cpu = smp_processor_id();
120
121                         cpumask_clear_cpu(cpu, tick_get_broadcast_mask());
122                         tick_broadcast_clear_oneshot(cpu);
123                 }
124         }
125         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
126         return ret;
127 }
128
129 /*
130  * Broadcast the event to the cpus, which are set in the mask (mangled).
131  */
132 static void tick_do_broadcast(struct cpumask *mask)
133 {
134         int cpu = smp_processor_id();
135         struct tick_device *td;
136
137         /*
138          * Check, if the current cpu is in the mask
139          */
140         if (cpumask_test_cpu(cpu, mask)) {
141                 cpumask_clear_cpu(cpu, mask);
142                 td = &per_cpu(tick_cpu_device, cpu);
143                 td->evtdev->event_handler(td->evtdev);
144         }
145
146         if (!cpumask_empty(mask)) {
147                 /*
148                  * It might be necessary to actually check whether the devices
149                  * have different broadcast functions. For now, just use the
150                  * one of the first device. This works as long as we have this
151                  * misfeature only on x86 (lapic)
152                  */
153                 td = &per_cpu(tick_cpu_device, cpumask_first(mask));
154                 td->evtdev->broadcast(mask);
155         }
156 }
157
158 /*
159  * Periodic broadcast:
160  * - invoke the broadcast handlers
161  */
162 static void tick_do_periodic_broadcast(void)
163 {
164         spin_lock(&tick_broadcast_lock);
165
166         cpumask_and(to_cpumask(tmpmask),
167                     cpu_online_mask, tick_get_broadcast_mask());
168         tick_do_broadcast(to_cpumask(tmpmask));
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(unsigned long *reason)
209 {
210         struct clock_event_device *bc, *dev;
211         struct tick_device *td;
212         unsigned long flags;
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 = cpumask_empty(tick_get_broadcast_mask());
232
233         switch (*reason) {
234         case CLOCK_EVT_NOTIFY_BROADCAST_ON:
235         case CLOCK_EVT_NOTIFY_BROADCAST_FORCE:
236                 if (!cpumask_test_cpu(cpu, tick_get_broadcast_mask())) {
237                         cpumask_set_cpu(cpu, tick_get_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                     cpumask_test_cpu(cpu, tick_get_broadcast_mask())) {
248                         cpumask_clear_cpu(cpu, tick_get_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 (cpumask_empty(tick_get_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 (!cpumask_test_cpu(*oncpu, cpu_online_mask))
276                 printk(KERN_ERR "tick-broadcast: ignoring broadcast for "
277                        "offline CPU #%d\n", *oncpu);
278         else
279                 tick_do_broadcast_on_off(&reason);
280 }
281
282 /*
283  * Set the periodic handler depending on broadcast on/off
284  */
285 void tick_set_periodic_handler(struct clock_event_device *dev, int broadcast)
286 {
287         if (!broadcast)
288                 dev->event_handler = tick_handle_periodic;
289         else
290                 dev->event_handler = tick_handle_periodic_broadcast;
291 }
292
293 /*
294  * Remove a CPU from broadcasting
295  */
296 void tick_shutdown_broadcast(unsigned int *cpup)
297 {
298         struct clock_event_device *bc;
299         unsigned long flags;
300         unsigned int cpu = *cpup;
301
302         spin_lock_irqsave(&tick_broadcast_lock, flags);
303
304         bc = tick_broadcast_device.evtdev;
305         cpumask_clear_cpu(cpu, tick_get_broadcast_mask());
306
307         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC) {
308                 if (bc && cpumask_empty(tick_get_broadcast_mask()))
309                         clockevents_shutdown(bc);
310         }
311
312         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
313 }
314
315 void tick_suspend_broadcast(void)
316 {
317         struct clock_event_device *bc;
318         unsigned long flags;
319
320         spin_lock_irqsave(&tick_broadcast_lock, flags);
321
322         bc = tick_broadcast_device.evtdev;
323         if (bc)
324                 clockevents_shutdown(bc);
325
326         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
327 }
328
329 int tick_resume_broadcast(void)
330 {
331         struct clock_event_device *bc;
332         unsigned long flags;
333         int broadcast = 0;
334
335         spin_lock_irqsave(&tick_broadcast_lock, flags);
336
337         bc = tick_broadcast_device.evtdev;
338
339         if (bc) {
340                 clockevents_set_mode(bc, CLOCK_EVT_MODE_RESUME);
341
342                 switch (tick_broadcast_device.mode) {
343                 case TICKDEV_MODE_PERIODIC:
344                         if (!cpumask_empty(tick_get_broadcast_mask()))
345                                 tick_broadcast_start_periodic(bc);
346                         broadcast = cpumask_test_cpu(smp_processor_id(),
347                                                      tick_get_broadcast_mask());
348                         break;
349                 case TICKDEV_MODE_ONESHOT:
350                         broadcast = tick_resume_broadcast_oneshot(bc);
351                         break;
352                 }
353         }
354         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
355
356         return broadcast;
357 }
358
359
360 #ifdef CONFIG_TICK_ONESHOT
361
362 /* FIXME: use cpumask_var_t. */
363 static DECLARE_BITMAP(tick_broadcast_oneshot_mask, NR_CPUS);
364
365 /*
366  * Exposed for debugging: see timer_list.c
367  */
368 struct cpumask *tick_get_broadcast_oneshot_mask(void)
369 {
370         return to_cpumask(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  * Called from irq_enter() when idle was interrupted to reenable the
388  * per cpu device.
389  */
390 void tick_check_oneshot_broadcast(int cpu)
391 {
392         if (cpumask_test_cpu(cpu, to_cpumask(tick_broadcast_oneshot_mask))) {
393                 struct tick_device *td = &per_cpu(tick_cpu_device, cpu);
394
395                 clockevents_set_mode(td->evtdev, CLOCK_EVT_MODE_ONESHOT);
396         }
397 }
398
399 /*
400  * Handle oneshot mode broadcasting
401  */
402 static void tick_handle_oneshot_broadcast(struct clock_event_device *dev)
403 {
404         struct tick_device *td;
405         ktime_t now, next_event;
406         int cpu;
407
408         spin_lock(&tick_broadcast_lock);
409 again:
410         dev->next_event.tv64 = KTIME_MAX;
411         next_event.tv64 = KTIME_MAX;
412         cpumask_clear(to_cpumask(tmpmask));
413         now = ktime_get();
414         /* Find all expired events */
415         for_each_cpu(cpu, tick_get_broadcast_oneshot_mask()) {
416                 td = &per_cpu(tick_cpu_device, cpu);
417                 if (td->evtdev->next_event.tv64 <= now.tv64)
418                         cpumask_set_cpu(cpu, to_cpumask(tmpmask));
419                 else if (td->evtdev->next_event.tv64 < next_event.tv64)
420                         next_event.tv64 = td->evtdev->next_event.tv64;
421         }
422
423         /*
424          * Wakeup the cpus which have an expired event.
425          */
426         tick_do_broadcast(to_cpumask(tmpmask));
427
428         /*
429          * Two reasons for reprogram:
430          *
431          * - The global event did not expire any CPU local
432          * events. This happens in dyntick mode, as the maximum PIT
433          * delta is quite small.
434          *
435          * - There are pending events on sleeping CPUs which were not
436          * in the event mask
437          */
438         if (next_event.tv64 != KTIME_MAX) {
439                 /*
440                  * Rearm the broadcast device. If event expired,
441                  * repeat the above
442                  */
443                 if (tick_broadcast_set_event(next_event, 0))
444                         goto again;
445         }
446         spin_unlock(&tick_broadcast_lock);
447 }
448
449 /*
450  * Powerstate information: The system enters/leaves a state, where
451  * affected devices might stop
452  */
453 void tick_broadcast_oneshot_control(unsigned long reason)
454 {
455         struct clock_event_device *bc, *dev;
456         struct tick_device *td;
457         unsigned long flags;
458         int cpu;
459
460         spin_lock_irqsave(&tick_broadcast_lock, flags);
461
462         /*
463          * Periodic mode does not care about the enter/exit of power
464          * states
465          */
466         if (tick_broadcast_device.mode == TICKDEV_MODE_PERIODIC)
467                 goto out;
468
469         bc = tick_broadcast_device.evtdev;
470         cpu = smp_processor_id();
471         td = &per_cpu(tick_cpu_device, cpu);
472         dev = td->evtdev;
473
474         if (!(dev->features & CLOCK_EVT_FEAT_C3STOP))
475                 goto out;
476
477         if (reason == CLOCK_EVT_NOTIFY_BROADCAST_ENTER) {
478                 if (!cpumask_test_cpu(cpu, tick_get_broadcast_oneshot_mask())) {
479                         cpumask_set_cpu(cpu, tick_get_broadcast_oneshot_mask());
480                         clockevents_set_mode(dev, CLOCK_EVT_MODE_SHUTDOWN);
481                         if (dev->next_event.tv64 < bc->next_event.tv64)
482                                 tick_broadcast_set_event(dev->next_event, 1);
483                 }
484         } else {
485                 if (cpumask_test_cpu(cpu, tick_get_broadcast_oneshot_mask())) {
486                         cpumask_clear_cpu(cpu,
487                                           tick_get_broadcast_oneshot_mask());
488                         clockevents_set_mode(dev, CLOCK_EVT_MODE_ONESHOT);
489                         if (dev->next_event.tv64 != KTIME_MAX)
490                                 tick_program_event(dev->next_event, 1);
491                 }
492         }
493
494 out:
495         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
496 }
497
498 /*
499  * Reset the one shot broadcast for a cpu
500  *
501  * Called with tick_broadcast_lock held
502  */
503 static void tick_broadcast_clear_oneshot(int cpu)
504 {
505         cpumask_clear_cpu(cpu, tick_get_broadcast_oneshot_mask());
506 }
507
508 static void tick_broadcast_init_next_event(struct cpumask *mask,
509                                            ktime_t expires)
510 {
511         struct tick_device *td;
512         int cpu;
513
514         for_each_cpu(cpu, mask) {
515                 td = &per_cpu(tick_cpu_device, cpu);
516                 if (td->evtdev)
517                         td->evtdev->next_event = expires;
518         }
519 }
520
521 /**
522  * tick_broadcast_setup_oneshot - setup the broadcast device
523  */
524 void tick_broadcast_setup_oneshot(struct clock_event_device *bc)
525 {
526         /* Set it up only once ! */
527         if (bc->event_handler != tick_handle_oneshot_broadcast) {
528                 int was_periodic = bc->mode == CLOCK_EVT_MODE_PERIODIC;
529                 int cpu = smp_processor_id();
530
531                 bc->event_handler = tick_handle_oneshot_broadcast;
532                 clockevents_set_mode(bc, CLOCK_EVT_MODE_ONESHOT);
533
534                 /* Take the do_timer update */
535                 tick_do_timer_cpu = cpu;
536
537                 /*
538                  * We must be careful here. There might be other CPUs
539                  * waiting for periodic broadcast. We need to set the
540                  * oneshot_mask bits for those and program the
541                  * broadcast device to fire.
542                  */
543                 cpumask_copy(to_cpumask(tmpmask), tick_get_broadcast_mask());
544                 cpumask_clear_cpu(cpu, to_cpumask(tmpmask));
545                 cpumask_or(tick_get_broadcast_oneshot_mask(),
546                            tick_get_broadcast_oneshot_mask(),
547                            to_cpumask(tmpmask));
548
549                 if (was_periodic && !cpumask_empty(to_cpumask(tmpmask))) {
550                         tick_broadcast_init_next_event(to_cpumask(tmpmask),
551                                                        tick_next_period);
552                         tick_broadcast_set_event(tick_next_period, 1);
553                 } else
554                         bc->next_event.tv64 = KTIME_MAX;
555         }
556 }
557
558 /*
559  * Select oneshot operating mode for the broadcast device
560  */
561 void tick_broadcast_switch_to_oneshot(void)
562 {
563         struct clock_event_device *bc;
564         unsigned long flags;
565
566         spin_lock_irqsave(&tick_broadcast_lock, flags);
567
568         tick_broadcast_device.mode = TICKDEV_MODE_ONESHOT;
569         bc = tick_broadcast_device.evtdev;
570         if (bc)
571                 tick_broadcast_setup_oneshot(bc);
572         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
573 }
574
575
576 /*
577  * Remove a dead CPU from broadcasting
578  */
579 void tick_shutdown_broadcast_oneshot(unsigned int *cpup)
580 {
581         unsigned long flags;
582         unsigned int cpu = *cpup;
583
584         spin_lock_irqsave(&tick_broadcast_lock, flags);
585
586         /*
587          * Clear the broadcast mask flag for the dead cpu, but do not
588          * stop the broadcast device!
589          */
590         cpumask_clear_cpu(cpu, tick_get_broadcast_oneshot_mask());
591
592         spin_unlock_irqrestore(&tick_broadcast_lock, flags);
593 }
594
595 /*
596  * Check, whether the broadcast device is in one shot mode
597  */
598 int tick_broadcast_oneshot_active(void)
599 {
600         return tick_broadcast_device.mode == TICKDEV_MODE_ONESHOT;
601 }
602
603 #endif