0a7840aeb0fb9efbc18a6e6e8e6f01de17ed91cd
[linux-2.6.git] / kernel / irq / manage.c
1 /*
2  * linux/kernel/irq/manage.c
3  *
4  * Copyright (C) 1992, 1998-2006 Linus Torvalds, Ingo Molnar
5  * Copyright (C) 2005-2006 Thomas Gleixner
6  *
7  * This file contains driver APIs to the irq subsystem.
8  */
9
10 #include <linux/irq.h>
11 #include <linux/kthread.h>
12 #include <linux/module.h>
13 #include <linux/random.h>
14 #include <linux/interrupt.h>
15 #include <linux/slab.h>
16 #include <linux/sched.h>
17
18 #include "internals.h"
19
20 #ifdef CONFIG_IRQ_FORCED_THREADING
21 __read_mostly bool force_irqthreads;
22
23 static int __init setup_forced_irqthreads(char *arg)
24 {
25         force_irqthreads = true;
26         return 0;
27 }
28 early_param("threadirqs", setup_forced_irqthreads);
29 #endif
30
31 /**
32  *      synchronize_irq - wait for pending IRQ handlers (on other CPUs)
33  *      @irq: interrupt number to wait for
34  *
35  *      This function waits for any pending IRQ handlers for this interrupt
36  *      to complete before returning. If you use this function while
37  *      holding a resource the IRQ handler may need you will deadlock.
38  *
39  *      This function may be called - with care - from IRQ context.
40  */
41 void synchronize_irq(unsigned int irq)
42 {
43         struct irq_desc *desc = irq_to_desc(irq);
44         bool inprogress;
45
46         if (!desc)
47                 return;
48
49         do {
50                 unsigned long flags;
51
52                 /*
53                  * Wait until we're out of the critical section.  This might
54                  * give the wrong answer due to the lack of memory barriers.
55                  */
56                 while (irqd_irq_inprogress(&desc->irq_data))
57                         cpu_relax();
58
59                 /* Ok, that indicated we're done: double-check carefully. */
60                 raw_spin_lock_irqsave(&desc->lock, flags);
61                 inprogress = irqd_irq_inprogress(&desc->irq_data);
62                 raw_spin_unlock_irqrestore(&desc->lock, flags);
63
64                 /* Oops, that failed? */
65         } while (inprogress);
66
67         /*
68          * We made sure that no hardirq handler is running. Now verify
69          * that no threaded handlers are active.
70          */
71         wait_event(desc->wait_for_threads, !atomic_read(&desc->threads_active));
72 }
73 EXPORT_SYMBOL(synchronize_irq);
74
75 #ifdef CONFIG_SMP
76 cpumask_var_t irq_default_affinity;
77
78 /**
79  *      irq_can_set_affinity - Check if the affinity of a given irq can be set
80  *      @irq:           Interrupt to check
81  *
82  */
83 int irq_can_set_affinity(unsigned int irq)
84 {
85         struct irq_desc *desc = irq_to_desc(irq);
86
87         if (!desc || !irqd_can_balance(&desc->irq_data) ||
88             !desc->irq_data.chip || !desc->irq_data.chip->irq_set_affinity)
89                 return 0;
90
91         return 1;
92 }
93
94 /**
95  *      irq_set_thread_affinity - Notify irq threads to adjust affinity
96  *      @desc:          irq descriptor which has affitnity changed
97  *
98  *      We just set IRQTF_AFFINITY and delegate the affinity setting
99  *      to the interrupt thread itself. We can not call
100  *      set_cpus_allowed_ptr() here as we hold desc->lock and this
101  *      code can be called from hard interrupt context.
102  */
103 void irq_set_thread_affinity(struct irq_desc *desc)
104 {
105         struct irqaction *action = desc->action;
106
107         while (action) {
108                 if (action->thread)
109                         set_bit(IRQTF_AFFINITY, &action->thread_flags);
110                 action = action->next;
111         }
112 }
113
114 #ifdef CONFIG_GENERIC_PENDING_IRQ
115 static inline bool irq_can_move_pcntxt(struct irq_data *data)
116 {
117         return irqd_can_move_in_process_context(data);
118 }
119 static inline bool irq_move_pending(struct irq_data *data)
120 {
121         return irqd_is_setaffinity_pending(data);
122 }
123 static inline void
124 irq_copy_pending(struct irq_desc *desc, const struct cpumask *mask)
125 {
126         cpumask_copy(desc->pending_mask, mask);
127 }
128 static inline void
129 irq_get_pending(struct cpumask *mask, struct irq_desc *desc)
130 {
131         cpumask_copy(mask, desc->pending_mask);
132 }
133 #else
134 static inline bool irq_can_move_pcntxt(struct irq_data *data) { return true; }
135 static inline bool irq_move_pending(struct irq_data *data) { return false; }
136 static inline void
137 irq_copy_pending(struct irq_desc *desc, const struct cpumask *mask) { }
138 static inline void
139 irq_get_pending(struct cpumask *mask, struct irq_desc *desc) { }
140 #endif
141
142 int __irq_set_affinity_locked(struct irq_data *data, const struct cpumask *mask)
143 {
144         struct irq_chip *chip = irq_data_get_irq_chip(data);
145         struct irq_desc *desc = irq_data_to_desc(data);
146         int ret = 0;
147
148         if (!chip || !chip->irq_set_affinity)
149                 return -EINVAL;
150
151         if (irq_can_move_pcntxt(data)) {
152                 ret = chip->irq_set_affinity(data, mask, false);
153                 switch (ret) {
154                 case IRQ_SET_MASK_OK:
155                         cpumask_copy(data->affinity, mask);
156                 case IRQ_SET_MASK_OK_NOCOPY:
157                         irq_set_thread_affinity(desc);
158                         ret = 0;
159                 }
160         } else {
161                 irqd_set_move_pending(data);
162                 irq_copy_pending(desc, mask);
163         }
164
165         if (desc->affinity_notify) {
166                 kref_get(&desc->affinity_notify->kref);
167                 schedule_work(&desc->affinity_notify->work);
168         }
169         irqd_set(data, IRQD_AFFINITY_SET);
170
171         return ret;
172 }
173
174 /**
175  *      irq_set_affinity - Set the irq affinity of a given irq
176  *      @irq:           Interrupt to set affinity
177  *      @mask:          cpumask
178  *
179  */
180 int irq_set_affinity(unsigned int irq, const struct cpumask *mask)
181 {
182         struct irq_desc *desc = irq_to_desc(irq);
183         unsigned long flags;
184         int ret;
185
186         if (!desc)
187                 return -EINVAL;
188
189         raw_spin_lock_irqsave(&desc->lock, flags);
190         ret =  __irq_set_affinity_locked(irq_desc_get_irq_data(desc), mask);
191         raw_spin_unlock_irqrestore(&desc->lock, flags);
192         return ret;
193 }
194
195 int irq_set_affinity_hint(unsigned int irq, const struct cpumask *m)
196 {
197         unsigned long flags;
198         struct irq_desc *desc = irq_get_desc_lock(irq, &flags);
199
200         if (!desc)
201                 return -EINVAL;
202         desc->affinity_hint = m;
203         irq_put_desc_unlock(desc, flags);
204         return 0;
205 }
206 EXPORT_SYMBOL_GPL(irq_set_affinity_hint);
207
208 static void irq_affinity_notify(struct work_struct *work)
209 {
210         struct irq_affinity_notify *notify =
211                 container_of(work, struct irq_affinity_notify, work);
212         struct irq_desc *desc = irq_to_desc(notify->irq);
213         cpumask_var_t cpumask;
214         unsigned long flags;
215
216         if (!desc || !alloc_cpumask_var(&cpumask, GFP_KERNEL))
217                 goto out;
218
219         raw_spin_lock_irqsave(&desc->lock, flags);
220         if (irq_move_pending(&desc->irq_data))
221                 irq_get_pending(cpumask, desc);
222         else
223                 cpumask_copy(cpumask, desc->irq_data.affinity);
224         raw_spin_unlock_irqrestore(&desc->lock, flags);
225
226         notify->notify(notify, cpumask);
227
228         free_cpumask_var(cpumask);
229 out:
230         kref_put(&notify->kref, notify->release);
231 }
232
233 /**
234  *      irq_set_affinity_notifier - control notification of IRQ affinity changes
235  *      @irq:           Interrupt for which to enable/disable notification
236  *      @notify:        Context for notification, or %NULL to disable
237  *                      notification.  Function pointers must be initialised;
238  *                      the other fields will be initialised by this function.
239  *
240  *      Must be called in process context.  Notification may only be enabled
241  *      after the IRQ is allocated and must be disabled before the IRQ is
242  *      freed using free_irq().
243  */
244 int
245 irq_set_affinity_notifier(unsigned int irq, struct irq_affinity_notify *notify)
246 {
247         struct irq_desc *desc = irq_to_desc(irq);
248         struct irq_affinity_notify *old_notify;
249         unsigned long flags;
250
251         /* The release function is promised process context */
252         might_sleep();
253
254         if (!desc)
255                 return -EINVAL;
256
257         /* Complete initialisation of *notify */
258         if (notify) {
259                 notify->irq = irq;
260                 kref_init(&notify->kref);
261                 INIT_WORK(&notify->work, irq_affinity_notify);
262         }
263
264         raw_spin_lock_irqsave(&desc->lock, flags);
265         old_notify = desc->affinity_notify;
266         desc->affinity_notify = notify;
267         raw_spin_unlock_irqrestore(&desc->lock, flags);
268
269         if (old_notify)
270                 kref_put(&old_notify->kref, old_notify->release);
271
272         return 0;
273 }
274 EXPORT_SYMBOL_GPL(irq_set_affinity_notifier);
275
276 #ifndef CONFIG_AUTO_IRQ_AFFINITY
277 /*
278  * Generic version of the affinity autoselector.
279  */
280 static int
281 setup_affinity(unsigned int irq, struct irq_desc *desc, struct cpumask *mask)
282 {
283         struct irq_chip *chip = irq_desc_get_chip(desc);
284         struct cpumask *set = irq_default_affinity;
285         int ret;
286
287         /* Excludes PER_CPU and NO_BALANCE interrupts */
288         if (!irq_can_set_affinity(irq))
289                 return 0;
290
291         /*
292          * Preserve an userspace affinity setup, but make sure that
293          * one of the targets is online.
294          */
295         if (irqd_has_set(&desc->irq_data, IRQD_AFFINITY_SET)) {
296                 if (cpumask_intersects(desc->irq_data.affinity,
297                                        cpu_online_mask))
298                         set = desc->irq_data.affinity;
299                 else
300                         irqd_clear(&desc->irq_data, IRQD_AFFINITY_SET);
301         }
302
303         cpumask_and(mask, cpu_online_mask, set);
304         ret = chip->irq_set_affinity(&desc->irq_data, mask, false);
305         switch (ret) {
306         case IRQ_SET_MASK_OK:
307                 cpumask_copy(desc->irq_data.affinity, mask);
308         case IRQ_SET_MASK_OK_NOCOPY:
309                 irq_set_thread_affinity(desc);
310         }
311         return 0;
312 }
313 #else
314 static inline int
315 setup_affinity(unsigned int irq, struct irq_desc *d, struct cpumask *mask)
316 {
317         return irq_select_affinity(irq);
318 }
319 #endif
320
321 /*
322  * Called when affinity is set via /proc/irq
323  */
324 int irq_select_affinity_usr(unsigned int irq, struct cpumask *mask)
325 {
326         struct irq_desc *desc = irq_to_desc(irq);
327         unsigned long flags;
328         int ret;
329
330         raw_spin_lock_irqsave(&desc->lock, flags);
331         ret = setup_affinity(irq, desc, mask);
332         raw_spin_unlock_irqrestore(&desc->lock, flags);
333         return ret;
334 }
335
336 #else
337 static inline int
338 setup_affinity(unsigned int irq, struct irq_desc *desc, struct cpumask *mask)
339 {
340         return 0;
341 }
342 #endif
343
344 void __disable_irq(struct irq_desc *desc, unsigned int irq, bool suspend)
345 {
346         if (suspend) {
347                 if (!desc->action || (desc->action->flags & IRQF_NO_SUSPEND))
348                         return;
349                 desc->istate |= IRQS_SUSPENDED;
350         }
351
352         if (!desc->depth++)
353                 irq_disable(desc);
354 }
355
356 static int __disable_irq_nosync(unsigned int irq)
357 {
358         unsigned long flags;
359         struct irq_desc *desc = irq_get_desc_buslock(irq, &flags);
360
361         if (!desc)
362                 return -EINVAL;
363         __disable_irq(desc, irq, false);
364         irq_put_desc_busunlock(desc, flags);
365         return 0;
366 }
367
368 /**
369  *      disable_irq_nosync - disable an irq without waiting
370  *      @irq: Interrupt to disable
371  *
372  *      Disable the selected interrupt line.  Disables and Enables are
373  *      nested.
374  *      Unlike disable_irq(), this function does not ensure existing
375  *      instances of the IRQ handler have completed before returning.
376  *
377  *      This function may be called from IRQ context.
378  */
379 void disable_irq_nosync(unsigned int irq)
380 {
381         __disable_irq_nosync(irq);
382 }
383 EXPORT_SYMBOL(disable_irq_nosync);
384
385 /**
386  *      disable_irq - disable an irq and wait for completion
387  *      @irq: Interrupt to disable
388  *
389  *      Disable the selected interrupt line.  Enables and Disables are
390  *      nested.
391  *      This function waits for any pending IRQ handlers for this interrupt
392  *      to complete before returning. If you use this function while
393  *      holding a resource the IRQ handler may need you will deadlock.
394  *
395  *      This function may be called - with care - from IRQ context.
396  */
397 void disable_irq(unsigned int irq)
398 {
399         if (!__disable_irq_nosync(irq))
400                 synchronize_irq(irq);
401 }
402 EXPORT_SYMBOL(disable_irq);
403
404 void __enable_irq(struct irq_desc *desc, unsigned int irq, bool resume)
405 {
406         if (resume) {
407                 if (!(desc->istate & IRQS_SUSPENDED)) {
408                         if (!desc->action)
409                                 return;
410                         if (!(desc->action->flags & IRQF_FORCE_RESUME))
411                                 return;
412                         /* Pretend that it got disabled ! */
413                         desc->depth++;
414                 }
415                 desc->istate &= ~IRQS_SUSPENDED;
416         }
417
418         switch (desc->depth) {
419         case 0:
420  err_out:
421                 WARN(1, KERN_WARNING "Unbalanced enable for IRQ %d\n", irq);
422                 break;
423         case 1: {
424                 if (desc->istate & IRQS_SUSPENDED)
425                         goto err_out;
426                 /* Prevent probing on this irq: */
427                 irq_settings_set_noprobe(desc);
428                 irq_enable(desc);
429                 check_irq_resend(desc, irq);
430                 /* fall-through */
431         }
432         default:
433                 desc->depth--;
434         }
435 }
436
437 /**
438  *      enable_irq - enable handling of an irq
439  *      @irq: Interrupt to enable
440  *
441  *      Undoes the effect of one call to disable_irq().  If this
442  *      matches the last disable, processing of interrupts on this
443  *      IRQ line is re-enabled.
444  *
445  *      This function may be called from IRQ context only when
446  *      desc->irq_data.chip->bus_lock and desc->chip->bus_sync_unlock are NULL !
447  */
448 void enable_irq(unsigned int irq)
449 {
450         unsigned long flags;
451         struct irq_desc *desc = irq_get_desc_buslock(irq, &flags);
452
453         if (!desc)
454                 return;
455         if (WARN(!desc->irq_data.chip,
456                  KERN_ERR "enable_irq before setup/request_irq: irq %u\n", irq))
457                 goto out;
458
459         __enable_irq(desc, irq, false);
460 out:
461         irq_put_desc_busunlock(desc, flags);
462 }
463 EXPORT_SYMBOL(enable_irq);
464
465 static int set_irq_wake_real(unsigned int irq, unsigned int on)
466 {
467         struct irq_desc *desc = irq_to_desc(irq);
468         int ret = -ENXIO;
469
470         if (desc->irq_data.chip->irq_set_wake)
471                 ret = desc->irq_data.chip->irq_set_wake(&desc->irq_data, on);
472
473         return ret;
474 }
475
476 /**
477  *      irq_set_irq_wake - control irq power management wakeup
478  *      @irq:   interrupt to control
479  *      @on:    enable/disable power management wakeup
480  *
481  *      Enable/disable power management wakeup mode, which is
482  *      disabled by default.  Enables and disables must match,
483  *      just as they match for non-wakeup mode support.
484  *
485  *      Wakeup mode lets this IRQ wake the system from sleep
486  *      states like "suspend to RAM".
487  */
488 int irq_set_irq_wake(unsigned int irq, unsigned int on)
489 {
490         unsigned long flags;
491         struct irq_desc *desc = irq_get_desc_buslock(irq, &flags);
492         int ret = 0;
493
494         if (!desc)
495                 return -EINVAL;
496
497         /* wakeup-capable irqs can be shared between drivers that
498          * don't need to have the same sleep mode behaviors.
499          */
500         if (on) {
501                 if (desc->wake_depth++ == 0) {
502                         ret = set_irq_wake_real(irq, on);
503                         if (ret)
504                                 desc->wake_depth = 0;
505                         else
506                                 irqd_set(&desc->irq_data, IRQD_WAKEUP_STATE);
507                 }
508         } else {
509                 if (desc->wake_depth == 0) {
510                         WARN(1, "Unbalanced IRQ %d wake disable\n", irq);
511                 } else if (--desc->wake_depth == 0) {
512                         ret = set_irq_wake_real(irq, on);
513                         if (ret)
514                                 desc->wake_depth = 1;
515                         else
516                                 irqd_clear(&desc->irq_data, IRQD_WAKEUP_STATE);
517                 }
518         }
519         irq_put_desc_busunlock(desc, flags);
520         return ret;
521 }
522 EXPORT_SYMBOL(irq_set_irq_wake);
523
524 /*
525  * Internal function that tells the architecture code whether a
526  * particular irq has been exclusively allocated or is available
527  * for driver use.
528  */
529 int can_request_irq(unsigned int irq, unsigned long irqflags)
530 {
531         unsigned long flags;
532         struct irq_desc *desc = irq_get_desc_lock(irq, &flags);
533         int canrequest = 0;
534
535         if (!desc)
536                 return 0;
537
538         if (irq_settings_can_request(desc)) {
539                 if (desc->action)
540                         if (irqflags & desc->action->flags & IRQF_SHARED)
541                                 canrequest =1;
542         }
543         irq_put_desc_unlock(desc, flags);
544         return canrequest;
545 }
546
547 int __irq_set_trigger(struct irq_desc *desc, unsigned int irq,
548                       unsigned long flags)
549 {
550         struct irq_chip *chip = desc->irq_data.chip;
551         int ret, unmask = 0;
552
553         if (!chip || !chip->irq_set_type) {
554                 /*
555                  * IRQF_TRIGGER_* but the PIC does not support multiple
556                  * flow-types?
557                  */
558                 pr_debug("No set_type function for IRQ %d (%s)\n", irq,
559                                 chip ? (chip->name ? : "unknown") : "unknown");
560                 return 0;
561         }
562
563         flags &= IRQ_TYPE_SENSE_MASK;
564
565         if (chip->flags & IRQCHIP_SET_TYPE_MASKED) {
566                 if (!irqd_irq_masked(&desc->irq_data))
567                         mask_irq(desc);
568                 if (!irqd_irq_disabled(&desc->irq_data))
569                         unmask = 1;
570         }
571
572         /* caller masked out all except trigger mode flags */
573         ret = chip->irq_set_type(&desc->irq_data, flags);
574
575         switch (ret) {
576         case IRQ_SET_MASK_OK:
577                 irqd_clear(&desc->irq_data, IRQD_TRIGGER_MASK);
578                 irqd_set(&desc->irq_data, flags);
579
580         case IRQ_SET_MASK_OK_NOCOPY:
581                 flags = irqd_get_trigger_type(&desc->irq_data);
582                 irq_settings_set_trigger_mask(desc, flags);
583                 irqd_clear(&desc->irq_data, IRQD_LEVEL);
584                 irq_settings_clr_level(desc);
585                 if (flags & IRQ_TYPE_LEVEL_MASK) {
586                         irq_settings_set_level(desc);
587                         irqd_set(&desc->irq_data, IRQD_LEVEL);
588                 }
589
590                 ret = 0;
591                 break;
592         default:
593                 pr_err("setting trigger mode %lu for irq %u failed (%pF)\n",
594                        flags, irq, chip->irq_set_type);
595         }
596         if (unmask)
597                 unmask_irq(desc);
598         return ret;
599 }
600
601 /*
602  * Default primary interrupt handler for threaded interrupts. Is
603  * assigned as primary handler when request_threaded_irq is called
604  * with handler == NULL. Useful for oneshot interrupts.
605  */
606 static irqreturn_t irq_default_primary_handler(int irq, void *dev_id)
607 {
608         return IRQ_WAKE_THREAD;
609 }
610
611 /*
612  * Primary handler for nested threaded interrupts. Should never be
613  * called.
614  */
615 static irqreturn_t irq_nested_primary_handler(int irq, void *dev_id)
616 {
617         WARN(1, "Primary handler called for nested irq %d\n", irq);
618         return IRQ_NONE;
619 }
620
621 static int irq_wait_for_interrupt(struct irqaction *action)
622 {
623         while (!kthread_should_stop()) {
624                 set_current_state(TASK_INTERRUPTIBLE);
625
626                 if (test_and_clear_bit(IRQTF_RUNTHREAD,
627                                        &action->thread_flags)) {
628                         __set_current_state(TASK_RUNNING);
629                         return 0;
630                 }
631                 schedule();
632         }
633         return -1;
634 }
635
636 /*
637  * Oneshot interrupts keep the irq line masked until the threaded
638  * handler finished. unmask if the interrupt has not been disabled and
639  * is marked MASKED.
640  */
641 static void irq_finalize_oneshot(struct irq_desc *desc,
642                                  struct irqaction *action, bool force)
643 {
644         if (!(desc->istate & IRQS_ONESHOT))
645                 return;
646 again:
647         chip_bus_lock(desc);
648         raw_spin_lock_irq(&desc->lock);
649
650         /*
651          * Implausible though it may be we need to protect us against
652          * the following scenario:
653          *
654          * The thread is faster done than the hard interrupt handler
655          * on the other CPU. If we unmask the irq line then the
656          * interrupt can come in again and masks the line, leaves due
657          * to IRQS_INPROGRESS and the irq line is masked forever.
658          *
659          * This also serializes the state of shared oneshot handlers
660          * versus "desc->threads_onehsot |= action->thread_mask;" in
661          * irq_wake_thread(). See the comment there which explains the
662          * serialization.
663          */
664         if (unlikely(irqd_irq_inprogress(&desc->irq_data))) {
665                 raw_spin_unlock_irq(&desc->lock);
666                 chip_bus_sync_unlock(desc);
667                 cpu_relax();
668                 goto again;
669         }
670
671         /*
672          * Now check again, whether the thread should run. Otherwise
673          * we would clear the threads_oneshot bit of this thread which
674          * was just set.
675          */
676         if (!force && test_bit(IRQTF_RUNTHREAD, &action->thread_flags))
677                 goto out_unlock;
678
679         desc->threads_oneshot &= ~action->thread_mask;
680
681         if (!desc->threads_oneshot && !irqd_irq_disabled(&desc->irq_data) &&
682             irqd_irq_masked(&desc->irq_data))
683                 unmask_irq(desc);
684
685 out_unlock:
686         raw_spin_unlock_irq(&desc->lock);
687         chip_bus_sync_unlock(desc);
688 }
689
690 #ifdef CONFIG_SMP
691 /*
692  * Check whether we need to chasnge the affinity of the interrupt thread.
693  */
694 static void
695 irq_thread_check_affinity(struct irq_desc *desc, struct irqaction *action)
696 {
697         cpumask_var_t mask;
698
699         if (!test_and_clear_bit(IRQTF_AFFINITY, &action->thread_flags))
700                 return;
701
702         /*
703          * In case we are out of memory we set IRQTF_AFFINITY again and
704          * try again next time
705          */
706         if (!alloc_cpumask_var(&mask, GFP_KERNEL)) {
707                 set_bit(IRQTF_AFFINITY, &action->thread_flags);
708                 return;
709         }
710
711         raw_spin_lock_irq(&desc->lock);
712         cpumask_copy(mask, desc->irq_data.affinity);
713         raw_spin_unlock_irq(&desc->lock);
714
715         set_cpus_allowed_ptr(current, mask);
716         free_cpumask_var(mask);
717 }
718 #else
719 static inline void
720 irq_thread_check_affinity(struct irq_desc *desc, struct irqaction *action) { }
721 #endif
722
723 /*
724  * Interrupts which are not explicitely requested as threaded
725  * interrupts rely on the implicit bh/preempt disable of the hard irq
726  * context. So we need to disable bh here to avoid deadlocks and other
727  * side effects.
728  */
729 static irqreturn_t
730 irq_forced_thread_fn(struct irq_desc *desc, struct irqaction *action)
731 {
732         irqreturn_t ret;
733
734         local_bh_disable();
735         ret = action->thread_fn(action->irq, action->dev_id);
736         irq_finalize_oneshot(desc, action, false);
737         local_bh_enable();
738         return ret;
739 }
740
741 /*
742  * Interrupts explicitely requested as threaded interupts want to be
743  * preemtible - many of them need to sleep and wait for slow busses to
744  * complete.
745  */
746 static irqreturn_t irq_thread_fn(struct irq_desc *desc,
747                 struct irqaction *action)
748 {
749         irqreturn_t ret;
750
751         ret = action->thread_fn(action->irq, action->dev_id);
752         irq_finalize_oneshot(desc, action, false);
753         return ret;
754 }
755
756 /*
757  * Interrupt handler thread
758  */
759 static int irq_thread(void *data)
760 {
761         static const struct sched_param param = {
762                 .sched_priority = MAX_USER_RT_PRIO/2,
763         };
764         struct irqaction *action = data;
765         struct irq_desc *desc = irq_to_desc(action->irq);
766         irqreturn_t (*handler_fn)(struct irq_desc *desc,
767                         struct irqaction *action);
768         int wake;
769
770         if (force_irqthreads & test_bit(IRQTF_FORCED_THREAD,
771                                         &action->thread_flags))
772                 handler_fn = irq_forced_thread_fn;
773         else
774                 handler_fn = irq_thread_fn;
775
776         sched_setscheduler(current, SCHED_FIFO, &param);
777         current->irqaction = action;
778
779         while (!irq_wait_for_interrupt(action)) {
780
781                 irq_thread_check_affinity(desc, action);
782
783                 atomic_inc(&desc->threads_active);
784
785                 raw_spin_lock_irq(&desc->lock);
786                 if (unlikely(irqd_irq_disabled(&desc->irq_data))) {
787                         /*
788                          * CHECKME: We might need a dedicated
789                          * IRQ_THREAD_PENDING flag here, which
790                          * retriggers the thread in check_irq_resend()
791                          * but AFAICT IRQS_PENDING should be fine as it
792                          * retriggers the interrupt itself --- tglx
793                          */
794                         desc->istate |= IRQS_PENDING;
795                         raw_spin_unlock_irq(&desc->lock);
796                 } else {
797                         irqreturn_t action_ret;
798
799                         raw_spin_unlock_irq(&desc->lock);
800                         action_ret = handler_fn(desc, action);
801                         if (!noirqdebug)
802                                 note_interrupt(action->irq, desc, action_ret);
803                 }
804
805                 wake = atomic_dec_and_test(&desc->threads_active);
806
807                 if (wake && waitqueue_active(&desc->wait_for_threads))
808                         wake_up(&desc->wait_for_threads);
809         }
810
811         /* Prevent a stale desc->threads_oneshot */
812         irq_finalize_oneshot(desc, action, true);
813
814         /*
815          * Clear irqaction. Otherwise exit_irq_thread() would make
816          * fuzz about an active irq thread going into nirvana.
817          */
818         current->irqaction = NULL;
819         return 0;
820 }
821
822 /*
823  * Called from do_exit()
824  */
825 void exit_irq_thread(void)
826 {
827         struct task_struct *tsk = current;
828         struct irq_desc *desc;
829
830         if (!tsk->irqaction)
831                 return;
832
833         printk(KERN_ERR
834                "exiting task \"%s\" (%d) is an active IRQ thread (irq %d)\n",
835                tsk->comm ? tsk->comm : "", tsk->pid, tsk->irqaction->irq);
836
837         desc = irq_to_desc(tsk->irqaction->irq);
838
839         /*
840          * Prevent a stale desc->threads_oneshot. Must be called
841          * before setting the IRQTF_DIED flag.
842          */
843         irq_finalize_oneshot(desc, tsk->irqaction, true);
844
845         /*
846          * Set the THREAD DIED flag to prevent further wakeups of the
847          * soon to be gone threaded handler.
848          */
849         set_bit(IRQTF_DIED, &tsk->irqaction->flags);
850 }
851
852 static void irq_setup_forced_threading(struct irqaction *new)
853 {
854         if (!force_irqthreads)
855                 return;
856         if (new->flags & (IRQF_NO_THREAD | IRQF_PERCPU | IRQF_ONESHOT))
857                 return;
858
859         new->flags |= IRQF_ONESHOT;
860
861         if (!new->thread_fn) {
862                 set_bit(IRQTF_FORCED_THREAD, &new->thread_flags);
863                 new->thread_fn = new->handler;
864                 new->handler = irq_default_primary_handler;
865         }
866 }
867
868 /*
869  * Internal function to register an irqaction - typically used to
870  * allocate special interrupts that are part of the architecture.
871  */
872 static int
873 __setup_irq(unsigned int irq, struct irq_desc *desc, struct irqaction *new)
874 {
875         struct irqaction *old, **old_ptr;
876         const char *old_name = NULL;
877         unsigned long flags, thread_mask = 0;
878         int ret, nested, shared = 0;
879         cpumask_var_t mask;
880
881         if (!desc)
882                 return -EINVAL;
883
884         if (desc->irq_data.chip == &no_irq_chip)
885                 return -ENOSYS;
886         /*
887          * Some drivers like serial.c use request_irq() heavily,
888          * so we have to be careful not to interfere with a
889          * running system.
890          */
891         if (new->flags & IRQF_SAMPLE_RANDOM) {
892                 /*
893                  * This function might sleep, we want to call it first,
894                  * outside of the atomic block.
895                  * Yes, this might clear the entropy pool if the wrong
896                  * driver is attempted to be loaded, without actually
897                  * installing a new handler, but is this really a problem,
898                  * only the sysadmin is able to do this.
899                  */
900                 rand_initialize_irq(irq);
901         }
902
903         /*
904          * Check whether the interrupt nests into another interrupt
905          * thread.
906          */
907         nested = irq_settings_is_nested_thread(desc);
908         if (nested) {
909                 if (!new->thread_fn)
910                         return -EINVAL;
911                 /*
912                  * Replace the primary handler which was provided from
913                  * the driver for non nested interrupt handling by the
914                  * dummy function which warns when called.
915                  */
916                 new->handler = irq_nested_primary_handler;
917         } else {
918                 if (irq_settings_can_thread(desc))
919                         irq_setup_forced_threading(new);
920         }
921
922         /*
923          * Create a handler thread when a thread function is supplied
924          * and the interrupt does not nest into another interrupt
925          * thread.
926          */
927         if (new->thread_fn && !nested) {
928                 struct task_struct *t;
929
930                 t = kthread_create(irq_thread, new, "irq/%d-%s", irq,
931                                    new->name);
932                 if (IS_ERR(t))
933                         return PTR_ERR(t);
934                 /*
935                  * We keep the reference to the task struct even if
936                  * the thread dies to avoid that the interrupt code
937                  * references an already freed task_struct.
938                  */
939                 get_task_struct(t);
940                 new->thread = t;
941         }
942
943         if (!alloc_cpumask_var(&mask, GFP_KERNEL)) {
944                 ret = -ENOMEM;
945                 goto out_thread;
946         }
947
948         /*
949          * The following block of code has to be executed atomically
950          */
951         raw_spin_lock_irqsave(&desc->lock, flags);
952         old_ptr = &desc->action;
953         old = *old_ptr;
954         if (old) {
955                 /*
956                  * Can't share interrupts unless both agree to and are
957                  * the same type (level, edge, polarity). So both flag
958                  * fields must have IRQF_SHARED set and the bits which
959                  * set the trigger type must match. Also all must
960                  * agree on ONESHOT.
961                  */
962                 if (!((old->flags & new->flags) & IRQF_SHARED) ||
963                     ((old->flags ^ new->flags) & IRQF_TRIGGER_MASK) ||
964                     ((old->flags ^ new->flags) & IRQF_ONESHOT)) {
965                         old_name = old->name;
966                         goto mismatch;
967                 }
968
969                 /* All handlers must agree on per-cpuness */
970                 if ((old->flags & IRQF_PERCPU) !=
971                     (new->flags & IRQF_PERCPU))
972                         goto mismatch;
973
974                 /* add new interrupt at end of irq queue */
975                 do {
976                         thread_mask |= old->thread_mask;
977                         old_ptr = &old->next;
978                         old = *old_ptr;
979                 } while (old);
980                 shared = 1;
981         }
982
983         /*
984          * Setup the thread mask for this irqaction. Unlikely to have
985          * 32 resp 64 irqs sharing one line, but who knows.
986          */
987         if (new->flags & IRQF_ONESHOT && thread_mask == ~0UL) {
988                 ret = -EBUSY;
989                 goto out_mask;
990         }
991         new->thread_mask = 1 << ffz(thread_mask);
992
993         if (!shared) {
994                 init_waitqueue_head(&desc->wait_for_threads);
995
996                 /* Setup the type (level, edge polarity) if configured: */
997                 if (new->flags & IRQF_TRIGGER_MASK) {
998                         ret = __irq_set_trigger(desc, irq,
999                                         new->flags & IRQF_TRIGGER_MASK);
1000
1001                         if (ret)
1002                                 goto out_mask;
1003                 }
1004
1005                 desc->istate &= ~(IRQS_AUTODETECT | IRQS_SPURIOUS_DISABLED | \
1006                                   IRQS_ONESHOT | IRQS_WAITING);
1007                 irqd_clear(&desc->irq_data, IRQD_IRQ_INPROGRESS);
1008
1009                 if (new->flags & IRQF_PERCPU) {
1010                         irqd_set(&desc->irq_data, IRQD_PER_CPU);
1011                         irq_settings_set_per_cpu(desc);
1012                 }
1013
1014                 if (new->flags & IRQF_ONESHOT)
1015                         desc->istate |= IRQS_ONESHOT;
1016
1017                 if (irq_settings_can_autoenable(desc))
1018                         irq_startup(desc);
1019                 else
1020                         /* Undo nested disables: */
1021                         desc->depth = 1;
1022
1023                 /* Exclude IRQ from balancing if requested */
1024                 if (new->flags & IRQF_NOBALANCING) {
1025                         irq_settings_set_no_balancing(desc);
1026                         irqd_set(&desc->irq_data, IRQD_NO_BALANCING);
1027                 }
1028
1029                 /* Set default affinity mask once everything is setup */
1030                 setup_affinity(irq, desc, mask);
1031
1032         } else if (new->flags & IRQF_TRIGGER_MASK) {
1033                 unsigned int nmsk = new->flags & IRQF_TRIGGER_MASK;
1034                 unsigned int omsk = irq_settings_get_trigger_mask(desc);
1035
1036                 if (nmsk != omsk)
1037                         /* hope the handler works with current  trigger mode */
1038                         pr_warning("IRQ %d uses trigger mode %u; requested %u\n",
1039                                    irq, nmsk, omsk);
1040         }
1041
1042         new->irq = irq;
1043         *old_ptr = new;
1044
1045         /* Reset broken irq detection when installing new handler */
1046         desc->irq_count = 0;
1047         desc->irqs_unhandled = 0;
1048
1049         /*
1050          * Check whether we disabled the irq via the spurious handler
1051          * before. Reenable it and give it another chance.
1052          */
1053         if (shared && (desc->istate & IRQS_SPURIOUS_DISABLED)) {
1054                 desc->istate &= ~IRQS_SPURIOUS_DISABLED;
1055                 __enable_irq(desc, irq, false);
1056         }
1057
1058         raw_spin_unlock_irqrestore(&desc->lock, flags);
1059
1060         /*
1061          * Strictly no need to wake it up, but hung_task complains
1062          * when no hard interrupt wakes the thread up.
1063          */
1064         if (new->thread)
1065                 wake_up_process(new->thread);
1066
1067         register_irq_proc(irq, desc);
1068         new->dir = NULL;
1069         register_handler_proc(irq, new);
1070         free_cpumask_var(mask);
1071
1072         return 0;
1073
1074 mismatch:
1075 #ifdef CONFIG_DEBUG_SHIRQ
1076         if (!(new->flags & IRQF_PROBE_SHARED)) {
1077                 printk(KERN_ERR "IRQ handler type mismatch for IRQ %d\n", irq);
1078                 if (old_name)
1079                         printk(KERN_ERR "current handler: %s\n", old_name);
1080                 dump_stack();
1081         }
1082 #endif
1083         ret = -EBUSY;
1084
1085 out_mask:
1086         raw_spin_unlock_irqrestore(&desc->lock, flags);
1087         free_cpumask_var(mask);
1088
1089 out_thread:
1090         if (new->thread) {
1091                 struct task_struct *t = new->thread;
1092
1093                 new->thread = NULL;
1094                 if (likely(!test_bit(IRQTF_DIED, &new->thread_flags)))
1095                         kthread_stop(t);
1096                 put_task_struct(t);
1097         }
1098         return ret;
1099 }
1100
1101 /**
1102  *      setup_irq - setup an interrupt
1103  *      @irq: Interrupt line to setup
1104  *      @act: irqaction for the interrupt
1105  *
1106  * Used to statically setup interrupts in the early boot process.
1107  */
1108 int setup_irq(unsigned int irq, struct irqaction *act)
1109 {
1110         int retval;
1111         struct irq_desc *desc = irq_to_desc(irq);
1112
1113         chip_bus_lock(desc);
1114         retval = __setup_irq(irq, desc, act);
1115         chip_bus_sync_unlock(desc);
1116
1117         return retval;
1118 }
1119 EXPORT_SYMBOL_GPL(setup_irq);
1120
1121  /*
1122  * Internal function to unregister an irqaction - used to free
1123  * regular and special interrupts that are part of the architecture.
1124  */
1125 static struct irqaction *__free_irq(unsigned int irq, void *dev_id)
1126 {
1127         struct irq_desc *desc = irq_to_desc(irq);
1128         struct irqaction *action, **action_ptr;
1129         unsigned long flags;
1130
1131         WARN(in_interrupt(), "Trying to free IRQ %d from IRQ context!\n", irq);
1132
1133         if (!desc)
1134                 return NULL;
1135
1136         raw_spin_lock_irqsave(&desc->lock, flags);
1137
1138         /*
1139          * There can be multiple actions per IRQ descriptor, find the right
1140          * one based on the dev_id:
1141          */
1142         action_ptr = &desc->action;
1143         for (;;) {
1144                 action = *action_ptr;
1145
1146                 if (!action) {
1147                         WARN(1, "Trying to free already-free IRQ %d\n", irq);
1148                         raw_spin_unlock_irqrestore(&desc->lock, flags);
1149
1150                         return NULL;
1151                 }
1152
1153                 if (action->dev_id == dev_id)
1154                         break;
1155                 action_ptr = &action->next;
1156         }
1157
1158         /* Found it - now remove it from the list of entries: */
1159         *action_ptr = action->next;
1160
1161         /* Currently used only by UML, might disappear one day: */
1162 #ifdef CONFIG_IRQ_RELEASE_METHOD
1163         if (desc->irq_data.chip->release)
1164                 desc->irq_data.chip->release(irq, dev_id);
1165 #endif
1166
1167         /* If this was the last handler, shut down the IRQ line: */
1168         if (!desc->action)
1169                 irq_shutdown(desc);
1170
1171 #ifdef CONFIG_SMP
1172         /* make sure affinity_hint is cleaned up */
1173         if (WARN_ON_ONCE(desc->affinity_hint))
1174                 desc->affinity_hint = NULL;
1175 #endif
1176
1177         raw_spin_unlock_irqrestore(&desc->lock, flags);
1178
1179         unregister_handler_proc(irq, action);
1180
1181         /* Make sure it's not being used on another CPU: */
1182         synchronize_irq(irq);
1183
1184 #ifdef CONFIG_DEBUG_SHIRQ
1185         /*
1186          * It's a shared IRQ -- the driver ought to be prepared for an IRQ
1187          * event to happen even now it's being freed, so let's make sure that
1188          * is so by doing an extra call to the handler ....
1189          *
1190          * ( We do this after actually deregistering it, to make sure that a
1191          *   'real' IRQ doesn't run in * parallel with our fake. )
1192          */
1193         if (action->flags & IRQF_SHARED) {
1194                 local_irq_save(flags);
1195                 action->handler(irq, dev_id);
1196                 local_irq_restore(flags);
1197         }
1198 #endif
1199
1200         if (action->thread) {
1201                 if (!test_bit(IRQTF_DIED, &action->thread_flags))
1202                         kthread_stop(action->thread);
1203                 put_task_struct(action->thread);
1204         }
1205
1206         return action;
1207 }
1208
1209 /**
1210  *      remove_irq - free an interrupt
1211  *      @irq: Interrupt line to free
1212  *      @act: irqaction for the interrupt
1213  *
1214  * Used to remove interrupts statically setup by the early boot process.
1215  */
1216 void remove_irq(unsigned int irq, struct irqaction *act)
1217 {
1218         __free_irq(irq, act->dev_id);
1219 }
1220 EXPORT_SYMBOL_GPL(remove_irq);
1221
1222 /**
1223  *      free_irq - free an interrupt allocated with request_irq
1224  *      @irq: Interrupt line to free
1225  *      @dev_id: Device identity to free
1226  *
1227  *      Remove an interrupt handler. The handler is removed and if the
1228  *      interrupt line is no longer in use by any driver it is disabled.
1229  *      On a shared IRQ the caller must ensure the interrupt is disabled
1230  *      on the card it drives before calling this function. The function
1231  *      does not return until any executing interrupts for this IRQ
1232  *      have completed.
1233  *
1234  *      This function must not be called from interrupt context.
1235  */
1236 void free_irq(unsigned int irq, void *dev_id)
1237 {
1238         struct irq_desc *desc = irq_to_desc(irq);
1239
1240         if (!desc)
1241                 return;
1242
1243 #ifdef CONFIG_SMP
1244         if (WARN_ON(desc->affinity_notify))
1245                 desc->affinity_notify = NULL;
1246 #endif
1247
1248         chip_bus_lock(desc);
1249         kfree(__free_irq(irq, dev_id));
1250         chip_bus_sync_unlock(desc);
1251 }
1252 EXPORT_SYMBOL(free_irq);
1253
1254 /**
1255  *      request_threaded_irq - allocate an interrupt line
1256  *      @irq: Interrupt line to allocate
1257  *      @handler: Function to be called when the IRQ occurs.
1258  *                Primary handler for threaded interrupts
1259  *                If NULL and thread_fn != NULL the default
1260  *                primary handler is installed
1261  *      @thread_fn: Function called from the irq handler thread
1262  *                  If NULL, no irq thread is created
1263  *      @irqflags: Interrupt type flags
1264  *      @devname: An ascii name for the claiming device
1265  *      @dev_id: A cookie passed back to the handler function
1266  *
1267  *      This call allocates interrupt resources and enables the
1268  *      interrupt line and IRQ handling. From the point this
1269  *      call is made your handler function may be invoked. Since
1270  *      your handler function must clear any interrupt the board
1271  *      raises, you must take care both to initialise your hardware
1272  *      and to set up the interrupt handler in the right order.
1273  *
1274  *      If you want to set up a threaded irq handler for your device
1275  *      then you need to supply @handler and @thread_fn. @handler ist
1276  *      still called in hard interrupt context and has to check
1277  *      whether the interrupt originates from the device. If yes it
1278  *      needs to disable the interrupt on the device and return
1279  *      IRQ_WAKE_THREAD which will wake up the handler thread and run
1280  *      @thread_fn. This split handler design is necessary to support
1281  *      shared interrupts.
1282  *
1283  *      Dev_id must be globally unique. Normally the address of the
1284  *      device data structure is used as the cookie. Since the handler
1285  *      receives this value it makes sense to use it.
1286  *
1287  *      If your interrupt is shared you must pass a non NULL dev_id
1288  *      as this is required when freeing the interrupt.
1289  *
1290  *      Flags:
1291  *
1292  *      IRQF_SHARED             Interrupt is shared
1293  *      IRQF_SAMPLE_RANDOM      The interrupt can be used for entropy
1294  *      IRQF_TRIGGER_*          Specify active edge(s) or level
1295  *
1296  */
1297 int request_threaded_irq(unsigned int irq, irq_handler_t handler,
1298                          irq_handler_t thread_fn, unsigned long irqflags,
1299                          const char *devname, void *dev_id)
1300 {
1301         struct irqaction *action;
1302         struct irq_desc *desc;
1303         int retval;
1304
1305         /*
1306          * Sanity-check: shared interrupts must pass in a real dev-ID,
1307          * otherwise we'll have trouble later trying to figure out
1308          * which interrupt is which (messes up the interrupt freeing
1309          * logic etc).
1310          */
1311         if ((irqflags & IRQF_SHARED) && !dev_id)
1312                 return -EINVAL;
1313
1314         desc = irq_to_desc(irq);
1315         if (!desc)
1316                 return -EINVAL;
1317
1318         if (!irq_settings_can_request(desc))
1319                 return -EINVAL;
1320
1321         if (!handler) {
1322                 if (!thread_fn)
1323                         return -EINVAL;
1324                 handler = irq_default_primary_handler;
1325         }
1326
1327         action = kzalloc(sizeof(struct irqaction), GFP_KERNEL);
1328         if (!action)
1329                 return -ENOMEM;
1330
1331         action->handler = handler;
1332         action->thread_fn = thread_fn;
1333         action->flags = irqflags;
1334         action->name = devname;
1335         action->dev_id = dev_id;
1336
1337         chip_bus_lock(desc);
1338         retval = __setup_irq(irq, desc, action);
1339         chip_bus_sync_unlock(desc);
1340
1341         if (retval)
1342                 kfree(action);
1343
1344 #ifdef CONFIG_DEBUG_SHIRQ_FIXME
1345         if (!retval && (irqflags & IRQF_SHARED)) {
1346                 /*
1347                  * It's a shared IRQ -- the driver ought to be prepared for it
1348                  * to happen immediately, so let's make sure....
1349                  * We disable the irq to make sure that a 'real' IRQ doesn't
1350                  * run in parallel with our fake.
1351                  */
1352                 unsigned long flags;
1353
1354                 disable_irq(irq);
1355                 local_irq_save(flags);
1356
1357                 handler(irq, dev_id);
1358
1359                 local_irq_restore(flags);
1360                 enable_irq(irq);
1361         }
1362 #endif
1363         return retval;
1364 }
1365 EXPORT_SYMBOL(request_threaded_irq);
1366
1367 /**
1368  *      request_any_context_irq - allocate an interrupt line
1369  *      @irq: Interrupt line to allocate
1370  *      @handler: Function to be called when the IRQ occurs.
1371  *                Threaded handler for threaded interrupts.
1372  *      @flags: Interrupt type flags
1373  *      @name: An ascii name for the claiming device
1374  *      @dev_id: A cookie passed back to the handler function
1375  *
1376  *      This call allocates interrupt resources and enables the
1377  *      interrupt line and IRQ handling. It selects either a
1378  *      hardirq or threaded handling method depending on the
1379  *      context.
1380  *
1381  *      On failure, it returns a negative value. On success,
1382  *      it returns either IRQC_IS_HARDIRQ or IRQC_IS_NESTED.
1383  */
1384 int request_any_context_irq(unsigned int irq, irq_handler_t handler,
1385                             unsigned long flags, const char *name, void *dev_id)
1386 {
1387         struct irq_desc *desc = irq_to_desc(irq);
1388         int ret;
1389
1390         if (!desc)
1391                 return -EINVAL;
1392
1393         if (irq_settings_is_nested_thread(desc)) {
1394                 ret = request_threaded_irq(irq, NULL, handler,
1395                                            flags, name, dev_id);
1396                 return !ret ? IRQC_IS_NESTED : ret;
1397         }
1398
1399         ret = request_irq(irq, handler, flags, name, dev_id);
1400         return !ret ? IRQC_IS_HARDIRQ : ret;
1401 }
1402 EXPORT_SYMBOL_GPL(request_any_context_irq);