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