Merge branch 'for-linus' of git://neil.brown.name/md
[sfrench/cifs-2.6.git] / kernel / irq / handle.c
1 /*
2  * linux/kernel/irq/handle.c
3  *
4  * Copyright (C) 1992, 1998-2006 Linus Torvalds, Ingo Molnar
5  * Copyright (C) 2005-2006, Thomas Gleixner, Russell King
6  *
7  * This file contains the core interrupt handling code.
8  *
9  * Detailed information is available in Documentation/DocBook/genericirq
10  *
11  */
12
13 #include <linux/irq.h>
14 #include <linux/module.h>
15 #include <linux/random.h>
16 #include <linux/interrupt.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/rculist.h>
19 #include <linux/hash.h>
20 #include <trace/irq.h>
21 #include <linux/bootmem.h>
22
23 #include "internals.h"
24
25 /*
26  * lockdep: we want to handle all irq_desc locks as a single lock-class:
27  */
28 struct lock_class_key irq_desc_lock_class;
29
30 /**
31  * handle_bad_irq - handle spurious and unhandled irqs
32  * @irq:       the interrupt number
33  * @desc:      description of the interrupt
34  *
35  * Handles spurious and unhandled IRQ's. It also prints a debugmessage.
36  */
37 void handle_bad_irq(unsigned int irq, struct irq_desc *desc)
38 {
39         print_irq_desc(irq, desc);
40         kstat_incr_irqs_this_cpu(irq, desc);
41         ack_bad_irq(irq);
42 }
43
44 #if defined(CONFIG_SMP) && defined(CONFIG_GENERIC_HARDIRQS)
45 static void __init init_irq_default_affinity(void)
46 {
47         alloc_bootmem_cpumask_var(&irq_default_affinity);
48         cpumask_setall(irq_default_affinity);
49 }
50 #else
51 static void __init init_irq_default_affinity(void)
52 {
53 }
54 #endif
55
56 /*
57  * Linux has a controller-independent interrupt architecture.
58  * Every controller has a 'controller-template', that is used
59  * by the main code to do the right thing. Each driver-visible
60  * interrupt source is transparently wired to the appropriate
61  * controller. Thus drivers need not be aware of the
62  * interrupt-controller.
63  *
64  * The code is designed to be easily extended with new/different
65  * interrupt controllers, without having to do assembly magic or
66  * having to touch the generic code.
67  *
68  * Controller mappings for all interrupt sources:
69  */
70 int nr_irqs = NR_IRQS;
71 EXPORT_SYMBOL_GPL(nr_irqs);
72
73 #ifdef CONFIG_SPARSE_IRQ
74
75 static struct irq_desc irq_desc_init = {
76         .irq        = -1,
77         .status     = IRQ_DISABLED,
78         .chip       = &no_irq_chip,
79         .handle_irq = handle_bad_irq,
80         .depth      = 1,
81         .lock       = __SPIN_LOCK_UNLOCKED(irq_desc_init.lock),
82 };
83
84 void init_kstat_irqs(struct irq_desc *desc, int cpu, int nr)
85 {
86         int node;
87         void *ptr;
88
89         node = cpu_to_node(cpu);
90         ptr = kzalloc_node(nr * sizeof(*desc->kstat_irqs), GFP_ATOMIC, node);
91
92         /*
93          * don't overwite if can not get new one
94          * init_copy_kstat_irqs() could still use old one
95          */
96         if (ptr) {
97                 printk(KERN_DEBUG "  alloc kstat_irqs on cpu %d node %d\n",
98                          cpu, node);
99                 desc->kstat_irqs = ptr;
100         }
101 }
102
103 static void init_one_irq_desc(int irq, struct irq_desc *desc, int cpu)
104 {
105         memcpy(desc, &irq_desc_init, sizeof(struct irq_desc));
106
107         spin_lock_init(&desc->lock);
108         desc->irq = irq;
109 #ifdef CONFIG_SMP
110         desc->cpu = cpu;
111 #endif
112         lockdep_set_class(&desc->lock, &irq_desc_lock_class);
113         init_kstat_irqs(desc, cpu, nr_cpu_ids);
114         if (!desc->kstat_irqs) {
115                 printk(KERN_ERR "can not alloc kstat_irqs\n");
116                 BUG_ON(1);
117         }
118         if (!init_alloc_desc_masks(desc, cpu, false)) {
119                 printk(KERN_ERR "can not alloc irq_desc cpumasks\n");
120                 BUG_ON(1);
121         }
122         arch_init_chip_data(desc, cpu);
123 }
124
125 /*
126  * Protect the sparse_irqs:
127  */
128 DEFINE_SPINLOCK(sparse_irq_lock);
129
130 struct irq_desc **irq_desc_ptrs __read_mostly;
131
132 static struct irq_desc irq_desc_legacy[NR_IRQS_LEGACY] __cacheline_aligned_in_smp = {
133         [0 ... NR_IRQS_LEGACY-1] = {
134                 .irq        = -1,
135                 .status     = IRQ_DISABLED,
136                 .chip       = &no_irq_chip,
137                 .handle_irq = handle_bad_irq,
138                 .depth      = 1,
139                 .lock       = __SPIN_LOCK_UNLOCKED(irq_desc_init.lock),
140         }
141 };
142
143 static unsigned int *kstat_irqs_legacy;
144
145 int __init early_irq_init(void)
146 {
147         struct irq_desc *desc;
148         int legacy_count;
149         int i;
150
151         init_irq_default_affinity();
152
153          /* initialize nr_irqs based on nr_cpu_ids */
154         arch_probe_nr_irqs();
155         printk(KERN_INFO "NR_IRQS:%d nr_irqs:%d\n", NR_IRQS, nr_irqs);
156
157         desc = irq_desc_legacy;
158         legacy_count = ARRAY_SIZE(irq_desc_legacy);
159
160         /* allocate irq_desc_ptrs array based on nr_irqs */
161         irq_desc_ptrs = alloc_bootmem(nr_irqs * sizeof(void *));
162
163         /* allocate based on nr_cpu_ids */
164         /* FIXME: invert kstat_irgs, and it'd be a per_cpu_alloc'd thing */
165         kstat_irqs_legacy = alloc_bootmem(NR_IRQS_LEGACY * nr_cpu_ids *
166                                           sizeof(int));
167
168         for (i = 0; i < legacy_count; i++) {
169                 desc[i].irq = i;
170                 desc[i].kstat_irqs = kstat_irqs_legacy + i * nr_cpu_ids;
171                 lockdep_set_class(&desc[i].lock, &irq_desc_lock_class);
172                 init_alloc_desc_masks(&desc[i], 0, true);
173                 irq_desc_ptrs[i] = desc + i;
174         }
175
176         for (i = legacy_count; i < nr_irqs; i++)
177                 irq_desc_ptrs[i] = NULL;
178
179         return arch_early_irq_init();
180 }
181
182 struct irq_desc *irq_to_desc(unsigned int irq)
183 {
184         if (irq_desc_ptrs && irq < nr_irqs)
185                 return irq_desc_ptrs[irq];
186
187         return NULL;
188 }
189
190 struct irq_desc *irq_to_desc_alloc_cpu(unsigned int irq, int cpu)
191 {
192         struct irq_desc *desc;
193         unsigned long flags;
194         int node;
195
196         if (irq >= nr_irqs) {
197                 WARN(1, "irq (%d) >= nr_irqs (%d) in irq_to_desc_alloc\n",
198                         irq, nr_irqs);
199                 return NULL;
200         }
201
202         desc = irq_desc_ptrs[irq];
203         if (desc)
204                 return desc;
205
206         spin_lock_irqsave(&sparse_irq_lock, flags);
207
208         /* We have to check it to avoid races with another CPU */
209         desc = irq_desc_ptrs[irq];
210         if (desc)
211                 goto out_unlock;
212
213         node = cpu_to_node(cpu);
214         desc = kzalloc_node(sizeof(*desc), GFP_ATOMIC, node);
215         printk(KERN_DEBUG "  alloc irq_desc for %d on cpu %d node %d\n",
216                  irq, cpu, node);
217         if (!desc) {
218                 printk(KERN_ERR "can not alloc irq_desc\n");
219                 BUG_ON(1);
220         }
221         init_one_irq_desc(irq, desc, cpu);
222
223         irq_desc_ptrs[irq] = desc;
224
225 out_unlock:
226         spin_unlock_irqrestore(&sparse_irq_lock, flags);
227
228         return desc;
229 }
230
231 #else /* !CONFIG_SPARSE_IRQ */
232
233 struct irq_desc irq_desc[NR_IRQS] __cacheline_aligned_in_smp = {
234         [0 ... NR_IRQS-1] = {
235                 .status = IRQ_DISABLED,
236                 .chip = &no_irq_chip,
237                 .handle_irq = handle_bad_irq,
238                 .depth = 1,
239                 .lock = __SPIN_LOCK_UNLOCKED(irq_desc->lock),
240         }
241 };
242
243 static unsigned int kstat_irqs_all[NR_IRQS][NR_CPUS];
244 int __init early_irq_init(void)
245 {
246         struct irq_desc *desc;
247         int count;
248         int i;
249
250         init_irq_default_affinity();
251
252         printk(KERN_INFO "NR_IRQS:%d\n", NR_IRQS);
253
254         desc = irq_desc;
255         count = ARRAY_SIZE(irq_desc);
256
257         for (i = 0; i < count; i++) {
258                 desc[i].irq = i;
259                 init_alloc_desc_masks(&desc[i], 0, true);
260                 desc[i].kstat_irqs = kstat_irqs_all[i];
261         }
262         return arch_early_irq_init();
263 }
264
265 struct irq_desc *irq_to_desc(unsigned int irq)
266 {
267         return (irq < NR_IRQS) ? irq_desc + irq : NULL;
268 }
269
270 struct irq_desc *irq_to_desc_alloc_cpu(unsigned int irq, int cpu)
271 {
272         return irq_to_desc(irq);
273 }
274 #endif /* !CONFIG_SPARSE_IRQ */
275
276 void clear_kstat_irqs(struct irq_desc *desc)
277 {
278         memset(desc->kstat_irqs, 0, nr_cpu_ids * sizeof(*(desc->kstat_irqs)));
279 }
280
281 /*
282  * What should we do if we get a hw irq event on an illegal vector?
283  * Each architecture has to answer this themself.
284  */
285 static void ack_bad(unsigned int irq)
286 {
287         struct irq_desc *desc = irq_to_desc(irq);
288
289         print_irq_desc(irq, desc);
290         ack_bad_irq(irq);
291 }
292
293 /*
294  * NOP functions
295  */
296 static void noop(unsigned int irq)
297 {
298 }
299
300 static unsigned int noop_ret(unsigned int irq)
301 {
302         return 0;
303 }
304
305 /*
306  * Generic no controller implementation
307  */
308 struct irq_chip no_irq_chip = {
309         .name           = "none",
310         .startup        = noop_ret,
311         .shutdown       = noop,
312         .enable         = noop,
313         .disable        = noop,
314         .ack            = ack_bad,
315         .end            = noop,
316 };
317
318 /*
319  * Generic dummy implementation which can be used for
320  * real dumb interrupt sources
321  */
322 struct irq_chip dummy_irq_chip = {
323         .name           = "dummy",
324         .startup        = noop_ret,
325         .shutdown       = noop,
326         .enable         = noop,
327         .disable        = noop,
328         .ack            = noop,
329         .mask           = noop,
330         .unmask         = noop,
331         .end            = noop,
332 };
333
334 /*
335  * Special, empty irq handler:
336  */
337 irqreturn_t no_action(int cpl, void *dev_id)
338 {
339         return IRQ_NONE;
340 }
341
342 DEFINE_TRACE(irq_handler_entry);
343 DEFINE_TRACE(irq_handler_exit);
344
345 /**
346  * handle_IRQ_event - irq action chain handler
347  * @irq:        the interrupt number
348  * @action:     the interrupt action chain for this irq
349  *
350  * Handles the action chain of an irq event
351  */
352 irqreturn_t handle_IRQ_event(unsigned int irq, struct irqaction *action)
353 {
354         irqreturn_t ret, retval = IRQ_NONE;
355         unsigned int status = 0;
356
357         WARN_ONCE(!in_irq(), "BUG: IRQ handler called from non-hardirq context!");
358
359         if (!(action->flags & IRQF_DISABLED))
360                 local_irq_enable_in_hardirq();
361
362         do {
363                 trace_irq_handler_entry(irq, action);
364                 ret = action->handler(irq, action->dev_id);
365                 trace_irq_handler_exit(irq, action, ret);
366                 if (ret == IRQ_HANDLED)
367                         status |= action->flags;
368                 retval |= ret;
369                 action = action->next;
370         } while (action);
371
372         if (status & IRQF_SAMPLE_RANDOM)
373                 add_interrupt_randomness(irq);
374         local_irq_disable();
375
376         return retval;
377 }
378
379 #ifndef CONFIG_GENERIC_HARDIRQS_NO__DO_IRQ
380
381 #ifdef CONFIG_ENABLE_WARN_DEPRECATED
382 # warning __do_IRQ is deprecated. Please convert to proper flow handlers
383 #endif
384
385 /**
386  * __do_IRQ - original all in one highlevel IRQ handler
387  * @irq:        the interrupt number
388  *
389  * __do_IRQ handles all normal device IRQ's (the special
390  * SMP cross-CPU interrupts have their own specific
391  * handlers).
392  *
393  * This is the original x86 implementation which is used for every
394  * interrupt type.
395  */
396 unsigned int __do_IRQ(unsigned int irq)
397 {
398         struct irq_desc *desc = irq_to_desc(irq);
399         struct irqaction *action;
400         unsigned int status;
401
402         kstat_incr_irqs_this_cpu(irq, desc);
403
404         if (CHECK_IRQ_PER_CPU(desc->status)) {
405                 irqreturn_t action_ret;
406
407                 /*
408                  * No locking required for CPU-local interrupts:
409                  */
410                 if (desc->chip->ack) {
411                         desc->chip->ack(irq);
412                         /* get new one */
413                         desc = irq_remap_to_desc(irq, desc);
414                 }
415                 if (likely(!(desc->status & IRQ_DISABLED))) {
416                         action_ret = handle_IRQ_event(irq, desc->action);
417                         if (!noirqdebug)
418                                 note_interrupt(irq, desc, action_ret);
419                 }
420                 desc->chip->end(irq);
421                 return 1;
422         }
423
424         spin_lock(&desc->lock);
425         if (desc->chip->ack) {
426                 desc->chip->ack(irq);
427                 desc = irq_remap_to_desc(irq, desc);
428         }
429         /*
430          * REPLAY is when Linux resends an IRQ that was dropped earlier
431          * WAITING is used by probe to mark irqs that are being tested
432          */
433         status = desc->status & ~(IRQ_REPLAY | IRQ_WAITING);
434         status |= IRQ_PENDING; /* we _want_ to handle it */
435
436         /*
437          * If the IRQ is disabled for whatever reason, we cannot
438          * use the action we have.
439          */
440         action = NULL;
441         if (likely(!(status & (IRQ_DISABLED | IRQ_INPROGRESS)))) {
442                 action = desc->action;
443                 status &= ~IRQ_PENDING; /* we commit to handling */
444                 status |= IRQ_INPROGRESS; /* we are handling it */
445         }
446         desc->status = status;
447
448         /*
449          * If there is no IRQ handler or it was disabled, exit early.
450          * Since we set PENDING, if another processor is handling
451          * a different instance of this same irq, the other processor
452          * will take care of it.
453          */
454         if (unlikely(!action))
455                 goto out;
456
457         /*
458          * Edge triggered interrupts need to remember
459          * pending events.
460          * This applies to any hw interrupts that allow a second
461          * instance of the same irq to arrive while we are in do_IRQ
462          * or in the handler. But the code here only handles the _second_
463          * instance of the irq, not the third or fourth. So it is mostly
464          * useful for irq hardware that does not mask cleanly in an
465          * SMP environment.
466          */
467         for (;;) {
468                 irqreturn_t action_ret;
469
470                 spin_unlock(&desc->lock);
471
472                 action_ret = handle_IRQ_event(irq, action);
473                 if (!noirqdebug)
474                         note_interrupt(irq, desc, action_ret);
475
476                 spin_lock(&desc->lock);
477                 if (likely(!(desc->status & IRQ_PENDING)))
478                         break;
479                 desc->status &= ~IRQ_PENDING;
480         }
481         desc->status &= ~IRQ_INPROGRESS;
482
483 out:
484         /*
485          * The ->end() handler has to deal with interrupts which got
486          * disabled while the handler was running.
487          */
488         desc->chip->end(irq);
489         spin_unlock(&desc->lock);
490
491         return 1;
492 }
493 #endif
494
495 void early_init_irq_lock_class(void)
496 {
497         struct irq_desc *desc;
498         int i;
499
500         for_each_irq_desc(i, desc) {
501                 lockdep_set_class(&desc->lock, &irq_desc_lock_class);
502         }
503 }
504
505 unsigned int kstat_irqs_cpu(unsigned int irq, int cpu)
506 {
507         struct irq_desc *desc = irq_to_desc(irq);
508         return desc ? desc->kstat_irqs[cpu] : 0;
509 }
510 EXPORT_SYMBOL(kstat_irqs_cpu);
511