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 <linux/bootmem.h>
  21
  22 #include "internals.h"
  23
  24 /*
  25  * lockdep: we want to handle all irq_desc locks as a single lock-class:
  26  */
  27 struct lock_class_key irq_desc_lock_class;
  28
  29 /**
  30  * handle_bad_irq - handle spurious and unhandled irqs
  31  * @irq:       the interrupt number
  32  * @desc:      description of the interrupt
  33  *
  34  * Handles spurious and unhandled IRQ's. It also prints a debugmessage.
  35  */
  36 void handle_bad_irq(unsigned int irq, struct irq_desc *desc)
  37 {
  38         print_irq_desc(irq, desc);
  39         kstat_incr_irqs_this_cpu(irq, desc);
  40         ack_bad_irq(irq);
  41 }
  42
  43 #if defined(CONFIG_SMP) && defined(CONFIG_GENERIC_HARDIRQS)
  44 static void __init init_irq_default_affinity(void)
  45 {
  46         alloc_bootmem_cpumask_var(&irq_default_affinity);
  47         cpumask_setall(irq_default_affinity);
  48 }
  49 #else
  50 static void __init init_irq_default_affinity(void)
  51 {
  52 }
  53 #endif
  54
  55 /*
  56  * Linux has a controller-independent interrupt architecture.
  57  * Every controller has a 'controller-template', that is used
  58  * by the main code to do the right thing. Each driver-visible
  59  * interrupt source is transparently wired to the appropriate
  60  * controller. Thus drivers need not be aware of the
  61  * interrupt-controller.
  62  *
  63  * The code is designed to be easily extended with new/different
  64  * interrupt controllers, without having to do assembly magic or
  65  * having to touch the generic code.
  66  *
  67  * Controller mappings for all interrupt sources:
  68  */
  69 int nr_irqs = NR_IRQS;
  70 EXPORT_SYMBOL_GPL(nr_irqs);
  71
  72 #ifdef CONFIG_SPARSE_IRQ
  73
  74 static struct irq_desc irq_desc_init = {
  75         .irq        = -1,
  76         .status     = IRQ_DISABLED,
  77         .chip       = &no_irq_chip,
  78         .handle_irq = handle_bad_irq,
  79         .depth      = 1,
  80         .lock       = __SPIN_LOCK_UNLOCKED(irq_desc_init.lock),
  81 };
  82
  83 void init_kstat_irqs(struct irq_desc *desc, int cpu, int nr)
  84 {
  85         int node;
  86         void *ptr;
  87
  88         node = cpu_to_node(cpu);
  89         ptr = kzalloc_node(nr * sizeof(*desc->kstat_irqs), GFP_ATOMIC, node);
  90
  91         /*
  92          * don't overwite if can not get new one
  93          * init_copy_kstat_irqs() could still use old one
  94          */
  95         if (ptr) {
  96                 printk(KERN_DEBUG "  alloc kstat_irqs on cpu %d node %d\n",
  97                          cpu, node);
  98                 desc->kstat_irqs = ptr;
  99         }
 100 }
 101
 102 static void init_one_irq_desc(int irq, struct irq_desc *desc, int cpu)
 103 {
 104         memcpy(desc, &irq_desc_init, sizeof(struct irq_desc));
 105
 106         spin_lock_init(&desc->lock);
 107         desc->irq = irq;
 108 #ifdef CONFIG_SMP
 109         desc->cpu = cpu;
 110 #endif
 111         lockdep_set_class(&desc->lock, &irq_desc_lock_class);
 112         init_kstat_irqs(desc, cpu, nr_cpu_ids);
 113         if (!desc->kstat_irqs) {
 114                 printk(KERN_ERR "can not alloc kstat_irqs\n");
 115                 BUG_ON(1);
 116         }
 117         if (!init_alloc_desc_masks(desc, cpu, false)) {
 118                 printk(KERN_ERR "can not alloc irq_desc cpumasks\n");
 119                 BUG_ON(1);
 120         }
 121         arch_init_chip_data(desc, cpu);
 122 }
 123
 124 /*
 125  * Protect the sparse_irqs:
 126  */
 127 DEFINE_SPINLOCK(sparse_irq_lock);
 128
 129 struct irq_desc **irq_desc_ptrs __read_mostly;
 130
 131 static struct irq_desc irq_desc_legacy[NR_IRQS_LEGACY] __cacheline_aligned_in_smp = {
 132         [0 ... NR_IRQS_LEGACY-1] = {
 133                 .irq        = -1,
 134                 .status     = IRQ_DISABLED,
 135                 .chip       = &no_irq_chip,
 136                 .handle_irq = handle_bad_irq,
 137                 .depth      = 1,
 138                 .lock       = __SPIN_LOCK_UNLOCKED(irq_desc_init.lock),
 139         }
 140 };
 141
 142 static unsigned int *kstat_irqs_legacy;
 143
 144 int __init early_irq_init(void)
 145 {
 146         struct irq_desc *desc;
 147         int legacy_count;
 148         int i;
 149
 150         init_irq_default_affinity();
 151
 152          /* initialize nr_irqs based on nr_cpu_ids */
 153         arch_probe_nr_irqs();
 154         printk(KERN_INFO "NR_IRQS:%d nr_irqs:%d\n", NR_IRQS, nr_irqs);
 155
 156         desc = irq_desc_legacy;
 157         legacy_count = ARRAY_SIZE(irq_desc_legacy);
 158
 159         /* allocate irq_desc_ptrs array based on nr_irqs */
 160         irq_desc_ptrs = alloc_bootmem(nr_irqs * sizeof(void *));
 161
 162         /* allocate based on nr_cpu_ids */
 163         /* FIXME: invert kstat_irgs, and it'd be a per_cpu_alloc'd thing */
 164         kstat_irqs_legacy = alloc_bootmem(NR_IRQS_LEGACY * nr_cpu_ids *
 165                                           sizeof(int));
 166
 167         for (i = 0; i < legacy_count; i++) {
 168                 desc[i].irq = i;
 169                 desc[i].kstat_irqs = kstat_irqs_legacy + i * nr_cpu_ids;
 170                 lockdep_set_class(&desc[i].lock, &irq_desc_lock_class);
 171                 init_alloc_desc_masks(&desc[i], 0, true);
 172                 irq_desc_ptrs[i] = desc + i;
 173         }
 174
 175         for (i = legacy_count; i < nr_irqs; i++)
 176                 irq_desc_ptrs[i] = NULL;
 177
 178         return arch_early_irq_init();
 179 }
 180
 181 struct irq_desc *irq_to_desc(unsigned int irq)
 182 {
 183         if (irq_desc_ptrs && irq < nr_irqs)
 184                 return irq_desc_ptrs[irq];
 185
 186         return NULL;
 187 }
 188
 189 struct irq_desc *irq_to_desc_alloc_cpu(unsigned int irq, int cpu)
 190 {
 191         struct irq_desc *desc;
 192         unsigned long flags;
 193         int node;
 194
 195         if (irq >= nr_irqs) {
 196                 WARN(1, "irq (%d) >= nr_irqs (%d) in irq_to_desc_alloc\n",
 197                         irq, nr_irqs);
 198                 return NULL;
 199         }
 200
 201         desc = irq_desc_ptrs[irq];
 202         if (desc)
 203                 return desc;
 204
 205         spin_lock_irqsave(&sparse_irq_lock, flags);
 206
 207         /* We have to check it to avoid races with another CPU */
 208         desc = irq_desc_ptrs[irq];
 209         if (desc)
 210                 goto out_unlock;
 211
 212         node = cpu_to_node(cpu);
 213         desc = kzalloc_node(sizeof(*desc), GFP_ATOMIC, node);
 214         printk(KERN_DEBUG "  alloc irq_desc for %d on cpu %d node %d\n",
 215                  irq, cpu, node);
 216         if (!desc) {
 217                 printk(KERN_ERR "can not alloc irq_desc\n");
 218                 BUG_ON(1);
 219         }
 220         init_one_irq_desc(irq, desc, cpu);
 221
 222         irq_desc_ptrs[irq] = desc;
 223
 224 out_unlock:
 225         spin_unlock_irqrestore(&sparse_irq_lock, flags);
 226
 227         return desc;
 228 }
 229
 230 #else /* !CONFIG_SPARSE_IRQ */
 231
 232 struct irq_desc irq_desc[NR_IRQS] __cacheline_aligned_in_smp = {
 233         [0 ... NR_IRQS-1] = {
 234                 .status = IRQ_DISABLED,
 235                 .chip = &no_irq_chip,
 236                 .handle_irq = handle_bad_irq,
 237                 .depth = 1,
 238                 .lock = __SPIN_LOCK_UNLOCKED(irq_desc->lock),
 239         }
 240 };
 241
 242 static unsigned int kstat_irqs_all[NR_IRQS][NR_CPUS];
 243 int __init early_irq_init(void)
 244 {
 245         struct irq_desc *desc;
 246         int count;
 247         int i;
 248
 249         init_irq_default_affinity();
 250
 251         printk(KERN_INFO "NR_IRQS:%d\n", NR_IRQS);
 252
 253         desc = irq_desc;
 254         count = ARRAY_SIZE(irq_desc);
 255
 256         for (i = 0; i < count; i++) {
 257                 desc[i].irq = i;
 258                 init_alloc_desc_masks(&desc[i], 0, true);
 259                 desc[i].kstat_irqs = kstat_irqs_all[i];
 260         }
 261         return arch_early_irq_init();
 262 }
 263
 264 struct irq_desc *irq_to_desc(unsigned int irq)
 265 {
 266         return (irq < NR_IRQS) ? irq_desc + irq : NULL;
 267 }
 268
 269 struct irq_desc *irq_to_desc_alloc_cpu(unsigned int irq, int cpu)
 270 {
 271         return irq_to_desc(irq);
 272 }
 273 #endif /* !CONFIG_SPARSE_IRQ */
 274
 275 void clear_kstat_irqs(struct irq_desc *desc)
 276 {
 277         memset(desc->kstat_irqs, 0, nr_cpu_ids * sizeof(*(desc->kstat_irqs)));
 278 }
 279
 280 /*
 281  * What should we do if we get a hw irq event on an illegal vector?
 282  * Each architecture has to answer this themself.
 283  */
 284 static void ack_bad(unsigned int irq)
 285 {
 286         struct irq_desc *desc = irq_to_desc(irq);
 287
 288         print_irq_desc(irq, desc);
 289         ack_bad_irq(irq);
 290 }
 291
 292 /*
 293  * NOP functions
 294  */
 295 static void noop(unsigned int irq)
 296 {
 297 }
 298
 299 static unsigned int noop_ret(unsigned int irq)
 300 {
 301         return 0;
 302 }
 303
 304 /*
 305  * Generic no controller implementation
 306  */
 307 struct irq_chip no_irq_chip = {
 308         .name           = "none",
 309         .startup        = noop_ret,
 310         .shutdown       = noop,
 311         .enable         = noop,
 312         .disable        = noop,
 313         .ack            = ack_bad,
 314         .end            = noop,
 315 };
 316
 317 /*
 318  * Generic dummy implementation which can be used for
 319  * real dumb interrupt sources
 320  */
 321 struct irq_chip dummy_irq_chip = {
 322         .name           = "dummy",
 323         .startup        = noop_ret,
 324         .shutdown       = noop,
 325         .enable         = noop,
 326         .disable        = noop,
 327         .ack            = noop,
 328         .mask           = noop,
 329         .unmask         = noop,
 330         .end            = noop,
 331 };
 332
 333 /*
 334  * Special, empty irq handler:
 335  */
 336 irqreturn_t no_action(int cpl, void *dev_id)
 337 {
 338         return IRQ_NONE;
 339 }
 340
 341 static void warn_no_thread(unsigned int irq, struct irqaction *action)
 342 {
 343         if (test_and_set_bit(IRQTF_WARNED, &action->thread_flags))
 344                 return;
 345
 346         printk(KERN_WARNING "IRQ %d device %s returned IRQ_WAKE_THREAD "
 347                "but no thread function available.", irq, action->name);
 348 }
 349
 350 /**
 351  * handle_IRQ_event - irq action chain handler
 352  * @irq:        the interrupt number
 353  * @action:     the interrupt action chain for this irq
 354  *
 355  * Handles the action chain of an irq event
 356  */
 357 irqreturn_t handle_IRQ_event(unsigned int irq, struct irqaction *action)
 358 {
 359         irqreturn_t ret, retval = IRQ_NONE;
 360         unsigned int status = 0;
 361
 362         WARN_ONCE(!in_irq(), "BUG: IRQ handler called from non-hardirq context!");
 363
 364         if (!(action->flags & IRQF_DISABLED))
 365                 local_irq_enable_in_hardirq();
 366
 367         do {
 368                 ret = action->handler(irq, action->dev_id);
 369
 370                 switch (ret) {
 371                 case IRQ_WAKE_THREAD:
 372                         /*
 373                          * Set result to handled so the spurious check
 374                          * does not trigger.
 375                          */
 376                         ret = IRQ_HANDLED;
 377
 378                         /*
 379                          * Catch drivers which return WAKE_THREAD but
 380                          * did not set up a thread function
 381                          */
 382                         if (unlikely(!action->thread_fn)) {
 383                                 warn_no_thread(irq, action);
 384                                 break;
 385                         }
 386
 387                         /*
 388                          * Wake up the handler thread for this
 389                          * action. In case the thread crashed and was
 390                          * killed we just pretend that we handled the
 391                          * interrupt. The hardirq handler above has
 392                          * disabled the device interrupt, so no irq
 393                          * storm is lurking.
 394                          */
 395                         if (likely(!test_bit(IRQTF_DIED,
 396                                              &action->thread_flags))) {
 397                                 set_bit(IRQTF_RUNTHREAD, &action->thread_flags);
 398                                 wake_up_process(action->thread);
 399                         }
 400
 401                         /* Fall through to add to randomness */
 402                 case IRQ_HANDLED:
 403                         status |= action->flags;
 404                         break;
 405
 406                 default:
 407                         break;
 408                 }
 409
 410                 retval |= ret;
 411                 action = action->next;
 412         } while (action);
 413
 414         if (status & IRQF_SAMPLE_RANDOM)
 415                 add_interrupt_randomness(irq);
 416         local_irq_disable();
 417
 418         return retval;
 419 }
 420
 421 #ifndef CONFIG_GENERIC_HARDIRQS_NO__DO_IRQ
 422
 423 #ifdef CONFIG_ENABLE_WARN_DEPRECATED
 424 # warning __do_IRQ is deprecated. Please convert to proper flow handlers
 425 #endif
 426
 427 /**
 428  * __do_IRQ - original all in one highlevel IRQ handler
 429  * @irq:        the interrupt number
 430  *
 431  * __do_IRQ handles all normal device IRQ's (the special
 432  * SMP cross-CPU interrupts have their own specific
 433  * handlers).
 434  *
 435  * This is the original x86 implementation which is used for every
 436  * interrupt type.
 437  */
 438 unsigned int __do_IRQ(unsigned int irq)
 439 {
 440         struct irq_desc *desc = irq_to_desc(irq);
 441         struct irqaction *action;
 442         unsigned int status;
 443
 444         kstat_incr_irqs_this_cpu(irq, desc);
 445
 446         if (CHECK_IRQ_PER_CPU(desc->status)) {
 447                 irqreturn_t action_ret;
 448
 449                 /*
 450                  * No locking required for CPU-local interrupts:
 451                  */
 452                 if (desc->chip->ack) {
 453                         desc->chip->ack(irq);
 454                         /* get new one */
 455                         desc = irq_remap_to_desc(irq, desc);
 456                 }
 457                 if (likely(!(desc->status & IRQ_DISABLED))) {
 458                         action_ret = handle_IRQ_event(irq, desc->action);
 459                         if (!noirqdebug)
 460                                 note_interrupt(irq, desc, action_ret);
 461                 }
 462                 desc->chip->end(irq);
 463                 return 1;
 464         }
 465
 466         spin_lock(&desc->lock);
 467         if (desc->chip->ack) {
 468                 desc->chip->ack(irq);
 469                 desc = irq_remap_to_desc(irq, desc);
 470         }
 471         /*
 472          * REPLAY is when Linux resends an IRQ that was dropped earlier
 473          * WAITING is used by probe to mark irqs that are being tested
 474          */
 475         status = desc->status & ~(IRQ_REPLAY | IRQ_WAITING);
 476         status |= IRQ_PENDING; /* we _want_ to handle it */
 477
 478         /*
 479          * If the IRQ is disabled for whatever reason, we cannot
 480          * use the action we have.
 481          */
 482         action = NULL;
 483         if (likely(!(status & (IRQ_DISABLED | IRQ_INPROGRESS)))) {
 484                 action = desc->action;
 485                 status &= ~IRQ_PENDING; /* we commit to handling */
 486                 status |= IRQ_INPROGRESS; /* we are handling it */
 487         }
 488         desc->status = status;
 489
 490         /*
 491          * If there is no IRQ handler or it was disabled, exit early.
 492          * Since we set PENDING, if another processor is handling
 493          * a different instance of this same irq, the other processor
 494          * will take care of it.
 495          */
 496         if (unlikely(!action))
 497                 goto out;
 498
 499         /*
 500          * Edge triggered interrupts need to remember
 501          * pending events.
 502          * This applies to any hw interrupts that allow a second
 503          * instance of the same irq to arrive while we are in do_IRQ
 504          * or in the handler. But the code here only handles the _second_
 505          * instance of the irq, not the third or fourth. So it is mostly
 506          * useful for irq hardware that does not mask cleanly in an
 507          * SMP environment.
 508          */
 509         for (;;) {
 510                 irqreturn_t action_ret;
 511
 512                 spin_unlock(&desc->lock);
 513
 514                 action_ret = handle_IRQ_event(irq, action);
 515                 if (!noirqdebug)
 516                         note_interrupt(irq, desc, action_ret);
 517
 518                 spin_lock(&desc->lock);
 519                 if (likely(!(desc->status & IRQ_PENDING)))
 520                         break;
 521                 desc->status &= ~IRQ_PENDING;
 522         }
 523         desc->status &= ~IRQ_INPROGRESS;
 524
 525 out:
 526         /*
 527          * The ->end() handler has to deal with interrupts which got
 528          * disabled while the handler was running.
 529          */
 530         desc->chip->end(irq);
 531         spin_unlock(&desc->lock);
 532
 533         return 1;
 534 }
 535 #endif
 536
 537 void early_init_irq_lock_class(void)
 538 {
 539         struct irq_desc *desc;
 540         int i;
 541
 542         for_each_irq_desc(i, desc) {
 543                 lockdep_set_class(&desc->lock, &irq_desc_lock_class);
 544         }
 545 }
 546
 547 unsigned int kstat_irqs_cpu(unsigned int irq, int cpu)
 548 {
 549         struct irq_desc *desc = irq_to_desc(irq);
 550         return desc ? desc->kstat_irqs[cpu] : 0;
 551 }
 552 EXPORT_SYMBOL(kstat_irqs_cpu);
 553