arch/i386/kernel/irq.c

   1 /*
   2  *      linux/arch/i386/kernel/irq.c
   3  *
   4  *      Copyright (C) 1992, 1998 Linus Torvalds, Ingo Molnar
   5  *
   6  * This file contains the code used by various IRQ handling routines:
   7  * asking for different IRQ's should be done through these routines
   8  * instead of just grabbing them. Thus setups with different IRQ numbers
   9  * shouldn't result in any weird surprises, and installing new handlers
  10  * should be easier.
  11  */
  12
  13 /*
  14  * (mostly architecture independent, will move to kernel/irq.c in 2.5.)
  15  *
  16  * IRQs are in fact implemented a bit like signal handlers for the kernel.
  17  * Naturally it's not a 1:1 relation, but there are similarities.
  18  */
  19
  20 #include <linux/config.h>
  21 #include <linux/errno.h>
  22 #include <linux/module.h>
  23 #include <linux/signal.h>
  24 #include <linux/sched.h>
  25 #include <linux/ioport.h>
  26 #include <linux/interrupt.h>
  27 #include <linux/timex.h>
  28 #include <linux/slab.h>
  29 #include <linux/random.h>
  30 #include <linux/smp_lock.h>
  31 #include <linux/init.h>
  32 #include <linux/kernel_stat.h>
  33 #include <linux/irq.h>
  34 #include <linux/proc_fs.h>
  35 #include <linux/seq_file.h>
  36 #include <linux/kallsyms.h>
  37
  38 #include <asm/atomic.h>
  39 #include <asm/io.h>
  40 #include <asm/smp.h>
  41 #include <asm/system.h>
  42 #include <asm/bitops.h>
  43 #include <asm/uaccess.h>
  44 #include <asm/delay.h>
  45 #include <asm/desc.h>
  46 #include <asm/irq.h>
  47
  48 /*
  49  * Linux has a controller-independent x86 interrupt architecture.
  50  * every controller has a 'controller-template', that is used
  51  * by the main code to do the right thing. Each driver-visible
  52  * interrupt source is transparently wired to the apropriate
  53  * controller. Thus drivers need not be aware of the
  54  * interrupt-controller.
  55  *
  56  * Various interrupt controllers we handle: 8259 PIC, SMP IO-APIC,
  57  * PIIX4's internal 8259 PIC and SGI's Visual Workstation Cobalt (IO-)APIC.
  58  * (IO-APICs assumed to be messaging to Pentium local-APICs)
  59  *
  60  * the code is designed to be easily extended with new/different
  61  * interrupt controllers, without having to do assembly magic.
  62  */
  63
  64 /*
  65  * Controller mappings for all interrupt sources:
  66  */
  67 irq_desc_t irq_desc[NR_IRQS] __cacheline_aligned = {
  68         [0 ... NR_IRQS-1] = {
  69                 .handler = &no_irq_type,
  70                 .lock = SPIN_LOCK_UNLOCKED
  71         }
  72 };
  73
  74 static void register_irq_proc (unsigned int irq);
  75
  76 /*
  77  * per-CPU IRQ handling stacks
  78  */
  79 #ifdef CONFIG_IRQSTACKS
  80 union irq_ctx *hardirq_ctx[NR_CPUS];
  81 union irq_ctx *softirq_ctx[NR_CPUS];
  82 #endif
  83
  84 /*
  85  * Special irq handlers.
  86  */
  87
  88 irqreturn_t no_action(int cpl, void *dev_id, struct pt_regs *regs)
  89 { return IRQ_NONE; }
  90
  91 /*
  92  * Generic no controller code
  93  */
  94
  95 static void enable_none(unsigned int irq) { }
  96 static unsigned int startup_none(unsigned int irq) { return 0; }
  97 static void disable_none(unsigned int irq) { }
  98 static void ack_none(unsigned int irq)
  99 {
 100 /*
 101  * 'what should we do if we get a hw irq event on an illegal vector'.
 102  * each architecture has to answer this themselves, it doesn't deserve
 103  * a generic callback i think.
 104  */
 105 #ifdef CONFIG_X86
 106         printk("unexpected IRQ trap at vector %02x\n", irq);
 107 #ifdef CONFIG_X86_LOCAL_APIC
 108         /*
 109          * Currently unexpected vectors happen only on SMP and APIC.
 110          * We _must_ ack these because every local APIC has only N
 111          * irq slots per priority level, and a 'hanging, unacked' IRQ
 112          * holds up an irq slot - in excessive cases (when multiple
 113          * unexpected vectors occur) that might lock up the APIC
 114          * completely.
 115          */
 116         ack_APIC_irq();
 117 #endif
 118 #endif
 119 }
 120
 121 /* startup is the same as "enable", shutdown is same as "disable" */
 122 #define shutdown_none   disable_none
 123 #define end_none        enable_none
 124
 125 struct hw_interrupt_type no_irq_type = {
 126         "none",
 127         startup_none,
 128         shutdown_none,
 129         enable_none,
 130         disable_none,
 131         ack_none,
 132         end_none
 133 };
 134
 135 atomic_t irq_err_count;
 136 #if defined(CONFIG_X86_IO_APIC) && defined(APIC_MISMATCH_DEBUG)
 137 atomic_t irq_mis_count;
 138 #endif
 139
 140 /*
 141  * Generic, controller-independent functions:
 142  */
 143
 144 int show_interrupts(struct seq_file *p, void *v)
 145 {
 146         int i = *(loff_t *) v, j;
 147         struct irqaction * action;
 148         unsigned long flags;
 149
 150         if (i == 0) {
 151                 seq_printf(p, "           ");
 152                 for (j=0; j<NR_CPUS; j++)
 153                         if (cpu_online(j))
 154                                 seq_printf(p, "CPU%d       ",j);
 155                 seq_putc(p, '\n');
 156         }
 157
 158         if (i < NR_IRQS) {
 159                 spin_lock_irqsave(&irq_desc[i].lock, flags);
 160                 action = irq_desc[i].action;
 161                 if (!action)
 162                         goto skip;
 163                 seq_printf(p, "%3d: ",i);
 164 #ifndef CONFIG_SMP
 165                 seq_printf(p, "%10u ", kstat_irqs(i));
 166 #else
 167                 for (j = 0; j < NR_CPUS; j++)
 168                         if (cpu_online(j))
 169                                 seq_printf(p, "%10u ", kstat_cpu(j).irqs[i]);
 170 #endif
 171                 seq_printf(p, " %14s", irq_desc[i].handler->typename);
 172                 seq_printf(p, "  %s", action->name);
 173
 174                 for (action=action->next; action; action = action->next)
 175                         seq_printf(p, ", %s", action->name);
 176
 177                 seq_putc(p, '\n');
 178 skip:
 179                 spin_unlock_irqrestore(&irq_desc[i].lock, flags);
 180         } else if (i == NR_IRQS) {
 181                 seq_printf(p, "NMI: ");
 182                 for (j = 0; j < NR_CPUS; j++)
 183                         if (cpu_online(j))
 184                                 seq_printf(p, "%10u ", nmi_count(j));
 185                 seq_putc(p, '\n');
 186 #ifdef CONFIG_X86_LOCAL_APIC
 187                 seq_printf(p, "LOC: ");
 188                 for (j = 0; j < NR_CPUS; j++)
 189                         if (cpu_online(j))
 190                                 seq_printf(p, "%10u ", irq_stat[j].apic_timer_irqs);
 191                 seq_putc(p, '\n');
 192 #endif
 193                 seq_printf(p, "ERR: %10u\n", atomic_read(&irq_err_count));
 194 #if defined(CONFIG_X86_IO_APIC) && defined(APIC_MISMATCH_DEBUG)
 195                 seq_printf(p, "MIS: %10u\n", atomic_read(&irq_mis_count));
 196 #endif
 197         }
 198         return 0;
 199 }
 200
 201
 202
 203
 204 #ifdef CONFIG_SMP
 205 inline void synchronize_irq(unsigned int irq)
 206 {
 207         while (irq_desc[irq].status & IRQ_INPROGRESS)
 208                 cpu_relax();
 209 }
 210 #endif
 211
 212 /*
 213  * This should really return information about whether
 214  * we should do bottom half handling etc. Right now we
 215  * end up _always_ checking the bottom half, which is a
 216  * waste of time and is not what some drivers would
 217  * prefer.
 218  */
 219 asmlinkage int handle_IRQ_event(unsigned int irq,
 220                 struct pt_regs *regs, struct irqaction *action)
 221 {
 222         int status = 1; /* Force the "do bottom halves" bit */
 223         int ret, retval = 0;
 224
 225         if (!(action->flags & SA_INTERRUPT))
 226                 local_irq_enable();
 227
 228         if (!(action->flags & SA_INTERRUPT))
 229                 local_irq_enable();
 230
 231         do {
 232                 ret = action->handler(irq, action->dev_id, regs);
 233                 if (ret == IRQ_HANDLED)
 234                         status |= action->flags;
 235                 retval |= ret;
 236                 action = action->next;
 237         } while (action);
 238         if (status & SA_SAMPLE_RANDOM)
 239                 add_interrupt_randomness(irq);
 240         local_irq_disable();
 241         return retval;
 242 }
 243
 244 static void __report_bad_irq(int irq, irq_desc_t *desc, irqreturn_t action_ret)
 245 {
 246         struct irqaction *action;
 247
 248         if (action_ret != IRQ_HANDLED && action_ret != IRQ_NONE) {
 249                 printk(KERN_ERR "irq event %d: bogus return value %x\n",
 250                                 irq, action_ret);
 251         } else {
 252                 printk(KERN_ERR "irq %d: nobody cared! (screaming interrupt?)\n", irq);
 253                 printk(KERN_ERR "irq %d: Please try booting with acpi=off and report a bug\n", irq);
 254         }
 255         dump_stack();
 256         printk(KERN_ERR "handlers:\n");
 257         action = desc->action;
 258         do {
 259                 printk(KERN_ERR "[<%p>]", action->handler);
 260                 print_symbol(" (%s)",
 261                         (unsigned long)action->handler);
 262                 printk("\n");
 263                 action = action->next;
 264         } while (action);
 265 }
 266
 267 static void report_bad_irq(int irq, irq_desc_t *desc, irqreturn_t action_ret)
 268 {
 269         static int count = 100;
 270
 271         if (count) {
 272                 count--;
 273                 __report_bad_irq(irq, desc, action_ret);
 274         }
 275 }
 276
 277 static int noirqdebug;
 278
 279 int __init noirqdebug_setup(char *str)
 280 {
 281         noirqdebug = 1;
 282         printk(KERN_INFO "IRQ lockup detection disabled\n");
 283         return 1;
 284 }
 285
 286 __setup("noirqdebug", noirqdebug_setup);
 287
 288 static int irqfixup;
 289
 290 static int __init irqfixup_setup(char *str)
 291 {
 292         irqfixup = 1;
 293         printk(KERN_WARNING "Misrouted IRQ fixup support enabled.\n");
 294         printk(KERN_WARNING "This may impact system performance.\n");
 295         return 1;
 296 }
 297
 298 __setup("irqfixup", irqfixup_setup);
 299
 300 static int __init irqpoll_setup(char *str)
 301 {
 302         irqfixup = 2;
 303         printk(KERN_WARNING "Misrouted IRQ fixup and polling support enabled.\n");
 304         printk(KERN_WARNING "This may significantly impact system performance.\n");
 305         return 1;
 306 }
 307
 308 __setup("irqpoll", irqpoll_setup);
 309
 310 /*
 311  *      Recovery handler for misrouted interrupts
 312  */
 313
 314 static asmlinkage int misrouted_irq(int irq, struct pt_regs *regs)
 315 {
 316         int i;
 317         irq_desc_t *desc;
 318         int ok = 0;
 319         int work = 0;   /* Did we do work for a real IRQ */
 320         for(i = 1; i < NR_IRQS; i++)
 321         {
 322                 struct irqaction *action;
 323                 if(i == irq)    /* Already tried */
 324                         continue;
 325                 desc = &irq_desc[i];
 326                 spin_lock(&desc->lock);
 327                 action = desc->action;
 328                 /* Already running on another processor */
 329                 if(desc->status & IRQ_INPROGRESS)
 330                 {
 331                         /* Already running: If it is shared get the other
 332                            CPU to go looking for our mystery interrupt too */
 333                         if(desc->action && (desc->action->flags & SA_SHIRQ))
 334                                 desc->status |= IRQ_PENDING;
 335                         spin_unlock(&desc->lock);
 336                         continue;
 337                 }
 338                 /* Honour the normal IRQ locking */
 339                 desc->status |= IRQ_INPROGRESS;
 340                 spin_unlock(&desc->lock);
 341                 while(action)
 342                 {
 343                         /* Only shared IRQ handlers are safe to call */
 344                         if(action->flags & SA_SHIRQ)
 345                         {
 346                                 if(action->handler(i, action->dev_id, regs) == IRQ_HANDLED)
 347                                         ok = 1;
 348                         }
 349                         action = action->next;
 350                 }
 351                 local_irq_disable();
 352                 /* Now clean up the flags */
 353                 spin_lock(&desc->lock);
 354                 action = desc->action;
 355
 356                 /* While we were looking for a fixup someone queued a real
 357                    IRQ clashing with our walk */
 358
 359                 while((desc->status & IRQ_PENDING) && action)
 360                 {
 361                         /* Perform real IRQ processing for the IRQ we deferred */
 362                         work = 1;
 363                         spin_unlock(&desc->lock);
 364                         handle_IRQ_event(i, regs, action);
 365                         spin_lock(&desc->lock);
 366                         desc->status &= ~IRQ_PENDING;
 367                 }
 368                 desc->status &= ~IRQ_INPROGRESS;
 369                 /* If we did actual work for the real IRQ line we must
 370                    let the IRQ controller clean up too */
 371                 if(work)
 372                         desc->handler->end(i);
 373                 spin_unlock(&desc->lock);
 374         }
 375         /* So the caller can adjust the irq error counts */
 376         return ok;
 377 }
 378
 379 /*
 380  * If 99,900 of the previous 100,000 interrupts have not been handled then
 381  * assume that the IRQ is stuck in some manner.  Drop a diagnostic and try to
 382  * turn the IRQ off.
 383  *
 384  * (The other 100-of-100,000 interrupts may have been a correctly-functioning
 385  *  device sharing an IRQ with the failing one)
 386  *
 387  * Called under desc->lock
 388  */
 389 static void note_interrupt(int irq, irq_desc_t *desc, irqreturn_t action_ret, struct pt_regs *regs)
 390 {
 391         if (action_ret != IRQ_HANDLED) {
 392                 desc->irqs_unhandled++;
 393                 if (action_ret != IRQ_NONE)
 394                         report_bad_irq(irq, desc, action_ret);
 395         }
 396         if(unlikely(irqfixup))  /* Don't punish working computers */
 397         {
 398                 if((irqfixup == 2 && irq == 0) || action_ret == IRQ_NONE)
 399                 {
 400                         int ok;
 401                         u32 *isp;
 402                         union irq_ctx * curctx;
 403                         union irq_ctx * irqctx;
 404
 405                         curctx = (union irq_ctx *) current_thread_info();
 406                         irqctx = hardirq_ctx[smp_processor_id()];
 407
 408                         spin_unlock(&desc->lock);
 409
 410                         /*
 411                          * this is where we switch to the IRQ stack. However, if we are already using
 412                          * the IRQ stack (because we interrupted a hardirq handler) we can't do that
 413                          * and just have to keep using the current stack (which is the irq stack already
 414                          * after all)
 415                          */
 416
 417                         if (curctx == irqctx)
 418                                 ok = misrouted_irq(irq, regs);
 419                         else {
 420                                 /* build the stack frame on the IRQ stack */
 421                                 isp = (u32*) ((char*)irqctx + sizeof(*irqctx));
 422                                 irqctx->tinfo.task = curctx->tinfo.task;
 423                                 irqctx->tinfo.previous_esp = current_stack_pointer();
 424
 425                                 *--isp = (u32) regs;
 426                                 *--isp = (u32) irq;
 427
 428                                 asm volatile(
 429                                 "       xchgl   %%ebx,%%esp     \n"
 430                                 "       call    misrouted_irq   \n"
 431                                 "       xchgl   %%ebx,%%esp     \n"
 432                                 : "=a"(ok)
 433                                 : "b"(isp)
 434                                 : "memory", "cc", "edx", "ecx"
 435                                 );
 436                         }
 437                         spin_lock(&desc->lock);
 438                         if (curctx != irqctx)
 439                                 irqctx->tinfo.task = NULL;
 440                         if(action_ret == IRQ_NONE)
 441                                 desc->irqs_unhandled -= ok;
 442                 }
 443         }
 444
 445         desc->irq_count++;
 446         if (desc->irq_count < 100000)
 447                 return;
 448
 449         desc->irq_count = 0;
 450         if (desc->irqs_unhandled > 99900) {
 451                 /*
 452                  * The interrupt is stuck
 453                  */
 454                 __report_bad_irq(irq, desc, action_ret);
 455                 /*
 456                  * Now kill the IRQ
 457                  */
 458                 printk(KERN_EMERG "Disabling IRQ #%d\n", irq);
 459                 desc->status |= IRQ_DISABLED;
 460                 desc->handler->disable(irq);
 461         }
 462         desc->irqs_unhandled = 0;
 463 }
 464
 465 /*
 466  * Generic enable/disable code: this just calls
 467  * down into the PIC-specific version for the actual
 468  * hardware disable after having gotten the irq
 469  * controller lock.
 470  */
 471
 472 /**
 473  *      disable_irq_nosync - disable an irq without waiting
 474  *      @irq: Interrupt to disable
 475  *
 476  *      Disable the selected interrupt line.  Disables and Enables are
 477  *      nested.
 478  *      Unlike disable_irq(), this function does not ensure existing
 479  *      instances of the IRQ handler have completed before returning.
 480  *
 481  *      This function may be called from IRQ context.
 482  */
 483
 484 inline void disable_irq_nosync(unsigned int irq)
 485 {
 486         irq_desc_t *desc = irq_desc + irq;
 487         unsigned long flags;
 488
 489         spin_lock_irqsave(&desc->lock, flags);
 490         if (!desc->depth++) {
 491                 desc->status |= IRQ_DISABLED;
 492                 desc->handler->disable(irq);
 493         }
 494         spin_unlock_irqrestore(&desc->lock, flags);
 495 }
 496
 497 /**
 498  *      disable_irq - disable an irq and wait for completion
 499  *      @irq: Interrupt to disable
 500  *
 501  *      Disable the selected interrupt line.  Enables and Disables are
 502  *      nested.
 503  *      This function waits for any pending IRQ handlers for this interrupt
 504  *      to complete before returning. If you use this function while
 505  *      holding a resource the IRQ handler may need you will deadlock.
 506  *
 507  *      This function may be called - with care - from IRQ context.
 508  */
 509
 510 void disable_irq(unsigned int irq)
 511 {
 512         irq_desc_t *desc = irq_desc + irq;
 513         disable_irq_nosync(irq);
 514         if (desc->action)
 515                 synchronize_irq(irq);
 516 }
 517
 518 /**
 519  *      enable_irq - enable handling of an irq
 520  *      @irq: Interrupt to enable
 521  *
 522  *      Undoes the effect of one call to disable_irq().  If this
 523  *      matches the last disable, processing of interrupts on this
 524  *      IRQ line is re-enabled.
 525  *
 526  *      This function may be called from IRQ context.
 527  */
 528
 529 void enable_irq(unsigned int irq)
 530 {
 531         irq_desc_t *desc = irq_desc + irq;
 532         unsigned long flags;
 533
 534         spin_lock_irqsave(&desc->lock, flags);
 535         switch (desc->depth) {
 536         case 1: {
 537                 unsigned int status = desc->status & ~IRQ_DISABLED;
 538                 desc->status = status;
 539                 if ((status & (IRQ_PENDING | IRQ_REPLAY)) == IRQ_PENDING) {
 540                         desc->status = status | IRQ_REPLAY;
 541                         hw_resend_irq(desc->handler,irq);
 542                 }
 543                 desc->handler->enable(irq);
 544                 /* fall-through */
 545         }
 546         default:
 547                 desc->depth--;
 548                 break;
 549         case 0:
 550                 printk("enable_irq(%u) unbalanced from %p\n", irq,
 551                        __builtin_return_address(0));
 552         }
 553         spin_unlock_irqrestore(&desc->lock, flags);
 554 }
 555
 556 /*
 557  * do_IRQ handles all normal device IRQ's (the special
 558  * SMP cross-CPU interrupts have their own specific
 559  * handlers).
 560  */
 561 asmlinkage unsigned int do_IRQ(struct pt_regs regs)
 562 {
 563         /*
 564          * We ack quickly, we don't want the irq controller
 565          * thinking we're snobs just because some other CPU has
 566          * disabled global interrupts (we have already done the
 567          * INT_ACK cycles, it's too late to try to pretend to the
 568          * controller that we aren't taking the interrupt).
 569          *
 570          * 0 return value means that this irq is already being
 571          * handled by some other CPU. (or is disabled)
 572          */
 573         int irq = regs.orig_eax & 0xff; /* high bits used in ret_from_ code  */
 574         irq_desc_t *desc = irq_desc + irq;
 575         struct irqaction * action;
 576         unsigned int status;
 577
 578         irq_enter();
 579
 580 #ifdef CONFIG_DEBUG_STACKOVERFLOW
 581         /* Debugging check for stack overflow: is there less than 1KB free? */
 582         {
 583                 long esp;
 584
 585                 __asm__ __volatile__("andl %%esp,%0" :
 586                                         "=r" (esp) : "0" (THREAD_SIZE - 1));
 587                 if (unlikely(esp < (sizeof(struct thread_info) + STACK_WARN))) {
 588                         printk("do_IRQ: stack overflow: %ld\n",
 589                                 esp - sizeof(struct thread_info));
 590                         dump_stack();
 591                 }
 592         }
 593 #endif
 594         kstat_this_cpu.irqs[irq]++;
 595         spin_lock(&desc->lock);
 596         desc->handler->ack(irq);
 597         /*
 598            REPLAY is when Linux resends an IRQ that was dropped earlier
 599            WAITING is used by probe to mark irqs that are being tested
 600            */
 601         status = desc->status & ~(IRQ_REPLAY | IRQ_WAITING);
 602         status |= IRQ_PENDING; /* we _want_ to handle it */
 603
 604         /*
 605          * If the IRQ is disabled for whatever reason, we cannot
 606          * use the action we have.
 607          */
 608         action = NULL;
 609         if (likely(!(status & (IRQ_DISABLED | IRQ_INPROGRESS)))) {
 610                 action = desc->action;
 611                 status &= ~IRQ_PENDING; /* we commit to handling */
 612                 status |= IRQ_INPROGRESS; /* we are handling it */
 613         }
 614         desc->status = status;
 615
 616         /*
 617          * If there is no IRQ handler or it was disabled, exit early.
 618            Since we set PENDING, if another processor is handling
 619            a different instance of this same irq, the other processor
 620            will take care of it.
 621          */
 622         if (unlikely(!action))
 623                 goto out;
 624
 625         /*
 626          * Edge triggered interrupts need to remember
 627          * pending events.
 628          * This applies to any hw interrupts that allow a second
 629          * instance of the same irq to arrive while we are in do_IRQ
 630          * or in the handler. But the code here only handles the _second_
 631          * instance of the irq, not the third or fourth. So it is mostly
 632          * useful for irq hardware that does not mask cleanly in an
 633          * SMP environment.
 634          */
 635         for (;;) {
 636                 irqreturn_t action_ret;
 637                 u32 *isp;
 638                 union irq_ctx * curctx;
 639                 union irq_ctx * irqctx;
 640 #ifdef CONFIG_IRQSTACKS
 641                 curctx = (union irq_ctx *) current_thread_info();
 642                 irqctx = hardirq_ctx[smp_processor_id()];
 643 #else
 644                 curctx = irqctx = (union irq_ctx *)0;
 645 #endif
 646                 spin_unlock(&desc->lock);
 647
 648                 /*
 649                  * this is where we switch to the IRQ stack. However, if we are already using
 650                  * the IRQ stack (because we interrupted a hardirq handler) we can't do that
 651                  * and just have to keep using the current stack (which is the irq stack already
 652                  * after all)
 653                  */
 654
 655                 if (curctx == irqctx)
 656                         action_ret = handle_IRQ_event(irq, &regs, action);
 657                 else {
 658                         /* build the stack frame on the IRQ stack */
 659                         isp = (u32*) ((char*)irqctx + sizeof(*irqctx));
 660                         irqctx->tinfo.task = curctx->tinfo.task;
 661                         irqctx->tinfo.previous_esp = current_stack_pointer();
 662
 663                         *--isp = (u32) action;
 664                         *--isp = (u32) &regs;
 665                         *--isp = (u32) irq;
 666
 667                         asm volatile(
 668                                 "       xchgl   %%ebx,%%esp     \n"
 669                                 "       call    handle_IRQ_event \n"
 670                                 "       xchgl   %%ebx,%%esp     \n"
 671                                 : "=a"(action_ret)
 672                                 : "b"(isp)
 673                                 : "memory", "cc", "edx", "ecx"
 674                         );
 675
 676
 677                 }
 678                 spin_lock(&desc->lock);
 679                 if (!noirqdebug)
 680                         note_interrupt(irq, desc, action_ret, &regs);
 681                 if (curctx != irqctx)
 682                         irqctx->tinfo.task = NULL;
 683                 if (likely(!(desc->status & IRQ_PENDING)))
 684                         break;
 685                 desc->status &= ~IRQ_PENDING;
 686         }
 687         desc->status &= ~IRQ_INPROGRESS;
 688
 689 out:
 690         /*
 691          * The ->end() handler has to deal with interrupts which got
 692          * disabled while the handler was running.
 693          */
 694         desc->handler->end(irq);
 695         spin_unlock(&desc->lock);
 696
 697         irq_exit();
 698
 699         return 1;
 700 }
 701
 702 int can_request_irq(unsigned int irq, unsigned long irqflags)
 703 {
 704         struct irqaction *action;
 705
 706         if (irq >= NR_IRQS)
 707                 return 0;
 708         action = irq_desc[irq].action;
 709         if (action) {
 710                 if (irqflags & action->flags & SA_SHIRQ)
 711                         action = NULL;
 712         }
 713         return !action;
 714 }
 715
 716 /**
 717  *      request_irq - allocate an interrupt line
 718  *      @irq: Interrupt line to allocate
 719  *      @handler: Function to be called when the IRQ occurs
 720  *      @irqflags: Interrupt type flags
 721  *      @devname: An ascii name for the claiming device
 722  *      @dev_id: A cookie passed back to the handler function
 723  *
 724  *      This call allocates interrupt resources and enables the
 725  *      interrupt line and IRQ handling. From the point this
 726  *      call is made your handler function may be invoked. Since
 727  *      your handler function must clear any interrupt the board
 728  *      raises, you must take care both to initialise your hardware
 729  *      and to set up the interrupt handler in the right order.
 730  *
 731  *      Dev_id must be globally unique. Normally the address of the
 732  *      device data structure is used as the cookie. Since the handler
 733  *      receives this value it makes sense to use it.
 734  *
 735  *      If your interrupt is shared you must pass a non NULL dev_id
 736  *      as this is required when freeing the interrupt.
 737  *
 738  *      Flags:
 739  *
 740  *      SA_SHIRQ                Interrupt is shared
 741  *
 742  *      SA_INTERRUPT            Disable local interrupts while processing
 743  *
 744  *      SA_SAMPLE_RANDOM        The interrupt can be used for entropy
 745  *
 746  */
 747
 748 int request_irq(unsigned int irq,
 749                 irqreturn_t (*handler)(int, void *, struct pt_regs *),
 750                 unsigned long irqflags,
 751                 const char * devname,
 752                 void *dev_id)
 753 {
 754         int retval;
 755         struct irqaction * action;
 756
 757 #if 1
 758         /*
 759          * Sanity-check: shared interrupts should REALLY pass in
 760          * a real dev-ID, otherwise we'll have trouble later trying
 761          * to figure out which interrupt is which (messes up the
 762          * interrupt freeing logic etc).
 763          */
 764         if (irqflags & SA_SHIRQ) {
 765                 if (!dev_id)
 766                         printk("Bad boy: %s (at 0x%x) called us without a dev_id!\n", devname, (&irq)[-1]);
 767         }
 768 #endif
 769
 770         if (irq >= NR_IRQS)
 771                 return -EINVAL;
 772         if (!handler)
 773                 return -EINVAL;
 774
 775         action = (struct irqaction *)
 776                         kmalloc(sizeof(struct irqaction), GFP_ATOMIC);
 777         if (!action)
 778                 return -ENOMEM;
 779
 780         action->handler = handler;
 781         action->flags = irqflags;
 782         cpus_clear(action->mask);
 783         action->name = devname;
 784         action->next = NULL;
 785         action->dev_id = dev_id;
 786
 787         retval = setup_irq(irq, action);
 788         if (retval)
 789                 kfree(action);
 790         return retval;
 791 }
 792
 793 EXPORT_SYMBOL(request_irq);
 794
 795 /**
 796  *      free_irq - free an interrupt
 797  *      @irq: Interrupt line to free
 798  *      @dev_id: Device identity to free
 799  *
 800  *      Remove an interrupt handler. The handler is removed and if the
 801  *      interrupt line is no longer in use by any driver it is disabled.
 802  *      On a shared IRQ the caller must ensure the interrupt is disabled
 803  *      on the card it drives before calling this function. The function
 804  *      does not return until any executing interrupts for this IRQ
 805  *      have completed.
 806  *
 807  *      This function must not be called from interrupt context.
 808  */
 809
 810 void free_irq(unsigned int irq, void *dev_id)
 811 {
 812         irq_desc_t *desc;
 813         struct irqaction **p;
 814         unsigned long flags;
 815
 816         if (irq >= NR_IRQS)
 817                 return;
 818
 819         desc = irq_desc + irq;
 820         spin_lock_irqsave(&desc->lock,flags);
 821         p = &desc->action;
 822         for (;;) {
 823                 struct irqaction * action = *p;
 824                 if (action) {
 825                         struct irqaction **pp = p;
 826                         p = &action->next;
 827                         if (action->dev_id != dev_id)
 828                                 continue;
 829
 830                         /* Found it - now remove it from the list of entries */
 831                         *pp = action->next;
 832                         if (!desc->action) {
 833                                 desc->status |= IRQ_DISABLED;
 834                                 desc->handler->shutdown(irq);
 835                         }
 836                         spin_unlock_irqrestore(&desc->lock,flags);
 837
 838                         /* Wait to make sure it's not being used on another CPU */
 839                         synchronize_irq(irq);
 840                         kfree(action);
 841                         return;
 842                 }
 843                 printk("Trying to free free IRQ%d\n",irq);
 844                 spin_unlock_irqrestore(&desc->lock,flags);
 845                 return;
 846         }
 847 }
 848
 849 EXPORT_SYMBOL(free_irq);
 850
 851 /*
 852  * IRQ autodetection code..
 853  *
 854  * This depends on the fact that any interrupt that
 855  * comes in on to an unassigned handler will get stuck
 856  * with "IRQ_WAITING" cleared and the interrupt
 857  * disabled.
 858  */
 859
 860 static DECLARE_MUTEX(probe_sem);
 861
 862 /**
 863  *      probe_irq_on    - begin an interrupt autodetect
 864  *
 865  *      Commence probing for an interrupt. The interrupts are scanned
 866  *      and a mask of potential interrupt lines is returned.
 867  *
 868  */
 869
 870 unsigned long probe_irq_on(void)
 871 {
 872         unsigned int i;
 873         irq_desc_t *desc;
 874         unsigned long val;
 875         unsigned long delay;
 876
 877         down(&probe_sem);
 878         /*
 879          * something may have generated an irq long ago and we want to
 880          * flush such a longstanding irq before considering it as spurious.
 881          */
 882         for (i = NR_IRQS-1; i > 0; i--)  {
 883                 desc = irq_desc + i;
 884
 885                 spin_lock_irq(&desc->lock);
 886                 if (!irq_desc[i].action)
 887                         irq_desc[i].handler->startup(i);
 888                 spin_unlock_irq(&desc->lock);
 889         }
 890
 891         /* Wait for longstanding interrupts to trigger. */
 892         for (delay = jiffies + HZ/50; time_after(delay, jiffies); )
 893                 /* about 20ms delay */ barrier();
 894
 895         /*
 896          * enable any unassigned irqs
 897          * (we must startup again here because if a longstanding irq
 898          * happened in the previous stage, it may have masked itself)
 899          */
 900         for (i = NR_IRQS-1; i > 0; i--) {
 901                 desc = irq_desc + i;
 902
 903                 spin_lock_irq(&desc->lock);
 904                 if (!desc->action) {
 905                         desc->status |= IRQ_AUTODETECT | IRQ_WAITING;
 906                         if (desc->handler->startup(i))
 907                                 desc->status |= IRQ_PENDING;
 908                 }
 909                 spin_unlock_irq(&desc->lock);
 910         }
 911
 912         /*
 913          * Wait for spurious interrupts to trigger
 914          */
 915         for (delay = jiffies + HZ/10; time_after(delay, jiffies); )
 916                 /* about 100ms delay */ barrier();
 917
 918         /*
 919          * Now filter out any obviously spurious interrupts
 920          */
 921         val = 0;
 922         for (i = 0; i < NR_IRQS; i++) {
 923                 irq_desc_t *desc = irq_desc + i;
 924                 unsigned int status;
 925
 926                 spin_lock_irq(&desc->lock);
 927                 status = desc->status;
 928
 929                 if (status & IRQ_AUTODETECT) {
 930                         /* It triggered already - consider it spurious. */
 931                         if (!(status & IRQ_WAITING)) {
 932                                 desc->status = status & ~IRQ_AUTODETECT;
 933                                 desc->handler->shutdown(i);
 934                         } else
 935                                 if (i < 32)
 936                                         val |= 1 << i;
 937                 }
 938                 spin_unlock_irq(&desc->lock);
 939         }
 940
 941         return val;
 942 }
 943
 944 EXPORT_SYMBOL(probe_irq_on);
 945
 946 /*
 947  * Return a mask of triggered interrupts (this
 948  * can handle only legacy ISA interrupts).
 949  */
 950
 951 /**
 952  *      probe_irq_mask - scan a bitmap of interrupt lines
 953  *      @val:   mask of interrupts to consider
 954  *
 955  *      Scan the ISA bus interrupt lines and return a bitmap of
 956  *      active interrupts. The interrupt probe logic state is then
 957  *      returned to its previous value.
 958  *
 959  *      Note: we need to scan all the irq's even though we will
 960  *      only return ISA irq numbers - just so that we reset them
 961  *      all to a known state.
 962  */
 963 unsigned int probe_irq_mask(unsigned long val)
 964 {
 965         int i;
 966         unsigned int mask;
 967
 968         mask = 0;
 969         for (i = 0; i < NR_IRQS; i++) {
 970                 irq_desc_t *desc = irq_desc + i;
 971                 unsigned int status;
 972
 973                 spin_lock_irq(&desc->lock);
 974                 status = desc->status;
 975
 976                 if (status & IRQ_AUTODETECT) {
 977                         if (i < 16 && !(status & IRQ_WAITING))
 978                                 mask |= 1 << i;
 979
 980                         desc->status = status & ~IRQ_AUTODETECT;
 981                         desc->handler->shutdown(i);
 982                 }
 983                 spin_unlock_irq(&desc->lock);
 984         }
 985         up(&probe_sem);
 986
 987         return mask & val;
 988 }
 989
 990 /*
 991  * Return the one interrupt that triggered (this can
 992  * handle any interrupt source).
 993  */
 994
 995 /**
 996  *      probe_irq_off   - end an interrupt autodetect
 997  *      @val: mask of potential interrupts (unused)
 998  *
 999  *      Scans the unused interrupt lines and returns the line which
1000  *      appears to have triggered the interrupt. If no interrupt was
1001  *      found then zero is returned. If more than one interrupt is
1002  *      found then minus the first candidate is returned to indicate
1003  *      their is doubt.
1004  *
1005  *      The interrupt probe logic state is returned to its previous
1006  *      value.
1007  *
1008  *      BUGS: When used in a module (which arguably shouldnt happen)
1009  *      nothing prevents two IRQ probe callers from overlapping. The
1010  *      results of this are non-optimal.
1011  */
1012
1013 int probe_irq_off(unsigned long val)
1014 {
1015         int i, irq_found, nr_irqs;
1016
1017         nr_irqs = 0;
1018         irq_found = 0;
1019         for (i = 0; i < NR_IRQS; i++) {
1020                 irq_desc_t *desc = irq_desc + i;
1021                 unsigned int status;
1022
1023                 spin_lock_irq(&desc->lock);
1024                 status = desc->status;
1025
1026                 if (status & IRQ_AUTODETECT) {
1027                         if (!(status & IRQ_WAITING)) {
1028                                 if (!nr_irqs)
1029                                         irq_found = i;
1030                                 nr_irqs++;
1031                         }
1032                         desc->status = status & ~IRQ_AUTODETECT;
1033                         desc->handler->shutdown(i);
1034                 }
1035                 spin_unlock_irq(&desc->lock);
1036         }
1037         up(&probe_sem);
1038
1039         if (nr_irqs > 1)
1040                 irq_found = -irq_found;
1041         return irq_found;
1042 }
1043
1044 EXPORT_SYMBOL(probe_irq_off);
1045
1046 /* this was setup_x86_irq but it seems pretty generic */
1047 int setup_irq(unsigned int irq, struct irqaction * new)
1048 {
1049         int shared = 0;
1050         unsigned long flags;
1051         struct irqaction *old, **p;
1052         irq_desc_t *desc = irq_desc + irq;
1053
1054         if (desc->handler == &no_irq_type)
1055                 return -ENOSYS;
1056         /*
1057          * Some drivers like serial.c use request_irq() heavily,
1058          * so we have to be careful not to interfere with a
1059          * running system.
1060          */
1061         if (new->flags & SA_SAMPLE_RANDOM) {
1062                 /*
1063                  * This function might sleep, we want to call it first,
1064                  * outside of the atomic block.
1065                  * Yes, this might clear the entropy pool if the wrong
1066                  * driver is attempted to be loaded, without actually
1067                  * installing a new handler, but is this really a problem,
1068                  * only the sysadmin is able to do this.
1069                  */
1070                 rand_initialize_irq(irq);
1071         }
1072
1073         /*
1074          * The following block of code has to be executed atomically
1075          */
1076         spin_lock_irqsave(&desc->lock,flags);
1077         p = &desc->action;
1078         if ((old = *p) != NULL) {
1079                 /* Can't share interrupts unless both agree to */
1080                 if (!(old->flags & new->flags & SA_SHIRQ)) {
1081                         spin_unlock_irqrestore(&desc->lock,flags);
1082                         return -EBUSY;
1083                 }
1084
1085                 /* add new interrupt at end of irq queue */
1086                 do {
1087                         p = &old->next;
1088                         old = *p;
1089                 } while (old);
1090                 shared = 1;
1091         }
1092
1093         *p = new;
1094
1095         if (!shared) {
1096                 desc->depth = 0;
1097                 desc->status &= ~(IRQ_DISABLED | IRQ_AUTODETECT | IRQ_WAITING | IRQ_INPROGRESS);
1098                 desc->handler->startup(irq);
1099         }
1100         spin_unlock_irqrestore(&desc->lock,flags);
1101
1102         register_irq_proc(irq);
1103         return 0;
1104 }
1105
1106 static struct proc_dir_entry * root_irq_dir;
1107 static struct proc_dir_entry * irq_dir [NR_IRQS];
1108
1109 #ifdef CONFIG_SMP
1110
1111 static struct proc_dir_entry *smp_affinity_entry[NR_IRQS];
1112
1113 cpumask_t irq_affinity[NR_IRQS] = { [0 ... NR_IRQS-1] = CPU_MASK_ALL };
1114
1115 static int irq_affinity_read_proc(char *page, char **start, off_t off,
1116                         int count, int *eof, void *data)
1117 {
1118         int len = cpumask_scnprintf(page, count, irq_affinity[(long)data]);
1119         if (count - len < 2)
1120                 return -EINVAL;
1121         len += sprintf(page + len, "\n");
1122         return len;
1123 }
1124
1125 int no_irq_affinity;
1126
1127 static int irq_affinity_write_proc(struct file *file, const char __user *buffer,
1128                                         unsigned long count, void *data)
1129 {
1130         int irq = (long)data, full_count = count, err;
1131         cpumask_t new_value, tmp;
1132
1133         if (!irq_desc[irq].handler->set_affinity || no_irq_affinity)
1134                 return -EIO;
1135
1136         err = cpumask_parse(buffer, count, new_value);
1137         if (err)
1138                 return err;
1139
1140         /*
1141          * Do not allow disabling IRQs completely - it's a too easy
1142          * way to make the system unusable accidentally :-) At least
1143          * one online CPU still has to be targeted.
1144          */
1145         cpus_and(tmp, new_value, cpu_online_map);
1146         if (cpus_empty(tmp))
1147                 return -EINVAL;
1148
1149         irq_affinity[irq] = new_value;
1150         irq_desc[irq].handler->set_affinity(irq,
1151                                         cpumask_of_cpu(first_cpu(new_value)));
1152
1153         return full_count;
1154 }
1155
1156 #endif
1157 #define MAX_NAMELEN 10
1158
1159 static void register_irq_proc (unsigned int irq)
1160 {
1161         char name [MAX_NAMELEN];
1162
1163         if (!root_irq_dir || (irq_desc[irq].handler == &no_irq_type) ||
1164                         irq_dir[irq])
1165                 return;
1166
1167         memset(name, 0, MAX_NAMELEN);
1168         sprintf(name, "%d", irq);
1169
1170         /* create /proc/irq/1234 */
1171         irq_dir[irq] = proc_mkdir(name, root_irq_dir);
1172
1173 #ifdef CONFIG_SMP
1174         {
1175                 struct proc_dir_entry *entry;
1176
1177                 /* create /proc/irq/1234/smp_affinity */
1178                 entry = create_proc_entry("smp_affinity", 0600, irq_dir[irq]);
1179
1180                 if (entry) {
1181                         entry->nlink = 1;
1182                         entry->data = (void *)(long)irq;
1183                         entry->read_proc = irq_affinity_read_proc;
1184                         entry->write_proc = irq_affinity_write_proc;
1185                 }
1186
1187                 smp_affinity_entry[irq] = entry;
1188         }
1189 #endif
1190 }
1191
1192 void init_irq_proc (void)
1193 {
1194         int i;
1195
1196         /* create /proc/irq */
1197         root_irq_dir = proc_mkdir("irq", NULL);
1198         create_prof_cpu_mask(root_irq_dir);
1199         /*
1200          * Create entries for all existing IRQs.
1201          */
1202         for (i = 0; i < NR_IRQS; i++)
1203                 register_irq_proc(i);
1204 }
1205
1206
1207 #ifdef CONFIG_IRQSTACKS
1208 /*
1209  * These should really be __section__(".bss.page_aligned") as well, but
1210  * gcc's 3.0 and earlier don't handle that correctly.
1211  */
1212 static char softirq_stack[NR_CPUS * THREAD_SIZE]  __attribute__((__aligned__(THREAD_SIZE)));
1213 static char hardirq_stack[NR_CPUS * THREAD_SIZE]  __attribute__((__aligned__(THREAD_SIZE)));
1214
1215 /*
1216  * allocate per-cpu stacks for hardirq and for softirq processing
1217  */
1218 void irq_ctx_init(int cpu)
1219 {
1220         union irq_ctx *irqctx;
1221
1222         if (hardirq_ctx[cpu])
1223                 return;
1224
1225         irqctx = (union irq_ctx*) &hardirq_stack[cpu*THREAD_SIZE];
1226         irqctx->tinfo.task              = NULL;
1227         irqctx->tinfo.exec_domain       = NULL;
1228         irqctx->tinfo.cpu               = cpu;
1229         irqctx->tinfo.preempt_count     = HARDIRQ_OFFSET;
1230         irqctx->tinfo.addr_limit        = MAKE_MM_SEG(0);
1231
1232         hardirq_ctx[cpu] = irqctx;
1233
1234         irqctx = (union irq_ctx*) &softirq_stack[cpu*THREAD_SIZE];
1235         irqctx->tinfo.task              = NULL;
1236         irqctx->tinfo.exec_domain       = NULL;
1237         irqctx->tinfo.cpu               = cpu;
1238         irqctx->tinfo.preempt_count     = SOFTIRQ_OFFSET;
1239         irqctx->tinfo.addr_limit        = MAKE_MM_SEG(0);
1240
1241         softirq_ctx[cpu] = irqctx;
1242
1243         printk("CPU %u irqstacks, hard=%p soft=%p\n",
1244                 cpu,hardirq_ctx[cpu],softirq_ctx[cpu]);
1245 }
1246
1247 extern asmlinkage void __do_softirq(void);
1248
1249 asmlinkage void do_softirq(void)
1250 {
1251         unsigned long flags;
1252         struct thread_info *curctx;
1253         union irq_ctx *irqctx;
1254         u32 *isp;
1255
1256         if (in_interrupt())
1257                 return;
1258
1259         local_irq_save(flags);
1260
1261         if (local_softirq_pending()) {
1262                 curctx = current_thread_info();
1263                 irqctx = softirq_ctx[smp_processor_id()];
1264                 irqctx->tinfo.task = curctx->task;
1265                 irqctx->tinfo.previous_esp = current_stack_pointer();
1266
1267                 /* build the stack frame on the softirq stack */
1268                 isp = (u32*) ((char*)irqctx + sizeof(*irqctx));
1269
1270
1271                 asm volatile(
1272                         "       xchgl   %%ebx,%%esp     \n"
1273                         "       call    __do_softirq    \n"
1274                         "       movl    %%ebx,%%esp     \n"
1275                         : "=b"(isp)
1276                         : "0"(isp)
1277                         : "memory", "cc", "edx", "ecx", "eax"
1278                 );
1279         }
1280
1281         local_irq_restore(flags);
1282 }
1283
1284 EXPORT_SYMBOL(do_softirq);
1285 #endif