SAFE public projects git trees. - safe/jmp/linux-2.6/blob - arch/sparc64/kernel/irq.c

   1 /* $Id: irq.c,v 1.114 2002/01/11 08:45:38 davem Exp $
   2  * irq.c: UltraSparc IRQ handling/init/registry.
   3  *
   4  * Copyright (C) 1997  David S. Miller  (davem@caip.rutgers.edu)
   5  * Copyright (C) 1998  Eddie C. Dost    (ecd@skynet.be)
   6  * Copyright (C) 1998  Jakub Jelinek    (jj@ultra.linux.cz)
   7  */
   8
   9 #include <linux/config.h>
  10 #include <linux/module.h>
  11 #include <linux/sched.h>
  12 #include <linux/ptrace.h>
  13 #include <linux/errno.h>
  14 #include <linux/kernel_stat.h>
  15 #include <linux/signal.h>
  16 #include <linux/mm.h>
  17 #include <linux/interrupt.h>
  18 #include <linux/slab.h>
  19 #include <linux/random.h>
  20 #include <linux/init.h>
  21 #include <linux/delay.h>
  22 #include <linux/proc_fs.h>
  23 #include <linux/seq_file.h>
  24 #include <linux/bootmem.h>
  25
  26 #include <asm/ptrace.h>
  27 #include <asm/processor.h>
  28 #include <asm/atomic.h>
  29 #include <asm/system.h>
  30 #include <asm/irq.h>
  31 #include <asm/io.h>
  32 #include <asm/sbus.h>
  33 #include <asm/iommu.h>
  34 #include <asm/upa.h>
  35 #include <asm/oplib.h>
  36 #include <asm/timer.h>
  37 #include <asm/smp.h>
  38 #include <asm/starfire.h>
  39 #include <asm/uaccess.h>
  40 #include <asm/cache.h>
  41 #include <asm/cpudata.h>
  42 #include <asm/auxio.h>
  43 #include <asm/head.h>
  44
  45 #ifdef CONFIG_SMP
  46 static void distribute_irqs(void);
  47 #endif
  48
  49 /* UPA nodes send interrupt packet to UltraSparc with first data reg
  50  * value low 5 (7 on Starfire) bits holding the IRQ identifier being
  51  * delivered.  We must translate this into a non-vector IRQ so we can
  52  * set the softint on this cpu.
  53  *
  54  * To make processing these packets efficient and race free we use
  55  * an array of irq buckets below.  The interrupt vector handler in
  56  * entry.S feeds incoming packets into per-cpu pil-indexed lists.
  57  * The IVEC handler does not need to act atomically, the PIL dispatch
  58  * code uses CAS to get an atomic snapshot of the list and clear it
  59  * at the same time.
  60  */
  61
  62 struct ino_bucket ivector_table[NUM_IVECS] __attribute__ ((aligned (SMP_CACHE_BYTES)));
  63
  64 /* This has to be in the main kernel image, it cannot be
  65  * turned into per-cpu data.  The reason is that the main
  66  * kernel image is locked into the TLB and this structure
  67  * is accessed from the vectored interrupt trap handler.  If
  68  * access to this structure takes a TLB miss it could cause
  69  * the 5-level sparc v9 trap stack to overflow.
  70  */
  71 #define irq_work(__cpu) &(trap_block[(__cpu)].irq_worklist)
  72
  73 static struct irqaction timer_irq_action = {
  74         .name = "timer",
  75 };
  76 static struct irqaction *irq_action[NR_IRQS] = { &timer_irq_action, };
  77
  78 /* This only synchronizes entities which modify IRQ handler
  79  * state and some selected user-level spots that want to
  80  * read things in the table.  IRQ handler processing orders
  81  * its' accesses such that no locking is needed.
  82  */
  83 static DEFINE_SPINLOCK(irq_action_lock);
  84
  85 static unsigned int virt_to_real_irq_table[NR_IRQS];
  86 static unsigned char virt_irq_cur = 1;
  87
  88 static unsigned char virt_irq_alloc(unsigned int real_irq)
  89 {
  90         unsigned char ent;
  91
  92         BUILD_BUG_ON(NR_IRQS >= 256);
  93
  94         ent = virt_irq_cur;
  95         if (ent >= NR_IRQS) {
  96                 printk(KERN_ERR "IRQ: Out of virtual IRQs.\n");
  97                 return 0;
  98         }
  99
 100         virt_irq_cur = ent + 1;
 101         virt_to_real_irq_table[ent] = real_irq;
 102
 103         return ent;
 104 }
 105
 106 #if 0 /* Currently unused. */
 107 static unsigned char real_to_virt_irq(unsigned int real_irq)
 108 {
 109         struct ino_bucket *bucket = __bucket(real_irq);
 110
 111         return bucket->virt_irq;
 112 }
 113 #endif
 114
 115 static unsigned int virt_to_real_irq(unsigned char virt_irq)
 116 {
 117         return virt_to_real_irq_table[virt_irq];
 118 }
 119
 120 void irq_install_pre_handler(int virt_irq,
 121                              void (*func)(struct ino_bucket *, void *, void *),
 122                              void *arg1, void *arg2)
 123 {
 124         unsigned int real_irq = virt_to_real_irq(virt_irq);
 125         struct ino_bucket *bucket;
 126         struct irq_desc *d;
 127
 128         if (unlikely(!real_irq))
 129                 return;
 130
 131         bucket = __bucket(real_irq);
 132         d = bucket->irq_info;
 133         d->pre_handler = func;
 134         d->pre_handler_arg1 = arg1;
 135         d->pre_handler_arg2 = arg2;
 136 }
 137
 138 static void register_irq_proc (unsigned int irq);
 139
 140 /*
 141  * Upper 2b of irqaction->flags holds the ino.
 142  * irqaction->mask holds the smp affinity information.
 143  */
 144 #define put_ino_in_irqaction(action, irq) \
 145         action->flags &= 0xffffffffffffUL; \
 146         action->flags |= __irq_ino(irq) << 48;
 147
 148 #define get_ino_in_irqaction(action)    (action->flags >> 48)
 149
 150 #define put_smpaff_in_irqaction(action, smpaff) (action)->mask = (smpaff)
 151 #define get_smpaff_in_irqaction(action)         ((action)->mask)
 152
 153 int show_interrupts(struct seq_file *p, void *v)
 154 {
 155         unsigned long flags;
 156         int i = *(loff_t *) v;
 157         struct irqaction *action;
 158 #ifdef CONFIG_SMP
 159         int j;
 160 #endif
 161
 162         spin_lock_irqsave(&irq_action_lock, flags);
 163         if (i <= NR_IRQS) {
 164                 if (!(action = *(i + irq_action)))
 165                         goto out_unlock;
 166                 seq_printf(p, "%3d: ", i);
 167 #ifndef CONFIG_SMP
 168                 seq_printf(p, "%10u ", kstat_irqs(i));
 169 #else
 170                 for_each_online_cpu(j) {
 171                         seq_printf(p, "%10u ",
 172                                    kstat_cpu(j).irqs[i]);
 173                 }
 174 #endif
 175                 seq_printf(p, " %s", action->name);
 176                 for (action = action->next; action; action = action->next)
 177                         seq_printf(p, ", %s", action->name);
 178                 seq_putc(p, '\n');
 179         }
 180 out_unlock:
 181         spin_unlock_irqrestore(&irq_action_lock, flags);
 182
 183         return 0;
 184 }
 185
 186 extern unsigned long real_hard_smp_processor_id(void);
 187
 188 static unsigned int sun4u_compute_tid(unsigned long imap, unsigned long cpuid)
 189 {
 190         unsigned int tid;
 191
 192         if (this_is_starfire) {
 193                 tid = starfire_translate(imap, cpuid);
 194                 tid <<= IMAP_TID_SHIFT;
 195                 tid &= IMAP_TID_UPA;
 196         } else {
 197                 if (tlb_type == cheetah || tlb_type == cheetah_plus) {
 198                         unsigned long ver;
 199
 200                         __asm__ ("rdpr %%ver, %0" : "=r" (ver));
 201                         if ((ver >> 32UL) == __JALAPENO_ID ||
 202                             (ver >> 32UL) == __SERRANO_ID) {
 203                                 tid = cpuid << IMAP_TID_SHIFT;
 204                                 tid &= IMAP_TID_JBUS;
 205                         } else {
 206                                 unsigned int a = cpuid & 0x1f;
 207                                 unsigned int n = (cpuid >> 5) & 0x1f;
 208
 209                                 tid = ((a << IMAP_AID_SHIFT) |
 210                                        (n << IMAP_NID_SHIFT));
 211                                 tid &= (IMAP_AID_SAFARI |
 212                                         IMAP_NID_SAFARI);;
 213                         }
 214                 } else {
 215                         tid = cpuid << IMAP_TID_SHIFT;
 216                         tid &= IMAP_TID_UPA;
 217                 }
 218         }
 219
 220         return tid;
 221 }
 222
 223 void enable_irq(unsigned int virt_irq)
 224 {
 225         unsigned int real_irq = virt_to_real_irq(virt_irq);
 226         struct ino_bucket *bucket;
 227         unsigned long imap, cpuid;
 228
 229         if (unlikely(!real_irq))
 230                 return;
 231
 232         bucket = __bucket(real_irq);
 233         imap = bucket->imap;
 234         if (unlikely(imap == 0UL))
 235                 return;
 236
 237         preempt_disable();
 238
 239         /* This gets the physical processor ID, even on uniprocessor,
 240          * so we can always program the interrupt target correctly.
 241          */
 242         cpuid = real_hard_smp_processor_id();
 243
 244         if (tlb_type == hypervisor) {
 245                 unsigned int ino = __irq_ino(real_irq);
 246                 int err;
 247
 248                 err = sun4v_intr_settarget(ino, cpuid);
 249                 if (err != HV_EOK)
 250                         printk("sun4v_intr_settarget(%x,%lu): err(%d)\n",
 251                                ino, cpuid, err);
 252                 err = sun4v_intr_setenabled(ino, HV_INTR_ENABLED);
 253                 if (err != HV_EOK)
 254                         printk("sun4v_intr_setenabled(%x): err(%d)\n",
 255                                ino, err);
 256         } else {
 257                 unsigned int tid = sun4u_compute_tid(imap, cpuid);
 258
 259                 /* NOTE NOTE NOTE, IGN and INO are read-only, IGN is a product
 260                  * of this SYSIO's preconfigured IGN in the SYSIO Control
 261                  * Register, the hardware just mirrors that value here.
 262                  * However for Graphics and UPA Slave devices the full
 263                  * IMAP_INR field can be set by the programmer here.
 264                  *
 265                  * Things like FFB can now be handled via the new IRQ
 266                  * mechanism.
 267                  */
 268                 upa_writel(tid | IMAP_VALID, imap);
 269         }
 270
 271         preempt_enable();
 272 }
 273
 274 void disable_irq(unsigned int virt_irq)
 275 {
 276         unsigned int real_irq = virt_to_real_irq(virt_irq);
 277         struct ino_bucket *bucket;
 278         unsigned long imap;
 279
 280         if (unlikely(!real_irq))
 281                 return;
 282
 283         bucket = __bucket(real_irq);
 284         imap = bucket->imap;
 285         if (unlikely(imap == 0UL))
 286                 return;
 287
 288         if (tlb_type == hypervisor) {
 289                 unsigned int ino = __irq_ino(real_irq);
 290                 int err;
 291
 292                 err = sun4v_intr_setenabled(ino, HV_INTR_DISABLED);
 293                 if (err != HV_EOK)
 294                         printk("sun4v_intr_setenabled(%x): "
 295                                "err(%d)\n", ino, err);
 296         } else {
 297                 u32 tmp;
 298
 299                 /* NOTE: We do not want to futz with the IRQ clear registers
 300                  *       and move the state to IDLE, the SCSI code does call
 301                  *       disable_irq() to assure atomicity in the queue cmd
 302                  *       SCSI adapter driver code.  Thus we'd lose interrupts.
 303                  */
 304                 tmp = upa_readl(imap);
 305                 tmp &= ~IMAP_VALID;
 306                 upa_writel(tmp, imap);
 307         }
 308 }
 309
 310 static void build_irq_error(const char *msg, unsigned int ino, int inofixup,
 311                             unsigned long iclr, unsigned long imap,
 312                             struct ino_bucket *bucket)
 313 {
 314         prom_printf("IRQ: INO %04x (%016lx:%016lx) --> "
 315                     "(%d:%016lx:%016lx), halting...\n",
 316                     ino, bucket->iclr, bucket->imap,
 317                     inofixup, iclr, imap);
 318         prom_halt();
 319 }
 320
 321 unsigned int build_irq(int inofixup, unsigned long iclr, unsigned long imap, unsigned char flags)
 322 {
 323         struct ino_bucket *bucket;
 324         int ino;
 325
 326         BUG_ON(tlb_type == hypervisor);
 327
 328         /* RULE: Both must be specified. */
 329         if (iclr == 0UL || imap == 0UL) {
 330                 prom_printf("Invalid build_irq %d %016lx %016lx\n",
 331                             inofixup, iclr, imap);
 332                 prom_halt();
 333         }
 334
 335         ino = (upa_readl(imap) & (IMAP_IGN | IMAP_INO)) + inofixup;
 336         if (ino > NUM_IVECS) {
 337                 prom_printf("Invalid INO %04x (%d:%016lx:%016lx)\n",
 338                             ino, inofixup, iclr, imap);
 339                 prom_halt();
 340         }
 341
 342         bucket = &ivector_table[ino];
 343         if (bucket->flags & IBF_ACTIVE)
 344                 build_irq_error("IRQ: Trying to build active INO bucket.\n",
 345                                 ino, inofixup, iclr, imap, bucket);
 346
 347         if (bucket->irq_info) {
 348                 if (bucket->imap != imap || bucket->iclr != iclr)
 349                         build_irq_error("IRQ: Trying to reinit INO bucket.\n",
 350                                         ino, inofixup, iclr, imap, bucket);
 351
 352                 goto out;
 353         }
 354
 355         bucket->irq_info = kzalloc(sizeof(struct irq_desc), GFP_ATOMIC);
 356         if (!bucket->irq_info) {
 357                 prom_printf("IRQ: Error, kmalloc(irq_desc) failed.\n");
 358                 prom_halt();
 359         }
 360
 361         /* Ok, looks good, set it up.  Don't touch the irq_chain or
 362          * the pending flag.
 363          */
 364         bucket->imap  = imap;
 365         bucket->iclr  = iclr;
 366         if (!bucket->virt_irq)
 367                 bucket->virt_irq = virt_irq_alloc(__irq(bucket));
 368         bucket->flags = flags;
 369
 370 out:
 371         return bucket->virt_irq;
 372 }
 373
 374 unsigned int sun4v_build_irq(u32 devhandle, unsigned int devino, unsigned char flags)
 375 {
 376         struct ino_bucket *bucket;
 377         unsigned long sysino;
 378
 379         sysino = sun4v_devino_to_sysino(devhandle, devino);
 380
 381         bucket = &ivector_table[sysino];
 382
 383         /* Catch accidental accesses to these things.  IMAP/ICLR handling
 384          * is done by hypervisor calls on sun4v platforms, not by direct
 385          * register accesses.
 386          *
 387          * But we need to make them look unique for the disable_irq() logic
 388          * in free_irq().
 389          */
 390         bucket->imap = ~0UL - sysino;
 391         bucket->iclr = ~0UL - sysino;
 392         if (!bucket->virt_irq)
 393                 bucket->virt_irq = virt_irq_alloc(__irq(bucket));
 394         bucket->flags = flags;
 395
 396         bucket->irq_info = kzalloc(sizeof(struct irq_desc), GFP_ATOMIC);
 397         if (!bucket->irq_info) {
 398                 prom_printf("IRQ: Error, kmalloc(irq_desc) failed.\n");
 399                 prom_halt();
 400         }
 401
 402         return bucket->virt_irq;
 403 }
 404
 405 static void atomic_bucket_insert(struct ino_bucket *bucket)
 406 {
 407         unsigned long pstate;
 408         unsigned int *ent;
 409
 410         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 411         __asm__ __volatile__("wrpr %0, %1, %%pstate"
 412                              : : "r" (pstate), "i" (PSTATE_IE));
 413         ent = irq_work(smp_processor_id());
 414         bucket->irq_chain = *ent;
 415         *ent = __irq(bucket);
 416         __asm__ __volatile__("wrpr %0, 0x0, %%pstate" : : "r" (pstate));
 417 }
 418
 419 static int check_irq_sharing(int pil, unsigned long irqflags)
 420 {
 421         struct irqaction *action;
 422
 423         action = *(irq_action + pil);
 424         if (action) {
 425                 if (!(action->flags & SA_SHIRQ) || !(irqflags & SA_SHIRQ))
 426                         return -EBUSY;
 427         }
 428         return 0;
 429 }
 430
 431 static void append_irq_action(int pil, struct irqaction *action)
 432 {
 433         struct irqaction **pp = irq_action + pil;
 434
 435         while (*pp)
 436                 pp = &((*pp)->next);
 437         *pp = action;
 438 }
 439
 440 static struct irqaction *get_action_slot(struct ino_bucket *bucket)
 441 {
 442         struct irq_desc *desc = bucket->irq_info;
 443         int max_irq, i;
 444
 445         max_irq = 1;
 446         if (bucket->flags & IBF_PCI)
 447                 max_irq = MAX_IRQ_DESC_ACTION;
 448         for (i = 0; i < max_irq; i++) {
 449                 struct irqaction *p = &desc->action[i];
 450                 u32 mask = (1 << i);
 451
 452                 if (desc->action_active_mask & mask)
 453                         continue;
 454
 455                 desc->action_active_mask |= mask;
 456                 return p;
 457         }
 458         return NULL;
 459 }
 460
 461 int request_irq(unsigned int virt_irq,
 462                 irqreturn_t (*handler)(int, void *, struct pt_regs *),
 463                 unsigned long irqflags, const char *name, void *dev_id)
 464 {
 465         struct irqaction *action;
 466         struct ino_bucket *bucket;
 467         unsigned long flags;
 468         unsigned int real_irq;
 469         int pending = 0;
 470
 471         real_irq = virt_to_real_irq(virt_irq);
 472         if (unlikely(!real_irq))
 473                 return -EINVAL;
 474
 475         if (unlikely(!handler))
 476                 return -EINVAL;
 477
 478         bucket = __bucket(real_irq);
 479         if (unlikely(!bucket->irq_info))
 480                 return -ENODEV;
 481
 482         if (irqflags & SA_SAMPLE_RANDOM) {
 483                 /*
 484                  * This function might sleep, we want to call it first,
 485                  * outside of the atomic block.
 486                  * Yes, this might clear the entropy pool if the wrong
 487                  * driver is attempted to be loaded, without actually
 488                  * installing a new handler, but is this really a problem,
 489                  * only the sysadmin is able to do this.
 490                  */
 491                 rand_initialize_irq(virt_irq);
 492         }
 493
 494         spin_lock_irqsave(&irq_action_lock, flags);
 495
 496         if (check_irq_sharing(virt_irq, irqflags)) {
 497                 spin_unlock_irqrestore(&irq_action_lock, flags);
 498                 return -EBUSY;
 499         }
 500
 501         action = get_action_slot(bucket);
 502         if (!action) {
 503                 spin_unlock_irqrestore(&irq_action_lock, flags);
 504                 return -ENOMEM;
 505         }
 506
 507         bucket->flags |= IBF_ACTIVE;
 508         pending = bucket->pending;
 509         if (pending)
 510                 bucket->pending = 0;
 511
 512         action->handler = handler;
 513         action->flags = irqflags;
 514         action->name = name;
 515         action->next = NULL;
 516         action->dev_id = dev_id;
 517         put_ino_in_irqaction(action, __irq_ino(real_irq));
 518         put_smpaff_in_irqaction(action, CPU_MASK_NONE);
 519
 520         append_irq_action(virt_irq, action);
 521
 522         enable_irq(virt_irq);
 523
 524         /* We ate the IVEC already, this makes sure it does not get lost. */
 525         if (pending) {
 526                 atomic_bucket_insert(bucket);
 527                 set_softint(1 << PIL_DEVICE_IRQ);
 528         }
 529
 530         spin_unlock_irqrestore(&irq_action_lock, flags);
 531
 532         register_irq_proc(virt_irq);
 533
 534 #ifdef CONFIG_SMP
 535         distribute_irqs();
 536 #endif
 537         return 0;
 538 }
 539
 540 EXPORT_SYMBOL(request_irq);
 541
 542 static struct irqaction *unlink_irq_action(unsigned int virt_irq, void *dev_id)
 543 {
 544         struct irqaction *action, **pp;
 545
 546         pp = irq_action + virt_irq;
 547         action = *pp;
 548         if (unlikely(!action))
 549                 return NULL;
 550
 551         if (unlikely(!action->handler)) {
 552                 printk("Freeing free IRQ %d\n", virt_irq);
 553                 return NULL;
 554         }
 555
 556         while (action && action->dev_id != dev_id) {
 557                 pp = &action->next;
 558                 action = *pp;
 559         }
 560
 561         if (likely(action))
 562                 *pp = action->next;
 563
 564         return action;
 565 }
 566
 567 void free_irq(unsigned int virt_irq, void *dev_id)
 568 {
 569         struct irqaction *action;
 570         struct ino_bucket *bucket;
 571         struct irq_desc *desc;
 572         unsigned long flags;
 573         unsigned int real_irq;
 574         int ent, i;
 575
 576         real_irq = virt_to_real_irq(virt_irq);
 577         if (unlikely(!real_irq))
 578                 return;
 579
 580         spin_lock_irqsave(&irq_action_lock, flags);
 581
 582         action = unlink_irq_action(virt_irq, dev_id);
 583
 584         spin_unlock_irqrestore(&irq_action_lock, flags);
 585
 586         if (unlikely(!action))
 587                 return;
 588
 589         synchronize_irq(virt_irq);
 590
 591         spin_lock_irqsave(&irq_action_lock, flags);
 592
 593         bucket = __bucket(real_irq);
 594         desc = bucket->irq_info;
 595
 596         for (i = 0; i < MAX_IRQ_DESC_ACTION; i++) {
 597                 struct irqaction *p = &desc->action[i];
 598
 599                 if (p == action) {
 600                         desc->action_active_mask &= ~(1 << i);
 601                         break;
 602                 }
 603         }
 604
 605         if (!desc->action_active_mask) {
 606                 unsigned long imap = bucket->imap;
 607
 608                 /* This unique interrupt source is now inactive. */
 609                 bucket->flags &= ~IBF_ACTIVE;
 610
 611                 /* See if any other buckets share this bucket's IMAP
 612                  * and are still active.
 613                  */
 614                 for (ent = 0; ent < NUM_IVECS; ent++) {
 615                         struct ino_bucket *bp = &ivector_table[ent];
 616                         if (bp != bucket        &&
 617                             bp->imap == imap    &&
 618                             (bp->flags & IBF_ACTIVE) != 0)
 619                                 break;
 620                 }
 621
 622                 /* Only disable when no other sub-irq levels of
 623                  * the same IMAP are active.
 624                  */
 625                 if (ent == NUM_IVECS)
 626                         disable_irq(virt_irq);
 627         }
 628
 629         spin_unlock_irqrestore(&irq_action_lock, flags);
 630 }
 631
 632 EXPORT_SYMBOL(free_irq);
 633
 634 #ifdef CONFIG_SMP
 635 void synchronize_irq(unsigned int virt_irq)
 636 {
 637         unsigned int real_irq = virt_to_real_irq(virt_irq);
 638         struct ino_bucket *bucket;
 639
 640         if (unlikely(!real_irq))
 641                 return;
 642
 643         bucket = __bucket(real_irq);
 644 #if 0
 645         /* The following is how I wish I could implement this.
 646          * Unfortunately the ICLR registers are read-only, you can
 647          * only write ICLR_foo values to them.  To get the current
 648          * IRQ status you would need to get at the IRQ diag registers
 649          * in the PCI/SBUS controller and the layout of those vary
 650          * from one controller to the next, sigh... -DaveM
 651          */
 652         unsigned long iclr = bucket->iclr;
 653
 654         while (1) {
 655                 u32 tmp = upa_readl(iclr);
 656
 657                 if (tmp == ICLR_TRANSMIT ||
 658                     tmp == ICLR_PENDING) {
 659                         cpu_relax();
 660                         continue;
 661                 }
 662                 break;
 663         }
 664 #else
 665         /* So we have to do this with a INPROGRESS bit just like x86.  */
 666         while (bucket->flags & IBF_INPROGRESS)
 667                 cpu_relax();
 668 #endif
 669 }
 670 #endif /* CONFIG_SMP */
 671
 672 static void process_bucket(struct ino_bucket *bp, struct pt_regs *regs)
 673 {
 674         struct irq_desc *desc = bp->irq_info;
 675         unsigned char flags = bp->flags;
 676         u32 action_mask, i;
 677         int random;
 678
 679         bp->flags |= IBF_INPROGRESS;
 680
 681         if (unlikely(!(flags & IBF_ACTIVE))) {
 682                 bp->pending = 1;
 683                 goto out;
 684         }
 685
 686         if (desc->pre_handler)
 687                 desc->pre_handler(bp,
 688                                   desc->pre_handler_arg1,
 689                                   desc->pre_handler_arg2);
 690
 691         action_mask = desc->action_active_mask;
 692         random = 0;
 693         for (i = 0; i < MAX_IRQ_DESC_ACTION; i++) {
 694                 struct irqaction *p = &desc->action[i];
 695                 u32 mask = (1 << i);
 696
 697                 if (!(action_mask & mask))
 698                         continue;
 699
 700                 action_mask &= ~mask;
 701
 702                 if (p->handler(bp->virt_irq, p->dev_id, regs) == IRQ_HANDLED)
 703                         random |= p->flags;
 704
 705                 if (!action_mask)
 706                         break;
 707         }
 708
 709         if (tlb_type == hypervisor) {
 710                 unsigned int ino = __irq_ino(bp);
 711                 int err;
 712
 713                 err = sun4v_intr_setstate(ino, HV_INTR_STATE_IDLE);
 714                 if (err != HV_EOK)
 715                         printk("sun4v_intr_setstate(%x): "
 716                                "err(%d)\n", ino, err);
 717         } else {
 718                 upa_writel(ICLR_IDLE, bp->iclr);
 719         }
 720
 721         /* Test and add entropy */
 722         if (random & SA_SAMPLE_RANDOM)
 723                 add_interrupt_randomness(bp->virt_irq);
 724 out:
 725         bp->flags &= ~IBF_INPROGRESS;
 726 }
 727
 728 #ifndef CONFIG_SMP
 729 extern irqreturn_t timer_interrupt(int, void *, struct pt_regs *);
 730
 731 void timer_irq(int irq, struct pt_regs *regs)
 732 {
 733         unsigned long clr_mask = 1 << irq;
 734         unsigned long tick_mask = tick_ops->softint_mask;
 735
 736         if (get_softint() & tick_mask) {
 737                 irq = 0;
 738                 clr_mask = tick_mask;
 739         }
 740         clear_softint(clr_mask);
 741
 742         irq_enter();
 743         kstat_this_cpu.irqs[0]++;
 744         timer_interrupt(irq, NULL, regs);
 745         irq_exit();
 746 }
 747 #endif
 748
 749 void handler_irq(int irq, struct pt_regs *regs)
 750 {
 751         struct ino_bucket *bp;
 752         int cpu = smp_processor_id();
 753
 754         /* XXX at this point we should be able to assert that
 755          * XXX irq is PIL_DEVICE_IRQ...
 756          */
 757         clear_softint(1 << irq);
 758
 759         irq_enter();
 760
 761         /* Sliiiick... */
 762         bp = __bucket(xchg32(irq_work(cpu), 0));
 763         while (bp) {
 764                 struct ino_bucket *nbp = __bucket(bp->irq_chain);
 765
 766                 kstat_this_cpu.irqs[bp->virt_irq]++;
 767
 768                 bp->irq_chain = 0;
 769                 process_bucket(bp, regs);
 770                 bp = nbp;
 771         }
 772         irq_exit();
 773 }
 774
 775 #ifdef CONFIG_BLK_DEV_FD
 776 extern irqreturn_t floppy_interrupt(int, void *, struct pt_regs *);
 777
 778 /* XXX No easy way to include asm/floppy.h XXX */
 779 extern unsigned char *pdma_vaddr;
 780 extern unsigned long pdma_size;
 781 extern volatile int doing_pdma;
 782 extern unsigned long fdc_status;
 783
 784 irqreturn_t sparc_floppy_irq(int irq, void *dev_cookie, struct pt_regs *regs)
 785 {
 786         if (likely(doing_pdma)) {
 787                 void __iomem *stat = (void __iomem *) fdc_status;
 788                 unsigned char *vaddr = pdma_vaddr;
 789                 unsigned long size = pdma_size;
 790                 u8 val;
 791
 792                 while (size) {
 793                         val = readb(stat);
 794                         if (unlikely(!(val & 0x80))) {
 795                                 pdma_vaddr = vaddr;
 796                                 pdma_size = size;
 797                                 return IRQ_HANDLED;
 798                         }
 799                         if (unlikely(!(val & 0x20))) {
 800                                 pdma_vaddr = vaddr;
 801                                 pdma_size = size;
 802                                 doing_pdma = 0;
 803                                 goto main_interrupt;
 804                         }
 805                         if (val & 0x40) {
 806                                 /* read */
 807                                 *vaddr++ = readb(stat + 1);
 808                         } else {
 809                                 unsigned char data = *vaddr++;
 810
 811                                 /* write */
 812                                 writeb(data, stat + 1);
 813                         }
 814                         size--;
 815                 }
 816
 817                 pdma_vaddr = vaddr;
 818                 pdma_size = size;
 819
 820                 /* Send Terminal Count pulse to floppy controller. */
 821                 val = readb(auxio_register);
 822                 val |= AUXIO_AUX1_FTCNT;
 823                 writeb(val, auxio_register);
 824                 val &= ~AUXIO_AUX1_FTCNT;
 825                 writeb(val, auxio_register);
 826
 827                 doing_pdma = 0;
 828         }
 829
 830 main_interrupt:
 831         return floppy_interrupt(irq, dev_cookie, regs);
 832 }
 833 EXPORT_SYMBOL(sparc_floppy_irq);
 834 #endif
 835
 836 /* We really don't need these at all on the Sparc.  We only have
 837  * stubs here because they are exported to modules.
 838  */
 839 unsigned long probe_irq_on(void)
 840 {
 841         return 0;
 842 }
 843
 844 EXPORT_SYMBOL(probe_irq_on);
 845
 846 int probe_irq_off(unsigned long mask)
 847 {
 848         return 0;
 849 }
 850
 851 EXPORT_SYMBOL(probe_irq_off);
 852
 853 #ifdef CONFIG_SMP
 854 static int retarget_one_irq(struct irqaction *p, int goal_cpu)
 855 {
 856         struct ino_bucket *bucket = get_ino_in_irqaction(p) + ivector_table;
 857
 858         while (!cpu_online(goal_cpu)) {
 859                 if (++goal_cpu >= NR_CPUS)
 860                         goal_cpu = 0;
 861         }
 862
 863         if (tlb_type == hypervisor) {
 864                 unsigned int ino = __irq_ino(bucket);
 865
 866                 sun4v_intr_settarget(ino, goal_cpu);
 867                 sun4v_intr_setenabled(ino, HV_INTR_ENABLED);
 868         } else {
 869                 unsigned long imap = bucket->imap;
 870                 unsigned int tid = sun4u_compute_tid(imap, goal_cpu);
 871
 872                 upa_writel(tid | IMAP_VALID, imap);
 873         }
 874
 875         do {
 876                 if (++goal_cpu >= NR_CPUS)
 877                         goal_cpu = 0;
 878         } while (!cpu_online(goal_cpu));
 879
 880         return goal_cpu;
 881 }
 882
 883 /* Called from request_irq. */
 884 static void distribute_irqs(void)
 885 {
 886         unsigned long flags;
 887         int cpu, level;
 888
 889         spin_lock_irqsave(&irq_action_lock, flags);
 890         cpu = 0;
 891
 892         for (level = 1; level < NR_IRQS; level++) {
 893                 struct irqaction *p = irq_action[level];
 894
 895                 while(p) {
 896                         cpu = retarget_one_irq(p, cpu);
 897                         p = p->next;
 898                 }
 899         }
 900         spin_unlock_irqrestore(&irq_action_lock, flags);
 901 }
 902 #endif
 903
 904 struct sun5_timer {
 905         u64     count0;
 906         u64     limit0;
 907         u64     count1;
 908         u64     limit1;
 909 };
 910
 911 static struct sun5_timer *prom_timers;
 912 static u64 prom_limit0, prom_limit1;
 913
 914 static void map_prom_timers(void)
 915 {
 916         unsigned int addr[3];
 917         int tnode, err;
 918
 919         /* PROM timer node hangs out in the top level of device siblings... */
 920         tnode = prom_finddevice("/counter-timer");
 921
 922         /* Assume if node is not present, PROM uses different tick mechanism
 923          * which we should not care about.
 924          */
 925         if (tnode == 0 || tnode == -1) {
 926                 prom_timers = (struct sun5_timer *) 0;
 927                 return;
 928         }
 929
 930         /* If PROM is really using this, it must be mapped by him. */
 931         err = prom_getproperty(tnode, "address", (char *)addr, sizeof(addr));
 932         if (err == -1) {
 933                 prom_printf("PROM does not have timer mapped, trying to continue.\n");
 934                 prom_timers = (struct sun5_timer *) 0;
 935                 return;
 936         }
 937         prom_timers = (struct sun5_timer *) ((unsigned long)addr[0]);
 938 }
 939
 940 static void kill_prom_timer(void)
 941 {
 942         if (!prom_timers)
 943                 return;
 944
 945         /* Save them away for later. */
 946         prom_limit0 = prom_timers->limit0;
 947         prom_limit1 = prom_timers->limit1;
 948
 949         /* Just as in sun4c/sun4m PROM uses timer which ticks at IRQ 14.
 950          * We turn both off here just to be paranoid.
 951          */
 952         prom_timers->limit0 = 0;
 953         prom_timers->limit1 = 0;
 954
 955         /* Wheee, eat the interrupt packet too... */
 956         __asm__ __volatile__(
 957 "       mov     0x40, %%g2\n"
 958 "       ldxa    [%%g0] %0, %%g1\n"
 959 "       ldxa    [%%g2] %1, %%g1\n"
 960 "       stxa    %%g0, [%%g0] %0\n"
 961 "       membar  #Sync\n"
 962         : /* no outputs */
 963         : "i" (ASI_INTR_RECEIVE), "i" (ASI_INTR_R)
 964         : "g1", "g2");
 965 }
 966
 967 void init_irqwork_curcpu(void)
 968 {
 969         int cpu = hard_smp_processor_id();
 970
 971         trap_block[cpu].irq_worklist = 0;
 972 }
 973
 974 static void __cpuinit register_one_mondo(unsigned long paddr, unsigned long type)
 975 {
 976         unsigned long num_entries = 128;
 977         unsigned long status;
 978
 979         status = sun4v_cpu_qconf(type, paddr, num_entries);
 980         if (status != HV_EOK) {
 981                 prom_printf("SUN4V: sun4v_cpu_qconf(%lu:%lx:%lu) failed, "
 982                             "err %lu\n", type, paddr, num_entries, status);
 983                 prom_halt();
 984         }
 985 }
 986
 987 static void __cpuinit sun4v_register_mondo_queues(int this_cpu)
 988 {
 989         struct trap_per_cpu *tb = &trap_block[this_cpu];
 990
 991         register_one_mondo(tb->cpu_mondo_pa, HV_CPU_QUEUE_CPU_MONDO);
 992         register_one_mondo(tb->dev_mondo_pa, HV_CPU_QUEUE_DEVICE_MONDO);
 993         register_one_mondo(tb->resum_mondo_pa, HV_CPU_QUEUE_RES_ERROR);
 994         register_one_mondo(tb->nonresum_mondo_pa, HV_CPU_QUEUE_NONRES_ERROR);
 995 }
 996
 997 static void __cpuinit alloc_one_mondo(unsigned long *pa_ptr, int use_bootmem)
 998 {
 999         void *page;
1000
1001         if (use_bootmem)
1002                 page = alloc_bootmem_low_pages(PAGE_SIZE);
1003         else
1004                 page = (void *) get_zeroed_page(GFP_ATOMIC);
1005
1006         if (!page) {
1007                 prom_printf("SUN4V: Error, cannot allocate mondo queue.\n");
1008                 prom_halt();
1009         }
1010
1011         *pa_ptr = __pa(page);
1012 }
1013
1014 static void __cpuinit alloc_one_kbuf(unsigned long *pa_ptr, int use_bootmem)
1015 {
1016         void *page;
1017
1018         if (use_bootmem)
1019                 page = alloc_bootmem_low_pages(PAGE_SIZE);
1020         else
1021                 page = (void *) get_zeroed_page(GFP_ATOMIC);
1022
1023         if (!page) {
1024                 prom_printf("SUN4V: Error, cannot allocate kbuf page.\n");
1025                 prom_halt();
1026         }
1027
1028         *pa_ptr = __pa(page);
1029 }
1030
1031 static void __cpuinit init_cpu_send_mondo_info(struct trap_per_cpu *tb, int use_bootmem)
1032 {
1033 #ifdef CONFIG_SMP
1034         void *page;
1035
1036         BUILD_BUG_ON((NR_CPUS * sizeof(u16)) > (PAGE_SIZE - 64));
1037
1038         if (use_bootmem)
1039                 page = alloc_bootmem_low_pages(PAGE_SIZE);
1040         else
1041                 page = (void *) get_zeroed_page(GFP_ATOMIC);
1042
1043         if (!page) {
1044                 prom_printf("SUN4V: Error, cannot allocate cpu mondo page.\n");
1045                 prom_halt();
1046         }
1047
1048         tb->cpu_mondo_block_pa = __pa(page);
1049         tb->cpu_list_pa = __pa(page + 64);
1050 #endif
1051 }
1052
1053 /* Allocate and register the mondo and error queues for this cpu.  */
1054 void __cpuinit sun4v_init_mondo_queues(int use_bootmem, int cpu, int alloc, int load)
1055 {
1056         struct trap_per_cpu *tb = &trap_block[cpu];
1057
1058         if (alloc) {
1059                 alloc_one_mondo(&tb->cpu_mondo_pa, use_bootmem);
1060                 alloc_one_mondo(&tb->dev_mondo_pa, use_bootmem);
1061                 alloc_one_mondo(&tb->resum_mondo_pa, use_bootmem);
1062                 alloc_one_kbuf(&tb->resum_kernel_buf_pa, use_bootmem);
1063                 alloc_one_mondo(&tb->nonresum_mondo_pa, use_bootmem);
1064                 alloc_one_kbuf(&tb->nonresum_kernel_buf_pa, use_bootmem);
1065
1066                 init_cpu_send_mondo_info(tb, use_bootmem);
1067         }
1068
1069         if (load) {
1070                 if (cpu != hard_smp_processor_id()) {
1071                         prom_printf("SUN4V: init mondo on cpu %d not %d\n",
1072                                     cpu, hard_smp_processor_id());
1073                         prom_halt();
1074                 }
1075                 sun4v_register_mondo_queues(cpu);
1076         }
1077 }
1078
1079 /* Only invoked on boot processor. */
1080 void __init init_IRQ(void)
1081 {
1082         map_prom_timers();
1083         kill_prom_timer();
1084         memset(&ivector_table[0], 0, sizeof(ivector_table));
1085
1086         if (tlb_type == hypervisor)
1087                 sun4v_init_mondo_queues(1, hard_smp_processor_id(), 1, 1);
1088
1089         /* We need to clear any IRQ's pending in the soft interrupt
1090          * registers, a spurious one could be left around from the
1091          * PROM timer which we just disabled.
1092          */
1093         clear_softint(get_softint());
1094
1095         /* Now that ivector table is initialized, it is safe
1096          * to receive IRQ vector traps.  We will normally take
1097          * one or two right now, in case some device PROM used
1098          * to boot us wants to speak to us.  We just ignore them.
1099          */
1100         __asm__ __volatile__("rdpr      %%pstate, %%g1\n\t"
1101                              "or        %%g1, %0, %%g1\n\t"
1102                              "wrpr      %%g1, 0x0, %%pstate"
1103                              : /* No outputs */
1104                              : "i" (PSTATE_IE)
1105                              : "g1");
1106 }
1107
1108 static struct proc_dir_entry *root_irq_dir;
1109 static struct proc_dir_entry *irq_dir[NR_IRQS];
1110
1111 #ifdef CONFIG_SMP
1112
1113 static int irq_affinity_read_proc(char *page, char **start, off_t off,
1114                                   int count, int *eof, void *data)
1115 {
1116         struct ino_bucket *bp = ivector_table + (long)data;
1117         struct irq_desc *desc = bp->irq_info;
1118         struct irqaction *ap = desc->action;
1119         cpumask_t mask;
1120         int len;
1121
1122         mask = get_smpaff_in_irqaction(ap);
1123         if (cpus_empty(mask))
1124                 mask = cpu_online_map;
1125
1126         len = cpumask_scnprintf(page, count, mask);
1127         if (count - len < 2)
1128                 return -EINVAL;
1129         len += sprintf(page + len, "\n");
1130         return len;
1131 }
1132
1133 static inline void set_intr_affinity(int virt_irq, cpumask_t hw_aff)
1134 {
1135         struct ino_bucket *bp;
1136         struct irq_desc *desc;
1137         struct irqaction *ap;
1138         unsigned int real_irq;
1139
1140         real_irq = virt_to_real_irq(virt_irq);
1141         if (unlikely(!real_irq))
1142                 return;
1143
1144         bp = __bucket(real_irq);
1145         desc = bp->irq_info;
1146         ap = desc->action;
1147
1148         /* Users specify affinity in terms of hw cpu ids.
1149          * As soon as we do this, handler_irq() might see and take action.
1150          */
1151         put_smpaff_in_irqaction(ap, hw_aff);
1152
1153         /* Migration is simply done by the next cpu to service this
1154          * interrupt.
1155          *
1156          * XXX Broken, this doesn't happen anymore...
1157          */
1158 }
1159
1160 static int irq_affinity_write_proc(struct file *file,
1161                                    const char __user *buffer,
1162                                    unsigned long count, void *data)
1163 {
1164         int virt_irq = (long) data, full_count = count, err;
1165         cpumask_t new_value;
1166
1167         err = cpumask_parse(buffer, count, new_value);
1168
1169         /*
1170          * Do not allow disabling IRQs completely - it's a too easy
1171          * way to make the system unusable accidentally :-) At least
1172          * one online CPU still has to be targeted.
1173          */
1174         cpus_and(new_value, new_value, cpu_online_map);
1175         if (cpus_empty(new_value))
1176                 return -EINVAL;
1177
1178         set_intr_affinity(virt_irq, new_value);
1179
1180         return full_count;
1181 }
1182
1183 #endif
1184
1185 #define MAX_NAMELEN 10
1186
1187 static void register_irq_proc(unsigned int virt_irq)
1188 {
1189         char name [MAX_NAMELEN];
1190
1191         if (!root_irq_dir || irq_dir[virt_irq])
1192                 return;
1193
1194         memset(name, 0, MAX_NAMELEN);
1195         sprintf(name, "%d", virt_irq);
1196
1197         /* create /proc/irq/1234 */
1198         irq_dir[virt_irq] = proc_mkdir(name, root_irq_dir);
1199
1200 #ifdef CONFIG_SMP
1201         /* XXX SMP affinity not supported on starfire yet. */
1202         if (this_is_starfire == 0) {
1203                 struct proc_dir_entry *entry;
1204
1205                 /* create /proc/irq/1234/smp_affinity */
1206                 entry = create_proc_entry("smp_affinity", 0600, irq_dir[irq]);
1207
1208                 if (entry) {
1209                         entry->nlink = 1;
1210                         entry->data = (void *)(long)virt_irq;
1211                         entry->read_proc = irq_affinity_read_proc;
1212                         entry->write_proc = irq_affinity_write_proc;
1213                 }
1214         }
1215 #endif
1216 }
1217
1218 void init_irq_proc(void)
1219 {
1220         /* create /proc/irq */
1221         root_irq_dir = proc_mkdir("irq", NULL);
1222 }
1223