git.oblomov.eu Git - 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 struct irq_work_struct {
  72         unsigned int    irq_worklists[16];
  73 };
  74 struct irq_work_struct __irq_work[NR_CPUS];
  75 #define irq_work(__cpu, __pil)  &(__irq_work[(__cpu)].irq_worklists[(__pil)])
  76
  77 static struct irqaction *irq_action[NR_IRQS+1];
  78
  79 /* This only synchronizes entities which modify IRQ handler
  80  * state and some selected user-level spots that want to
  81  * read things in the table.  IRQ handler processing orders
  82  * its' accesses such that no locking is needed.
  83  */
  84 static DEFINE_SPINLOCK(irq_action_lock);
  85
  86 static void register_irq_proc (unsigned int irq);
  87
  88 /*
  89  * Upper 2b of irqaction->flags holds the ino.
  90  * irqaction->mask holds the smp affinity information.
  91  */
  92 #define put_ino_in_irqaction(action, irq) \
  93         action->flags &= 0xffffffffffffUL; \
  94         if (__bucket(irq) == &pil0_dummy_bucket) \
  95                 action->flags |= 0xdeadUL << 48;  \
  96         else \
  97                 action->flags |= __irq_ino(irq) << 48;
  98 #define get_ino_in_irqaction(action)    (action->flags >> 48)
  99
 100 #define put_smpaff_in_irqaction(action, smpaff) (action)->mask = (smpaff)
 101 #define get_smpaff_in_irqaction(action)         ((action)->mask)
 102
 103 int show_interrupts(struct seq_file *p, void *v)
 104 {
 105         unsigned long flags;
 106         int i = *(loff_t *) v;
 107         struct irqaction *action;
 108 #ifdef CONFIG_SMP
 109         int j;
 110 #endif
 111
 112         spin_lock_irqsave(&irq_action_lock, flags);
 113         if (i <= NR_IRQS) {
 114                 if (!(action = *(i + irq_action)))
 115                         goto out_unlock;
 116                 seq_printf(p, "%3d: ", i);
 117 #ifndef CONFIG_SMP
 118                 seq_printf(p, "%10u ", kstat_irqs(i));
 119 #else
 120                 for_each_online_cpu(j) {
 121                         seq_printf(p, "%10u ",
 122                                    kstat_cpu(j).irqs[i]);
 123                 }
 124 #endif
 125                 seq_printf(p, " %s:%lx", action->name,
 126                            get_ino_in_irqaction(action));
 127                 for (action = action->next; action; action = action->next) {
 128                         seq_printf(p, ", %s:%lx", action->name,
 129                                    get_ino_in_irqaction(action));
 130                 }
 131                 seq_putc(p, '\n');
 132         }
 133 out_unlock:
 134         spin_unlock_irqrestore(&irq_action_lock, flags);
 135
 136         return 0;
 137 }
 138
 139 extern unsigned long real_hard_smp_processor_id(void);
 140
 141 static unsigned int sun4u_compute_tid(unsigned long imap, unsigned long cpuid)
 142 {
 143         unsigned int tid;
 144
 145         if (this_is_starfire) {
 146                 tid = starfire_translate(imap, cpuid);
 147                 tid <<= IMAP_TID_SHIFT;
 148                 tid &= IMAP_TID_UPA;
 149         } else {
 150                 if (tlb_type == cheetah || tlb_type == cheetah_plus) {
 151                         unsigned long ver;
 152
 153                         __asm__ ("rdpr %%ver, %0" : "=r" (ver));
 154                         if ((ver >> 32UL) == __JALAPENO_ID ||
 155                             (ver >> 32UL) == __SERRANO_ID) {
 156                                 tid = cpuid << IMAP_TID_SHIFT;
 157                                 tid &= IMAP_TID_JBUS;
 158                         } else {
 159                                 unsigned int a = cpuid & 0x1f;
 160                                 unsigned int n = (cpuid >> 5) & 0x1f;
 161
 162                                 tid = ((a << IMAP_AID_SHIFT) |
 163                                        (n << IMAP_NID_SHIFT));
 164                                 tid &= (IMAP_AID_SAFARI |
 165                                         IMAP_NID_SAFARI);;
 166                         }
 167                 } else {
 168                         tid = cpuid << IMAP_TID_SHIFT;
 169                         tid &= IMAP_TID_UPA;
 170                 }
 171         }
 172
 173         return tid;
 174 }
 175
 176 /* Now these are always passed a true fully specified sun4u INO. */
 177 void enable_irq(unsigned int irq)
 178 {
 179         struct ino_bucket *bucket = __bucket(irq);
 180         unsigned long imap, cpuid;
 181
 182         imap = bucket->imap;
 183         if (imap == 0UL)
 184                 return;
 185
 186         preempt_disable();
 187
 188         /* This gets the physical processor ID, even on uniprocessor,
 189          * so we can always program the interrupt target correctly.
 190          */
 191         cpuid = real_hard_smp_processor_id();
 192
 193         if (tlb_type == hypervisor) {
 194                 unsigned int ino = __irq_ino(irq);
 195                 int err;
 196
 197                 err = sun4v_intr_settarget(ino, cpuid);
 198                 if (err != HV_EOK)
 199                         printk("sun4v_intr_settarget(%x,%lu): err(%d)\n",
 200                                ino, cpuid, err);
 201                 err = sun4v_intr_setenabled(ino, HV_INTR_ENABLED);
 202                 if (err != HV_EOK)
 203                         printk("sun4v_intr_setenabled(%x): err(%d)\n",
 204                                ino, err);
 205         } else {
 206                 unsigned int tid = sun4u_compute_tid(imap, cpuid);
 207
 208                 /* NOTE NOTE NOTE, IGN and INO are read-only, IGN is a product
 209                  * of this SYSIO's preconfigured IGN in the SYSIO Control
 210                  * Register, the hardware just mirrors that value here.
 211                  * However for Graphics and UPA Slave devices the full
 212                  * IMAP_INR field can be set by the programmer here.
 213                  *
 214                  * Things like FFB can now be handled via the new IRQ
 215                  * mechanism.
 216                  */
 217                 upa_writel(tid | IMAP_VALID, imap);
 218         }
 219
 220         preempt_enable();
 221 }
 222
 223 /* This now gets passed true ino's as well. */
 224 void disable_irq(unsigned int irq)
 225 {
 226         struct ino_bucket *bucket = __bucket(irq);
 227         unsigned long imap;
 228
 229         imap = bucket->imap;
 230         if (imap != 0UL) {
 231                 if (tlb_type == hypervisor) {
 232                         unsigned int ino = __irq_ino(irq);
 233                         int err;
 234
 235                         err = sun4v_intr_setenabled(ino, HV_INTR_DISABLED);
 236                         if (err != HV_EOK)
 237                                 printk("sun4v_intr_setenabled(%x): "
 238                                        "err(%d)\n", ino, err);
 239                 } else {
 240                         u32 tmp;
 241
 242                         /* NOTE: We do not want to futz with the IRQ clear registers
 243                          *       and move the state to IDLE, the SCSI code does call
 244                          *       disable_irq() to assure atomicity in the queue cmd
 245                          *       SCSI adapter driver code.  Thus we'd lose interrupts.
 246                          */
 247                         tmp = upa_readl(imap);
 248                         tmp &= ~IMAP_VALID;
 249                         upa_writel(tmp, imap);
 250                 }
 251         }
 252 }
 253
 254 /* The timer is the one "weird" interrupt which is generated by
 255  * the CPU %tick register and not by some normal vectored interrupt
 256  * source.  To handle this special case, we use this dummy INO bucket.
 257  */
 258 static struct irq_desc pil0_dummy_desc;
 259 static struct ino_bucket pil0_dummy_bucket = {
 260         .irq_info       =       &pil0_dummy_desc,
 261 };
 262
 263 static void build_irq_error(const char *msg, unsigned int ino, int pil, int inofixup,
 264                             unsigned long iclr, unsigned long imap,
 265                             struct ino_bucket *bucket)
 266 {
 267         prom_printf("IRQ: INO %04x (%d:%016lx:%016lx) --> "
 268                     "(%d:%d:%016lx:%016lx), halting...\n",
 269                     ino, bucket->pil, bucket->iclr, bucket->imap,
 270                     pil, inofixup, iclr, imap);
 271         prom_halt();
 272 }
 273
 274 unsigned int build_irq(int pil, int inofixup, unsigned long iclr, unsigned long imap)
 275 {
 276         struct ino_bucket *bucket;
 277         int ino;
 278
 279         if (pil == 0) {
 280                 if (iclr != 0UL || imap != 0UL) {
 281                         prom_printf("Invalid dummy bucket for PIL0 (%lx:%lx)\n",
 282                                     iclr, imap);
 283                         prom_halt();
 284                 }
 285                 return __irq(&pil0_dummy_bucket);
 286         }
 287
 288         BUG_ON(tlb_type == hypervisor);
 289
 290         /* RULE: Both must be specified in all other cases. */
 291         if (iclr == 0UL || imap == 0UL) {
 292                 prom_printf("Invalid build_irq %d %d %016lx %016lx\n",
 293                             pil, inofixup, iclr, imap);
 294                 prom_halt();
 295         }
 296
 297         ino = (upa_readl(imap) & (IMAP_IGN | IMAP_INO)) + inofixup;
 298         if (ino > NUM_IVECS) {
 299                 prom_printf("Invalid INO %04x (%d:%d:%016lx:%016lx)\n",
 300                             ino, pil, inofixup, iclr, imap);
 301                 prom_halt();
 302         }
 303
 304         bucket = &ivector_table[ino];
 305         if (bucket->flags & IBF_ACTIVE)
 306                 build_irq_error("IRQ: Trying to build active INO bucket.\n",
 307                                 ino, pil, inofixup, iclr, imap, bucket);
 308
 309         if (bucket->irq_info) {
 310                 if (bucket->imap != imap || bucket->iclr != iclr)
 311                         build_irq_error("IRQ: Trying to reinit INO bucket.\n",
 312                                         ino, pil, inofixup, iclr, imap, bucket);
 313
 314                 goto out;
 315         }
 316
 317         bucket->irq_info = kzalloc(sizeof(struct irq_desc), GFP_ATOMIC);
 318         if (!bucket->irq_info) {
 319                 prom_printf("IRQ: Error, kmalloc(irq_desc) failed.\n");
 320                 prom_halt();
 321         }
 322
 323         /* Ok, looks good, set it up.  Don't touch the irq_chain or
 324          * the pending flag.
 325          */
 326         bucket->imap  = imap;
 327         bucket->iclr  = iclr;
 328         bucket->pil   = pil;
 329         bucket->flags = 0;
 330
 331 out:
 332         return __irq(bucket);
 333 }
 334
 335 unsigned int sun4v_build_irq(u32 devhandle, unsigned int devino, int pil, unsigned char flags)
 336 {
 337         struct ino_bucket *bucket;
 338         unsigned long sysino;
 339
 340         sysino = sun4v_devino_to_sysino(devhandle, devino);
 341
 342         bucket = &ivector_table[sysino];
 343
 344         /* Catch accidental accesses to these things.  IMAP/ICLR handling
 345          * is done by hypervisor calls on sun4v platforms, not by direct
 346          * register accesses.
 347          *
 348          * But we need to make them look unique for the disable_irq() logic
 349          * in free_irq().
 350          */
 351         bucket->imap = ~0UL - sysino;
 352         bucket->iclr = ~0UL - sysino;
 353
 354         bucket->pil = pil;
 355         bucket->flags = flags;
 356
 357         bucket->irq_info = kzalloc(sizeof(struct irq_desc), GFP_ATOMIC);
 358         if (!bucket->irq_info) {
 359                 prom_printf("IRQ: Error, kmalloc(irq_desc) failed.\n");
 360                 prom_halt();
 361         }
 362
 363         return __irq(bucket);
 364 }
 365
 366 static void atomic_bucket_insert(struct ino_bucket *bucket)
 367 {
 368         unsigned long pstate;
 369         unsigned int *ent;
 370
 371         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 372         __asm__ __volatile__("wrpr %0, %1, %%pstate"
 373                              : : "r" (pstate), "i" (PSTATE_IE));
 374         ent = irq_work(smp_processor_id(), bucket->pil);
 375         bucket->irq_chain = *ent;
 376         *ent = __irq(bucket);
 377         __asm__ __volatile__("wrpr %0, 0x0, %%pstate" : : "r" (pstate));
 378 }
 379
 380 static int check_irq_sharing(int pil, unsigned long irqflags)
 381 {
 382         struct irqaction *action, *tmp;
 383
 384         action = *(irq_action + pil);
 385         if (action) {
 386                 if ((action->flags & SA_SHIRQ) && (irqflags & SA_SHIRQ)) {
 387                         for (tmp = action; tmp->next; tmp = tmp->next)
 388                                 ;
 389                 } else {
 390                         return -EBUSY;
 391                 }
 392         }
 393         return 0;
 394 }
 395
 396 static void append_irq_action(int pil, struct irqaction *action)
 397 {
 398         struct irqaction **pp = irq_action + pil;
 399
 400         while (*pp)
 401                 pp = &((*pp)->next);
 402         *pp = action;
 403 }
 404
 405 static struct irqaction *get_action_slot(struct ino_bucket *bucket)
 406 {
 407         struct irq_desc *desc = bucket->irq_info;
 408         int max_irq, i;
 409
 410         max_irq = 1;
 411         if (bucket->flags & IBF_PCI)
 412                 max_irq = MAX_IRQ_DESC_ACTION;
 413         for (i = 0; i < max_irq; i++) {
 414                 struct irqaction *p = &desc->action[i];
 415                 u32 mask = (1 << i);
 416
 417                 if (desc->action_active_mask & mask)
 418                         continue;
 419
 420                 desc->action_active_mask |= mask;
 421                 return p;
 422         }
 423         return NULL;
 424 }
 425
 426 int request_irq(unsigned int irq, irqreturn_t (*handler)(int, void *, struct pt_regs *),
 427                 unsigned long irqflags, const char *name, void *dev_id)
 428 {
 429         struct irqaction *action;
 430         struct ino_bucket *bucket = __bucket(irq);
 431         unsigned long flags;
 432         int pending = 0;
 433
 434         if (unlikely(!handler))
 435                 return -EINVAL;
 436
 437         if (unlikely(!bucket->irq_info))
 438                 return -ENODEV;
 439
 440         if ((bucket != &pil0_dummy_bucket) && (irqflags & SA_SAMPLE_RANDOM)) {
 441                 /*
 442                  * This function might sleep, we want to call it first,
 443                  * outside of the atomic block. In SA_STATIC_ALLOC case,
 444                  * random driver's kmalloc will fail, but it is safe.
 445                  * If already initialized, random driver will not reinit.
 446                  * Yes, this might clear the entropy pool if the wrong
 447                  * driver is attempted to be loaded, without actually
 448                  * installing a new handler, but is this really a problem,
 449                  * only the sysadmin is able to do this.
 450                  */
 451                 rand_initialize_irq(irq);
 452         }
 453
 454         spin_lock_irqsave(&irq_action_lock, flags);
 455
 456         if (check_irq_sharing(bucket->pil, irqflags)) {
 457                 spin_unlock_irqrestore(&irq_action_lock, flags);
 458                 return -EBUSY;
 459         }
 460
 461         action = get_action_slot(bucket);
 462         if (!action) {
 463                 spin_unlock_irqrestore(&irq_action_lock, flags);
 464                 return -ENOMEM;
 465         }
 466
 467         bucket->flags |= IBF_ACTIVE;
 468         pending = 0;
 469         if (bucket != &pil0_dummy_bucket) {
 470                 pending = bucket->pending;
 471                 if (pending)
 472                         bucket->pending = 0;
 473         }
 474
 475         action->handler = handler;
 476         action->flags = irqflags;
 477         action->name = name;
 478         action->next = NULL;
 479         action->dev_id = dev_id;
 480         put_ino_in_irqaction(action, irq);
 481         put_smpaff_in_irqaction(action, CPU_MASK_NONE);
 482
 483         append_irq_action(bucket->pil, action);
 484
 485         enable_irq(irq);
 486
 487         /* We ate the IVEC already, this makes sure it does not get lost. */
 488         if (pending) {
 489                 atomic_bucket_insert(bucket);
 490                 set_softint(1 << bucket->pil);
 491         }
 492
 493         spin_unlock_irqrestore(&irq_action_lock, flags);
 494
 495         if (bucket != &pil0_dummy_bucket)
 496                 register_irq_proc(__irq_ino(irq));
 497
 498 #ifdef CONFIG_SMP
 499         distribute_irqs();
 500 #endif
 501         return 0;
 502 }
 503
 504 EXPORT_SYMBOL(request_irq);
 505
 506 static struct irqaction *unlink_irq_action(unsigned int irq, void *dev_id)
 507 {
 508         struct ino_bucket *bucket = __bucket(irq);
 509         struct irqaction *action, **pp;
 510
 511         pp = irq_action + bucket->pil;
 512         action = *pp;
 513         if (unlikely(!action))
 514                 return NULL;
 515
 516         if (unlikely(!action->handler)) {
 517                 printk("Freeing free IRQ %d\n", bucket->pil);
 518                 return NULL;
 519         }
 520
 521         while (action && action->dev_id != dev_id) {
 522                 pp = &action->next;
 523                 action = *pp;
 524         }
 525
 526         if (likely(action))
 527                 *pp = action->next;
 528
 529         return action;
 530 }
 531
 532 void free_irq(unsigned int irq, void *dev_id)
 533 {
 534         struct irqaction *action;
 535         struct ino_bucket *bucket;
 536         unsigned long flags;
 537
 538         spin_lock_irqsave(&irq_action_lock, flags);
 539
 540         action = unlink_irq_action(irq, dev_id);
 541
 542         spin_unlock_irqrestore(&irq_action_lock, flags);
 543
 544         if (unlikely(!action))
 545                 return;
 546
 547         synchronize_irq(irq);
 548
 549         spin_lock_irqsave(&irq_action_lock, flags);
 550
 551         bucket = __bucket(irq);
 552         if (bucket != &pil0_dummy_bucket) {
 553                 struct irq_desc *desc = bucket->irq_info;
 554                 int ent, i;
 555
 556                 for (i = 0; i < MAX_IRQ_DESC_ACTION; i++) {
 557                         struct irqaction *p = &desc->action[i];
 558
 559                         if (p == action) {
 560                                 desc->action_active_mask &= ~(1 << i);
 561                                 break;
 562                         }
 563                 }
 564
 565                 if (!desc->action_active_mask) {
 566                         unsigned long imap = bucket->imap;
 567
 568                         /* This unique interrupt source is now inactive. */
 569                         bucket->flags &= ~IBF_ACTIVE;
 570
 571                         /* See if any other buckets share this bucket's IMAP
 572                          * and are still active.
 573                          */
 574                         for (ent = 0; ent < NUM_IVECS; ent++) {
 575                                 struct ino_bucket *bp = &ivector_table[ent];
 576                                 if (bp != bucket        &&
 577                                     bp->imap == imap    &&
 578                                     (bp->flags & IBF_ACTIVE) != 0)
 579                                         break;
 580                         }
 581
 582                         /* Only disable when no other sub-irq levels of
 583                          * the same IMAP are active.
 584                          */
 585                         if (ent == NUM_IVECS)
 586                                 disable_irq(irq);
 587                 }
 588         }
 589
 590         spin_unlock_irqrestore(&irq_action_lock, flags);
 591 }
 592
 593 EXPORT_SYMBOL(free_irq);
 594
 595 #ifdef CONFIG_SMP
 596 void synchronize_irq(unsigned int irq)
 597 {
 598         struct ino_bucket *bucket = __bucket(irq);
 599
 600 #if 0
 601         /* The following is how I wish I could implement this.
 602          * Unfortunately the ICLR registers are read-only, you can
 603          * only write ICLR_foo values to them.  To get the current
 604          * IRQ status you would need to get at the IRQ diag registers
 605          * in the PCI/SBUS controller and the layout of those vary
 606          * from one controller to the next, sigh... -DaveM
 607          */
 608         unsigned long iclr = bucket->iclr;
 609
 610         while (1) {
 611                 u32 tmp = upa_readl(iclr);
 612
 613                 if (tmp == ICLR_TRANSMIT ||
 614                     tmp == ICLR_PENDING) {
 615                         cpu_relax();
 616                         continue;
 617                 }
 618                 break;
 619         }
 620 #else
 621         /* So we have to do this with a INPROGRESS bit just like x86.  */
 622         while (bucket->flags & IBF_INPROGRESS)
 623                 cpu_relax();
 624 #endif
 625 }
 626 #endif /* CONFIG_SMP */
 627
 628 static void process_bucket(int irq, struct ino_bucket *bp, struct pt_regs *regs)
 629 {
 630         struct irq_desc *desc = bp->irq_info;
 631         unsigned char flags = bp->flags;
 632         u32 action_mask, i;
 633         int random;
 634
 635         bp->flags |= IBF_INPROGRESS;
 636
 637         if (unlikely(!(flags & IBF_ACTIVE))) {
 638                 bp->pending = 1;
 639                 goto out;
 640         }
 641
 642         if (desc->pre_handler)
 643                 desc->pre_handler(bp,
 644                                   desc->pre_handler_arg1,
 645                                   desc->pre_handler_arg2);
 646
 647         action_mask = desc->action_active_mask;
 648         random = 0;
 649         for (i = 0; i < MAX_IRQ_DESC_ACTION; i++) {
 650                 struct irqaction *p = &desc->action[i];
 651                 u32 mask = (1 << i);
 652
 653                 if (!(action_mask & mask))
 654                         continue;
 655
 656                 action_mask &= ~mask;
 657
 658                 if (p->handler(__irq(bp), p->dev_id, regs) == IRQ_HANDLED)
 659                         random |= p->flags;
 660
 661                 if (!action_mask)
 662                         break;
 663         }
 664         if (bp->pil != 0) {
 665                 if (tlb_type == hypervisor) {
 666                         unsigned int ino = __irq_ino(bp);
 667                         int err;
 668
 669                         err = sun4v_intr_setstate(ino, HV_INTR_STATE_IDLE);
 670                         if (err != HV_EOK)
 671                                 printk("sun4v_intr_setstate(%x): "
 672                                        "err(%d)\n", ino, err);
 673                 } else {
 674                         upa_writel(ICLR_IDLE, bp->iclr);
 675                 }
 676
 677                 /* Test and add entropy */
 678                 if (random & SA_SAMPLE_RANDOM)
 679                         add_interrupt_randomness(irq);
 680         }
 681 out:
 682         bp->flags &= ~IBF_INPROGRESS;
 683 }
 684
 685 void handler_irq(int irq, struct pt_regs *regs)
 686 {
 687         struct ino_bucket *bp;
 688         int cpu = smp_processor_id();
 689
 690 #ifndef CONFIG_SMP
 691         /*
 692          * Check for TICK_INT on level 14 softint.
 693          */
 694         {
 695                 unsigned long clr_mask = 1 << irq;
 696                 unsigned long tick_mask = tick_ops->softint_mask;
 697
 698                 if ((irq == 14) && (get_softint() & tick_mask)) {
 699                         irq = 0;
 700                         clr_mask = tick_mask;
 701                 }
 702                 clear_softint(clr_mask);
 703         }
 704 #else
 705         clear_softint(1 << irq);
 706 #endif
 707
 708         irq_enter();
 709         kstat_this_cpu.irqs[irq]++;
 710
 711         /* Sliiiick... */
 712 #ifndef CONFIG_SMP
 713         bp = ((irq != 0) ?
 714               __bucket(xchg32(irq_work(cpu, irq), 0)) :
 715               &pil0_dummy_bucket);
 716 #else
 717         bp = __bucket(xchg32(irq_work(cpu, irq), 0));
 718 #endif
 719         while (bp) {
 720                 struct ino_bucket *nbp = __bucket(bp->irq_chain);
 721
 722                 bp->irq_chain = 0;
 723                 process_bucket(irq, bp, regs);
 724                 bp = nbp;
 725         }
 726         irq_exit();
 727 }
 728
 729 #ifdef CONFIG_BLK_DEV_FD
 730 extern irqreturn_t floppy_interrupt(int, void *, struct pt_regs *);
 731
 732 /* XXX No easy way to include asm/floppy.h XXX */
 733 extern unsigned char *pdma_vaddr;
 734 extern unsigned long pdma_size;
 735 extern volatile int doing_pdma;
 736 extern unsigned long fdc_status;
 737
 738 irqreturn_t sparc_floppy_irq(int irq, void *dev_cookie, struct pt_regs *regs)
 739 {
 740         if (likely(doing_pdma)) {
 741                 void __iomem *stat = (void __iomem *) fdc_status;
 742                 unsigned char *vaddr = pdma_vaddr;
 743                 unsigned long size = pdma_size;
 744                 u8 val;
 745
 746                 while (size) {
 747                         val = readb(stat);
 748                         if (unlikely(!(val & 0x80))) {
 749                                 pdma_vaddr = vaddr;
 750                                 pdma_size = size;
 751                                 return IRQ_HANDLED;
 752                         }
 753                         if (unlikely(!(val & 0x20))) {
 754                                 pdma_vaddr = vaddr;
 755                                 pdma_size = size;
 756                                 doing_pdma = 0;
 757                                 goto main_interrupt;
 758                         }
 759                         if (val & 0x40) {
 760                                 /* read */
 761                                 *vaddr++ = readb(stat + 1);
 762                         } else {
 763                                 unsigned char data = *vaddr++;
 764
 765                                 /* write */
 766                                 writeb(data, stat + 1);
 767                         }
 768                         size--;
 769                 }
 770
 771                 pdma_vaddr = vaddr;
 772                 pdma_size = size;
 773
 774                 /* Send Terminal Count pulse to floppy controller. */
 775                 val = readb(auxio_register);
 776                 val |= AUXIO_AUX1_FTCNT;
 777                 writeb(val, auxio_register);
 778                 val &= ~AUXIO_AUX1_FTCNT;
 779                 writeb(val, auxio_register);
 780
 781                 doing_pdma = 0;
 782         }
 783
 784 main_interrupt:
 785         return floppy_interrupt(irq, dev_cookie, regs);
 786 }
 787 EXPORT_SYMBOL(sparc_floppy_irq);
 788 #endif
 789
 790 /* We really don't need these at all on the Sparc.  We only have
 791  * stubs here because they are exported to modules.
 792  */
 793 unsigned long probe_irq_on(void)
 794 {
 795         return 0;
 796 }
 797
 798 EXPORT_SYMBOL(probe_irq_on);
 799
 800 int probe_irq_off(unsigned long mask)
 801 {
 802         return 0;
 803 }
 804
 805 EXPORT_SYMBOL(probe_irq_off);
 806
 807 #ifdef CONFIG_SMP
 808 static int retarget_one_irq(struct irqaction *p, int goal_cpu)
 809 {
 810         struct ino_bucket *bucket = get_ino_in_irqaction(p) + ivector_table;
 811
 812         while (!cpu_online(goal_cpu)) {
 813                 if (++goal_cpu >= NR_CPUS)
 814                         goal_cpu = 0;
 815         }
 816
 817         if (tlb_type == hypervisor) {
 818                 unsigned int ino = __irq_ino(bucket);
 819
 820                 sun4v_intr_settarget(ino, goal_cpu);
 821                 sun4v_intr_setenabled(ino, HV_INTR_ENABLED);
 822         } else {
 823                 unsigned long imap = bucket->imap;
 824                 unsigned int tid = sun4u_compute_tid(imap, goal_cpu);
 825
 826                 upa_writel(tid | IMAP_VALID, imap);
 827         }
 828
 829         do {
 830                 if (++goal_cpu >= NR_CPUS)
 831                         goal_cpu = 0;
 832         } while (!cpu_online(goal_cpu));
 833
 834         return goal_cpu;
 835 }
 836
 837 /* Called from request_irq. */
 838 static void distribute_irqs(void)
 839 {
 840         unsigned long flags;
 841         int cpu, level;
 842
 843         spin_lock_irqsave(&irq_action_lock, flags);
 844         cpu = 0;
 845
 846         /*
 847          * Skip the timer at [0], and very rare error/power intrs at [15].
 848          * Also level [12], it causes problems on Ex000 systems.
 849          */
 850         for (level = 1; level < NR_IRQS; level++) {
 851                 struct irqaction *p = irq_action[level];
 852
 853                 if (level == 12)
 854                         continue;
 855
 856                 while(p) {
 857                         cpu = retarget_one_irq(p, cpu);
 858                         p = p->next;
 859                 }
 860         }
 861         spin_unlock_irqrestore(&irq_action_lock, flags);
 862 }
 863 #endif
 864
 865 struct sun5_timer {
 866         u64     count0;
 867         u64     limit0;
 868         u64     count1;
 869         u64     limit1;
 870 };
 871
 872 static struct sun5_timer *prom_timers;
 873 static u64 prom_limit0, prom_limit1;
 874
 875 static void map_prom_timers(void)
 876 {
 877         unsigned int addr[3];
 878         int tnode, err;
 879
 880         /* PROM timer node hangs out in the top level of device siblings... */
 881         tnode = prom_finddevice("/counter-timer");
 882
 883         /* Assume if node is not present, PROM uses different tick mechanism
 884          * which we should not care about.
 885          */
 886         if (tnode == 0 || tnode == -1) {
 887                 prom_timers = (struct sun5_timer *) 0;
 888                 return;
 889         }
 890
 891         /* If PROM is really using this, it must be mapped by him. */
 892         err = prom_getproperty(tnode, "address", (char *)addr, sizeof(addr));
 893         if (err == -1) {
 894                 prom_printf("PROM does not have timer mapped, trying to continue.\n");
 895                 prom_timers = (struct sun5_timer *) 0;
 896                 return;
 897         }
 898         prom_timers = (struct sun5_timer *) ((unsigned long)addr[0]);
 899 }
 900
 901 static void kill_prom_timer(void)
 902 {
 903         if (!prom_timers)
 904                 return;
 905
 906         /* Save them away for later. */
 907         prom_limit0 = prom_timers->limit0;
 908         prom_limit1 = prom_timers->limit1;
 909
 910         /* Just as in sun4c/sun4m PROM uses timer which ticks at IRQ 14.
 911          * We turn both off here just to be paranoid.
 912          */
 913         prom_timers->limit0 = 0;
 914         prom_timers->limit1 = 0;
 915
 916         /* Wheee, eat the interrupt packet too... */
 917         __asm__ __volatile__(
 918 "       mov     0x40, %%g2\n"
 919 "       ldxa    [%%g0] %0, %%g1\n"
 920 "       ldxa    [%%g2] %1, %%g1\n"
 921 "       stxa    %%g0, [%%g0] %0\n"
 922 "       membar  #Sync\n"
 923         : /* no outputs */
 924         : "i" (ASI_INTR_RECEIVE), "i" (ASI_INTR_R)
 925         : "g1", "g2");
 926 }
 927
 928 void init_irqwork_curcpu(void)
 929 {
 930         int cpu = hard_smp_processor_id();
 931
 932         memset(__irq_work + cpu, 0, sizeof(struct irq_work_struct));
 933 }
 934
 935 static void __cpuinit register_one_mondo(unsigned long paddr, unsigned long type)
 936 {
 937         unsigned long num_entries = 128;
 938         unsigned long status;
 939
 940         status = sun4v_cpu_qconf(type, paddr, num_entries);
 941         if (status != HV_EOK) {
 942                 prom_printf("SUN4V: sun4v_cpu_qconf(%lu:%lx:%lu) failed, "
 943                             "err %lu\n", type, paddr, num_entries, status);
 944                 prom_halt();
 945         }
 946 }
 947
 948 static void __cpuinit sun4v_register_mondo_queues(int this_cpu)
 949 {
 950         struct trap_per_cpu *tb = &trap_block[this_cpu];
 951
 952         register_one_mondo(tb->cpu_mondo_pa, HV_CPU_QUEUE_CPU_MONDO);
 953         register_one_mondo(tb->dev_mondo_pa, HV_CPU_QUEUE_DEVICE_MONDO);
 954         register_one_mondo(tb->resum_mondo_pa, HV_CPU_QUEUE_RES_ERROR);
 955         register_one_mondo(tb->nonresum_mondo_pa, HV_CPU_QUEUE_NONRES_ERROR);
 956 }
 957
 958 static void __cpuinit alloc_one_mondo(unsigned long *pa_ptr, int use_bootmem)
 959 {
 960         void *page;
 961
 962         if (use_bootmem)
 963                 page = alloc_bootmem_low_pages(PAGE_SIZE);
 964         else
 965                 page = (void *) get_zeroed_page(GFP_ATOMIC);
 966
 967         if (!page) {
 968                 prom_printf("SUN4V: Error, cannot allocate mondo queue.\n");
 969                 prom_halt();
 970         }
 971
 972         *pa_ptr = __pa(page);
 973 }
 974
 975 static void __cpuinit alloc_one_kbuf(unsigned long *pa_ptr, int use_bootmem)
 976 {
 977         void *page;
 978
 979         if (use_bootmem)
 980                 page = alloc_bootmem_low_pages(PAGE_SIZE);
 981         else
 982                 page = (void *) get_zeroed_page(GFP_ATOMIC);
 983
 984         if (!page) {
 985                 prom_printf("SUN4V: Error, cannot allocate kbuf page.\n");
 986                 prom_halt();
 987         }
 988
 989         *pa_ptr = __pa(page);
 990 }
 991
 992 static void __cpuinit init_cpu_send_mondo_info(struct trap_per_cpu *tb, int use_bootmem)
 993 {
 994 #ifdef CONFIG_SMP
 995         void *page;
 996
 997         BUILD_BUG_ON((NR_CPUS * sizeof(u16)) > (PAGE_SIZE - 64));
 998
 999         if (use_bootmem)
1000                 page = alloc_bootmem_low_pages(PAGE_SIZE);
1001         else
1002                 page = (void *) get_zeroed_page(GFP_ATOMIC);
1003
1004         if (!page) {
1005                 prom_printf("SUN4V: Error, cannot allocate cpu mondo page.\n");
1006                 prom_halt();
1007         }
1008
1009         tb->cpu_mondo_block_pa = __pa(page);
1010         tb->cpu_list_pa = __pa(page + 64);
1011 #endif
1012 }
1013
1014 /* Allocate and register the mondo and error queues for this cpu.  */
1015 void __cpuinit sun4v_init_mondo_queues(int use_bootmem, int cpu, int alloc, int load)
1016 {
1017         struct trap_per_cpu *tb = &trap_block[cpu];
1018
1019         if (alloc) {
1020                 alloc_one_mondo(&tb->cpu_mondo_pa, use_bootmem);
1021                 alloc_one_mondo(&tb->dev_mondo_pa, use_bootmem);
1022                 alloc_one_mondo(&tb->resum_mondo_pa, use_bootmem);
1023                 alloc_one_kbuf(&tb->resum_kernel_buf_pa, use_bootmem);
1024                 alloc_one_mondo(&tb->nonresum_mondo_pa, use_bootmem);
1025                 alloc_one_kbuf(&tb->nonresum_kernel_buf_pa, use_bootmem);
1026
1027                 init_cpu_send_mondo_info(tb, use_bootmem);
1028         }
1029
1030         if (load) {
1031                 if (cpu != hard_smp_processor_id()) {
1032                         prom_printf("SUN4V: init mondo on cpu %d not %d\n",
1033                                     cpu, hard_smp_processor_id());
1034                         prom_halt();
1035                 }
1036                 sun4v_register_mondo_queues(cpu);
1037         }
1038 }
1039
1040 /* Only invoked on boot processor. */
1041 void __init init_IRQ(void)
1042 {
1043         map_prom_timers();
1044         kill_prom_timer();
1045         memset(&ivector_table[0], 0, sizeof(ivector_table));
1046
1047         if (tlb_type == hypervisor)
1048                 sun4v_init_mondo_queues(1, hard_smp_processor_id(), 1, 1);
1049
1050         /* We need to clear any IRQ's pending in the soft interrupt
1051          * registers, a spurious one could be left around from the
1052          * PROM timer which we just disabled.
1053          */
1054         clear_softint(get_softint());
1055
1056         /* Now that ivector table is initialized, it is safe
1057          * to receive IRQ vector traps.  We will normally take
1058          * one or two right now, in case some device PROM used
1059          * to boot us wants to speak to us.  We just ignore them.
1060          */
1061         __asm__ __volatile__("rdpr      %%pstate, %%g1\n\t"
1062                              "or        %%g1, %0, %%g1\n\t"
1063                              "wrpr      %%g1, 0x0, %%pstate"
1064                              : /* No outputs */
1065                              : "i" (PSTATE_IE)
1066                              : "g1");
1067 }
1068
1069 static struct proc_dir_entry * root_irq_dir;
1070 static struct proc_dir_entry * irq_dir [NUM_IVECS];
1071
1072 #ifdef CONFIG_SMP
1073
1074 static int irq_affinity_read_proc (char *page, char **start, off_t off,
1075                         int count, int *eof, void *data)
1076 {
1077         struct ino_bucket *bp = ivector_table + (long)data;
1078         struct irq_desc *desc = bp->irq_info;
1079         struct irqaction *ap = desc->action;
1080         cpumask_t mask;
1081         int len;
1082
1083         mask = get_smpaff_in_irqaction(ap);
1084         if (cpus_empty(mask))
1085                 mask = cpu_online_map;
1086
1087         len = cpumask_scnprintf(page, count, mask);
1088         if (count - len < 2)
1089                 return -EINVAL;
1090         len += sprintf(page + len, "\n");
1091         return len;
1092 }
1093
1094 static inline void set_intr_affinity(int irq, cpumask_t hw_aff)
1095 {
1096         struct ino_bucket *bp = ivector_table + irq;
1097         struct irq_desc *desc = bp->irq_info;
1098         struct irqaction *ap = desc->action;
1099
1100         /* Users specify affinity in terms of hw cpu ids.
1101          * As soon as we do this, handler_irq() might see and take action.
1102          */
1103         put_smpaff_in_irqaction(ap, hw_aff);
1104
1105         /* Migration is simply done by the next cpu to service this
1106          * interrupt.
1107          */
1108 }
1109
1110 static int irq_affinity_write_proc (struct file *file, const char __user *buffer,
1111                                         unsigned long count, void *data)
1112 {
1113         int irq = (long) data, full_count = count, err;
1114         cpumask_t new_value;
1115
1116         err = cpumask_parse(buffer, count, new_value);
1117
1118         /*
1119          * Do not allow disabling IRQs completely - it's a too easy
1120          * way to make the system unusable accidentally :-) At least
1121          * one online CPU still has to be targeted.
1122          */
1123         cpus_and(new_value, new_value, cpu_online_map);
1124         if (cpus_empty(new_value))
1125                 return -EINVAL;
1126
1127         set_intr_affinity(irq, new_value);
1128
1129         return full_count;
1130 }
1131
1132 #endif
1133
1134 #define MAX_NAMELEN 10
1135
1136 static void register_irq_proc (unsigned int irq)
1137 {
1138         char name [MAX_NAMELEN];
1139
1140         if (!root_irq_dir || irq_dir[irq])
1141                 return;
1142
1143         memset(name, 0, MAX_NAMELEN);
1144         sprintf(name, "%x", irq);
1145
1146         /* create /proc/irq/1234 */
1147         irq_dir[irq] = proc_mkdir(name, root_irq_dir);
1148
1149 #ifdef CONFIG_SMP
1150         /* XXX SMP affinity not supported on starfire yet. */
1151         if (this_is_starfire == 0) {
1152                 struct proc_dir_entry *entry;
1153
1154                 /* create /proc/irq/1234/smp_affinity */
1155                 entry = create_proc_entry("smp_affinity", 0600, irq_dir[irq]);
1156
1157                 if (entry) {
1158                         entry->nlink = 1;
1159                         entry->data = (void *)(long)irq;
1160                         entry->read_proc = irq_affinity_read_proc;
1161                         entry->write_proc = irq_affinity_write_proc;
1162                 }
1163         }
1164 #endif
1165 }
1166
1167 void init_irq_proc (void)
1168 {
1169         /* create /proc/irq */
1170         root_irq_dir = proc_mkdir("irq", NULL);
1171 }
1172