drivers/xen/events.c

   1 /*
   2  * Xen event channels
   3  *
   4  * Xen models interrupts with abstract event channels.  Because each
   5  * domain gets 1024 event channels, but NR_IRQ is not that large, we
   6  * must dynamically map irqs<->event channels.  The event channels
   7  * interface with the rest of the kernel by defining a xen interrupt
   8  * chip.  When an event is recieved, it is mapped to an irq and sent
   9  * through the normal interrupt processing path.
  10  *
  11  * There are four kinds of events which can be mapped to an event
  12  * channel:
  13  *
  14  * 1. Inter-domain notifications.  This includes all the virtual
  15  *    device events, since they're driven by front-ends in another domain
  16  *    (typically dom0).
  17  * 2. VIRQs, typically used for timers.  These are per-cpu events.
  18  * 3. IPIs.
  19  * 4. PIRQs - Hardware interrupts.
  20  *
  21  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
  22  */
  23
  24 #include <linux/linkage.h>
  25 #include <linux/interrupt.h>
  26 #include <linux/irq.h>
  27 #include <linux/module.h>
  28 #include <linux/string.h>
  29 #include <linux/bootmem.h>
  30 #include <linux/slab.h>
  31 #include <linux/irqnr.h>
  32
  33 #include <asm/desc.h>
  34 #include <asm/ptrace.h>
  35 #include <asm/irq.h>
  36 #include <asm/idle.h>
  37 #include <asm/io_apic.h>
  38 #include <asm/sync_bitops.h>
  39 #include <asm/xen/hypercall.h>
  40 #include <asm/xen/hypervisor.h>
  41
  42 #include <xen/xen.h>
  43 #include <xen/hvm.h>
  44 #include <xen/xen-ops.h>
  45 #include <xen/events.h>
  46 #include <xen/interface/xen.h>
  47 #include <xen/interface/event_channel.h>
  48 #include <xen/interface/hvm/hvm_op.h>
  49 #include <xen/interface/hvm/params.h>
  50
  51 /*
  52  * This lock protects updates to the following mapping and reference-count
  53  * arrays. The lock does not need to be acquired to read the mapping tables.
  54  */
  55 static DEFINE_SPINLOCK(irq_mapping_update_lock);
  56
  57 /* IRQ <-> VIRQ mapping. */
  58 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
  59
  60 /* IRQ <-> IPI mapping */
  61 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
  62
  63 /* Interrupt types. */
  64 enum xen_irq_type {
  65         IRQT_UNBOUND = 0,
  66         IRQT_PIRQ,
  67         IRQT_VIRQ,
  68         IRQT_IPI,
  69         IRQT_EVTCHN
  70 };
  71
  72 /*
  73  * Packed IRQ information:
  74  * type - enum xen_irq_type
  75  * event channel - irq->event channel mapping
  76  * cpu - cpu this event channel is bound to
  77  * index - type-specific information:
  78  *    PIRQ - vector, with MSB being "needs EIO"
  79  *    VIRQ - virq number
  80  *    IPI - IPI vector
  81  *    EVTCHN -
  82  */
  83 struct irq_info
  84 {
  85         enum xen_irq_type type; /* type */
  86         unsigned short evtchn;  /* event channel */
  87         unsigned short cpu;     /* cpu bound */
  88
  89         union {
  90                 unsigned short virq;
  91                 enum ipi_vector ipi;
  92                 struct {
  93                         unsigned short gsi;
  94                         unsigned char vector;
  95                         unsigned char flags;
  96                 } pirq;
  97         } u;
  98 };
  99 #define PIRQ_NEEDS_EOI  (1 << 0)
 100
 101 static struct irq_info *irq_info;
 102
 103 static int *evtchn_to_irq;
 104 struct cpu_evtchn_s {
 105         unsigned long bits[NR_EVENT_CHANNELS/BITS_PER_LONG];
 106 };
 107 static struct cpu_evtchn_s *cpu_evtchn_mask_p;
 108 static inline unsigned long *cpu_evtchn_mask(int cpu)
 109 {
 110         return cpu_evtchn_mask_p[cpu].bits;
 111 }
 112
 113 /* Xen will never allocate port zero for any purpose. */
 114 #define VALID_EVTCHN(chn)       ((chn) != 0)
 115
 116 static struct irq_chip xen_dynamic_chip;
 117 static struct irq_chip xen_percpu_chip;
 118 static struct irq_chip xen_pirq_chip;
 119
 120 /* Constructor for packed IRQ information. */
 121 static struct irq_info mk_unbound_info(void)
 122 {
 123         return (struct irq_info) { .type = IRQT_UNBOUND };
 124 }
 125
 126 static struct irq_info mk_evtchn_info(unsigned short evtchn)
 127 {
 128         return (struct irq_info) { .type = IRQT_EVTCHN, .evtchn = evtchn,
 129                         .cpu = 0 };
 130 }
 131
 132 static struct irq_info mk_ipi_info(unsigned short evtchn, enum ipi_vector ipi)
 133 {
 134         return (struct irq_info) { .type = IRQT_IPI, .evtchn = evtchn,
 135                         .cpu = 0, .u.ipi = ipi };
 136 }
 137
 138 static struct irq_info mk_virq_info(unsigned short evtchn, unsigned short virq)
 139 {
 140         return (struct irq_info) { .type = IRQT_VIRQ, .evtchn = evtchn,
 141                         .cpu = 0, .u.virq = virq };
 142 }
 143
 144 static struct irq_info mk_pirq_info(unsigned short evtchn,
 145                                     unsigned short gsi, unsigned short vector)
 146 {
 147         return (struct irq_info) { .type = IRQT_PIRQ, .evtchn = evtchn,
 148                         .cpu = 0, .u.pirq = { .gsi = gsi, .vector = vector } };
 149 }
 150
 151 /*
 152  * Accessors for packed IRQ information.
 153  */
 154 static struct irq_info *info_for_irq(unsigned irq)
 155 {
 156         return &irq_info[irq];
 157 }
 158
 159 static unsigned int evtchn_from_irq(unsigned irq)
 160 {
 161         return info_for_irq(irq)->evtchn;
 162 }
 163
 164 unsigned irq_from_evtchn(unsigned int evtchn)
 165 {
 166         return evtchn_to_irq[evtchn];
 167 }
 168 EXPORT_SYMBOL_GPL(irq_from_evtchn);
 169
 170 static enum ipi_vector ipi_from_irq(unsigned irq)
 171 {
 172         struct irq_info *info = info_for_irq(irq);
 173
 174         BUG_ON(info == NULL);
 175         BUG_ON(info->type != IRQT_IPI);
 176
 177         return info->u.ipi;
 178 }
 179
 180 static unsigned virq_from_irq(unsigned irq)
 181 {
 182         struct irq_info *info = info_for_irq(irq);
 183
 184         BUG_ON(info == NULL);
 185         BUG_ON(info->type != IRQT_VIRQ);
 186
 187         return info->u.virq;
 188 }
 189
 190 static unsigned gsi_from_irq(unsigned irq)
 191 {
 192         struct irq_info *info = info_for_irq(irq);
 193
 194         BUG_ON(info == NULL);
 195         BUG_ON(info->type != IRQT_PIRQ);
 196
 197         return info->u.pirq.gsi;
 198 }
 199
 200 static unsigned vector_from_irq(unsigned irq)
 201 {
 202         struct irq_info *info = info_for_irq(irq);
 203
 204         BUG_ON(info == NULL);
 205         BUG_ON(info->type != IRQT_PIRQ);
 206
 207         return info->u.pirq.vector;
 208 }
 209
 210 static enum xen_irq_type type_from_irq(unsigned irq)
 211 {
 212         return info_for_irq(irq)->type;
 213 }
 214
 215 static unsigned cpu_from_irq(unsigned irq)
 216 {
 217         return info_for_irq(irq)->cpu;
 218 }
 219
 220 static unsigned int cpu_from_evtchn(unsigned int evtchn)
 221 {
 222         int irq = evtchn_to_irq[evtchn];
 223         unsigned ret = 0;
 224
 225         if (irq != -1)
 226                 ret = cpu_from_irq(irq);
 227
 228         return ret;
 229 }
 230
 231 static bool pirq_needs_eoi(unsigned irq)
 232 {
 233         struct irq_info *info = info_for_irq(irq);
 234
 235         BUG_ON(info->type != IRQT_PIRQ);
 236
 237         return info->u.pirq.flags & PIRQ_NEEDS_EOI;
 238 }
 239
 240 static inline unsigned long active_evtchns(unsigned int cpu,
 241                                            struct shared_info *sh,
 242                                            unsigned int idx)
 243 {
 244         return (sh->evtchn_pending[idx] &
 245                 cpu_evtchn_mask(cpu)[idx] &
 246                 ~sh->evtchn_mask[idx]);
 247 }
 248
 249 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
 250 {
 251         int irq = evtchn_to_irq[chn];
 252
 253         BUG_ON(irq == -1);
 254 #ifdef CONFIG_SMP
 255         cpumask_copy(irq_to_desc(irq)->affinity, cpumask_of(cpu));
 256 #endif
 257
 258         __clear_bit(chn, cpu_evtchn_mask(cpu_from_irq(irq)));
 259         __set_bit(chn, cpu_evtchn_mask(cpu));
 260
 261         irq_info[irq].cpu = cpu;
 262 }
 263
 264 static void init_evtchn_cpu_bindings(void)
 265 {
 266 #ifdef CONFIG_SMP
 267         struct irq_desc *desc;
 268         int i;
 269
 270         /* By default all event channels notify CPU#0. */
 271         for_each_irq_desc(i, desc) {
 272                 cpumask_copy(desc->affinity, cpumask_of(0));
 273         }
 274 #endif
 275
 276         memset(cpu_evtchn_mask(0), ~0, sizeof(cpu_evtchn_mask(0)));
 277 }
 278
 279 static inline void clear_evtchn(int port)
 280 {
 281         struct shared_info *s = HYPERVISOR_shared_info;
 282         sync_clear_bit(port, &s->evtchn_pending[0]);
 283 }
 284
 285 static inline void set_evtchn(int port)
 286 {
 287         struct shared_info *s = HYPERVISOR_shared_info;
 288         sync_set_bit(port, &s->evtchn_pending[0]);
 289 }
 290
 291 static inline int test_evtchn(int port)
 292 {
 293         struct shared_info *s = HYPERVISOR_shared_info;
 294         return sync_test_bit(port, &s->evtchn_pending[0]);
 295 }
 296
 297
 298 /**
 299  * notify_remote_via_irq - send event to remote end of event channel via irq
 300  * @irq: irq of event channel to send event to
 301  *
 302  * Unlike notify_remote_via_evtchn(), this is safe to use across
 303  * save/restore. Notifications on a broken connection are silently
 304  * dropped.
 305  */
 306 void notify_remote_via_irq(int irq)
 307 {
 308         int evtchn = evtchn_from_irq(irq);
 309
 310         if (VALID_EVTCHN(evtchn))
 311                 notify_remote_via_evtchn(evtchn);
 312 }
 313 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
 314
 315 static void mask_evtchn(int port)
 316 {
 317         struct shared_info *s = HYPERVISOR_shared_info;
 318         sync_set_bit(port, &s->evtchn_mask[0]);
 319 }
 320
 321 static void unmask_evtchn(int port)
 322 {
 323         struct shared_info *s = HYPERVISOR_shared_info;
 324         unsigned int cpu = get_cpu();
 325
 326         BUG_ON(!irqs_disabled());
 327
 328         /* Slow path (hypercall) if this is a non-local port. */
 329         if (unlikely(cpu != cpu_from_evtchn(port))) {
 330                 struct evtchn_unmask unmask = { .port = port };
 331                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 332         } else {
 333                 struct vcpu_info *vcpu_info = __get_cpu_var(xen_vcpu);
 334
 335                 sync_clear_bit(port, &s->evtchn_mask[0]);
 336
 337                 /*
 338                  * The following is basically the equivalent of
 339                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
 340                  * the interrupt edge' if the channel is masked.
 341                  */
 342                 if (sync_test_bit(port, &s->evtchn_pending[0]) &&
 343                     !sync_test_and_set_bit(port / BITS_PER_LONG,
 344                                            &vcpu_info->evtchn_pending_sel))
 345                         vcpu_info->evtchn_upcall_pending = 1;
 346         }
 347
 348         put_cpu();
 349 }
 350
 351 static int get_nr_hw_irqs(void)
 352 {
 353         int ret = 1;
 354
 355 #ifdef CONFIG_X86_IO_APIC
 356         ret = get_nr_irqs_gsi();
 357 #endif
 358
 359         return ret;
 360 }
 361
 362 static int find_unbound_irq(void)
 363 {
 364         struct irq_data *data;
 365         int irq, res;
 366
 367         for (irq = 0; irq < nr_irqs; irq++) {
 368                 data = irq_get_irq_data(irq);
 369                 /* only 0->15 have init'd desc; handle irq > 16 */
 370                 if (!data)
 371                         break;
 372                 if (data->chip == &no_irq_chip)
 373                         break;
 374                 if (data->chip != &xen_dynamic_chip)
 375                         continue;
 376                 if (irq_info[irq].type == IRQT_UNBOUND)
 377                         return irq;
 378         }
 379
 380         if (irq == nr_irqs)
 381                 panic("No available IRQ to bind to: increase nr_irqs!\n");
 382
 383         res = irq_alloc_desc_at(irq, 0);
 384
 385         if (WARN_ON(res != irq))
 386                 return -1;
 387
 388         return irq;
 389 }
 390
 391 static bool identity_mapped_irq(unsigned irq)
 392 {
 393         /* identity map all the hardware irqs */
 394         return irq < get_nr_hw_irqs();
 395 }
 396
 397 static void pirq_unmask_notify(int irq)
 398 {
 399         struct physdev_eoi eoi = { .irq = irq };
 400
 401         if (unlikely(pirq_needs_eoi(irq))) {
 402                 int rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
 403                 WARN_ON(rc);
 404         }
 405 }
 406
 407 static void pirq_query_unmask(int irq)
 408 {
 409         struct physdev_irq_status_query irq_status;
 410         struct irq_info *info = info_for_irq(irq);
 411
 412         BUG_ON(info->type != IRQT_PIRQ);
 413
 414         irq_status.irq = irq;
 415         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
 416                 irq_status.flags = 0;
 417
 418         info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
 419         if (irq_status.flags & XENIRQSTAT_needs_eoi)
 420                 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
 421 }
 422
 423 static bool probing_irq(int irq)
 424 {
 425         struct irq_desc *desc = irq_to_desc(irq);
 426
 427         return desc && desc->action == NULL;
 428 }
 429
 430 static unsigned int startup_pirq(unsigned int irq)
 431 {
 432         struct evtchn_bind_pirq bind_pirq;
 433         struct irq_info *info = info_for_irq(irq);
 434         int evtchn = evtchn_from_irq(irq);
 435
 436         BUG_ON(info->type != IRQT_PIRQ);
 437
 438         if (VALID_EVTCHN(evtchn))
 439                 goto out;
 440
 441         bind_pirq.pirq = irq;
 442         /* NB. We are happy to share unless we are probing. */
 443         bind_pirq.flags = probing_irq(irq) ? 0 : BIND_PIRQ__WILL_SHARE;
 444         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq) != 0) {
 445                 if (!probing_irq(irq))
 446                         printk(KERN_INFO "Failed to obtain physical IRQ %d\n",
 447                                irq);
 448                 return 0;
 449         }
 450         evtchn = bind_pirq.port;
 451
 452         pirq_query_unmask(irq);
 453
 454         evtchn_to_irq[evtchn] = irq;
 455         bind_evtchn_to_cpu(evtchn, 0);
 456         info->evtchn = evtchn;
 457
 458 out:
 459         unmask_evtchn(evtchn);
 460         pirq_unmask_notify(irq);
 461
 462         return 0;
 463 }
 464
 465 static void shutdown_pirq(unsigned int irq)
 466 {
 467         struct evtchn_close close;
 468         struct irq_info *info = info_for_irq(irq);
 469         int evtchn = evtchn_from_irq(irq);
 470
 471         BUG_ON(info->type != IRQT_PIRQ);
 472
 473         if (!VALID_EVTCHN(evtchn))
 474                 return;
 475
 476         mask_evtchn(evtchn);
 477
 478         close.port = evtchn;
 479         if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 480                 BUG();
 481
 482         bind_evtchn_to_cpu(evtchn, 0);
 483         evtchn_to_irq[evtchn] = -1;
 484         info->evtchn = 0;
 485 }
 486
 487 static void enable_pirq(unsigned int irq)
 488 {
 489         startup_pirq(irq);
 490 }
 491
 492 static void disable_pirq(unsigned int irq)
 493 {
 494 }
 495
 496 static void ack_pirq(unsigned int irq)
 497 {
 498         int evtchn = evtchn_from_irq(irq);
 499
 500         move_native_irq(irq);
 501
 502         if (VALID_EVTCHN(evtchn)) {
 503                 mask_evtchn(evtchn);
 504                 clear_evtchn(evtchn);
 505         }
 506 }
 507
 508 static void end_pirq(unsigned int irq)
 509 {
 510         int evtchn = evtchn_from_irq(irq);
 511         struct irq_desc *desc = irq_to_desc(irq);
 512
 513         if (WARN_ON(!desc))
 514                 return;
 515
 516         if ((desc->status & (IRQ_DISABLED|IRQ_PENDING)) ==
 517             (IRQ_DISABLED|IRQ_PENDING)) {
 518                 shutdown_pirq(irq);
 519         } else if (VALID_EVTCHN(evtchn)) {
 520                 unmask_evtchn(evtchn);
 521                 pirq_unmask_notify(irq);
 522         }
 523 }
 524
 525 static int find_irq_by_gsi(unsigned gsi)
 526 {
 527         int irq;
 528
 529         for (irq = 0; irq < nr_irqs; irq++) {
 530                 struct irq_info *info = info_for_irq(irq);
 531
 532                 if (info == NULL || info->type != IRQT_PIRQ)
 533                         continue;
 534
 535                 if (gsi_from_irq(irq) == gsi)
 536                         return irq;
 537         }
 538
 539         return -1;
 540 }
 541
 542 /*
 543  * Allocate a physical irq, along with a vector.  We don't assign an
 544  * event channel until the irq actually started up.  Return an
 545  * existing irq if we've already got one for the gsi.
 546  */
 547 int xen_allocate_pirq(unsigned gsi, char *name)
 548 {
 549         int irq;
 550         struct physdev_irq irq_op;
 551
 552         spin_lock(&irq_mapping_update_lock);
 553
 554         irq = find_irq_by_gsi(gsi);
 555         if (irq != -1) {
 556                 printk(KERN_INFO "xen_allocate_pirq: returning irq %d for gsi %u\n",
 557                        irq, gsi);
 558                 goto out;       /* XXX need refcount? */
 559         }
 560
 561         if (identity_mapped_irq(gsi)) {
 562                 irq = gsi;
 563                 irq_to_desc_alloc_node(irq, 0);
 564                 dynamic_irq_init(irq);
 565         } else
 566                 irq = find_unbound_irq();
 567
 568         set_irq_chip_and_handler_name(irq, &xen_pirq_chip,
 569                                       handle_level_irq, name);
 570
 571         irq_op.irq = irq;
 572         if (HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
 573                 dynamic_irq_cleanup(irq);
 574                 irq = -ENOSPC;
 575                 goto out;
 576         }
 577
 578         irq_info[irq] = mk_pirq_info(0, gsi, irq_op.vector);
 579
 580 out:
 581         spin_unlock(&irq_mapping_update_lock);
 582
 583         return irq;
 584 }
 585
 586 int xen_vector_from_irq(unsigned irq)
 587 {
 588         return vector_from_irq(irq);
 589 }
 590
 591 int xen_gsi_from_irq(unsigned irq)
 592 {
 593         return gsi_from_irq(irq);
 594 }
 595
 596 int bind_evtchn_to_irq(unsigned int evtchn)
 597 {
 598         int irq;
 599
 600         spin_lock(&irq_mapping_update_lock);
 601
 602         irq = evtchn_to_irq[evtchn];
 603
 604         if (irq == -1) {
 605                 irq = find_unbound_irq();
 606
 607                 set_irq_chip_and_handler_name(irq, &xen_dynamic_chip,
 608                                               handle_edge_irq, "event");
 609
 610                 evtchn_to_irq[evtchn] = irq;
 611                 irq_info[irq] = mk_evtchn_info(evtchn);
 612         }
 613
 614         spin_unlock(&irq_mapping_update_lock);
 615
 616         return irq;
 617 }
 618 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
 619
 620 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
 621 {
 622         struct evtchn_bind_ipi bind_ipi;
 623         int evtchn, irq;
 624
 625         spin_lock(&irq_mapping_update_lock);
 626
 627         irq = per_cpu(ipi_to_irq, cpu)[ipi];
 628
 629         if (irq == -1) {
 630                 irq = find_unbound_irq();
 631                 if (irq < 0)
 632                         goto out;
 633
 634                 set_irq_chip_and_handler_name(irq, &xen_percpu_chip,
 635                                               handle_percpu_irq, "ipi");
 636
 637                 bind_ipi.vcpu = cpu;
 638                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
 639                                                 &bind_ipi) != 0)
 640                         BUG();
 641                 evtchn = bind_ipi.port;
 642
 643                 evtchn_to_irq[evtchn] = irq;
 644                 irq_info[irq] = mk_ipi_info(evtchn, ipi);
 645                 per_cpu(ipi_to_irq, cpu)[ipi] = irq;
 646
 647                 bind_evtchn_to_cpu(evtchn, cpu);
 648         }
 649
 650  out:
 651         spin_unlock(&irq_mapping_update_lock);
 652         return irq;
 653 }
 654
 655
 656 static int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
 657 {
 658         struct evtchn_bind_virq bind_virq;
 659         int evtchn, irq;
 660
 661         spin_lock(&irq_mapping_update_lock);
 662
 663         irq = per_cpu(virq_to_irq, cpu)[virq];
 664
 665         if (irq == -1) {
 666                 bind_virq.virq = virq;
 667                 bind_virq.vcpu = cpu;
 668                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 669                                                 &bind_virq) != 0)
 670                         BUG();
 671                 evtchn = bind_virq.port;
 672
 673                 irq = find_unbound_irq();
 674
 675                 set_irq_chip_and_handler_name(irq, &xen_percpu_chip,
 676                                               handle_percpu_irq, "virq");
 677
 678                 evtchn_to_irq[evtchn] = irq;
 679                 irq_info[irq] = mk_virq_info(evtchn, virq);
 680
 681                 per_cpu(virq_to_irq, cpu)[virq] = irq;
 682
 683                 bind_evtchn_to_cpu(evtchn, cpu);
 684         }
 685
 686         spin_unlock(&irq_mapping_update_lock);
 687
 688         return irq;
 689 }
 690
 691 static void unbind_from_irq(unsigned int irq)
 692 {
 693         struct evtchn_close close;
 694         int evtchn = evtchn_from_irq(irq);
 695
 696         spin_lock(&irq_mapping_update_lock);
 697
 698         if (VALID_EVTCHN(evtchn)) {
 699                 close.port = evtchn;
 700                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 701                         BUG();
 702
 703                 switch (type_from_irq(irq)) {
 704                 case IRQT_VIRQ:
 705                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
 706                                 [virq_from_irq(irq)] = -1;
 707                         break;
 708                 case IRQT_IPI:
 709                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
 710                                 [ipi_from_irq(irq)] = -1;
 711                         break;
 712                 default:
 713                         break;
 714                 }
 715
 716                 /* Closed ports are implicitly re-bound to VCPU0. */
 717                 bind_evtchn_to_cpu(evtchn, 0);
 718
 719                 evtchn_to_irq[evtchn] = -1;
 720         }
 721
 722         if (irq_info[irq].type != IRQT_UNBOUND) {
 723                 irq_info[irq] = mk_unbound_info();
 724
 725                 irq_free_desc(irq);
 726         }
 727
 728         spin_unlock(&irq_mapping_update_lock);
 729 }
 730
 731 int bind_evtchn_to_irqhandler(unsigned int evtchn,
 732                               irq_handler_t handler,
 733                               unsigned long irqflags,
 734                               const char *devname, void *dev_id)
 735 {
 736         unsigned int irq;
 737         int retval;
 738
 739         irq = bind_evtchn_to_irq(evtchn);
 740         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 741         if (retval != 0) {
 742                 unbind_from_irq(irq);
 743                 return retval;
 744         }
 745
 746         return irq;
 747 }
 748 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
 749
 750 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
 751                             irq_handler_t handler,
 752                             unsigned long irqflags, const char *devname, void *dev_id)
 753 {
 754         unsigned int irq;
 755         int retval;
 756
 757         irq = bind_virq_to_irq(virq, cpu);
 758         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 759         if (retval != 0) {
 760                 unbind_from_irq(irq);
 761                 return retval;
 762         }
 763
 764         return irq;
 765 }
 766 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
 767
 768 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
 769                            unsigned int cpu,
 770                            irq_handler_t handler,
 771                            unsigned long irqflags,
 772                            const char *devname,
 773                            void *dev_id)
 774 {
 775         int irq, retval;
 776
 777         irq = bind_ipi_to_irq(ipi, cpu);
 778         if (irq < 0)
 779                 return irq;
 780
 781         irqflags |= IRQF_NO_SUSPEND;
 782         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 783         if (retval != 0) {
 784                 unbind_from_irq(irq);
 785                 return retval;
 786         }
 787
 788         return irq;
 789 }
 790
 791 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
 792 {
 793         free_irq(irq, dev_id);
 794         unbind_from_irq(irq);
 795 }
 796 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
 797
 798 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
 799 {
 800         int irq = per_cpu(ipi_to_irq, cpu)[vector];
 801         BUG_ON(irq < 0);
 802         notify_remote_via_irq(irq);
 803 }
 804
 805 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
 806 {
 807         struct shared_info *sh = HYPERVISOR_shared_info;
 808         int cpu = smp_processor_id();
 809         int i;
 810         unsigned long flags;
 811         static DEFINE_SPINLOCK(debug_lock);
 812
 813         spin_lock_irqsave(&debug_lock, flags);
 814
 815         printk("vcpu %d\n  ", cpu);
 816
 817         for_each_online_cpu(i) {
 818                 struct vcpu_info *v = per_cpu(xen_vcpu, i);
 819                 printk("%d: masked=%d pending=%d event_sel %08lx\n  ", i,
 820                         (get_irq_regs() && i == cpu) ? xen_irqs_disabled(get_irq_regs()) : v->evtchn_upcall_mask,
 821                         v->evtchn_upcall_pending,
 822                         v->evtchn_pending_sel);
 823         }
 824         printk("pending:\n   ");
 825         for(i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
 826                 printk("%08lx%s", sh->evtchn_pending[i],
 827                         i % 8 == 0 ? "\n   " : " ");
 828         printk("\nmasks:\n   ");
 829         for(i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
 830                 printk("%08lx%s", sh->evtchn_mask[i],
 831                         i % 8 == 0 ? "\n   " : " ");
 832
 833         printk("\nunmasked:\n   ");
 834         for(i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
 835                 printk("%08lx%s", sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
 836                         i % 8 == 0 ? "\n   " : " ");
 837
 838         printk("\npending list:\n");
 839         for(i = 0; i < NR_EVENT_CHANNELS; i++) {
 840                 if (sync_test_bit(i, sh->evtchn_pending)) {
 841                         printk("  %d: event %d -> irq %d\n",
 842                                cpu_from_evtchn(i), i,
 843                                evtchn_to_irq[i]);
 844                 }
 845         }
 846
 847         spin_unlock_irqrestore(&debug_lock, flags);
 848
 849         return IRQ_HANDLED;
 850 }
 851
 852 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
 853
 854 /*
 855  * Search the CPUs pending events bitmasks.  For each one found, map
 856  * the event number to an irq, and feed it into do_IRQ() for
 857  * handling.
 858  *
 859  * Xen uses a two-level bitmap to speed searching.  The first level is
 860  * a bitset of words which contain pending event bits.  The second
 861  * level is a bitset of pending events themselves.
 862  */
 863 static void __xen_evtchn_do_upcall(void)
 864 {
 865         int cpu = get_cpu();
 866         struct shared_info *s = HYPERVISOR_shared_info;
 867         struct vcpu_info *vcpu_info = __get_cpu_var(xen_vcpu);
 868         unsigned count;
 869
 870         do {
 871                 unsigned long pending_words;
 872
 873                 vcpu_info->evtchn_upcall_pending = 0;
 874
 875                 if (__get_cpu_var(xed_nesting_count)++)
 876                         goto out;
 877
 878 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
 879                 /* Clear master flag /before/ clearing selector flag. */
 880                 wmb();
 881 #endif
 882                 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
 883                 while (pending_words != 0) {
 884                         unsigned long pending_bits;
 885                         int word_idx = __ffs(pending_words);
 886                         pending_words &= ~(1UL << word_idx);
 887
 888                         while ((pending_bits = active_evtchns(cpu, s, word_idx)) != 0) {
 889                                 int bit_idx = __ffs(pending_bits);
 890                                 int port = (word_idx * BITS_PER_LONG) + bit_idx;
 891                                 int irq = evtchn_to_irq[port];
 892                                 struct irq_desc *desc;
 893
 894                                 if (irq != -1) {
 895                                         desc = irq_to_desc(irq);
 896                                         if (desc)
 897                                                 generic_handle_irq_desc(irq, desc);
 898                                 }
 899                         }
 900                 }
 901
 902                 BUG_ON(!irqs_disabled());
 903
 904                 count = __get_cpu_var(xed_nesting_count);
 905                 __get_cpu_var(xed_nesting_count) = 0;
 906         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
 907
 908 out:
 909
 910         put_cpu();
 911 }
 912
 913 void xen_evtchn_do_upcall(struct pt_regs *regs)
 914 {
 915         struct pt_regs *old_regs = set_irq_regs(regs);
 916
 917         exit_idle();
 918         irq_enter();
 919
 920         __xen_evtchn_do_upcall();
 921
 922         irq_exit();
 923         set_irq_regs(old_regs);
 924 }
 925
 926 void xen_hvm_evtchn_do_upcall(void)
 927 {
 928         __xen_evtchn_do_upcall();
 929 }
 930 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
 931
 932 /* Rebind a new event channel to an existing irq. */
 933 void rebind_evtchn_irq(int evtchn, int irq)
 934 {
 935         struct irq_info *info = info_for_irq(irq);
 936
 937         /* Make sure the irq is masked, since the new event channel
 938            will also be masked. */
 939         disable_irq(irq);
 940
 941         spin_lock(&irq_mapping_update_lock);
 942
 943         /* After resume the irq<->evtchn mappings are all cleared out */
 944         BUG_ON(evtchn_to_irq[evtchn] != -1);
 945         /* Expect irq to have been bound before,
 946            so there should be a proper type */
 947         BUG_ON(info->type == IRQT_UNBOUND);
 948
 949         evtchn_to_irq[evtchn] = irq;
 950         irq_info[irq] = mk_evtchn_info(evtchn);
 951
 952         spin_unlock(&irq_mapping_update_lock);
 953
 954         /* new event channels are always bound to cpu 0 */
 955         irq_set_affinity(irq, cpumask_of(0));
 956
 957         /* Unmask the event channel. */
 958         enable_irq(irq);
 959 }
 960
 961 /* Rebind an evtchn so that it gets delivered to a specific cpu */
 962 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
 963 {
 964         struct evtchn_bind_vcpu bind_vcpu;
 965         int evtchn = evtchn_from_irq(irq);
 966
 967         /* events delivered via platform PCI interrupts are always
 968          * routed to vcpu 0 */
 969         if (!VALID_EVTCHN(evtchn) ||
 970                 (xen_hvm_domain() && !xen_have_vector_callback))
 971                 return -1;
 972
 973         /* Send future instances of this interrupt to other vcpu. */
 974         bind_vcpu.port = evtchn;
 975         bind_vcpu.vcpu = tcpu;
 976
 977         /*
 978          * If this fails, it usually just indicates that we're dealing with a
 979          * virq or IPI channel, which don't actually need to be rebound. Ignore
 980          * it, but don't do the xenlinux-level rebind in that case.
 981          */
 982         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
 983                 bind_evtchn_to_cpu(evtchn, tcpu);
 984
 985         return 0;
 986 }
 987
 988 static int set_affinity_irq(unsigned irq, const struct cpumask *dest)
 989 {
 990         unsigned tcpu = cpumask_first(dest);
 991
 992         return rebind_irq_to_cpu(irq, tcpu);
 993 }
 994
 995 int resend_irq_on_evtchn(unsigned int irq)
 996 {
 997         int masked, evtchn = evtchn_from_irq(irq);
 998         struct shared_info *s = HYPERVISOR_shared_info;
 999
1000         if (!VALID_EVTCHN(evtchn))
1001                 return 1;
1002
1003         masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
1004         sync_set_bit(evtchn, s->evtchn_pending);
1005         if (!masked)
1006                 unmask_evtchn(evtchn);
1007
1008         return 1;
1009 }
1010
1011 static void enable_dynirq(unsigned int irq)
1012 {
1013         int evtchn = evtchn_from_irq(irq);
1014
1015         if (VALID_EVTCHN(evtchn))
1016                 unmask_evtchn(evtchn);
1017 }
1018
1019 static void disable_dynirq(unsigned int irq)
1020 {
1021         int evtchn = evtchn_from_irq(irq);
1022
1023         if (VALID_EVTCHN(evtchn))
1024                 mask_evtchn(evtchn);
1025 }
1026
1027 static void ack_dynirq(unsigned int irq)
1028 {
1029         int evtchn = evtchn_from_irq(irq);
1030
1031         move_native_irq(irq);
1032
1033         if (VALID_EVTCHN(evtchn))
1034                 clear_evtchn(evtchn);
1035 }
1036
1037 static int retrigger_dynirq(unsigned int irq)
1038 {
1039         int evtchn = evtchn_from_irq(irq);
1040         struct shared_info *sh = HYPERVISOR_shared_info;
1041         int ret = 0;
1042
1043         if (VALID_EVTCHN(evtchn)) {
1044                 int masked;
1045
1046                 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
1047                 sync_set_bit(evtchn, sh->evtchn_pending);
1048                 if (!masked)
1049                         unmask_evtchn(evtchn);
1050                 ret = 1;
1051         }
1052
1053         return ret;
1054 }
1055
1056 static void restore_cpu_virqs(unsigned int cpu)
1057 {
1058         struct evtchn_bind_virq bind_virq;
1059         int virq, irq, evtchn;
1060
1061         for (virq = 0; virq < NR_VIRQS; virq++) {
1062                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1063                         continue;
1064
1065                 BUG_ON(virq_from_irq(irq) != virq);
1066
1067                 /* Get a new binding from Xen. */
1068                 bind_virq.virq = virq;
1069                 bind_virq.vcpu = cpu;
1070                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1071                                                 &bind_virq) != 0)
1072                         BUG();
1073                 evtchn = bind_virq.port;
1074
1075                 /* Record the new mapping. */
1076                 evtchn_to_irq[evtchn] = irq;
1077                 irq_info[irq] = mk_virq_info(evtchn, virq);
1078                 bind_evtchn_to_cpu(evtchn, cpu);
1079
1080                 /* Ready for use. */
1081                 unmask_evtchn(evtchn);
1082         }
1083 }
1084
1085 static void restore_cpu_ipis(unsigned int cpu)
1086 {
1087         struct evtchn_bind_ipi bind_ipi;
1088         int ipi, irq, evtchn;
1089
1090         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1091                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1092                         continue;
1093
1094                 BUG_ON(ipi_from_irq(irq) != ipi);
1095
1096                 /* Get a new binding from Xen. */
1097                 bind_ipi.vcpu = cpu;
1098                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1099                                                 &bind_ipi) != 0)
1100                         BUG();
1101                 evtchn = bind_ipi.port;
1102
1103                 /* Record the new mapping. */
1104                 evtchn_to_irq[evtchn] = irq;
1105                 irq_info[irq] = mk_ipi_info(evtchn, ipi);
1106                 bind_evtchn_to_cpu(evtchn, cpu);
1107
1108                 /* Ready for use. */
1109                 unmask_evtchn(evtchn);
1110
1111         }
1112 }
1113
1114 /* Clear an irq's pending state, in preparation for polling on it */
1115 void xen_clear_irq_pending(int irq)
1116 {
1117         int evtchn = evtchn_from_irq(irq);
1118
1119         if (VALID_EVTCHN(evtchn))
1120                 clear_evtchn(evtchn);
1121 }
1122
1123 void xen_set_irq_pending(int irq)
1124 {
1125         int evtchn = evtchn_from_irq(irq);
1126
1127         if (VALID_EVTCHN(evtchn))
1128                 set_evtchn(evtchn);
1129 }
1130
1131 bool xen_test_irq_pending(int irq)
1132 {
1133         int evtchn = evtchn_from_irq(irq);
1134         bool ret = false;
1135
1136         if (VALID_EVTCHN(evtchn))
1137                 ret = test_evtchn(evtchn);
1138
1139         return ret;
1140 }
1141
1142 /* Poll waiting for an irq to become pending.  In the usual case, the
1143    irq will be disabled so it won't deliver an interrupt. */
1144 void xen_poll_irq(int irq)
1145 {
1146         evtchn_port_t evtchn = evtchn_from_irq(irq);
1147
1148         if (VALID_EVTCHN(evtchn)) {
1149                 struct sched_poll poll;
1150
1151                 poll.nr_ports = 1;
1152                 poll.timeout = 0;
1153                 set_xen_guest_handle(poll.ports, &evtchn);
1154
1155                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1156                         BUG();
1157         }
1158 }
1159
1160 void xen_irq_resume(void)
1161 {
1162         unsigned int cpu, irq, evtchn;
1163
1164         init_evtchn_cpu_bindings();
1165
1166         /* New event-channel space is not 'live' yet. */
1167         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1168                 mask_evtchn(evtchn);
1169
1170         /* No IRQ <-> event-channel mappings. */
1171         for (irq = 0; irq < nr_irqs; irq++)
1172                 irq_info[irq].evtchn = 0; /* zap event-channel binding */
1173
1174         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1175                 evtchn_to_irq[evtchn] = -1;
1176
1177         for_each_possible_cpu(cpu) {
1178                 restore_cpu_virqs(cpu);
1179                 restore_cpu_ipis(cpu);
1180         }
1181 }
1182
1183 static struct irq_chip xen_dynamic_chip __read_mostly = {
1184         .name           = "xen-dyn",
1185
1186         .disable        = disable_dynirq,
1187         .mask           = disable_dynirq,
1188         .unmask         = enable_dynirq,
1189
1190         .ack            = ack_dynirq,
1191         .set_affinity   = set_affinity_irq,
1192         .retrigger      = retrigger_dynirq,
1193 };
1194
1195 static struct irq_chip xen_pirq_chip __read_mostly = {
1196         .name           = "xen-pirq",
1197
1198         .startup        = startup_pirq,
1199         .shutdown       = shutdown_pirq,
1200
1201         .enable         = enable_pirq,
1202         .unmask         = enable_pirq,
1203
1204         .disable        = disable_pirq,
1205         .mask           = disable_pirq,
1206
1207         .ack            = ack_pirq,
1208         .end            = end_pirq,
1209
1210         .set_affinity   = set_affinity_irq,
1211
1212         .retrigger      = retrigger_dynirq,
1213 };
1214
1215 static struct irq_chip xen_percpu_chip __read_mostly = {
1216         .name           = "xen-percpu",
1217
1218         .disable        = disable_dynirq,
1219         .mask           = disable_dynirq,
1220         .unmask         = enable_dynirq,
1221
1222         .ack            = ack_dynirq,
1223 };
1224
1225 int xen_set_callback_via(uint64_t via)
1226 {
1227         struct xen_hvm_param a;
1228         a.domid = DOMID_SELF;
1229         a.index = HVM_PARAM_CALLBACK_IRQ;
1230         a.value = via;
1231         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1232 }
1233 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1234
1235 #ifdef CONFIG_XEN_PVHVM
1236 /* Vector callbacks are better than PCI interrupts to receive event
1237  * channel notifications because we can receive vector callbacks on any
1238  * vcpu and we don't need PCI support or APIC interactions. */
1239 void xen_callback_vector(void)
1240 {
1241         int rc;
1242         uint64_t callback_via;
1243         if (xen_have_vector_callback) {
1244                 callback_via = HVM_CALLBACK_VECTOR(XEN_HVM_EVTCHN_CALLBACK);
1245                 rc = xen_set_callback_via(callback_via);
1246                 if (rc) {
1247                         printk(KERN_ERR "Request for Xen HVM callback vector"
1248                                         " failed.\n");
1249                         xen_have_vector_callback = 0;
1250                         return;
1251                 }
1252                 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1253                                 "enabled\n");
1254                 /* in the restore case the vector has already been allocated */
1255                 if (!test_bit(XEN_HVM_EVTCHN_CALLBACK, used_vectors))
1256                         alloc_intr_gate(XEN_HVM_EVTCHN_CALLBACK, xen_hvm_callback_vector);
1257         }
1258 }
1259 #else
1260 void xen_callback_vector(void) {}
1261 #endif
1262
1263 void __init xen_init_IRQ(void)
1264 {
1265         int i;
1266
1267         cpu_evtchn_mask_p = kcalloc(nr_cpu_ids, sizeof(struct cpu_evtchn_s),
1268                                     GFP_KERNEL);
1269         irq_info = kcalloc(nr_irqs, sizeof(*irq_info), GFP_KERNEL);
1270
1271         evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1272                                     GFP_KERNEL);
1273         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1274                 evtchn_to_irq[i] = -1;
1275
1276         init_evtchn_cpu_bindings();
1277
1278         /* No event channels are 'live' right now. */
1279         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1280                 mask_evtchn(i);
1281
1282         if (xen_hvm_domain()) {
1283                 xen_callback_vector();
1284                 native_init_IRQ();
1285         } else {
1286                 irq_ctx_init(smp_processor_id());
1287         }
1288 }