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