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 received, 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 #include <linux/pci.h>
  33
  34 #include <asm/desc.h>
  35 #include <asm/ptrace.h>
  36 #include <asm/irq.h>
  37 #include <asm/idle.h>
  38 #include <asm/io_apic.h>
  39 #include <asm/sync_bitops.h>
  40 #include <asm/xen/page.h>
  41 #include <asm/xen/pci.h>
  42 #include <asm/xen/hypercall.h>
  43 #include <asm/xen/hypervisor.h>
  44
  45 #include <xen/xen.h>
  46 #include <xen/hvm.h>
  47 #include <xen/xen-ops.h>
  48 #include <xen/events.h>
  49 #include <xen/interface/xen.h>
  50 #include <xen/interface/event_channel.h>
  51 #include <xen/interface/hvm/hvm_op.h>
  52 #include <xen/interface/hvm/params.h>
  53
  54 /*
  55  * This lock protects updates to the following mapping and reference-count
  56  * arrays. The lock does not need to be acquired to read the mapping tables.
  57  */
  58 static DEFINE_MUTEX(irq_mapping_update_lock);
  59
  60 static LIST_HEAD(xen_irq_list_head);
  61
  62 /* IRQ <-> VIRQ mapping. */
  63 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
  64
  65 /* IRQ <-> IPI mapping */
  66 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
  67
  68 /* Interrupt types. */
  69 enum xen_irq_type {
  70         IRQT_UNBOUND = 0,
  71         IRQT_PIRQ,
  72         IRQT_VIRQ,
  73         IRQT_IPI,
  74         IRQT_EVTCHN
  75 };
  76
  77 /*
  78  * Packed IRQ information:
  79  * type - enum xen_irq_type
  80  * event channel - irq->event channel mapping
  81  * cpu - cpu this event channel is bound to
  82  * index - type-specific information:
  83  *    PIRQ - vector, with MSB being "needs EIO", or physical IRQ of the HVM
  84  *           guest, or GSI (real passthrough IRQ) of the device.
  85  *    VIRQ - virq number
  86  *    IPI - IPI vector
  87  *    EVTCHN -
  88  */
  89 struct irq_info {
  90         struct list_head list;
  91         int refcnt;
  92         enum xen_irq_type type; /* type */
  93         unsigned irq;
  94         unsigned short evtchn;  /* event channel */
  95         unsigned short cpu;     /* cpu bound */
  96
  97         union {
  98                 unsigned short virq;
  99                 enum ipi_vector ipi;
 100                 struct {
 101                         unsigned short pirq;
 102                         unsigned short gsi;
 103                         unsigned char vector;
 104                         unsigned char flags;
 105                         uint16_t domid;
 106                 } pirq;
 107         } u;
 108 };
 109 #define PIRQ_NEEDS_EOI  (1 << 0)
 110 #define PIRQ_SHAREABLE  (1 << 1)
 111
 112 static int *evtchn_to_irq;
 113 static unsigned long *pirq_eoi_map;
 114 static bool (*pirq_needs_eoi)(unsigned irq);
 115
 116 static DEFINE_PER_CPU(unsigned long [NR_EVENT_CHANNELS/BITS_PER_LONG],
 117                       cpu_evtchn_mask);
 118
 119 /* Xen will never allocate port zero for any purpose. */
 120 #define VALID_EVTCHN(chn)       ((chn) != 0)
 121
 122 static struct irq_chip xen_dynamic_chip;
 123 static struct irq_chip xen_percpu_chip;
 124 static struct irq_chip xen_pirq_chip;
 125 static void enable_dynirq(struct irq_data *data);
 126 static void disable_dynirq(struct irq_data *data);
 127
 128 /* Get info for IRQ */
 129 static struct irq_info *info_for_irq(unsigned irq)
 130 {
 131         return irq_get_handler_data(irq);
 132 }
 133
 134 /* Constructors for packed IRQ information. */
 135 static void xen_irq_info_common_init(struct irq_info *info,
 136                                      unsigned irq,
 137                                      enum xen_irq_type type,
 138                                      unsigned short evtchn,
 139                                      unsigned short cpu)
 140 {
 141
 142         BUG_ON(info->type != IRQT_UNBOUND && info->type != type);
 143
 144         info->type = type;
 145         info->irq = irq;
 146         info->evtchn = evtchn;
 147         info->cpu = cpu;
 148
 149         evtchn_to_irq[evtchn] = irq;
 150 }
 151
 152 static void xen_irq_info_evtchn_init(unsigned irq,
 153                                      unsigned short evtchn)
 154 {
 155         struct irq_info *info = info_for_irq(irq);
 156
 157         xen_irq_info_common_init(info, irq, IRQT_EVTCHN, evtchn, 0);
 158 }
 159
 160 static void xen_irq_info_ipi_init(unsigned cpu,
 161                                   unsigned irq,
 162                                   unsigned short evtchn,
 163                                   enum ipi_vector ipi)
 164 {
 165         struct irq_info *info = info_for_irq(irq);
 166
 167         xen_irq_info_common_init(info, irq, IRQT_IPI, evtchn, 0);
 168
 169         info->u.ipi = ipi;
 170
 171         per_cpu(ipi_to_irq, cpu)[ipi] = irq;
 172 }
 173
 174 static void xen_irq_info_virq_init(unsigned cpu,
 175                                    unsigned irq,
 176                                    unsigned short evtchn,
 177                                    unsigned short virq)
 178 {
 179         struct irq_info *info = info_for_irq(irq);
 180
 181         xen_irq_info_common_init(info, irq, IRQT_VIRQ, evtchn, 0);
 182
 183         info->u.virq = virq;
 184
 185         per_cpu(virq_to_irq, cpu)[virq] = irq;
 186 }
 187
 188 static void xen_irq_info_pirq_init(unsigned irq,
 189                                    unsigned short evtchn,
 190                                    unsigned short pirq,
 191                                    unsigned short gsi,
 192                                    unsigned short vector,
 193                                    uint16_t domid,
 194                                    unsigned char flags)
 195 {
 196         struct irq_info *info = info_for_irq(irq);
 197
 198         xen_irq_info_common_init(info, irq, IRQT_PIRQ, evtchn, 0);
 199
 200         info->u.pirq.pirq = pirq;
 201         info->u.pirq.gsi = gsi;
 202         info->u.pirq.vector = vector;
 203         info->u.pirq.domid = domid;
 204         info->u.pirq.flags = flags;
 205 }
 206
 207 /*
 208  * Accessors for packed IRQ information.
 209  */
 210 static unsigned int evtchn_from_irq(unsigned irq)
 211 {
 212         if (unlikely(WARN(irq < 0 || irq >= nr_irqs, "Invalid irq %d!\n", irq)))
 213                 return 0;
 214
 215         return info_for_irq(irq)->evtchn;
 216 }
 217
 218 unsigned irq_from_evtchn(unsigned int evtchn)
 219 {
 220         return evtchn_to_irq[evtchn];
 221 }
 222 EXPORT_SYMBOL_GPL(irq_from_evtchn);
 223
 224 static enum ipi_vector ipi_from_irq(unsigned irq)
 225 {
 226         struct irq_info *info = info_for_irq(irq);
 227
 228         BUG_ON(info == NULL);
 229         BUG_ON(info->type != IRQT_IPI);
 230
 231         return info->u.ipi;
 232 }
 233
 234 static unsigned virq_from_irq(unsigned irq)
 235 {
 236         struct irq_info *info = info_for_irq(irq);
 237
 238         BUG_ON(info == NULL);
 239         BUG_ON(info->type != IRQT_VIRQ);
 240
 241         return info->u.virq;
 242 }
 243
 244 static unsigned pirq_from_irq(unsigned irq)
 245 {
 246         struct irq_info *info = info_for_irq(irq);
 247
 248         BUG_ON(info == NULL);
 249         BUG_ON(info->type != IRQT_PIRQ);
 250
 251         return info->u.pirq.pirq;
 252 }
 253
 254 static enum xen_irq_type type_from_irq(unsigned irq)
 255 {
 256         return info_for_irq(irq)->type;
 257 }
 258
 259 static unsigned cpu_from_irq(unsigned irq)
 260 {
 261         return info_for_irq(irq)->cpu;
 262 }
 263
 264 static unsigned int cpu_from_evtchn(unsigned int evtchn)
 265 {
 266         int irq = evtchn_to_irq[evtchn];
 267         unsigned ret = 0;
 268
 269         if (irq != -1)
 270                 ret = cpu_from_irq(irq);
 271
 272         return ret;
 273 }
 274
 275 static bool pirq_check_eoi_map(unsigned irq)
 276 {
 277         return test_bit(pirq_from_irq(irq), pirq_eoi_map);
 278 }
 279
 280 static bool pirq_needs_eoi_flag(unsigned irq)
 281 {
 282         struct irq_info *info = info_for_irq(irq);
 283         BUG_ON(info->type != IRQT_PIRQ);
 284
 285         return info->u.pirq.flags & PIRQ_NEEDS_EOI;
 286 }
 287
 288 static inline unsigned long active_evtchns(unsigned int cpu,
 289                                            struct shared_info *sh,
 290                                            unsigned int idx)
 291 {
 292         return sh->evtchn_pending[idx] &
 293                 per_cpu(cpu_evtchn_mask, cpu)[idx] &
 294                 ~sh->evtchn_mask[idx];
 295 }
 296
 297 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
 298 {
 299         int irq = evtchn_to_irq[chn];
 300
 301         BUG_ON(irq == -1);
 302 #ifdef CONFIG_SMP
 303         cpumask_copy(irq_to_desc(irq)->irq_data.affinity, cpumask_of(cpu));
 304 #endif
 305
 306         clear_bit(chn, per_cpu(cpu_evtchn_mask, cpu_from_irq(irq)));
 307         set_bit(chn, per_cpu(cpu_evtchn_mask, cpu));
 308
 309         info_for_irq(irq)->cpu = cpu;
 310 }
 311
 312 static void init_evtchn_cpu_bindings(void)
 313 {
 314         int i;
 315 #ifdef CONFIG_SMP
 316         struct irq_info *info;
 317
 318         /* By default all event channels notify CPU#0. */
 319         list_for_each_entry(info, &xen_irq_list_head, list) {
 320                 struct irq_desc *desc = irq_to_desc(info->irq);
 321                 cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 322         }
 323 #endif
 324
 325         for_each_possible_cpu(i)
 326                 memset(per_cpu(cpu_evtchn_mask, i),
 327                        (i == 0) ? ~0 : 0, sizeof(*per_cpu(cpu_evtchn_mask, i)));
 328 }
 329
 330 static inline void clear_evtchn(int port)
 331 {
 332         struct shared_info *s = HYPERVISOR_shared_info;
 333         sync_clear_bit(port, &s->evtchn_pending[0]);
 334 }
 335
 336 static inline void set_evtchn(int port)
 337 {
 338         struct shared_info *s = HYPERVISOR_shared_info;
 339         sync_set_bit(port, &s->evtchn_pending[0]);
 340 }
 341
 342 static inline int test_evtchn(int port)
 343 {
 344         struct shared_info *s = HYPERVISOR_shared_info;
 345         return sync_test_bit(port, &s->evtchn_pending[0]);
 346 }
 347
 348
 349 /**
 350  * notify_remote_via_irq - send event to remote end of event channel via irq
 351  * @irq: irq of event channel to send event to
 352  *
 353  * Unlike notify_remote_via_evtchn(), this is safe to use across
 354  * save/restore. Notifications on a broken connection are silently
 355  * dropped.
 356  */
 357 void notify_remote_via_irq(int irq)
 358 {
 359         int evtchn = evtchn_from_irq(irq);
 360
 361         if (VALID_EVTCHN(evtchn))
 362                 notify_remote_via_evtchn(evtchn);
 363 }
 364 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
 365
 366 static void mask_evtchn(int port)
 367 {
 368         struct shared_info *s = HYPERVISOR_shared_info;
 369         sync_set_bit(port, &s->evtchn_mask[0]);
 370 }
 371
 372 static void unmask_evtchn(int port)
 373 {
 374         struct shared_info *s = HYPERVISOR_shared_info;
 375         unsigned int cpu = get_cpu();
 376
 377         BUG_ON(!irqs_disabled());
 378
 379         /* Slow path (hypercall) if this is a non-local port. */
 380         if (unlikely(cpu != cpu_from_evtchn(port))) {
 381                 struct evtchn_unmask unmask = { .port = port };
 382                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 383         } else {
 384                 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
 385
 386                 sync_clear_bit(port, &s->evtchn_mask[0]);
 387
 388                 /*
 389                  * The following is basically the equivalent of
 390                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
 391                  * the interrupt edge' if the channel is masked.
 392                  */
 393                 if (sync_test_bit(port, &s->evtchn_pending[0]) &&
 394                     !sync_test_and_set_bit(port / BITS_PER_LONG,
 395                                            &vcpu_info->evtchn_pending_sel))
 396                         vcpu_info->evtchn_upcall_pending = 1;
 397         }
 398
 399         put_cpu();
 400 }
 401
 402 static void xen_irq_init(unsigned irq)
 403 {
 404         struct irq_info *info;
 405 #ifdef CONFIG_SMP
 406         struct irq_desc *desc = irq_to_desc(irq);
 407
 408         /* By default all event channels notify CPU#0. */
 409         cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 410 #endif
 411
 412         info = kzalloc(sizeof(*info), GFP_KERNEL);
 413         if (info == NULL)
 414                 panic("Unable to allocate metadata for IRQ%d\n", irq);
 415
 416         info->type = IRQT_UNBOUND;
 417         info->refcnt = -1;
 418
 419         irq_set_handler_data(irq, info);
 420
 421         list_add_tail(&info->list, &xen_irq_list_head);
 422 }
 423
 424 static int __must_check xen_allocate_irq_dynamic(void)
 425 {
 426         int first = 0;
 427         int irq;
 428
 429 #ifdef CONFIG_X86_IO_APIC
 430         /*
 431          * For an HVM guest or domain 0 which see "real" (emulated or
 432          * actual respectively) GSIs we allocate dynamic IRQs
 433          * e.g. those corresponding to event channels or MSIs
 434          * etc. from the range above those "real" GSIs to avoid
 435          * collisions.
 436          */
 437         if (xen_initial_domain() || xen_hvm_domain())
 438                 first = get_nr_irqs_gsi();
 439 #endif
 440
 441         irq = irq_alloc_desc_from(first, -1);
 442
 443         if (irq >= 0)
 444                 xen_irq_init(irq);
 445
 446         return irq;
 447 }
 448
 449 static int __must_check xen_allocate_irq_gsi(unsigned gsi)
 450 {
 451         int irq;
 452
 453         /*
 454          * A PV guest has no concept of a GSI (since it has no ACPI
 455          * nor access to/knowledge of the physical APICs). Therefore
 456          * all IRQs are dynamically allocated from the entire IRQ
 457          * space.
 458          */
 459         if (xen_pv_domain() && !xen_initial_domain())
 460                 return xen_allocate_irq_dynamic();
 461
 462         /* Legacy IRQ descriptors are already allocated by the arch. */
 463         if (gsi < NR_IRQS_LEGACY)
 464                 irq = gsi;
 465         else
 466                 irq = irq_alloc_desc_at(gsi, -1);
 467
 468         xen_irq_init(irq);
 469
 470         return irq;
 471 }
 472
 473 static void xen_free_irq(unsigned irq)
 474 {
 475         struct irq_info *info = irq_get_handler_data(irq);
 476
 477         list_del(&info->list);
 478
 479         irq_set_handler_data(irq, NULL);
 480
 481         WARN_ON(info->refcnt > 0);
 482
 483         kfree(info);
 484
 485         /* Legacy IRQ descriptors are managed by the arch. */
 486         if (irq < NR_IRQS_LEGACY)
 487                 return;
 488
 489         irq_free_desc(irq);
 490 }
 491
 492 static void pirq_query_unmask(int irq)
 493 {
 494         struct physdev_irq_status_query irq_status;
 495         struct irq_info *info = info_for_irq(irq);
 496
 497         BUG_ON(info->type != IRQT_PIRQ);
 498
 499         irq_status.irq = pirq_from_irq(irq);
 500         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
 501                 irq_status.flags = 0;
 502
 503         info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
 504         if (irq_status.flags & XENIRQSTAT_needs_eoi)
 505                 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
 506 }
 507
 508 static bool probing_irq(int irq)
 509 {
 510         struct irq_desc *desc = irq_to_desc(irq);
 511
 512         return desc && desc->action == NULL;
 513 }
 514
 515 static void eoi_pirq(struct irq_data *data)
 516 {
 517         int evtchn = evtchn_from_irq(data->irq);
 518         struct physdev_eoi eoi = { .irq = pirq_from_irq(data->irq) };
 519         int rc = 0;
 520
 521         irq_move_irq(data);
 522
 523         if (VALID_EVTCHN(evtchn))
 524                 clear_evtchn(evtchn);
 525
 526         if (pirq_needs_eoi(data->irq)) {
 527                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
 528                 WARN_ON(rc);
 529         }
 530 }
 531
 532 static void mask_ack_pirq(struct irq_data *data)
 533 {
 534         disable_dynirq(data);
 535         eoi_pirq(data);
 536 }
 537
 538 static unsigned int __startup_pirq(unsigned int irq)
 539 {
 540         struct evtchn_bind_pirq bind_pirq;
 541         struct irq_info *info = info_for_irq(irq);
 542         int evtchn = evtchn_from_irq(irq);
 543         int rc;
 544
 545         BUG_ON(info->type != IRQT_PIRQ);
 546
 547         if (VALID_EVTCHN(evtchn))
 548                 goto out;
 549
 550         bind_pirq.pirq = pirq_from_irq(irq);
 551         /* NB. We are happy to share unless we are probing. */
 552         bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
 553                                         BIND_PIRQ__WILL_SHARE : 0;
 554         rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
 555         if (rc != 0) {
 556                 if (!probing_irq(irq))
 557                         printk(KERN_INFO "Failed to obtain physical IRQ %d\n",
 558                                irq);
 559                 return 0;
 560         }
 561         evtchn = bind_pirq.port;
 562
 563         pirq_query_unmask(irq);
 564
 565         evtchn_to_irq[evtchn] = irq;
 566         bind_evtchn_to_cpu(evtchn, 0);
 567         info->evtchn = evtchn;
 568
 569 out:
 570         unmask_evtchn(evtchn);
 571         eoi_pirq(irq_get_irq_data(irq));
 572
 573         return 0;
 574 }
 575
 576 static unsigned int startup_pirq(struct irq_data *data)
 577 {
 578         return __startup_pirq(data->irq);
 579 }
 580
 581 static void shutdown_pirq(struct irq_data *data)
 582 {
 583         struct evtchn_close close;
 584         unsigned int irq = data->irq;
 585         struct irq_info *info = info_for_irq(irq);
 586         int evtchn = evtchn_from_irq(irq);
 587
 588         BUG_ON(info->type != IRQT_PIRQ);
 589
 590         if (!VALID_EVTCHN(evtchn))
 591                 return;
 592
 593         mask_evtchn(evtchn);
 594
 595         close.port = evtchn;
 596         if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 597                 BUG();
 598
 599         bind_evtchn_to_cpu(evtchn, 0);
 600         evtchn_to_irq[evtchn] = -1;
 601         info->evtchn = 0;
 602 }
 603
 604 static void enable_pirq(struct irq_data *data)
 605 {
 606         startup_pirq(data);
 607 }
 608
 609 static void disable_pirq(struct irq_data *data)
 610 {
 611         disable_dynirq(data);
 612 }
 613
 614 int xen_irq_from_gsi(unsigned gsi)
 615 {
 616         struct irq_info *info;
 617
 618         list_for_each_entry(info, &xen_irq_list_head, list) {
 619                 if (info->type != IRQT_PIRQ)
 620                         continue;
 621
 622                 if (info->u.pirq.gsi == gsi)
 623                         return info->irq;
 624         }
 625
 626         return -1;
 627 }
 628 EXPORT_SYMBOL_GPL(xen_irq_from_gsi);
 629
 630 /*
 631  * Do not make any assumptions regarding the relationship between the
 632  * IRQ number returned here and the Xen pirq argument.
 633  *
 634  * Note: We don't assign an event channel until the irq actually started
 635  * up.  Return an existing irq if we've already got one for the gsi.
 636  *
 637  * Shareable implies level triggered, not shareable implies edge
 638  * triggered here.
 639  */
 640 int xen_bind_pirq_gsi_to_irq(unsigned gsi,
 641                              unsigned pirq, int shareable, char *name)
 642 {
 643         int irq = -1;
 644         struct physdev_irq irq_op;
 645
 646         mutex_lock(&irq_mapping_update_lock);
 647
 648         irq = xen_irq_from_gsi(gsi);
 649         if (irq != -1) {
 650                 printk(KERN_INFO "xen_map_pirq_gsi: returning irq %d for gsi %u\n",
 651                        irq, gsi);
 652                 goto out;
 653         }
 654
 655         irq = xen_allocate_irq_gsi(gsi);
 656         if (irq < 0)
 657                 goto out;
 658
 659         irq_op.irq = irq;
 660         irq_op.vector = 0;
 661
 662         /* Only the privileged domain can do this. For non-priv, the pcifront
 663          * driver provides a PCI bus that does the call to do exactly
 664          * this in the priv domain. */
 665         if (xen_initial_domain() &&
 666             HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
 667                 xen_free_irq(irq);
 668                 irq = -ENOSPC;
 669                 goto out;
 670         }
 671
 672         xen_irq_info_pirq_init(irq, 0, pirq, gsi, irq_op.vector, DOMID_SELF,
 673                                shareable ? PIRQ_SHAREABLE : 0);
 674
 675         pirq_query_unmask(irq);
 676         /* We try to use the handler with the appropriate semantic for the
 677          * type of interrupt: if the interrupt is an edge triggered
 678          * interrupt we use handle_edge_irq.
 679          *
 680          * On the other hand if the interrupt is level triggered we use
 681          * handle_fasteoi_irq like the native code does for this kind of
 682          * interrupts.
 683          *
 684          * Depending on the Xen version, pirq_needs_eoi might return true
 685          * not only for level triggered interrupts but for edge triggered
 686          * interrupts too. In any case Xen always honors the eoi mechanism,
 687          * not injecting any more pirqs of the same kind if the first one
 688          * hasn't received an eoi yet. Therefore using the fasteoi handler
 689          * is the right choice either way.
 690          */
 691         if (shareable)
 692                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 693                                 handle_fasteoi_irq, name);
 694         else
 695                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 696                                 handle_edge_irq, name);
 697
 698 out:
 699         mutex_unlock(&irq_mapping_update_lock);
 700
 701         return irq;
 702 }
 703
 704 #ifdef CONFIG_PCI_MSI
 705 int xen_allocate_pirq_msi(struct pci_dev *dev, struct msi_desc *msidesc)
 706 {
 707         int rc;
 708         struct physdev_get_free_pirq op_get_free_pirq;
 709
 710         op_get_free_pirq.type = MAP_PIRQ_TYPE_MSI;
 711         rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
 712
 713         WARN_ONCE(rc == -ENOSYS,
 714                   "hypervisor does not support the PHYSDEVOP_get_free_pirq interface\n");
 715
 716         return rc ? -1 : op_get_free_pirq.pirq;
 717 }
 718
 719 int xen_bind_pirq_msi_to_irq(struct pci_dev *dev, struct msi_desc *msidesc,
 720                              int pirq, int vector, const char *name,
 721                              domid_t domid)
 722 {
 723         int irq, ret;
 724
 725         mutex_lock(&irq_mapping_update_lock);
 726
 727         irq = xen_allocate_irq_dynamic();
 728         if (irq < 0)
 729                 goto out;
 730
 731         irq_set_chip_and_handler_name(irq, &xen_pirq_chip, handle_edge_irq,
 732                         name);
 733
 734         xen_irq_info_pirq_init(irq, 0, pirq, 0, vector, domid, 0);
 735         ret = irq_set_msi_desc(irq, msidesc);
 736         if (ret < 0)
 737                 goto error_irq;
 738 out:
 739         mutex_unlock(&irq_mapping_update_lock);
 740         return irq;
 741 error_irq:
 742         mutex_unlock(&irq_mapping_update_lock);
 743         xen_free_irq(irq);
 744         return ret;
 745 }
 746 #endif
 747
 748 int xen_destroy_irq(int irq)
 749 {
 750         struct irq_desc *desc;
 751         struct physdev_unmap_pirq unmap_irq;
 752         struct irq_info *info = info_for_irq(irq);
 753         int rc = -ENOENT;
 754
 755         mutex_lock(&irq_mapping_update_lock);
 756
 757         desc = irq_to_desc(irq);
 758         if (!desc)
 759                 goto out;
 760
 761         if (xen_initial_domain()) {
 762                 unmap_irq.pirq = info->u.pirq.pirq;
 763                 unmap_irq.domid = info->u.pirq.domid;
 764                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
 765                 /* If another domain quits without making the pci_disable_msix
 766                  * call, the Xen hypervisor takes care of freeing the PIRQs
 767                  * (free_domain_pirqs).
 768                  */
 769                 if ((rc == -ESRCH && info->u.pirq.domid != DOMID_SELF))
 770                         printk(KERN_INFO "domain %d does not have %d anymore\n",
 771                                 info->u.pirq.domid, info->u.pirq.pirq);
 772                 else if (rc) {
 773                         printk(KERN_WARNING "unmap irq failed %d\n", rc);
 774                         goto out;
 775                 }
 776         }
 777
 778         xen_free_irq(irq);
 779
 780 out:
 781         mutex_unlock(&irq_mapping_update_lock);
 782         return rc;
 783 }
 784
 785 int xen_irq_from_pirq(unsigned pirq)
 786 {
 787         int irq;
 788
 789         struct irq_info *info;
 790
 791         mutex_lock(&irq_mapping_update_lock);
 792
 793         list_for_each_entry(info, &xen_irq_list_head, list) {
 794                 if (info->type != IRQT_PIRQ)
 795                         continue;
 796                 irq = info->irq;
 797                 if (info->u.pirq.pirq == pirq)
 798                         goto out;
 799         }
 800         irq = -1;
 801 out:
 802         mutex_unlock(&irq_mapping_update_lock);
 803
 804         return irq;
 805 }
 806
 807
 808 int xen_pirq_from_irq(unsigned irq)
 809 {
 810         return pirq_from_irq(irq);
 811 }
 812 EXPORT_SYMBOL_GPL(xen_pirq_from_irq);
 813 int bind_evtchn_to_irq(unsigned int evtchn)
 814 {
 815         int irq;
 816
 817         mutex_lock(&irq_mapping_update_lock);
 818
 819         irq = evtchn_to_irq[evtchn];
 820
 821         if (irq == -1) {
 822                 irq = xen_allocate_irq_dynamic();
 823                 if (irq == -1)
 824                         goto out;
 825
 826                 irq_set_chip_and_handler_name(irq, &xen_dynamic_chip,
 827                                               handle_edge_irq, "event");
 828
 829                 xen_irq_info_evtchn_init(irq, evtchn);
 830         }
 831
 832 out:
 833         mutex_unlock(&irq_mapping_update_lock);
 834
 835         return irq;
 836 }
 837 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
 838
 839 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
 840 {
 841         struct evtchn_bind_ipi bind_ipi;
 842         int evtchn, irq;
 843
 844         mutex_lock(&irq_mapping_update_lock);
 845
 846         irq = per_cpu(ipi_to_irq, cpu)[ipi];
 847
 848         if (irq == -1) {
 849                 irq = xen_allocate_irq_dynamic();
 850                 if (irq < 0)
 851                         goto out;
 852
 853                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 854                                               handle_percpu_irq, "ipi");
 855
 856                 bind_ipi.vcpu = cpu;
 857                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
 858                                                 &bind_ipi) != 0)
 859                         BUG();
 860                 evtchn = bind_ipi.port;
 861
 862                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
 863
 864                 bind_evtchn_to_cpu(evtchn, cpu);
 865         }
 866
 867  out:
 868         mutex_unlock(&irq_mapping_update_lock);
 869         return irq;
 870 }
 871
 872 static int bind_interdomain_evtchn_to_irq(unsigned int remote_domain,
 873                                           unsigned int remote_port)
 874 {
 875         struct evtchn_bind_interdomain bind_interdomain;
 876         int err;
 877
 878         bind_interdomain.remote_dom  = remote_domain;
 879         bind_interdomain.remote_port = remote_port;
 880
 881         err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
 882                                           &bind_interdomain);
 883
 884         return err ? : bind_evtchn_to_irq(bind_interdomain.local_port);
 885 }
 886
 887 static int find_virq(unsigned int virq, unsigned int cpu)
 888 {
 889         struct evtchn_status status;
 890         int port, rc = -ENOENT;
 891
 892         memset(&status, 0, sizeof(status));
 893         for (port = 0; port <= NR_EVENT_CHANNELS; port++) {
 894                 status.dom = DOMID_SELF;
 895                 status.port = port;
 896                 rc = HYPERVISOR_event_channel_op(EVTCHNOP_status, &status);
 897                 if (rc < 0)
 898                         continue;
 899                 if (status.status != EVTCHNSTAT_virq)
 900                         continue;
 901                 if (status.u.virq == virq && status.vcpu == cpu) {
 902                         rc = port;
 903                         break;
 904                 }
 905         }
 906         return rc;
 907 }
 908
 909 int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
 910 {
 911         struct evtchn_bind_virq bind_virq;
 912         int evtchn, irq, ret;
 913
 914         mutex_lock(&irq_mapping_update_lock);
 915
 916         irq = per_cpu(virq_to_irq, cpu)[virq];
 917
 918         if (irq == -1) {
 919                 irq = xen_allocate_irq_dynamic();
 920                 if (irq == -1)
 921                         goto out;
 922
 923                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 924                                               handle_percpu_irq, "virq");
 925
 926                 bind_virq.virq = virq;
 927                 bind_virq.vcpu = cpu;
 928                 ret = HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 929                                                 &bind_virq);
 930                 if (ret == 0)
 931                         evtchn = bind_virq.port;
 932                 else {
 933                         if (ret == -EEXIST)
 934                                 ret = find_virq(virq, cpu);
 935                         BUG_ON(ret < 0);
 936                         evtchn = ret;
 937                 }
 938
 939                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
 940
 941                 bind_evtchn_to_cpu(evtchn, cpu);
 942         }
 943
 944 out:
 945         mutex_unlock(&irq_mapping_update_lock);
 946
 947         return irq;
 948 }
 949
 950 static void unbind_from_irq(unsigned int irq)
 951 {
 952         struct evtchn_close close;
 953         int evtchn = evtchn_from_irq(irq);
 954         struct irq_info *info = irq_get_handler_data(irq);
 955
 956         mutex_lock(&irq_mapping_update_lock);
 957
 958         if (info->refcnt > 0) {
 959                 info->refcnt--;
 960                 if (info->refcnt != 0)
 961                         goto done;
 962         }
 963
 964         if (VALID_EVTCHN(evtchn)) {
 965                 close.port = evtchn;
 966                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 967                         BUG();
 968
 969                 switch (type_from_irq(irq)) {
 970                 case IRQT_VIRQ:
 971                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
 972                                 [virq_from_irq(irq)] = -1;
 973                         break;
 974                 case IRQT_IPI:
 975                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
 976                                 [ipi_from_irq(irq)] = -1;
 977                         break;
 978                 default:
 979                         break;
 980                 }
 981
 982                 /* Closed ports are implicitly re-bound to VCPU0. */
 983                 bind_evtchn_to_cpu(evtchn, 0);
 984
 985                 evtchn_to_irq[evtchn] = -1;
 986         }
 987
 988         BUG_ON(info_for_irq(irq)->type == IRQT_UNBOUND);
 989
 990         xen_free_irq(irq);
 991
 992  done:
 993         mutex_unlock(&irq_mapping_update_lock);
 994 }
 995
 996 int bind_evtchn_to_irqhandler(unsigned int evtchn,
 997                               irq_handler_t handler,
 998                               unsigned long irqflags,
 999                               const char *devname, void *dev_id)
1000 {
1001         int irq, retval;
1002
1003         irq = bind_evtchn_to_irq(evtchn);
1004         if (irq < 0)
1005                 return irq;
1006         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1007         if (retval != 0) {
1008                 unbind_from_irq(irq);
1009                 return retval;
1010         }
1011
1012         return irq;
1013 }
1014 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
1015
1016 int bind_interdomain_evtchn_to_irqhandler(unsigned int remote_domain,
1017                                           unsigned int remote_port,
1018                                           irq_handler_t handler,
1019                                           unsigned long irqflags,
1020                                           const char *devname,
1021                                           void *dev_id)
1022 {
1023         int irq, retval;
1024
1025         irq = bind_interdomain_evtchn_to_irq(remote_domain, remote_port);
1026         if (irq < 0)
1027                 return irq;
1028
1029         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1030         if (retval != 0) {
1031                 unbind_from_irq(irq);
1032                 return retval;
1033         }
1034
1035         return irq;
1036 }
1037 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler);
1038
1039 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
1040                             irq_handler_t handler,
1041                             unsigned long irqflags, const char *devname, void *dev_id)
1042 {
1043         int irq, retval;
1044
1045         irq = bind_virq_to_irq(virq, cpu);
1046         if (irq < 0)
1047                 return irq;
1048         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1049         if (retval != 0) {
1050                 unbind_from_irq(irq);
1051                 return retval;
1052         }
1053
1054         return irq;
1055 }
1056 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1057
1058 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1059                            unsigned int cpu,
1060                            irq_handler_t handler,
1061                            unsigned long irqflags,
1062                            const char *devname,
1063                            void *dev_id)
1064 {
1065         int irq, retval;
1066
1067         irq = bind_ipi_to_irq(ipi, cpu);
1068         if (irq < 0)
1069                 return irq;
1070
1071         irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME | IRQF_EARLY_RESUME;
1072         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1073         if (retval != 0) {
1074                 unbind_from_irq(irq);
1075                 return retval;
1076         }
1077
1078         return irq;
1079 }
1080
1081 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1082 {
1083         free_irq(irq, dev_id);
1084         unbind_from_irq(irq);
1085 }
1086 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1087
1088 int evtchn_make_refcounted(unsigned int evtchn)
1089 {
1090         int irq = evtchn_to_irq[evtchn];
1091         struct irq_info *info;
1092
1093         if (irq == -1)
1094                 return -ENOENT;
1095
1096         info = irq_get_handler_data(irq);
1097
1098         if (!info)
1099                 return -ENOENT;
1100
1101         WARN_ON(info->refcnt != -1);
1102
1103         info->refcnt = 1;
1104
1105         return 0;
1106 }
1107 EXPORT_SYMBOL_GPL(evtchn_make_refcounted);
1108
1109 int evtchn_get(unsigned int evtchn)
1110 {
1111         int irq;
1112         struct irq_info *info;
1113         int err = -ENOENT;
1114
1115         if (evtchn >= NR_EVENT_CHANNELS)
1116                 return -EINVAL;
1117
1118         mutex_lock(&irq_mapping_update_lock);
1119
1120         irq = evtchn_to_irq[evtchn];
1121         if (irq == -1)
1122                 goto done;
1123
1124         info = irq_get_handler_data(irq);
1125
1126         if (!info)
1127                 goto done;
1128
1129         err = -EINVAL;
1130         if (info->refcnt <= 0)
1131                 goto done;
1132
1133         info->refcnt++;
1134         err = 0;
1135  done:
1136         mutex_unlock(&irq_mapping_update_lock);
1137
1138         return err;
1139 }
1140 EXPORT_SYMBOL_GPL(evtchn_get);
1141
1142 void evtchn_put(unsigned int evtchn)
1143 {
1144         int irq = evtchn_to_irq[evtchn];
1145         if (WARN_ON(irq == -1))
1146                 return;
1147         unbind_from_irq(irq);
1148 }
1149 EXPORT_SYMBOL_GPL(evtchn_put);
1150
1151 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1152 {
1153         int irq = per_cpu(ipi_to_irq, cpu)[vector];
1154         BUG_ON(irq < 0);
1155         notify_remote_via_irq(irq);
1156 }
1157
1158 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
1159 {
1160         struct shared_info *sh = HYPERVISOR_shared_info;
1161         int cpu = smp_processor_id();
1162         unsigned long *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu);
1163         int i;
1164         unsigned long flags;
1165         static DEFINE_SPINLOCK(debug_lock);
1166         struct vcpu_info *v;
1167
1168         spin_lock_irqsave(&debug_lock, flags);
1169
1170         printk("\nvcpu %d\n  ", cpu);
1171
1172         for_each_online_cpu(i) {
1173                 int pending;
1174                 v = per_cpu(xen_vcpu, i);
1175                 pending = (get_irq_regs() && i == cpu)
1176                         ? xen_irqs_disabled(get_irq_regs())
1177                         : v->evtchn_upcall_mask;
1178                 printk("%d: masked=%d pending=%d event_sel %0*lx\n  ", i,
1179                        pending, v->evtchn_upcall_pending,
1180                        (int)(sizeof(v->evtchn_pending_sel)*2),
1181                        v->evtchn_pending_sel);
1182         }
1183         v = per_cpu(xen_vcpu, cpu);
1184
1185         printk("\npending:\n   ");
1186         for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
1187                 printk("%0*lx%s", (int)sizeof(sh->evtchn_pending[0])*2,
1188                        sh->evtchn_pending[i],
1189                        i % 8 == 0 ? "\n   " : " ");
1190         printk("\nglobal mask:\n   ");
1191         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1192                 printk("%0*lx%s",
1193                        (int)(sizeof(sh->evtchn_mask[0])*2),
1194                        sh->evtchn_mask[i],
1195                        i % 8 == 0 ? "\n   " : " ");
1196
1197         printk("\nglobally unmasked:\n   ");
1198         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1199                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1200                        sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
1201                        i % 8 == 0 ? "\n   " : " ");
1202
1203         printk("\nlocal cpu%d mask:\n   ", cpu);
1204         for (i = (NR_EVENT_CHANNELS/BITS_PER_LONG)-1; i >= 0; i--)
1205                 printk("%0*lx%s", (int)(sizeof(cpu_evtchn[0])*2),
1206                        cpu_evtchn[i],
1207                        i % 8 == 0 ? "\n   " : " ");
1208
1209         printk("\nlocally unmasked:\n   ");
1210         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
1211                 unsigned long pending = sh->evtchn_pending[i]
1212                         & ~sh->evtchn_mask[i]
1213                         & cpu_evtchn[i];
1214                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1215                        pending, i % 8 == 0 ? "\n   " : " ");
1216         }
1217
1218         printk("\npending list:\n");
1219         for (i = 0; i < NR_EVENT_CHANNELS; i++) {
1220                 if (sync_test_bit(i, sh->evtchn_pending)) {
1221                         int word_idx = i / BITS_PER_LONG;
1222                         printk("  %d: event %d -> irq %d%s%s%s\n",
1223                                cpu_from_evtchn(i), i,
1224                                evtchn_to_irq[i],
1225                                sync_test_bit(word_idx, &v->evtchn_pending_sel)
1226                                              ? "" : " l2-clear",
1227                                !sync_test_bit(i, sh->evtchn_mask)
1228                                              ? "" : " globally-masked",
1229                                sync_test_bit(i, cpu_evtchn)
1230                                              ? "" : " locally-masked");
1231                 }
1232         }
1233
1234         spin_unlock_irqrestore(&debug_lock, flags);
1235
1236         return IRQ_HANDLED;
1237 }
1238
1239 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
1240 static DEFINE_PER_CPU(unsigned int, current_word_idx);
1241 static DEFINE_PER_CPU(unsigned int, current_bit_idx);
1242
1243 /*
1244  * Mask out the i least significant bits of w
1245  */
1246 #define MASK_LSBS(w, i) (w & ((~0UL) << i))
1247
1248 /*
1249  * Search the CPUs pending events bitmasks.  For each one found, map
1250  * the event number to an irq, and feed it into do_IRQ() for
1251  * handling.
1252  *
1253  * Xen uses a two-level bitmap to speed searching.  The first level is
1254  * a bitset of words which contain pending event bits.  The second
1255  * level is a bitset of pending events themselves.
1256  */
1257 static void __xen_evtchn_do_upcall(void)
1258 {
1259         int start_word_idx, start_bit_idx;
1260         int word_idx, bit_idx;
1261         int i;
1262         int cpu = get_cpu();
1263         struct shared_info *s = HYPERVISOR_shared_info;
1264         struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1265         unsigned count;
1266
1267         do {
1268                 unsigned long pending_words;
1269
1270                 vcpu_info->evtchn_upcall_pending = 0;
1271
1272                 if (__this_cpu_inc_return(xed_nesting_count) - 1)
1273                         goto out;
1274
1275 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
1276                 /* Clear master flag /before/ clearing selector flag. */
1277                 wmb();
1278 #endif
1279                 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
1280
1281                 start_word_idx = __this_cpu_read(current_word_idx);
1282                 start_bit_idx = __this_cpu_read(current_bit_idx);
1283
1284                 word_idx = start_word_idx;
1285
1286                 for (i = 0; pending_words != 0; i++) {
1287                         unsigned long pending_bits;
1288                         unsigned long words;
1289
1290                         words = MASK_LSBS(pending_words, word_idx);
1291
1292                         /*
1293                          * If we masked out all events, wrap to beginning.
1294                          */
1295                         if (words == 0) {
1296                                 word_idx = 0;
1297                                 bit_idx = 0;
1298                                 continue;
1299                         }
1300                         word_idx = __ffs(words);
1301
1302                         pending_bits = active_evtchns(cpu, s, word_idx);
1303                         bit_idx = 0; /* usually scan entire word from start */
1304                         if (word_idx == start_word_idx) {
1305                                 /* We scan the starting word in two parts */
1306                                 if (i == 0)
1307                                         /* 1st time: start in the middle */
1308                                         bit_idx = start_bit_idx;
1309                                 else
1310                                         /* 2nd time: mask bits done already */
1311                                         bit_idx &= (1UL << start_bit_idx) - 1;
1312                         }
1313
1314                         do {
1315                                 unsigned long bits;
1316                                 int port, irq;
1317                                 struct irq_desc *desc;
1318
1319                                 bits = MASK_LSBS(pending_bits, bit_idx);
1320
1321                                 /* If we masked out all events, move on. */
1322                                 if (bits == 0)
1323                                         break;
1324
1325                                 bit_idx = __ffs(bits);
1326
1327                                 /* Process port. */
1328                                 port = (word_idx * BITS_PER_LONG) + bit_idx;
1329                                 irq = evtchn_to_irq[port];
1330
1331                                 if (irq != -1) {
1332                                         desc = irq_to_desc(irq);
1333                                         if (desc)
1334                                                 generic_handle_irq_desc(irq, desc);
1335                                 }
1336
1337                                 bit_idx = (bit_idx + 1) % BITS_PER_LONG;
1338
1339                                 /* Next caller starts at last processed + 1 */
1340                                 __this_cpu_write(current_word_idx,
1341                                                  bit_idx ? word_idx :
1342                                                  (word_idx+1) % BITS_PER_LONG);
1343                                 __this_cpu_write(current_bit_idx, bit_idx);
1344                         } while (bit_idx != 0);
1345
1346                         /* Scan start_l1i twice; all others once. */
1347                         if ((word_idx != start_word_idx) || (i != 0))
1348                                 pending_words &= ~(1UL << word_idx);
1349
1350                         word_idx = (word_idx + 1) % BITS_PER_LONG;
1351                 }
1352
1353                 BUG_ON(!irqs_disabled());
1354
1355                 count = __this_cpu_read(xed_nesting_count);
1356                 __this_cpu_write(xed_nesting_count, 0);
1357         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
1358
1359 out:
1360
1361         put_cpu();
1362 }
1363
1364 void xen_evtchn_do_upcall(struct pt_regs *regs)
1365 {
1366         struct pt_regs *old_regs = set_irq_regs(regs);
1367
1368         exit_idle();
1369         irq_enter();
1370
1371         __xen_evtchn_do_upcall();
1372
1373         irq_exit();
1374         set_irq_regs(old_regs);
1375 }
1376
1377 void xen_hvm_evtchn_do_upcall(void)
1378 {
1379         __xen_evtchn_do_upcall();
1380 }
1381 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1382
1383 /* Rebind a new event channel to an existing irq. */
1384 void rebind_evtchn_irq(int evtchn, int irq)
1385 {
1386         struct irq_info *info = info_for_irq(irq);
1387
1388         /* Make sure the irq is masked, since the new event channel
1389            will also be masked. */
1390         disable_irq(irq);
1391
1392         mutex_lock(&irq_mapping_update_lock);
1393
1394         /* After resume the irq<->evtchn mappings are all cleared out */
1395         BUG_ON(evtchn_to_irq[evtchn] != -1);
1396         /* Expect irq to have been bound before,
1397            so there should be a proper type */
1398         BUG_ON(info->type == IRQT_UNBOUND);
1399
1400         xen_irq_info_evtchn_init(irq, evtchn);
1401
1402         mutex_unlock(&irq_mapping_update_lock);
1403
1404         /* new event channels are always bound to cpu 0 */
1405         irq_set_affinity(irq, cpumask_of(0));
1406
1407         /* Unmask the event channel. */
1408         enable_irq(irq);
1409 }
1410
1411 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1412 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1413 {
1414         struct evtchn_bind_vcpu bind_vcpu;
1415         int evtchn = evtchn_from_irq(irq);
1416
1417         if (!VALID_EVTCHN(evtchn))
1418                 return -1;
1419
1420         /*
1421          * Events delivered via platform PCI interrupts are always
1422          * routed to vcpu 0 and hence cannot be rebound.
1423          */
1424         if (xen_hvm_domain() && !xen_have_vector_callback)
1425                 return -1;
1426
1427         /* Send future instances of this interrupt to other vcpu. */
1428         bind_vcpu.port = evtchn;
1429         bind_vcpu.vcpu = tcpu;
1430
1431         /*
1432          * If this fails, it usually just indicates that we're dealing with a
1433          * virq or IPI channel, which don't actually need to be rebound. Ignore
1434          * it, but don't do the xenlinux-level rebind in that case.
1435          */
1436         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1437                 bind_evtchn_to_cpu(evtchn, tcpu);
1438
1439         return 0;
1440 }
1441
1442 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1443                             bool force)
1444 {
1445         unsigned tcpu = cpumask_first(dest);
1446
1447         return rebind_irq_to_cpu(data->irq, tcpu);
1448 }
1449
1450 int resend_irq_on_evtchn(unsigned int irq)
1451 {
1452         int masked, evtchn = evtchn_from_irq(irq);
1453         struct shared_info *s = HYPERVISOR_shared_info;
1454
1455         if (!VALID_EVTCHN(evtchn))
1456                 return 1;
1457
1458         masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
1459         sync_set_bit(evtchn, s->evtchn_pending);
1460         if (!masked)
1461                 unmask_evtchn(evtchn);
1462
1463         return 1;
1464 }
1465
1466 static void enable_dynirq(struct irq_data *data)
1467 {
1468         int evtchn = evtchn_from_irq(data->irq);
1469
1470         if (VALID_EVTCHN(evtchn))
1471                 unmask_evtchn(evtchn);
1472 }
1473
1474 static void disable_dynirq(struct irq_data *data)
1475 {
1476         int evtchn = evtchn_from_irq(data->irq);
1477
1478         if (VALID_EVTCHN(evtchn))
1479                 mask_evtchn(evtchn);
1480 }
1481
1482 static void ack_dynirq(struct irq_data *data)
1483 {
1484         int evtchn = evtchn_from_irq(data->irq);
1485
1486         irq_move_irq(data);
1487
1488         if (VALID_EVTCHN(evtchn))
1489                 clear_evtchn(evtchn);
1490 }
1491
1492 static void mask_ack_dynirq(struct irq_data *data)
1493 {
1494         disable_dynirq(data);
1495         ack_dynirq(data);
1496 }
1497
1498 static int retrigger_dynirq(struct irq_data *data)
1499 {
1500         int evtchn = evtchn_from_irq(data->irq);
1501         struct shared_info *sh = HYPERVISOR_shared_info;
1502         int ret = 0;
1503
1504         if (VALID_EVTCHN(evtchn)) {
1505                 int masked;
1506
1507                 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
1508                 sync_set_bit(evtchn, sh->evtchn_pending);
1509                 if (!masked)
1510                         unmask_evtchn(evtchn);
1511                 ret = 1;
1512         }
1513
1514         return ret;
1515 }
1516
1517 static void restore_pirqs(void)
1518 {
1519         int pirq, rc, irq, gsi;
1520         struct physdev_map_pirq map_irq;
1521         struct irq_info *info;
1522
1523         list_for_each_entry(info, &xen_irq_list_head, list) {
1524                 if (info->type != IRQT_PIRQ)
1525                         continue;
1526
1527                 pirq = info->u.pirq.pirq;
1528                 gsi = info->u.pirq.gsi;
1529                 irq = info->irq;
1530
1531                 /* save/restore of PT devices doesn't work, so at this point the
1532                  * only devices present are GSI based emulated devices */
1533                 if (!gsi)
1534                         continue;
1535
1536                 map_irq.domid = DOMID_SELF;
1537                 map_irq.type = MAP_PIRQ_TYPE_GSI;
1538                 map_irq.index = gsi;
1539                 map_irq.pirq = pirq;
1540
1541                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1542                 if (rc) {
1543                         printk(KERN_WARNING "xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1544                                         gsi, irq, pirq, rc);
1545                         xen_free_irq(irq);
1546                         continue;
1547                 }
1548
1549                 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1550
1551                 __startup_pirq(irq);
1552         }
1553 }
1554
1555 static void restore_cpu_virqs(unsigned int cpu)
1556 {
1557         struct evtchn_bind_virq bind_virq;
1558         int virq, irq, evtchn;
1559
1560         for (virq = 0; virq < NR_VIRQS; virq++) {
1561                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1562                         continue;
1563
1564                 BUG_ON(virq_from_irq(irq) != virq);
1565
1566                 /* Get a new binding from Xen. */
1567                 bind_virq.virq = virq;
1568                 bind_virq.vcpu = cpu;
1569                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1570                                                 &bind_virq) != 0)
1571                         BUG();
1572                 evtchn = bind_virq.port;
1573
1574                 /* Record the new mapping. */
1575                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
1576                 bind_evtchn_to_cpu(evtchn, cpu);
1577         }
1578 }
1579
1580 static void restore_cpu_ipis(unsigned int cpu)
1581 {
1582         struct evtchn_bind_ipi bind_ipi;
1583         int ipi, irq, evtchn;
1584
1585         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1586                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1587                         continue;
1588
1589                 BUG_ON(ipi_from_irq(irq) != ipi);
1590
1591                 /* Get a new binding from Xen. */
1592                 bind_ipi.vcpu = cpu;
1593                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1594                                                 &bind_ipi) != 0)
1595                         BUG();
1596                 evtchn = bind_ipi.port;
1597
1598                 /* Record the new mapping. */
1599                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
1600                 bind_evtchn_to_cpu(evtchn, cpu);
1601         }
1602 }
1603
1604 /* Clear an irq's pending state, in preparation for polling on it */
1605 void xen_clear_irq_pending(int irq)
1606 {
1607         int evtchn = evtchn_from_irq(irq);
1608
1609         if (VALID_EVTCHN(evtchn))
1610                 clear_evtchn(evtchn);
1611 }
1612 EXPORT_SYMBOL(xen_clear_irq_pending);
1613 void xen_set_irq_pending(int irq)
1614 {
1615         int evtchn = evtchn_from_irq(irq);
1616
1617         if (VALID_EVTCHN(evtchn))
1618                 set_evtchn(evtchn);
1619 }
1620
1621 bool xen_test_irq_pending(int irq)
1622 {
1623         int evtchn = evtchn_from_irq(irq);
1624         bool ret = false;
1625
1626         if (VALID_EVTCHN(evtchn))
1627                 ret = test_evtchn(evtchn);
1628
1629         return ret;
1630 }
1631
1632 /* Poll waiting for an irq to become pending with timeout.  In the usual case,
1633  * the irq will be disabled so it won't deliver an interrupt. */
1634 void xen_poll_irq_timeout(int irq, u64 timeout)
1635 {
1636         evtchn_port_t evtchn = evtchn_from_irq(irq);
1637
1638         if (VALID_EVTCHN(evtchn)) {
1639                 struct sched_poll poll;
1640
1641                 poll.nr_ports = 1;
1642                 poll.timeout = timeout;
1643                 set_xen_guest_handle(poll.ports, &evtchn);
1644
1645                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1646                         BUG();
1647         }
1648 }
1649 EXPORT_SYMBOL(xen_poll_irq_timeout);
1650 /* Poll waiting for an irq to become pending.  In the usual case, the
1651  * irq will be disabled so it won't deliver an interrupt. */
1652 void xen_poll_irq(int irq)
1653 {
1654         xen_poll_irq_timeout(irq, 0 /* no timeout */);
1655 }
1656
1657 /* Check whether the IRQ line is shared with other guests. */
1658 int xen_test_irq_shared(int irq)
1659 {
1660         struct irq_info *info = info_for_irq(irq);
1661         struct physdev_irq_status_query irq_status = { .irq = info->u.pirq.pirq };
1662
1663         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
1664                 return 0;
1665         return !(irq_status.flags & XENIRQSTAT_shared);
1666 }
1667 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
1668
1669 void xen_irq_resume(void)
1670 {
1671         unsigned int cpu, evtchn;
1672         struct irq_info *info;
1673
1674         init_evtchn_cpu_bindings();
1675
1676         /* New event-channel space is not 'live' yet. */
1677         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1678                 mask_evtchn(evtchn);
1679
1680         /* No IRQ <-> event-channel mappings. */
1681         list_for_each_entry(info, &xen_irq_list_head, list)
1682                 info->evtchn = 0; /* zap event-channel binding */
1683
1684         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1685                 evtchn_to_irq[evtchn] = -1;
1686
1687         for_each_possible_cpu(cpu) {
1688                 restore_cpu_virqs(cpu);
1689                 restore_cpu_ipis(cpu);
1690         }
1691
1692         restore_pirqs();
1693 }
1694
1695 static struct irq_chip xen_dynamic_chip __read_mostly = {
1696         .name                   = "xen-dyn",
1697
1698         .irq_disable            = disable_dynirq,
1699         .irq_mask               = disable_dynirq,
1700         .irq_unmask             = enable_dynirq,
1701
1702         .irq_ack                = ack_dynirq,
1703         .irq_mask_ack           = mask_ack_dynirq,
1704
1705         .irq_set_affinity       = set_affinity_irq,
1706         .irq_retrigger          = retrigger_dynirq,
1707 };
1708
1709 static struct irq_chip xen_pirq_chip __read_mostly = {
1710         .name                   = "xen-pirq",
1711
1712         .irq_startup            = startup_pirq,
1713         .irq_shutdown           = shutdown_pirq,
1714         .irq_enable             = enable_pirq,
1715         .irq_disable            = disable_pirq,
1716
1717         .irq_mask               = disable_dynirq,
1718         .irq_unmask             = enable_dynirq,
1719
1720         .irq_ack                = eoi_pirq,
1721         .irq_eoi                = eoi_pirq,
1722         .irq_mask_ack           = mask_ack_pirq,
1723
1724         .irq_set_affinity       = set_affinity_irq,
1725
1726         .irq_retrigger          = retrigger_dynirq,
1727 };
1728
1729 static struct irq_chip xen_percpu_chip __read_mostly = {
1730         .name                   = "xen-percpu",
1731
1732         .irq_disable            = disable_dynirq,
1733         .irq_mask               = disable_dynirq,
1734         .irq_unmask             = enable_dynirq,
1735
1736         .irq_ack                = ack_dynirq,
1737 };
1738
1739 int xen_set_callback_via(uint64_t via)
1740 {
1741         struct xen_hvm_param a;
1742         a.domid = DOMID_SELF;
1743         a.index = HVM_PARAM_CALLBACK_IRQ;
1744         a.value = via;
1745         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1746 }
1747 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1748
1749 #ifdef CONFIG_XEN_PVHVM
1750 /* Vector callbacks are better than PCI interrupts to receive event
1751  * channel notifications because we can receive vector callbacks on any
1752  * vcpu and we don't need PCI support or APIC interactions. */
1753 void xen_callback_vector(void)
1754 {
1755         int rc;
1756         uint64_t callback_via;
1757         if (xen_have_vector_callback) {
1758                 callback_via = HVM_CALLBACK_VECTOR(XEN_HVM_EVTCHN_CALLBACK);
1759                 rc = xen_set_callback_via(callback_via);
1760                 if (rc) {
1761                         printk(KERN_ERR "Request for Xen HVM callback vector"
1762                                         " failed.\n");
1763                         xen_have_vector_callback = 0;
1764                         return;
1765                 }
1766                 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1767                                 "enabled\n");
1768                 /* in the restore case the vector has already been allocated */
1769                 if (!test_bit(XEN_HVM_EVTCHN_CALLBACK, used_vectors))
1770                         alloc_intr_gate(XEN_HVM_EVTCHN_CALLBACK, xen_hvm_callback_vector);
1771         }
1772 }
1773 #else
1774 void xen_callback_vector(void) {}
1775 #endif
1776
1777 void __init xen_init_IRQ(void)
1778 {
1779         int i, rc;
1780
1781         evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1782                                     GFP_KERNEL);
1783         BUG_ON(!evtchn_to_irq);
1784         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1785                 evtchn_to_irq[i] = -1;
1786
1787         init_evtchn_cpu_bindings();
1788
1789         /* No event channels are 'live' right now. */
1790         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1791                 mask_evtchn(i);
1792
1793         pirq_needs_eoi = pirq_needs_eoi_flag;
1794
1795         if (xen_hvm_domain()) {
1796                 xen_callback_vector();
1797                 native_init_IRQ();
1798                 /* pci_xen_hvm_init must be called after native_init_IRQ so that
1799                  * __acpi_register_gsi can point at the right function */
1800                 pci_xen_hvm_init();
1801         } else {
1802                 struct physdev_pirq_eoi_gmfn eoi_gmfn;
1803
1804                 irq_ctx_init(smp_processor_id());
1805                 if (xen_initial_domain())
1806                         pci_xen_initial_domain();
1807
1808                 pirq_eoi_map = (void *)__get_free_page(GFP_KERNEL|__GFP_ZERO);
1809                 eoi_gmfn.gmfn = virt_to_mfn(pirq_eoi_map);
1810                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_pirq_eoi_gmfn_v2, &eoi_gmfn);
1811                 if (rc != 0) {
1812                         free_page((unsigned long) pirq_eoi_map);
1813                         pirq_eoi_map = NULL;
1814                 } else
1815                         pirq_needs_eoi = pirq_check_eoi_map;
1816         }
1817 }