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