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_MUTEX(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, NR_EVENT_CHANNELS/8);
 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         if (irq >= 0)
 435                 xen_irq_init(irq);
 436
 437         return irq;
 438 }
 439
 440 static int __must_check xen_allocate_irq_gsi(unsigned gsi)
 441 {
 442         int irq;
 443
 444         /*
 445          * A PV guest has no concept of a GSI (since it has no ACPI
 446          * nor access to/knowledge of the physical APICs). Therefore
 447          * all IRQs are dynamically allocated from the entire IRQ
 448          * space.
 449          */
 450         if (xen_pv_domain() && !xen_initial_domain())
 451                 return xen_allocate_irq_dynamic();
 452
 453         /* Legacy IRQ descriptors are already allocated by the arch. */
 454         if (gsi < NR_IRQS_LEGACY)
 455                 irq = gsi;
 456         else
 457                 irq = irq_alloc_desc_at(gsi, -1);
 458
 459         xen_irq_init(irq);
 460
 461         return irq;
 462 }
 463
 464 static void xen_free_irq(unsigned irq)
 465 {
 466         struct irq_info *info = irq_get_handler_data(irq);
 467
 468         list_del(&info->list);
 469
 470         irq_set_handler_data(irq, NULL);
 471
 472         kfree(info);
 473
 474         /* Legacy IRQ descriptors are managed by the arch. */
 475         if (irq < NR_IRQS_LEGACY)
 476                 return;
 477
 478         irq_free_desc(irq);
 479 }
 480
 481 static void pirq_query_unmask(int irq)
 482 {
 483         struct physdev_irq_status_query irq_status;
 484         struct irq_info *info = info_for_irq(irq);
 485
 486         BUG_ON(info->type != IRQT_PIRQ);
 487
 488         irq_status.irq = pirq_from_irq(irq);
 489         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
 490                 irq_status.flags = 0;
 491
 492         info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
 493         if (irq_status.flags & XENIRQSTAT_needs_eoi)
 494                 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
 495 }
 496
 497 static bool probing_irq(int irq)
 498 {
 499         struct irq_desc *desc = irq_to_desc(irq);
 500
 501         return desc && desc->action == NULL;
 502 }
 503
 504 static void eoi_pirq(struct irq_data *data)
 505 {
 506         int evtchn = evtchn_from_irq(data->irq);
 507         struct physdev_eoi eoi = { .irq = pirq_from_irq(data->irq) };
 508         int rc = 0;
 509
 510         irq_move_irq(data);
 511
 512         if (VALID_EVTCHN(evtchn))
 513                 clear_evtchn(evtchn);
 514
 515         if (pirq_needs_eoi(data->irq)) {
 516                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
 517                 WARN_ON(rc);
 518         }
 519 }
 520
 521 static void mask_ack_pirq(struct irq_data *data)
 522 {
 523         disable_dynirq(data);
 524         eoi_pirq(data);
 525 }
 526
 527 static unsigned int __startup_pirq(unsigned int irq)
 528 {
 529         struct evtchn_bind_pirq bind_pirq;
 530         struct irq_info *info = info_for_irq(irq);
 531         int evtchn = evtchn_from_irq(irq);
 532         int rc;
 533
 534         BUG_ON(info->type != IRQT_PIRQ);
 535
 536         if (VALID_EVTCHN(evtchn))
 537                 goto out;
 538
 539         bind_pirq.pirq = pirq_from_irq(irq);
 540         /* NB. We are happy to share unless we are probing. */
 541         bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
 542                                         BIND_PIRQ__WILL_SHARE : 0;
 543         rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
 544         if (rc != 0) {
 545                 if (!probing_irq(irq))
 546                         printk(KERN_INFO "Failed to obtain physical IRQ %d\n",
 547                                irq);
 548                 return 0;
 549         }
 550         evtchn = bind_pirq.port;
 551
 552         pirq_query_unmask(irq);
 553
 554         evtchn_to_irq[evtchn] = irq;
 555         bind_evtchn_to_cpu(evtchn, 0);
 556         info->evtchn = evtchn;
 557
 558 out:
 559         unmask_evtchn(evtchn);
 560         eoi_pirq(irq_get_irq_data(irq));
 561
 562         return 0;
 563 }
 564
 565 static unsigned int startup_pirq(struct irq_data *data)
 566 {
 567         return __startup_pirq(data->irq);
 568 }
 569
 570 static void shutdown_pirq(struct irq_data *data)
 571 {
 572         struct evtchn_close close;
 573         unsigned int irq = data->irq;
 574         struct irq_info *info = info_for_irq(irq);
 575         int evtchn = evtchn_from_irq(irq);
 576
 577         BUG_ON(info->type != IRQT_PIRQ);
 578
 579         if (!VALID_EVTCHN(evtchn))
 580                 return;
 581
 582         mask_evtchn(evtchn);
 583
 584         close.port = evtchn;
 585         if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 586                 BUG();
 587
 588         bind_evtchn_to_cpu(evtchn, 0);
 589         evtchn_to_irq[evtchn] = -1;
 590         info->evtchn = 0;
 591 }
 592
 593 static void enable_pirq(struct irq_data *data)
 594 {
 595         startup_pirq(data);
 596 }
 597
 598 static void disable_pirq(struct irq_data *data)
 599 {
 600         disable_dynirq(data);
 601 }
 602
 603 int xen_irq_from_gsi(unsigned gsi)
 604 {
 605         struct irq_info *info;
 606
 607         list_for_each_entry(info, &xen_irq_list_head, list) {
 608                 if (info->type != IRQT_PIRQ)
 609                         continue;
 610
 611                 if (info->u.pirq.gsi == gsi)
 612                         return info->irq;
 613         }
 614
 615         return -1;
 616 }
 617 EXPORT_SYMBOL_GPL(xen_irq_from_gsi);
 618
 619 /*
 620  * Do not make any assumptions regarding the relationship between the
 621  * IRQ number returned here and the Xen pirq argument.
 622  *
 623  * Note: We don't assign an event channel until the irq actually started
 624  * up.  Return an existing irq if we've already got one for the gsi.
 625  *
 626  * Shareable implies level triggered, not shareable implies edge
 627  * triggered here.
 628  */
 629 int xen_bind_pirq_gsi_to_irq(unsigned gsi,
 630                              unsigned pirq, int shareable, char *name)
 631 {
 632         int irq = -1;
 633         struct physdev_irq irq_op;
 634
 635         mutex_lock(&irq_mapping_update_lock);
 636
 637         irq = xen_irq_from_gsi(gsi);
 638         if (irq != -1) {
 639                 printk(KERN_INFO "xen_map_pirq_gsi: returning irq %d for gsi %u\n",
 640                        irq, gsi);
 641                 goto out;       /* XXX need refcount? */
 642         }
 643
 644         irq = xen_allocate_irq_gsi(gsi);
 645         if (irq < 0)
 646                 goto out;
 647
 648         irq_op.irq = irq;
 649         irq_op.vector = 0;
 650
 651         /* Only the privileged domain can do this. For non-priv, the pcifront
 652          * driver provides a PCI bus that does the call to do exactly
 653          * this in the priv domain. */
 654         if (xen_initial_domain() &&
 655             HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
 656                 xen_free_irq(irq);
 657                 irq = -ENOSPC;
 658                 goto out;
 659         }
 660
 661         xen_irq_info_pirq_init(irq, 0, pirq, gsi, irq_op.vector, DOMID_SELF,
 662                                shareable ? PIRQ_SHAREABLE : 0);
 663
 664         pirq_query_unmask(irq);
 665         /* We try to use the handler with the appropriate semantic for the
 666          * type of interrupt: if the interrupt is an edge triggered
 667          * interrupt we use handle_edge_irq.
 668          *
 669          * On the other hand if the interrupt is level triggered we use
 670          * handle_fasteoi_irq like the native code does for this kind of
 671          * interrupts.
 672          *
 673          * Depending on the Xen version, pirq_needs_eoi might return true
 674          * not only for level triggered interrupts but for edge triggered
 675          * interrupts too. In any case Xen always honors the eoi mechanism,
 676          * not injecting any more pirqs of the same kind if the first one
 677          * hasn't received an eoi yet. Therefore using the fasteoi handler
 678          * is the right choice either way.
 679          */
 680         if (shareable)
 681                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 682                                 handle_fasteoi_irq, name);
 683         else
 684                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 685                                 handle_edge_irq, name);
 686
 687 out:
 688         mutex_unlock(&irq_mapping_update_lock);
 689
 690         return irq;
 691 }
 692
 693 #ifdef CONFIG_PCI_MSI
 694 int xen_allocate_pirq_msi(struct pci_dev *dev, struct msi_desc *msidesc)
 695 {
 696         int rc;
 697         struct physdev_get_free_pirq op_get_free_pirq;
 698
 699         op_get_free_pirq.type = MAP_PIRQ_TYPE_MSI;
 700         rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
 701
 702         WARN_ONCE(rc == -ENOSYS,
 703                   "hypervisor does not support the PHYSDEVOP_get_free_pirq interface\n");
 704
 705         return rc ? -1 : op_get_free_pirq.pirq;
 706 }
 707
 708 int xen_bind_pirq_msi_to_irq(struct pci_dev *dev, struct msi_desc *msidesc,
 709                              int pirq, int vector, const char *name,
 710                              domid_t domid)
 711 {
 712         int irq, ret;
 713
 714         mutex_lock(&irq_mapping_update_lock);
 715
 716         irq = xen_allocate_irq_dynamic();
 717         if (irq < 0)
 718                 goto out;
 719
 720         irq_set_chip_and_handler_name(irq, &xen_pirq_chip, handle_edge_irq,
 721                         name);
 722
 723         xen_irq_info_pirq_init(irq, 0, pirq, 0, vector, domid, 0);
 724         ret = irq_set_msi_desc(irq, msidesc);
 725         if (ret < 0)
 726                 goto error_irq;
 727 out:
 728         mutex_unlock(&irq_mapping_update_lock);
 729         return irq;
 730 error_irq:
 731         mutex_unlock(&irq_mapping_update_lock);
 732         xen_free_irq(irq);
 733         return ret;
 734 }
 735 #endif
 736
 737 int xen_destroy_irq(int irq)
 738 {
 739         struct irq_desc *desc;
 740         struct physdev_unmap_pirq unmap_irq;
 741         struct irq_info *info = info_for_irq(irq);
 742         int rc = -ENOENT;
 743
 744         mutex_lock(&irq_mapping_update_lock);
 745
 746         desc = irq_to_desc(irq);
 747         if (!desc)
 748                 goto out;
 749
 750         if (xen_initial_domain()) {
 751                 unmap_irq.pirq = info->u.pirq.pirq;
 752                 unmap_irq.domid = info->u.pirq.domid;
 753                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
 754                 /* If another domain quits without making the pci_disable_msix
 755                  * call, the Xen hypervisor takes care of freeing the PIRQs
 756                  * (free_domain_pirqs).
 757                  */
 758                 if ((rc == -ESRCH && info->u.pirq.domid != DOMID_SELF))
 759                         printk(KERN_INFO "domain %d does not have %d anymore\n",
 760                                 info->u.pirq.domid, info->u.pirq.pirq);
 761                 else if (rc) {
 762                         printk(KERN_WARNING "unmap irq failed %d\n", rc);
 763                         goto out;
 764                 }
 765         }
 766
 767         xen_free_irq(irq);
 768
 769 out:
 770         mutex_unlock(&irq_mapping_update_lock);
 771         return rc;
 772 }
 773
 774 int xen_irq_from_pirq(unsigned pirq)
 775 {
 776         int irq;
 777
 778         struct irq_info *info;
 779
 780         mutex_lock(&irq_mapping_update_lock);
 781
 782         list_for_each_entry(info, &xen_irq_list_head, list) {
 783                 if (info->type != IRQT_PIRQ)
 784                         continue;
 785                 irq = info->irq;
 786                 if (info->u.pirq.pirq == pirq)
 787                         goto out;
 788         }
 789         irq = -1;
 790 out:
 791         mutex_unlock(&irq_mapping_update_lock);
 792
 793         return irq;
 794 }
 795
 796
 797 int xen_pirq_from_irq(unsigned irq)
 798 {
 799         return pirq_from_irq(irq);
 800 }
 801 EXPORT_SYMBOL_GPL(xen_pirq_from_irq);
 802 int bind_evtchn_to_irq(unsigned int evtchn)
 803 {
 804         int irq;
 805
 806         mutex_lock(&irq_mapping_update_lock);
 807
 808         irq = evtchn_to_irq[evtchn];
 809
 810         if (irq == -1) {
 811                 irq = xen_allocate_irq_dynamic();
 812                 if (irq == -1)
 813                         goto out;
 814
 815                 irq_set_chip_and_handler_name(irq, &xen_dynamic_chip,
 816                                               handle_edge_irq, "event");
 817
 818                 xen_irq_info_evtchn_init(irq, evtchn);
 819         }
 820
 821 out:
 822         mutex_unlock(&irq_mapping_update_lock);
 823
 824         return irq;
 825 }
 826 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
 827
 828 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
 829 {
 830         struct evtchn_bind_ipi bind_ipi;
 831         int evtchn, irq;
 832
 833         mutex_lock(&irq_mapping_update_lock);
 834
 835         irq = per_cpu(ipi_to_irq, cpu)[ipi];
 836
 837         if (irq == -1) {
 838                 irq = xen_allocate_irq_dynamic();
 839                 if (irq < 0)
 840                         goto out;
 841
 842                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 843                                               handle_percpu_irq, "ipi");
 844
 845                 bind_ipi.vcpu = cpu;
 846                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
 847                                                 &bind_ipi) != 0)
 848                         BUG();
 849                 evtchn = bind_ipi.port;
 850
 851                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
 852
 853                 bind_evtchn_to_cpu(evtchn, cpu);
 854         }
 855
 856  out:
 857         mutex_unlock(&irq_mapping_update_lock);
 858         return irq;
 859 }
 860
 861 static int bind_interdomain_evtchn_to_irq(unsigned int remote_domain,
 862                                           unsigned int remote_port)
 863 {
 864         struct evtchn_bind_interdomain bind_interdomain;
 865         int err;
 866
 867         bind_interdomain.remote_dom  = remote_domain;
 868         bind_interdomain.remote_port = remote_port;
 869
 870         err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
 871                                           &bind_interdomain);
 872
 873         return err ? : bind_evtchn_to_irq(bind_interdomain.local_port);
 874 }
 875
 876 static int find_virq(unsigned int virq, unsigned int cpu)
 877 {
 878         struct evtchn_status status;
 879         int port, rc = -ENOENT;
 880
 881         memset(&status, 0, sizeof(status));
 882         for (port = 0; port <= NR_EVENT_CHANNELS; port++) {
 883                 status.dom = DOMID_SELF;
 884                 status.port = port;
 885                 rc = HYPERVISOR_event_channel_op(EVTCHNOP_status, &status);
 886                 if (rc < 0)
 887                         continue;
 888                 if (status.status != EVTCHNSTAT_virq)
 889                         continue;
 890                 if (status.u.virq == virq && status.vcpu == cpu) {
 891                         rc = port;
 892                         break;
 893                 }
 894         }
 895         return rc;
 896 }
 897
 898 int bind_virq_to_irq(unsigned int virq, unsigned int cpu, bool percpu)
 899 {
 900         struct evtchn_bind_virq bind_virq;
 901         int evtchn, irq, ret;
 902
 903         mutex_lock(&irq_mapping_update_lock);
 904
 905         irq = per_cpu(virq_to_irq, cpu)[virq];
 906
 907         if (irq == -1) {
 908                 irq = xen_allocate_irq_dynamic();
 909                 if (irq == -1)
 910                         goto out;
 911
 912                 if (percpu)
 913                         irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 914                                                       handle_percpu_irq, "virq");
 915                 else
 916                         irq_set_chip_and_handler_name(irq, &xen_dynamic_chip,
 917                                                       handle_edge_irq, "virq");
 918
 919                 bind_virq.virq = virq;
 920                 bind_virq.vcpu = cpu;
 921                 ret = HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 922                                                 &bind_virq);
 923                 if (ret == 0)
 924                         evtchn = bind_virq.port;
 925                 else {
 926                         if (ret == -EEXIST)
 927                                 ret = find_virq(virq, cpu);
 928                         BUG_ON(ret < 0);
 929                         evtchn = ret;
 930                 }
 931
 932                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
 933
 934                 bind_evtchn_to_cpu(evtchn, cpu);
 935         }
 936
 937 out:
 938         mutex_unlock(&irq_mapping_update_lock);
 939
 940         return irq;
 941 }
 942
 943 static void unbind_from_irq(unsigned int irq)
 944 {
 945         struct evtchn_close close;
 946         int evtchn = evtchn_from_irq(irq);
 947
 948         mutex_lock(&irq_mapping_update_lock);
 949
 950         if (VALID_EVTCHN(evtchn)) {
 951                 close.port = evtchn;
 952                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 953                         BUG();
 954
 955                 switch (type_from_irq(irq)) {
 956                 case IRQT_VIRQ:
 957                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
 958                                 [virq_from_irq(irq)] = -1;
 959                         break;
 960                 case IRQT_IPI:
 961                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
 962                                 [ipi_from_irq(irq)] = -1;
 963                         break;
 964                 default:
 965                         break;
 966                 }
 967
 968                 /* Closed ports are implicitly re-bound to VCPU0. */
 969                 bind_evtchn_to_cpu(evtchn, 0);
 970
 971                 evtchn_to_irq[evtchn] = -1;
 972         }
 973
 974         BUG_ON(info_for_irq(irq)->type == IRQT_UNBOUND);
 975
 976         xen_free_irq(irq);
 977
 978         mutex_unlock(&irq_mapping_update_lock);
 979 }
 980
 981 int bind_evtchn_to_irqhandler(unsigned int evtchn,
 982                               irq_handler_t handler,
 983                               unsigned long irqflags,
 984                               const char *devname, void *dev_id)
 985 {
 986         int irq, retval;
 987
 988         irq = bind_evtchn_to_irq(evtchn);
 989         if (irq < 0)
 990                 return irq;
 991         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 992         if (retval != 0) {
 993                 unbind_from_irq(irq);
 994                 return retval;
 995         }
 996
 997         return irq;
 998 }
 999 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
1000
1001 int bind_interdomain_evtchn_to_irqhandler(unsigned int remote_domain,
1002                                           unsigned int remote_port,
1003                                           irq_handler_t handler,
1004                                           unsigned long irqflags,
1005                                           const char *devname,
1006                                           void *dev_id)
1007 {
1008         int irq, retval;
1009
1010         irq = bind_interdomain_evtchn_to_irq(remote_domain, remote_port);
1011         if (irq < 0)
1012                 return irq;
1013
1014         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1015         if (retval != 0) {
1016                 unbind_from_irq(irq);
1017                 return retval;
1018         }
1019
1020         return irq;
1021 }
1022 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler);
1023
1024 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
1025                             irq_handler_t handler,
1026                             unsigned long irqflags, const char *devname, void *dev_id)
1027 {
1028         int irq, retval;
1029
1030         irq = bind_virq_to_irq(virq, cpu, irqflags & IRQF_PERCPU);
1031         if (irq < 0)
1032                 return irq;
1033         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1034         if (retval != 0) {
1035                 unbind_from_irq(irq);
1036                 return retval;
1037         }
1038
1039         return irq;
1040 }
1041 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1042
1043 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1044                            unsigned int cpu,
1045                            irq_handler_t handler,
1046                            unsigned long irqflags,
1047                            const char *devname,
1048                            void *dev_id)
1049 {
1050         int irq, retval;
1051
1052         irq = bind_ipi_to_irq(ipi, cpu);
1053         if (irq < 0)
1054                 return irq;
1055
1056         irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME | IRQF_EARLY_RESUME;
1057         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1058         if (retval != 0) {
1059                 unbind_from_irq(irq);
1060                 return retval;
1061         }
1062
1063         return irq;
1064 }
1065
1066 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1067 {
1068         free_irq(irq, dev_id);
1069         unbind_from_irq(irq);
1070 }
1071 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1072
1073 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1074 {
1075         int irq = per_cpu(ipi_to_irq, cpu)[vector];
1076         BUG_ON(irq < 0);
1077         notify_remote_via_irq(irq);
1078 }
1079
1080 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
1081 {
1082         struct shared_info *sh = HYPERVISOR_shared_info;
1083         int cpu = smp_processor_id();
1084         unsigned long *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu);
1085         int i;
1086         unsigned long flags;
1087         static DEFINE_SPINLOCK(debug_lock);
1088         struct vcpu_info *v;
1089
1090         spin_lock_irqsave(&debug_lock, flags);
1091
1092         printk("\nvcpu %d\n  ", cpu);
1093
1094         for_each_online_cpu(i) {
1095                 int pending;
1096                 v = per_cpu(xen_vcpu, i);
1097                 pending = (get_irq_regs() && i == cpu)
1098                         ? xen_irqs_disabled(get_irq_regs())
1099                         : v->evtchn_upcall_mask;
1100                 printk("%d: masked=%d pending=%d event_sel %0*lx\n  ", i,
1101                        pending, v->evtchn_upcall_pending,
1102                        (int)(sizeof(v->evtchn_pending_sel)*2),
1103                        v->evtchn_pending_sel);
1104         }
1105         v = per_cpu(xen_vcpu, cpu);
1106
1107         printk("\npending:\n   ");
1108         for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
1109                 printk("%0*lx%s", (int)sizeof(sh->evtchn_pending[0])*2,
1110                        sh->evtchn_pending[i],
1111                        i % 8 == 0 ? "\n   " : " ");
1112         printk("\nglobal mask:\n   ");
1113         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1114                 printk("%0*lx%s",
1115                        (int)(sizeof(sh->evtchn_mask[0])*2),
1116                        sh->evtchn_mask[i],
1117                        i % 8 == 0 ? "\n   " : " ");
1118
1119         printk("\nglobally unmasked:\n   ");
1120         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1121                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1122                        sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
1123                        i % 8 == 0 ? "\n   " : " ");
1124
1125         printk("\nlocal cpu%d mask:\n   ", cpu);
1126         for (i = (NR_EVENT_CHANNELS/BITS_PER_LONG)-1; i >= 0; i--)
1127                 printk("%0*lx%s", (int)(sizeof(cpu_evtchn[0])*2),
1128                        cpu_evtchn[i],
1129                        i % 8 == 0 ? "\n   " : " ");
1130
1131         printk("\nlocally unmasked:\n   ");
1132         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
1133                 unsigned long pending = sh->evtchn_pending[i]
1134                         & ~sh->evtchn_mask[i]
1135                         & cpu_evtchn[i];
1136                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1137                        pending, i % 8 == 0 ? "\n   " : " ");
1138         }
1139
1140         printk("\npending list:\n");
1141         for (i = 0; i < NR_EVENT_CHANNELS; i++) {
1142                 if (sync_test_bit(i, sh->evtchn_pending)) {
1143                         int word_idx = i / BITS_PER_LONG;
1144                         printk("  %d: event %d -> irq %d%s%s%s\n",
1145                                cpu_from_evtchn(i), i,
1146                                evtchn_to_irq[i],
1147                                sync_test_bit(word_idx, &v->evtchn_pending_sel)
1148                                              ? "" : " l2-clear",
1149                                !sync_test_bit(i, sh->evtchn_mask)
1150                                              ? "" : " globally-masked",
1151                                sync_test_bit(i, cpu_evtchn)
1152                                              ? "" : " locally-masked");
1153                 }
1154         }
1155
1156         spin_unlock_irqrestore(&debug_lock, flags);
1157
1158         return IRQ_HANDLED;
1159 }
1160
1161 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
1162 static DEFINE_PER_CPU(unsigned int, current_word_idx);
1163 static DEFINE_PER_CPU(unsigned int, current_bit_idx);
1164
1165 /*
1166  * Mask out the i least significant bits of w
1167  */
1168 #define MASK_LSBS(w, i) (w & ((~0UL) << i))
1169
1170 /*
1171  * Search the CPUs pending events bitmasks.  For each one found, map
1172  * the event number to an irq, and feed it into do_IRQ() for
1173  * handling.
1174  *
1175  * Xen uses a two-level bitmap to speed searching.  The first level is
1176  * a bitset of words which contain pending event bits.  The second
1177  * level is a bitset of pending events themselves.
1178  */
1179 static void __xen_evtchn_do_upcall(void)
1180 {
1181         int start_word_idx, start_bit_idx;
1182         int word_idx, bit_idx;
1183         int i, irq;
1184         int cpu = get_cpu();
1185         struct shared_info *s = HYPERVISOR_shared_info;
1186         struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1187         unsigned count;
1188
1189         do {
1190                 unsigned long pending_words;
1191                 unsigned long pending_bits;
1192                 struct irq_desc *desc;
1193
1194                 vcpu_info->evtchn_upcall_pending = 0;
1195
1196                 if (__this_cpu_inc_return(xed_nesting_count) - 1)
1197                         goto out;
1198
1199 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
1200                 /* Clear master flag /before/ clearing selector flag. */
1201                 wmb();
1202 #endif
1203                 if ((irq = per_cpu(virq_to_irq, cpu)[VIRQ_TIMER]) != -1) {
1204                         int evtchn = evtchn_from_irq(irq);
1205                         word_idx = evtchn / BITS_PER_LONG;
1206                         pending_bits = evtchn % BITS_PER_LONG;
1207                         if (active_evtchns(cpu, s, word_idx) & (1ULL << pending_bits)) {
1208                                 desc = irq_to_desc(irq);
1209                                 if (desc)
1210                                         generic_handle_irq_desc(irq, desc);
1211                         }
1212                 }
1213
1214                 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
1215
1216                 start_word_idx = __this_cpu_read(current_word_idx);
1217                 start_bit_idx = __this_cpu_read(current_bit_idx);
1218
1219                 word_idx = start_word_idx;
1220
1221                 for (i = 0; pending_words != 0; i++) {
1222                         unsigned long words;
1223
1224                         words = MASK_LSBS(pending_words, word_idx);
1225
1226                         /*
1227                          * If we masked out all events, wrap to beginning.
1228                          */
1229                         if (words == 0) {
1230                                 word_idx = 0;
1231                                 bit_idx = 0;
1232                                 continue;
1233                         }
1234                         word_idx = __ffs(words);
1235
1236                         pending_bits = active_evtchns(cpu, s, word_idx);
1237                         bit_idx = 0; /* usually scan entire word from start */
1238                         if (word_idx == start_word_idx) {
1239                                 /* We scan the starting word in two parts */
1240                                 if (i == 0)
1241                                         /* 1st time: start in the middle */
1242                                         bit_idx = start_bit_idx;
1243                                 else
1244                                         /* 2nd time: mask bits done already */
1245                                         bit_idx &= (1UL << start_bit_idx) - 1;
1246                         }
1247
1248                         do {
1249                                 unsigned long bits;
1250                                 int port;
1251
1252                                 bits = MASK_LSBS(pending_bits, bit_idx);
1253
1254                                 /* If we masked out all events, move on. */
1255                                 if (bits == 0)
1256                                         break;
1257
1258                                 bit_idx = __ffs(bits);
1259
1260                                 /* Process port. */
1261                                 port = (word_idx * BITS_PER_LONG) + bit_idx;
1262                                 irq = evtchn_to_irq[port];
1263
1264                                 if (irq != -1) {
1265                                         desc = irq_to_desc(irq);
1266                                         if (desc)
1267                                                 generic_handle_irq_desc(irq, desc);
1268                                 }
1269
1270                                 bit_idx = (bit_idx + 1) % BITS_PER_LONG;
1271
1272                                 /* Next caller starts at last processed + 1 */
1273                                 __this_cpu_write(current_word_idx,
1274                                                  bit_idx ? word_idx :
1275                                                  (word_idx+1) % BITS_PER_LONG);
1276                                 __this_cpu_write(current_bit_idx, bit_idx);
1277                         } while (bit_idx != 0);
1278
1279                         /* Scan start_l1i twice; all others once. */
1280                         if ((word_idx != start_word_idx) || (i != 0))
1281                                 pending_words &= ~(1UL << word_idx);
1282
1283                         word_idx = (word_idx + 1) % BITS_PER_LONG;
1284                 }
1285
1286                 BUG_ON(!irqs_disabled());
1287
1288                 count = __this_cpu_read(xed_nesting_count);
1289                 __this_cpu_write(xed_nesting_count, 0);
1290         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
1291
1292 out:
1293
1294         put_cpu();
1295 }
1296
1297 void xen_evtchn_do_upcall(struct pt_regs *regs)
1298 {
1299         struct pt_regs *old_regs = set_irq_regs(regs);
1300
1301         exit_idle();
1302         irq_enter();
1303
1304         __xen_evtchn_do_upcall();
1305
1306         irq_exit();
1307         set_irq_regs(old_regs);
1308 }
1309
1310 void xen_hvm_evtchn_do_upcall(void)
1311 {
1312         __xen_evtchn_do_upcall();
1313 }
1314 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1315
1316 /* Rebind a new event channel to an existing irq. */
1317 void rebind_evtchn_irq(int evtchn, int irq)
1318 {
1319         struct irq_info *info = info_for_irq(irq);
1320
1321         /* Make sure the irq is masked, since the new event channel
1322            will also be masked. */
1323         disable_irq(irq);
1324
1325         mutex_lock(&irq_mapping_update_lock);
1326
1327         /* After resume the irq<->evtchn mappings are all cleared out */
1328         BUG_ON(evtchn_to_irq[evtchn] != -1);
1329         /* Expect irq to have been bound before,
1330            so there should be a proper type */
1331         BUG_ON(info->type == IRQT_UNBOUND);
1332
1333         xen_irq_info_evtchn_init(irq, evtchn);
1334
1335         mutex_unlock(&irq_mapping_update_lock);
1336
1337         /* new event channels are always bound to cpu 0 */
1338         irq_set_affinity(irq, cpumask_of(0));
1339
1340         /* Unmask the event channel. */
1341         enable_irq(irq);
1342 }
1343
1344 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1345 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1346 {
1347         struct shared_info *s = HYPERVISOR_shared_info;
1348         struct evtchn_bind_vcpu bind_vcpu;
1349         int evtchn = evtchn_from_irq(irq);
1350         int masked;
1351
1352         if (!VALID_EVTCHN(evtchn))
1353                 return -1;
1354
1355         /*
1356          * Events delivered via platform PCI interrupts are always
1357          * routed to vcpu 0 and hence cannot be rebound.
1358          */
1359         if (xen_hvm_domain() && !xen_have_vector_callback)
1360                 return -1;
1361
1362         /* Send future instances of this interrupt to other vcpu. */
1363         bind_vcpu.port = evtchn;
1364         bind_vcpu.vcpu = tcpu;
1365
1366         /*
1367          * Mask the event while changing the VCPU binding to prevent
1368          * it being delivered on an unexpected VCPU.
1369          */
1370         masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
1371
1372         /*
1373          * If this fails, it usually just indicates that we're dealing with a
1374          * virq or IPI channel, which don't actually need to be rebound. Ignore
1375          * it, but don't do the xenlinux-level rebind in that case.
1376          */
1377         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1378                 bind_evtchn_to_cpu(evtchn, tcpu);
1379
1380         if (!masked)
1381                 unmask_evtchn(evtchn);
1382
1383         return 0;
1384 }
1385
1386 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1387                             bool force)
1388 {
1389         unsigned tcpu = cpumask_first(dest);
1390
1391         return rebind_irq_to_cpu(data->irq, tcpu);
1392 }
1393
1394 int resend_irq_on_evtchn(unsigned int irq)
1395 {
1396         int masked, evtchn = evtchn_from_irq(irq);
1397         struct shared_info *s = HYPERVISOR_shared_info;
1398
1399         if (!VALID_EVTCHN(evtchn))
1400                 return 1;
1401
1402         masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
1403         sync_set_bit(evtchn, s->evtchn_pending);
1404         if (!masked)
1405                 unmask_evtchn(evtchn);
1406
1407         return 1;
1408 }
1409
1410 static void enable_dynirq(struct irq_data *data)
1411 {
1412         int evtchn = evtchn_from_irq(data->irq);
1413
1414         if (VALID_EVTCHN(evtchn))
1415                 unmask_evtchn(evtchn);
1416 }
1417
1418 static void disable_dynirq(struct irq_data *data)
1419 {
1420         int evtchn = evtchn_from_irq(data->irq);
1421
1422         if (VALID_EVTCHN(evtchn))
1423                 mask_evtchn(evtchn);
1424 }
1425
1426 static void ack_dynirq(struct irq_data *data)
1427 {
1428         int evtchn = evtchn_from_irq(data->irq);
1429
1430         irq_move_irq(data);
1431
1432         if (VALID_EVTCHN(evtchn))
1433                 clear_evtchn(evtchn);
1434 }
1435
1436 static void mask_ack_dynirq(struct irq_data *data)
1437 {
1438         disable_dynirq(data);
1439         ack_dynirq(data);
1440 }
1441
1442 static int retrigger_dynirq(struct irq_data *data)
1443 {
1444         int evtchn = evtchn_from_irq(data->irq);
1445         struct shared_info *sh = HYPERVISOR_shared_info;
1446         int ret = 0;
1447
1448         if (VALID_EVTCHN(evtchn)) {
1449                 int masked;
1450
1451                 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
1452                 sync_set_bit(evtchn, sh->evtchn_pending);
1453                 if (!masked)
1454                         unmask_evtchn(evtchn);
1455                 ret = 1;
1456         }
1457
1458         return ret;
1459 }
1460
1461 static void restore_pirqs(void)
1462 {
1463         int pirq, rc, irq, gsi;
1464         struct physdev_map_pirq map_irq;
1465         struct irq_info *info;
1466
1467         list_for_each_entry(info, &xen_irq_list_head, list) {
1468                 if (info->type != IRQT_PIRQ)
1469                         continue;
1470
1471                 pirq = info->u.pirq.pirq;
1472                 gsi = info->u.pirq.gsi;
1473                 irq = info->irq;
1474
1475                 /* save/restore of PT devices doesn't work, so at this point the
1476                  * only devices present are GSI based emulated devices */
1477                 if (!gsi)
1478                         continue;
1479
1480                 map_irq.domid = DOMID_SELF;
1481                 map_irq.type = MAP_PIRQ_TYPE_GSI;
1482                 map_irq.index = gsi;
1483                 map_irq.pirq = pirq;
1484
1485                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1486                 if (rc) {
1487                         printk(KERN_WARNING "xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1488                                         gsi, irq, pirq, rc);
1489                         xen_free_irq(irq);
1490                         continue;
1491                 }
1492
1493                 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1494
1495                 __startup_pirq(irq);
1496         }
1497 }
1498
1499 static void restore_cpu_virqs(unsigned int cpu)
1500 {
1501         struct evtchn_bind_virq bind_virq;
1502         int virq, irq, evtchn;
1503
1504         for (virq = 0; virq < NR_VIRQS; virq++) {
1505                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1506                         continue;
1507
1508                 BUG_ON(virq_from_irq(irq) != virq);
1509
1510                 /* Get a new binding from Xen. */
1511                 bind_virq.virq = virq;
1512                 bind_virq.vcpu = cpu;
1513                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1514                                                 &bind_virq) != 0)
1515                         BUG();
1516                 evtchn = bind_virq.port;
1517
1518                 /* Record the new mapping. */
1519                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
1520                 bind_evtchn_to_cpu(evtchn, cpu);
1521         }
1522 }
1523
1524 static void restore_cpu_ipis(unsigned int cpu)
1525 {
1526         struct evtchn_bind_ipi bind_ipi;
1527         int ipi, irq, evtchn;
1528
1529         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1530                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1531                         continue;
1532
1533                 BUG_ON(ipi_from_irq(irq) != ipi);
1534
1535                 /* Get a new binding from Xen. */
1536                 bind_ipi.vcpu = cpu;
1537                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1538                                                 &bind_ipi) != 0)
1539                         BUG();
1540                 evtchn = bind_ipi.port;
1541
1542                 /* Record the new mapping. */
1543                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
1544                 bind_evtchn_to_cpu(evtchn, cpu);
1545         }
1546 }
1547
1548 /* Clear an irq's pending state, in preparation for polling on it */
1549 void xen_clear_irq_pending(int irq)
1550 {
1551         int evtchn = evtchn_from_irq(irq);
1552
1553         if (VALID_EVTCHN(evtchn))
1554                 clear_evtchn(evtchn);
1555 }
1556 EXPORT_SYMBOL(xen_clear_irq_pending);
1557 void xen_set_irq_pending(int irq)
1558 {
1559         int evtchn = evtchn_from_irq(irq);
1560
1561         if (VALID_EVTCHN(evtchn))
1562                 set_evtchn(evtchn);
1563 }
1564
1565 bool xen_test_irq_pending(int irq)
1566 {
1567         int evtchn = evtchn_from_irq(irq);
1568         bool ret = false;
1569
1570         if (VALID_EVTCHN(evtchn))
1571                 ret = test_evtchn(evtchn);
1572
1573         return ret;
1574 }
1575
1576 /* Poll waiting for an irq to become pending with timeout.  In the usual case,
1577  * the irq will be disabled so it won't deliver an interrupt. */
1578 void xen_poll_irq_timeout(int irq, u64 timeout)
1579 {
1580         evtchn_port_t evtchn = evtchn_from_irq(irq);
1581
1582         if (VALID_EVTCHN(evtchn)) {
1583                 struct sched_poll poll;
1584
1585                 poll.nr_ports = 1;
1586                 poll.timeout = timeout;
1587                 set_xen_guest_handle(poll.ports, &evtchn);
1588
1589                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1590                         BUG();
1591         }
1592 }
1593 EXPORT_SYMBOL(xen_poll_irq_timeout);
1594 /* Poll waiting for an irq to become pending.  In the usual case, the
1595  * irq will be disabled so it won't deliver an interrupt. */
1596 void xen_poll_irq(int irq)
1597 {
1598         xen_poll_irq_timeout(irq, 0 /* no timeout */);
1599 }
1600
1601 /* Check whether the IRQ line is shared with other guests. */
1602 int xen_test_irq_shared(int irq)
1603 {
1604         struct irq_info *info = info_for_irq(irq);
1605         struct physdev_irq_status_query irq_status = { .irq = info->u.pirq.pirq };
1606
1607         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
1608                 return 0;
1609         return !(irq_status.flags & XENIRQSTAT_shared);
1610 }
1611 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
1612
1613 void xen_irq_resume(void)
1614 {
1615         unsigned int cpu, evtchn;
1616         struct irq_info *info;
1617
1618         init_evtchn_cpu_bindings();
1619
1620         /* New event-channel space is not 'live' yet. */
1621         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1622                 mask_evtchn(evtchn);
1623
1624         /* No IRQ <-> event-channel mappings. */
1625         list_for_each_entry(info, &xen_irq_list_head, list)
1626                 info->evtchn = 0; /* zap event-channel binding */
1627
1628         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1629                 evtchn_to_irq[evtchn] = -1;
1630
1631         for_each_possible_cpu(cpu) {
1632                 restore_cpu_virqs(cpu);
1633                 restore_cpu_ipis(cpu);
1634         }
1635
1636         restore_pirqs();
1637 }
1638
1639 static struct irq_chip xen_dynamic_chip __read_mostly = {
1640         .name                   = "xen-dyn",
1641
1642         .irq_disable            = disable_dynirq,
1643         .irq_mask               = disable_dynirq,
1644         .irq_unmask             = enable_dynirq,
1645
1646         .irq_ack                = ack_dynirq,
1647         .irq_mask_ack           = mask_ack_dynirq,
1648
1649         .irq_set_affinity       = set_affinity_irq,
1650         .irq_retrigger          = retrigger_dynirq,
1651 };
1652
1653 static struct irq_chip xen_pirq_chip __read_mostly = {
1654         .name                   = "xen-pirq",
1655
1656         .irq_startup            = startup_pirq,
1657         .irq_shutdown           = shutdown_pirq,
1658         .irq_enable             = enable_pirq,
1659         .irq_disable            = disable_pirq,
1660
1661         .irq_mask               = disable_dynirq,
1662         .irq_unmask             = enable_dynirq,
1663
1664         .irq_ack                = eoi_pirq,
1665         .irq_eoi                = eoi_pirq,
1666         .irq_mask_ack           = mask_ack_pirq,
1667
1668         .irq_set_affinity       = set_affinity_irq,
1669
1670         .irq_retrigger          = retrigger_dynirq,
1671 };
1672
1673 static struct irq_chip xen_percpu_chip __read_mostly = {
1674         .name                   = "xen-percpu",
1675
1676         .irq_disable            = disable_dynirq,
1677         .irq_mask               = disable_dynirq,
1678         .irq_unmask             = enable_dynirq,
1679
1680         .irq_ack                = ack_dynirq,
1681 };
1682
1683 int xen_set_callback_via(uint64_t via)
1684 {
1685         struct xen_hvm_param a;
1686         a.domid = DOMID_SELF;
1687         a.index = HVM_PARAM_CALLBACK_IRQ;
1688         a.value = via;
1689         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1690 }
1691 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1692
1693 #ifdef CONFIG_XEN_PVHVM
1694 /* Vector callbacks are better than PCI interrupts to receive event
1695  * channel notifications because we can receive vector callbacks on any
1696  * vcpu and we don't need PCI support or APIC interactions. */
1697 void xen_callback_vector(void)
1698 {
1699         int rc;
1700         uint64_t callback_via;
1701         if (xen_have_vector_callback) {
1702                 callback_via = HVM_CALLBACK_VECTOR(XEN_HVM_EVTCHN_CALLBACK);
1703                 rc = xen_set_callback_via(callback_via);
1704                 if (rc) {
1705                         printk(KERN_ERR "Request for Xen HVM callback vector"
1706                                         " failed.\n");
1707                         xen_have_vector_callback = 0;
1708                         return;
1709                 }
1710                 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1711                                 "enabled\n");
1712                 /* in the restore case the vector has already been allocated */
1713                 if (!test_bit(XEN_HVM_EVTCHN_CALLBACK, used_vectors))
1714                         alloc_intr_gate(XEN_HVM_EVTCHN_CALLBACK, xen_hvm_callback_vector);
1715         }
1716 }
1717 #else
1718 void xen_callback_vector(void) {}
1719 #endif
1720
1721 void __init xen_init_IRQ(void)
1722 {
1723         int i;
1724
1725         evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1726                                     GFP_KERNEL);
1727         BUG_ON(!evtchn_to_irq);
1728         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1729                 evtchn_to_irq[i] = -1;
1730
1731         init_evtchn_cpu_bindings();
1732
1733         /* No event channels are 'live' right now. */
1734         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1735                 mask_evtchn(i);
1736
1737         if (xen_hvm_domain()) {
1738                 xen_callback_vector();
1739                 native_init_IRQ();
1740                 /* pci_xen_hvm_init must be called after native_init_IRQ so that
1741                  * __acpi_register_gsi can point at the right function */
1742                 pci_xen_hvm_init();
1743         } else {
1744                 irq_ctx_init(smp_processor_id());
1745                 if (xen_initial_domain())
1746                         pci_xen_initial_domain();
1747         }
1748 }