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 {
  90         struct list_head list;
  91         enum xen_irq_type type; /* type */
  92         unsigned irq;
  93         unsigned short evtchn;  /* event channel */
  94         unsigned short cpu;     /* cpu bound */
  95
  96         union {
  97                 unsigned short virq;
  98                 enum ipi_vector ipi;
  99                 struct {
 100                         unsigned short pirq;
 101                         unsigned short gsi;
 102                         unsigned char vector;
 103                         unsigned char flags;
 104                         uint16_t domid;
 105                 } pirq;
 106         } u;
 107 };
 108 #define PIRQ_NEEDS_EOI  (1 << 0)
 109 #define PIRQ_SHAREABLE  (1 << 1)
 110
 111 static int *evtchn_to_irq;
 112
 113 static DEFINE_PER_CPU(unsigned long [NR_EVENT_CHANNELS/BITS_PER_LONG],
 114                       cpu_evtchn_mask);
 115
 116 /* Xen will never allocate port zero for any purpose. */
 117 #define VALID_EVTCHN(chn)       ((chn) != 0)
 118
 119 static struct irq_chip xen_dynamic_chip;
 120 static struct irq_chip xen_percpu_chip;
 121 static struct irq_chip xen_pirq_chip;
 122 static void enable_dynirq(struct irq_data *data);
 123 static void disable_dynirq(struct irq_data *data);
 124
 125 /* Get info for IRQ */
 126 static struct irq_info *info_for_irq(unsigned irq)
 127 {
 128         return irq_get_handler_data(irq);
 129 }
 130
 131 /* Constructors for packed IRQ information. */
 132 static void xen_irq_info_common_init(struct irq_info *info,
 133                                      unsigned irq,
 134                                      enum xen_irq_type type,
 135                                      unsigned short evtchn,
 136                                      unsigned short cpu)
 137 {
 138
 139         BUG_ON(info->type != IRQT_UNBOUND && info->type != type);
 140
 141         info->type = type;
 142         info->irq = irq;
 143         info->evtchn = evtchn;
 144         info->cpu = cpu;
 145
 146         evtchn_to_irq[evtchn] = irq;
 147 }
 148
 149 static void xen_irq_info_evtchn_init(unsigned irq,
 150                                      unsigned short evtchn)
 151 {
 152         struct irq_info *info = info_for_irq(irq);
 153
 154         xen_irq_info_common_init(info, irq, IRQT_EVTCHN, evtchn, 0);
 155 }
 156
 157 static void xen_irq_info_ipi_init(unsigned cpu,
 158                                   unsigned irq,
 159                                   unsigned short evtchn,
 160                                   enum ipi_vector ipi)
 161 {
 162         struct irq_info *info = info_for_irq(irq);
 163
 164         xen_irq_info_common_init(info, irq, IRQT_IPI, evtchn, 0);
 165
 166         info->u.ipi = ipi;
 167
 168         per_cpu(ipi_to_irq, cpu)[ipi] = irq;
 169 }
 170
 171 static void xen_irq_info_virq_init(unsigned cpu,
 172                                    unsigned irq,
 173                                    unsigned short evtchn,
 174                                    unsigned short virq)
 175 {
 176         struct irq_info *info = info_for_irq(irq);
 177
 178         xen_irq_info_common_init(info, irq, IRQT_VIRQ, evtchn, 0);
 179
 180         info->u.virq = virq;
 181
 182         per_cpu(virq_to_irq, cpu)[virq] = irq;
 183 }
 184
 185 static void xen_irq_info_pirq_init(unsigned irq,
 186                                    unsigned short evtchn,
 187                                    unsigned short pirq,
 188                                    unsigned short gsi,
 189                                    unsigned short vector,
 190                                    uint16_t domid,
 191                                    unsigned char flags)
 192 {
 193         struct irq_info *info = info_for_irq(irq);
 194
 195         xen_irq_info_common_init(info, irq, IRQT_PIRQ, evtchn, 0);
 196
 197         info->u.pirq.pirq = pirq;
 198         info->u.pirq.gsi = gsi;
 199         info->u.pirq.vector = vector;
 200         info->u.pirq.domid = domid;
 201         info->u.pirq.flags = flags;
 202 }
 203
 204 /*
 205  * Accessors for packed IRQ information.
 206  */
 207 static unsigned int evtchn_from_irq(unsigned irq)
 208 {
 209         if (unlikely(WARN(irq < 0 || irq >= nr_irqs, "Invalid irq %d!\n", irq)))
 210                 return 0;
 211
 212         return info_for_irq(irq)->evtchn;
 213 }
 214
 215 unsigned irq_from_evtchn(unsigned int evtchn)
 216 {
 217         return evtchn_to_irq[evtchn];
 218 }
 219 EXPORT_SYMBOL_GPL(irq_from_evtchn);
 220
 221 static enum ipi_vector ipi_from_irq(unsigned irq)
 222 {
 223         struct irq_info *info = info_for_irq(irq);
 224
 225         BUG_ON(info == NULL);
 226         BUG_ON(info->type != IRQT_IPI);
 227
 228         return info->u.ipi;
 229 }
 230
 231 static unsigned virq_from_irq(unsigned irq)
 232 {
 233         struct irq_info *info = info_for_irq(irq);
 234
 235         BUG_ON(info == NULL);
 236         BUG_ON(info->type != IRQT_VIRQ);
 237
 238         return info->u.virq;
 239 }
 240
 241 static unsigned pirq_from_irq(unsigned irq)
 242 {
 243         struct irq_info *info = info_for_irq(irq);
 244
 245         BUG_ON(info == NULL);
 246         BUG_ON(info->type != IRQT_PIRQ);
 247
 248         return info->u.pirq.pirq;
 249 }
 250
 251 static enum xen_irq_type type_from_irq(unsigned irq)
 252 {
 253         return info_for_irq(irq)->type;
 254 }
 255
 256 static unsigned cpu_from_irq(unsigned irq)
 257 {
 258         return info_for_irq(irq)->cpu;
 259 }
 260
 261 static unsigned int cpu_from_evtchn(unsigned int evtchn)
 262 {
 263         int irq = evtchn_to_irq[evtchn];
 264         unsigned ret = 0;
 265
 266         if (irq != -1)
 267                 ret = cpu_from_irq(irq);
 268
 269         return ret;
 270 }
 271
 272 static bool pirq_needs_eoi(unsigned irq)
 273 {
 274         struct irq_info *info = info_for_irq(irq);
 275
 276         BUG_ON(info->type != IRQT_PIRQ);
 277
 278         return info->u.pirq.flags & PIRQ_NEEDS_EOI;
 279 }
 280
 281 static inline unsigned long active_evtchns(unsigned int cpu,
 282                                            struct shared_info *sh,
 283                                            unsigned int idx)
 284 {
 285         return (sh->evtchn_pending[idx] &
 286                 per_cpu(cpu_evtchn_mask, cpu)[idx] &
 287                 ~sh->evtchn_mask[idx]);
 288 }
 289
 290 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
 291 {
 292         int irq = evtchn_to_irq[chn];
 293
 294         BUG_ON(irq == -1);
 295 #ifdef CONFIG_SMP
 296         cpumask_copy(irq_to_desc(irq)->irq_data.affinity, cpumask_of(cpu));
 297 #endif
 298
 299         clear_bit(chn, per_cpu(cpu_evtchn_mask, cpu_from_irq(irq)));
 300         set_bit(chn, per_cpu(cpu_evtchn_mask, cpu));
 301
 302         info_for_irq(irq)->cpu = cpu;
 303 }
 304
 305 static void init_evtchn_cpu_bindings(void)
 306 {
 307         int i;
 308 #ifdef CONFIG_SMP
 309         struct irq_info *info;
 310
 311         /* By default all event channels notify CPU#0. */
 312         list_for_each_entry(info, &xen_irq_list_head, list) {
 313                 struct irq_desc *desc = irq_to_desc(info->irq);
 314                 cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 315         }
 316 #endif
 317
 318         for_each_possible_cpu(i)
 319                 memset(per_cpu(cpu_evtchn_mask, i),
 320                        (i == 0) ? ~0 : 0, sizeof(*per_cpu(cpu_evtchn_mask, i)));
 321 }
 322
 323 static inline void clear_evtchn(int port)
 324 {
 325         struct shared_info *s = HYPERVISOR_shared_info;
 326         sync_clear_bit(port, &s->evtchn_pending[0]);
 327 }
 328
 329 static inline void set_evtchn(int port)
 330 {
 331         struct shared_info *s = HYPERVISOR_shared_info;
 332         sync_set_bit(port, &s->evtchn_pending[0]);
 333 }
 334
 335 static inline int test_evtchn(int port)
 336 {
 337         struct shared_info *s = HYPERVISOR_shared_info;
 338         return sync_test_bit(port, &s->evtchn_pending[0]);
 339 }
 340
 341
 342 /**
 343  * notify_remote_via_irq - send event to remote end of event channel via irq
 344  * @irq: irq of event channel to send event to
 345  *
 346  * Unlike notify_remote_via_evtchn(), this is safe to use across
 347  * save/restore. Notifications on a broken connection are silently
 348  * dropped.
 349  */
 350 void notify_remote_via_irq(int irq)
 351 {
 352         int evtchn = evtchn_from_irq(irq);
 353
 354         if (VALID_EVTCHN(evtchn))
 355                 notify_remote_via_evtchn(evtchn);
 356 }
 357 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
 358
 359 static void mask_evtchn(int port)
 360 {
 361         struct shared_info *s = HYPERVISOR_shared_info;
 362         sync_set_bit(port, &s->evtchn_mask[0]);
 363 }
 364
 365 static void unmask_evtchn(int port)
 366 {
 367         struct shared_info *s = HYPERVISOR_shared_info;
 368         unsigned int cpu = get_cpu();
 369
 370         BUG_ON(!irqs_disabled());
 371
 372         /* Slow path (hypercall) if this is a non-local port. */
 373         if (unlikely(cpu != cpu_from_evtchn(port))) {
 374                 struct evtchn_unmask unmask = { .port = port };
 375                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 376         } else {
 377                 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
 378
 379                 sync_clear_bit(port, &s->evtchn_mask[0]);
 380
 381                 /*
 382                  * The following is basically the equivalent of
 383                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
 384                  * the interrupt edge' if the channel is masked.
 385                  */
 386                 if (sync_test_bit(port, &s->evtchn_pending[0]) &&
 387                     !sync_test_and_set_bit(port / BITS_PER_LONG,
 388                                            &vcpu_info->evtchn_pending_sel))
 389                         vcpu_info->evtchn_upcall_pending = 1;
 390         }
 391
 392         put_cpu();
 393 }
 394
 395 static void xen_irq_init(unsigned irq)
 396 {
 397         struct irq_info *info;
 398 #ifdef CONFIG_SMP
 399         struct irq_desc *desc = irq_to_desc(irq);
 400
 401         /* By default all event channels notify CPU#0. */
 402         cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 403 #endif
 404
 405         info = kzalloc(sizeof(*info), GFP_KERNEL);
 406         if (info == NULL)
 407                 panic("Unable to allocate metadata for IRQ%d\n", irq);
 408
 409         info->type = IRQT_UNBOUND;
 410
 411         irq_set_handler_data(irq, info);
 412
 413         list_add_tail(&info->list, &xen_irq_list_head);
 414 }
 415
 416 static int __must_check xen_allocate_irq_dynamic(void)
 417 {
 418         int first = 0;
 419         int irq;
 420
 421 #ifdef CONFIG_X86_IO_APIC
 422         /*
 423          * For an HVM guest or domain 0 which see "real" (emulated or
 424          * actual respectively) GSIs we allocate dynamic IRQs
 425          * e.g. those corresponding to event channels or MSIs
 426          * etc. from the range above those "real" GSIs to avoid
 427          * collisions.
 428          */
 429         if (xen_initial_domain() || xen_hvm_domain())
 430                 first = get_nr_irqs_gsi();
 431 #endif
 432
 433         irq = irq_alloc_desc_from(first, -1);
 434
 435         if (irq >= 0)
 436                 xen_irq_init(irq);
 437
 438         return irq;
 439 }
 440
 441 static int __must_check xen_allocate_irq_gsi(unsigned gsi)
 442 {
 443         int irq;
 444
 445         /*
 446          * A PV guest has no concept of a GSI (since it has no ACPI
 447          * nor access to/knowledge of the physical APICs). Therefore
 448          * all IRQs are dynamically allocated from the entire IRQ
 449          * space.
 450          */
 451         if (xen_pv_domain() && !xen_initial_domain())
 452                 return xen_allocate_irq_dynamic();
 453
 454         /* Legacy IRQ descriptors are already allocated by the arch. */
 455         if (gsi < NR_IRQS_LEGACY)
 456                 irq = gsi;
 457         else
 458                 irq = irq_alloc_desc_at(gsi, -1);
 459
 460         xen_irq_init(irq);
 461
 462         return irq;
 463 }
 464
 465 static void xen_free_irq(unsigned irq)
 466 {
 467         struct irq_info *info = irq_get_handler_data(irq);
 468
 469         list_del(&info->list);
 470
 471         irq_set_handler_data(irq, NULL);
 472
 473         kfree(info);
 474
 475         /* Legacy IRQ descriptors are managed by the arch. */
 476         if (irq < NR_IRQS_LEGACY)
 477                 return;
 478
 479         irq_free_desc(irq);
 480 }
 481
 482 static void pirq_query_unmask(int irq)
 483 {
 484         struct physdev_irq_status_query irq_status;
 485         struct irq_info *info = info_for_irq(irq);
 486
 487         BUG_ON(info->type != IRQT_PIRQ);
 488
 489         irq_status.irq = pirq_from_irq(irq);
 490         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
 491                 irq_status.flags = 0;
 492
 493         info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
 494         if (irq_status.flags & XENIRQSTAT_needs_eoi)
 495                 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
 496 }
 497
 498 static bool probing_irq(int irq)
 499 {
 500         struct irq_desc *desc = irq_to_desc(irq);
 501
 502         return desc && desc->action == NULL;
 503 }
 504
 505 static void eoi_pirq(struct irq_data *data)
 506 {
 507         int evtchn = evtchn_from_irq(data->irq);
 508         struct physdev_eoi eoi = { .irq = pirq_from_irq(data->irq) };
 509         int rc = 0;
 510
 511         irq_move_irq(data);
 512
 513         if (VALID_EVTCHN(evtchn))
 514                 clear_evtchn(evtchn);
 515
 516         if (pirq_needs_eoi(data->irq)) {
 517                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
 518                 WARN_ON(rc);
 519         }
 520 }
 521
 522 static void mask_ack_pirq(struct irq_data *data)
 523 {
 524         disable_dynirq(data);
 525         eoi_pirq(data);
 526 }
 527
 528 static unsigned int __startup_pirq(unsigned int irq)
 529 {
 530         struct evtchn_bind_pirq bind_pirq;
 531         struct irq_info *info = info_for_irq(irq);
 532         int evtchn = evtchn_from_irq(irq);
 533         int rc;
 534
 535         BUG_ON(info->type != IRQT_PIRQ);
 536
 537         if (VALID_EVTCHN(evtchn))
 538                 goto out;
 539
 540         bind_pirq.pirq = pirq_from_irq(irq);
 541         /* NB. We are happy to share unless we are probing. */
 542         bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
 543                                         BIND_PIRQ__WILL_SHARE : 0;
 544         rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
 545         if (rc != 0) {
 546                 if (!probing_irq(irq))
 547                         printk(KERN_INFO "Failed to obtain physical IRQ %d\n",
 548                                irq);
 549                 return 0;
 550         }
 551         evtchn = bind_pirq.port;
 552
 553         pirq_query_unmask(irq);
 554
 555         evtchn_to_irq[evtchn] = irq;
 556         bind_evtchn_to_cpu(evtchn, 0);
 557         info->evtchn = evtchn;
 558
 559 out:
 560         unmask_evtchn(evtchn);
 561         eoi_pirq(irq_get_irq_data(irq));
 562
 563         return 0;
 564 }
 565
 566 static unsigned int startup_pirq(struct irq_data *data)
 567 {
 568         return __startup_pirq(data->irq);
 569 }
 570
 571 static void shutdown_pirq(struct irq_data *data)
 572 {
 573         struct evtchn_close close;
 574         unsigned int irq = data->irq;
 575         struct irq_info *info = info_for_irq(irq);
 576         int evtchn = evtchn_from_irq(irq);
 577
 578         BUG_ON(info->type != IRQT_PIRQ);
 579
 580         if (!VALID_EVTCHN(evtchn))
 581                 return;
 582
 583         mask_evtchn(evtchn);
 584
 585         close.port = evtchn;
 586         if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 587                 BUG();
 588
 589         bind_evtchn_to_cpu(evtchn, 0);
 590         evtchn_to_irq[evtchn] = -1;
 591         info->evtchn = 0;
 592 }
 593
 594 static void enable_pirq(struct irq_data *data)
 595 {
 596         startup_pirq(data);
 597 }
 598
 599 static void disable_pirq(struct irq_data *data)
 600 {
 601         disable_dynirq(data);
 602 }
 603
 604 static int find_irq_by_gsi(unsigned gsi)
 605 {
 606         struct irq_info *info;
 607
 608         list_for_each_entry(info, &xen_irq_list_head, list) {
 609                 if (info->type != IRQT_PIRQ)
 610                         continue;
 611
 612                 if (info->u.pirq.gsi == gsi)
 613                         return info->irq;
 614         }
 615
 616         return -1;
 617 }
 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 = find_irq_by_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)
 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                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 913                                               handle_percpu_irq, "virq");
 914
 915                 bind_virq.virq = virq;
 916                 bind_virq.vcpu = cpu;
 917                 ret = HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 918                                                 &bind_virq);
 919                 if (ret == 0)
 920                         evtchn = bind_virq.port;
 921                 else {
 922                         if (ret == -EEXIST)
 923                                 ret = find_virq(virq, cpu);
 924                         BUG_ON(ret < 0);
 925                         evtchn = ret;
 926                 }
 927
 928                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
 929
 930                 bind_evtchn_to_cpu(evtchn, cpu);
 931         }
 932
 933 out:
 934         mutex_unlock(&irq_mapping_update_lock);
 935
 936         return irq;
 937 }
 938
 939 static void unbind_from_irq(unsigned int irq)
 940 {
 941         struct evtchn_close close;
 942         int evtchn = evtchn_from_irq(irq);
 943
 944         mutex_lock(&irq_mapping_update_lock);
 945
 946         if (VALID_EVTCHN(evtchn)) {
 947                 close.port = evtchn;
 948                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 949                         BUG();
 950
 951                 switch (type_from_irq(irq)) {
 952                 case IRQT_VIRQ:
 953                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
 954                                 [virq_from_irq(irq)] = -1;
 955                         break;
 956                 case IRQT_IPI:
 957                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
 958                                 [ipi_from_irq(irq)] = -1;
 959                         break;
 960                 default:
 961                         break;
 962                 }
 963
 964                 /* Closed ports are implicitly re-bound to VCPU0. */
 965                 bind_evtchn_to_cpu(evtchn, 0);
 966
 967                 evtchn_to_irq[evtchn] = -1;
 968         }
 969
 970         BUG_ON(info_for_irq(irq)->type == IRQT_UNBOUND);
 971
 972         xen_free_irq(irq);
 973
 974         mutex_unlock(&irq_mapping_update_lock);
 975 }
 976
 977 int bind_evtchn_to_irqhandler(unsigned int evtchn,
 978                               irq_handler_t handler,
 979                               unsigned long irqflags,
 980                               const char *devname, void *dev_id)
 981 {
 982         int irq, retval;
 983
 984         irq = bind_evtchn_to_irq(evtchn);
 985         if (irq < 0)
 986                 return irq;
 987         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 988         if (retval != 0) {
 989                 unbind_from_irq(irq);
 990                 return retval;
 991         }
 992
 993         return irq;
 994 }
 995 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
 996
 997 int bind_interdomain_evtchn_to_irqhandler(unsigned int remote_domain,
 998                                           unsigned int remote_port,
 999                                           irq_handler_t handler,
1000                                           unsigned long irqflags,
1001                                           const char *devname,
1002                                           void *dev_id)
1003 {
1004         int irq, retval;
1005
1006         irq = bind_interdomain_evtchn_to_irq(remote_domain, remote_port);
1007         if (irq < 0)
1008                 return irq;
1009
1010         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1011         if (retval != 0) {
1012                 unbind_from_irq(irq);
1013                 return retval;
1014         }
1015
1016         return irq;
1017 }
1018 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler);
1019
1020 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
1021                             irq_handler_t handler,
1022                             unsigned long irqflags, const char *devname, void *dev_id)
1023 {
1024         int irq, retval;
1025
1026         irq = bind_virq_to_irq(virq, cpu);
1027         if (irq < 0)
1028                 return irq;
1029         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1030         if (retval != 0) {
1031                 unbind_from_irq(irq);
1032                 return retval;
1033         }
1034
1035         return irq;
1036 }
1037 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1038
1039 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1040                            unsigned int cpu,
1041                            irq_handler_t handler,
1042                            unsigned long irqflags,
1043                            const char *devname,
1044                            void *dev_id)
1045 {
1046         int irq, retval;
1047
1048         irq = bind_ipi_to_irq(ipi, cpu);
1049         if (irq < 0)
1050                 return irq;
1051
1052         irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME | IRQF_EARLY_RESUME;
1053         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1054         if (retval != 0) {
1055                 unbind_from_irq(irq);
1056                 return retval;
1057         }
1058
1059         return irq;
1060 }
1061
1062 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1063 {
1064         free_irq(irq, dev_id);
1065         unbind_from_irq(irq);
1066 }
1067 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1068
1069 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1070 {
1071         int irq = per_cpu(ipi_to_irq, cpu)[vector];
1072         BUG_ON(irq < 0);
1073         notify_remote_via_irq(irq);
1074 }
1075
1076 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
1077 {
1078         struct shared_info *sh = HYPERVISOR_shared_info;
1079         int cpu = smp_processor_id();
1080         unsigned long *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu);
1081         int i;
1082         unsigned long flags;
1083         static DEFINE_SPINLOCK(debug_lock);
1084         struct vcpu_info *v;
1085
1086         spin_lock_irqsave(&debug_lock, flags);
1087
1088         printk("\nvcpu %d\n  ", cpu);
1089
1090         for_each_online_cpu(i) {
1091                 int pending;
1092                 v = per_cpu(xen_vcpu, i);
1093                 pending = (get_irq_regs() && i == cpu)
1094                         ? xen_irqs_disabled(get_irq_regs())
1095                         : v->evtchn_upcall_mask;
1096                 printk("%d: masked=%d pending=%d event_sel %0*lx\n  ", i,
1097                        pending, v->evtchn_upcall_pending,
1098                        (int)(sizeof(v->evtchn_pending_sel)*2),
1099                        v->evtchn_pending_sel);
1100         }
1101         v = per_cpu(xen_vcpu, cpu);
1102
1103         printk("\npending:\n   ");
1104         for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
1105                 printk("%0*lx%s", (int)sizeof(sh->evtchn_pending[0])*2,
1106                        sh->evtchn_pending[i],
1107                        i % 8 == 0 ? "\n   " : " ");
1108         printk("\nglobal mask:\n   ");
1109         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1110                 printk("%0*lx%s",
1111                        (int)(sizeof(sh->evtchn_mask[0])*2),
1112                        sh->evtchn_mask[i],
1113                        i % 8 == 0 ? "\n   " : " ");
1114
1115         printk("\nglobally unmasked:\n   ");
1116         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1117                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1118                        sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
1119                        i % 8 == 0 ? "\n   " : " ");
1120
1121         printk("\nlocal cpu%d mask:\n   ", cpu);
1122         for (i = (NR_EVENT_CHANNELS/BITS_PER_LONG)-1; i >= 0; i--)
1123                 printk("%0*lx%s", (int)(sizeof(cpu_evtchn[0])*2),
1124                        cpu_evtchn[i],
1125                        i % 8 == 0 ? "\n   " : " ");
1126
1127         printk("\nlocally unmasked:\n   ");
1128         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
1129                 unsigned long pending = sh->evtchn_pending[i]
1130                         & ~sh->evtchn_mask[i]
1131                         & cpu_evtchn[i];
1132                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1133                        pending, i % 8 == 0 ? "\n   " : " ");
1134         }
1135
1136         printk("\npending list:\n");
1137         for (i = 0; i < NR_EVENT_CHANNELS; i++) {
1138                 if (sync_test_bit(i, sh->evtchn_pending)) {
1139                         int word_idx = i / BITS_PER_LONG;
1140                         printk("  %d: event %d -> irq %d%s%s%s\n",
1141                                cpu_from_evtchn(i), i,
1142                                evtchn_to_irq[i],
1143                                sync_test_bit(word_idx, &v->evtchn_pending_sel)
1144                                              ? "" : " l2-clear",
1145                                !sync_test_bit(i, sh->evtchn_mask)
1146                                              ? "" : " globally-masked",
1147                                sync_test_bit(i, cpu_evtchn)
1148                                              ? "" : " locally-masked");
1149                 }
1150         }
1151
1152         spin_unlock_irqrestore(&debug_lock, flags);
1153
1154         return IRQ_HANDLED;
1155 }
1156
1157 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
1158 static DEFINE_PER_CPU(unsigned int, current_word_idx);
1159 static DEFINE_PER_CPU(unsigned int, current_bit_idx);
1160
1161 /*
1162  * Mask out the i least significant bits of w
1163  */
1164 #define MASK_LSBS(w, i) (w & ((~0UL) << i))
1165
1166 /*
1167  * Search the CPUs pending events bitmasks.  For each one found, map
1168  * the event number to an irq, and feed it into do_IRQ() for
1169  * handling.
1170  *
1171  * Xen uses a two-level bitmap to speed searching.  The first level is
1172  * a bitset of words which contain pending event bits.  The second
1173  * level is a bitset of pending events themselves.
1174  */
1175 static void __xen_evtchn_do_upcall(void)
1176 {
1177         int start_word_idx, start_bit_idx;
1178         int word_idx, bit_idx;
1179         int i;
1180         int cpu = get_cpu();
1181         struct shared_info *s = HYPERVISOR_shared_info;
1182         struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1183         unsigned count;
1184
1185         do {
1186                 unsigned long pending_words;
1187
1188                 vcpu_info->evtchn_upcall_pending = 0;
1189
1190                 if (__this_cpu_inc_return(xed_nesting_count) - 1)
1191                         goto out;
1192
1193 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
1194                 /* Clear master flag /before/ clearing selector flag. */
1195                 wmb();
1196 #endif
1197                 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
1198
1199                 start_word_idx = __this_cpu_read(current_word_idx);
1200                 start_bit_idx = __this_cpu_read(current_bit_idx);
1201
1202                 word_idx = start_word_idx;
1203
1204                 for (i = 0; pending_words != 0; i++) {
1205                         unsigned long pending_bits;
1206                         unsigned long words;
1207
1208                         words = MASK_LSBS(pending_words, word_idx);
1209
1210                         /*
1211                          * If we masked out all events, wrap to beginning.
1212                          */
1213                         if (words == 0) {
1214                                 word_idx = 0;
1215                                 bit_idx = 0;
1216                                 continue;
1217                         }
1218                         word_idx = __ffs(words);
1219
1220                         pending_bits = active_evtchns(cpu, s, word_idx);
1221                         bit_idx = 0; /* usually scan entire word from start */
1222                         if (word_idx == start_word_idx) {
1223                                 /* We scan the starting word in two parts */
1224                                 if (i == 0)
1225                                         /* 1st time: start in the middle */
1226                                         bit_idx = start_bit_idx;
1227                                 else
1228                                         /* 2nd time: mask bits done already */
1229                                         bit_idx &= (1UL << start_bit_idx) - 1;
1230                         }
1231
1232                         do {
1233                                 unsigned long bits;
1234                                 int port, irq;
1235                                 struct irq_desc *desc;
1236
1237                                 bits = MASK_LSBS(pending_bits, bit_idx);
1238
1239                                 /* If we masked out all events, move on. */
1240                                 if (bits == 0)
1241                                         break;
1242
1243                                 bit_idx = __ffs(bits);
1244
1245                                 /* Process port. */
1246                                 port = (word_idx * BITS_PER_LONG) + bit_idx;
1247                                 irq = evtchn_to_irq[port];
1248
1249                                 if (irq != -1) {
1250                                         desc = irq_to_desc(irq);
1251                                         if (desc)
1252                                                 generic_handle_irq_desc(irq, desc);
1253                                 }
1254
1255                                 bit_idx = (bit_idx + 1) % BITS_PER_LONG;
1256
1257                                 /* Next caller starts at last processed + 1 */
1258                                 __this_cpu_write(current_word_idx,
1259                                                  bit_idx ? word_idx :
1260                                                  (word_idx+1) % BITS_PER_LONG);
1261                                 __this_cpu_write(current_bit_idx, bit_idx);
1262                         } while (bit_idx != 0);
1263
1264                         /* Scan start_l1i twice; all others once. */
1265                         if ((word_idx != start_word_idx) || (i != 0))
1266                                 pending_words &= ~(1UL << word_idx);
1267
1268                         word_idx = (word_idx + 1) % BITS_PER_LONG;
1269                 }
1270
1271                 BUG_ON(!irqs_disabled());
1272
1273                 count = __this_cpu_read(xed_nesting_count);
1274                 __this_cpu_write(xed_nesting_count, 0);
1275         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
1276
1277 out:
1278
1279         put_cpu();
1280 }
1281
1282 void xen_evtchn_do_upcall(struct pt_regs *regs)
1283 {
1284         struct pt_regs *old_regs = set_irq_regs(regs);
1285
1286         exit_idle();
1287         irq_enter();
1288
1289         __xen_evtchn_do_upcall();
1290
1291         irq_exit();
1292         set_irq_regs(old_regs);
1293 }
1294
1295 void xen_hvm_evtchn_do_upcall(void)
1296 {
1297         __xen_evtchn_do_upcall();
1298 }
1299 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1300
1301 /* Rebind a new event channel to an existing irq. */
1302 void rebind_evtchn_irq(int evtchn, int irq)
1303 {
1304         struct irq_info *info = info_for_irq(irq);
1305
1306         /* Make sure the irq is masked, since the new event channel
1307            will also be masked. */
1308         disable_irq(irq);
1309
1310         mutex_lock(&irq_mapping_update_lock);
1311
1312         /* After resume the irq<->evtchn mappings are all cleared out */
1313         BUG_ON(evtchn_to_irq[evtchn] != -1);
1314         /* Expect irq to have been bound before,
1315            so there should be a proper type */
1316         BUG_ON(info->type == IRQT_UNBOUND);
1317
1318         xen_irq_info_evtchn_init(irq, evtchn);
1319
1320         mutex_unlock(&irq_mapping_update_lock);
1321
1322         /* new event channels are always bound to cpu 0 */
1323         irq_set_affinity(irq, cpumask_of(0));
1324
1325         /* Unmask the event channel. */
1326         enable_irq(irq);
1327 }
1328
1329 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1330 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1331 {
1332         struct evtchn_bind_vcpu bind_vcpu;
1333         int evtchn = evtchn_from_irq(irq);
1334
1335         if (!VALID_EVTCHN(evtchn))
1336                 return -1;
1337
1338         /*
1339          * Events delivered via platform PCI interrupts are always
1340          * routed to vcpu 0 and hence cannot be rebound.
1341          */
1342         if (xen_hvm_domain() && !xen_have_vector_callback)
1343                 return -1;
1344
1345         /* Send future instances of this interrupt to other vcpu. */
1346         bind_vcpu.port = evtchn;
1347         bind_vcpu.vcpu = tcpu;
1348
1349         /*
1350          * If this fails, it usually just indicates that we're dealing with a
1351          * virq or IPI channel, which don't actually need to be rebound. Ignore
1352          * it, but don't do the xenlinux-level rebind in that case.
1353          */
1354         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1355                 bind_evtchn_to_cpu(evtchn, tcpu);
1356
1357         return 0;
1358 }
1359
1360 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1361                             bool force)
1362 {
1363         unsigned tcpu = cpumask_first(dest);
1364
1365         return rebind_irq_to_cpu(data->irq, tcpu);
1366 }
1367
1368 int resend_irq_on_evtchn(unsigned int irq)
1369 {
1370         int masked, evtchn = evtchn_from_irq(irq);
1371         struct shared_info *s = HYPERVISOR_shared_info;
1372
1373         if (!VALID_EVTCHN(evtchn))
1374                 return 1;
1375
1376         masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
1377         sync_set_bit(evtchn, s->evtchn_pending);
1378         if (!masked)
1379                 unmask_evtchn(evtchn);
1380
1381         return 1;
1382 }
1383
1384 static void enable_dynirq(struct irq_data *data)
1385 {
1386         int evtchn = evtchn_from_irq(data->irq);
1387
1388         if (VALID_EVTCHN(evtchn))
1389                 unmask_evtchn(evtchn);
1390 }
1391
1392 static void disable_dynirq(struct irq_data *data)
1393 {
1394         int evtchn = evtchn_from_irq(data->irq);
1395
1396         if (VALID_EVTCHN(evtchn))
1397                 mask_evtchn(evtchn);
1398 }
1399
1400 static void ack_dynirq(struct irq_data *data)
1401 {
1402         int evtchn = evtchn_from_irq(data->irq);
1403
1404         irq_move_irq(data);
1405
1406         if (VALID_EVTCHN(evtchn))
1407                 clear_evtchn(evtchn);
1408 }
1409
1410 static void mask_ack_dynirq(struct irq_data *data)
1411 {
1412         disable_dynirq(data);
1413         ack_dynirq(data);
1414 }
1415
1416 static int retrigger_dynirq(struct irq_data *data)
1417 {
1418         int evtchn = evtchn_from_irq(data->irq);
1419         struct shared_info *sh = HYPERVISOR_shared_info;
1420         int ret = 0;
1421
1422         if (VALID_EVTCHN(evtchn)) {
1423                 int masked;
1424
1425                 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
1426                 sync_set_bit(evtchn, sh->evtchn_pending);
1427                 if (!masked)
1428                         unmask_evtchn(evtchn);
1429                 ret = 1;
1430         }
1431
1432         return ret;
1433 }
1434
1435 static void restore_pirqs(void)
1436 {
1437         int pirq, rc, irq, gsi;
1438         struct physdev_map_pirq map_irq;
1439         struct irq_info *info;
1440
1441         list_for_each_entry(info, &xen_irq_list_head, list) {
1442                 if (info->type != IRQT_PIRQ)
1443                         continue;
1444
1445                 pirq = info->u.pirq.pirq;
1446                 gsi = info->u.pirq.gsi;
1447                 irq = info->irq;
1448
1449                 /* save/restore of PT devices doesn't work, so at this point the
1450                  * only devices present are GSI based emulated devices */
1451                 if (!gsi)
1452                         continue;
1453
1454                 map_irq.domid = DOMID_SELF;
1455                 map_irq.type = MAP_PIRQ_TYPE_GSI;
1456                 map_irq.index = gsi;
1457                 map_irq.pirq = pirq;
1458
1459                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1460                 if (rc) {
1461                         printk(KERN_WARNING "xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1462                                         gsi, irq, pirq, rc);
1463                         xen_free_irq(irq);
1464                         continue;
1465                 }
1466
1467                 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1468
1469                 __startup_pirq(irq);
1470         }
1471 }
1472
1473 static void restore_cpu_virqs(unsigned int cpu)
1474 {
1475         struct evtchn_bind_virq bind_virq;
1476         int virq, irq, evtchn;
1477
1478         for (virq = 0; virq < NR_VIRQS; virq++) {
1479                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1480                         continue;
1481
1482                 BUG_ON(virq_from_irq(irq) != virq);
1483
1484                 /* Get a new binding from Xen. */
1485                 bind_virq.virq = virq;
1486                 bind_virq.vcpu = cpu;
1487                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1488                                                 &bind_virq) != 0)
1489                         BUG();
1490                 evtchn = bind_virq.port;
1491
1492                 /* Record the new mapping. */
1493                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
1494                 bind_evtchn_to_cpu(evtchn, cpu);
1495         }
1496 }
1497
1498 static void restore_cpu_ipis(unsigned int cpu)
1499 {
1500         struct evtchn_bind_ipi bind_ipi;
1501         int ipi, irq, evtchn;
1502
1503         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1504                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1505                         continue;
1506
1507                 BUG_ON(ipi_from_irq(irq) != ipi);
1508
1509                 /* Get a new binding from Xen. */
1510                 bind_ipi.vcpu = cpu;
1511                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1512                                                 &bind_ipi) != 0)
1513                         BUG();
1514                 evtchn = bind_ipi.port;
1515
1516                 /* Record the new mapping. */
1517                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
1518                 bind_evtchn_to_cpu(evtchn, cpu);
1519         }
1520 }
1521
1522 /* Clear an irq's pending state, in preparation for polling on it */
1523 void xen_clear_irq_pending(int irq)
1524 {
1525         int evtchn = evtchn_from_irq(irq);
1526
1527         if (VALID_EVTCHN(evtchn))
1528                 clear_evtchn(evtchn);
1529 }
1530 EXPORT_SYMBOL(xen_clear_irq_pending);
1531 void xen_set_irq_pending(int irq)
1532 {
1533         int evtchn = evtchn_from_irq(irq);
1534
1535         if (VALID_EVTCHN(evtchn))
1536                 set_evtchn(evtchn);
1537 }
1538
1539 bool xen_test_irq_pending(int irq)
1540 {
1541         int evtchn = evtchn_from_irq(irq);
1542         bool ret = false;
1543
1544         if (VALID_EVTCHN(evtchn))
1545                 ret = test_evtchn(evtchn);
1546
1547         return ret;
1548 }
1549
1550 /* Poll waiting for an irq to become pending with timeout.  In the usual case,
1551  * the irq will be disabled so it won't deliver an interrupt. */
1552 void xen_poll_irq_timeout(int irq, u64 timeout)
1553 {
1554         evtchn_port_t evtchn = evtchn_from_irq(irq);
1555
1556         if (VALID_EVTCHN(evtchn)) {
1557                 struct sched_poll poll;
1558
1559                 poll.nr_ports = 1;
1560                 poll.timeout = timeout;
1561                 set_xen_guest_handle(poll.ports, &evtchn);
1562
1563                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1564                         BUG();
1565         }
1566 }
1567 EXPORT_SYMBOL(xen_poll_irq_timeout);
1568 /* Poll waiting for an irq to become pending.  In the usual case, the
1569  * irq will be disabled so it won't deliver an interrupt. */
1570 void xen_poll_irq(int irq)
1571 {
1572         xen_poll_irq_timeout(irq, 0 /* no timeout */);
1573 }
1574
1575 /* Check whether the IRQ line is shared with other guests. */
1576 int xen_test_irq_shared(int irq)
1577 {
1578         struct irq_info *info = info_for_irq(irq);
1579         struct physdev_irq_status_query irq_status = { .irq = info->u.pirq.pirq };
1580
1581         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
1582                 return 0;
1583         return !(irq_status.flags & XENIRQSTAT_shared);
1584 }
1585 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
1586
1587 void xen_irq_resume(void)
1588 {
1589         unsigned int cpu, evtchn;
1590         struct irq_info *info;
1591
1592         init_evtchn_cpu_bindings();
1593
1594         /* New event-channel space is not 'live' yet. */
1595         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1596                 mask_evtchn(evtchn);
1597
1598         /* No IRQ <-> event-channel mappings. */
1599         list_for_each_entry(info, &xen_irq_list_head, list)
1600                 info->evtchn = 0; /* zap event-channel binding */
1601
1602         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1603                 evtchn_to_irq[evtchn] = -1;
1604
1605         for_each_possible_cpu(cpu) {
1606                 restore_cpu_virqs(cpu);
1607                 restore_cpu_ipis(cpu);
1608         }
1609
1610         restore_pirqs();
1611 }
1612
1613 static struct irq_chip xen_dynamic_chip __read_mostly = {
1614         .name                   = "xen-dyn",
1615
1616         .irq_disable            = disable_dynirq,
1617         .irq_mask               = disable_dynirq,
1618         .irq_unmask             = enable_dynirq,
1619
1620         .irq_ack                = ack_dynirq,
1621         .irq_mask_ack           = mask_ack_dynirq,
1622
1623         .irq_set_affinity       = set_affinity_irq,
1624         .irq_retrigger          = retrigger_dynirq,
1625 };
1626
1627 static struct irq_chip xen_pirq_chip __read_mostly = {
1628         .name                   = "xen-pirq",
1629
1630         .irq_startup            = startup_pirq,
1631         .irq_shutdown           = shutdown_pirq,
1632         .irq_enable             = enable_pirq,
1633         .irq_disable            = disable_pirq,
1634
1635         .irq_mask               = disable_dynirq,
1636         .irq_unmask             = enable_dynirq,
1637
1638         .irq_ack                = eoi_pirq,
1639         .irq_eoi                = eoi_pirq,
1640         .irq_mask_ack           = mask_ack_pirq,
1641
1642         .irq_set_affinity       = set_affinity_irq,
1643
1644         .irq_retrigger          = retrigger_dynirq,
1645 };
1646
1647 static struct irq_chip xen_percpu_chip __read_mostly = {
1648         .name                   = "xen-percpu",
1649
1650         .irq_disable            = disable_dynirq,
1651         .irq_mask               = disable_dynirq,
1652         .irq_unmask             = enable_dynirq,
1653
1654         .irq_ack                = ack_dynirq,
1655 };
1656
1657 int xen_set_callback_via(uint64_t via)
1658 {
1659         struct xen_hvm_param a;
1660         a.domid = DOMID_SELF;
1661         a.index = HVM_PARAM_CALLBACK_IRQ;
1662         a.value = via;
1663         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1664 }
1665 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1666
1667 #ifdef CONFIG_XEN_PVHVM
1668 /* Vector callbacks are better than PCI interrupts to receive event
1669  * channel notifications because we can receive vector callbacks on any
1670  * vcpu and we don't need PCI support or APIC interactions. */
1671 void xen_callback_vector(void)
1672 {
1673         int rc;
1674         uint64_t callback_via;
1675         if (xen_have_vector_callback) {
1676                 callback_via = HVM_CALLBACK_VECTOR(XEN_HVM_EVTCHN_CALLBACK);
1677                 rc = xen_set_callback_via(callback_via);
1678                 if (rc) {
1679                         printk(KERN_ERR "Request for Xen HVM callback vector"
1680                                         " failed.\n");
1681                         xen_have_vector_callback = 0;
1682                         return;
1683                 }
1684                 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1685                                 "enabled\n");
1686                 /* in the restore case the vector has already been allocated */
1687                 if (!test_bit(XEN_HVM_EVTCHN_CALLBACK, used_vectors))
1688                         alloc_intr_gate(XEN_HVM_EVTCHN_CALLBACK, xen_hvm_callback_vector);
1689         }
1690 }
1691 #else
1692 void xen_callback_vector(void) {}
1693 #endif
1694
1695 void __init xen_init_IRQ(void)
1696 {
1697         int i;
1698
1699         evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1700                                     GFP_KERNEL);
1701         BUG_ON(!evtchn_to_irq);
1702         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1703                 evtchn_to_irq[i] = -1;
1704
1705         init_evtchn_cpu_bindings();
1706
1707         /* No event channels are 'live' right now. */
1708         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1709                 mask_evtchn(i);
1710
1711         if (xen_hvm_domain()) {
1712                 xen_callback_vector();
1713                 native_init_IRQ();
1714                 /* pci_xen_hvm_init must be called after native_init_IRQ so that
1715                  * __acpi_register_gsi can point at the right function */
1716                 pci_xen_hvm_init();
1717         } else {
1718                 irq_ctx_init(smp_processor_id());
1719                 if (xen_initial_domain())
1720                         pci_xen_initial_domain();
1721         }
1722 }