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, sizeof(*per_cpu(cpu_evtchn_mask, i)));
 320 }
 321
 322 static inline void clear_evtchn(int port)
 323 {
 324         struct shared_info *s = HYPERVISOR_shared_info;
 325         sync_clear_bit(port, &s->evtchn_pending[0]);
 326 }
 327
 328 static inline void set_evtchn(int port)
 329 {
 330         struct shared_info *s = HYPERVISOR_shared_info;
 331         sync_set_bit(port, &s->evtchn_pending[0]);
 332 }
 333
 334 static inline int test_evtchn(int port)
 335 {
 336         struct shared_info *s = HYPERVISOR_shared_info;
 337         return sync_test_bit(port, &s->evtchn_pending[0]);
 338 }
 339
 340
 341 /**
 342  * notify_remote_via_irq - send event to remote end of event channel via irq
 343  * @irq: irq of event channel to send event to
 344  *
 345  * Unlike notify_remote_via_evtchn(), this is safe to use across
 346  * save/restore. Notifications on a broken connection are silently
 347  * dropped.
 348  */
 349 void notify_remote_via_irq(int irq)
 350 {
 351         int evtchn = evtchn_from_irq(irq);
 352
 353         if (VALID_EVTCHN(evtchn))
 354                 notify_remote_via_evtchn(evtchn);
 355 }
 356 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
 357
 358 static void mask_evtchn(int port)
 359 {
 360         struct shared_info *s = HYPERVISOR_shared_info;
 361         sync_set_bit(port, &s->evtchn_mask[0]);
 362 }
 363
 364 static void unmask_evtchn(int port)
 365 {
 366         struct shared_info *s = HYPERVISOR_shared_info;
 367         unsigned int cpu = get_cpu();
 368
 369         BUG_ON(!irqs_disabled());
 370
 371         /* Slow path (hypercall) if this is a non-local port. */
 372         if (unlikely(cpu != cpu_from_evtchn(port))) {
 373                 struct evtchn_unmask unmask = { .port = port };
 374                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 375         } else {
 376                 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
 377
 378                 sync_clear_bit(port, &s->evtchn_mask[0]);
 379
 380                 /*
 381                  * The following is basically the equivalent of
 382                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
 383                  * the interrupt edge' if the channel is masked.
 384                  */
 385                 if (sync_test_bit(port, &s->evtchn_pending[0]) &&
 386                     !sync_test_and_set_bit(port / BITS_PER_LONG,
 387                                            &vcpu_info->evtchn_pending_sel))
 388                         vcpu_info->evtchn_upcall_pending = 1;
 389         }
 390
 391         put_cpu();
 392 }
 393
 394 static void xen_irq_init(unsigned irq)
 395 {
 396         struct irq_info *info;
 397 #ifdef CONFIG_SMP
 398         struct irq_desc *desc = irq_to_desc(irq);
 399
 400         /* By default all event channels notify CPU#0. */
 401         cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 402 #endif
 403
 404         info = kzalloc(sizeof(*info), GFP_KERNEL);
 405         if (info == NULL)
 406                 panic("Unable to allocate metadata for IRQ%d\n", irq);
 407
 408         info->type = IRQT_UNBOUND;
 409
 410         irq_set_handler_data(irq, info);
 411
 412         list_add_tail(&info->list, &xen_irq_list_head);
 413 }
 414
 415 static int __must_check xen_allocate_irq_dynamic(void)
 416 {
 417         int first = 0;
 418         int irq;
 419
 420 #ifdef CONFIG_X86_IO_APIC
 421         /*
 422          * For an HVM guest or domain 0 which see "real" (emulated or
 423          * actual respectively) GSIs we allocate dynamic IRQs
 424          * e.g. those corresponding to event channels or MSIs
 425          * etc. from the range above those "real" GSIs to avoid
 426          * collisions.
 427          */
 428         if (xen_initial_domain() || xen_hvm_domain())
 429                 first = get_nr_irqs_gsi();
 430 #endif
 431
 432         irq = irq_alloc_desc_from(first, -1);
 433
 434         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 static int find_irq_by_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
 618 /*
 619  * Do not make any assumptions regarding the relationship between the
 620  * IRQ number returned here and the Xen pirq argument.
 621  *
 622  * Note: We don't assign an event channel until the irq actually started
 623  * up.  Return an existing irq if we've already got one for the gsi.
 624  *
 625  * Shareable implies level triggered, not shareable implies edge
 626  * triggered here.
 627  */
 628 int xen_bind_pirq_gsi_to_irq(unsigned gsi,
 629                              unsigned pirq, int shareable, char *name)
 630 {
 631         int irq = -1;
 632         struct physdev_irq irq_op;
 633
 634         mutex_lock(&irq_mapping_update_lock);
 635
 636         irq = find_irq_by_gsi(gsi);
 637         if (irq != -1) {
 638                 printk(KERN_INFO "xen_map_pirq_gsi: returning irq %d for gsi %u\n",
 639                        irq, gsi);
 640                 goto out;       /* XXX need refcount? */
 641         }
 642
 643         irq = xen_allocate_irq_gsi(gsi);
 644         if (irq < 0)
 645                 goto out;
 646
 647         irq_op.irq = irq;
 648         irq_op.vector = 0;
 649
 650         /* Only the privileged domain can do this. For non-priv, the pcifront
 651          * driver provides a PCI bus that does the call to do exactly
 652          * this in the priv domain. */
 653         if (xen_initial_domain() &&
 654             HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
 655                 xen_free_irq(irq);
 656                 irq = -ENOSPC;
 657                 goto out;
 658         }
 659
 660         xen_irq_info_pirq_init(irq, 0, pirq, gsi, irq_op.vector, DOMID_SELF,
 661                                shareable ? PIRQ_SHAREABLE : 0);
 662
 663         pirq_query_unmask(irq);
 664         /* We try to use the handler with the appropriate semantic for the
 665          * type of interrupt: if the interrupt is an edge triggered
 666          * interrupt we use handle_edge_irq.
 667          *
 668          * On the other hand if the interrupt is level triggered we use
 669          * handle_fasteoi_irq like the native code does for this kind of
 670          * interrupts.
 671          *
 672          * Depending on the Xen version, pirq_needs_eoi might return true
 673          * not only for level triggered interrupts but for edge triggered
 674          * interrupts too. In any case Xen always honors the eoi mechanism,
 675          * not injecting any more pirqs of the same kind if the first one
 676          * hasn't received an eoi yet. Therefore using the fasteoi handler
 677          * is the right choice either way.
 678          */
 679         if (shareable)
 680                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 681                                 handle_fasteoi_irq, name);
 682         else
 683                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 684                                 handle_edge_irq, name);
 685
 686 out:
 687         mutex_unlock(&irq_mapping_update_lock);
 688
 689         return irq;
 690 }
 691
 692 #ifdef CONFIG_PCI_MSI
 693 int xen_allocate_pirq_msi(struct pci_dev *dev, struct msi_desc *msidesc)
 694 {
 695         int rc;
 696         struct physdev_get_free_pirq op_get_free_pirq;
 697
 698         op_get_free_pirq.type = MAP_PIRQ_TYPE_MSI;
 699         rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
 700
 701         WARN_ONCE(rc == -ENOSYS,
 702                   "hypervisor does not support the PHYSDEVOP_get_free_pirq interface\n");
 703
 704         return rc ? -1 : op_get_free_pirq.pirq;
 705 }
 706
 707 int xen_bind_pirq_msi_to_irq(struct pci_dev *dev, struct msi_desc *msidesc,
 708                              int pirq, int vector, const char *name,
 709                              domid_t domid)
 710 {
 711         int irq, ret;
 712
 713         mutex_lock(&irq_mapping_update_lock);
 714
 715         irq = xen_allocate_irq_dynamic();
 716         if (irq < 0)
 717                 goto out;
 718
 719         irq_set_chip_and_handler_name(irq, &xen_pirq_chip, handle_edge_irq,
 720                         name);
 721
 722         xen_irq_info_pirq_init(irq, 0, pirq, 0, vector, domid, 0);
 723         ret = irq_set_msi_desc(irq, msidesc);
 724         if (ret < 0)
 725                 goto error_irq;
 726 out:
 727         mutex_unlock(&irq_mapping_update_lock);
 728         return irq;
 729 error_irq:
 730         mutex_unlock(&irq_mapping_update_lock);
 731         xen_free_irq(irq);
 732         return ret;
 733 }
 734 #endif
 735
 736 int xen_destroy_irq(int irq)
 737 {
 738         struct irq_desc *desc;
 739         struct physdev_unmap_pirq unmap_irq;
 740         struct irq_info *info = info_for_irq(irq);
 741         int rc = -ENOENT;
 742
 743         mutex_lock(&irq_mapping_update_lock);
 744
 745         desc = irq_to_desc(irq);
 746         if (!desc)
 747                 goto out;
 748
 749         if (xen_initial_domain()) {
 750                 unmap_irq.pirq = info->u.pirq.pirq;
 751                 unmap_irq.domid = info->u.pirq.domid;
 752                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
 753                 /* If another domain quits without making the pci_disable_msix
 754                  * call, the Xen hypervisor takes care of freeing the PIRQs
 755                  * (free_domain_pirqs).
 756                  */
 757                 if ((rc == -ESRCH && info->u.pirq.domid != DOMID_SELF))
 758                         printk(KERN_INFO "domain %d does not have %d anymore\n",
 759                                 info->u.pirq.domid, info->u.pirq.pirq);
 760                 else if (rc) {
 761                         printk(KERN_WARNING "unmap irq failed %d\n", rc);
 762                         goto out;
 763                 }
 764         }
 765
 766         xen_free_irq(irq);
 767
 768 out:
 769         mutex_unlock(&irq_mapping_update_lock);
 770         return rc;
 771 }
 772
 773 int xen_irq_from_pirq(unsigned pirq)
 774 {
 775         int irq;
 776
 777         struct irq_info *info;
 778
 779         mutex_lock(&irq_mapping_update_lock);
 780
 781         list_for_each_entry(info, &xen_irq_list_head, list) {
 782                 if (info->type != IRQT_PIRQ)
 783                         continue;
 784                 irq = info->irq;
 785                 if (info->u.pirq.pirq == pirq)
 786                         goto out;
 787         }
 788         irq = -1;
 789 out:
 790         mutex_unlock(&irq_mapping_update_lock);
 791
 792         return irq;
 793 }
 794
 795
 796 int xen_pirq_from_irq(unsigned irq)
 797 {
 798         return pirq_from_irq(irq);
 799 }
 800 EXPORT_SYMBOL_GPL(xen_pirq_from_irq);
 801 int bind_evtchn_to_irq(unsigned int evtchn)
 802 {
 803         int irq;
 804
 805         mutex_lock(&irq_mapping_update_lock);
 806
 807         irq = evtchn_to_irq[evtchn];
 808
 809         if (irq == -1) {
 810                 irq = xen_allocate_irq_dynamic();
 811                 if (irq == -1)
 812                         goto out;
 813
 814                 irq_set_chip_and_handler_name(irq, &xen_dynamic_chip,
 815                                               handle_edge_irq, "event");
 816
 817                 xen_irq_info_evtchn_init(irq, evtchn);
 818         }
 819
 820 out:
 821         mutex_unlock(&irq_mapping_update_lock);
 822
 823         return irq;
 824 }
 825 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
 826
 827 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
 828 {
 829         struct evtchn_bind_ipi bind_ipi;
 830         int evtchn, irq;
 831
 832         mutex_lock(&irq_mapping_update_lock);
 833
 834         irq = per_cpu(ipi_to_irq, cpu)[ipi];
 835
 836         if (irq == -1) {
 837                 irq = xen_allocate_irq_dynamic();
 838                 if (irq < 0)
 839                         goto out;
 840
 841                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 842                                               handle_percpu_irq, "ipi");
 843
 844                 bind_ipi.vcpu = cpu;
 845                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
 846                                                 &bind_ipi) != 0)
 847                         BUG();
 848                 evtchn = bind_ipi.port;
 849
 850                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
 851
 852                 bind_evtchn_to_cpu(evtchn, cpu);
 853         }
 854
 855  out:
 856         mutex_unlock(&irq_mapping_update_lock);
 857         return irq;
 858 }
 859
 860 static int bind_interdomain_evtchn_to_irq(unsigned int remote_domain,
 861                                           unsigned int remote_port)
 862 {
 863         struct evtchn_bind_interdomain bind_interdomain;
 864         int err;
 865
 866         bind_interdomain.remote_dom  = remote_domain;
 867         bind_interdomain.remote_port = remote_port;
 868
 869         err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
 870                                           &bind_interdomain);
 871
 872         return err ? : bind_evtchn_to_irq(bind_interdomain.local_port);
 873 }
 874
 875 static int find_virq(unsigned int virq, unsigned int cpu)
 876 {
 877         struct evtchn_status status;
 878         int port, rc = -ENOENT;
 879
 880         memset(&status, 0, sizeof(status));
 881         for (port = 0; port <= NR_EVENT_CHANNELS; port++) {
 882                 status.dom = DOMID_SELF;
 883                 status.port = port;
 884                 rc = HYPERVISOR_event_channel_op(EVTCHNOP_status, &status);
 885                 if (rc < 0)
 886                         continue;
 887                 if (status.status != EVTCHNSTAT_virq)
 888                         continue;
 889                 if (status.u.virq == virq && status.vcpu == cpu) {
 890                         rc = port;
 891                         break;
 892                 }
 893         }
 894         return rc;
 895 }
 896
 897 int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
 898 {
 899         struct evtchn_bind_virq bind_virq;
 900         int evtchn, irq, ret;
 901
 902         mutex_lock(&irq_mapping_update_lock);
 903
 904         irq = per_cpu(virq_to_irq, cpu)[virq];
 905
 906         if (irq == -1) {
 907                 irq = xen_allocate_irq_dynamic();
 908                 if (irq == -1)
 909                         goto out;
 910
 911                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 912                                               handle_percpu_irq, "virq");
 913
 914                 bind_virq.virq = virq;
 915                 bind_virq.vcpu = cpu;
 916                 ret = HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 917                                                 &bind_virq);
 918                 if (ret == 0)
 919                         evtchn = bind_virq.port;
 920                 else {
 921                         if (ret == -EEXIST)
 922                                 ret = find_virq(virq, cpu);
 923                         BUG_ON(ret < 0);
 924                         evtchn = ret;
 925                 }
 926
 927                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
 928
 929                 bind_evtchn_to_cpu(evtchn, cpu);
 930         }
 931
 932 out:
 933         mutex_unlock(&irq_mapping_update_lock);
 934
 935         return irq;
 936 }
 937
 938 static void unbind_from_irq(unsigned int irq)
 939 {
 940         struct evtchn_close close;
 941         int evtchn = evtchn_from_irq(irq);
 942
 943         mutex_lock(&irq_mapping_update_lock);
 944
 945         if (VALID_EVTCHN(evtchn)) {
 946                 close.port = evtchn;
 947                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 948                         BUG();
 949
 950                 switch (type_from_irq(irq)) {
 951                 case IRQT_VIRQ:
 952                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
 953                                 [virq_from_irq(irq)] = -1;
 954                         break;
 955                 case IRQT_IPI:
 956                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
 957                                 [ipi_from_irq(irq)] = -1;
 958                         break;
 959                 default:
 960                         break;
 961                 }
 962
 963                 /* Closed ports are implicitly re-bound to VCPU0. */
 964                 bind_evtchn_to_cpu(evtchn, 0);
 965
 966                 evtchn_to_irq[evtchn] = -1;
 967         }
 968
 969         BUG_ON(info_for_irq(irq)->type == IRQT_UNBOUND);
 970
 971         xen_free_irq(irq);
 972
 973         mutex_unlock(&irq_mapping_update_lock);
 974 }
 975
 976 int bind_evtchn_to_irqhandler(unsigned int evtchn,
 977                               irq_handler_t handler,
 978                               unsigned long irqflags,
 979                               const char *devname, void *dev_id)
 980 {
 981         int irq, retval;
 982
 983         irq = bind_evtchn_to_irq(evtchn);
 984         if (irq < 0)
 985                 return irq;
 986         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 987         if (retval != 0) {
 988                 unbind_from_irq(irq);
 989                 return retval;
 990         }
 991
 992         return irq;
 993 }
 994 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
 995
 996 int bind_interdomain_evtchn_to_irqhandler(unsigned int remote_domain,
 997                                           unsigned int remote_port,
 998                                           irq_handler_t handler,
 999                                           unsigned long irqflags,
1000                                           const char *devname,
1001                                           void *dev_id)
1002 {
1003         int irq, retval;
1004
1005         irq = bind_interdomain_evtchn_to_irq(remote_domain, remote_port);
1006         if (irq < 0)
1007                 return irq;
1008
1009         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1010         if (retval != 0) {
1011                 unbind_from_irq(irq);
1012                 return retval;
1013         }
1014
1015         return irq;
1016 }
1017 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler);
1018
1019 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
1020                             irq_handler_t handler,
1021                             unsigned long irqflags, const char *devname, void *dev_id)
1022 {
1023         int irq, retval;
1024
1025         irq = bind_virq_to_irq(virq, cpu);
1026         if (irq < 0)
1027                 return irq;
1028         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1029         if (retval != 0) {
1030                 unbind_from_irq(irq);
1031                 return retval;
1032         }
1033
1034         return irq;
1035 }
1036 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1037
1038 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1039                            unsigned int cpu,
1040                            irq_handler_t handler,
1041                            unsigned long irqflags,
1042                            const char *devname,
1043                            void *dev_id)
1044 {
1045         int irq, retval;
1046
1047         irq = bind_ipi_to_irq(ipi, cpu);
1048         if (irq < 0)
1049                 return irq;
1050
1051         irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME | IRQF_EARLY_RESUME;
1052         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1053         if (retval != 0) {
1054                 unbind_from_irq(irq);
1055                 return retval;
1056         }
1057
1058         return irq;
1059 }
1060
1061 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1062 {
1063         free_irq(irq, dev_id);
1064         unbind_from_irq(irq);
1065 }
1066 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1067
1068 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1069 {
1070         int irq = per_cpu(ipi_to_irq, cpu)[vector];
1071         BUG_ON(irq < 0);
1072         notify_remote_via_irq(irq);
1073 }
1074
1075 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
1076 {
1077         struct shared_info *sh = HYPERVISOR_shared_info;
1078         int cpu = smp_processor_id();
1079         unsigned long *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu);
1080         int i;
1081         unsigned long flags;
1082         static DEFINE_SPINLOCK(debug_lock);
1083         struct vcpu_info *v;
1084
1085         spin_lock_irqsave(&debug_lock, flags);
1086
1087         printk("\nvcpu %d\n  ", cpu);
1088
1089         for_each_online_cpu(i) {
1090                 int pending;
1091                 v = per_cpu(xen_vcpu, i);
1092                 pending = (get_irq_regs() && i == cpu)
1093                         ? xen_irqs_disabled(get_irq_regs())
1094                         : v->evtchn_upcall_mask;
1095                 printk("%d: masked=%d pending=%d event_sel %0*lx\n  ", i,
1096                        pending, v->evtchn_upcall_pending,
1097                        (int)(sizeof(v->evtchn_pending_sel)*2),
1098                        v->evtchn_pending_sel);
1099         }
1100         v = per_cpu(xen_vcpu, cpu);
1101
1102         printk("\npending:\n   ");
1103         for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
1104                 printk("%0*lx%s", (int)sizeof(sh->evtchn_pending[0])*2,
1105                        sh->evtchn_pending[i],
1106                        i % 8 == 0 ? "\n   " : " ");
1107         printk("\nglobal mask:\n   ");
1108         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1109                 printk("%0*lx%s",
1110                        (int)(sizeof(sh->evtchn_mask[0])*2),
1111                        sh->evtchn_mask[i],
1112                        i % 8 == 0 ? "\n   " : " ");
1113
1114         printk("\nglobally unmasked:\n   ");
1115         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1116                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1117                        sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
1118                        i % 8 == 0 ? "\n   " : " ");
1119
1120         printk("\nlocal cpu%d mask:\n   ", cpu);
1121         for (i = (NR_EVENT_CHANNELS/BITS_PER_LONG)-1; i >= 0; i--)
1122                 printk("%0*lx%s", (int)(sizeof(cpu_evtchn[0])*2),
1123                        cpu_evtchn[i],
1124                        i % 8 == 0 ? "\n   " : " ");
1125
1126         printk("\nlocally unmasked:\n   ");
1127         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
1128                 unsigned long pending = sh->evtchn_pending[i]
1129                         & ~sh->evtchn_mask[i]
1130                         & cpu_evtchn[i];
1131                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1132                        pending, i % 8 == 0 ? "\n   " : " ");
1133         }
1134
1135         printk("\npending list:\n");
1136         for (i = 0; i < NR_EVENT_CHANNELS; i++) {
1137                 if (sync_test_bit(i, sh->evtchn_pending)) {
1138                         int word_idx = i / BITS_PER_LONG;
1139                         printk("  %d: event %d -> irq %d%s%s%s\n",
1140                                cpu_from_evtchn(i), i,
1141                                evtchn_to_irq[i],
1142                                sync_test_bit(word_idx, &v->evtchn_pending_sel)
1143                                              ? "" : " l2-clear",
1144                                !sync_test_bit(i, sh->evtchn_mask)
1145                                              ? "" : " globally-masked",
1146                                sync_test_bit(i, cpu_evtchn)
1147                                              ? "" : " locally-masked");
1148                 }
1149         }
1150
1151         spin_unlock_irqrestore(&debug_lock, flags);
1152
1153         return IRQ_HANDLED;
1154 }
1155
1156 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
1157 static DEFINE_PER_CPU(unsigned int, current_word_idx);
1158 static DEFINE_PER_CPU(unsigned int, current_bit_idx);
1159
1160 /*
1161  * Mask out the i least significant bits of w
1162  */
1163 #define MASK_LSBS(w, i) (w & ((~0UL) << i))
1164
1165 /*
1166  * Search the CPUs pending events bitmasks.  For each one found, map
1167  * the event number to an irq, and feed it into do_IRQ() for
1168  * handling.
1169  *
1170  * Xen uses a two-level bitmap to speed searching.  The first level is
1171  * a bitset of words which contain pending event bits.  The second
1172  * level is a bitset of pending events themselves.
1173  */
1174 static void __xen_evtchn_do_upcall(void)
1175 {
1176         int start_word_idx, start_bit_idx;
1177         int word_idx, bit_idx;
1178         int i;
1179         int cpu = get_cpu();
1180         struct shared_info *s = HYPERVISOR_shared_info;
1181         struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1182         unsigned count;
1183
1184         do {
1185                 unsigned long pending_words;
1186
1187                 vcpu_info->evtchn_upcall_pending = 0;
1188
1189                 if (__this_cpu_inc_return(xed_nesting_count) - 1)
1190                         goto out;
1191
1192 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
1193                 /* Clear master flag /before/ clearing selector flag. */
1194                 wmb();
1195 #endif
1196                 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
1197
1198                 start_word_idx = __this_cpu_read(current_word_idx);
1199                 start_bit_idx = __this_cpu_read(current_bit_idx);
1200
1201                 word_idx = start_word_idx;
1202
1203                 for (i = 0; pending_words != 0; i++) {
1204                         unsigned long pending_bits;
1205                         unsigned long words;
1206
1207                         words = MASK_LSBS(pending_words, word_idx);
1208
1209                         /*
1210                          * If we masked out all events, wrap to beginning.
1211                          */
1212                         if (words == 0) {
1213                                 word_idx = 0;
1214                                 bit_idx = 0;
1215                                 continue;
1216                         }
1217                         word_idx = __ffs(words);
1218
1219                         pending_bits = active_evtchns(cpu, s, word_idx);
1220                         bit_idx = 0; /* usually scan entire word from start */
1221                         if (word_idx == start_word_idx) {
1222                                 /* We scan the starting word in two parts */
1223                                 if (i == 0)
1224                                         /* 1st time: start in the middle */
1225                                         bit_idx = start_bit_idx;
1226                                 else
1227                                         /* 2nd time: mask bits done already */
1228                                         bit_idx &= (1UL << start_bit_idx) - 1;
1229                         }
1230
1231                         do {
1232                                 unsigned long bits;
1233                                 int port, irq;
1234                                 struct irq_desc *desc;
1235
1236                                 bits = MASK_LSBS(pending_bits, bit_idx);
1237
1238                                 /* If we masked out all events, move on. */
1239                                 if (bits == 0)
1240                                         break;
1241
1242                                 bit_idx = __ffs(bits);
1243
1244                                 /* Process port. */
1245                                 port = (word_idx * BITS_PER_LONG) + bit_idx;
1246                                 irq = evtchn_to_irq[port];
1247
1248                                 if (irq != -1) {
1249                                         desc = irq_to_desc(irq);
1250                                         if (desc)
1251                                                 generic_handle_irq_desc(irq, desc);
1252                                 }
1253
1254                                 bit_idx = (bit_idx + 1) % BITS_PER_LONG;
1255
1256                                 /* Next caller starts at last processed + 1 */
1257                                 __this_cpu_write(current_word_idx,
1258                                                  bit_idx ? word_idx :
1259                                                  (word_idx+1) % BITS_PER_LONG);
1260                                 __this_cpu_write(current_bit_idx, bit_idx);
1261                         } while (bit_idx != 0);
1262
1263                         /* Scan start_l1i twice; all others once. */
1264                         if ((word_idx != start_word_idx) || (i != 0))
1265                                 pending_words &= ~(1UL << word_idx);
1266
1267                         word_idx = (word_idx + 1) % BITS_PER_LONG;
1268                 }
1269
1270                 BUG_ON(!irqs_disabled());
1271
1272                 count = __this_cpu_read(xed_nesting_count);
1273                 __this_cpu_write(xed_nesting_count, 0);
1274         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
1275
1276 out:
1277
1278         put_cpu();
1279 }
1280
1281 void xen_evtchn_do_upcall(struct pt_regs *regs)
1282 {
1283         struct pt_regs *old_regs = set_irq_regs(regs);
1284
1285         exit_idle();
1286         irq_enter();
1287
1288         __xen_evtchn_do_upcall();
1289
1290         irq_exit();
1291         set_irq_regs(old_regs);
1292 }
1293
1294 void xen_hvm_evtchn_do_upcall(void)
1295 {
1296         __xen_evtchn_do_upcall();
1297 }
1298 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1299
1300 /* Rebind a new event channel to an existing irq. */
1301 void rebind_evtchn_irq(int evtchn, int irq)
1302 {
1303         struct irq_info *info = info_for_irq(irq);
1304
1305         /* Make sure the irq is masked, since the new event channel
1306            will also be masked. */
1307         disable_irq(irq);
1308
1309         mutex_lock(&irq_mapping_update_lock);
1310
1311         /* After resume the irq<->evtchn mappings are all cleared out */
1312         BUG_ON(evtchn_to_irq[evtchn] != -1);
1313         /* Expect irq to have been bound before,
1314            so there should be a proper type */
1315         BUG_ON(info->type == IRQT_UNBOUND);
1316
1317         xen_irq_info_evtchn_init(irq, evtchn);
1318
1319         mutex_unlock(&irq_mapping_update_lock);
1320
1321         /* new event channels are always bound to cpu 0 */
1322         irq_set_affinity(irq, cpumask_of(0));
1323
1324         /* Unmask the event channel. */
1325         enable_irq(irq);
1326 }
1327
1328 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1329 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1330 {
1331         struct evtchn_bind_vcpu bind_vcpu;
1332         int evtchn = evtchn_from_irq(irq);
1333
1334         if (!VALID_EVTCHN(evtchn))
1335                 return -1;
1336
1337         /*
1338          * Events delivered via platform PCI interrupts are always
1339          * routed to vcpu 0 and hence cannot be rebound.
1340          */
1341         if (xen_hvm_domain() && !xen_have_vector_callback)
1342                 return -1;
1343
1344         /* Send future instances of this interrupt to other vcpu. */
1345         bind_vcpu.port = evtchn;
1346         bind_vcpu.vcpu = tcpu;
1347
1348         /*
1349          * If this fails, it usually just indicates that we're dealing with a
1350          * virq or IPI channel, which don't actually need to be rebound. Ignore
1351          * it, but don't do the xenlinux-level rebind in that case.
1352          */
1353         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1354                 bind_evtchn_to_cpu(evtchn, tcpu);
1355
1356         return 0;
1357 }
1358
1359 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1360                             bool force)
1361 {
1362         unsigned tcpu = cpumask_first(dest);
1363
1364         return rebind_irq_to_cpu(data->irq, tcpu);
1365 }
1366
1367 int resend_irq_on_evtchn(unsigned int irq)
1368 {
1369         int masked, evtchn = evtchn_from_irq(irq);
1370         struct shared_info *s = HYPERVISOR_shared_info;
1371
1372         if (!VALID_EVTCHN(evtchn))
1373                 return 1;
1374
1375         masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
1376         sync_set_bit(evtchn, s->evtchn_pending);
1377         if (!masked)
1378                 unmask_evtchn(evtchn);
1379
1380         return 1;
1381 }
1382
1383 static void enable_dynirq(struct irq_data *data)
1384 {
1385         int evtchn = evtchn_from_irq(data->irq);
1386
1387         if (VALID_EVTCHN(evtchn))
1388                 unmask_evtchn(evtchn);
1389 }
1390
1391 static void disable_dynirq(struct irq_data *data)
1392 {
1393         int evtchn = evtchn_from_irq(data->irq);
1394
1395         if (VALID_EVTCHN(evtchn))
1396                 mask_evtchn(evtchn);
1397 }
1398
1399 static void ack_dynirq(struct irq_data *data)
1400 {
1401         int evtchn = evtchn_from_irq(data->irq);
1402
1403         irq_move_irq(data);
1404
1405         if (VALID_EVTCHN(evtchn))
1406                 clear_evtchn(evtchn);
1407 }
1408
1409 static void mask_ack_dynirq(struct irq_data *data)
1410 {
1411         disable_dynirq(data);
1412         ack_dynirq(data);
1413 }
1414
1415 static int retrigger_dynirq(struct irq_data *data)
1416 {
1417         int evtchn = evtchn_from_irq(data->irq);
1418         struct shared_info *sh = HYPERVISOR_shared_info;
1419         int ret = 0;
1420
1421         if (VALID_EVTCHN(evtchn)) {
1422                 int masked;
1423
1424                 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
1425                 sync_set_bit(evtchn, sh->evtchn_pending);
1426                 if (!masked)
1427                         unmask_evtchn(evtchn);
1428                 ret = 1;
1429         }
1430
1431         return ret;
1432 }
1433
1434 static void restore_pirqs(void)
1435 {
1436         int pirq, rc, irq, gsi;
1437         struct physdev_map_pirq map_irq;
1438         struct irq_info *info;
1439
1440         list_for_each_entry(info, &xen_irq_list_head, list) {
1441                 if (info->type != IRQT_PIRQ)
1442                         continue;
1443
1444                 pirq = info->u.pirq.pirq;
1445                 gsi = info->u.pirq.gsi;
1446                 irq = info->irq;
1447
1448                 /* save/restore of PT devices doesn't work, so at this point the
1449                  * only devices present are GSI based emulated devices */
1450                 if (!gsi)
1451                         continue;
1452
1453                 map_irq.domid = DOMID_SELF;
1454                 map_irq.type = MAP_PIRQ_TYPE_GSI;
1455                 map_irq.index = gsi;
1456                 map_irq.pirq = pirq;
1457
1458                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1459                 if (rc) {
1460                         printk(KERN_WARNING "xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1461                                         gsi, irq, pirq, rc);
1462                         xen_free_irq(irq);
1463                         continue;
1464                 }
1465
1466                 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1467
1468                 __startup_pirq(irq);
1469         }
1470 }
1471
1472 static void restore_cpu_virqs(unsigned int cpu)
1473 {
1474         struct evtchn_bind_virq bind_virq;
1475         int virq, irq, evtchn;
1476
1477         for (virq = 0; virq < NR_VIRQS; virq++) {
1478                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1479                         continue;
1480
1481                 BUG_ON(virq_from_irq(irq) != virq);
1482
1483                 /* Get a new binding from Xen. */
1484                 bind_virq.virq = virq;
1485                 bind_virq.vcpu = cpu;
1486                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1487                                                 &bind_virq) != 0)
1488                         BUG();
1489                 evtchn = bind_virq.port;
1490
1491                 /* Record the new mapping. */
1492                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
1493                 bind_evtchn_to_cpu(evtchn, cpu);
1494         }
1495 }
1496
1497 static void restore_cpu_ipis(unsigned int cpu)
1498 {
1499         struct evtchn_bind_ipi bind_ipi;
1500         int ipi, irq, evtchn;
1501
1502         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1503                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1504                         continue;
1505
1506                 BUG_ON(ipi_from_irq(irq) != ipi);
1507
1508                 /* Get a new binding from Xen. */
1509                 bind_ipi.vcpu = cpu;
1510                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1511                                                 &bind_ipi) != 0)
1512                         BUG();
1513                 evtchn = bind_ipi.port;
1514
1515                 /* Record the new mapping. */
1516                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
1517                 bind_evtchn_to_cpu(evtchn, cpu);
1518         }
1519 }
1520
1521 /* Clear an irq's pending state, in preparation for polling on it */
1522 void xen_clear_irq_pending(int irq)
1523 {
1524         int evtchn = evtchn_from_irq(irq);
1525
1526         if (VALID_EVTCHN(evtchn))
1527                 clear_evtchn(evtchn);
1528 }
1529 EXPORT_SYMBOL(xen_clear_irq_pending);
1530 void xen_set_irq_pending(int irq)
1531 {
1532         int evtchn = evtchn_from_irq(irq);
1533
1534         if (VALID_EVTCHN(evtchn))
1535                 set_evtchn(evtchn);
1536 }
1537
1538 bool xen_test_irq_pending(int irq)
1539 {
1540         int evtchn = evtchn_from_irq(irq);
1541         bool ret = false;
1542
1543         if (VALID_EVTCHN(evtchn))
1544                 ret = test_evtchn(evtchn);
1545
1546         return ret;
1547 }
1548
1549 /* Poll waiting for an irq to become pending with timeout.  In the usual case,
1550  * the irq will be disabled so it won't deliver an interrupt. */
1551 void xen_poll_irq_timeout(int irq, u64 timeout)
1552 {
1553         evtchn_port_t evtchn = evtchn_from_irq(irq);
1554
1555         if (VALID_EVTCHN(evtchn)) {
1556                 struct sched_poll poll;
1557
1558                 poll.nr_ports = 1;
1559                 poll.timeout = timeout;
1560                 set_xen_guest_handle(poll.ports, &evtchn);
1561
1562                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1563                         BUG();
1564         }
1565 }
1566 EXPORT_SYMBOL(xen_poll_irq_timeout);
1567 /* Poll waiting for an irq to become pending.  In the usual case, the
1568  * irq will be disabled so it won't deliver an interrupt. */
1569 void xen_poll_irq(int irq)
1570 {
1571         xen_poll_irq_timeout(irq, 0 /* no timeout */);
1572 }
1573
1574 /* Check whether the IRQ line is shared with other guests. */
1575 int xen_test_irq_shared(int irq)
1576 {
1577         struct irq_info *info = info_for_irq(irq);
1578         struct physdev_irq_status_query irq_status = { .irq = info->u.pirq.pirq };
1579
1580         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
1581                 return 0;
1582         return !(irq_status.flags & XENIRQSTAT_shared);
1583 }
1584 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
1585
1586 void xen_irq_resume(void)
1587 {
1588         unsigned int cpu, evtchn;
1589         struct irq_info *info;
1590
1591         init_evtchn_cpu_bindings();
1592
1593         /* New event-channel space is not 'live' yet. */
1594         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1595                 mask_evtchn(evtchn);
1596
1597         /* No IRQ <-> event-channel mappings. */
1598         list_for_each_entry(info, &xen_irq_list_head, list)
1599                 info->evtchn = 0; /* zap event-channel binding */
1600
1601         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1602                 evtchn_to_irq[evtchn] = -1;
1603
1604         for_each_possible_cpu(cpu) {
1605                 restore_cpu_virqs(cpu);
1606                 restore_cpu_ipis(cpu);
1607         }
1608
1609         restore_pirqs();
1610 }
1611
1612 static struct irq_chip xen_dynamic_chip __read_mostly = {
1613         .name                   = "xen-dyn",
1614
1615         .irq_disable            = disable_dynirq,
1616         .irq_mask               = disable_dynirq,
1617         .irq_unmask             = enable_dynirq,
1618
1619         .irq_ack                = ack_dynirq,
1620         .irq_mask_ack           = mask_ack_dynirq,
1621
1622         .irq_set_affinity       = set_affinity_irq,
1623         .irq_retrigger          = retrigger_dynirq,
1624 };
1625
1626 static struct irq_chip xen_pirq_chip __read_mostly = {
1627         .name                   = "xen-pirq",
1628
1629         .irq_startup            = startup_pirq,
1630         .irq_shutdown           = shutdown_pirq,
1631         .irq_enable             = enable_pirq,
1632         .irq_disable            = disable_pirq,
1633
1634         .irq_mask               = disable_dynirq,
1635         .irq_unmask             = enable_dynirq,
1636
1637         .irq_ack                = eoi_pirq,
1638         .irq_eoi                = eoi_pirq,
1639         .irq_mask_ack           = mask_ack_pirq,
1640
1641         .irq_set_affinity       = set_affinity_irq,
1642
1643         .irq_retrigger          = retrigger_dynirq,
1644 };
1645
1646 static struct irq_chip xen_percpu_chip __read_mostly = {
1647         .name                   = "xen-percpu",
1648
1649         .irq_disable            = disable_dynirq,
1650         .irq_mask               = disable_dynirq,
1651         .irq_unmask             = enable_dynirq,
1652
1653         .irq_ack                = ack_dynirq,
1654 };
1655
1656 int xen_set_callback_via(uint64_t via)
1657 {
1658         struct xen_hvm_param a;
1659         a.domid = DOMID_SELF;
1660         a.index = HVM_PARAM_CALLBACK_IRQ;
1661         a.value = via;
1662         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1663 }
1664 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1665
1666 #ifdef CONFIG_XEN_PVHVM
1667 /* Vector callbacks are better than PCI interrupts to receive event
1668  * channel notifications because we can receive vector callbacks on any
1669  * vcpu and we don't need PCI support or APIC interactions. */
1670 void xen_callback_vector(void)
1671 {
1672         int rc;
1673         uint64_t callback_via;
1674         if (xen_have_vector_callback) {
1675                 callback_via = HVM_CALLBACK_VECTOR(XEN_HVM_EVTCHN_CALLBACK);
1676                 rc = xen_set_callback_via(callback_via);
1677                 if (rc) {
1678                         printk(KERN_ERR "Request for Xen HVM callback vector"
1679                                         " failed.\n");
1680                         xen_have_vector_callback = 0;
1681                         return;
1682                 }
1683                 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1684                                 "enabled\n");
1685                 /* in the restore case the vector has already been allocated */
1686                 if (!test_bit(XEN_HVM_EVTCHN_CALLBACK, used_vectors))
1687                         alloc_intr_gate(XEN_HVM_EVTCHN_CALLBACK, xen_hvm_callback_vector);
1688         }
1689 }
1690 #else
1691 void xen_callback_vector(void) {}
1692 #endif
1693
1694 void __init xen_init_IRQ(void)
1695 {
1696         int i;
1697
1698         evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1699                                     GFP_KERNEL);
1700         BUG_ON(!evtchn_to_irq);
1701         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1702                 evtchn_to_irq[i] = -1;
1703
1704         init_evtchn_cpu_bindings();
1705
1706         /* No event channels are 'live' right now. */
1707         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1708                 mask_evtchn(i);
1709
1710         if (xen_hvm_domain()) {
1711                 xen_callback_vector();
1712                 native_init_IRQ();
1713                 /* pci_xen_hvm_init must be called after native_init_IRQ so that
1714                  * __acpi_register_gsi can point at the right function */
1715                 pci_xen_hvm_init();
1716         } else {
1717                 irq_ctx_init(smp_processor_id());
1718                 if (xen_initial_domain())
1719                         pci_xen_initial_domain();
1720         }
1721 }