drivers/xen/events.c

   1 /*
   2  * Xen event channels
   3  *
   4  * Xen models interrupts with abstract event channels.  Because each
   5  * domain gets 1024 event channels, but NR_IRQ is not that large, we
   6  * must dynamically map irqs<->event channels.  The event channels
   7  * interface with the rest of the kernel by defining a xen interrupt
   8  * chip.  When an event is received, it is mapped to an irq and sent
   9  * through the normal interrupt processing path.
  10  *
  11  * There are four kinds of events which can be mapped to an event
  12  * channel:
  13  *
  14  * 1. Inter-domain notifications.  This includes all the virtual
  15  *    device events, since they're driven by front-ends in another domain
  16  *    (typically dom0).
  17  * 2. VIRQs, typically used for timers.  These are per-cpu events.
  18  * 3. IPIs.
  19  * 4. PIRQs - Hardware interrupts.
  20  *
  21  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
  22  */
  23
  24 #include <linux/linkage.h>
  25 #include <linux/interrupt.h>
  26 #include <linux/irq.h>
  27 #include <linux/module.h>
  28 #include <linux/string.h>
  29 #include <linux/bootmem.h>
  30 #include <linux/slab.h>
  31 #include <linux/irqnr.h>
  32 #include <linux/pci.h>
  33
  34 #include <asm/desc.h>
  35 #include <asm/ptrace.h>
  36 #include <asm/irq.h>
  37 #include <asm/idle.h>
  38 #include <asm/io_apic.h>
  39 #include <asm/sync_bitops.h>
  40 #include <asm/xen/page.h>
  41 #include <asm/xen/pci.h>
  42 #include <asm/xen/hypercall.h>
  43 #include <asm/xen/hypervisor.h>
  44
  45 #include <xen/xen.h>
  46 #include <xen/hvm.h>
  47 #include <xen/xen-ops.h>
  48 #include <xen/events.h>
  49 #include <xen/interface/xen.h>
  50 #include <xen/interface/event_channel.h>
  51 #include <xen/interface/hvm/hvm_op.h>
  52 #include <xen/interface/hvm/params.h>
  53
  54 /*
  55  * This lock protects updates to the following mapping and reference-count
  56  * arrays. The lock does not need to be acquired to read the mapping tables.
  57  */
  58 static DEFINE_MUTEX(irq_mapping_update_lock);
  59
  60 static LIST_HEAD(xen_irq_list_head);
  61
  62 /* IRQ <-> VIRQ mapping. */
  63 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
  64
  65 /* IRQ <-> IPI mapping */
  66 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
  67
  68 /* Interrupt types. */
  69 enum xen_irq_type {
  70         IRQT_UNBOUND = 0,
  71         IRQT_PIRQ,
  72         IRQT_VIRQ,
  73         IRQT_IPI,
  74         IRQT_EVTCHN
  75 };
  76
  77 /*
  78  * Packed IRQ information:
  79  * type - enum xen_irq_type
  80  * event channel - irq->event channel mapping
  81  * cpu - cpu this event channel is bound to
  82  * index - type-specific information:
  83  *    PIRQ - vector, with MSB being "needs EIO", or physical IRQ of the HVM
  84  *           guest, or GSI (real passthrough IRQ) of the device.
  85  *    VIRQ - virq number
  86  *    IPI - IPI vector
  87  *    EVTCHN -
  88  */
  89 struct irq_info {
  90         struct list_head list;
  91         int refcnt;
  92         enum xen_irq_type type; /* type */
  93         unsigned irq;
  94         unsigned short evtchn;  /* event channel */
  95         unsigned short cpu;     /* cpu bound */
  96
  97         union {
  98                 unsigned short virq;
  99                 enum ipi_vector ipi;
 100                 struct {
 101                         unsigned short pirq;
 102                         unsigned short gsi;
 103                         unsigned char vector;
 104                         unsigned char flags;
 105                         uint16_t domid;
 106                 } pirq;
 107         } u;
 108 };
 109 #define PIRQ_NEEDS_EOI  (1 << 0)
 110 #define PIRQ_SHAREABLE  (1 << 1)
 111
 112 static int *evtchn_to_irq;
 113 static unsigned long *pirq_eoi_map;
 114 static bool (*pirq_needs_eoi)(unsigned irq);
 115
 116 static DEFINE_PER_CPU(unsigned long [NR_EVENT_CHANNELS/BITS_PER_LONG],
 117                       cpu_evtchn_mask);
 118
 119 /* Xen will never allocate port zero for any purpose. */
 120 #define VALID_EVTCHN(chn)       ((chn) != 0)
 121
 122 static struct irq_chip xen_dynamic_chip;
 123 static struct irq_chip xen_percpu_chip;
 124 static struct irq_chip xen_pirq_chip;
 125 static void enable_dynirq(struct irq_data *data);
 126 static void disable_dynirq(struct irq_data *data);
 127
 128 /* Get info for IRQ */
 129 static struct irq_info *info_for_irq(unsigned irq)
 130 {
 131         return irq_get_handler_data(irq);
 132 }
 133
 134 /* Constructors for packed IRQ information. */
 135 static void xen_irq_info_common_init(struct irq_info *info,
 136                                      unsigned irq,
 137                                      enum xen_irq_type type,
 138                                      unsigned short evtchn,
 139                                      unsigned short cpu)
 140 {
 141
 142         BUG_ON(info->type != IRQT_UNBOUND && info->type != type);
 143
 144         info->type = type;
 145         info->irq = irq;
 146         info->evtchn = evtchn;
 147         info->cpu = cpu;
 148
 149         evtchn_to_irq[evtchn] = irq;
 150 }
 151
 152 static void xen_irq_info_evtchn_init(unsigned irq,
 153                                      unsigned short evtchn)
 154 {
 155         struct irq_info *info = info_for_irq(irq);
 156
 157         xen_irq_info_common_init(info, irq, IRQT_EVTCHN, evtchn, 0);
 158 }
 159
 160 static void xen_irq_info_ipi_init(unsigned cpu,
 161                                   unsigned irq,
 162                                   unsigned short evtchn,
 163                                   enum ipi_vector ipi)
 164 {
 165         struct irq_info *info = info_for_irq(irq);
 166
 167         xen_irq_info_common_init(info, irq, IRQT_IPI, evtchn, 0);
 168
 169         info->u.ipi = ipi;
 170
 171         per_cpu(ipi_to_irq, cpu)[ipi] = irq;
 172 }
 173
 174 static void xen_irq_info_virq_init(unsigned cpu,
 175                                    unsigned irq,
 176                                    unsigned short evtchn,
 177                                    unsigned short virq)
 178 {
 179         struct irq_info *info = info_for_irq(irq);
 180
 181         xen_irq_info_common_init(info, irq, IRQT_VIRQ, evtchn, 0);
 182
 183         info->u.virq = virq;
 184
 185         per_cpu(virq_to_irq, cpu)[virq] = irq;
 186 }
 187
 188 static void xen_irq_info_pirq_init(unsigned irq,
 189                                    unsigned short evtchn,
 190                                    unsigned short pirq,
 191                                    unsigned short gsi,
 192                                    unsigned short vector,
 193                                    uint16_t domid,
 194                                    unsigned char flags)
 195 {
 196         struct irq_info *info = info_for_irq(irq);
 197
 198         xen_irq_info_common_init(info, irq, IRQT_PIRQ, evtchn, 0);
 199
 200         info->u.pirq.pirq = pirq;
 201         info->u.pirq.gsi = gsi;
 202         info->u.pirq.vector = vector;
 203         info->u.pirq.domid = domid;
 204         info->u.pirq.flags = flags;
 205 }
 206
 207 /*
 208  * Accessors for packed IRQ information.
 209  */
 210 static unsigned int evtchn_from_irq(unsigned irq)
 211 {
 212         if (unlikely(WARN(irq < 0 || irq >= nr_irqs, "Invalid irq %d!\n", irq)))
 213                 return 0;
 214
 215         return info_for_irq(irq)->evtchn;
 216 }
 217
 218 unsigned irq_from_evtchn(unsigned int evtchn)
 219 {
 220         return evtchn_to_irq[evtchn];
 221 }
 222 EXPORT_SYMBOL_GPL(irq_from_evtchn);
 223
 224 static enum ipi_vector ipi_from_irq(unsigned irq)
 225 {
 226         struct irq_info *info = info_for_irq(irq);
 227
 228         BUG_ON(info == NULL);
 229         BUG_ON(info->type != IRQT_IPI);
 230
 231         return info->u.ipi;
 232 }
 233
 234 static unsigned virq_from_irq(unsigned irq)
 235 {
 236         struct irq_info *info = info_for_irq(irq);
 237
 238         BUG_ON(info == NULL);
 239         BUG_ON(info->type != IRQT_VIRQ);
 240
 241         return info->u.virq;
 242 }
 243
 244 static unsigned pirq_from_irq(unsigned irq)
 245 {
 246         struct irq_info *info = info_for_irq(irq);
 247
 248         BUG_ON(info == NULL);
 249         BUG_ON(info->type != IRQT_PIRQ);
 250
 251         return info->u.pirq.pirq;
 252 }
 253
 254 static enum xen_irq_type type_from_irq(unsigned irq)
 255 {
 256         return info_for_irq(irq)->type;
 257 }
 258
 259 static unsigned cpu_from_irq(unsigned irq)
 260 {
 261         return info_for_irq(irq)->cpu;
 262 }
 263
 264 static unsigned int cpu_from_evtchn(unsigned int evtchn)
 265 {
 266         int irq = evtchn_to_irq[evtchn];
 267         unsigned ret = 0;
 268
 269         if (irq != -1)
 270                 ret = cpu_from_irq(irq);
 271
 272         return ret;
 273 }
 274
 275 static bool pirq_check_eoi_map(unsigned irq)
 276 {
 277         return test_bit(pirq_from_irq(irq), pirq_eoi_map);
 278 }
 279
 280 static bool pirq_needs_eoi_flag(unsigned irq)
 281 {
 282         struct irq_info *info = info_for_irq(irq);
 283         BUG_ON(info->type != IRQT_PIRQ);
 284
 285         return info->u.pirq.flags & PIRQ_NEEDS_EOI;
 286 }
 287
 288 static inline unsigned long active_evtchns(unsigned int cpu,
 289                                            struct shared_info *sh,
 290                                            unsigned int idx)
 291 {
 292         return sh->evtchn_pending[idx] &
 293                 per_cpu(cpu_evtchn_mask, cpu)[idx] &
 294                 ~sh->evtchn_mask[idx];
 295 }
 296
 297 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
 298 {
 299         int irq = evtchn_to_irq[chn];
 300
 301         BUG_ON(irq == -1);
 302 #ifdef CONFIG_SMP
 303         cpumask_copy(irq_to_desc(irq)->irq_data.affinity, cpumask_of(cpu));
 304 #endif
 305
 306         clear_bit(chn, per_cpu(cpu_evtchn_mask, cpu_from_irq(irq)));
 307         set_bit(chn, per_cpu(cpu_evtchn_mask, cpu));
 308
 309         info_for_irq(irq)->cpu = cpu;
 310 }
 311
 312 static void init_evtchn_cpu_bindings(void)
 313 {
 314         int i;
 315 #ifdef CONFIG_SMP
 316         struct irq_info *info;
 317
 318         /* By default all event channels notify CPU#0. */
 319         list_for_each_entry(info, &xen_irq_list_head, list) {
 320                 struct irq_desc *desc = irq_to_desc(info->irq);
 321                 cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 322         }
 323 #endif
 324
 325         for_each_possible_cpu(i)
 326                 memset(per_cpu(cpu_evtchn_mask, i),
 327                        (i == 0) ? ~0 : 0, sizeof(*per_cpu(cpu_evtchn_mask, i)));
 328 }
 329
 330 static inline void clear_evtchn(int port)
 331 {
 332         struct shared_info *s = HYPERVISOR_shared_info;
 333         sync_clear_bit(port, &s->evtchn_pending[0]);
 334 }
 335
 336 static inline void set_evtchn(int port)
 337 {
 338         struct shared_info *s = HYPERVISOR_shared_info;
 339         sync_set_bit(port, &s->evtchn_pending[0]);
 340 }
 341
 342 static inline int test_evtchn(int port)
 343 {
 344         struct shared_info *s = HYPERVISOR_shared_info;
 345         return sync_test_bit(port, &s->evtchn_pending[0]);
 346 }
 347
 348
 349 /**
 350  * notify_remote_via_irq - send event to remote end of event channel via irq
 351  * @irq: irq of event channel to send event to
 352  *
 353  * Unlike notify_remote_via_evtchn(), this is safe to use across
 354  * save/restore. Notifications on a broken connection are silently
 355  * dropped.
 356  */
 357 void notify_remote_via_irq(int irq)
 358 {
 359         int evtchn = evtchn_from_irq(irq);
 360
 361         if (VALID_EVTCHN(evtchn))
 362                 notify_remote_via_evtchn(evtchn);
 363 }
 364 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
 365
 366 static void mask_evtchn(int port)
 367 {
 368         struct shared_info *s = HYPERVISOR_shared_info;
 369         sync_set_bit(port, &s->evtchn_mask[0]);
 370 }
 371
 372 static void unmask_evtchn(int port)
 373 {
 374         struct shared_info *s = HYPERVISOR_shared_info;
 375         unsigned int cpu = get_cpu();
 376
 377         BUG_ON(!irqs_disabled());
 378
 379         /* Slow path (hypercall) if this is a non-local port. */
 380         if (unlikely(cpu != cpu_from_evtchn(port))) {
 381                 struct evtchn_unmask unmask = { .port = port };
 382                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 383         } else {
 384                 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
 385
 386                 sync_clear_bit(port, &s->evtchn_mask[0]);
 387
 388                 /*
 389                  * The following is basically the equivalent of
 390                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
 391                  * the interrupt edge' if the channel is masked.
 392                  */
 393                 if (sync_test_bit(port, &s->evtchn_pending[0]) &&
 394                     !sync_test_and_set_bit(port / BITS_PER_LONG,
 395                                            &vcpu_info->evtchn_pending_sel))
 396                         vcpu_info->evtchn_upcall_pending = 1;
 397         }
 398
 399         put_cpu();
 400 }
 401
 402 static void xen_irq_init(unsigned irq)
 403 {
 404         struct irq_info *info;
 405 #ifdef CONFIG_SMP
 406         struct irq_desc *desc = irq_to_desc(irq);
 407
 408         /* By default all event channels notify CPU#0. */
 409         cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 410 #endif
 411
 412         info = kzalloc(sizeof(*info), GFP_KERNEL);
 413         if (info == NULL)
 414                 panic("Unable to allocate metadata for IRQ%d\n", irq);
 415
 416         info->type = IRQT_UNBOUND;
 417         info->refcnt = -1;
 418
 419         irq_set_handler_data(irq, info);
 420
 421         list_add_tail(&info->list, &xen_irq_list_head);
 422 }
 423
 424 static int __must_check xen_allocate_irq_dynamic(void)
 425 {
 426         int first = 0;
 427         int irq;
 428
 429 #ifdef CONFIG_X86_IO_APIC
 430         /*
 431          * For an HVM guest or domain 0 which see "real" (emulated or
 432          * actual respectively) GSIs we allocate dynamic IRQs
 433          * e.g. those corresponding to event channels or MSIs
 434          * etc. from the range above those "real" GSIs to avoid
 435          * collisions.
 436          */
 437         if (xen_initial_domain() || xen_hvm_domain())
 438                 first = get_nr_irqs_gsi();
 439 #endif
 440
 441         irq = irq_alloc_desc_from(first, -1);
 442
 443         if (irq >= 0)
 444                 xen_irq_init(irq);
 445
 446         return irq;
 447 }
 448
 449 static int __must_check xen_allocate_irq_gsi(unsigned gsi)
 450 {
 451         int irq;
 452
 453         /*
 454          * A PV guest has no concept of a GSI (since it has no ACPI
 455          * nor access to/knowledge of the physical APICs). Therefore
 456          * all IRQs are dynamically allocated from the entire IRQ
 457          * space.
 458          */
 459         if (xen_pv_domain() && !xen_initial_domain())
 460                 return xen_allocate_irq_dynamic();
 461
 462         /* Legacy IRQ descriptors are already allocated by the arch. */
 463         if (gsi < NR_IRQS_LEGACY)
 464                 irq = gsi;
 465         else
 466                 irq = irq_alloc_desc_at(gsi, -1);
 467
 468         xen_irq_init(irq);
 469
 470         return irq;
 471 }
 472
 473 static void xen_free_irq(unsigned irq)
 474 {
 475         struct irq_info *info = irq_get_handler_data(irq);
 476
 477         list_del(&info->list);
 478
 479         irq_set_handler_data(irq, NULL);
 480
 481         WARN_ON(info->refcnt > 0);
 482
 483         kfree(info);
 484
 485         /* Legacy IRQ descriptors are managed by the arch. */
 486         if (irq < NR_IRQS_LEGACY)
 487                 return;
 488
 489         irq_free_desc(irq);
 490 }
 491
 492 static void pirq_query_unmask(int irq)
 493 {
 494         struct physdev_irq_status_query irq_status;
 495         struct irq_info *info = info_for_irq(irq);
 496
 497         BUG_ON(info->type != IRQT_PIRQ);
 498
 499         irq_status.irq = pirq_from_irq(irq);
 500         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
 501                 irq_status.flags = 0;
 502
 503         info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
 504         if (irq_status.flags & XENIRQSTAT_needs_eoi)
 505                 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
 506 }
 507
 508 static bool probing_irq(int irq)
 509 {
 510         struct irq_desc *desc = irq_to_desc(irq);
 511
 512         return desc && desc->action == NULL;
 513 }
 514
 515 static void eoi_pirq(struct irq_data *data)
 516 {
 517         int evtchn = evtchn_from_irq(data->irq);
 518         struct physdev_eoi eoi = { .irq = pirq_from_irq(data->irq) };
 519         int rc = 0;
 520
 521         irq_move_irq(data);
 522
 523         if (VALID_EVTCHN(evtchn))
 524                 clear_evtchn(evtchn);
 525
 526         if (pirq_needs_eoi(data->irq)) {
 527                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
 528                 WARN_ON(rc);
 529         }
 530 }
 531
 532 static void mask_ack_pirq(struct irq_data *data)
 533 {
 534         disable_dynirq(data);
 535         eoi_pirq(data);
 536 }
 537
 538 static unsigned int __startup_pirq(unsigned int irq)
 539 {
 540         struct evtchn_bind_pirq bind_pirq;
 541         struct irq_info *info = info_for_irq(irq);
 542         int evtchn = evtchn_from_irq(irq);
 543         int rc;
 544
 545         BUG_ON(info->type != IRQT_PIRQ);
 546
 547         if (VALID_EVTCHN(evtchn))
 548                 goto out;
 549
 550         bind_pirq.pirq = pirq_from_irq(irq);
 551         /* NB. We are happy to share unless we are probing. */
 552         bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
 553                                         BIND_PIRQ__WILL_SHARE : 0;
 554         rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
 555         if (rc != 0) {
 556                 if (!probing_irq(irq))
 557                         printk(KERN_INFO "Failed to obtain physical IRQ %d\n",
 558                                irq);
 559                 return 0;
 560         }
 561         evtchn = bind_pirq.port;
 562
 563         pirq_query_unmask(irq);
 564
 565         evtchn_to_irq[evtchn] = irq;
 566         bind_evtchn_to_cpu(evtchn, 0);
 567         info->evtchn = evtchn;
 568
 569 out:
 570         unmask_evtchn(evtchn);
 571         eoi_pirq(irq_get_irq_data(irq));
 572
 573         return 0;
 574 }
 575
 576 static unsigned int startup_pirq(struct irq_data *data)
 577 {
 578         return __startup_pirq(data->irq);
 579 }
 580
 581 static void shutdown_pirq(struct irq_data *data)
 582 {
 583         struct evtchn_close close;
 584         unsigned int irq = data->irq;
 585         struct irq_info *info = info_for_irq(irq);
 586         int evtchn = evtchn_from_irq(irq);
 587
 588         BUG_ON(info->type != IRQT_PIRQ);
 589
 590         if (!VALID_EVTCHN(evtchn))
 591                 return;
 592
 593         mask_evtchn(evtchn);
 594
 595         close.port = evtchn;
 596         if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 597                 BUG();
 598
 599         bind_evtchn_to_cpu(evtchn, 0);
 600         evtchn_to_irq[evtchn] = -1;
 601         info->evtchn = 0;
 602 }
 603
 604 static void enable_pirq(struct irq_data *data)
 605 {
 606         startup_pirq(data);
 607 }
 608
 609 static void disable_pirq(struct irq_data *data)
 610 {
 611         disable_dynirq(data);
 612 }
 613
 614 static int find_irq_by_gsi(unsigned gsi)
 615 {
 616         struct irq_info *info;
 617
 618         list_for_each_entry(info, &xen_irq_list_head, list) {
 619                 if (info->type != IRQT_PIRQ)
 620                         continue;
 621
 622                 if (info->u.pirq.gsi == gsi)
 623                         return info->irq;
 624         }
 625
 626         return -1;
 627 }
 628
 629 /*
 630  * Do not make any assumptions regarding the relationship between the
 631  * IRQ number returned here and the Xen pirq argument.
 632  *
 633  * Note: We don't assign an event channel until the irq actually started
 634  * up.  Return an existing irq if we've already got one for the gsi.
 635  *
 636  * Shareable implies level triggered, not shareable implies edge
 637  * triggered here.
 638  */
 639 int xen_bind_pirq_gsi_to_irq(unsigned gsi,
 640                              unsigned pirq, int shareable, char *name)
 641 {
 642         int irq = -1;
 643         struct physdev_irq irq_op;
 644
 645         mutex_lock(&irq_mapping_update_lock);
 646
 647         irq = find_irq_by_gsi(gsi);
 648         if (irq != -1) {
 649                 printk(KERN_INFO "xen_map_pirq_gsi: returning irq %d for gsi %u\n",
 650                        irq, gsi);
 651                 goto out;
 652         }
 653
 654         irq = xen_allocate_irq_gsi(gsi);
 655         if (irq < 0)
 656                 goto out;
 657
 658         irq_op.irq = irq;
 659         irq_op.vector = 0;
 660
 661         /* Only the privileged domain can do this. For non-priv, the pcifront
 662          * driver provides a PCI bus that does the call to do exactly
 663          * this in the priv domain. */
 664         if (xen_initial_domain() &&
 665             HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
 666                 xen_free_irq(irq);
 667                 irq = -ENOSPC;
 668                 goto out;
 669         }
 670
 671         xen_irq_info_pirq_init(irq, 0, pirq, gsi, irq_op.vector, DOMID_SELF,
 672                                shareable ? PIRQ_SHAREABLE : 0);
 673
 674         pirq_query_unmask(irq);
 675         /* We try to use the handler with the appropriate semantic for the
 676          * type of interrupt: if the interrupt is an edge triggered
 677          * interrupt we use handle_edge_irq.
 678          *
 679          * On the other hand if the interrupt is level triggered we use
 680          * handle_fasteoi_irq like the native code does for this kind of
 681          * interrupts.
 682          *
 683          * Depending on the Xen version, pirq_needs_eoi might return true
 684          * not only for level triggered interrupts but for edge triggered
 685          * interrupts too. In any case Xen always honors the eoi mechanism,
 686          * not injecting any more pirqs of the same kind if the first one
 687          * hasn't received an eoi yet. Therefore using the fasteoi handler
 688          * is the right choice either way.
 689          */
 690         if (shareable)
 691                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 692                                 handle_fasteoi_irq, name);
 693         else
 694                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 695                                 handle_edge_irq, name);
 696
 697 out:
 698         mutex_unlock(&irq_mapping_update_lock);
 699
 700         return irq;
 701 }
 702
 703 #ifdef CONFIG_PCI_MSI
 704 int xen_allocate_pirq_msi(struct pci_dev *dev, struct msi_desc *msidesc)
 705 {
 706         int rc;
 707         struct physdev_get_free_pirq op_get_free_pirq;
 708
 709         op_get_free_pirq.type = MAP_PIRQ_TYPE_MSI;
 710         rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
 711
 712         WARN_ONCE(rc == -ENOSYS,
 713                   "hypervisor does not support the PHYSDEVOP_get_free_pirq interface\n");
 714
 715         return rc ? -1 : op_get_free_pirq.pirq;
 716 }
 717
 718 int xen_bind_pirq_msi_to_irq(struct pci_dev *dev, struct msi_desc *msidesc,
 719                              int pirq, int vector, const char *name,
 720                              domid_t domid)
 721 {
 722         int irq, ret;
 723
 724         mutex_lock(&irq_mapping_update_lock);
 725
 726         irq = xen_allocate_irq_dynamic();
 727         if (irq < 0)
 728                 goto out;
 729
 730         irq_set_chip_and_handler_name(irq, &xen_pirq_chip, handle_edge_irq,
 731                         name);
 732
 733         xen_irq_info_pirq_init(irq, 0, pirq, 0, vector, domid, 0);
 734         ret = irq_set_msi_desc(irq, msidesc);
 735         if (ret < 0)
 736                 goto error_irq;
 737 out:
 738         mutex_unlock(&irq_mapping_update_lock);
 739         return irq;
 740 error_irq:
 741         mutex_unlock(&irq_mapping_update_lock);
 742         xen_free_irq(irq);
 743         return ret;
 744 }
 745 #endif
 746
 747 int xen_destroy_irq(int irq)
 748 {
 749         struct irq_desc *desc;
 750         struct physdev_unmap_pirq unmap_irq;
 751         struct irq_info *info = info_for_irq(irq);
 752         int rc = -ENOENT;
 753
 754         mutex_lock(&irq_mapping_update_lock);
 755
 756         desc = irq_to_desc(irq);
 757         if (!desc)
 758                 goto out;
 759
 760         if (xen_initial_domain()) {
 761                 unmap_irq.pirq = info->u.pirq.pirq;
 762                 unmap_irq.domid = info->u.pirq.domid;
 763                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
 764                 /* If another domain quits without making the pci_disable_msix
 765                  * call, the Xen hypervisor takes care of freeing the PIRQs
 766                  * (free_domain_pirqs).
 767                  */
 768                 if ((rc == -ESRCH && info->u.pirq.domid != DOMID_SELF))
 769                         printk(KERN_INFO "domain %d does not have %d anymore\n",
 770                                 info->u.pirq.domid, info->u.pirq.pirq);
 771                 else if (rc) {
 772                         printk(KERN_WARNING "unmap irq failed %d\n", rc);
 773                         goto out;
 774                 }
 775         }
 776
 777         xen_free_irq(irq);
 778
 779 out:
 780         mutex_unlock(&irq_mapping_update_lock);
 781         return rc;
 782 }
 783
 784 int xen_irq_from_pirq(unsigned pirq)
 785 {
 786         int irq;
 787
 788         struct irq_info *info;
 789
 790         mutex_lock(&irq_mapping_update_lock);
 791
 792         list_for_each_entry(info, &xen_irq_list_head, list) {
 793                 if (info->type != IRQT_PIRQ)
 794                         continue;
 795                 irq = info->irq;
 796                 if (info->u.pirq.pirq == pirq)
 797                         goto out;
 798         }
 799         irq = -1;
 800 out:
 801         mutex_unlock(&irq_mapping_update_lock);
 802
 803         return irq;
 804 }
 805
 806
 807 int xen_pirq_from_irq(unsigned irq)
 808 {
 809         return pirq_from_irq(irq);
 810 }
 811 EXPORT_SYMBOL_GPL(xen_pirq_from_irq);
 812 int bind_evtchn_to_irq(unsigned int evtchn)
 813 {
 814         int irq;
 815
 816         mutex_lock(&irq_mapping_update_lock);
 817
 818         irq = evtchn_to_irq[evtchn];
 819
 820         if (irq == -1) {
 821                 irq = xen_allocate_irq_dynamic();
 822                 if (irq == -1)
 823                         goto out;
 824
 825                 irq_set_chip_and_handler_name(irq, &xen_dynamic_chip,
 826                                               handle_edge_irq, "event");
 827
 828                 xen_irq_info_evtchn_init(irq, evtchn);
 829         }
 830
 831 out:
 832         mutex_unlock(&irq_mapping_update_lock);
 833
 834         return irq;
 835 }
 836 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
 837
 838 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
 839 {
 840         struct evtchn_bind_ipi bind_ipi;
 841         int evtchn, irq;
 842
 843         mutex_lock(&irq_mapping_update_lock);
 844
 845         irq = per_cpu(ipi_to_irq, cpu)[ipi];
 846
 847         if (irq == -1) {
 848                 irq = xen_allocate_irq_dynamic();
 849                 if (irq < 0)
 850                         goto out;
 851
 852                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 853                                               handle_percpu_irq, "ipi");
 854
 855                 bind_ipi.vcpu = cpu;
 856                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
 857                                                 &bind_ipi) != 0)
 858                         BUG();
 859                 evtchn = bind_ipi.port;
 860
 861                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
 862
 863                 bind_evtchn_to_cpu(evtchn, cpu);
 864         }
 865
 866  out:
 867         mutex_unlock(&irq_mapping_update_lock);
 868         return irq;
 869 }
 870
 871 static int bind_interdomain_evtchn_to_irq(unsigned int remote_domain,
 872                                           unsigned int remote_port)
 873 {
 874         struct evtchn_bind_interdomain bind_interdomain;
 875         int err;
 876
 877         bind_interdomain.remote_dom  = remote_domain;
 878         bind_interdomain.remote_port = remote_port;
 879
 880         err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
 881                                           &bind_interdomain);
 882
 883         return err ? : bind_evtchn_to_irq(bind_interdomain.local_port);
 884 }
 885
 886 static int find_virq(unsigned int virq, unsigned int cpu)
 887 {
 888         struct evtchn_status status;
 889         int port, rc = -ENOENT;
 890
 891         memset(&status, 0, sizeof(status));
 892         for (port = 0; port <= NR_EVENT_CHANNELS; port++) {
 893                 status.dom = DOMID_SELF;
 894                 status.port = port;
 895                 rc = HYPERVISOR_event_channel_op(EVTCHNOP_status, &status);
 896                 if (rc < 0)
 897                         continue;
 898                 if (status.status != EVTCHNSTAT_virq)
 899                         continue;
 900                 if (status.u.virq == virq && status.vcpu == cpu) {
 901                         rc = port;
 902                         break;
 903                 }
 904         }
 905         return rc;
 906 }
 907
 908 int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
 909 {
 910         struct evtchn_bind_virq bind_virq;
 911         int evtchn, irq, ret;
 912
 913         mutex_lock(&irq_mapping_update_lock);
 914
 915         irq = per_cpu(virq_to_irq, cpu)[virq];
 916
 917         if (irq == -1) {
 918                 irq = xen_allocate_irq_dynamic();
 919                 if (irq == -1)
 920                         goto out;
 921
 922                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 923                                               handle_percpu_irq, "virq");
 924
 925                 bind_virq.virq = virq;
 926                 bind_virq.vcpu = cpu;
 927                 ret = HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 928                                                 &bind_virq);
 929                 if (ret == 0)
 930                         evtchn = bind_virq.port;
 931                 else {
 932                         if (ret == -EEXIST)
 933                                 ret = find_virq(virq, cpu);
 934                         BUG_ON(ret < 0);
 935                         evtchn = ret;
 936                 }
 937
 938                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
 939
 940                 bind_evtchn_to_cpu(evtchn, cpu);
 941         }
 942
 943 out:
 944         mutex_unlock(&irq_mapping_update_lock);
 945
 946         return irq;
 947 }
 948
 949 static void unbind_from_irq(unsigned int irq)
 950 {
 951         struct evtchn_close close;
 952         int evtchn = evtchn_from_irq(irq);
 953         struct irq_info *info = irq_get_handler_data(irq);
 954
 955         mutex_lock(&irq_mapping_update_lock);
 956
 957         if (info->refcnt > 0) {
 958                 info->refcnt--;
 959                 if (info->refcnt != 0)
 960                         goto done;
 961         }
 962
 963         if (VALID_EVTCHN(evtchn)) {
 964                 close.port = evtchn;
 965                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 966                         BUG();
 967
 968                 switch (type_from_irq(irq)) {
 969                 case IRQT_VIRQ:
 970                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
 971                                 [virq_from_irq(irq)] = -1;
 972                         break;
 973                 case IRQT_IPI:
 974                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
 975                                 [ipi_from_irq(irq)] = -1;
 976                         break;
 977                 default:
 978                         break;
 979                 }
 980
 981                 /* Closed ports are implicitly re-bound to VCPU0. */
 982                 bind_evtchn_to_cpu(evtchn, 0);
 983
 984                 evtchn_to_irq[evtchn] = -1;
 985         }
 986
 987         BUG_ON(info_for_irq(irq)->type == IRQT_UNBOUND);
 988
 989         xen_free_irq(irq);
 990
 991  done:
 992         mutex_unlock(&irq_mapping_update_lock);
 993 }
 994
 995 int bind_evtchn_to_irqhandler(unsigned int evtchn,
 996                               irq_handler_t handler,
 997                               unsigned long irqflags,
 998                               const char *devname, void *dev_id)
 999 {
1000         int irq, retval;
1001
1002         irq = bind_evtchn_to_irq(evtchn);
1003         if (irq < 0)
1004                 return irq;
1005         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1006         if (retval != 0) {
1007                 unbind_from_irq(irq);
1008                 return retval;
1009         }
1010
1011         return irq;
1012 }
1013 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
1014
1015 int bind_interdomain_evtchn_to_irqhandler(unsigned int remote_domain,
1016                                           unsigned int remote_port,
1017                                           irq_handler_t handler,
1018                                           unsigned long irqflags,
1019                                           const char *devname,
1020                                           void *dev_id)
1021 {
1022         int irq, retval;
1023
1024         irq = bind_interdomain_evtchn_to_irq(remote_domain, remote_port);
1025         if (irq < 0)
1026                 return irq;
1027
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_interdomain_evtchn_to_irqhandler);
1037
1038 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
1039                             irq_handler_t handler,
1040                             unsigned long irqflags, const char *devname, void *dev_id)
1041 {
1042         int irq, retval;
1043
1044         irq = bind_virq_to_irq(virq, cpu);
1045         if (irq < 0)
1046                 return irq;
1047         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1048         if (retval != 0) {
1049                 unbind_from_irq(irq);
1050                 return retval;
1051         }
1052
1053         return irq;
1054 }
1055 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1056
1057 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1058                            unsigned int cpu,
1059                            irq_handler_t handler,
1060                            unsigned long irqflags,
1061                            const char *devname,
1062                            void *dev_id)
1063 {
1064         int irq, retval;
1065
1066         irq = bind_ipi_to_irq(ipi, cpu);
1067         if (irq < 0)
1068                 return irq;
1069
1070         irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME | IRQF_EARLY_RESUME;
1071         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1072         if (retval != 0) {
1073                 unbind_from_irq(irq);
1074                 return retval;
1075         }
1076
1077         return irq;
1078 }
1079
1080 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1081 {
1082         free_irq(irq, dev_id);
1083         unbind_from_irq(irq);
1084 }
1085 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1086
1087 int evtchn_make_refcounted(unsigned int evtchn)
1088 {
1089         int irq = evtchn_to_irq[evtchn];
1090         struct irq_info *info;
1091
1092         if (irq == -1)
1093                 return -ENOENT;
1094
1095         info = irq_get_handler_data(irq);
1096
1097         if (!info)
1098                 return -ENOENT;
1099
1100         WARN_ON(info->refcnt != -1);
1101
1102         info->refcnt = 1;
1103
1104         return 0;
1105 }
1106 EXPORT_SYMBOL_GPL(evtchn_make_refcounted);
1107
1108 int evtchn_get(unsigned int evtchn)
1109 {
1110         int irq;
1111         struct irq_info *info;
1112         int err = -ENOENT;
1113
1114         if (evtchn >= NR_EVENT_CHANNELS)
1115                 return -EINVAL;
1116
1117         mutex_lock(&irq_mapping_update_lock);
1118
1119         irq = evtchn_to_irq[evtchn];
1120         if (irq == -1)
1121                 goto done;
1122
1123         info = irq_get_handler_data(irq);
1124
1125         if (!info)
1126                 goto done;
1127
1128         err = -EINVAL;
1129         if (info->refcnt <= 0)
1130                 goto done;
1131
1132         info->refcnt++;
1133         err = 0;
1134  done:
1135         mutex_unlock(&irq_mapping_update_lock);
1136
1137         return err;
1138 }
1139 EXPORT_SYMBOL_GPL(evtchn_get);
1140
1141 void evtchn_put(unsigned int evtchn)
1142 {
1143         int irq = evtchn_to_irq[evtchn];
1144         if (WARN_ON(irq == -1))
1145                 return;
1146         unbind_from_irq(irq);
1147 }
1148 EXPORT_SYMBOL_GPL(evtchn_put);
1149
1150 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1151 {
1152         int irq = per_cpu(ipi_to_irq, cpu)[vector];
1153         BUG_ON(irq < 0);
1154         notify_remote_via_irq(irq);
1155 }
1156
1157 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
1158 {
1159         struct shared_info *sh = HYPERVISOR_shared_info;
1160         int cpu = smp_processor_id();
1161         unsigned long *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu);
1162         int i;
1163         unsigned long flags;
1164         static DEFINE_SPINLOCK(debug_lock);
1165         struct vcpu_info *v;
1166
1167         spin_lock_irqsave(&debug_lock, flags);
1168
1169         printk("\nvcpu %d\n  ", cpu);
1170
1171         for_each_online_cpu(i) {
1172                 int pending;
1173                 v = per_cpu(xen_vcpu, i);
1174                 pending = (get_irq_regs() && i == cpu)
1175                         ? xen_irqs_disabled(get_irq_regs())
1176                         : v->evtchn_upcall_mask;
1177                 printk("%d: masked=%d pending=%d event_sel %0*lx\n  ", i,
1178                        pending, v->evtchn_upcall_pending,
1179                        (int)(sizeof(v->evtchn_pending_sel)*2),
1180                        v->evtchn_pending_sel);
1181         }
1182         v = per_cpu(xen_vcpu, cpu);
1183
1184         printk("\npending:\n   ");
1185         for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
1186                 printk("%0*lx%s", (int)sizeof(sh->evtchn_pending[0])*2,
1187                        sh->evtchn_pending[i],
1188                        i % 8 == 0 ? "\n   " : " ");
1189         printk("\nglobal mask:\n   ");
1190         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1191                 printk("%0*lx%s",
1192                        (int)(sizeof(sh->evtchn_mask[0])*2),
1193                        sh->evtchn_mask[i],
1194                        i % 8 == 0 ? "\n   " : " ");
1195
1196         printk("\nglobally unmasked:\n   ");
1197         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1198                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1199                        sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
1200                        i % 8 == 0 ? "\n   " : " ");
1201
1202         printk("\nlocal cpu%d mask:\n   ", cpu);
1203         for (i = (NR_EVENT_CHANNELS/BITS_PER_LONG)-1; i >= 0; i--)
1204                 printk("%0*lx%s", (int)(sizeof(cpu_evtchn[0])*2),
1205                        cpu_evtchn[i],
1206                        i % 8 == 0 ? "\n   " : " ");
1207
1208         printk("\nlocally unmasked:\n   ");
1209         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
1210                 unsigned long pending = sh->evtchn_pending[i]
1211                         & ~sh->evtchn_mask[i]
1212                         & cpu_evtchn[i];
1213                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1214                        pending, i % 8 == 0 ? "\n   " : " ");
1215         }
1216
1217         printk("\npending list:\n");
1218         for (i = 0; i < NR_EVENT_CHANNELS; i++) {
1219                 if (sync_test_bit(i, sh->evtchn_pending)) {
1220                         int word_idx = i / BITS_PER_LONG;
1221                         printk("  %d: event %d -> irq %d%s%s%s\n",
1222                                cpu_from_evtchn(i), i,
1223                                evtchn_to_irq[i],
1224                                sync_test_bit(word_idx, &v->evtchn_pending_sel)
1225                                              ? "" : " l2-clear",
1226                                !sync_test_bit(i, sh->evtchn_mask)
1227                                              ? "" : " globally-masked",
1228                                sync_test_bit(i, cpu_evtchn)
1229                                              ? "" : " locally-masked");
1230                 }
1231         }
1232
1233         spin_unlock_irqrestore(&debug_lock, flags);
1234
1235         return IRQ_HANDLED;
1236 }
1237
1238 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
1239 static DEFINE_PER_CPU(unsigned int, current_word_idx);
1240 static DEFINE_PER_CPU(unsigned int, current_bit_idx);
1241
1242 /*
1243  * Mask out the i least significant bits of w
1244  */
1245 #define MASK_LSBS(w, i) (w & ((~0UL) << i))
1246
1247 /*
1248  * Search the CPUs pending events bitmasks.  For each one found, map
1249  * the event number to an irq, and feed it into do_IRQ() for
1250  * handling.
1251  *
1252  * Xen uses a two-level bitmap to speed searching.  The first level is
1253  * a bitset of words which contain pending event bits.  The second
1254  * level is a bitset of pending events themselves.
1255  */
1256 static void __xen_evtchn_do_upcall(void)
1257 {
1258         int start_word_idx, start_bit_idx;
1259         int word_idx, bit_idx;
1260         int i;
1261         int cpu = get_cpu();
1262         struct shared_info *s = HYPERVISOR_shared_info;
1263         struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1264         unsigned count;
1265
1266         do {
1267                 unsigned long pending_words;
1268
1269                 vcpu_info->evtchn_upcall_pending = 0;
1270
1271                 if (__this_cpu_inc_return(xed_nesting_count) - 1)
1272                         goto out;
1273
1274 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
1275                 /* Clear master flag /before/ clearing selector flag. */
1276                 wmb();
1277 #endif
1278                 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
1279
1280                 start_word_idx = __this_cpu_read(current_word_idx);
1281                 start_bit_idx = __this_cpu_read(current_bit_idx);
1282
1283                 word_idx = start_word_idx;
1284
1285                 for (i = 0; pending_words != 0; i++) {
1286                         unsigned long pending_bits;
1287                         unsigned long words;
1288
1289                         words = MASK_LSBS(pending_words, word_idx);
1290
1291                         /*
1292                          * If we masked out all events, wrap to beginning.
1293                          */
1294                         if (words == 0) {
1295                                 word_idx = 0;
1296                                 bit_idx = 0;
1297                                 continue;
1298                         }
1299                         word_idx = __ffs(words);
1300
1301                         pending_bits = active_evtchns(cpu, s, word_idx);
1302                         bit_idx = 0; /* usually scan entire word from start */
1303                         if (word_idx == start_word_idx) {
1304                                 /* We scan the starting word in two parts */
1305                                 if (i == 0)
1306                                         /* 1st time: start in the middle */
1307                                         bit_idx = start_bit_idx;
1308                                 else
1309                                         /* 2nd time: mask bits done already */
1310                                         bit_idx &= (1UL << start_bit_idx) - 1;
1311                         }
1312
1313                         do {
1314                                 unsigned long bits;
1315                                 int port, irq;
1316                                 struct irq_desc *desc;
1317
1318                                 bits = MASK_LSBS(pending_bits, bit_idx);
1319
1320                                 /* If we masked out all events, move on. */
1321                                 if (bits == 0)
1322                                         break;
1323
1324                                 bit_idx = __ffs(bits);
1325
1326                                 /* Process port. */
1327                                 port = (word_idx * BITS_PER_LONG) + bit_idx;
1328                                 irq = evtchn_to_irq[port];
1329
1330                                 if (irq != -1) {
1331                                         desc = irq_to_desc(irq);
1332                                         if (desc)
1333                                                 generic_handle_irq_desc(irq, desc);
1334                                 }
1335
1336                                 bit_idx = (bit_idx + 1) % BITS_PER_LONG;
1337
1338                                 /* Next caller starts at last processed + 1 */
1339                                 __this_cpu_write(current_word_idx,
1340                                                  bit_idx ? word_idx :
1341                                                  (word_idx+1) % BITS_PER_LONG);
1342                                 __this_cpu_write(current_bit_idx, bit_idx);
1343                         } while (bit_idx != 0);
1344
1345                         /* Scan start_l1i twice; all others once. */
1346                         if ((word_idx != start_word_idx) || (i != 0))
1347                                 pending_words &= ~(1UL << word_idx);
1348
1349                         word_idx = (word_idx + 1) % BITS_PER_LONG;
1350                 }
1351
1352                 BUG_ON(!irqs_disabled());
1353
1354                 count = __this_cpu_read(xed_nesting_count);
1355                 __this_cpu_write(xed_nesting_count, 0);
1356         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
1357
1358 out:
1359
1360         put_cpu();
1361 }
1362
1363 void xen_evtchn_do_upcall(struct pt_regs *regs)
1364 {
1365         struct pt_regs *old_regs = set_irq_regs(regs);
1366
1367         exit_idle();
1368         irq_enter();
1369
1370         __xen_evtchn_do_upcall();
1371
1372         irq_exit();
1373         set_irq_regs(old_regs);
1374 }
1375
1376 void xen_hvm_evtchn_do_upcall(void)
1377 {
1378         __xen_evtchn_do_upcall();
1379 }
1380 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1381
1382 /* Rebind a new event channel to an existing irq. */
1383 void rebind_evtchn_irq(int evtchn, int irq)
1384 {
1385         struct irq_info *info = info_for_irq(irq);
1386
1387         /* Make sure the irq is masked, since the new event channel
1388            will also be masked. */
1389         disable_irq(irq);
1390
1391         mutex_lock(&irq_mapping_update_lock);
1392
1393         /* After resume the irq<->evtchn mappings are all cleared out */
1394         BUG_ON(evtchn_to_irq[evtchn] != -1);
1395         /* Expect irq to have been bound before,
1396            so there should be a proper type */
1397         BUG_ON(info->type == IRQT_UNBOUND);
1398
1399         xen_irq_info_evtchn_init(irq, evtchn);
1400
1401         mutex_unlock(&irq_mapping_update_lock);
1402
1403         /* new event channels are always bound to cpu 0 */
1404         irq_set_affinity(irq, cpumask_of(0));
1405
1406         /* Unmask the event channel. */
1407         enable_irq(irq);
1408 }
1409
1410 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1411 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1412 {
1413         struct evtchn_bind_vcpu bind_vcpu;
1414         int evtchn = evtchn_from_irq(irq);
1415
1416         if (!VALID_EVTCHN(evtchn))
1417                 return -1;
1418
1419         /*
1420          * Events delivered via platform PCI interrupts are always
1421          * routed to vcpu 0 and hence cannot be rebound.
1422          */
1423         if (xen_hvm_domain() && !xen_have_vector_callback)
1424                 return -1;
1425
1426         /* Send future instances of this interrupt to other vcpu. */
1427         bind_vcpu.port = evtchn;
1428         bind_vcpu.vcpu = tcpu;
1429
1430         /*
1431          * If this fails, it usually just indicates that we're dealing with a
1432          * virq or IPI channel, which don't actually need to be rebound. Ignore
1433          * it, but don't do the xenlinux-level rebind in that case.
1434          */
1435         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1436                 bind_evtchn_to_cpu(evtchn, tcpu);
1437
1438         return 0;
1439 }
1440
1441 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1442                             bool force)
1443 {
1444         unsigned tcpu = cpumask_first(dest);
1445
1446         return rebind_irq_to_cpu(data->irq, tcpu);
1447 }
1448
1449 int resend_irq_on_evtchn(unsigned int irq)
1450 {
1451         int masked, evtchn = evtchn_from_irq(irq);
1452         struct shared_info *s = HYPERVISOR_shared_info;
1453
1454         if (!VALID_EVTCHN(evtchn))
1455                 return 1;
1456
1457         masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
1458         sync_set_bit(evtchn, s->evtchn_pending);
1459         if (!masked)
1460                 unmask_evtchn(evtchn);
1461
1462         return 1;
1463 }
1464
1465 static void enable_dynirq(struct irq_data *data)
1466 {
1467         int evtchn = evtchn_from_irq(data->irq);
1468
1469         if (VALID_EVTCHN(evtchn))
1470                 unmask_evtchn(evtchn);
1471 }
1472
1473 static void disable_dynirq(struct irq_data *data)
1474 {
1475         int evtchn = evtchn_from_irq(data->irq);
1476
1477         if (VALID_EVTCHN(evtchn))
1478                 mask_evtchn(evtchn);
1479 }
1480
1481 static void ack_dynirq(struct irq_data *data)
1482 {
1483         int evtchn = evtchn_from_irq(data->irq);
1484
1485         irq_move_irq(data);
1486
1487         if (VALID_EVTCHN(evtchn))
1488                 clear_evtchn(evtchn);
1489 }
1490
1491 static void mask_ack_dynirq(struct irq_data *data)
1492 {
1493         disable_dynirq(data);
1494         ack_dynirq(data);
1495 }
1496
1497 static int retrigger_dynirq(struct irq_data *data)
1498 {
1499         int evtchn = evtchn_from_irq(data->irq);
1500         struct shared_info *sh = HYPERVISOR_shared_info;
1501         int ret = 0;
1502
1503         if (VALID_EVTCHN(evtchn)) {
1504                 int masked;
1505
1506                 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
1507                 sync_set_bit(evtchn, sh->evtchn_pending);
1508                 if (!masked)
1509                         unmask_evtchn(evtchn);
1510                 ret = 1;
1511         }
1512
1513         return ret;
1514 }
1515
1516 static void restore_pirqs(void)
1517 {
1518         int pirq, rc, irq, gsi;
1519         struct physdev_map_pirq map_irq;
1520         struct irq_info *info;
1521
1522         list_for_each_entry(info, &xen_irq_list_head, list) {
1523                 if (info->type != IRQT_PIRQ)
1524                         continue;
1525
1526                 pirq = info->u.pirq.pirq;
1527                 gsi = info->u.pirq.gsi;
1528                 irq = info->irq;
1529
1530                 /* save/restore of PT devices doesn't work, so at this point the
1531                  * only devices present are GSI based emulated devices */
1532                 if (!gsi)
1533                         continue;
1534
1535                 map_irq.domid = DOMID_SELF;
1536                 map_irq.type = MAP_PIRQ_TYPE_GSI;
1537                 map_irq.index = gsi;
1538                 map_irq.pirq = pirq;
1539
1540                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1541                 if (rc) {
1542                         printk(KERN_WARNING "xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1543                                         gsi, irq, pirq, rc);
1544                         xen_free_irq(irq);
1545                         continue;
1546                 }
1547
1548                 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1549
1550                 __startup_pirq(irq);
1551         }
1552 }
1553
1554 static void restore_cpu_virqs(unsigned int cpu)
1555 {
1556         struct evtchn_bind_virq bind_virq;
1557         int virq, irq, evtchn;
1558
1559         for (virq = 0; virq < NR_VIRQS; virq++) {
1560                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1561                         continue;
1562
1563                 BUG_ON(virq_from_irq(irq) != virq);
1564
1565                 /* Get a new binding from Xen. */
1566                 bind_virq.virq = virq;
1567                 bind_virq.vcpu = cpu;
1568                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1569                                                 &bind_virq) != 0)
1570                         BUG();
1571                 evtchn = bind_virq.port;
1572
1573                 /* Record the new mapping. */
1574                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
1575                 bind_evtchn_to_cpu(evtchn, cpu);
1576         }
1577 }
1578
1579 static void restore_cpu_ipis(unsigned int cpu)
1580 {
1581         struct evtchn_bind_ipi bind_ipi;
1582         int ipi, irq, evtchn;
1583
1584         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1585                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1586                         continue;
1587
1588                 BUG_ON(ipi_from_irq(irq) != ipi);
1589
1590                 /* Get a new binding from Xen. */
1591                 bind_ipi.vcpu = cpu;
1592                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1593                                                 &bind_ipi) != 0)
1594                         BUG();
1595                 evtchn = bind_ipi.port;
1596
1597                 /* Record the new mapping. */
1598                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
1599                 bind_evtchn_to_cpu(evtchn, cpu);
1600         }
1601 }
1602
1603 /* Clear an irq's pending state, in preparation for polling on it */
1604 void xen_clear_irq_pending(int irq)
1605 {
1606         int evtchn = evtchn_from_irq(irq);
1607
1608         if (VALID_EVTCHN(evtchn))
1609                 clear_evtchn(evtchn);
1610 }
1611 EXPORT_SYMBOL(xen_clear_irq_pending);
1612 void xen_set_irq_pending(int irq)
1613 {
1614         int evtchn = evtchn_from_irq(irq);
1615
1616         if (VALID_EVTCHN(evtchn))
1617                 set_evtchn(evtchn);
1618 }
1619
1620 bool xen_test_irq_pending(int irq)
1621 {
1622         int evtchn = evtchn_from_irq(irq);
1623         bool ret = false;
1624
1625         if (VALID_EVTCHN(evtchn))
1626                 ret = test_evtchn(evtchn);
1627
1628         return ret;
1629 }
1630
1631 /* Poll waiting for an irq to become pending with timeout.  In the usual case,
1632  * the irq will be disabled so it won't deliver an interrupt. */
1633 void xen_poll_irq_timeout(int irq, u64 timeout)
1634 {
1635         evtchn_port_t evtchn = evtchn_from_irq(irq);
1636
1637         if (VALID_EVTCHN(evtchn)) {
1638                 struct sched_poll poll;
1639
1640                 poll.nr_ports = 1;
1641                 poll.timeout = timeout;
1642                 set_xen_guest_handle(poll.ports, &evtchn);
1643
1644                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1645                         BUG();
1646         }
1647 }
1648 EXPORT_SYMBOL(xen_poll_irq_timeout);
1649 /* Poll waiting for an irq to become pending.  In the usual case, the
1650  * irq will be disabled so it won't deliver an interrupt. */
1651 void xen_poll_irq(int irq)
1652 {
1653         xen_poll_irq_timeout(irq, 0 /* no timeout */);
1654 }
1655
1656 /* Check whether the IRQ line is shared with other guests. */
1657 int xen_test_irq_shared(int irq)
1658 {
1659         struct irq_info *info = info_for_irq(irq);
1660         struct physdev_irq_status_query irq_status = { .irq = info->u.pirq.pirq };
1661
1662         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
1663                 return 0;
1664         return !(irq_status.flags & XENIRQSTAT_shared);
1665 }
1666 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
1667
1668 void xen_irq_resume(void)
1669 {
1670         unsigned int cpu, evtchn;
1671         struct irq_info *info;
1672
1673         init_evtchn_cpu_bindings();
1674
1675         /* New event-channel space is not 'live' yet. */
1676         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1677                 mask_evtchn(evtchn);
1678
1679         /* No IRQ <-> event-channel mappings. */
1680         list_for_each_entry(info, &xen_irq_list_head, list)
1681                 info->evtchn = 0; /* zap event-channel binding */
1682
1683         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1684                 evtchn_to_irq[evtchn] = -1;
1685
1686         for_each_possible_cpu(cpu) {
1687                 restore_cpu_virqs(cpu);
1688                 restore_cpu_ipis(cpu);
1689         }
1690
1691         restore_pirqs();
1692 }
1693
1694 static struct irq_chip xen_dynamic_chip __read_mostly = {
1695         .name                   = "xen-dyn",
1696
1697         .irq_disable            = disable_dynirq,
1698         .irq_mask               = disable_dynirq,
1699         .irq_unmask             = enable_dynirq,
1700
1701         .irq_ack                = ack_dynirq,
1702         .irq_mask_ack           = mask_ack_dynirq,
1703
1704         .irq_set_affinity       = set_affinity_irq,
1705         .irq_retrigger          = retrigger_dynirq,
1706 };
1707
1708 static struct irq_chip xen_pirq_chip __read_mostly = {
1709         .name                   = "xen-pirq",
1710
1711         .irq_startup            = startup_pirq,
1712         .irq_shutdown           = shutdown_pirq,
1713         .irq_enable             = enable_pirq,
1714         .irq_disable            = disable_pirq,
1715
1716         .irq_mask               = disable_dynirq,
1717         .irq_unmask             = enable_dynirq,
1718
1719         .irq_ack                = eoi_pirq,
1720         .irq_eoi                = eoi_pirq,
1721         .irq_mask_ack           = mask_ack_pirq,
1722
1723         .irq_set_affinity       = set_affinity_irq,
1724
1725         .irq_retrigger          = retrigger_dynirq,
1726 };
1727
1728 static struct irq_chip xen_percpu_chip __read_mostly = {
1729         .name                   = "xen-percpu",
1730
1731         .irq_disable            = disable_dynirq,
1732         .irq_mask               = disable_dynirq,
1733         .irq_unmask             = enable_dynirq,
1734
1735         .irq_ack                = ack_dynirq,
1736 };
1737
1738 int xen_set_callback_via(uint64_t via)
1739 {
1740         struct xen_hvm_param a;
1741         a.domid = DOMID_SELF;
1742         a.index = HVM_PARAM_CALLBACK_IRQ;
1743         a.value = via;
1744         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1745 }
1746 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1747
1748 #ifdef CONFIG_XEN_PVHVM
1749 /* Vector callbacks are better than PCI interrupts to receive event
1750  * channel notifications because we can receive vector callbacks on any
1751  * vcpu and we don't need PCI support or APIC interactions. */
1752 void xen_callback_vector(void)
1753 {
1754         int rc;
1755         uint64_t callback_via;
1756         if (xen_have_vector_callback) {
1757                 callback_via = HVM_CALLBACK_VECTOR(XEN_HVM_EVTCHN_CALLBACK);
1758                 rc = xen_set_callback_via(callback_via);
1759                 if (rc) {
1760                         printk(KERN_ERR "Request for Xen HVM callback vector"
1761                                         " failed.\n");
1762                         xen_have_vector_callback = 0;
1763                         return;
1764                 }
1765                 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1766                                 "enabled\n");
1767                 /* in the restore case the vector has already been allocated */
1768                 if (!test_bit(XEN_HVM_EVTCHN_CALLBACK, used_vectors))
1769                         alloc_intr_gate(XEN_HVM_EVTCHN_CALLBACK, xen_hvm_callback_vector);
1770         }
1771 }
1772 #else
1773 void xen_callback_vector(void) {}
1774 #endif
1775
1776 void __init xen_init_IRQ(void)
1777 {
1778         int i, rc;
1779
1780         evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1781                                     GFP_KERNEL);
1782         BUG_ON(!evtchn_to_irq);
1783         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1784                 evtchn_to_irq[i] = -1;
1785
1786         init_evtchn_cpu_bindings();
1787
1788         /* No event channels are 'live' right now. */
1789         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1790                 mask_evtchn(i);
1791
1792         pirq_needs_eoi = pirq_needs_eoi_flag;
1793
1794         if (xen_hvm_domain()) {
1795                 xen_callback_vector();
1796                 native_init_IRQ();
1797                 /* pci_xen_hvm_init must be called after native_init_IRQ so that
1798                  * __acpi_register_gsi can point at the right function */
1799                 pci_xen_hvm_init();
1800         } else {
1801                 struct physdev_pirq_eoi_gmfn eoi_gmfn;
1802
1803                 irq_ctx_init(smp_processor_id());
1804                 if (xen_initial_domain())
1805                         pci_xen_initial_domain();
1806
1807                 pirq_eoi_map = (void *)__get_free_page(GFP_KERNEL|__GFP_ZERO);
1808                 eoi_gmfn.gmfn = virt_to_mfn(pirq_eoi_map);
1809                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_pirq_eoi_gmfn_v2, &eoi_gmfn);
1810                 if (rc != 0) {
1811                         free_page((unsigned long) pirq_eoi_map);
1812                         pirq_eoi_map = NULL;
1813                 } else
1814                         pirq_needs_eoi = pirq_check_eoi_map;
1815         }
1816 }