drivers/xen/events.c

   1 /*
   2  * Xen event channels
   3  *
   4  * Xen models interrupts with abstract event channels.  Because each
   5  * domain gets 1024 event channels, but NR_IRQ is not that large, we
   6  * must dynamically map irqs<->event channels.  The event channels
   7  * interface with the rest of the kernel by defining a xen interrupt
   8  * chip.  When an event is received, it is mapped to an irq and sent
   9  * through the normal interrupt processing path.
  10  *
  11  * There are four kinds of events which can be mapped to an event
  12  * channel:
  13  *
  14  * 1. Inter-domain notifications.  This includes all the virtual
  15  *    device events, since they're driven by front-ends in another domain
  16  *    (typically dom0).
  17  * 2. VIRQs, typically used for timers.  These are per-cpu events.
  18  * 3. IPIs.
  19  * 4. PIRQs - Hardware interrupts.
  20  *
  21  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
  22  */
  23
  24 #include <linux/linkage.h>
  25 #include <linux/interrupt.h>
  26 #include <linux/irq.h>
  27 #include <linux/module.h>
  28 #include <linux/string.h>
  29 #include <linux/bootmem.h>
  30 #include <linux/slab.h>
  31 #include <linux/irqnr.h>
  32 #include <linux/pci.h>
  33
  34 #include <asm/desc.h>
  35 #include <asm/ptrace.h>
  36 #include <asm/irq.h>
  37 #include <asm/idle.h>
  38 #include <asm/io_apic.h>
  39 #include <asm/sync_bitops.h>
  40 #include <asm/xen/pci.h>
  41 #include <asm/xen/hypercall.h>
  42 #include <asm/xen/hypervisor.h>
  43
  44 #include <xen/xen.h>
  45 #include <xen/hvm.h>
  46 #include <xen/xen-ops.h>
  47 #include <xen/events.h>
  48 #include <xen/interface/xen.h>
  49 #include <xen/interface/event_channel.h>
  50 #include <xen/interface/hvm/hvm_op.h>
  51 #include <xen/interface/hvm/params.h>
  52
  53 /*
  54  * This lock protects updates to the following mapping and reference-count
  55  * arrays. The lock does not need to be acquired to read the mapping tables.
  56  */
  57 static DEFINE_SPINLOCK(irq_mapping_update_lock);
  58
  59 static LIST_HEAD(xen_irq_list_head);
  60
  61 /* IRQ <-> VIRQ mapping. */
  62 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
  63
  64 /* IRQ <-> IPI mapping */
  65 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
  66
  67 /* Interrupt types. */
  68 enum xen_irq_type {
  69         IRQT_UNBOUND = 0,
  70         IRQT_PIRQ,
  71         IRQT_VIRQ,
  72         IRQT_IPI,
  73         IRQT_EVTCHN
  74 };
  75
  76 /*
  77  * Packed IRQ information:
  78  * type - enum xen_irq_type
  79  * event channel - irq->event channel mapping
  80  * cpu - cpu this event channel is bound to
  81  * index - type-specific information:
  82  *    PIRQ - vector, with MSB being "needs EIO", or physical IRQ of the HVM
  83  *           guest, or GSI (real passthrough IRQ) of the device.
  84  *    VIRQ - virq number
  85  *    IPI - IPI vector
  86  *    EVTCHN -
  87  */
  88 struct irq_info
  89 {
  90         struct list_head list;
  91         enum xen_irq_type type; /* type */
  92         unsigned irq;
  93         unsigned short evtchn;  /* event channel */
  94         unsigned short cpu;     /* cpu bound */
  95
  96         union {
  97                 unsigned short virq;
  98                 enum ipi_vector ipi;
  99                 struct {
 100                         unsigned short pirq;
 101                         unsigned short gsi;
 102                         unsigned char vector;
 103                         unsigned char flags;
 104                         uint16_t domid;
 105                 } pirq;
 106         } u;
 107 };
 108 #define PIRQ_NEEDS_EOI  (1 << 0)
 109 #define PIRQ_SHAREABLE  (1 << 1)
 110
 111 static int *evtchn_to_irq;
 112
 113 static DEFINE_PER_CPU(unsigned long [NR_EVENT_CHANNELS/BITS_PER_LONG],
 114                       cpu_evtchn_mask);
 115
 116 /* Xen will never allocate port zero for any purpose. */
 117 #define VALID_EVTCHN(chn)       ((chn) != 0)
 118
 119 static struct irq_chip xen_dynamic_chip;
 120 static struct irq_chip xen_percpu_chip;
 121 static struct irq_chip xen_pirq_chip;
 122
 123 /* Get info for IRQ */
 124 static struct irq_info *info_for_irq(unsigned irq)
 125 {
 126         return irq_get_handler_data(irq);
 127 }
 128
 129 /* Constructors for packed IRQ information. */
 130 static void xen_irq_info_common_init(struct irq_info *info,
 131                                      unsigned irq,
 132                                      enum xen_irq_type type,
 133                                      unsigned short evtchn,
 134                                      unsigned short cpu)
 135 {
 136
 137         BUG_ON(info->type != IRQT_UNBOUND && info->type != type);
 138
 139         info->type = type;
 140         info->irq = irq;
 141         info->evtchn = evtchn;
 142         info->cpu = cpu;
 143
 144         evtchn_to_irq[evtchn] = irq;
 145 }
 146
 147 static void xen_irq_info_evtchn_init(unsigned irq,
 148                                      unsigned short evtchn)
 149 {
 150         struct irq_info *info = info_for_irq(irq);
 151
 152         xen_irq_info_common_init(info, irq, IRQT_EVTCHN, evtchn, 0);
 153 }
 154
 155 static void xen_irq_info_ipi_init(unsigned cpu,
 156                                   unsigned irq,
 157                                   unsigned short evtchn,
 158                                   enum ipi_vector ipi)
 159 {
 160         struct irq_info *info = info_for_irq(irq);
 161
 162         xen_irq_info_common_init(info, irq, IRQT_IPI, evtchn, 0);
 163
 164         info->u.ipi = ipi;
 165
 166         per_cpu(ipi_to_irq, cpu)[ipi] = irq;
 167 }
 168
 169 static void xen_irq_info_virq_init(unsigned cpu,
 170                                    unsigned irq,
 171                                    unsigned short evtchn,
 172                                    unsigned short virq)
 173 {
 174         struct irq_info *info = info_for_irq(irq);
 175
 176         xen_irq_info_common_init(info, irq, IRQT_VIRQ, evtchn, 0);
 177
 178         info->u.virq = virq;
 179
 180         per_cpu(virq_to_irq, cpu)[virq] = irq;
 181 }
 182
 183 static void xen_irq_info_pirq_init(unsigned irq,
 184                                    unsigned short evtchn,
 185                                    unsigned short pirq,
 186                                    unsigned short gsi,
 187                                    unsigned short vector,
 188                                    uint16_t domid,
 189                                    unsigned char flags)
 190 {
 191         struct irq_info *info = info_for_irq(irq);
 192
 193         xen_irq_info_common_init(info, irq, IRQT_PIRQ, evtchn, 0);
 194
 195         info->u.pirq.pirq = pirq;
 196         info->u.pirq.gsi = gsi;
 197         info->u.pirq.vector = vector;
 198         info->u.pirq.domid = domid;
 199         info->u.pirq.flags = flags;
 200 }
 201
 202 /*
 203  * Accessors for packed IRQ information.
 204  */
 205 static unsigned int evtchn_from_irq(unsigned irq)
 206 {
 207         if (unlikely(WARN(irq < 0 || irq >= nr_irqs, "Invalid irq %d!\n", irq)))
 208                 return 0;
 209
 210         return info_for_irq(irq)->evtchn;
 211 }
 212
 213 unsigned irq_from_evtchn(unsigned int evtchn)
 214 {
 215         return evtchn_to_irq[evtchn];
 216 }
 217 EXPORT_SYMBOL_GPL(irq_from_evtchn);
 218
 219 static enum ipi_vector ipi_from_irq(unsigned irq)
 220 {
 221         struct irq_info *info = info_for_irq(irq);
 222
 223         BUG_ON(info == NULL);
 224         BUG_ON(info->type != IRQT_IPI);
 225
 226         return info->u.ipi;
 227 }
 228
 229 static unsigned virq_from_irq(unsigned irq)
 230 {
 231         struct irq_info *info = info_for_irq(irq);
 232
 233         BUG_ON(info == NULL);
 234         BUG_ON(info->type != IRQT_VIRQ);
 235
 236         return info->u.virq;
 237 }
 238
 239 static unsigned pirq_from_irq(unsigned irq)
 240 {
 241         struct irq_info *info = info_for_irq(irq);
 242
 243         BUG_ON(info == NULL);
 244         BUG_ON(info->type != IRQT_PIRQ);
 245
 246         return info->u.pirq.pirq;
 247 }
 248
 249 static enum xen_irq_type type_from_irq(unsigned irq)
 250 {
 251         return info_for_irq(irq)->type;
 252 }
 253
 254 static unsigned cpu_from_irq(unsigned irq)
 255 {
 256         return info_for_irq(irq)->cpu;
 257 }
 258
 259 static unsigned int cpu_from_evtchn(unsigned int evtchn)
 260 {
 261         int irq = evtchn_to_irq[evtchn];
 262         unsigned ret = 0;
 263
 264         if (irq != -1)
 265                 ret = cpu_from_irq(irq);
 266
 267         return ret;
 268 }
 269
 270 static bool pirq_needs_eoi(unsigned irq)
 271 {
 272         struct irq_info *info = info_for_irq(irq);
 273
 274         BUG_ON(info->type != IRQT_PIRQ);
 275
 276         return info->u.pirq.flags & PIRQ_NEEDS_EOI;
 277 }
 278
 279 static inline unsigned long active_evtchns(unsigned int cpu,
 280                                            struct shared_info *sh,
 281                                            unsigned int idx)
 282 {
 283         return (sh->evtchn_pending[idx] &
 284                 per_cpu(cpu_evtchn_mask, cpu)[idx] &
 285                 ~sh->evtchn_mask[idx]);
 286 }
 287
 288 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
 289 {
 290         int irq = evtchn_to_irq[chn];
 291
 292         BUG_ON(irq == -1);
 293 #ifdef CONFIG_SMP
 294         cpumask_copy(irq_to_desc(irq)->irq_data.affinity, cpumask_of(cpu));
 295 #endif
 296
 297         clear_bit(chn, per_cpu(cpu_evtchn_mask, cpu_from_irq(irq)));
 298         set_bit(chn, per_cpu(cpu_evtchn_mask, cpu));
 299
 300         info_for_irq(irq)->cpu = cpu;
 301 }
 302
 303 static void init_evtchn_cpu_bindings(void)
 304 {
 305         int i;
 306 #ifdef CONFIG_SMP
 307         struct irq_info *info;
 308
 309         /* By default all event channels notify CPU#0. */
 310         list_for_each_entry(info, &xen_irq_list_head, list) {
 311                 struct irq_desc *desc = irq_to_desc(info->irq);
 312                 cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 313         }
 314 #endif
 315
 316         for_each_possible_cpu(i)
 317                 memset(per_cpu(cpu_evtchn_mask, i),
 318                        (i == 0) ? ~0 : 0, sizeof(*per_cpu(cpu_evtchn_mask, i)));
 319 }
 320
 321 static inline void clear_evtchn(int port)
 322 {
 323         struct shared_info *s = HYPERVISOR_shared_info;
 324         sync_clear_bit(port, &s->evtchn_pending[0]);
 325 }
 326
 327 static inline void set_evtchn(int port)
 328 {
 329         struct shared_info *s = HYPERVISOR_shared_info;
 330         sync_set_bit(port, &s->evtchn_pending[0]);
 331 }
 332
 333 static inline int test_evtchn(int port)
 334 {
 335         struct shared_info *s = HYPERVISOR_shared_info;
 336         return sync_test_bit(port, &s->evtchn_pending[0]);
 337 }
 338
 339
 340 /**
 341  * notify_remote_via_irq - send event to remote end of event channel via irq
 342  * @irq: irq of event channel to send event to
 343  *
 344  * Unlike notify_remote_via_evtchn(), this is safe to use across
 345  * save/restore. Notifications on a broken connection are silently
 346  * dropped.
 347  */
 348 void notify_remote_via_irq(int irq)
 349 {
 350         int evtchn = evtchn_from_irq(irq);
 351
 352         if (VALID_EVTCHN(evtchn))
 353                 notify_remote_via_evtchn(evtchn);
 354 }
 355 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
 356
 357 static void mask_evtchn(int port)
 358 {
 359         struct shared_info *s = HYPERVISOR_shared_info;
 360         sync_set_bit(port, &s->evtchn_mask[0]);
 361 }
 362
 363 static void unmask_evtchn(int port)
 364 {
 365         struct shared_info *s = HYPERVISOR_shared_info;
 366         unsigned int cpu = get_cpu();
 367
 368         BUG_ON(!irqs_disabled());
 369
 370         /* Slow path (hypercall) if this is a non-local port. */
 371         if (unlikely(cpu != cpu_from_evtchn(port))) {
 372                 struct evtchn_unmask unmask = { .port = port };
 373                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 374         } else {
 375                 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
 376
 377                 sync_clear_bit(port, &s->evtchn_mask[0]);
 378
 379                 /*
 380                  * The following is basically the equivalent of
 381                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
 382                  * the interrupt edge' if the channel is masked.
 383                  */
 384                 if (sync_test_bit(port, &s->evtchn_pending[0]) &&
 385                     !sync_test_and_set_bit(port / BITS_PER_LONG,
 386                                            &vcpu_info->evtchn_pending_sel))
 387                         vcpu_info->evtchn_upcall_pending = 1;
 388         }
 389
 390         put_cpu();
 391 }
 392
 393 static void xen_irq_init(unsigned irq)
 394 {
 395         struct irq_info *info;
 396         struct irq_desc *desc = irq_to_desc(irq);
 397
 398 #ifdef CONFIG_SMP
 399         /* By default all event channels notify CPU#0. */
 400         cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 401 #endif
 402
 403         info = kzalloc(sizeof(*info), GFP_KERNEL);
 404         if (info == NULL)
 405                 panic("Unable to allocate metadata for IRQ%d\n", irq);
 406
 407         info->type = IRQT_UNBOUND;
 408
 409         irq_set_handler_data(irq, info);
 410
 411         list_add_tail(&info->list, &xen_irq_list_head);
 412 }
 413
 414 static int __must_check xen_allocate_irq_dynamic(void)
 415 {
 416         int first = 0;
 417         int irq;
 418
 419 #ifdef CONFIG_X86_IO_APIC
 420         /*
 421          * For an HVM guest or domain 0 which see "real" (emulated or
 422          * actual respectively) GSIs we allocate dynamic IRQs
 423          * e.g. those corresponding to event channels or MSIs
 424          * etc. from the range above those "real" GSIs to avoid
 425          * collisions.
 426          */
 427         if (xen_initial_domain() || xen_hvm_domain())
 428                 first = get_nr_irqs_gsi();
 429 #endif
 430
 431         irq = irq_alloc_desc_from(first, -1);
 432
 433         xen_irq_init(irq);
 434
 435         return irq;
 436 }
 437
 438 static int __must_check xen_allocate_irq_gsi(unsigned gsi)
 439 {
 440         int irq;
 441
 442         /*
 443          * A PV guest has no concept of a GSI (since it has no ACPI
 444          * nor access to/knowledge of the physical APICs). Therefore
 445          * all IRQs are dynamically allocated from the entire IRQ
 446          * space.
 447          */
 448         if (xen_pv_domain() && !xen_initial_domain())
 449                 return xen_allocate_irq_dynamic();
 450
 451         /* Legacy IRQ descriptors are already allocated by the arch. */
 452         if (gsi < NR_IRQS_LEGACY)
 453                 irq = gsi;
 454         else
 455                 irq = irq_alloc_desc_at(gsi, -1);
 456
 457         xen_irq_init(irq);
 458
 459         return irq;
 460 }
 461
 462 static void xen_free_irq(unsigned irq)
 463 {
 464         struct irq_info *info = irq_get_handler_data(irq);
 465
 466         list_del(&info->list);
 467
 468         irq_set_handler_data(irq, NULL);
 469
 470         kfree(info);
 471
 472         /* Legacy IRQ descriptors are managed by the arch. */
 473         if (irq < NR_IRQS_LEGACY)
 474                 return;
 475
 476         irq_free_desc(irq);
 477 }
 478
 479 static void pirq_unmask_notify(int irq)
 480 {
 481         struct physdev_eoi eoi = { .irq = pirq_from_irq(irq) };
 482
 483         if (unlikely(pirq_needs_eoi(irq))) {
 484                 int rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
 485                 WARN_ON(rc);
 486         }
 487 }
 488
 489 static void pirq_query_unmask(int irq)
 490 {
 491         struct physdev_irq_status_query irq_status;
 492         struct irq_info *info = info_for_irq(irq);
 493
 494         BUG_ON(info->type != IRQT_PIRQ);
 495
 496         irq_status.irq = pirq_from_irq(irq);
 497         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
 498                 irq_status.flags = 0;
 499
 500         info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
 501         if (irq_status.flags & XENIRQSTAT_needs_eoi)
 502                 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
 503 }
 504
 505 static bool probing_irq(int irq)
 506 {
 507         struct irq_desc *desc = irq_to_desc(irq);
 508
 509         return desc && desc->action == NULL;
 510 }
 511
 512 static unsigned int __startup_pirq(unsigned int irq)
 513 {
 514         struct evtchn_bind_pirq bind_pirq;
 515         struct irq_info *info = info_for_irq(irq);
 516         int evtchn = evtchn_from_irq(irq);
 517         int rc;
 518
 519         BUG_ON(info->type != IRQT_PIRQ);
 520
 521         if (VALID_EVTCHN(evtchn))
 522                 goto out;
 523
 524         bind_pirq.pirq = pirq_from_irq(irq);
 525         /* NB. We are happy to share unless we are probing. */
 526         bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
 527                                         BIND_PIRQ__WILL_SHARE : 0;
 528         rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
 529         if (rc != 0) {
 530                 if (!probing_irq(irq))
 531                         printk(KERN_INFO "Failed to obtain physical IRQ %d\n",
 532                                irq);
 533                 return 0;
 534         }
 535         evtchn = bind_pirq.port;
 536
 537         pirq_query_unmask(irq);
 538
 539         evtchn_to_irq[evtchn] = irq;
 540         bind_evtchn_to_cpu(evtchn, 0);
 541         info->evtchn = evtchn;
 542
 543 out:
 544         unmask_evtchn(evtchn);
 545         pirq_unmask_notify(irq);
 546
 547         return 0;
 548 }
 549
 550 static unsigned int startup_pirq(struct irq_data *data)
 551 {
 552         return __startup_pirq(data->irq);
 553 }
 554
 555 static void shutdown_pirq(struct irq_data *data)
 556 {
 557         struct evtchn_close close;
 558         unsigned int irq = data->irq;
 559         struct irq_info *info = info_for_irq(irq);
 560         int evtchn = evtchn_from_irq(irq);
 561
 562         BUG_ON(info->type != IRQT_PIRQ);
 563
 564         if (!VALID_EVTCHN(evtchn))
 565                 return;
 566
 567         mask_evtchn(evtchn);
 568
 569         close.port = evtchn;
 570         if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 571                 BUG();
 572
 573         bind_evtchn_to_cpu(evtchn, 0);
 574         evtchn_to_irq[evtchn] = -1;
 575         info->evtchn = 0;
 576 }
 577
 578 static void enable_pirq(struct irq_data *data)
 579 {
 580         startup_pirq(data);
 581 }
 582
 583 static void disable_pirq(struct irq_data *data)
 584 {
 585 }
 586
 587 static void ack_pirq(struct irq_data *data)
 588 {
 589         int evtchn = evtchn_from_irq(data->irq);
 590
 591         irq_move_irq(data);
 592
 593         if (VALID_EVTCHN(evtchn)) {
 594                 mask_evtchn(evtchn);
 595                 clear_evtchn(evtchn);
 596         }
 597 }
 598
 599 static int find_irq_by_gsi(unsigned gsi)
 600 {
 601         struct irq_info *info;
 602
 603         list_for_each_entry(info, &xen_irq_list_head, list) {
 604                 if (info->type != IRQT_PIRQ)
 605                         continue;
 606
 607                 if (info->u.pirq.gsi == gsi)
 608                         return info->irq;
 609         }
 610
 611         return -1;
 612 }
 613
 614 int xen_allocate_pirq_gsi(unsigned gsi)
 615 {
 616         return gsi;
 617 }
 618
 619 /*
 620  * Do not make any assumptions regarding the relationship between the
 621  * IRQ number returned here and the Xen pirq argument.
 622  *
 623  * Note: We don't assign an event channel until the irq actually started
 624  * up.  Return an existing irq if we've already got one for the gsi.
 625  */
 626 int xen_bind_pirq_gsi_to_irq(unsigned gsi,
 627                              unsigned pirq, int shareable, char *name)
 628 {
 629         int irq = -1;
 630         struct physdev_irq irq_op;
 631
 632         spin_lock(&irq_mapping_update_lock);
 633
 634         irq = find_irq_by_gsi(gsi);
 635         if (irq != -1) {
 636                 printk(KERN_INFO "xen_map_pirq_gsi: returning irq %d for gsi %u\n",
 637                        irq, gsi);
 638                 goto out;       /* XXX need refcount? */
 639         }
 640
 641         irq = xen_allocate_irq_gsi(gsi);
 642         if (irq < 0)
 643                 goto out;
 644
 645         irq_set_chip_and_handler_name(irq, &xen_pirq_chip, handle_level_irq,
 646                                       name);
 647
 648         irq_op.irq = irq;
 649         irq_op.vector = 0;
 650
 651         /* Only the privileged domain can do this. For non-priv, the pcifront
 652          * driver provides a PCI bus that does the call to do exactly
 653          * this in the priv domain. */
 654         if (xen_initial_domain() &&
 655             HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
 656                 xen_free_irq(irq);
 657                 irq = -ENOSPC;
 658                 goto out;
 659         }
 660
 661         xen_irq_info_pirq_init(irq, 0, pirq, gsi, irq_op.vector, DOMID_SELF,
 662                                shareable ? PIRQ_SHAREABLE : 0);
 663
 664 out:
 665         spin_unlock(&irq_mapping_update_lock);
 666
 667         return irq;
 668 }
 669
 670 #ifdef CONFIG_PCI_MSI
 671 int xen_allocate_pirq_msi(struct pci_dev *dev, struct msi_desc *msidesc)
 672 {
 673         int rc;
 674         struct physdev_get_free_pirq op_get_free_pirq;
 675
 676         op_get_free_pirq.type = MAP_PIRQ_TYPE_MSI;
 677         rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
 678
 679         WARN_ONCE(rc == -ENOSYS,
 680                   "hypervisor does not support the PHYSDEVOP_get_free_pirq interface\n");
 681
 682         return rc ? -1 : op_get_free_pirq.pirq;
 683 }
 684
 685 int xen_bind_pirq_msi_to_irq(struct pci_dev *dev, struct msi_desc *msidesc,
 686                              int pirq, int vector, const char *name,
 687                              domid_t domid)
 688 {
 689         int irq, ret;
 690
 691         spin_lock(&irq_mapping_update_lock);
 692
 693         irq = xen_allocate_irq_dynamic();
 694         if (irq == -1)
 695                 goto out;
 696
 697         irq_set_chip_and_handler_name(irq, &xen_pirq_chip, handle_level_irq,
 698                                       name);
 699
 700         xen_irq_info_pirq_init(irq, 0, pirq, 0, vector, domid, 0);
 701         ret = irq_set_msi_desc(irq, msidesc);
 702         if (ret < 0)
 703                 goto error_irq;
 704 out:
 705         spin_unlock(&irq_mapping_update_lock);
 706         return irq;
 707 error_irq:
 708         spin_unlock(&irq_mapping_update_lock);
 709         xen_free_irq(irq);
 710         return -1;
 711 }
 712 #endif
 713
 714 int xen_destroy_irq(int irq)
 715 {
 716         struct irq_desc *desc;
 717         struct physdev_unmap_pirq unmap_irq;
 718         struct irq_info *info = info_for_irq(irq);
 719         int rc = -ENOENT;
 720
 721         spin_lock(&irq_mapping_update_lock);
 722
 723         desc = irq_to_desc(irq);
 724         if (!desc)
 725                 goto out;
 726
 727         if (xen_initial_domain()) {
 728                 unmap_irq.pirq = info->u.pirq.pirq;
 729                 unmap_irq.domid = info->u.pirq.domid;
 730                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
 731                 /* If another domain quits without making the pci_disable_msix
 732                  * call, the Xen hypervisor takes care of freeing the PIRQs
 733                  * (free_domain_pirqs).
 734                  */
 735                 if ((rc == -ESRCH && info->u.pirq.domid != DOMID_SELF))
 736                         printk(KERN_INFO "domain %d does not have %d anymore\n",
 737                                 info->u.pirq.domid, info->u.pirq.pirq);
 738                 else if (rc) {
 739                         printk(KERN_WARNING "unmap irq failed %d\n", rc);
 740                         goto out;
 741                 }
 742         }
 743
 744         xen_free_irq(irq);
 745
 746 out:
 747         spin_unlock(&irq_mapping_update_lock);
 748         return rc;
 749 }
 750
 751 int xen_irq_from_pirq(unsigned pirq)
 752 {
 753         int irq;
 754
 755         struct irq_info *info;
 756
 757         spin_lock(&irq_mapping_update_lock);
 758
 759         list_for_each_entry(info, &xen_irq_list_head, list) {
 760                 if (info == NULL || info->type != IRQT_PIRQ)
 761                         continue;
 762                 irq = info->irq;
 763                 if (info->u.pirq.pirq == pirq)
 764                         goto out;
 765         }
 766         irq = -1;
 767 out:
 768         spin_unlock(&irq_mapping_update_lock);
 769
 770         return irq;
 771 }
 772
 773
 774 int xen_pirq_from_irq(unsigned irq)
 775 {
 776         return pirq_from_irq(irq);
 777 }
 778 EXPORT_SYMBOL_GPL(xen_pirq_from_irq);
 779 int bind_evtchn_to_irq(unsigned int evtchn)
 780 {
 781         int irq;
 782
 783         spin_lock(&irq_mapping_update_lock);
 784
 785         irq = evtchn_to_irq[evtchn];
 786
 787         if (irq == -1) {
 788                 irq = xen_allocate_irq_dynamic();
 789                 if (irq == -1)
 790                         goto out;
 791
 792                 irq_set_chip_and_handler_name(irq, &xen_dynamic_chip,
 793                                               handle_fasteoi_irq, "event");
 794
 795                 xen_irq_info_evtchn_init(irq, evtchn);
 796         }
 797
 798 out:
 799         spin_unlock(&irq_mapping_update_lock);
 800
 801         return irq;
 802 }
 803 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
 804
 805 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
 806 {
 807         struct evtchn_bind_ipi bind_ipi;
 808         int evtchn, irq;
 809
 810         spin_lock(&irq_mapping_update_lock);
 811
 812         irq = per_cpu(ipi_to_irq, cpu)[ipi];
 813
 814         if (irq == -1) {
 815                 irq = xen_allocate_irq_dynamic();
 816                 if (irq < 0)
 817                         goto out;
 818
 819                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 820                                               handle_percpu_irq, "ipi");
 821
 822                 bind_ipi.vcpu = cpu;
 823                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
 824                                                 &bind_ipi) != 0)
 825                         BUG();
 826                 evtchn = bind_ipi.port;
 827
 828                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
 829
 830                 bind_evtchn_to_cpu(evtchn, cpu);
 831         }
 832
 833  out:
 834         spin_unlock(&irq_mapping_update_lock);
 835         return irq;
 836 }
 837
 838 static int bind_interdomain_evtchn_to_irq(unsigned int remote_domain,
 839                                           unsigned int remote_port)
 840 {
 841         struct evtchn_bind_interdomain bind_interdomain;
 842         int err;
 843
 844         bind_interdomain.remote_dom  = remote_domain;
 845         bind_interdomain.remote_port = remote_port;
 846
 847         err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
 848                                           &bind_interdomain);
 849
 850         return err ? : bind_evtchn_to_irq(bind_interdomain.local_port);
 851 }
 852
 853
 854 int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
 855 {
 856         struct evtchn_bind_virq bind_virq;
 857         int evtchn, irq;
 858
 859         spin_lock(&irq_mapping_update_lock);
 860
 861         irq = per_cpu(virq_to_irq, cpu)[virq];
 862
 863         if (irq == -1) {
 864                 irq = xen_allocate_irq_dynamic();
 865                 if (irq == -1)
 866                         goto out;
 867
 868                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 869                                               handle_percpu_irq, "virq");
 870
 871                 bind_virq.virq = virq;
 872                 bind_virq.vcpu = cpu;
 873                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 874                                                 &bind_virq) != 0)
 875                         BUG();
 876                 evtchn = bind_virq.port;
 877
 878                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
 879
 880                 bind_evtchn_to_cpu(evtchn, cpu);
 881         }
 882
 883 out:
 884         spin_unlock(&irq_mapping_update_lock);
 885
 886         return irq;
 887 }
 888
 889 static void unbind_from_irq(unsigned int irq)
 890 {
 891         struct evtchn_close close;
 892         int evtchn = evtchn_from_irq(irq);
 893
 894         spin_lock(&irq_mapping_update_lock);
 895
 896         if (VALID_EVTCHN(evtchn)) {
 897                 close.port = evtchn;
 898                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 899                         BUG();
 900
 901                 switch (type_from_irq(irq)) {
 902                 case IRQT_VIRQ:
 903                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
 904                                 [virq_from_irq(irq)] = -1;
 905                         break;
 906                 case IRQT_IPI:
 907                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
 908                                 [ipi_from_irq(irq)] = -1;
 909                         break;
 910                 default:
 911                         break;
 912                 }
 913
 914                 /* Closed ports are implicitly re-bound to VCPU0. */
 915                 bind_evtchn_to_cpu(evtchn, 0);
 916
 917                 evtchn_to_irq[evtchn] = -1;
 918         }
 919
 920         BUG_ON(info_for_irq(irq)->type == IRQT_UNBOUND);
 921
 922         xen_free_irq(irq);
 923
 924         spin_unlock(&irq_mapping_update_lock);
 925 }
 926
 927 int bind_evtchn_to_irqhandler(unsigned int evtchn,
 928                               irq_handler_t handler,
 929                               unsigned long irqflags,
 930                               const char *devname, void *dev_id)
 931 {
 932         int irq, retval;
 933
 934         irq = bind_evtchn_to_irq(evtchn);
 935         if (irq < 0)
 936                 return irq;
 937         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 938         if (retval != 0) {
 939                 unbind_from_irq(irq);
 940                 return retval;
 941         }
 942
 943         return irq;
 944 }
 945 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
 946
 947 int bind_interdomain_evtchn_to_irqhandler(unsigned int remote_domain,
 948                                           unsigned int remote_port,
 949                                           irq_handler_t handler,
 950                                           unsigned long irqflags,
 951                                           const char *devname,
 952                                           void *dev_id)
 953 {
 954         int irq, retval;
 955
 956         irq = bind_interdomain_evtchn_to_irq(remote_domain, remote_port);
 957         if (irq < 0)
 958                 return irq;
 959
 960         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 961         if (retval != 0) {
 962                 unbind_from_irq(irq);
 963                 return retval;
 964         }
 965
 966         return irq;
 967 }
 968 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler);
 969
 970 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
 971                             irq_handler_t handler,
 972                             unsigned long irqflags, const char *devname, void *dev_id)
 973 {
 974         int irq, retval;
 975
 976         irq = bind_virq_to_irq(virq, cpu);
 977         if (irq < 0)
 978                 return irq;
 979         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 980         if (retval != 0) {
 981                 unbind_from_irq(irq);
 982                 return retval;
 983         }
 984
 985         return irq;
 986 }
 987 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
 988
 989 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
 990                            unsigned int cpu,
 991                            irq_handler_t handler,
 992                            unsigned long irqflags,
 993                            const char *devname,
 994                            void *dev_id)
 995 {
 996         int irq, retval;
 997
 998         irq = bind_ipi_to_irq(ipi, cpu);
 999         if (irq < 0)
1000                 return irq;
1001
1002         irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME;
1003         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1004         if (retval != 0) {
1005                 unbind_from_irq(irq);
1006                 return retval;
1007         }
1008
1009         return irq;
1010 }
1011
1012 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1013 {
1014         free_irq(irq, dev_id);
1015         unbind_from_irq(irq);
1016 }
1017 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1018
1019 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1020 {
1021         int irq = per_cpu(ipi_to_irq, cpu)[vector];
1022         BUG_ON(irq < 0);
1023         notify_remote_via_irq(irq);
1024 }
1025
1026 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
1027 {
1028         struct shared_info *sh = HYPERVISOR_shared_info;
1029         int cpu = smp_processor_id();
1030         unsigned long *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu);
1031         int i;
1032         unsigned long flags;
1033         static DEFINE_SPINLOCK(debug_lock);
1034         struct vcpu_info *v;
1035
1036         spin_lock_irqsave(&debug_lock, flags);
1037
1038         printk("\nvcpu %d\n  ", cpu);
1039
1040         for_each_online_cpu(i) {
1041                 int pending;
1042                 v = per_cpu(xen_vcpu, i);
1043                 pending = (get_irq_regs() && i == cpu)
1044                         ? xen_irqs_disabled(get_irq_regs())
1045                         : v->evtchn_upcall_mask;
1046                 printk("%d: masked=%d pending=%d event_sel %0*lx\n  ", i,
1047                        pending, v->evtchn_upcall_pending,
1048                        (int)(sizeof(v->evtchn_pending_sel)*2),
1049                        v->evtchn_pending_sel);
1050         }
1051         v = per_cpu(xen_vcpu, cpu);
1052
1053         printk("\npending:\n   ");
1054         for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
1055                 printk("%0*lx%s", (int)sizeof(sh->evtchn_pending[0])*2,
1056                        sh->evtchn_pending[i],
1057                        i % 8 == 0 ? "\n   " : " ");
1058         printk("\nglobal mask:\n   ");
1059         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1060                 printk("%0*lx%s",
1061                        (int)(sizeof(sh->evtchn_mask[0])*2),
1062                        sh->evtchn_mask[i],
1063                        i % 8 == 0 ? "\n   " : " ");
1064
1065         printk("\nglobally unmasked:\n   ");
1066         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1067                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1068                        sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
1069                        i % 8 == 0 ? "\n   " : " ");
1070
1071         printk("\nlocal cpu%d mask:\n   ", cpu);
1072         for (i = (NR_EVENT_CHANNELS/BITS_PER_LONG)-1; i >= 0; i--)
1073                 printk("%0*lx%s", (int)(sizeof(cpu_evtchn[0])*2),
1074                        cpu_evtchn[i],
1075                        i % 8 == 0 ? "\n   " : " ");
1076
1077         printk("\nlocally unmasked:\n   ");
1078         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
1079                 unsigned long pending = sh->evtchn_pending[i]
1080                         & ~sh->evtchn_mask[i]
1081                         & cpu_evtchn[i];
1082                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1083                        pending, i % 8 == 0 ? "\n   " : " ");
1084         }
1085
1086         printk("\npending list:\n");
1087         for (i = 0; i < NR_EVENT_CHANNELS; i++) {
1088                 if (sync_test_bit(i, sh->evtchn_pending)) {
1089                         int word_idx = i / BITS_PER_LONG;
1090                         printk("  %d: event %d -> irq %d%s%s%s\n",
1091                                cpu_from_evtchn(i), i,
1092                                evtchn_to_irq[i],
1093                                sync_test_bit(word_idx, &v->evtchn_pending_sel)
1094                                              ? "" : " l2-clear",
1095                                !sync_test_bit(i, sh->evtchn_mask)
1096                                              ? "" : " globally-masked",
1097                                sync_test_bit(i, cpu_evtchn)
1098                                              ? "" : " locally-masked");
1099                 }
1100         }
1101
1102         spin_unlock_irqrestore(&debug_lock, flags);
1103
1104         return IRQ_HANDLED;
1105 }
1106
1107 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
1108 static DEFINE_PER_CPU(unsigned int, current_word_idx);
1109 static DEFINE_PER_CPU(unsigned int, current_bit_idx);
1110
1111 /*
1112  * Mask out the i least significant bits of w
1113  */
1114 #define MASK_LSBS(w, i) (w & ((~0UL) << i))
1115
1116 /*
1117  * Search the CPUs pending events bitmasks.  For each one found, map
1118  * the event number to an irq, and feed it into do_IRQ() for
1119  * handling.
1120  *
1121  * Xen uses a two-level bitmap to speed searching.  The first level is
1122  * a bitset of words which contain pending event bits.  The second
1123  * level is a bitset of pending events themselves.
1124  */
1125 static void __xen_evtchn_do_upcall(void)
1126 {
1127         int start_word_idx, start_bit_idx;
1128         int word_idx, bit_idx;
1129         int i;
1130         int cpu = get_cpu();
1131         struct shared_info *s = HYPERVISOR_shared_info;
1132         struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1133         unsigned count;
1134
1135         do {
1136                 unsigned long pending_words;
1137
1138                 vcpu_info->evtchn_upcall_pending = 0;
1139
1140                 if (__this_cpu_inc_return(xed_nesting_count) - 1)
1141                         goto out;
1142
1143 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
1144                 /* Clear master flag /before/ clearing selector flag. */
1145                 wmb();
1146 #endif
1147                 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
1148
1149                 start_word_idx = __this_cpu_read(current_word_idx);
1150                 start_bit_idx = __this_cpu_read(current_bit_idx);
1151
1152                 word_idx = start_word_idx;
1153
1154                 for (i = 0; pending_words != 0; i++) {
1155                         unsigned long pending_bits;
1156                         unsigned long words;
1157
1158                         words = MASK_LSBS(pending_words, word_idx);
1159
1160                         /*
1161                          * If we masked out all events, wrap to beginning.
1162                          */
1163                         if (words == 0) {
1164                                 word_idx = 0;
1165                                 bit_idx = 0;
1166                                 continue;
1167                         }
1168                         word_idx = __ffs(words);
1169
1170                         pending_bits = active_evtchns(cpu, s, word_idx);
1171                         bit_idx = 0; /* usually scan entire word from start */
1172                         if (word_idx == start_word_idx) {
1173                                 /* We scan the starting word in two parts */
1174                                 if (i == 0)
1175                                         /* 1st time: start in the middle */
1176                                         bit_idx = start_bit_idx;
1177                                 else
1178                                         /* 2nd time: mask bits done already */
1179                                         bit_idx &= (1UL << start_bit_idx) - 1;
1180                         }
1181
1182                         do {
1183                                 unsigned long bits;
1184                                 int port, irq;
1185                                 struct irq_desc *desc;
1186
1187                                 bits = MASK_LSBS(pending_bits, bit_idx);
1188
1189                                 /* If we masked out all events, move on. */
1190                                 if (bits == 0)
1191                                         break;
1192
1193                                 bit_idx = __ffs(bits);
1194
1195                                 /* Process port. */
1196                                 port = (word_idx * BITS_PER_LONG) + bit_idx;
1197                                 irq = evtchn_to_irq[port];
1198
1199                                 mask_evtchn(port);
1200                                 clear_evtchn(port);
1201
1202                                 if (irq != -1) {
1203                                         desc = irq_to_desc(irq);
1204                                         if (desc)
1205                                                 generic_handle_irq_desc(irq, desc);
1206                                 }
1207
1208                                 bit_idx = (bit_idx + 1) % BITS_PER_LONG;
1209
1210                                 /* Next caller starts at last processed + 1 */
1211                                 __this_cpu_write(current_word_idx,
1212                                                  bit_idx ? word_idx :
1213                                                  (word_idx+1) % BITS_PER_LONG);
1214                                 __this_cpu_write(current_bit_idx, bit_idx);
1215                         } while (bit_idx != 0);
1216
1217                         /* Scan start_l1i twice; all others once. */
1218                         if ((word_idx != start_word_idx) || (i != 0))
1219                                 pending_words &= ~(1UL << word_idx);
1220
1221                         word_idx = (word_idx + 1) % BITS_PER_LONG;
1222                 }
1223
1224                 BUG_ON(!irqs_disabled());
1225
1226                 count = __this_cpu_read(xed_nesting_count);
1227                 __this_cpu_write(xed_nesting_count, 0);
1228         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
1229
1230 out:
1231
1232         put_cpu();
1233 }
1234
1235 void xen_evtchn_do_upcall(struct pt_regs *regs)
1236 {
1237         struct pt_regs *old_regs = set_irq_regs(regs);
1238
1239         exit_idle();
1240         irq_enter();
1241
1242         __xen_evtchn_do_upcall();
1243
1244         irq_exit();
1245         set_irq_regs(old_regs);
1246 }
1247
1248 void xen_hvm_evtchn_do_upcall(void)
1249 {
1250         __xen_evtchn_do_upcall();
1251 }
1252 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1253
1254 /* Rebind a new event channel to an existing irq. */
1255 void rebind_evtchn_irq(int evtchn, int irq)
1256 {
1257         struct irq_info *info = info_for_irq(irq);
1258
1259         /* Make sure the irq is masked, since the new event channel
1260            will also be masked. */
1261         disable_irq(irq);
1262
1263         spin_lock(&irq_mapping_update_lock);
1264
1265         /* After resume the irq<->evtchn mappings are all cleared out */
1266         BUG_ON(evtchn_to_irq[evtchn] != -1);
1267         /* Expect irq to have been bound before,
1268            so there should be a proper type */
1269         BUG_ON(info->type == IRQT_UNBOUND);
1270
1271         xen_irq_info_evtchn_init(irq, evtchn);
1272
1273         spin_unlock(&irq_mapping_update_lock);
1274
1275         /* new event channels are always bound to cpu 0 */
1276         irq_set_affinity(irq, cpumask_of(0));
1277
1278         /* Unmask the event channel. */
1279         enable_irq(irq);
1280 }
1281
1282 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1283 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1284 {
1285         struct evtchn_bind_vcpu bind_vcpu;
1286         int evtchn = evtchn_from_irq(irq);
1287
1288         if (!VALID_EVTCHN(evtchn))
1289                 return -1;
1290
1291         /*
1292          * Events delivered via platform PCI interrupts are always
1293          * routed to vcpu 0 and hence cannot be rebound.
1294          */
1295         if (xen_hvm_domain() && !xen_have_vector_callback)
1296                 return -1;
1297
1298         /* Send future instances of this interrupt to other vcpu. */
1299         bind_vcpu.port = evtchn;
1300         bind_vcpu.vcpu = tcpu;
1301
1302         /*
1303          * If this fails, it usually just indicates that we're dealing with a
1304          * virq or IPI channel, which don't actually need to be rebound. Ignore
1305          * it, but don't do the xenlinux-level rebind in that case.
1306          */
1307         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1308                 bind_evtchn_to_cpu(evtchn, tcpu);
1309
1310         return 0;
1311 }
1312
1313 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1314                             bool force)
1315 {
1316         unsigned tcpu = cpumask_first(dest);
1317
1318         return rebind_irq_to_cpu(data->irq, tcpu);
1319 }
1320
1321 int resend_irq_on_evtchn(unsigned int irq)
1322 {
1323         int masked, evtchn = evtchn_from_irq(irq);
1324         struct shared_info *s = HYPERVISOR_shared_info;
1325
1326         if (!VALID_EVTCHN(evtchn))
1327                 return 1;
1328
1329         masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
1330         sync_set_bit(evtchn, s->evtchn_pending);
1331         if (!masked)
1332                 unmask_evtchn(evtchn);
1333
1334         return 1;
1335 }
1336
1337 static void enable_dynirq(struct irq_data *data)
1338 {
1339         int evtchn = evtchn_from_irq(data->irq);
1340
1341         if (VALID_EVTCHN(evtchn))
1342                 unmask_evtchn(evtchn);
1343 }
1344
1345 static void disable_dynirq(struct irq_data *data)
1346 {
1347         int evtchn = evtchn_from_irq(data->irq);
1348
1349         if (VALID_EVTCHN(evtchn))
1350                 mask_evtchn(evtchn);
1351 }
1352
1353 static void ack_dynirq(struct irq_data *data)
1354 {
1355         int evtchn = evtchn_from_irq(data->irq);
1356
1357         irq_move_masked_irq(data);
1358
1359         if (VALID_EVTCHN(evtchn))
1360                 unmask_evtchn(evtchn);
1361 }
1362
1363 static int retrigger_dynirq(struct irq_data *data)
1364 {
1365         int evtchn = evtchn_from_irq(data->irq);
1366         struct shared_info *sh = HYPERVISOR_shared_info;
1367         int ret = 0;
1368
1369         if (VALID_EVTCHN(evtchn)) {
1370                 int masked;
1371
1372                 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
1373                 sync_set_bit(evtchn, sh->evtchn_pending);
1374                 if (!masked)
1375                         unmask_evtchn(evtchn);
1376                 ret = 1;
1377         }
1378
1379         return ret;
1380 }
1381
1382 static void restore_pirqs(void)
1383 {
1384         int pirq, rc, irq, gsi;
1385         struct physdev_map_pirq map_irq;
1386         struct irq_info *info;
1387
1388         list_for_each_entry(info, &xen_irq_list_head, list) {
1389                 if (info->type != IRQT_PIRQ)
1390                         continue;
1391
1392                 pirq = info->u.pirq.pirq;
1393                 gsi = info->u.pirq.gsi;
1394                 irq = info->irq;
1395
1396                 /* save/restore of PT devices doesn't work, so at this point the
1397                  * only devices present are GSI based emulated devices */
1398                 if (!gsi)
1399                         continue;
1400
1401                 map_irq.domid = DOMID_SELF;
1402                 map_irq.type = MAP_PIRQ_TYPE_GSI;
1403                 map_irq.index = gsi;
1404                 map_irq.pirq = pirq;
1405
1406                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1407                 if (rc) {
1408                         printk(KERN_WARNING "xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1409                                         gsi, irq, pirq, rc);
1410                         xen_free_irq(irq);
1411                         continue;
1412                 }
1413
1414                 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1415
1416                 __startup_pirq(irq);
1417         }
1418 }
1419
1420 static void restore_cpu_virqs(unsigned int cpu)
1421 {
1422         struct evtchn_bind_virq bind_virq;
1423         int virq, irq, evtchn;
1424
1425         for (virq = 0; virq < NR_VIRQS; virq++) {
1426                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1427                         continue;
1428
1429                 BUG_ON(virq_from_irq(irq) != virq);
1430
1431                 /* Get a new binding from Xen. */
1432                 bind_virq.virq = virq;
1433                 bind_virq.vcpu = cpu;
1434                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1435                                                 &bind_virq) != 0)
1436                         BUG();
1437                 evtchn = bind_virq.port;
1438
1439                 /* Record the new mapping. */
1440                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
1441                 bind_evtchn_to_cpu(evtchn, cpu);
1442         }
1443 }
1444
1445 static void restore_cpu_ipis(unsigned int cpu)
1446 {
1447         struct evtchn_bind_ipi bind_ipi;
1448         int ipi, irq, evtchn;
1449
1450         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1451                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1452                         continue;
1453
1454                 BUG_ON(ipi_from_irq(irq) != ipi);
1455
1456                 /* Get a new binding from Xen. */
1457                 bind_ipi.vcpu = cpu;
1458                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1459                                                 &bind_ipi) != 0)
1460                         BUG();
1461                 evtchn = bind_ipi.port;
1462
1463                 /* Record the new mapping. */
1464                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
1465                 bind_evtchn_to_cpu(evtchn, cpu);
1466         }
1467 }
1468
1469 /* Clear an irq's pending state, in preparation for polling on it */
1470 void xen_clear_irq_pending(int irq)
1471 {
1472         int evtchn = evtchn_from_irq(irq);
1473
1474         if (VALID_EVTCHN(evtchn))
1475                 clear_evtchn(evtchn);
1476 }
1477 EXPORT_SYMBOL(xen_clear_irq_pending);
1478 void xen_set_irq_pending(int irq)
1479 {
1480         int evtchn = evtchn_from_irq(irq);
1481
1482         if (VALID_EVTCHN(evtchn))
1483                 set_evtchn(evtchn);
1484 }
1485
1486 bool xen_test_irq_pending(int irq)
1487 {
1488         int evtchn = evtchn_from_irq(irq);
1489         bool ret = false;
1490
1491         if (VALID_EVTCHN(evtchn))
1492                 ret = test_evtchn(evtchn);
1493
1494         return ret;
1495 }
1496
1497 /* Poll waiting for an irq to become pending with timeout.  In the usual case,
1498  * the irq will be disabled so it won't deliver an interrupt. */
1499 void xen_poll_irq_timeout(int irq, u64 timeout)
1500 {
1501         evtchn_port_t evtchn = evtchn_from_irq(irq);
1502
1503         if (VALID_EVTCHN(evtchn)) {
1504                 struct sched_poll poll;
1505
1506                 poll.nr_ports = 1;
1507                 poll.timeout = timeout;
1508                 set_xen_guest_handle(poll.ports, &evtchn);
1509
1510                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1511                         BUG();
1512         }
1513 }
1514 EXPORT_SYMBOL(xen_poll_irq_timeout);
1515 /* Poll waiting for an irq to become pending.  In the usual case, the
1516  * irq will be disabled so it won't deliver an interrupt. */
1517 void xen_poll_irq(int irq)
1518 {
1519         xen_poll_irq_timeout(irq, 0 /* no timeout */);
1520 }
1521
1522 /* Check whether the IRQ line is shared with other guests. */
1523 int xen_test_irq_shared(int irq)
1524 {
1525         struct irq_info *info = info_for_irq(irq);
1526         struct physdev_irq_status_query irq_status = { .irq = info->u.pirq.pirq };
1527
1528         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
1529                 return 0;
1530         return !(irq_status.flags & XENIRQSTAT_shared);
1531 }
1532 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
1533
1534 void xen_irq_resume(void)
1535 {
1536         unsigned int cpu, evtchn;
1537         struct irq_info *info;
1538
1539         init_evtchn_cpu_bindings();
1540
1541         /* New event-channel space is not 'live' yet. */
1542         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1543                 mask_evtchn(evtchn);
1544
1545         /* No IRQ <-> event-channel mappings. */
1546         list_for_each_entry(info, &xen_irq_list_head, list)
1547                 info->evtchn = 0; /* zap event-channel binding */
1548
1549         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1550                 evtchn_to_irq[evtchn] = -1;
1551
1552         for_each_possible_cpu(cpu) {
1553                 restore_cpu_virqs(cpu);
1554                 restore_cpu_ipis(cpu);
1555         }
1556
1557         restore_pirqs();
1558 }
1559
1560 static struct irq_chip xen_dynamic_chip __read_mostly = {
1561         .name                   = "xen-dyn",
1562
1563         .irq_disable            = disable_dynirq,
1564         .irq_mask               = disable_dynirq,
1565         .irq_unmask             = enable_dynirq,
1566
1567         .irq_eoi                = ack_dynirq,
1568         .irq_set_affinity       = set_affinity_irq,
1569         .irq_retrigger          = retrigger_dynirq,
1570 };
1571
1572 static struct irq_chip xen_pirq_chip __read_mostly = {
1573         .name                   = "xen-pirq",
1574
1575         .irq_startup            = startup_pirq,
1576         .irq_shutdown           = shutdown_pirq,
1577
1578         .irq_enable             = enable_pirq,
1579         .irq_unmask             = enable_pirq,
1580
1581         .irq_disable            = disable_pirq,
1582         .irq_mask               = disable_pirq,
1583
1584         .irq_ack                = ack_pirq,
1585
1586         .irq_set_affinity       = set_affinity_irq,
1587
1588         .irq_retrigger          = retrigger_dynirq,
1589 };
1590
1591 static struct irq_chip xen_percpu_chip __read_mostly = {
1592         .name                   = "xen-percpu",
1593
1594         .irq_disable            = disable_dynirq,
1595         .irq_mask               = disable_dynirq,
1596         .irq_unmask             = enable_dynirq,
1597
1598         .irq_ack                = ack_dynirq,
1599 };
1600
1601 int xen_set_callback_via(uint64_t via)
1602 {
1603         struct xen_hvm_param a;
1604         a.domid = DOMID_SELF;
1605         a.index = HVM_PARAM_CALLBACK_IRQ;
1606         a.value = via;
1607         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1608 }
1609 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1610
1611 #ifdef CONFIG_XEN_PVHVM
1612 /* Vector callbacks are better than PCI interrupts to receive event
1613  * channel notifications because we can receive vector callbacks on any
1614  * vcpu and we don't need PCI support or APIC interactions. */
1615 void xen_callback_vector(void)
1616 {
1617         int rc;
1618         uint64_t callback_via;
1619         if (xen_have_vector_callback) {
1620                 callback_via = HVM_CALLBACK_VECTOR(XEN_HVM_EVTCHN_CALLBACK);
1621                 rc = xen_set_callback_via(callback_via);
1622                 if (rc) {
1623                         printk(KERN_ERR "Request for Xen HVM callback vector"
1624                                         " failed.\n");
1625                         xen_have_vector_callback = 0;
1626                         return;
1627                 }
1628                 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1629                                 "enabled\n");
1630                 /* in the restore case the vector has already been allocated */
1631                 if (!test_bit(XEN_HVM_EVTCHN_CALLBACK, used_vectors))
1632                         alloc_intr_gate(XEN_HVM_EVTCHN_CALLBACK, xen_hvm_callback_vector);
1633         }
1634 }
1635 #else
1636 void xen_callback_vector(void) {}
1637 #endif
1638
1639 void __init xen_init_IRQ(void)
1640 {
1641         int i;
1642
1643         evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1644                                     GFP_KERNEL);
1645         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1646                 evtchn_to_irq[i] = -1;
1647
1648         init_evtchn_cpu_bindings();
1649
1650         /* No event channels are 'live' right now. */
1651         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1652                 mask_evtchn(i);
1653
1654         if (xen_hvm_domain()) {
1655                 xen_callback_vector();
1656                 native_init_IRQ();
1657                 /* pci_xen_hvm_init must be called after native_init_IRQ so that
1658                  * __acpi_register_gsi can point at the right function */
1659                 pci_xen_hvm_init();
1660         } else {
1661                 irq_ctx_init(smp_processor_id());
1662                 if (xen_initial_domain())
1663                         xen_setup_pirqs();
1664         }
1665 }