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 #ifdef CONFIG_X86
  35 #include <asm/desc.h>
  36 #include <asm/ptrace.h>
  37 #include <asm/irq.h>
  38 #include <asm/idle.h>
  39 #include <asm/io_apic.h>
  40 #include <asm/xen/page.h>
  41 #include <asm/xen/pci.h>
  42 #endif
  43 #include <asm/sync_bitops.h>
  44 #include <asm/xen/hypercall.h>
  45 #include <asm/xen/hypervisor.h>
  46
  47 #include <xen/xen.h>
  48 #include <xen/hvm.h>
  49 #include <xen/xen-ops.h>
  50 #include <xen/events.h>
  51 #include <xen/interface/xen.h>
  52 #include <xen/interface/event_channel.h>
  53 #include <xen/interface/hvm/hvm_op.h>
  54 #include <xen/interface/hvm/params.h>
  55 #include <xen/interface/physdev.h>
  56 #include <xen/interface/sched.h>
  57 #include <asm/hw_irq.h>
  58
  59 /*
  60  * This lock protects updates to the following mapping and reference-count
  61  * arrays. The lock does not need to be acquired to read the mapping tables.
  62  */
  63 static DEFINE_MUTEX(irq_mapping_update_lock);
  64
  65 static LIST_HEAD(xen_irq_list_head);
  66
  67 /* IRQ <-> VIRQ mapping. */
  68 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
  69
  70 /* IRQ <-> IPI mapping */
  71 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
  72
  73 /* Interrupt types. */
  74 enum xen_irq_type {
  75         IRQT_UNBOUND = 0,
  76         IRQT_PIRQ,
  77         IRQT_VIRQ,
  78         IRQT_IPI,
  79         IRQT_EVTCHN
  80 };
  81
  82 /*
  83  * Packed IRQ information:
  84  * type - enum xen_irq_type
  85  * event channel - irq->event channel mapping
  86  * cpu - cpu this event channel is bound to
  87  * index - type-specific information:
  88  *    PIRQ - vector, with MSB being "needs EIO", or physical IRQ of the HVM
  89  *           guest, or GSI (real passthrough IRQ) of the device.
  90  *    VIRQ - virq number
  91  *    IPI - IPI vector
  92  *    EVTCHN -
  93  */
  94 struct irq_info {
  95         struct list_head list;
  96         int refcnt;
  97         enum xen_irq_type type; /* type */
  98         unsigned irq;
  99         unsigned short evtchn;  /* event channel */
 100         unsigned short cpu;     /* cpu bound */
 101
 102         union {
 103                 unsigned short virq;
 104                 enum ipi_vector ipi;
 105                 struct {
 106                         unsigned short pirq;
 107                         unsigned short gsi;
 108                         unsigned char vector;
 109                         unsigned char flags;
 110                         uint16_t domid;
 111                 } pirq;
 112         } u;
 113 };
 114 #define PIRQ_NEEDS_EOI  (1 << 0)
 115 #define PIRQ_SHAREABLE  (1 << 1)
 116
 117 static int *evtchn_to_irq;
 118 #ifdef CONFIG_X86
 119 static unsigned long *pirq_eoi_map;
 120 #endif
 121 static bool (*pirq_needs_eoi)(unsigned irq);
 122
 123 /*
 124  * Note sizeof(xen_ulong_t) can be more than sizeof(unsigned long). Be
 125  * careful to only use bitops which allow for this (e.g
 126  * test_bit/find_first_bit and friends but not __ffs) and to pass
 127  * BITS_PER_EVTCHN_WORD as the bitmask length.
 128  */
 129 #define BITS_PER_EVTCHN_WORD (sizeof(xen_ulong_t)*8)
 130 /*
 131  * Make a bitmask (i.e. unsigned long *) of a xen_ulong_t
 132  * array. Primarily to avoid long lines (hence the terse name).
 133  */
 134 #define BM(x) (unsigned long *)(x)
 135 /* Find the first set bit in a evtchn mask */
 136 #define EVTCHN_FIRST_BIT(w) find_first_bit(BM(&(w)), BITS_PER_EVTCHN_WORD)
 137
 138 static DEFINE_PER_CPU(xen_ulong_t [NR_EVENT_CHANNELS/BITS_PER_EVTCHN_WORD],
 139                       cpu_evtchn_mask);
 140
 141 /* Xen will never allocate port zero for any purpose. */
 142 #define VALID_EVTCHN(chn)       ((chn) != 0)
 143
 144 static struct irq_chip xen_dynamic_chip;
 145 static struct irq_chip xen_percpu_chip;
 146 static struct irq_chip xen_pirq_chip;
 147 static void enable_dynirq(struct irq_data *data);
 148 static void disable_dynirq(struct irq_data *data);
 149
 150 /* Get info for IRQ */
 151 static struct irq_info *info_for_irq(unsigned irq)
 152 {
 153         return irq_get_handler_data(irq);
 154 }
 155
 156 /* Constructors for packed IRQ information. */
 157 static void xen_irq_info_common_init(struct irq_info *info,
 158                                      unsigned irq,
 159                                      enum xen_irq_type type,
 160                                      unsigned short evtchn,
 161                                      unsigned short cpu)
 162 {
 163
 164         BUG_ON(info->type != IRQT_UNBOUND && info->type != type);
 165
 166         info->type = type;
 167         info->irq = irq;
 168         info->evtchn = evtchn;
 169         info->cpu = cpu;
 170
 171         evtchn_to_irq[evtchn] = irq;
 172 }
 173
 174 static void xen_irq_info_evtchn_init(unsigned irq,
 175                                      unsigned short evtchn)
 176 {
 177         struct irq_info *info = info_for_irq(irq);
 178
 179         xen_irq_info_common_init(info, irq, IRQT_EVTCHN, evtchn, 0);
 180 }
 181
 182 static void xen_irq_info_ipi_init(unsigned cpu,
 183                                   unsigned irq,
 184                                   unsigned short evtchn,
 185                                   enum ipi_vector ipi)
 186 {
 187         struct irq_info *info = info_for_irq(irq);
 188
 189         xen_irq_info_common_init(info, irq, IRQT_IPI, evtchn, 0);
 190
 191         info->u.ipi = ipi;
 192
 193         per_cpu(ipi_to_irq, cpu)[ipi] = irq;
 194 }
 195
 196 static void xen_irq_info_virq_init(unsigned cpu,
 197                                    unsigned irq,
 198                                    unsigned short evtchn,
 199                                    unsigned short virq)
 200 {
 201         struct irq_info *info = info_for_irq(irq);
 202
 203         xen_irq_info_common_init(info, irq, IRQT_VIRQ, evtchn, 0);
 204
 205         info->u.virq = virq;
 206
 207         per_cpu(virq_to_irq, cpu)[virq] = irq;
 208 }
 209
 210 static void xen_irq_info_pirq_init(unsigned irq,
 211                                    unsigned short evtchn,
 212                                    unsigned short pirq,
 213                                    unsigned short gsi,
 214                                    unsigned short vector,
 215                                    uint16_t domid,
 216                                    unsigned char flags)
 217 {
 218         struct irq_info *info = info_for_irq(irq);
 219
 220         xen_irq_info_common_init(info, irq, IRQT_PIRQ, evtchn, 0);
 221
 222         info->u.pirq.pirq = pirq;
 223         info->u.pirq.gsi = gsi;
 224         info->u.pirq.vector = vector;
 225         info->u.pirq.domid = domid;
 226         info->u.pirq.flags = flags;
 227 }
 228
 229 /*
 230  * Accessors for packed IRQ information.
 231  */
 232 static unsigned int evtchn_from_irq(unsigned irq)
 233 {
 234         if (unlikely(WARN(irq < 0 || irq >= nr_irqs, "Invalid irq %d!\n", irq)))
 235                 return 0;
 236
 237         return info_for_irq(irq)->evtchn;
 238 }
 239
 240 unsigned irq_from_evtchn(unsigned int evtchn)
 241 {
 242         return evtchn_to_irq[evtchn];
 243 }
 244 EXPORT_SYMBOL_GPL(irq_from_evtchn);
 245
 246 static enum ipi_vector ipi_from_irq(unsigned irq)
 247 {
 248         struct irq_info *info = info_for_irq(irq);
 249
 250         BUG_ON(info == NULL);
 251         BUG_ON(info->type != IRQT_IPI);
 252
 253         return info->u.ipi;
 254 }
 255
 256 static unsigned virq_from_irq(unsigned irq)
 257 {
 258         struct irq_info *info = info_for_irq(irq);
 259
 260         BUG_ON(info == NULL);
 261         BUG_ON(info->type != IRQT_VIRQ);
 262
 263         return info->u.virq;
 264 }
 265
 266 static unsigned pirq_from_irq(unsigned irq)
 267 {
 268         struct irq_info *info = info_for_irq(irq);
 269
 270         BUG_ON(info == NULL);
 271         BUG_ON(info->type != IRQT_PIRQ);
 272
 273         return info->u.pirq.pirq;
 274 }
 275
 276 static enum xen_irq_type type_from_irq(unsigned irq)
 277 {
 278         return info_for_irq(irq)->type;
 279 }
 280
 281 static unsigned cpu_from_irq(unsigned irq)
 282 {
 283         return info_for_irq(irq)->cpu;
 284 }
 285
 286 static unsigned int cpu_from_evtchn(unsigned int evtchn)
 287 {
 288         int irq = evtchn_to_irq[evtchn];
 289         unsigned ret = 0;
 290
 291         if (irq != -1)
 292                 ret = cpu_from_irq(irq);
 293
 294         return ret;
 295 }
 296
 297 #ifdef CONFIG_X86
 298 static bool pirq_check_eoi_map(unsigned irq)
 299 {
 300         return test_bit(pirq_from_irq(irq), pirq_eoi_map);
 301 }
 302 #endif
 303
 304 static bool pirq_needs_eoi_flag(unsigned irq)
 305 {
 306         struct irq_info *info = info_for_irq(irq);
 307         BUG_ON(info->type != IRQT_PIRQ);
 308
 309         return info->u.pirq.flags & PIRQ_NEEDS_EOI;
 310 }
 311
 312 static inline xen_ulong_t active_evtchns(unsigned int cpu,
 313                                          struct shared_info *sh,
 314                                          unsigned int idx)
 315 {
 316         return sh->evtchn_pending[idx] &
 317                 per_cpu(cpu_evtchn_mask, cpu)[idx] &
 318                 ~sh->evtchn_mask[idx];
 319 }
 320
 321 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
 322 {
 323         int irq = evtchn_to_irq[chn];
 324
 325         BUG_ON(irq == -1);
 326 #ifdef CONFIG_SMP
 327         cpumask_copy(irq_to_desc(irq)->irq_data.affinity, cpumask_of(cpu));
 328 #endif
 329
 330         clear_bit(chn, BM(per_cpu(cpu_evtchn_mask, cpu_from_irq(irq))));
 331         set_bit(chn, BM(per_cpu(cpu_evtchn_mask, cpu)));
 332
 333         info_for_irq(irq)->cpu = cpu;
 334 }
 335
 336 static void init_evtchn_cpu_bindings(void)
 337 {
 338         int i;
 339 #ifdef CONFIG_SMP
 340         struct irq_info *info;
 341
 342         /* By default all event channels notify CPU#0. */
 343         list_for_each_entry(info, &xen_irq_list_head, list) {
 344                 struct irq_desc *desc = irq_to_desc(info->irq);
 345                 cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 346         }
 347 #endif
 348
 349         for_each_possible_cpu(i)
 350                 memset(per_cpu(cpu_evtchn_mask, i),
 351                        (i == 0) ? ~0 : 0, sizeof(*per_cpu(cpu_evtchn_mask, i)));
 352 }
 353
 354 static inline void clear_evtchn(int port)
 355 {
 356         struct shared_info *s = HYPERVISOR_shared_info;
 357         sync_clear_bit(port, BM(&s->evtchn_pending[0]));
 358 }
 359
 360 static inline void set_evtchn(int port)
 361 {
 362         struct shared_info *s = HYPERVISOR_shared_info;
 363         sync_set_bit(port, BM(&s->evtchn_pending[0]));
 364 }
 365
 366 static inline int test_evtchn(int port)
 367 {
 368         struct shared_info *s = HYPERVISOR_shared_info;
 369         return sync_test_bit(port, BM(&s->evtchn_pending[0]));
 370 }
 371
 372
 373 /**
 374  * notify_remote_via_irq - send event to remote end of event channel via irq
 375  * @irq: irq of event channel to send event to
 376  *
 377  * Unlike notify_remote_via_evtchn(), this is safe to use across
 378  * save/restore. Notifications on a broken connection are silently
 379  * dropped.
 380  */
 381 void notify_remote_via_irq(int irq)
 382 {
 383         int evtchn = evtchn_from_irq(irq);
 384
 385         if (VALID_EVTCHN(evtchn))
 386                 notify_remote_via_evtchn(evtchn);
 387 }
 388 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
 389
 390 static void mask_evtchn(int port)
 391 {
 392         struct shared_info *s = HYPERVISOR_shared_info;
 393         sync_set_bit(port, BM(&s->evtchn_mask[0]));
 394 }
 395
 396 static void unmask_evtchn(int port)
 397 {
 398         struct shared_info *s = HYPERVISOR_shared_info;
 399         unsigned int cpu = get_cpu();
 400         int do_hypercall = 0, evtchn_pending = 0;
 401
 402         BUG_ON(!irqs_disabled());
 403
 404         if (unlikely((cpu != cpu_from_evtchn(port))))
 405                 do_hypercall = 1;
 406         else
 407                 evtchn_pending = sync_test_bit(port, BM(&s->evtchn_pending[0]));
 408
 409         if (unlikely(evtchn_pending && xen_hvm_domain()))
 410                 do_hypercall = 1;
 411
 412         /* Slow path (hypercall) if this is a non-local port or if this is
 413          * an hvm domain and an event is pending (hvm domains don't have
 414          * their own implementation of irq_enable). */
 415         if (do_hypercall) {
 416                 struct evtchn_unmask unmask = { .port = port };
 417                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 418         } else {
 419                 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
 420
 421                 sync_clear_bit(port, BM(&s->evtchn_mask[0]));
 422
 423                 /*
 424                  * The following is basically the equivalent of
 425                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
 426                  * the interrupt edge' if the channel is masked.
 427                  */
 428                 if (evtchn_pending &&
 429                     !sync_test_and_set_bit(port / BITS_PER_EVTCHN_WORD,
 430                                            BM(&vcpu_info->evtchn_pending_sel)))
 431                         vcpu_info->evtchn_upcall_pending = 1;
 432         }
 433
 434         put_cpu();
 435 }
 436
 437 static void xen_irq_init(unsigned irq)
 438 {
 439         struct irq_info *info;
 440 #ifdef CONFIG_SMP
 441         struct irq_desc *desc = irq_to_desc(irq);
 442
 443         /* By default all event channels notify CPU#0. */
 444         cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 445 #endif
 446
 447         info = kzalloc(sizeof(*info), GFP_KERNEL);
 448         if (info == NULL)
 449                 panic("Unable to allocate metadata for IRQ%d\n", irq);
 450
 451         info->type = IRQT_UNBOUND;
 452         info->refcnt = -1;
 453
 454         irq_set_handler_data(irq, info);
 455
 456         list_add_tail(&info->list, &xen_irq_list_head);
 457 }
 458
 459 static int __must_check xen_allocate_irq_dynamic(void)
 460 {
 461         int first = 0;
 462         int irq;
 463
 464 #ifdef CONFIG_X86_IO_APIC
 465         /*
 466          * For an HVM guest or domain 0 which see "real" (emulated or
 467          * actual respectively) GSIs we allocate dynamic IRQs
 468          * e.g. those corresponding to event channels or MSIs
 469          * etc. from the range above those "real" GSIs to avoid
 470          * collisions.
 471          */
 472         if (xen_initial_domain() || xen_hvm_domain())
 473                 first = get_nr_irqs_gsi();
 474 #endif
 475
 476         irq = irq_alloc_desc_from(first, -1);
 477
 478         if (irq >= 0)
 479                 xen_irq_init(irq);
 480
 481         return irq;
 482 }
 483
 484 static int __must_check xen_allocate_irq_gsi(unsigned gsi)
 485 {
 486         int irq;
 487
 488         /*
 489          * A PV guest has no concept of a GSI (since it has no ACPI
 490          * nor access to/knowledge of the physical APICs). Therefore
 491          * all IRQs are dynamically allocated from the entire IRQ
 492          * space.
 493          */
 494         if (xen_pv_domain() && !xen_initial_domain())
 495                 return xen_allocate_irq_dynamic();
 496
 497         /* Legacy IRQ descriptors are already allocated by the arch. */
 498         if (gsi < NR_IRQS_LEGACY)
 499                 irq = gsi;
 500         else
 501                 irq = irq_alloc_desc_at(gsi, -1);
 502
 503         xen_irq_init(irq);
 504
 505         return irq;
 506 }
 507
 508 static void xen_free_irq(unsigned irq)
 509 {
 510         struct irq_info *info = irq_get_handler_data(irq);
 511
 512         list_del(&info->list);
 513
 514         irq_set_handler_data(irq, NULL);
 515
 516         WARN_ON(info->refcnt > 0);
 517
 518         kfree(info);
 519
 520         /* Legacy IRQ descriptors are managed by the arch. */
 521         if (irq < NR_IRQS_LEGACY)
 522                 return;
 523
 524         irq_free_desc(irq);
 525 }
 526
 527 static void pirq_query_unmask(int irq)
 528 {
 529         struct physdev_irq_status_query irq_status;
 530         struct irq_info *info = info_for_irq(irq);
 531
 532         BUG_ON(info->type != IRQT_PIRQ);
 533
 534         irq_status.irq = pirq_from_irq(irq);
 535         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
 536                 irq_status.flags = 0;
 537
 538         info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
 539         if (irq_status.flags & XENIRQSTAT_needs_eoi)
 540                 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
 541 }
 542
 543 static bool probing_irq(int irq)
 544 {
 545         struct irq_desc *desc = irq_to_desc(irq);
 546
 547         return desc && desc->action == NULL;
 548 }
 549
 550 static void eoi_pirq(struct irq_data *data)
 551 {
 552         int evtchn = evtchn_from_irq(data->irq);
 553         struct physdev_eoi eoi = { .irq = pirq_from_irq(data->irq) };
 554         int rc = 0;
 555
 556         irq_move_irq(data);
 557
 558         if (VALID_EVTCHN(evtchn))
 559                 clear_evtchn(evtchn);
 560
 561         if (pirq_needs_eoi(data->irq)) {
 562                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
 563                 WARN_ON(rc);
 564         }
 565 }
 566
 567 static void mask_ack_pirq(struct irq_data *data)
 568 {
 569         disable_dynirq(data);
 570         eoi_pirq(data);
 571 }
 572
 573 static unsigned int __startup_pirq(unsigned int irq)
 574 {
 575         struct evtchn_bind_pirq bind_pirq;
 576         struct irq_info *info = info_for_irq(irq);
 577         int evtchn = evtchn_from_irq(irq);
 578         int rc;
 579
 580         BUG_ON(info->type != IRQT_PIRQ);
 581
 582         if (VALID_EVTCHN(evtchn))
 583                 goto out;
 584
 585         bind_pirq.pirq = pirq_from_irq(irq);
 586         /* NB. We are happy to share unless we are probing. */
 587         bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
 588                                         BIND_PIRQ__WILL_SHARE : 0;
 589         rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
 590         if (rc != 0) {
 591                 if (!probing_irq(irq))
 592                         printk(KERN_INFO "Failed to obtain physical IRQ %d\n",
 593                                irq);
 594                 return 0;
 595         }
 596         evtchn = bind_pirq.port;
 597
 598         pirq_query_unmask(irq);
 599
 600         evtchn_to_irq[evtchn] = irq;
 601         bind_evtchn_to_cpu(evtchn, 0);
 602         info->evtchn = evtchn;
 603
 604 out:
 605         unmask_evtchn(evtchn);
 606         eoi_pirq(irq_get_irq_data(irq));
 607
 608         return 0;
 609 }
 610
 611 static unsigned int startup_pirq(struct irq_data *data)
 612 {
 613         return __startup_pirq(data->irq);
 614 }
 615
 616 static void shutdown_pirq(struct irq_data *data)
 617 {
 618         struct evtchn_close close;
 619         unsigned int irq = data->irq;
 620         struct irq_info *info = info_for_irq(irq);
 621         int evtchn = evtchn_from_irq(irq);
 622
 623         BUG_ON(info->type != IRQT_PIRQ);
 624
 625         if (!VALID_EVTCHN(evtchn))
 626                 return;
 627
 628         mask_evtchn(evtchn);
 629
 630         close.port = evtchn;
 631         if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 632                 BUG();
 633
 634         bind_evtchn_to_cpu(evtchn, 0);
 635         evtchn_to_irq[evtchn] = -1;
 636         info->evtchn = 0;
 637 }
 638
 639 static void enable_pirq(struct irq_data *data)
 640 {
 641         startup_pirq(data);
 642 }
 643
 644 static void disable_pirq(struct irq_data *data)
 645 {
 646         disable_dynirq(data);
 647 }
 648
 649 int xen_irq_from_gsi(unsigned gsi)
 650 {
 651         struct irq_info *info;
 652
 653         list_for_each_entry(info, &xen_irq_list_head, list) {
 654                 if (info->type != IRQT_PIRQ)
 655                         continue;
 656
 657                 if (info->u.pirq.gsi == gsi)
 658                         return info->irq;
 659         }
 660
 661         return -1;
 662 }
 663 EXPORT_SYMBOL_GPL(xen_irq_from_gsi);
 664
 665 /*
 666  * Do not make any assumptions regarding the relationship between the
 667  * IRQ number returned here and the Xen pirq argument.
 668  *
 669  * Note: We don't assign an event channel until the irq actually started
 670  * up.  Return an existing irq if we've already got one for the gsi.
 671  *
 672  * Shareable implies level triggered, not shareable implies edge
 673  * triggered here.
 674  */
 675 int xen_bind_pirq_gsi_to_irq(unsigned gsi,
 676                              unsigned pirq, int shareable, char *name)
 677 {
 678         int irq = -1;
 679         struct physdev_irq irq_op;
 680
 681         mutex_lock(&irq_mapping_update_lock);
 682
 683         irq = xen_irq_from_gsi(gsi);
 684         if (irq != -1) {
 685                 printk(KERN_INFO "xen_map_pirq_gsi: returning irq %d for gsi %u\n",
 686                        irq, gsi);
 687                 goto out;
 688         }
 689
 690         irq = xen_allocate_irq_gsi(gsi);
 691         if (irq < 0)
 692                 goto out;
 693
 694         irq_op.irq = irq;
 695         irq_op.vector = 0;
 696
 697         /* Only the privileged domain can do this. For non-priv, the pcifront
 698          * driver provides a PCI bus that does the call to do exactly
 699          * this in the priv domain. */
 700         if (xen_initial_domain() &&
 701             HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
 702                 xen_free_irq(irq);
 703                 irq = -ENOSPC;
 704                 goto out;
 705         }
 706
 707         xen_irq_info_pirq_init(irq, 0, pirq, gsi, irq_op.vector, DOMID_SELF,
 708                                shareable ? PIRQ_SHAREABLE : 0);
 709
 710         pirq_query_unmask(irq);
 711         /* We try to use the handler with the appropriate semantic for the
 712          * type of interrupt: if the interrupt is an edge triggered
 713          * interrupt we use handle_edge_irq.
 714          *
 715          * On the other hand if the interrupt is level triggered we use
 716          * handle_fasteoi_irq like the native code does for this kind of
 717          * interrupts.
 718          *
 719          * Depending on the Xen version, pirq_needs_eoi might return true
 720          * not only for level triggered interrupts but for edge triggered
 721          * interrupts too. In any case Xen always honors the eoi mechanism,
 722          * not injecting any more pirqs of the same kind if the first one
 723          * hasn't received an eoi yet. Therefore using the fasteoi handler
 724          * is the right choice either way.
 725          */
 726         if (shareable)
 727                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 728                                 handle_fasteoi_irq, name);
 729         else
 730                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 731                                 handle_edge_irq, name);
 732
 733 out:
 734         mutex_unlock(&irq_mapping_update_lock);
 735
 736         return irq;
 737 }
 738
 739 #ifdef CONFIG_PCI_MSI
 740 int xen_allocate_pirq_msi(struct pci_dev *dev, struct msi_desc *msidesc)
 741 {
 742         int rc;
 743         struct physdev_get_free_pirq op_get_free_pirq;
 744
 745         op_get_free_pirq.type = MAP_PIRQ_TYPE_MSI;
 746         rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
 747
 748         WARN_ONCE(rc == -ENOSYS,
 749                   "hypervisor does not support the PHYSDEVOP_get_free_pirq interface\n");
 750
 751         return rc ? -1 : op_get_free_pirq.pirq;
 752 }
 753
 754 int xen_bind_pirq_msi_to_irq(struct pci_dev *dev, struct msi_desc *msidesc,
 755                              int pirq, int vector, const char *name,
 756                              domid_t domid)
 757 {
 758         int irq, ret;
 759
 760         mutex_lock(&irq_mapping_update_lock);
 761
 762         irq = xen_allocate_irq_dynamic();
 763         if (irq < 0)
 764                 goto out;
 765
 766         irq_set_chip_and_handler_name(irq, &xen_pirq_chip, handle_edge_irq,
 767                         name);
 768
 769         xen_irq_info_pirq_init(irq, 0, pirq, 0, vector, domid, 0);
 770         ret = irq_set_msi_desc(irq, msidesc);
 771         if (ret < 0)
 772                 goto error_irq;
 773 out:
 774         mutex_unlock(&irq_mapping_update_lock);
 775         return irq;
 776 error_irq:
 777         mutex_unlock(&irq_mapping_update_lock);
 778         xen_free_irq(irq);
 779         return ret;
 780 }
 781 #endif
 782
 783 int xen_destroy_irq(int irq)
 784 {
 785         struct irq_desc *desc;
 786         struct physdev_unmap_pirq unmap_irq;
 787         struct irq_info *info = info_for_irq(irq);
 788         int rc = -ENOENT;
 789
 790         mutex_lock(&irq_mapping_update_lock);
 791
 792         desc = irq_to_desc(irq);
 793         if (!desc)
 794                 goto out;
 795
 796         if (xen_initial_domain()) {
 797                 unmap_irq.pirq = info->u.pirq.pirq;
 798                 unmap_irq.domid = info->u.pirq.domid;
 799                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
 800                 /* If another domain quits without making the pci_disable_msix
 801                  * call, the Xen hypervisor takes care of freeing the PIRQs
 802                  * (free_domain_pirqs).
 803                  */
 804                 if ((rc == -ESRCH && info->u.pirq.domid != DOMID_SELF))
 805                         printk(KERN_INFO "domain %d does not have %d anymore\n",
 806                                 info->u.pirq.domid, info->u.pirq.pirq);
 807                 else if (rc) {
 808                         printk(KERN_WARNING "unmap irq failed %d\n", rc);
 809                         goto out;
 810                 }
 811         }
 812
 813         xen_free_irq(irq);
 814
 815 out:
 816         mutex_unlock(&irq_mapping_update_lock);
 817         return rc;
 818 }
 819
 820 int xen_irq_from_pirq(unsigned pirq)
 821 {
 822         int irq;
 823
 824         struct irq_info *info;
 825
 826         mutex_lock(&irq_mapping_update_lock);
 827
 828         list_for_each_entry(info, &xen_irq_list_head, list) {
 829                 if (info->type != IRQT_PIRQ)
 830                         continue;
 831                 irq = info->irq;
 832                 if (info->u.pirq.pirq == pirq)
 833                         goto out;
 834         }
 835         irq = -1;
 836 out:
 837         mutex_unlock(&irq_mapping_update_lock);
 838
 839         return irq;
 840 }
 841
 842
 843 int xen_pirq_from_irq(unsigned irq)
 844 {
 845         return pirq_from_irq(irq);
 846 }
 847 EXPORT_SYMBOL_GPL(xen_pirq_from_irq);
 848 int bind_evtchn_to_irq(unsigned int evtchn)
 849 {
 850         int irq;
 851
 852         mutex_lock(&irq_mapping_update_lock);
 853
 854         irq = evtchn_to_irq[evtchn];
 855
 856         if (irq == -1) {
 857                 irq = xen_allocate_irq_dynamic();
 858                 if (irq < 0)
 859                         goto out;
 860
 861                 irq_set_chip_and_handler_name(irq, &xen_dynamic_chip,
 862                                               handle_edge_irq, "event");
 863
 864                 xen_irq_info_evtchn_init(irq, evtchn);
 865         } else {
 866                 struct irq_info *info = info_for_irq(irq);
 867                 WARN_ON(info == NULL || info->type != IRQT_EVTCHN);
 868         }
 869         irq_clear_status_flags(irq, IRQ_NOREQUEST|IRQ_NOAUTOEN);
 870
 871 out:
 872         mutex_unlock(&irq_mapping_update_lock);
 873
 874         return irq;
 875 }
 876 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
 877
 878 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
 879 {
 880         struct evtchn_bind_ipi bind_ipi;
 881         int evtchn, irq;
 882
 883         mutex_lock(&irq_mapping_update_lock);
 884
 885         irq = per_cpu(ipi_to_irq, cpu)[ipi];
 886
 887         if (irq == -1) {
 888                 irq = xen_allocate_irq_dynamic();
 889                 if (irq < 0)
 890                         goto out;
 891
 892                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 893                                               handle_percpu_irq, "ipi");
 894
 895                 bind_ipi.vcpu = cpu;
 896                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
 897                                                 &bind_ipi) != 0)
 898                         BUG();
 899                 evtchn = bind_ipi.port;
 900
 901                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
 902
 903                 bind_evtchn_to_cpu(evtchn, cpu);
 904         } else {
 905                 struct irq_info *info = info_for_irq(irq);
 906                 WARN_ON(info == NULL || info->type != IRQT_IPI);
 907         }
 908
 909  out:
 910         mutex_unlock(&irq_mapping_update_lock);
 911         return irq;
 912 }
 913
 914 static int bind_interdomain_evtchn_to_irq(unsigned int remote_domain,
 915                                           unsigned int remote_port)
 916 {
 917         struct evtchn_bind_interdomain bind_interdomain;
 918         int err;
 919
 920         bind_interdomain.remote_dom  = remote_domain;
 921         bind_interdomain.remote_port = remote_port;
 922
 923         err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
 924                                           &bind_interdomain);
 925
 926         return err ? : bind_evtchn_to_irq(bind_interdomain.local_port);
 927 }
 928
 929 static int find_virq(unsigned int virq, unsigned int cpu)
 930 {
 931         struct evtchn_status status;
 932         int port, rc = -ENOENT;
 933
 934         memset(&status, 0, sizeof(status));
 935         for (port = 0; port <= NR_EVENT_CHANNELS; port++) {
 936                 status.dom = DOMID_SELF;
 937                 status.port = port;
 938                 rc = HYPERVISOR_event_channel_op(EVTCHNOP_status, &status);
 939                 if (rc < 0)
 940                         continue;
 941                 if (status.status != EVTCHNSTAT_virq)
 942                         continue;
 943                 if (status.u.virq == virq && status.vcpu == cpu) {
 944                         rc = port;
 945                         break;
 946                 }
 947         }
 948         return rc;
 949 }
 950
 951 int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
 952 {
 953         struct evtchn_bind_virq bind_virq;
 954         int evtchn, irq, ret;
 955
 956         mutex_lock(&irq_mapping_update_lock);
 957
 958         irq = per_cpu(virq_to_irq, cpu)[virq];
 959
 960         if (irq == -1) {
 961                 irq = xen_allocate_irq_dynamic();
 962                 if (irq < 0)
 963                         goto out;
 964
 965                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 966                                               handle_percpu_irq, "virq");
 967
 968                 bind_virq.virq = virq;
 969                 bind_virq.vcpu = cpu;
 970                 ret = HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 971                                                 &bind_virq);
 972                 if (ret == 0)
 973                         evtchn = bind_virq.port;
 974                 else {
 975                         if (ret == -EEXIST)
 976                                 ret = find_virq(virq, cpu);
 977                         BUG_ON(ret < 0);
 978                         evtchn = ret;
 979                 }
 980
 981                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
 982
 983                 bind_evtchn_to_cpu(evtchn, cpu);
 984         } else {
 985                 struct irq_info *info = info_for_irq(irq);
 986                 WARN_ON(info == NULL || info->type != IRQT_VIRQ);
 987         }
 988
 989 out:
 990         mutex_unlock(&irq_mapping_update_lock);
 991
 992         return irq;
 993 }
 994
 995 static void unbind_from_irq(unsigned int irq)
 996 {
 997         struct evtchn_close close;
 998         int evtchn = evtchn_from_irq(irq);
 999         struct irq_info *info = irq_get_handler_data(irq);
1000
1001         mutex_lock(&irq_mapping_update_lock);
1002
1003         if (info->refcnt > 0) {
1004                 info->refcnt--;
1005                 if (info->refcnt != 0)
1006                         goto done;
1007         }
1008
1009         if (VALID_EVTCHN(evtchn)) {
1010                 close.port = evtchn;
1011                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
1012                         BUG();
1013
1014                 switch (type_from_irq(irq)) {
1015                 case IRQT_VIRQ:
1016                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
1017                                 [virq_from_irq(irq)] = -1;
1018                         break;
1019                 case IRQT_IPI:
1020                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
1021                                 [ipi_from_irq(irq)] = -1;
1022                         break;
1023                 default:
1024                         break;
1025                 }
1026
1027                 /* Closed ports are implicitly re-bound to VCPU0. */
1028                 bind_evtchn_to_cpu(evtchn, 0);
1029
1030                 evtchn_to_irq[evtchn] = -1;
1031         }
1032
1033         BUG_ON(info_for_irq(irq)->type == IRQT_UNBOUND);
1034
1035         xen_free_irq(irq);
1036
1037  done:
1038         mutex_unlock(&irq_mapping_update_lock);
1039 }
1040
1041 int bind_evtchn_to_irqhandler(unsigned int evtchn,
1042                               irq_handler_t handler,
1043                               unsigned long irqflags,
1044                               const char *devname, void *dev_id)
1045 {
1046         int irq, retval;
1047
1048         irq = bind_evtchn_to_irq(evtchn);
1049         if (irq < 0)
1050                 return irq;
1051         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1052         if (retval != 0) {
1053                 unbind_from_irq(irq);
1054                 return retval;
1055         }
1056
1057         return irq;
1058 }
1059 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
1060
1061 int bind_interdomain_evtchn_to_irqhandler(unsigned int remote_domain,
1062                                           unsigned int remote_port,
1063                                           irq_handler_t handler,
1064                                           unsigned long irqflags,
1065                                           const char *devname,
1066                                           void *dev_id)
1067 {
1068         int irq, retval;
1069
1070         irq = bind_interdomain_evtchn_to_irq(remote_domain, remote_port);
1071         if (irq < 0)
1072                 return irq;
1073
1074         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1075         if (retval != 0) {
1076                 unbind_from_irq(irq);
1077                 return retval;
1078         }
1079
1080         return irq;
1081 }
1082 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler);
1083
1084 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
1085                             irq_handler_t handler,
1086                             unsigned long irqflags, const char *devname, void *dev_id)
1087 {
1088         int irq, retval;
1089
1090         irq = bind_virq_to_irq(virq, cpu);
1091         if (irq < 0)
1092                 return irq;
1093         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1094         if (retval != 0) {
1095                 unbind_from_irq(irq);
1096                 return retval;
1097         }
1098
1099         return irq;
1100 }
1101 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1102
1103 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1104                            unsigned int cpu,
1105                            irq_handler_t handler,
1106                            unsigned long irqflags,
1107                            const char *devname,
1108                            void *dev_id)
1109 {
1110         int irq, retval;
1111
1112         irq = bind_ipi_to_irq(ipi, cpu);
1113         if (irq < 0)
1114                 return irq;
1115
1116         irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME | IRQF_EARLY_RESUME;
1117         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1118         if (retval != 0) {
1119                 unbind_from_irq(irq);
1120                 return retval;
1121         }
1122
1123         return irq;
1124 }
1125
1126 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1127 {
1128         free_irq(irq, dev_id);
1129         unbind_from_irq(irq);
1130 }
1131 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1132
1133 int evtchn_make_refcounted(unsigned int evtchn)
1134 {
1135         int irq = evtchn_to_irq[evtchn];
1136         struct irq_info *info;
1137
1138         if (irq == -1)
1139                 return -ENOENT;
1140
1141         info = irq_get_handler_data(irq);
1142
1143         if (!info)
1144                 return -ENOENT;
1145
1146         WARN_ON(info->refcnt != -1);
1147
1148         info->refcnt = 1;
1149
1150         return 0;
1151 }
1152 EXPORT_SYMBOL_GPL(evtchn_make_refcounted);
1153
1154 int evtchn_get(unsigned int evtchn)
1155 {
1156         int irq;
1157         struct irq_info *info;
1158         int err = -ENOENT;
1159
1160         if (evtchn >= NR_EVENT_CHANNELS)
1161                 return -EINVAL;
1162
1163         mutex_lock(&irq_mapping_update_lock);
1164
1165         irq = evtchn_to_irq[evtchn];
1166         if (irq == -1)
1167                 goto done;
1168
1169         info = irq_get_handler_data(irq);
1170
1171         if (!info)
1172                 goto done;
1173
1174         err = -EINVAL;
1175         if (info->refcnt <= 0)
1176                 goto done;
1177
1178         info->refcnt++;
1179         err = 0;
1180  done:
1181         mutex_unlock(&irq_mapping_update_lock);
1182
1183         return err;
1184 }
1185 EXPORT_SYMBOL_GPL(evtchn_get);
1186
1187 void evtchn_put(unsigned int evtchn)
1188 {
1189         int irq = evtchn_to_irq[evtchn];
1190         if (WARN_ON(irq == -1))
1191                 return;
1192         unbind_from_irq(irq);
1193 }
1194 EXPORT_SYMBOL_GPL(evtchn_put);
1195
1196 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1197 {
1198         int irq = per_cpu(ipi_to_irq, cpu)[vector];
1199         BUG_ON(irq < 0);
1200         notify_remote_via_irq(irq);
1201 }
1202
1203 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
1204 {
1205         struct shared_info *sh = HYPERVISOR_shared_info;
1206         int cpu = smp_processor_id();
1207         xen_ulong_t *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu);
1208         int i;
1209         unsigned long flags;
1210         static DEFINE_SPINLOCK(debug_lock);
1211         struct vcpu_info *v;
1212
1213         spin_lock_irqsave(&debug_lock, flags);
1214
1215         printk("\nvcpu %d\n  ", cpu);
1216
1217         for_each_online_cpu(i) {
1218                 int pending;
1219                 v = per_cpu(xen_vcpu, i);
1220                 pending = (get_irq_regs() && i == cpu)
1221                         ? xen_irqs_disabled(get_irq_regs())
1222                         : v->evtchn_upcall_mask;
1223                 printk("%d: masked=%d pending=%d event_sel %0*"PRI_xen_ulong"\n  ", i,
1224                        pending, v->evtchn_upcall_pending,
1225                        (int)(sizeof(v->evtchn_pending_sel)*2),
1226                        v->evtchn_pending_sel);
1227         }
1228         v = per_cpu(xen_vcpu, cpu);
1229
1230         printk("\npending:\n   ");
1231         for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
1232                 printk("%0*"PRI_xen_ulong"%s",
1233                        (int)sizeof(sh->evtchn_pending[0])*2,
1234                        sh->evtchn_pending[i],
1235                        i % 8 == 0 ? "\n   " : " ");
1236         printk("\nglobal mask:\n   ");
1237         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1238                 printk("%0*"PRI_xen_ulong"%s",
1239                        (int)(sizeof(sh->evtchn_mask[0])*2),
1240                        sh->evtchn_mask[i],
1241                        i % 8 == 0 ? "\n   " : " ");
1242
1243         printk("\nglobally unmasked:\n   ");
1244         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1245                 printk("%0*"PRI_xen_ulong"%s",
1246                        (int)(sizeof(sh->evtchn_mask[0])*2),
1247                        sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
1248                        i % 8 == 0 ? "\n   " : " ");
1249
1250         printk("\nlocal cpu%d mask:\n   ", cpu);
1251         for (i = (NR_EVENT_CHANNELS/BITS_PER_EVTCHN_WORD)-1; i >= 0; i--)
1252                 printk("%0*"PRI_xen_ulong"%s", (int)(sizeof(cpu_evtchn[0])*2),
1253                        cpu_evtchn[i],
1254                        i % 8 == 0 ? "\n   " : " ");
1255
1256         printk("\nlocally unmasked:\n   ");
1257         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
1258                 xen_ulong_t pending = sh->evtchn_pending[i]
1259                         & ~sh->evtchn_mask[i]
1260                         & cpu_evtchn[i];
1261                 printk("%0*"PRI_xen_ulong"%s",
1262                        (int)(sizeof(sh->evtchn_mask[0])*2),
1263                        pending, i % 8 == 0 ? "\n   " : " ");
1264         }
1265
1266         printk("\npending list:\n");
1267         for (i = 0; i < NR_EVENT_CHANNELS; i++) {
1268                 if (sync_test_bit(i, BM(sh->evtchn_pending))) {
1269                         int word_idx = i / BITS_PER_EVTCHN_WORD;
1270                         printk("  %d: event %d -> irq %d%s%s%s\n",
1271                                cpu_from_evtchn(i), i,
1272                                evtchn_to_irq[i],
1273                                sync_test_bit(word_idx, BM(&v->evtchn_pending_sel))
1274                                              ? "" : " l2-clear",
1275                                !sync_test_bit(i, BM(sh->evtchn_mask))
1276                                              ? "" : " globally-masked",
1277                                sync_test_bit(i, BM(cpu_evtchn))
1278                                              ? "" : " locally-masked");
1279                 }
1280         }
1281
1282         spin_unlock_irqrestore(&debug_lock, flags);
1283
1284         return IRQ_HANDLED;
1285 }
1286
1287 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
1288 static DEFINE_PER_CPU(unsigned int, current_word_idx);
1289 static DEFINE_PER_CPU(unsigned int, current_bit_idx);
1290
1291 /*
1292  * Mask out the i least significant bits of w
1293  */
1294 #define MASK_LSBS(w, i) (w & ((~((xen_ulong_t)0UL)) << i))
1295
1296 /*
1297  * Search the CPUs pending events bitmasks.  For each one found, map
1298  * the event number to an irq, and feed it into do_IRQ() for
1299  * handling.
1300  *
1301  * Xen uses a two-level bitmap to speed searching.  The first level is
1302  * a bitset of words which contain pending event bits.  The second
1303  * level is a bitset of pending events themselves.
1304  */
1305 static void __xen_evtchn_do_upcall(void)
1306 {
1307         int start_word_idx, start_bit_idx;
1308         int word_idx, bit_idx;
1309         int i;
1310         int cpu = get_cpu();
1311         struct shared_info *s = HYPERVISOR_shared_info;
1312         struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1313         unsigned count;
1314
1315         do {
1316                 xen_ulong_t pending_words;
1317
1318                 vcpu_info->evtchn_upcall_pending = 0;
1319
1320                 if (__this_cpu_inc_return(xed_nesting_count) - 1)
1321                         goto out;
1322
1323                 /*
1324                  * Master flag must be cleared /before/ clearing
1325                  * selector flag. xchg_xen_ulong must contain an
1326                  * appropriate barrier.
1327                  */
1328                 pending_words = xchg_xen_ulong(&vcpu_info->evtchn_pending_sel, 0);
1329
1330                 start_word_idx = __this_cpu_read(current_word_idx);
1331                 start_bit_idx = __this_cpu_read(current_bit_idx);
1332
1333                 word_idx = start_word_idx;
1334
1335                 for (i = 0; pending_words != 0; i++) {
1336                         xen_ulong_t pending_bits;
1337                         xen_ulong_t words;
1338
1339                         words = MASK_LSBS(pending_words, word_idx);
1340
1341                         /*
1342                          * If we masked out all events, wrap to beginning.
1343                          */
1344                         if (words == 0) {
1345                                 word_idx = 0;
1346                                 bit_idx = 0;
1347                                 continue;
1348                         }
1349                         word_idx = EVTCHN_FIRST_BIT(words);
1350
1351                         pending_bits = active_evtchns(cpu, s, word_idx);
1352                         bit_idx = 0; /* usually scan entire word from start */
1353                         if (word_idx == start_word_idx) {
1354                                 /* We scan the starting word in two parts */
1355                                 if (i == 0)
1356                                         /* 1st time: start in the middle */
1357                                         bit_idx = start_bit_idx;
1358                                 else
1359                                         /* 2nd time: mask bits done already */
1360                                         bit_idx &= (1UL << start_bit_idx) - 1;
1361                         }
1362
1363                         do {
1364                                 xen_ulong_t bits;
1365                                 int port, irq;
1366                                 struct irq_desc *desc;
1367
1368                                 bits = MASK_LSBS(pending_bits, bit_idx);
1369
1370                                 /* If we masked out all events, move on. */
1371                                 if (bits == 0)
1372                                         break;
1373
1374                                 bit_idx = EVTCHN_FIRST_BIT(bits);
1375
1376                                 /* Process port. */
1377                                 port = (word_idx * BITS_PER_EVTCHN_WORD) + bit_idx;
1378                                 irq = evtchn_to_irq[port];
1379
1380                                 if (irq != -1) {
1381                                         desc = irq_to_desc(irq);
1382                                         if (desc)
1383                                                 generic_handle_irq_desc(irq, desc);
1384                                 }
1385
1386                                 bit_idx = (bit_idx + 1) % BITS_PER_EVTCHN_WORD;
1387
1388                                 /* Next caller starts at last processed + 1 */
1389                                 __this_cpu_write(current_word_idx,
1390                                                  bit_idx ? word_idx :
1391                                                  (word_idx+1) % BITS_PER_EVTCHN_WORD);
1392                                 __this_cpu_write(current_bit_idx, bit_idx);
1393                         } while (bit_idx != 0);
1394
1395                         /* Scan start_l1i twice; all others once. */
1396                         if ((word_idx != start_word_idx) || (i != 0))
1397                                 pending_words &= ~(1UL << word_idx);
1398
1399                         word_idx = (word_idx + 1) % BITS_PER_EVTCHN_WORD;
1400                 }
1401
1402                 BUG_ON(!irqs_disabled());
1403
1404                 count = __this_cpu_read(xed_nesting_count);
1405                 __this_cpu_write(xed_nesting_count, 0);
1406         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
1407
1408 out:
1409
1410         put_cpu();
1411 }
1412
1413 void xen_evtchn_do_upcall(struct pt_regs *regs)
1414 {
1415         struct pt_regs *old_regs = set_irq_regs(regs);
1416
1417         irq_enter();
1418 #ifdef CONFIG_X86
1419         exit_idle();
1420 #endif
1421
1422         __xen_evtchn_do_upcall();
1423
1424         irq_exit();
1425         set_irq_regs(old_regs);
1426 }
1427
1428 void xen_hvm_evtchn_do_upcall(void)
1429 {
1430         __xen_evtchn_do_upcall();
1431 }
1432 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1433
1434 /* Rebind a new event channel to an existing irq. */
1435 void rebind_evtchn_irq(int evtchn, int irq)
1436 {
1437         struct irq_info *info = info_for_irq(irq);
1438
1439         /* Make sure the irq is masked, since the new event channel
1440            will also be masked. */
1441         disable_irq(irq);
1442
1443         mutex_lock(&irq_mapping_update_lock);
1444
1445         /* After resume the irq<->evtchn mappings are all cleared out */
1446         BUG_ON(evtchn_to_irq[evtchn] != -1);
1447         /* Expect irq to have been bound before,
1448            so there should be a proper type */
1449         BUG_ON(info->type == IRQT_UNBOUND);
1450
1451         xen_irq_info_evtchn_init(irq, evtchn);
1452
1453         mutex_unlock(&irq_mapping_update_lock);
1454
1455         /* new event channels are always bound to cpu 0 */
1456         irq_set_affinity(irq, cpumask_of(0));
1457
1458         /* Unmask the event channel. */
1459         enable_irq(irq);
1460 }
1461
1462 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1463 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1464 {
1465         struct evtchn_bind_vcpu bind_vcpu;
1466         int evtchn = evtchn_from_irq(irq);
1467
1468         if (!VALID_EVTCHN(evtchn))
1469                 return -1;
1470
1471         /*
1472          * Events delivered via platform PCI interrupts are always
1473          * routed to vcpu 0 and hence cannot be rebound.
1474          */
1475         if (xen_hvm_domain() && !xen_have_vector_callback)
1476                 return -1;
1477
1478         /* Send future instances of this interrupt to other vcpu. */
1479         bind_vcpu.port = evtchn;
1480         bind_vcpu.vcpu = tcpu;
1481
1482         /*
1483          * If this fails, it usually just indicates that we're dealing with a
1484          * virq or IPI channel, which don't actually need to be rebound. Ignore
1485          * it, but don't do the xenlinux-level rebind in that case.
1486          */
1487         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1488                 bind_evtchn_to_cpu(evtchn, tcpu);
1489
1490         return 0;
1491 }
1492
1493 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1494                             bool force)
1495 {
1496         unsigned tcpu = cpumask_first(dest);
1497
1498         return rebind_irq_to_cpu(data->irq, tcpu);
1499 }
1500
1501 int resend_irq_on_evtchn(unsigned int irq)
1502 {
1503         int masked, evtchn = evtchn_from_irq(irq);
1504         struct shared_info *s = HYPERVISOR_shared_info;
1505
1506         if (!VALID_EVTCHN(evtchn))
1507                 return 1;
1508
1509         masked = sync_test_and_set_bit(evtchn, BM(s->evtchn_mask));
1510         sync_set_bit(evtchn, BM(s->evtchn_pending));
1511         if (!masked)
1512                 unmask_evtchn(evtchn);
1513
1514         return 1;
1515 }
1516
1517 static void enable_dynirq(struct irq_data *data)
1518 {
1519         int evtchn = evtchn_from_irq(data->irq);
1520
1521         if (VALID_EVTCHN(evtchn))
1522                 unmask_evtchn(evtchn);
1523 }
1524
1525 static void disable_dynirq(struct irq_data *data)
1526 {
1527         int evtchn = evtchn_from_irq(data->irq);
1528
1529         if (VALID_EVTCHN(evtchn))
1530                 mask_evtchn(evtchn);
1531 }
1532
1533 static void ack_dynirq(struct irq_data *data)
1534 {
1535         int evtchn = evtchn_from_irq(data->irq);
1536
1537         irq_move_irq(data);
1538
1539         if (VALID_EVTCHN(evtchn))
1540                 clear_evtchn(evtchn);
1541 }
1542
1543 static void mask_ack_dynirq(struct irq_data *data)
1544 {
1545         disable_dynirq(data);
1546         ack_dynirq(data);
1547 }
1548
1549 static int retrigger_dynirq(struct irq_data *data)
1550 {
1551         int evtchn = evtchn_from_irq(data->irq);
1552         struct shared_info *sh = HYPERVISOR_shared_info;
1553         int ret = 0;
1554
1555         if (VALID_EVTCHN(evtchn)) {
1556                 int masked;
1557
1558                 masked = sync_test_and_set_bit(evtchn, BM(sh->evtchn_mask));
1559                 sync_set_bit(evtchn, BM(sh->evtchn_pending));
1560                 if (!masked)
1561                         unmask_evtchn(evtchn);
1562                 ret = 1;
1563         }
1564
1565         return ret;
1566 }
1567
1568 static void restore_pirqs(void)
1569 {
1570         int pirq, rc, irq, gsi;
1571         struct physdev_map_pirq map_irq;
1572         struct irq_info *info;
1573
1574         list_for_each_entry(info, &xen_irq_list_head, list) {
1575                 if (info->type != IRQT_PIRQ)
1576                         continue;
1577
1578                 pirq = info->u.pirq.pirq;
1579                 gsi = info->u.pirq.gsi;
1580                 irq = info->irq;
1581
1582                 /* save/restore of PT devices doesn't work, so at this point the
1583                  * only devices present are GSI based emulated devices */
1584                 if (!gsi)
1585                         continue;
1586
1587                 map_irq.domid = DOMID_SELF;
1588                 map_irq.type = MAP_PIRQ_TYPE_GSI;
1589                 map_irq.index = gsi;
1590                 map_irq.pirq = pirq;
1591
1592                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1593                 if (rc) {
1594                         printk(KERN_WARNING "xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1595                                         gsi, irq, pirq, rc);
1596                         xen_free_irq(irq);
1597                         continue;
1598                 }
1599
1600                 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1601
1602                 __startup_pirq(irq);
1603         }
1604 }
1605
1606 static void restore_cpu_virqs(unsigned int cpu)
1607 {
1608         struct evtchn_bind_virq bind_virq;
1609         int virq, irq, evtchn;
1610
1611         for (virq = 0; virq < NR_VIRQS; virq++) {
1612                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1613                         continue;
1614
1615                 BUG_ON(virq_from_irq(irq) != virq);
1616
1617                 /* Get a new binding from Xen. */
1618                 bind_virq.virq = virq;
1619                 bind_virq.vcpu = cpu;
1620                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1621                                                 &bind_virq) != 0)
1622                         BUG();
1623                 evtchn = bind_virq.port;
1624
1625                 /* Record the new mapping. */
1626                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
1627                 bind_evtchn_to_cpu(evtchn, cpu);
1628         }
1629 }
1630
1631 static void restore_cpu_ipis(unsigned int cpu)
1632 {
1633         struct evtchn_bind_ipi bind_ipi;
1634         int ipi, irq, evtchn;
1635
1636         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1637                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1638                         continue;
1639
1640                 BUG_ON(ipi_from_irq(irq) != ipi);
1641
1642                 /* Get a new binding from Xen. */
1643                 bind_ipi.vcpu = cpu;
1644                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1645                                                 &bind_ipi) != 0)
1646                         BUG();
1647                 evtchn = bind_ipi.port;
1648
1649                 /* Record the new mapping. */
1650                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
1651                 bind_evtchn_to_cpu(evtchn, cpu);
1652         }
1653 }
1654
1655 /* Clear an irq's pending state, in preparation for polling on it */
1656 void xen_clear_irq_pending(int irq)
1657 {
1658         int evtchn = evtchn_from_irq(irq);
1659
1660         if (VALID_EVTCHN(evtchn))
1661                 clear_evtchn(evtchn);
1662 }
1663 EXPORT_SYMBOL(xen_clear_irq_pending);
1664 void xen_set_irq_pending(int irq)
1665 {
1666         int evtchn = evtchn_from_irq(irq);
1667
1668         if (VALID_EVTCHN(evtchn))
1669                 set_evtchn(evtchn);
1670 }
1671
1672 bool xen_test_irq_pending(int irq)
1673 {
1674         int evtchn = evtchn_from_irq(irq);
1675         bool ret = false;
1676
1677         if (VALID_EVTCHN(evtchn))
1678                 ret = test_evtchn(evtchn);
1679
1680         return ret;
1681 }
1682
1683 /* Poll waiting for an irq to become pending with timeout.  In the usual case,
1684  * the irq will be disabled so it won't deliver an interrupt. */
1685 void xen_poll_irq_timeout(int irq, u64 timeout)
1686 {
1687         evtchn_port_t evtchn = evtchn_from_irq(irq);
1688
1689         if (VALID_EVTCHN(evtchn)) {
1690                 struct sched_poll poll;
1691
1692                 poll.nr_ports = 1;
1693                 poll.timeout = timeout;
1694                 set_xen_guest_handle(poll.ports, &evtchn);
1695
1696                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1697                         BUG();
1698         }
1699 }
1700 EXPORT_SYMBOL(xen_poll_irq_timeout);
1701 /* Poll waiting for an irq to become pending.  In the usual case, the
1702  * irq will be disabled so it won't deliver an interrupt. */
1703 void xen_poll_irq(int irq)
1704 {
1705         xen_poll_irq_timeout(irq, 0 /* no timeout */);
1706 }
1707
1708 /* Check whether the IRQ line is shared with other guests. */
1709 int xen_test_irq_shared(int irq)
1710 {
1711         struct irq_info *info = info_for_irq(irq);
1712         struct physdev_irq_status_query irq_status = { .irq = info->u.pirq.pirq };
1713
1714         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
1715                 return 0;
1716         return !(irq_status.flags & XENIRQSTAT_shared);
1717 }
1718 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
1719
1720 void xen_irq_resume(void)
1721 {
1722         unsigned int cpu, evtchn;
1723         struct irq_info *info;
1724
1725         init_evtchn_cpu_bindings();
1726
1727         /* New event-channel space is not 'live' yet. */
1728         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1729                 mask_evtchn(evtchn);
1730
1731         /* No IRQ <-> event-channel mappings. */
1732         list_for_each_entry(info, &xen_irq_list_head, list)
1733                 info->evtchn = 0; /* zap event-channel binding */
1734
1735         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1736                 evtchn_to_irq[evtchn] = -1;
1737
1738         for_each_possible_cpu(cpu) {
1739                 restore_cpu_virqs(cpu);
1740                 restore_cpu_ipis(cpu);
1741         }
1742
1743         restore_pirqs();
1744 }
1745
1746 static struct irq_chip xen_dynamic_chip __read_mostly = {
1747         .name                   = "xen-dyn",
1748
1749         .irq_disable            = disable_dynirq,
1750         .irq_mask               = disable_dynirq,
1751         .irq_unmask             = enable_dynirq,
1752
1753         .irq_ack                = ack_dynirq,
1754         .irq_mask_ack           = mask_ack_dynirq,
1755
1756         .irq_set_affinity       = set_affinity_irq,
1757         .irq_retrigger          = retrigger_dynirq,
1758 };
1759
1760 static struct irq_chip xen_pirq_chip __read_mostly = {
1761         .name                   = "xen-pirq",
1762
1763         .irq_startup            = startup_pirq,
1764         .irq_shutdown           = shutdown_pirq,
1765         .irq_enable             = enable_pirq,
1766         .irq_disable            = disable_pirq,
1767
1768         .irq_mask               = disable_dynirq,
1769         .irq_unmask             = enable_dynirq,
1770
1771         .irq_ack                = eoi_pirq,
1772         .irq_eoi                = eoi_pirq,
1773         .irq_mask_ack           = mask_ack_pirq,
1774
1775         .irq_set_affinity       = set_affinity_irq,
1776
1777         .irq_retrigger          = retrigger_dynirq,
1778 };
1779
1780 static struct irq_chip xen_percpu_chip __read_mostly = {
1781         .name                   = "xen-percpu",
1782
1783         .irq_disable            = disable_dynirq,
1784         .irq_mask               = disable_dynirq,
1785         .irq_unmask             = enable_dynirq,
1786
1787         .irq_ack                = ack_dynirq,
1788 };
1789
1790 int xen_set_callback_via(uint64_t via)
1791 {
1792         struct xen_hvm_param a;
1793         a.domid = DOMID_SELF;
1794         a.index = HVM_PARAM_CALLBACK_IRQ;
1795         a.value = via;
1796         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1797 }
1798 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1799
1800 #ifdef CONFIG_XEN_PVHVM
1801 /* Vector callbacks are better than PCI interrupts to receive event
1802  * channel notifications because we can receive vector callbacks on any
1803  * vcpu and we don't need PCI support or APIC interactions. */
1804 void xen_callback_vector(void)
1805 {
1806         int rc;
1807         uint64_t callback_via;
1808         if (xen_have_vector_callback) {
1809                 callback_via = HVM_CALLBACK_VECTOR(HYPERVISOR_CALLBACK_VECTOR);
1810                 rc = xen_set_callback_via(callback_via);
1811                 if (rc) {
1812                         printk(KERN_ERR "Request for Xen HVM callback vector"
1813                                         " failed.\n");
1814                         xen_have_vector_callback = 0;
1815                         return;
1816                 }
1817                 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1818                                 "enabled\n");
1819                 /* in the restore case the vector has already been allocated */
1820                 if (!test_bit(HYPERVISOR_CALLBACK_VECTOR, used_vectors))
1821                         alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR,
1822                                         xen_hvm_callback_vector);
1823         }
1824 }
1825 #else
1826 void xen_callback_vector(void) {}
1827 #endif
1828
1829 void __init xen_init_IRQ(void)
1830 {
1831         int i;
1832
1833         evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1834                                     GFP_KERNEL);
1835         BUG_ON(!evtchn_to_irq);
1836         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1837                 evtchn_to_irq[i] = -1;
1838
1839         init_evtchn_cpu_bindings();
1840
1841         /* No event channels are 'live' right now. */
1842         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1843                 mask_evtchn(i);
1844
1845         pirq_needs_eoi = pirq_needs_eoi_flag;
1846
1847 #ifdef CONFIG_X86
1848         if (xen_hvm_domain()) {
1849                 xen_callback_vector();
1850                 native_init_IRQ();
1851                 /* pci_xen_hvm_init must be called after native_init_IRQ so that
1852                  * __acpi_register_gsi can point at the right function */
1853                 pci_xen_hvm_init();
1854         } else {
1855                 int rc;
1856                 struct physdev_pirq_eoi_gmfn eoi_gmfn;
1857
1858                 irq_ctx_init(smp_processor_id());
1859                 if (xen_initial_domain())
1860                         pci_xen_initial_domain();
1861
1862                 pirq_eoi_map = (void *)__get_free_page(GFP_KERNEL|__GFP_ZERO);
1863                 eoi_gmfn.gmfn = virt_to_mfn(pirq_eoi_map);
1864                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_pirq_eoi_gmfn_v2, &eoi_gmfn);
1865                 if (rc != 0) {
1866                         free_page((unsigned long) pirq_eoi_map);
1867                         pirq_eoi_map = NULL;
1868                 } else
1869                         pirq_needs_eoi = pirq_check_eoi_map;
1870         }
1871 #endif
1872 }