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 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
  25
  26 #include <linux/linkage.h>
  27 #include <linux/interrupt.h>
  28 #include <linux/irq.h>
  29 #include <linux/module.h>
  30 #include <linux/string.h>
  31 #include <linux/bootmem.h>
  32 #include <linux/slab.h>
  33 #include <linux/irqnr.h>
  34 #include <linux/pci.h>
  35
  36 #ifdef CONFIG_X86
  37 #include <asm/desc.h>
  38 #include <asm/ptrace.h>
  39 #include <asm/irq.h>
  40 #include <asm/idle.h>
  41 #include <asm/io_apic.h>
  42 #include <asm/xen/page.h>
  43 #include <asm/xen/pci.h>
  44 #endif
  45 #include <asm/sync_bitops.h>
  46 #include <asm/xen/hypercall.h>
  47 #include <asm/xen/hypervisor.h>
  48
  49 #include <xen/xen.h>
  50 #include <xen/hvm.h>
  51 #include <xen/xen-ops.h>
  52 #include <xen/events.h>
  53 #include <xen/interface/xen.h>
  54 #include <xen/interface/event_channel.h>
  55 #include <xen/interface/hvm/hvm_op.h>
  56 #include <xen/interface/hvm/params.h>
  57 #include <xen/interface/physdev.h>
  58 #include <xen/interface/sched.h>
  59 #include <xen/interface/vcpu.h>
  60 #include <asm/hw_irq.h>
  61
  62 /*
  63  * This lock protects updates to the following mapping and reference-count
  64  * arrays. The lock does not need to be acquired to read the mapping tables.
  65  */
  66 static DEFINE_MUTEX(irq_mapping_update_lock);
  67
  68 static LIST_HEAD(xen_irq_list_head);
  69
  70 /* IRQ <-> VIRQ mapping. */
  71 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
  72
  73 /* IRQ <-> IPI mapping */
  74 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
  75
  76 /* Interrupt types. */
  77 enum xen_irq_type {
  78         IRQT_UNBOUND = 0,
  79         IRQT_PIRQ,
  80         IRQT_VIRQ,
  81         IRQT_IPI,
  82         IRQT_EVTCHN
  83 };
  84
  85 /*
  86  * Packed IRQ information:
  87  * type - enum xen_irq_type
  88  * event channel - irq->event channel mapping
  89  * cpu - cpu this event channel is bound to
  90  * index - type-specific information:
  91  *    PIRQ - physical IRQ, GSI, flags, and owner domain
  92  *    VIRQ - virq number
  93  *    IPI - IPI vector
  94  *    EVTCHN -
  95  */
  96 struct irq_info {
  97         struct list_head list;
  98         int refcnt;
  99         enum xen_irq_type type; /* type */
 100         unsigned irq;
 101         unsigned short evtchn;  /* event channel */
 102         unsigned short cpu;     /* cpu bound */
 103
 104         union {
 105                 unsigned short virq;
 106                 enum ipi_vector ipi;
 107                 struct {
 108                         unsigned short pirq;
 109                         unsigned short gsi;
 110                         unsigned char flags;
 111                         uint16_t domid;
 112                 } pirq;
 113         } u;
 114 };
 115 #define PIRQ_NEEDS_EOI  (1 << 0)
 116 #define PIRQ_SHAREABLE  (1 << 1)
 117
 118 static int *evtchn_to_irq;
 119 #ifdef CONFIG_X86
 120 static unsigned long *pirq_eoi_map;
 121 #endif
 122 static bool (*pirq_needs_eoi)(unsigned irq);
 123
 124 /*
 125  * Note sizeof(xen_ulong_t) can be more than sizeof(unsigned long). Be
 126  * careful to only use bitops which allow for this (e.g
 127  * test_bit/find_first_bit and friends but not __ffs) and to pass
 128  * BITS_PER_EVTCHN_WORD as the bitmask length.
 129  */
 130 #define BITS_PER_EVTCHN_WORD (sizeof(xen_ulong_t)*8)
 131 /*
 132  * Make a bitmask (i.e. unsigned long *) of a xen_ulong_t
 133  * array. Primarily to avoid long lines (hence the terse name).
 134  */
 135 #define BM(x) (unsigned long *)(x)
 136 /* Find the first set bit in a evtchn mask */
 137 #define EVTCHN_FIRST_BIT(w) find_first_bit(BM(&(w)), BITS_PER_EVTCHN_WORD)
 138
 139 static DEFINE_PER_CPU(xen_ulong_t [NR_EVENT_CHANNELS/BITS_PER_EVTCHN_WORD],
 140                       cpu_evtchn_mask);
 141
 142 /* Xen will never allocate port zero for any purpose. */
 143 #define VALID_EVTCHN(chn)       ((chn) != 0)
 144
 145 static struct irq_chip xen_dynamic_chip;
 146 static struct irq_chip xen_percpu_chip;
 147 static struct irq_chip xen_pirq_chip;
 148 static void enable_dynirq(struct irq_data *data);
 149 static void disable_dynirq(struct irq_data *data);
 150
 151 /* Get info for IRQ */
 152 static struct irq_info *info_for_irq(unsigned irq)
 153 {
 154         return irq_get_handler_data(irq);
 155 }
 156
 157 /* Constructors for packed IRQ information. */
 158 static void xen_irq_info_common_init(struct irq_info *info,
 159                                      unsigned irq,
 160                                      enum xen_irq_type type,
 161                                      unsigned short evtchn,
 162                                      unsigned short cpu)
 163 {
 164
 165         BUG_ON(info->type != IRQT_UNBOUND && info->type != type);
 166
 167         info->type = type;
 168         info->irq = irq;
 169         info->evtchn = evtchn;
 170         info->cpu = cpu;
 171
 172         evtchn_to_irq[evtchn] = irq;
 173
 174         irq_clear_status_flags(irq, IRQ_NOREQUEST|IRQ_NOAUTOEN);
 175 }
 176
 177 static void xen_irq_info_evtchn_init(unsigned irq,
 178                                      unsigned short evtchn)
 179 {
 180         struct irq_info *info = info_for_irq(irq);
 181
 182         xen_irq_info_common_init(info, irq, IRQT_EVTCHN, evtchn, 0);
 183 }
 184
 185 static void xen_irq_info_ipi_init(unsigned cpu,
 186                                   unsigned irq,
 187                                   unsigned short evtchn,
 188                                   enum ipi_vector ipi)
 189 {
 190         struct irq_info *info = info_for_irq(irq);
 191
 192         xen_irq_info_common_init(info, irq, IRQT_IPI, evtchn, 0);
 193
 194         info->u.ipi = ipi;
 195
 196         per_cpu(ipi_to_irq, cpu)[ipi] = irq;
 197 }
 198
 199 static void xen_irq_info_virq_init(unsigned cpu,
 200                                    unsigned irq,
 201                                    unsigned short evtchn,
 202                                    unsigned short virq)
 203 {
 204         struct irq_info *info = info_for_irq(irq);
 205
 206         xen_irq_info_common_init(info, irq, IRQT_VIRQ, evtchn, 0);
 207
 208         info->u.virq = virq;
 209
 210         per_cpu(virq_to_irq, cpu)[virq] = irq;
 211 }
 212
 213 static void xen_irq_info_pirq_init(unsigned irq,
 214                                    unsigned short evtchn,
 215                                    unsigned short pirq,
 216                                    unsigned short gsi,
 217                                    uint16_t domid,
 218                                    unsigned char flags)
 219 {
 220         struct irq_info *info = info_for_irq(irq);
 221
 222         xen_irq_info_common_init(info, irq, IRQT_PIRQ, evtchn, 0);
 223
 224         info->u.pirq.pirq = pirq;
 225         info->u.pirq.gsi = gsi;
 226         info->u.pirq.domid = domid;
 227         info->u.pirq.flags = flags;
 228 }
 229
 230 /*
 231  * Accessors for packed IRQ information.
 232  */
 233 static unsigned int evtchn_from_irq(unsigned irq)
 234 {
 235         if (unlikely(WARN(irq < 0 || irq >= nr_irqs, "Invalid irq %d!\n", irq)))
 236                 return 0;
 237
 238         return info_for_irq(irq)->evtchn;
 239 }
 240
 241 unsigned irq_from_evtchn(unsigned int evtchn)
 242 {
 243         return evtchn_to_irq[evtchn];
 244 }
 245 EXPORT_SYMBOL_GPL(irq_from_evtchn);
 246
 247 static enum ipi_vector ipi_from_irq(unsigned irq)
 248 {
 249         struct irq_info *info = info_for_irq(irq);
 250
 251         BUG_ON(info == NULL);
 252         BUG_ON(info->type != IRQT_IPI);
 253
 254         return info->u.ipi;
 255 }
 256
 257 static unsigned virq_from_irq(unsigned irq)
 258 {
 259         struct irq_info *info = info_for_irq(irq);
 260
 261         BUG_ON(info == NULL);
 262         BUG_ON(info->type != IRQT_VIRQ);
 263
 264         return info->u.virq;
 265 }
 266
 267 static unsigned pirq_from_irq(unsigned irq)
 268 {
 269         struct irq_info *info = info_for_irq(irq);
 270
 271         BUG_ON(info == NULL);
 272         BUG_ON(info->type != IRQT_PIRQ);
 273
 274         return info->u.pirq.pirq;
 275 }
 276
 277 static enum xen_irq_type type_from_irq(unsigned irq)
 278 {
 279         return info_for_irq(irq)->type;
 280 }
 281
 282 static unsigned cpu_from_irq(unsigned irq)
 283 {
 284         return info_for_irq(irq)->cpu;
 285 }
 286
 287 static unsigned int cpu_from_evtchn(unsigned int evtchn)
 288 {
 289         int irq = evtchn_to_irq[evtchn];
 290         unsigned ret = 0;
 291
 292         if (irq != -1)
 293                 ret = cpu_from_irq(irq);
 294
 295         return ret;
 296 }
 297
 298 #ifdef CONFIG_X86
 299 static bool pirq_check_eoi_map(unsigned irq)
 300 {
 301         return test_bit(pirq_from_irq(irq), pirq_eoi_map);
 302 }
 303 #endif
 304
 305 static bool pirq_needs_eoi_flag(unsigned irq)
 306 {
 307         struct irq_info *info = info_for_irq(irq);
 308         BUG_ON(info->type != IRQT_PIRQ);
 309
 310         return info->u.pirq.flags & PIRQ_NEEDS_EOI;
 311 }
 312
 313 static inline xen_ulong_t active_evtchns(unsigned int cpu,
 314                                          struct shared_info *sh,
 315                                          unsigned int idx)
 316 {
 317         return sh->evtchn_pending[idx] &
 318                 per_cpu(cpu_evtchn_mask, cpu)[idx] &
 319                 ~sh->evtchn_mask[idx];
 320 }
 321
 322 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
 323 {
 324         int irq = evtchn_to_irq[chn];
 325
 326         BUG_ON(irq == -1);
 327 #ifdef CONFIG_SMP
 328         cpumask_copy(irq_to_desc(irq)->irq_data.affinity, cpumask_of(cpu));
 329 #endif
 330
 331         clear_bit(chn, BM(per_cpu(cpu_evtchn_mask, cpu_from_irq(irq))));
 332         set_bit(chn, BM(per_cpu(cpu_evtchn_mask, cpu)));
 333
 334         info_for_irq(irq)->cpu = cpu;
 335 }
 336
 337 static void init_evtchn_cpu_bindings(void)
 338 {
 339         int i;
 340 #ifdef CONFIG_SMP
 341         struct irq_info *info;
 342
 343         /* By default all event channels notify CPU#0. */
 344         list_for_each_entry(info, &xen_irq_list_head, list) {
 345                 struct irq_desc *desc = irq_to_desc(info->irq);
 346                 cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 347         }
 348 #endif
 349
 350         for_each_possible_cpu(i)
 351                 memset(per_cpu(cpu_evtchn_mask, i),
 352                        (i == 0) ? ~0 : 0, NR_EVENT_CHANNELS/8);
 353 }
 354
 355 static inline void clear_evtchn(int port)
 356 {
 357         struct shared_info *s = HYPERVISOR_shared_info;
 358         sync_clear_bit(port, BM(&s->evtchn_pending[0]));
 359 }
 360
 361 static inline void set_evtchn(int port)
 362 {
 363         struct shared_info *s = HYPERVISOR_shared_info;
 364         sync_set_bit(port, BM(&s->evtchn_pending[0]));
 365 }
 366
 367 static inline int test_evtchn(int port)
 368 {
 369         struct shared_info *s = HYPERVISOR_shared_info;
 370         return sync_test_bit(port, BM(&s->evtchn_pending[0]));
 371 }
 372
 373
 374 /**
 375  * notify_remote_via_irq - send event to remote end of event channel via irq
 376  * @irq: irq of event channel to send event to
 377  *
 378  * Unlike notify_remote_via_evtchn(), this is safe to use across
 379  * save/restore. Notifications on a broken connection are silently
 380  * dropped.
 381  */
 382 void notify_remote_via_irq(int irq)
 383 {
 384         int evtchn = evtchn_from_irq(irq);
 385
 386         if (VALID_EVTCHN(evtchn))
 387                 notify_remote_via_evtchn(evtchn);
 388 }
 389 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
 390
 391 static void mask_evtchn(int port)
 392 {
 393         struct shared_info *s = HYPERVISOR_shared_info;
 394         sync_set_bit(port, BM(&s->evtchn_mask[0]));
 395 }
 396
 397 static void unmask_evtchn(int port)
 398 {
 399         struct shared_info *s = HYPERVISOR_shared_info;
 400         unsigned int cpu = get_cpu();
 401         int do_hypercall = 0, evtchn_pending = 0;
 402
 403         BUG_ON(!irqs_disabled());
 404
 405         if (unlikely((cpu != cpu_from_evtchn(port))))
 406                 do_hypercall = 1;
 407         else {
 408                 /*
 409                  * Need to clear the mask before checking pending to
 410                  * avoid a race with an event becoming pending.
 411                  *
 412                  * EVTCHNOP_unmask will only trigger an upcall if the
 413                  * mask bit was set, so if a hypercall is needed
 414                  * remask the event.
 415                  */
 416                 sync_clear_bit(port, BM(&s->evtchn_mask[0]));
 417                 evtchn_pending = sync_test_bit(port, BM(&s->evtchn_pending[0]));
 418
 419                 if (unlikely(evtchn_pending && xen_hvm_domain())) {
 420                         sync_set_bit(port, BM(&s->evtchn_mask[0]));
 421                         do_hypercall = 1;
 422                 }
 423         }
 424
 425         /* Slow path (hypercall) if this is a non-local port or if this is
 426          * an hvm domain and an event is pending (hvm domains don't have
 427          * their own implementation of irq_enable). */
 428         if (do_hypercall) {
 429                 struct evtchn_unmask unmask = { .port = port };
 430                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 431         } else {
 432                 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
 433
 434                 /*
 435                  * The following is basically the equivalent of
 436                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
 437                  * the interrupt edge' if the channel is masked.
 438                  */
 439                 if (evtchn_pending &&
 440                     !sync_test_and_set_bit(port / BITS_PER_EVTCHN_WORD,
 441                                            BM(&vcpu_info->evtchn_pending_sel)))
 442                         vcpu_info->evtchn_upcall_pending = 1;
 443         }
 444
 445         put_cpu();
 446 }
 447
 448 static void xen_irq_init(unsigned irq)
 449 {
 450         struct irq_info *info;
 451 #ifdef CONFIG_SMP
 452         struct irq_desc *desc = irq_to_desc(irq);
 453
 454         /* By default all event channels notify CPU#0. */
 455         cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 456 #endif
 457
 458         info = kzalloc(sizeof(*info), GFP_KERNEL);
 459         if (info == NULL)
 460                 panic("Unable to allocate metadata for IRQ%d\n", irq);
 461
 462         info->type = IRQT_UNBOUND;
 463         info->refcnt = -1;
 464
 465         irq_set_handler_data(irq, info);
 466
 467         list_add_tail(&info->list, &xen_irq_list_head);
 468 }
 469
 470 static int __must_check xen_allocate_irq_dynamic(void)
 471 {
 472         int first = 0;
 473         int irq;
 474
 475 #ifdef CONFIG_X86_IO_APIC
 476         /*
 477          * For an HVM guest or domain 0 which see "real" (emulated or
 478          * actual respectively) GSIs we allocate dynamic IRQs
 479          * e.g. those corresponding to event channels or MSIs
 480          * etc. from the range above those "real" GSIs to avoid
 481          * collisions.
 482          */
 483         if (xen_initial_domain() || xen_hvm_domain())
 484                 first = get_nr_irqs_gsi();
 485 #endif
 486
 487         irq = irq_alloc_desc_from(first, -1);
 488
 489         if (irq >= 0)
 490                 xen_irq_init(irq);
 491
 492         return irq;
 493 }
 494
 495 static int __must_check xen_allocate_irq_gsi(unsigned gsi)
 496 {
 497         int irq;
 498
 499         /*
 500          * A PV guest has no concept of a GSI (since it has no ACPI
 501          * nor access to/knowledge of the physical APICs). Therefore
 502          * all IRQs are dynamically allocated from the entire IRQ
 503          * space.
 504          */
 505         if (xen_pv_domain() && !xen_initial_domain())
 506                 return xen_allocate_irq_dynamic();
 507
 508         /* Legacy IRQ descriptors are already allocated by the arch. */
 509         if (gsi < NR_IRQS_LEGACY)
 510                 irq = gsi;
 511         else
 512                 irq = irq_alloc_desc_at(gsi, -1);
 513
 514         xen_irq_init(irq);
 515
 516         return irq;
 517 }
 518
 519 static void xen_free_irq(unsigned irq)
 520 {
 521         struct irq_info *info = irq_get_handler_data(irq);
 522
 523         if (WARN_ON(!info))
 524                 return;
 525
 526         list_del(&info->list);
 527
 528         irq_set_handler_data(irq, NULL);
 529
 530         WARN_ON(info->refcnt > 0);
 531
 532         kfree(info);
 533
 534         /* Legacy IRQ descriptors are managed by the arch. */
 535         if (irq < NR_IRQS_LEGACY)
 536                 return;
 537
 538         irq_free_desc(irq);
 539 }
 540
 541 static void pirq_query_unmask(int irq)
 542 {
 543         struct physdev_irq_status_query irq_status;
 544         struct irq_info *info = info_for_irq(irq);
 545
 546         BUG_ON(info->type != IRQT_PIRQ);
 547
 548         irq_status.irq = pirq_from_irq(irq);
 549         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
 550                 irq_status.flags = 0;
 551
 552         info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
 553         if (irq_status.flags & XENIRQSTAT_needs_eoi)
 554                 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
 555 }
 556
 557 static bool probing_irq(int irq)
 558 {
 559         struct irq_desc *desc = irq_to_desc(irq);
 560
 561         return desc && desc->action == NULL;
 562 }
 563
 564 static void eoi_pirq(struct irq_data *data)
 565 {
 566         int evtchn = evtchn_from_irq(data->irq);
 567         struct physdev_eoi eoi = { .irq = pirq_from_irq(data->irq) };
 568         int rc = 0;
 569
 570         irq_move_irq(data);
 571
 572         if (VALID_EVTCHN(evtchn))
 573                 clear_evtchn(evtchn);
 574
 575         if (pirq_needs_eoi(data->irq)) {
 576                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
 577                 WARN_ON(rc);
 578         }
 579 }
 580
 581 static void mask_ack_pirq(struct irq_data *data)
 582 {
 583         disable_dynirq(data);
 584         eoi_pirq(data);
 585 }
 586
 587 static unsigned int __startup_pirq(unsigned int irq)
 588 {
 589         struct evtchn_bind_pirq bind_pirq;
 590         struct irq_info *info = info_for_irq(irq);
 591         int evtchn = evtchn_from_irq(irq);
 592         int rc;
 593
 594         BUG_ON(info->type != IRQT_PIRQ);
 595
 596         if (VALID_EVTCHN(evtchn))
 597                 goto out;
 598
 599         bind_pirq.pirq = pirq_from_irq(irq);
 600         /* NB. We are happy to share unless we are probing. */
 601         bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
 602                                         BIND_PIRQ__WILL_SHARE : 0;
 603         rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
 604         if (rc != 0) {
 605                 if (!probing_irq(irq))
 606                         pr_info("Failed to obtain physical IRQ %d\n", irq);
 607                 return 0;
 608         }
 609         evtchn = bind_pirq.port;
 610
 611         pirq_query_unmask(irq);
 612
 613         evtchn_to_irq[evtchn] = irq;
 614         bind_evtchn_to_cpu(evtchn, 0);
 615         info->evtchn = evtchn;
 616
 617 out:
 618         unmask_evtchn(evtchn);
 619         eoi_pirq(irq_get_irq_data(irq));
 620
 621         return 0;
 622 }
 623
 624 static unsigned int startup_pirq(struct irq_data *data)
 625 {
 626         return __startup_pirq(data->irq);
 627 }
 628
 629 static void shutdown_pirq(struct irq_data *data)
 630 {
 631         struct evtchn_close close;
 632         unsigned int irq = data->irq;
 633         struct irq_info *info = info_for_irq(irq);
 634         int evtchn = evtchn_from_irq(irq);
 635
 636         BUG_ON(info->type != IRQT_PIRQ);
 637
 638         if (!VALID_EVTCHN(evtchn))
 639                 return;
 640
 641         mask_evtchn(evtchn);
 642
 643         close.port = evtchn;
 644         if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 645                 BUG();
 646
 647         bind_evtchn_to_cpu(evtchn, 0);
 648         evtchn_to_irq[evtchn] = -1;
 649         info->evtchn = 0;
 650 }
 651
 652 static void enable_pirq(struct irq_data *data)
 653 {
 654         startup_pirq(data);
 655 }
 656
 657 static void disable_pirq(struct irq_data *data)
 658 {
 659         disable_dynirq(data);
 660 }
 661
 662 int xen_irq_from_gsi(unsigned gsi)
 663 {
 664         struct irq_info *info;
 665
 666         list_for_each_entry(info, &xen_irq_list_head, list) {
 667                 if (info->type != IRQT_PIRQ)
 668                         continue;
 669
 670                 if (info->u.pirq.gsi == gsi)
 671                         return info->irq;
 672         }
 673
 674         return -1;
 675 }
 676 EXPORT_SYMBOL_GPL(xen_irq_from_gsi);
 677
 678 /*
 679  * Do not make any assumptions regarding the relationship between the
 680  * IRQ number returned here and the Xen pirq argument.
 681  *
 682  * Note: We don't assign an event channel until the irq actually started
 683  * up.  Return an existing irq if we've already got one for the gsi.
 684  *
 685  * Shareable implies level triggered, not shareable implies edge
 686  * triggered here.
 687  */
 688 int xen_bind_pirq_gsi_to_irq(unsigned gsi,
 689                              unsigned pirq, int shareable, char *name)
 690 {
 691         int irq = -1;
 692         struct physdev_irq irq_op;
 693
 694         mutex_lock(&irq_mapping_update_lock);
 695
 696         irq = xen_irq_from_gsi(gsi);
 697         if (irq != -1) {
 698                 pr_info("%s: returning irq %d for gsi %u\n",
 699                         __func__, irq, gsi);
 700                 goto out;
 701         }
 702
 703         irq = xen_allocate_irq_gsi(gsi);
 704         if (irq < 0)
 705                 goto out;
 706
 707         irq_op.irq = irq;
 708         irq_op.vector = 0;
 709
 710         /* Only the privileged domain can do this. For non-priv, the pcifront
 711          * driver provides a PCI bus that does the call to do exactly
 712          * this in the priv domain. */
 713         if (xen_initial_domain() &&
 714             HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
 715                 xen_free_irq(irq);
 716                 irq = -ENOSPC;
 717                 goto out;
 718         }
 719
 720         xen_irq_info_pirq_init(irq, 0, pirq, gsi, DOMID_SELF,
 721                                shareable ? PIRQ_SHAREABLE : 0);
 722
 723         pirq_query_unmask(irq);
 724         /* We try to use the handler with the appropriate semantic for the
 725          * type of interrupt: if the interrupt is an edge triggered
 726          * interrupt we use handle_edge_irq.
 727          *
 728          * On the other hand if the interrupt is level triggered we use
 729          * handle_fasteoi_irq like the native code does for this kind of
 730          * interrupts.
 731          *
 732          * Depending on the Xen version, pirq_needs_eoi might return true
 733          * not only for level triggered interrupts but for edge triggered
 734          * interrupts too. In any case Xen always honors the eoi mechanism,
 735          * not injecting any more pirqs of the same kind if the first one
 736          * hasn't received an eoi yet. Therefore using the fasteoi handler
 737          * is the right choice either way.
 738          */
 739         if (shareable)
 740                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 741                                 handle_fasteoi_irq, name);
 742         else
 743                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 744                                 handle_edge_irq, name);
 745
 746 out:
 747         mutex_unlock(&irq_mapping_update_lock);
 748
 749         return irq;
 750 }
 751
 752 #ifdef CONFIG_PCI_MSI
 753 int xen_allocate_pirq_msi(struct pci_dev *dev, struct msi_desc *msidesc)
 754 {
 755         int rc;
 756         struct physdev_get_free_pirq op_get_free_pirq;
 757
 758         op_get_free_pirq.type = MAP_PIRQ_TYPE_MSI;
 759         rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
 760
 761         WARN_ONCE(rc == -ENOSYS,
 762                   "hypervisor does not support the PHYSDEVOP_get_free_pirq interface\n");
 763
 764         return rc ? -1 : op_get_free_pirq.pirq;
 765 }
 766
 767 int xen_bind_pirq_msi_to_irq(struct pci_dev *dev, struct msi_desc *msidesc,
 768                              int pirq, const char *name, domid_t domid)
 769 {
 770         int irq, ret;
 771
 772         mutex_lock(&irq_mapping_update_lock);
 773
 774         irq = xen_allocate_irq_dynamic();
 775         if (irq < 0)
 776                 goto out;
 777
 778         irq_set_chip_and_handler_name(irq, &xen_pirq_chip, handle_edge_irq,
 779                         name);
 780
 781         xen_irq_info_pirq_init(irq, 0, pirq, 0, domid, 0);
 782         ret = irq_set_msi_desc(irq, msidesc);
 783         if (ret < 0)
 784                 goto error_irq;
 785 out:
 786         mutex_unlock(&irq_mapping_update_lock);
 787         return irq;
 788 error_irq:
 789         mutex_unlock(&irq_mapping_update_lock);
 790         xen_free_irq(irq);
 791         return ret;
 792 }
 793 #endif
 794
 795 int xen_destroy_irq(int irq)
 796 {
 797         struct irq_desc *desc;
 798         struct physdev_unmap_pirq unmap_irq;
 799         struct irq_info *info = info_for_irq(irq);
 800         int rc = -ENOENT;
 801
 802         mutex_lock(&irq_mapping_update_lock);
 803
 804         desc = irq_to_desc(irq);
 805         if (!desc)
 806                 goto out;
 807
 808         if (xen_initial_domain()) {
 809                 unmap_irq.pirq = info->u.pirq.pirq;
 810                 unmap_irq.domid = info->u.pirq.domid;
 811                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
 812                 /* If another domain quits without making the pci_disable_msix
 813                  * call, the Xen hypervisor takes care of freeing the PIRQs
 814                  * (free_domain_pirqs).
 815                  */
 816                 if ((rc == -ESRCH && info->u.pirq.domid != DOMID_SELF))
 817                         pr_info("domain %d does not have %d anymore\n",
 818                                 info->u.pirq.domid, info->u.pirq.pirq);
 819                 else if (rc) {
 820                         pr_warn("unmap irq failed %d\n", rc);
 821                         goto out;
 822                 }
 823         }
 824
 825         xen_free_irq(irq);
 826
 827 out:
 828         mutex_unlock(&irq_mapping_update_lock);
 829         return rc;
 830 }
 831
 832 int xen_irq_from_pirq(unsigned pirq)
 833 {
 834         int irq;
 835
 836         struct irq_info *info;
 837
 838         mutex_lock(&irq_mapping_update_lock);
 839
 840         list_for_each_entry(info, &xen_irq_list_head, list) {
 841                 if (info->type != IRQT_PIRQ)
 842                         continue;
 843                 irq = info->irq;
 844                 if (info->u.pirq.pirq == pirq)
 845                         goto out;
 846         }
 847         irq = -1;
 848 out:
 849         mutex_unlock(&irq_mapping_update_lock);
 850
 851         return irq;
 852 }
 853
 854
 855 int xen_pirq_from_irq(unsigned irq)
 856 {
 857         return pirq_from_irq(irq);
 858 }
 859 EXPORT_SYMBOL_GPL(xen_pirq_from_irq);
 860 int bind_evtchn_to_irq(unsigned int evtchn)
 861 {
 862         int irq;
 863
 864         mutex_lock(&irq_mapping_update_lock);
 865
 866         irq = evtchn_to_irq[evtchn];
 867
 868         if (irq == -1) {
 869                 irq = xen_allocate_irq_dynamic();
 870                 if (irq < 0)
 871                         goto out;
 872
 873                 irq_set_chip_and_handler_name(irq, &xen_dynamic_chip,
 874                                               handle_edge_irq, "event");
 875
 876                 xen_irq_info_evtchn_init(irq, evtchn);
 877         } else {
 878                 struct irq_info *info = info_for_irq(irq);
 879                 WARN_ON(info == NULL || info->type != IRQT_EVTCHN);
 880         }
 881
 882 out:
 883         mutex_unlock(&irq_mapping_update_lock);
 884
 885         return irq;
 886 }
 887 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
 888
 889 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
 890 {
 891         struct evtchn_bind_ipi bind_ipi;
 892         int evtchn, irq;
 893
 894         mutex_lock(&irq_mapping_update_lock);
 895
 896         irq = per_cpu(ipi_to_irq, cpu)[ipi];
 897
 898         if (irq == -1) {
 899                 irq = xen_allocate_irq_dynamic();
 900                 if (irq < 0)
 901                         goto out;
 902
 903                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 904                                               handle_percpu_irq, "ipi");
 905
 906                 bind_ipi.vcpu = cpu;
 907                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
 908                                                 &bind_ipi) != 0)
 909                         BUG();
 910                 evtchn = bind_ipi.port;
 911
 912                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
 913
 914                 bind_evtchn_to_cpu(evtchn, cpu);
 915         } else {
 916                 struct irq_info *info = info_for_irq(irq);
 917                 WARN_ON(info == NULL || info->type != IRQT_IPI);
 918         }
 919
 920  out:
 921         mutex_unlock(&irq_mapping_update_lock);
 922         return irq;
 923 }
 924
 925 static int bind_interdomain_evtchn_to_irq(unsigned int remote_domain,
 926                                           unsigned int remote_port)
 927 {
 928         struct evtchn_bind_interdomain bind_interdomain;
 929         int err;
 930
 931         bind_interdomain.remote_dom  = remote_domain;
 932         bind_interdomain.remote_port = remote_port;
 933
 934         err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
 935                                           &bind_interdomain);
 936
 937         return err ? : bind_evtchn_to_irq(bind_interdomain.local_port);
 938 }
 939
 940 static int find_virq(unsigned int virq, unsigned int cpu)
 941 {
 942         struct evtchn_status status;
 943         int port, rc = -ENOENT;
 944
 945         memset(&status, 0, sizeof(status));
 946         for (port = 0; port <= NR_EVENT_CHANNELS; port++) {
 947                 status.dom = DOMID_SELF;
 948                 status.port = port;
 949                 rc = HYPERVISOR_event_channel_op(EVTCHNOP_status, &status);
 950                 if (rc < 0)
 951                         continue;
 952                 if (status.status != EVTCHNSTAT_virq)
 953                         continue;
 954                 if (status.u.virq == virq && status.vcpu == cpu) {
 955                         rc = port;
 956                         break;
 957                 }
 958         }
 959         return rc;
 960 }
 961
 962 int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
 963 {
 964         struct evtchn_bind_virq bind_virq;
 965         int evtchn, irq, ret;
 966
 967         mutex_lock(&irq_mapping_update_lock);
 968
 969         irq = per_cpu(virq_to_irq, cpu)[virq];
 970
 971         if (irq == -1) {
 972                 irq = xen_allocate_irq_dynamic();
 973                 if (irq < 0)
 974                         goto out;
 975
 976                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 977                                               handle_percpu_irq, "virq");
 978
 979                 bind_virq.virq = virq;
 980                 bind_virq.vcpu = cpu;
 981                 ret = HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 982                                                 &bind_virq);
 983                 if (ret == 0)
 984                         evtchn = bind_virq.port;
 985                 else {
 986                         if (ret == -EEXIST)
 987                                 ret = find_virq(virq, cpu);
 988                         BUG_ON(ret < 0);
 989                         evtchn = ret;
 990                 }
 991
 992                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
 993
 994                 bind_evtchn_to_cpu(evtchn, cpu);
 995         } else {
 996                 struct irq_info *info = info_for_irq(irq);
 997                 WARN_ON(info == NULL || info->type != IRQT_VIRQ);
 998         }
 999
1000 out:
1001         mutex_unlock(&irq_mapping_update_lock);
1002
1003         return irq;
1004 }
1005
1006 static void unbind_from_irq(unsigned int irq)
1007 {
1008         struct evtchn_close close;
1009         int evtchn = evtchn_from_irq(irq);
1010         struct irq_info *info = irq_get_handler_data(irq);
1011
1012         if (WARN_ON(!info))
1013                 return;
1014
1015         mutex_lock(&irq_mapping_update_lock);
1016
1017         if (info->refcnt > 0) {
1018                 info->refcnt--;
1019                 if (info->refcnt != 0)
1020                         goto done;
1021         }
1022
1023         if (VALID_EVTCHN(evtchn)) {
1024                 close.port = evtchn;
1025                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
1026                         BUG();
1027
1028                 switch (type_from_irq(irq)) {
1029                 case IRQT_VIRQ:
1030                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
1031                                 [virq_from_irq(irq)] = -1;
1032                         break;
1033                 case IRQT_IPI:
1034                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
1035                                 [ipi_from_irq(irq)] = -1;
1036                         break;
1037                 default:
1038                         break;
1039                 }
1040
1041                 /* Closed ports are implicitly re-bound to VCPU0. */
1042                 bind_evtchn_to_cpu(evtchn, 0);
1043
1044                 evtchn_to_irq[evtchn] = -1;
1045         }
1046
1047         BUG_ON(info_for_irq(irq)->type == IRQT_UNBOUND);
1048
1049         xen_free_irq(irq);
1050
1051  done:
1052         mutex_unlock(&irq_mapping_update_lock);
1053 }
1054
1055 int bind_evtchn_to_irqhandler(unsigned int evtchn,
1056                               irq_handler_t handler,
1057                               unsigned long irqflags,
1058                               const char *devname, void *dev_id)
1059 {
1060         int irq, retval;
1061
1062         irq = bind_evtchn_to_irq(evtchn);
1063         if (irq < 0)
1064                 return irq;
1065         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1066         if (retval != 0) {
1067                 unbind_from_irq(irq);
1068                 return retval;
1069         }
1070
1071         return irq;
1072 }
1073 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
1074
1075 int bind_interdomain_evtchn_to_irqhandler(unsigned int remote_domain,
1076                                           unsigned int remote_port,
1077                                           irq_handler_t handler,
1078                                           unsigned long irqflags,
1079                                           const char *devname,
1080                                           void *dev_id)
1081 {
1082         int irq, retval;
1083
1084         irq = bind_interdomain_evtchn_to_irq(remote_domain, remote_port);
1085         if (irq < 0)
1086                 return irq;
1087
1088         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1089         if (retval != 0) {
1090                 unbind_from_irq(irq);
1091                 return retval;
1092         }
1093
1094         return irq;
1095 }
1096 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler);
1097
1098 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
1099                             irq_handler_t handler,
1100                             unsigned long irqflags, const char *devname, void *dev_id)
1101 {
1102         int irq, retval;
1103
1104         irq = bind_virq_to_irq(virq, cpu);
1105         if (irq < 0)
1106                 return irq;
1107         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1108         if (retval != 0) {
1109                 unbind_from_irq(irq);
1110                 return retval;
1111         }
1112
1113         return irq;
1114 }
1115 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1116
1117 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1118                            unsigned int cpu,
1119                            irq_handler_t handler,
1120                            unsigned long irqflags,
1121                            const char *devname,
1122                            void *dev_id)
1123 {
1124         int irq, retval;
1125
1126         irq = bind_ipi_to_irq(ipi, cpu);
1127         if (irq < 0)
1128                 return irq;
1129
1130         irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME | IRQF_EARLY_RESUME;
1131         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1132         if (retval != 0) {
1133                 unbind_from_irq(irq);
1134                 return retval;
1135         }
1136
1137         return irq;
1138 }
1139
1140 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1141 {
1142         struct irq_info *info = irq_get_handler_data(irq);
1143
1144         if (WARN_ON(!info))
1145                 return;
1146         free_irq(irq, dev_id);
1147         unbind_from_irq(irq);
1148 }
1149 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1150
1151 int evtchn_make_refcounted(unsigned int evtchn)
1152 {
1153         int irq = evtchn_to_irq[evtchn];
1154         struct irq_info *info;
1155
1156         if (irq == -1)
1157                 return -ENOENT;
1158
1159         info = irq_get_handler_data(irq);
1160
1161         if (!info)
1162                 return -ENOENT;
1163
1164         WARN_ON(info->refcnt != -1);
1165
1166         info->refcnt = 1;
1167
1168         return 0;
1169 }
1170 EXPORT_SYMBOL_GPL(evtchn_make_refcounted);
1171
1172 int evtchn_get(unsigned int evtchn)
1173 {
1174         int irq;
1175         struct irq_info *info;
1176         int err = -ENOENT;
1177
1178         if (evtchn >= NR_EVENT_CHANNELS)
1179                 return -EINVAL;
1180
1181         mutex_lock(&irq_mapping_update_lock);
1182
1183         irq = evtchn_to_irq[evtchn];
1184         if (irq == -1)
1185                 goto done;
1186
1187         info = irq_get_handler_data(irq);
1188
1189         if (!info)
1190                 goto done;
1191
1192         err = -EINVAL;
1193         if (info->refcnt <= 0)
1194                 goto done;
1195
1196         info->refcnt++;
1197         err = 0;
1198  done:
1199         mutex_unlock(&irq_mapping_update_lock);
1200
1201         return err;
1202 }
1203 EXPORT_SYMBOL_GPL(evtchn_get);
1204
1205 void evtchn_put(unsigned int evtchn)
1206 {
1207         int irq = evtchn_to_irq[evtchn];
1208         if (WARN_ON(irq == -1))
1209                 return;
1210         unbind_from_irq(irq);
1211 }
1212 EXPORT_SYMBOL_GPL(evtchn_put);
1213
1214 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1215 {
1216         int irq;
1217
1218 #ifdef CONFIG_X86
1219         if (unlikely(vector == XEN_NMI_VECTOR)) {
1220                 int rc =  HYPERVISOR_vcpu_op(VCPUOP_send_nmi, cpu, NULL);
1221                 if (rc < 0)
1222                         printk(KERN_WARNING "Sending nmi to CPU%d failed (rc:%d)\n", cpu, rc);
1223                 return;
1224         }
1225 #endif
1226         irq = per_cpu(ipi_to_irq, cpu)[vector];
1227         BUG_ON(irq < 0);
1228         notify_remote_via_irq(irq);
1229 }
1230
1231 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
1232 {
1233         struct shared_info *sh = HYPERVISOR_shared_info;
1234         int cpu = smp_processor_id();
1235         xen_ulong_t *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu);
1236         int i;
1237         unsigned long flags;
1238         static DEFINE_SPINLOCK(debug_lock);
1239         struct vcpu_info *v;
1240
1241         spin_lock_irqsave(&debug_lock, flags);
1242
1243         printk("\nvcpu %d\n  ", cpu);
1244
1245         for_each_online_cpu(i) {
1246                 int pending;
1247                 v = per_cpu(xen_vcpu, i);
1248                 pending = (get_irq_regs() && i == cpu)
1249                         ? xen_irqs_disabled(get_irq_regs())
1250                         : v->evtchn_upcall_mask;
1251                 printk("%d: masked=%d pending=%d event_sel %0*"PRI_xen_ulong"\n  ", i,
1252                        pending, v->evtchn_upcall_pending,
1253                        (int)(sizeof(v->evtchn_pending_sel)*2),
1254                        v->evtchn_pending_sel);
1255         }
1256         v = per_cpu(xen_vcpu, cpu);
1257
1258         printk("\npending:\n   ");
1259         for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
1260                 printk("%0*"PRI_xen_ulong"%s",
1261                        (int)sizeof(sh->evtchn_pending[0])*2,
1262                        sh->evtchn_pending[i],
1263                        i % 8 == 0 ? "\n   " : " ");
1264         printk("\nglobal mask:\n   ");
1265         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1266                 printk("%0*"PRI_xen_ulong"%s",
1267                        (int)(sizeof(sh->evtchn_mask[0])*2),
1268                        sh->evtchn_mask[i],
1269                        i % 8 == 0 ? "\n   " : " ");
1270
1271         printk("\nglobally unmasked:\n   ");
1272         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1273                 printk("%0*"PRI_xen_ulong"%s",
1274                        (int)(sizeof(sh->evtchn_mask[0])*2),
1275                        sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
1276                        i % 8 == 0 ? "\n   " : " ");
1277
1278         printk("\nlocal cpu%d mask:\n   ", cpu);
1279         for (i = (NR_EVENT_CHANNELS/BITS_PER_EVTCHN_WORD)-1; i >= 0; i--)
1280                 printk("%0*"PRI_xen_ulong"%s", (int)(sizeof(cpu_evtchn[0])*2),
1281                        cpu_evtchn[i],
1282                        i % 8 == 0 ? "\n   " : " ");
1283
1284         printk("\nlocally unmasked:\n   ");
1285         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
1286                 xen_ulong_t pending = sh->evtchn_pending[i]
1287                         & ~sh->evtchn_mask[i]
1288                         & cpu_evtchn[i];
1289                 printk("%0*"PRI_xen_ulong"%s",
1290                        (int)(sizeof(sh->evtchn_mask[0])*2),
1291                        pending, i % 8 == 0 ? "\n   " : " ");
1292         }
1293
1294         printk("\npending list:\n");
1295         for (i = 0; i < NR_EVENT_CHANNELS; i++) {
1296                 if (sync_test_bit(i, BM(sh->evtchn_pending))) {
1297                         int word_idx = i / BITS_PER_EVTCHN_WORD;
1298                         printk("  %d: event %d -> irq %d%s%s%s\n",
1299                                cpu_from_evtchn(i), i,
1300                                evtchn_to_irq[i],
1301                                sync_test_bit(word_idx, BM(&v->evtchn_pending_sel))
1302                                              ? "" : " l2-clear",
1303                                !sync_test_bit(i, BM(sh->evtchn_mask))
1304                                              ? "" : " globally-masked",
1305                                sync_test_bit(i, BM(cpu_evtchn))
1306                                              ? "" : " locally-masked");
1307                 }
1308         }
1309
1310         spin_unlock_irqrestore(&debug_lock, flags);
1311
1312         return IRQ_HANDLED;
1313 }
1314
1315 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
1316 static DEFINE_PER_CPU(unsigned int, current_word_idx);
1317 static DEFINE_PER_CPU(unsigned int, current_bit_idx);
1318
1319 /*
1320  * Mask out the i least significant bits of w
1321  */
1322 #define MASK_LSBS(w, i) (w & ((~((xen_ulong_t)0UL)) << i))
1323
1324 /*
1325  * Search the CPUs pending events bitmasks.  For each one found, map
1326  * the event number to an irq, and feed it into do_IRQ() for
1327  * handling.
1328  *
1329  * Xen uses a two-level bitmap to speed searching.  The first level is
1330  * a bitset of words which contain pending event bits.  The second
1331  * level is a bitset of pending events themselves.
1332  */
1333 static void __xen_evtchn_do_upcall(void)
1334 {
1335         int start_word_idx, start_bit_idx;
1336         int word_idx, bit_idx;
1337         int i, irq;
1338         int cpu = get_cpu();
1339         struct shared_info *s = HYPERVISOR_shared_info;
1340         struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1341         unsigned count;
1342
1343         do {
1344                 xen_ulong_t pending_words;
1345                 xen_ulong_t pending_bits;
1346                 struct irq_desc *desc;
1347
1348                 vcpu_info->evtchn_upcall_pending = 0;
1349
1350                 if (__this_cpu_inc_return(xed_nesting_count) - 1)
1351                         goto out;
1352
1353                 /*
1354                  * Master flag must be cleared /before/ clearing
1355                  * selector flag. xchg_xen_ulong must contain an
1356                  * appropriate barrier.
1357                  */
1358                 if ((irq = per_cpu(virq_to_irq, cpu)[VIRQ_TIMER]) != -1) {
1359                         int evtchn = evtchn_from_irq(irq);
1360                         word_idx = evtchn / BITS_PER_LONG;
1361                         pending_bits = evtchn % BITS_PER_LONG;
1362                         if (active_evtchns(cpu, s, word_idx) & (1ULL << pending_bits)) {
1363                                 desc = irq_to_desc(irq);
1364                                 if (desc)
1365                                         generic_handle_irq_desc(irq, desc);
1366                         }
1367                 }
1368
1369                 pending_words = xchg_xen_ulong(&vcpu_info->evtchn_pending_sel, 0);
1370
1371                 start_word_idx = __this_cpu_read(current_word_idx);
1372                 start_bit_idx = __this_cpu_read(current_bit_idx);
1373
1374                 word_idx = start_word_idx;
1375
1376                 for (i = 0; pending_words != 0; i++) {
1377                         xen_ulong_t words;
1378
1379                         words = MASK_LSBS(pending_words, word_idx);
1380
1381                         /*
1382                          * If we masked out all events, wrap to beginning.
1383                          */
1384                         if (words == 0) {
1385                                 word_idx = 0;
1386                                 bit_idx = 0;
1387                                 continue;
1388                         }
1389                         word_idx = EVTCHN_FIRST_BIT(words);
1390
1391                         pending_bits = active_evtchns(cpu, s, word_idx);
1392                         bit_idx = 0; /* usually scan entire word from start */
1393                         /*
1394                          * We scan the starting word in two parts.
1395                          *
1396                          * 1st time: start in the middle, scanning the
1397                          * upper bits.
1398                          *
1399                          * 2nd time: scan the whole word (not just the
1400                          * parts skipped in the first pass) -- if an
1401                          * event in the previously scanned bits is
1402                          * pending again it would just be scanned on
1403                          * the next loop anyway.
1404                          */
1405                         if (word_idx == start_word_idx) {
1406                                 if (i == 0)
1407                                         bit_idx = start_bit_idx;
1408                         }
1409
1410                         do {
1411                                 xen_ulong_t bits;
1412                                 int port;
1413
1414                                 bits = MASK_LSBS(pending_bits, bit_idx);
1415
1416                                 /* If we masked out all events, move on. */
1417                                 if (bits == 0)
1418                                         break;
1419
1420                                 bit_idx = EVTCHN_FIRST_BIT(bits);
1421
1422                                 /* Process port. */
1423                                 port = (word_idx * BITS_PER_EVTCHN_WORD) + bit_idx;
1424                                 irq = evtchn_to_irq[port];
1425
1426                                 if (irq != -1) {
1427                                         desc = irq_to_desc(irq);
1428                                         if (desc)
1429                                                 generic_handle_irq_desc(irq, desc);
1430                                 }
1431
1432                                 bit_idx = (bit_idx + 1) % BITS_PER_EVTCHN_WORD;
1433
1434                                 /* Next caller starts at last processed + 1 */
1435                                 __this_cpu_write(current_word_idx,
1436                                                  bit_idx ? word_idx :
1437                                                  (word_idx+1) % BITS_PER_EVTCHN_WORD);
1438                                 __this_cpu_write(current_bit_idx, bit_idx);
1439                         } while (bit_idx != 0);
1440
1441                         /* Scan start_l1i twice; all others once. */
1442                         if ((word_idx != start_word_idx) || (i != 0))
1443                                 pending_words &= ~(1UL << word_idx);
1444
1445                         word_idx = (word_idx + 1) % BITS_PER_EVTCHN_WORD;
1446                 }
1447
1448                 BUG_ON(!irqs_disabled());
1449
1450                 count = __this_cpu_read(xed_nesting_count);
1451                 __this_cpu_write(xed_nesting_count, 0);
1452         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
1453
1454 out:
1455
1456         put_cpu();
1457 }
1458
1459 void xen_evtchn_do_upcall(struct pt_regs *regs)
1460 {
1461         struct pt_regs *old_regs = set_irq_regs(regs);
1462
1463         irq_enter();
1464 #ifdef CONFIG_X86
1465         exit_idle();
1466 #endif
1467
1468         __xen_evtchn_do_upcall();
1469
1470         irq_exit();
1471         set_irq_regs(old_regs);
1472 }
1473
1474 void xen_hvm_evtchn_do_upcall(void)
1475 {
1476         __xen_evtchn_do_upcall();
1477 }
1478 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1479
1480 /* Rebind a new event channel to an existing irq. */
1481 void rebind_evtchn_irq(int evtchn, int irq)
1482 {
1483         struct irq_info *info = info_for_irq(irq);
1484
1485         if (WARN_ON(!info))
1486                 return;
1487
1488         /* Make sure the irq is masked, since the new event channel
1489            will also be masked. */
1490         disable_irq(irq);
1491
1492         mutex_lock(&irq_mapping_update_lock);
1493
1494         /* After resume the irq<->evtchn mappings are all cleared out */
1495         BUG_ON(evtchn_to_irq[evtchn] != -1);
1496         /* Expect irq to have been bound before,
1497            so there should be a proper type */
1498         BUG_ON(info->type == IRQT_UNBOUND);
1499
1500         xen_irq_info_evtchn_init(irq, evtchn);
1501
1502         mutex_unlock(&irq_mapping_update_lock);
1503
1504         /* new event channels are always bound to cpu 0 */
1505         irq_set_affinity(irq, cpumask_of(0));
1506
1507         /* Unmask the event channel. */
1508         enable_irq(irq);
1509 }
1510
1511 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1512 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1513 {
1514         struct shared_info *s = HYPERVISOR_shared_info;
1515         struct evtchn_bind_vcpu bind_vcpu;
1516         int evtchn = evtchn_from_irq(irq);
1517         int masked;
1518
1519         if (!VALID_EVTCHN(evtchn))
1520                 return -1;
1521
1522         /*
1523          * Events delivered via platform PCI interrupts are always
1524          * routed to vcpu 0 and hence cannot be rebound.
1525          */
1526         if (xen_hvm_domain() && !xen_have_vector_callback)
1527                 return -1;
1528
1529         /* Send future instances of this interrupt to other vcpu. */
1530         bind_vcpu.port = evtchn;
1531         bind_vcpu.vcpu = tcpu;
1532
1533         /*
1534          * Mask the event while changing the VCPU binding to prevent
1535          * it being delivered on an unexpected VCPU.
1536          */
1537         masked = sync_test_and_set_bit(evtchn, BM(s->evtchn_mask));
1538
1539         /*
1540          * If this fails, it usually just indicates that we're dealing with a
1541          * virq or IPI channel, which don't actually need to be rebound. Ignore
1542          * it, but don't do the xenlinux-level rebind in that case.
1543          */
1544         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1545                 bind_evtchn_to_cpu(evtchn, tcpu);
1546
1547         if (!masked)
1548                 unmask_evtchn(evtchn);
1549
1550         return 0;
1551 }
1552
1553 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1554                             bool force)
1555 {
1556         unsigned tcpu = cpumask_first(dest);
1557
1558         return rebind_irq_to_cpu(data->irq, tcpu);
1559 }
1560
1561 int resend_irq_on_evtchn(unsigned int irq)
1562 {
1563         int masked, evtchn = evtchn_from_irq(irq);
1564         struct shared_info *s = HYPERVISOR_shared_info;
1565
1566         if (!VALID_EVTCHN(evtchn))
1567                 return 1;
1568
1569         masked = sync_test_and_set_bit(evtchn, BM(s->evtchn_mask));
1570         sync_set_bit(evtchn, BM(s->evtchn_pending));
1571         if (!masked)
1572                 unmask_evtchn(evtchn);
1573
1574         return 1;
1575 }
1576
1577 static void enable_dynirq(struct irq_data *data)
1578 {
1579         int evtchn = evtchn_from_irq(data->irq);
1580
1581         if (VALID_EVTCHN(evtchn))
1582                 unmask_evtchn(evtchn);
1583 }
1584
1585 static void disable_dynirq(struct irq_data *data)
1586 {
1587         int evtchn = evtchn_from_irq(data->irq);
1588
1589         if (VALID_EVTCHN(evtchn))
1590                 mask_evtchn(evtchn);
1591 }
1592
1593 static void ack_dynirq(struct irq_data *data)
1594 {
1595         int evtchn = evtchn_from_irq(data->irq);
1596
1597         irq_move_irq(data);
1598
1599         if (VALID_EVTCHN(evtchn))
1600                 clear_evtchn(evtchn);
1601 }
1602
1603 static void mask_ack_dynirq(struct irq_data *data)
1604 {
1605         disable_dynirq(data);
1606         ack_dynirq(data);
1607 }
1608
1609 static int retrigger_dynirq(struct irq_data *data)
1610 {
1611         int evtchn = evtchn_from_irq(data->irq);
1612         struct shared_info *sh = HYPERVISOR_shared_info;
1613         int ret = 0;
1614
1615         if (VALID_EVTCHN(evtchn)) {
1616                 int masked;
1617
1618                 masked = sync_test_and_set_bit(evtchn, BM(sh->evtchn_mask));
1619                 sync_set_bit(evtchn, BM(sh->evtchn_pending));
1620                 if (!masked)
1621                         unmask_evtchn(evtchn);
1622                 ret = 1;
1623         }
1624
1625         return ret;
1626 }
1627
1628 static void restore_pirqs(void)
1629 {
1630         int pirq, rc, irq, gsi;
1631         struct physdev_map_pirq map_irq;
1632         struct irq_info *info;
1633
1634         list_for_each_entry(info, &xen_irq_list_head, list) {
1635                 if (info->type != IRQT_PIRQ)
1636                         continue;
1637
1638                 pirq = info->u.pirq.pirq;
1639                 gsi = info->u.pirq.gsi;
1640                 irq = info->irq;
1641
1642                 /* save/restore of PT devices doesn't work, so at this point the
1643                  * only devices present are GSI based emulated devices */
1644                 if (!gsi)
1645                         continue;
1646
1647                 map_irq.domid = DOMID_SELF;
1648                 map_irq.type = MAP_PIRQ_TYPE_GSI;
1649                 map_irq.index = gsi;
1650                 map_irq.pirq = pirq;
1651
1652                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1653                 if (rc) {
1654                         pr_warn("xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1655                                 gsi, irq, pirq, rc);
1656                         xen_free_irq(irq);
1657                         continue;
1658                 }
1659
1660                 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1661
1662                 __startup_pirq(irq);
1663         }
1664 }
1665
1666 static void restore_cpu_virqs(unsigned int cpu)
1667 {
1668         struct evtchn_bind_virq bind_virq;
1669         int virq, irq, evtchn;
1670
1671         for (virq = 0; virq < NR_VIRQS; virq++) {
1672                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1673                         continue;
1674
1675                 BUG_ON(virq_from_irq(irq) != virq);
1676
1677                 /* Get a new binding from Xen. */
1678                 bind_virq.virq = virq;
1679                 bind_virq.vcpu = cpu;
1680                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1681                                                 &bind_virq) != 0)
1682                         BUG();
1683                 evtchn = bind_virq.port;
1684
1685                 /* Record the new mapping. */
1686                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
1687                 bind_evtchn_to_cpu(evtchn, cpu);
1688         }
1689 }
1690
1691 static void restore_cpu_ipis(unsigned int cpu)
1692 {
1693         struct evtchn_bind_ipi bind_ipi;
1694         int ipi, irq, evtchn;
1695
1696         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1697                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1698                         continue;
1699
1700                 BUG_ON(ipi_from_irq(irq) != ipi);
1701
1702                 /* Get a new binding from Xen. */
1703                 bind_ipi.vcpu = cpu;
1704                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1705                                                 &bind_ipi) != 0)
1706                         BUG();
1707                 evtchn = bind_ipi.port;
1708
1709                 /* Record the new mapping. */
1710                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
1711                 bind_evtchn_to_cpu(evtchn, cpu);
1712         }
1713 }
1714
1715 /* Clear an irq's pending state, in preparation for polling on it */
1716 void xen_clear_irq_pending(int irq)
1717 {
1718         int evtchn = evtchn_from_irq(irq);
1719
1720         if (VALID_EVTCHN(evtchn))
1721                 clear_evtchn(evtchn);
1722 }
1723 EXPORT_SYMBOL(xen_clear_irq_pending);
1724 void xen_set_irq_pending(int irq)
1725 {
1726         int evtchn = evtchn_from_irq(irq);
1727
1728         if (VALID_EVTCHN(evtchn))
1729                 set_evtchn(evtchn);
1730 }
1731
1732 bool xen_test_irq_pending(int irq)
1733 {
1734         int evtchn = evtchn_from_irq(irq);
1735         bool ret = false;
1736
1737         if (VALID_EVTCHN(evtchn))
1738                 ret = test_evtchn(evtchn);
1739
1740         return ret;
1741 }
1742
1743 /* Poll waiting for an irq to become pending with timeout.  In the usual case,
1744  * the irq will be disabled so it won't deliver an interrupt. */
1745 void xen_poll_irq_timeout(int irq, u64 timeout)
1746 {
1747         evtchn_port_t evtchn = evtchn_from_irq(irq);
1748
1749         if (VALID_EVTCHN(evtchn)) {
1750                 struct sched_poll poll;
1751
1752                 poll.nr_ports = 1;
1753                 poll.timeout = timeout;
1754                 set_xen_guest_handle(poll.ports, &evtchn);
1755
1756                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1757                         BUG();
1758         }
1759 }
1760 EXPORT_SYMBOL(xen_poll_irq_timeout);
1761 /* Poll waiting for an irq to become pending.  In the usual case, the
1762  * irq will be disabled so it won't deliver an interrupt. */
1763 void xen_poll_irq(int irq)
1764 {
1765         xen_poll_irq_timeout(irq, 0 /* no timeout */);
1766 }
1767
1768 /* Check whether the IRQ line is shared with other guests. */
1769 int xen_test_irq_shared(int irq)
1770 {
1771         struct irq_info *info = info_for_irq(irq);
1772         struct physdev_irq_status_query irq_status;
1773
1774         if (WARN_ON(!info))
1775                 return -ENOENT;
1776
1777         irq_status.irq = info->u.pirq.pirq;
1778
1779         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
1780                 return 0;
1781         return !(irq_status.flags & XENIRQSTAT_shared);
1782 }
1783 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
1784
1785 void xen_irq_resume(void)
1786 {
1787         unsigned int cpu, evtchn;
1788         struct irq_info *info;
1789
1790         init_evtchn_cpu_bindings();
1791
1792         /* New event-channel space is not 'live' yet. */
1793         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1794                 mask_evtchn(evtchn);
1795
1796         /* No IRQ <-> event-channel mappings. */
1797         list_for_each_entry(info, &xen_irq_list_head, list)
1798                 info->evtchn = 0; /* zap event-channel binding */
1799
1800         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1801                 evtchn_to_irq[evtchn] = -1;
1802
1803         for_each_possible_cpu(cpu) {
1804                 restore_cpu_virqs(cpu);
1805                 restore_cpu_ipis(cpu);
1806         }
1807
1808         restore_pirqs();
1809 }
1810
1811 static struct irq_chip xen_dynamic_chip __read_mostly = {
1812         .name                   = "xen-dyn",
1813
1814         .irq_disable            = disable_dynirq,
1815         .irq_mask               = disable_dynirq,
1816         .irq_unmask             = enable_dynirq,
1817
1818         .irq_ack                = ack_dynirq,
1819         .irq_mask_ack           = mask_ack_dynirq,
1820
1821         .irq_set_affinity       = set_affinity_irq,
1822         .irq_retrigger          = retrigger_dynirq,
1823 };
1824
1825 static struct irq_chip xen_pirq_chip __read_mostly = {
1826         .name                   = "xen-pirq",
1827
1828         .irq_startup            = startup_pirq,
1829         .irq_shutdown           = shutdown_pirq,
1830         .irq_enable             = enable_pirq,
1831         .irq_disable            = disable_pirq,
1832
1833         .irq_mask               = disable_dynirq,
1834         .irq_unmask             = enable_dynirq,
1835
1836         .irq_ack                = eoi_pirq,
1837         .irq_eoi                = eoi_pirq,
1838         .irq_mask_ack           = mask_ack_pirq,
1839
1840         .irq_set_affinity       = set_affinity_irq,
1841
1842         .irq_retrigger          = retrigger_dynirq,
1843 };
1844
1845 static struct irq_chip xen_percpu_chip __read_mostly = {
1846         .name                   = "xen-percpu",
1847
1848         .irq_disable            = disable_dynirq,
1849         .irq_mask               = disable_dynirq,
1850         .irq_unmask             = enable_dynirq,
1851
1852         .irq_ack                = ack_dynirq,
1853 };
1854
1855 int xen_set_callback_via(uint64_t via)
1856 {
1857         struct xen_hvm_param a;
1858         a.domid = DOMID_SELF;
1859         a.index = HVM_PARAM_CALLBACK_IRQ;
1860         a.value = via;
1861         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1862 }
1863 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1864
1865 #ifdef CONFIG_XEN_PVHVM
1866 /* Vector callbacks are better than PCI interrupts to receive event
1867  * channel notifications because we can receive vector callbacks on any
1868  * vcpu and we don't need PCI support or APIC interactions. */
1869 void xen_callback_vector(void)
1870 {
1871         int rc;
1872         uint64_t callback_via;
1873         if (xen_have_vector_callback) {
1874                 callback_via = HVM_CALLBACK_VECTOR(HYPERVISOR_CALLBACK_VECTOR);
1875                 rc = xen_set_callback_via(callback_via);
1876                 if (rc) {
1877                         pr_err("Request for Xen HVM callback vector failed\n");
1878                         xen_have_vector_callback = 0;
1879                         return;
1880                 }
1881                 pr_info("Xen HVM callback vector for event delivery is enabled\n");
1882                 /* in the restore case the vector has already been allocated */
1883                 if (!test_bit(HYPERVISOR_CALLBACK_VECTOR, used_vectors))
1884                         alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR,
1885                                         xen_hvm_callback_vector);
1886         }
1887 }
1888 #else
1889 void xen_callback_vector(void) {}
1890 #endif
1891
1892 void __init xen_init_IRQ(void)
1893 {
1894         int i;
1895
1896         evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1897                                     GFP_KERNEL);
1898         BUG_ON(!evtchn_to_irq);
1899         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1900                 evtchn_to_irq[i] = -1;
1901
1902         init_evtchn_cpu_bindings();
1903
1904         /* No event channels are 'live' right now. */
1905         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1906                 mask_evtchn(i);
1907
1908         pirq_needs_eoi = pirq_needs_eoi_flag;
1909
1910 #ifdef CONFIG_X86
1911         if (xen_hvm_domain()) {
1912                 xen_callback_vector();
1913                 native_init_IRQ();
1914                 /* pci_xen_hvm_init must be called after native_init_IRQ so that
1915                  * __acpi_register_gsi can point at the right function */
1916                 pci_xen_hvm_init();
1917         } else {
1918                 int rc;
1919                 struct physdev_pirq_eoi_gmfn eoi_gmfn;
1920
1921                 irq_ctx_init(smp_processor_id());
1922                 if (xen_initial_domain())
1923                         pci_xen_initial_domain();
1924
1925                 pirq_eoi_map = (void *)__get_free_page(GFP_KERNEL|__GFP_ZERO);
1926                 eoi_gmfn.gmfn = virt_to_mfn(pirq_eoi_map);
1927                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_pirq_eoi_gmfn_v2, &eoi_gmfn);
1928                 if (rc != 0) {
1929                         free_page((unsigned long) pirq_eoi_map);
1930                         pirq_eoi_map = NULL;
1931                 } else
1932                         pirq_needs_eoi = pirq_check_eoi_map;
1933         }
1934 #endif
1935 }