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