drivers/xen/events.c

   1 /*
   2  * Xen event channels
   3  *
   4  * Xen models interrupts with abstract event channels.  Because each
   5  * domain gets 1024 event channels, but NR_IRQ is not that large, we
   6  * must dynamically map irqs<->event channels.  The event channels
   7  * interface with the rest of the kernel by defining a xen interrupt
   8  * chip.  When an event is received, it is mapped to an irq and sent
   9  * through the normal interrupt processing path.
  10  *
  11  * There are four kinds of events which can be mapped to an event
  12  * channel:
  13  *
  14  * 1. Inter-domain notifications.  This includes all the virtual
  15  *    device events, since they're driven by front-ends in another domain
  16  *    (typically dom0).
  17  * 2. VIRQs, typically used for timers.  These are per-cpu events.
  18  * 3. IPIs.
  19  * 4. PIRQs - Hardware interrupts.
  20  *
  21  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
  22  */
  23
  24 #include <linux/linkage.h>
  25 #include <linux/interrupt.h>
  26 #include <linux/irq.h>
  27 #include <linux/module.h>
  28 #include <linux/string.h>
  29 #include <linux/bootmem.h>
  30 #include <linux/slab.h>
  31 #include <linux/irqnr.h>
  32 #include <linux/pci.h>
  33
  34 #ifdef CONFIG_X86
  35 #include <asm/desc.h>
  36 #include <asm/ptrace.h>
  37 #include <asm/irq.h>
  38 #include <asm/idle.h>
  39 #include <asm/io_apic.h>
  40 #include <asm/xen/page.h>
  41 #include <asm/xen/pci.h>
  42 #endif
  43 #include <asm/sync_bitops.h>
  44 #include <asm/xen/hypercall.h>
  45 #include <asm/xen/hypervisor.h>
  46
  47 #include <xen/xen.h>
  48 #include <xen/hvm.h>
  49 #include <xen/xen-ops.h>
  50 #include <xen/events.h>
  51 #include <xen/interface/xen.h>
  52 #include <xen/interface/event_channel.h>
  53 #include <xen/interface/hvm/hvm_op.h>
  54 #include <xen/interface/hvm/params.h>
  55 #include <xen/interface/physdev.h>
  56 #include <xen/interface/sched.h>
  57 #include <asm/hw_irq.h>
  58
  59 /*
  60  * This lock protects updates to the following mapping and reference-count
  61  * arrays. The lock does not need to be acquired to read the mapping tables.
  62  */
  63 static DEFINE_MUTEX(irq_mapping_update_lock);
  64
  65 static LIST_HEAD(xen_irq_list_head);
  66
  67 /* IRQ <-> VIRQ mapping. */
  68 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
  69
  70 /* IRQ <-> IPI mapping */
  71 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
  72
  73 /* Interrupt types. */
  74 enum xen_irq_type {
  75         IRQT_UNBOUND = 0,
  76         IRQT_PIRQ,
  77         IRQT_VIRQ,
  78         IRQT_IPI,
  79         IRQT_EVTCHN
  80 };
  81
  82 /*
  83  * Packed IRQ information:
  84  * type - enum xen_irq_type
  85  * event channel - irq->event channel mapping
  86  * cpu - cpu this event channel is bound to
  87  * index - type-specific information:
  88  *    PIRQ - physical IRQ, GSI, flags, and owner domain
  89  *    VIRQ - virq number
  90  *    IPI - IPI vector
  91  *    EVTCHN -
  92  */
  93 struct irq_info {
  94         struct list_head list;
  95         int refcnt;
  96         enum xen_irq_type type; /* type */
  97         unsigned irq;
  98         unsigned short evtchn;  /* event channel */
  99         unsigned short cpu;     /* cpu bound */
 100
 101         union {
 102                 unsigned short virq;
 103                 enum ipi_vector ipi;
 104                 struct {
 105                         unsigned short pirq;
 106                         unsigned short gsi;
 107                         unsigned char flags;
 108                         uint16_t domid;
 109                 } pirq;
 110         } u;
 111 };
 112 #define PIRQ_NEEDS_EOI  (1 << 0)
 113 #define PIRQ_SHAREABLE  (1 << 1)
 114
 115 static int *evtchn_to_irq;
 116 #ifdef CONFIG_X86
 117 static unsigned long *pirq_eoi_map;
 118 #endif
 119 static bool (*pirq_needs_eoi)(unsigned irq);
 120
 121 /*
 122  * Note sizeof(xen_ulong_t) can be more than sizeof(unsigned long). Be
 123  * careful to only use bitops which allow for this (e.g
 124  * test_bit/find_first_bit and friends but not __ffs) and to pass
 125  * BITS_PER_EVTCHN_WORD as the bitmask length.
 126  */
 127 #define BITS_PER_EVTCHN_WORD (sizeof(xen_ulong_t)*8)
 128 /*
 129  * Make a bitmask (i.e. unsigned long *) of a xen_ulong_t
 130  * array. Primarily to avoid long lines (hence the terse name).
 131  */
 132 #define BM(x) (unsigned long *)(x)
 133 /* Find the first set bit in a evtchn mask */
 134 #define EVTCHN_FIRST_BIT(w) find_first_bit(BM(&(w)), BITS_PER_EVTCHN_WORD)
 135
 136 static DEFINE_PER_CPU(xen_ulong_t [NR_EVENT_CHANNELS/BITS_PER_EVTCHN_WORD],
 137                       cpu_evtchn_mask);
 138
 139 /* Xen will never allocate port zero for any purpose. */
 140 #define VALID_EVTCHN(chn)       ((chn) != 0)
 141
 142 static struct irq_chip xen_dynamic_chip;
 143 static struct irq_chip xen_percpu_chip;
 144 static struct irq_chip xen_pirq_chip;
 145 static void enable_dynirq(struct irq_data *data);
 146 static void disable_dynirq(struct irq_data *data);
 147
 148 /* Get info for IRQ */
 149 static struct irq_info *info_for_irq(unsigned irq)
 150 {
 151         return irq_get_handler_data(irq);
 152 }
 153
 154 /* Constructors for packed IRQ information. */
 155 static void xen_irq_info_common_init(struct irq_info *info,
 156                                      unsigned irq,
 157                                      enum xen_irq_type type,
 158                                      unsigned short evtchn,
 159                                      unsigned short cpu)
 160 {
 161
 162         BUG_ON(info->type != IRQT_UNBOUND && info->type != type);
 163
 164         info->type = type;
 165         info->irq = irq;
 166         info->evtchn = evtchn;
 167         info->cpu = cpu;
 168
 169         evtchn_to_irq[evtchn] = irq;
 170
 171         irq_clear_status_flags(irq, IRQ_NOREQUEST|IRQ_NOAUTOEN);
 172 }
 173
 174 static void xen_irq_info_evtchn_init(unsigned irq,
 175                                      unsigned short evtchn)
 176 {
 177         struct irq_info *info = info_for_irq(irq);
 178
 179         xen_irq_info_common_init(info, irq, IRQT_EVTCHN, evtchn, 0);
 180 }
 181
 182 static void xen_irq_info_ipi_init(unsigned cpu,
 183                                   unsigned irq,
 184                                   unsigned short evtchn,
 185                                   enum ipi_vector ipi)
 186 {
 187         struct irq_info *info = info_for_irq(irq);
 188
 189         xen_irq_info_common_init(info, irq, IRQT_IPI, evtchn, 0);
 190
 191         info->u.ipi = ipi;
 192
 193         per_cpu(ipi_to_irq, cpu)[ipi] = irq;
 194 }
 195
 196 static void xen_irq_info_virq_init(unsigned cpu,
 197                                    unsigned irq,
 198                                    unsigned short evtchn,
 199                                    unsigned short virq)
 200 {
 201         struct irq_info *info = info_for_irq(irq);
 202
 203         xen_irq_info_common_init(info, irq, IRQT_VIRQ, evtchn, 0);
 204
 205         info->u.virq = virq;
 206
 207         per_cpu(virq_to_irq, cpu)[virq] = irq;
 208 }
 209
 210 static void xen_irq_info_pirq_init(unsigned irq,
 211                                    unsigned short evtchn,
 212                                    unsigned short pirq,
 213                                    unsigned short gsi,
 214                                    uint16_t domid,
 215                                    unsigned char flags)
 216 {
 217         struct irq_info *info = info_for_irq(irq);
 218
 219         xen_irq_info_common_init(info, irq, IRQT_PIRQ, evtchn, 0);
 220
 221         info->u.pirq.pirq = pirq;
 222         info->u.pirq.gsi = gsi;
 223         info->u.pirq.domid = domid;
 224         info->u.pirq.flags = flags;
 225 }
 226
 227 /*
 228  * Accessors for packed IRQ information.
 229  */
 230 static unsigned int evtchn_from_irq(unsigned irq)
 231 {
 232         if (unlikely(WARN(irq < 0 || irq >= nr_irqs, "Invalid irq %d!\n", irq)))
 233                 return 0;
 234
 235         return info_for_irq(irq)->evtchn;
 236 }
 237
 238 unsigned irq_from_evtchn(unsigned int evtchn)
 239 {
 240         return evtchn_to_irq[evtchn];
 241 }
 242 EXPORT_SYMBOL_GPL(irq_from_evtchn);
 243
 244 static enum ipi_vector ipi_from_irq(unsigned irq)
 245 {
 246         struct irq_info *info = info_for_irq(irq);
 247
 248         BUG_ON(info == NULL);
 249         BUG_ON(info->type != IRQT_IPI);
 250
 251         return info->u.ipi;
 252 }
 253
 254 static unsigned virq_from_irq(unsigned irq)
 255 {
 256         struct irq_info *info = info_for_irq(irq);
 257
 258         BUG_ON(info == NULL);
 259         BUG_ON(info->type != IRQT_VIRQ);
 260
 261         return info->u.virq;
 262 }
 263
 264 static unsigned pirq_from_irq(unsigned irq)
 265 {
 266         struct irq_info *info = info_for_irq(irq);
 267
 268         BUG_ON(info == NULL);
 269         BUG_ON(info->type != IRQT_PIRQ);
 270
 271         return info->u.pirq.pirq;
 272 }
 273
 274 static enum xen_irq_type type_from_irq(unsigned irq)
 275 {
 276         return info_for_irq(irq)->type;
 277 }
 278
 279 static unsigned cpu_from_irq(unsigned irq)
 280 {
 281         return info_for_irq(irq)->cpu;
 282 }
 283
 284 static unsigned int cpu_from_evtchn(unsigned int evtchn)
 285 {
 286         int irq = evtchn_to_irq[evtchn];
 287         unsigned ret = 0;
 288
 289         if (irq != -1)
 290                 ret = cpu_from_irq(irq);
 291
 292         return ret;
 293 }
 294
 295 #ifdef CONFIG_X86
 296 static bool pirq_check_eoi_map(unsigned irq)
 297 {
 298         return test_bit(pirq_from_irq(irq), pirq_eoi_map);
 299 }
 300 #endif
 301
 302 static bool pirq_needs_eoi_flag(unsigned irq)
 303 {
 304         struct irq_info *info = info_for_irq(irq);
 305         BUG_ON(info->type != IRQT_PIRQ);
 306
 307         return info->u.pirq.flags & PIRQ_NEEDS_EOI;
 308 }
 309
 310 static inline xen_ulong_t active_evtchns(unsigned int cpu,
 311                                          struct shared_info *sh,
 312                                          unsigned int idx)
 313 {
 314         return sh->evtchn_pending[idx] &
 315                 per_cpu(cpu_evtchn_mask, cpu)[idx] &
 316                 ~sh->evtchn_mask[idx];
 317 }
 318
 319 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
 320 {
 321         int irq = evtchn_to_irq[chn];
 322
 323         BUG_ON(irq == -1);
 324 #ifdef CONFIG_SMP
 325         cpumask_copy(irq_to_desc(irq)->irq_data.affinity, cpumask_of(cpu));
 326 #endif
 327
 328         clear_bit(chn, BM(per_cpu(cpu_evtchn_mask, cpu_from_irq(irq))));
 329         set_bit(chn, BM(per_cpu(cpu_evtchn_mask, cpu)));
 330
 331         info_for_irq(irq)->cpu = cpu;
 332 }
 333
 334 static void init_evtchn_cpu_bindings(void)
 335 {
 336         int i;
 337 #ifdef CONFIG_SMP
 338         struct irq_info *info;
 339
 340         /* By default all event channels notify CPU#0. */
 341         list_for_each_entry(info, &xen_irq_list_head, list) {
 342                 struct irq_desc *desc = irq_to_desc(info->irq);
 343                 cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 344         }
 345 #endif
 346
 347         for_each_possible_cpu(i)
 348                 memset(per_cpu(cpu_evtchn_mask, i),
 349                        (i == 0) ? ~0 : 0, NR_EVENT_CHANNELS/8);
 350 }
 351
 352 static inline void clear_evtchn(int port)
 353 {
 354         struct shared_info *s = HYPERVISOR_shared_info;
 355         sync_clear_bit(port, BM(&s->evtchn_pending[0]));
 356 }
 357
 358 static inline void set_evtchn(int port)
 359 {
 360         struct shared_info *s = HYPERVISOR_shared_info;
 361         sync_set_bit(port, BM(&s->evtchn_pending[0]));
 362 }
 363
 364 static inline int test_evtchn(int port)
 365 {
 366         struct shared_info *s = HYPERVISOR_shared_info;
 367         return sync_test_bit(port, BM(&s->evtchn_pending[0]));
 368 }
 369
 370
 371 /**
 372  * notify_remote_via_irq - send event to remote end of event channel via irq
 373  * @irq: irq of event channel to send event to
 374  *
 375  * Unlike notify_remote_via_evtchn(), this is safe to use across
 376  * save/restore. Notifications on a broken connection are silently
 377  * dropped.
 378  */
 379 void notify_remote_via_irq(int irq)
 380 {
 381         int evtchn = evtchn_from_irq(irq);
 382
 383         if (VALID_EVTCHN(evtchn))
 384                 notify_remote_via_evtchn(evtchn);
 385 }
 386 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
 387
 388 static void mask_evtchn(int port)
 389 {
 390         struct shared_info *s = HYPERVISOR_shared_info;
 391         sync_set_bit(port, BM(&s->evtchn_mask[0]));
 392 }
 393
 394 static void unmask_evtchn(int port)
 395 {
 396         struct shared_info *s = HYPERVISOR_shared_info;
 397         unsigned int cpu = get_cpu();
 398         int do_hypercall = 0, evtchn_pending = 0;
 399
 400         BUG_ON(!irqs_disabled());
 401
 402         if (unlikely((cpu != cpu_from_evtchn(port))))
 403                 do_hypercall = 1;
 404         else {
 405                 /*
 406                  * Need to clear the mask before checking pending to
 407                  * avoid a race with an event becoming pending.
 408                  *
 409                  * EVTCHNOP_unmask will only trigger an upcall if the
 410                  * mask bit was set, so if a hypercall is needed
 411                  * remask the event.
 412                  */
 413                 sync_clear_bit(port, BM(&s->evtchn_mask[0]));
 414                 evtchn_pending = sync_test_bit(port, BM(&s->evtchn_pending[0]));
 415
 416                 if (unlikely(evtchn_pending && xen_hvm_domain())) {
 417                         sync_set_bit(port, BM(&s->evtchn_mask[0]));
 418                         do_hypercall = 1;
 419                 }
 420         }
 421
 422         /* Slow path (hypercall) if this is a non-local port or if this is
 423          * an hvm domain and an event is pending (hvm domains don't have
 424          * their own implementation of irq_enable). */
 425         if (do_hypercall) {
 426                 struct evtchn_unmask unmask = { .port = port };
 427                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 428         } else {
 429                 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
 430
 431                 /*
 432                  * The following is basically the equivalent of
 433                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
 434                  * the interrupt edge' if the channel is masked.
 435                  */
 436                 if (evtchn_pending &&
 437                     !sync_test_and_set_bit(port / BITS_PER_EVTCHN_WORD,
 438                                            BM(&vcpu_info->evtchn_pending_sel)))
 439                         vcpu_info->evtchn_upcall_pending = 1;
 440         }
 441
 442         put_cpu();
 443 }
 444
 445 static void xen_irq_init(unsigned irq)
 446 {
 447         struct irq_info *info;
 448 #ifdef CONFIG_SMP
 449         struct irq_desc *desc = irq_to_desc(irq);
 450
 451         /* By default all event channels notify CPU#0. */
 452         cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 453 #endif
 454
 455         info = kzalloc(sizeof(*info), GFP_KERNEL);
 456         if (info == NULL)
 457                 panic("Unable to allocate metadata for IRQ%d\n", irq);
 458
 459         info->type = IRQT_UNBOUND;
 460         info->refcnt = -1;
 461
 462         irq_set_handler_data(irq, info);
 463
 464         list_add_tail(&info->list, &xen_irq_list_head);
 465 }
 466
 467 static int __must_check xen_allocate_irq_dynamic(void)
 468 {
 469         int first = 0;
 470         int irq;
 471
 472 #ifdef CONFIG_X86_IO_APIC
 473         /*
 474          * For an HVM guest or domain 0 which see "real" (emulated or
 475          * actual respectively) GSIs we allocate dynamic IRQs
 476          * e.g. those corresponding to event channels or MSIs
 477          * etc. from the range above those "real" GSIs to avoid
 478          * collisions.
 479          */
 480         if (xen_initial_domain() || xen_hvm_domain())
 481                 first = get_nr_irqs_gsi();
 482 #endif
 483
 484         irq = irq_alloc_desc_from(first, -1);
 485
 486         if (irq >= 0)
 487                 xen_irq_init(irq);
 488
 489         return irq;
 490 }
 491
 492 static int __must_check xen_allocate_irq_gsi(unsigned gsi)
 493 {
 494         int irq;
 495
 496         /*
 497          * A PV guest has no concept of a GSI (since it has no ACPI
 498          * nor access to/knowledge of the physical APICs). Therefore
 499          * all IRQs are dynamically allocated from the entire IRQ
 500          * space.
 501          */
 502         if (xen_pv_domain() && !xen_initial_domain())
 503                 return xen_allocate_irq_dynamic();
 504
 505         /* Legacy IRQ descriptors are already allocated by the arch. */
 506         if (gsi < NR_IRQS_LEGACY)
 507                 irq = gsi;
 508         else
 509                 irq = irq_alloc_desc_at(gsi, -1);
 510
 511         xen_irq_init(irq);
 512
 513         return irq;
 514 }
 515
 516 static void xen_free_irq(unsigned irq)
 517 {
 518         struct irq_info *info = irq_get_handler_data(irq);
 519
 520         if (WARN_ON(!info))
 521                 return;
 522
 523         list_del(&info->list);
 524
 525         irq_set_handler_data(irq, NULL);
 526
 527         WARN_ON(info->refcnt > 0);
 528
 529         kfree(info);
 530
 531         /* Legacy IRQ descriptors are managed by the arch. */
 532         if (irq < NR_IRQS_LEGACY)
 533                 return;
 534
 535         irq_free_desc(irq);
 536 }
 537
 538 static void pirq_query_unmask(int irq)
 539 {
 540         struct physdev_irq_status_query irq_status;
 541         struct irq_info *info = info_for_irq(irq);
 542
 543         BUG_ON(info->type != IRQT_PIRQ);
 544
 545         irq_status.irq = pirq_from_irq(irq);
 546         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
 547                 irq_status.flags = 0;
 548
 549         info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
 550         if (irq_status.flags & XENIRQSTAT_needs_eoi)
 551                 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
 552 }
 553
 554 static bool probing_irq(int irq)
 555 {
 556         struct irq_desc *desc = irq_to_desc(irq);
 557
 558         return desc && desc->action == NULL;
 559 }
 560
 561 static void eoi_pirq(struct irq_data *data)
 562 {
 563         int evtchn = evtchn_from_irq(data->irq);
 564         struct physdev_eoi eoi = { .irq = pirq_from_irq(data->irq) };
 565         int rc = 0;
 566
 567         irq_move_irq(data);
 568
 569         if (VALID_EVTCHN(evtchn))
 570                 clear_evtchn(evtchn);
 571
 572         if (pirq_needs_eoi(data->irq)) {
 573                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
 574                 WARN_ON(rc);
 575         }
 576 }
 577
 578 static void mask_ack_pirq(struct irq_data *data)
 579 {
 580         disable_dynirq(data);
 581         eoi_pirq(data);
 582 }
 583
 584 static unsigned int __startup_pirq(unsigned int irq)
 585 {
 586         struct evtchn_bind_pirq bind_pirq;
 587         struct irq_info *info = info_for_irq(irq);
 588         int evtchn = evtchn_from_irq(irq);
 589         int rc;
 590
 591         BUG_ON(info->type != IRQT_PIRQ);
 592
 593         if (VALID_EVTCHN(evtchn))
 594                 goto out;
 595
 596         bind_pirq.pirq = pirq_from_irq(irq);
 597         /* NB. We are happy to share unless we are probing. */
 598         bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
 599                                         BIND_PIRQ__WILL_SHARE : 0;
 600         rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
 601         if (rc != 0) {
 602                 if (!probing_irq(irq))
 603                         printk(KERN_INFO "Failed to obtain physical IRQ %d\n",
 604                                irq);
 605                 return 0;
 606         }
 607         evtchn = bind_pirq.port;
 608
 609         pirq_query_unmask(irq);
 610
 611         evtchn_to_irq[evtchn] = irq;
 612         bind_evtchn_to_cpu(evtchn, 0);
 613         info->evtchn = evtchn;
 614
 615 out:
 616         unmask_evtchn(evtchn);
 617         eoi_pirq(irq_get_irq_data(irq));
 618
 619         return 0;
 620 }
 621
 622 static unsigned int startup_pirq(struct irq_data *data)
 623 {
 624         return __startup_pirq(data->irq);
 625 }
 626
 627 static void shutdown_pirq(struct irq_data *data)
 628 {
 629         struct evtchn_close close;
 630         unsigned int irq = data->irq;
 631         struct irq_info *info = info_for_irq(irq);
 632         int evtchn = evtchn_from_irq(irq);
 633
 634         BUG_ON(info->type != IRQT_PIRQ);
 635
 636         if (!VALID_EVTCHN(evtchn))
 637                 return;
 638
 639         mask_evtchn(evtchn);
 640
 641         close.port = evtchn;
 642         if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 643                 BUG();
 644
 645         bind_evtchn_to_cpu(evtchn, 0);
 646         evtchn_to_irq[evtchn] = -1;
 647         info->evtchn = 0;
 648 }
 649
 650 static void enable_pirq(struct irq_data *data)
 651 {
 652         startup_pirq(data);
 653 }
 654
 655 static void disable_pirq(struct irq_data *data)
 656 {
 657         disable_dynirq(data);
 658 }
 659
 660 int xen_irq_from_gsi(unsigned gsi)
 661 {
 662         struct irq_info *info;
 663
 664         list_for_each_entry(info, &xen_irq_list_head, list) {
 665                 if (info->type != IRQT_PIRQ)
 666                         continue;
 667
 668                 if (info->u.pirq.gsi == gsi)
 669                         return info->irq;
 670         }
 671
 672         return -1;
 673 }
 674 EXPORT_SYMBOL_GPL(xen_irq_from_gsi);
 675
 676 /*
 677  * Do not make any assumptions regarding the relationship between the
 678  * IRQ number returned here and the Xen pirq argument.
 679  *
 680  * Note: We don't assign an event channel until the irq actually started
 681  * up.  Return an existing irq if we've already got one for the gsi.
 682  *
 683  * Shareable implies level triggered, not shareable implies edge
 684  * triggered here.
 685  */
 686 int xen_bind_pirq_gsi_to_irq(unsigned gsi,
 687                              unsigned pirq, int shareable, char *name)
 688 {
 689         int irq = -1;
 690         struct physdev_irq irq_op;
 691
 692         mutex_lock(&irq_mapping_update_lock);
 693
 694         irq = xen_irq_from_gsi(gsi);
 695         if (irq != -1) {
 696                 printk(KERN_INFO "xen_map_pirq_gsi: returning irq %d for gsi %u\n",
 697                        irq, gsi);
 698                 goto out;
 699         }
 700
 701         irq = xen_allocate_irq_gsi(gsi);
 702         if (irq < 0)
 703                 goto out;
 704
 705         irq_op.irq = irq;
 706         irq_op.vector = 0;
 707
 708         /* Only the privileged domain can do this. For non-priv, the pcifront
 709          * driver provides a PCI bus that does the call to do exactly
 710          * this in the priv domain. */
 711         if (xen_initial_domain() &&
 712             HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
 713                 xen_free_irq(irq);
 714                 irq = -ENOSPC;
 715                 goto out;
 716         }
 717
 718         xen_irq_info_pirq_init(irq, 0, pirq, gsi, DOMID_SELF,
 719                                shareable ? PIRQ_SHAREABLE : 0);
 720
 721         pirq_query_unmask(irq);
 722         /* We try to use the handler with the appropriate semantic for the
 723          * type of interrupt: if the interrupt is an edge triggered
 724          * interrupt we use handle_edge_irq.
 725          *
 726          * On the other hand if the interrupt is level triggered we use
 727          * handle_fasteoi_irq like the native code does for this kind of
 728          * interrupts.
 729          *
 730          * Depending on the Xen version, pirq_needs_eoi might return true
 731          * not only for level triggered interrupts but for edge triggered
 732          * interrupts too. In any case Xen always honors the eoi mechanism,
 733          * not injecting any more pirqs of the same kind if the first one
 734          * hasn't received an eoi yet. Therefore using the fasteoi handler
 735          * is the right choice either way.
 736          */
 737         if (shareable)
 738                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 739                                 handle_fasteoi_irq, name);
 740         else
 741                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 742                                 handle_edge_irq, name);
 743
 744 out:
 745         mutex_unlock(&irq_mapping_update_lock);
 746
 747         return irq;
 748 }
 749
 750 #ifdef CONFIG_PCI_MSI
 751 int xen_allocate_pirq_msi(struct pci_dev *dev, struct msi_desc *msidesc)
 752 {
 753         int rc;
 754         struct physdev_get_free_pirq op_get_free_pirq;
 755
 756         op_get_free_pirq.type = MAP_PIRQ_TYPE_MSI;
 757         rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
 758
 759         WARN_ONCE(rc == -ENOSYS,
 760                   "hypervisor does not support the PHYSDEVOP_get_free_pirq interface\n");
 761
 762         return rc ? -1 : op_get_free_pirq.pirq;
 763 }
 764
 765 int xen_bind_pirq_msi_to_irq(struct pci_dev *dev, struct msi_desc *msidesc,
 766                              int pirq, const char *name, domid_t domid)
 767 {
 768         int irq, ret;
 769
 770         mutex_lock(&irq_mapping_update_lock);
 771
 772         irq = xen_allocate_irq_dynamic();
 773         if (irq < 0)
 774                 goto out;
 775
 776         irq_set_chip_and_handler_name(irq, &xen_pirq_chip, handle_edge_irq,
 777                         name);
 778
 779         xen_irq_info_pirq_init(irq, 0, pirq, 0, domid, 0);
 780         ret = irq_set_msi_desc(irq, msidesc);
 781         if (ret < 0)
 782                 goto error_irq;
 783 out:
 784         mutex_unlock(&irq_mapping_update_lock);
 785         return irq;
 786 error_irq:
 787         mutex_unlock(&irq_mapping_update_lock);
 788         xen_free_irq(irq);
 789         return ret;
 790 }
 791 #endif
 792
 793 int xen_destroy_irq(int irq)
 794 {
 795         struct irq_desc *desc;
 796         struct physdev_unmap_pirq unmap_irq;
 797         struct irq_info *info = info_for_irq(irq);
 798         int rc = -ENOENT;
 799
 800         mutex_lock(&irq_mapping_update_lock);
 801
 802         desc = irq_to_desc(irq);
 803         if (!desc)
 804                 goto out;
 805
 806         if (xen_initial_domain()) {
 807                 unmap_irq.pirq = info->u.pirq.pirq;
 808                 unmap_irq.domid = info->u.pirq.domid;
 809                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
 810                 /* If another domain quits without making the pci_disable_msix
 811                  * call, the Xen hypervisor takes care of freeing the PIRQs
 812                  * (free_domain_pirqs).
 813                  */
 814                 if ((rc == -ESRCH && info->u.pirq.domid != DOMID_SELF))
 815                         printk(KERN_INFO "domain %d does not have %d anymore\n",
 816                                 info->u.pirq.domid, info->u.pirq.pirq);
 817                 else if (rc) {
 818                         printk(KERN_WARNING "unmap irq failed %d\n", rc);
 819                         goto out;
 820                 }
 821         }
 822
 823         xen_free_irq(irq);
 824
 825 out:
 826         mutex_unlock(&irq_mapping_update_lock);
 827         return rc;
 828 }
 829
 830 int xen_irq_from_pirq(unsigned pirq)
 831 {
 832         int irq;
 833
 834         struct irq_info *info;
 835
 836         mutex_lock(&irq_mapping_update_lock);
 837
 838         list_for_each_entry(info, &xen_irq_list_head, list) {
 839                 if (info->type != IRQT_PIRQ)
 840                         continue;
 841                 irq = info->irq;
 842                 if (info->u.pirq.pirq == pirq)
 843                         goto out;
 844         }
 845         irq = -1;
 846 out:
 847         mutex_unlock(&irq_mapping_update_lock);
 848
 849         return irq;
 850 }
 851
 852
 853 int xen_pirq_from_irq(unsigned irq)
 854 {
 855         return pirq_from_irq(irq);
 856 }
 857 EXPORT_SYMBOL_GPL(xen_pirq_from_irq);
 858 int bind_evtchn_to_irq(unsigned int evtchn)
 859 {
 860         int irq;
 861
 862         mutex_lock(&irq_mapping_update_lock);
 863
 864         irq = evtchn_to_irq[evtchn];
 865
 866         if (irq == -1) {
 867                 irq = xen_allocate_irq_dynamic();
 868                 if (irq < 0)
 869                         goto out;
 870
 871                 irq_set_chip_and_handler_name(irq, &xen_dynamic_chip,
 872                                               handle_edge_irq, "event");
 873
 874                 xen_irq_info_evtchn_init(irq, evtchn);
 875         } else {
 876                 struct irq_info *info = info_for_irq(irq);
 877                 WARN_ON(info == NULL || info->type != IRQT_EVTCHN);
 878         }
 879
 880 out:
 881         mutex_unlock(&irq_mapping_update_lock);
 882
 883         return irq;
 884 }
 885 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
 886
 887 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
 888 {
 889         struct evtchn_bind_ipi bind_ipi;
 890         int evtchn, irq;
 891
 892         mutex_lock(&irq_mapping_update_lock);
 893
 894         irq = per_cpu(ipi_to_irq, cpu)[ipi];
 895
 896         if (irq == -1) {
 897                 irq = xen_allocate_irq_dynamic();
 898                 if (irq < 0)
 899                         goto out;
 900
 901                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 902                                               handle_percpu_irq, "ipi");
 903
 904                 bind_ipi.vcpu = cpu;
 905                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
 906                                                 &bind_ipi) != 0)
 907                         BUG();
 908                 evtchn = bind_ipi.port;
 909
 910                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
 911
 912                 bind_evtchn_to_cpu(evtchn, cpu);
 913         } else {
 914                 struct irq_info *info = info_for_irq(irq);
 915                 WARN_ON(info == NULL || info->type != IRQT_IPI);
 916         }
 917
 918  out:
 919         mutex_unlock(&irq_mapping_update_lock);
 920         return irq;
 921 }
 922
 923 static int bind_interdomain_evtchn_to_irq(unsigned int remote_domain,
 924                                           unsigned int remote_port)
 925 {
 926         struct evtchn_bind_interdomain bind_interdomain;
 927         int err;
 928
 929         bind_interdomain.remote_dom  = remote_domain;
 930         bind_interdomain.remote_port = remote_port;
 931
 932         err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
 933                                           &bind_interdomain);
 934
 935         return err ? : bind_evtchn_to_irq(bind_interdomain.local_port);
 936 }
 937
 938 static int find_virq(unsigned int virq, unsigned int cpu)
 939 {
 940         struct evtchn_status status;
 941         int port, rc = -ENOENT;
 942
 943         memset(&status, 0, sizeof(status));
 944         for (port = 0; port <= NR_EVENT_CHANNELS; port++) {
 945                 status.dom = DOMID_SELF;
 946                 status.port = port;
 947                 rc = HYPERVISOR_event_channel_op(EVTCHNOP_status, &status);
 948                 if (rc < 0)
 949                         continue;
 950                 if (status.status != EVTCHNSTAT_virq)
 951                         continue;
 952                 if (status.u.virq == virq && status.vcpu == cpu) {
 953                         rc = port;
 954                         break;
 955                 }
 956         }
 957         return rc;
 958 }
 959
 960 int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
 961 {
 962         struct evtchn_bind_virq bind_virq;
 963         int evtchn, irq, ret;
 964
 965         mutex_lock(&irq_mapping_update_lock);
 966
 967         irq = per_cpu(virq_to_irq, cpu)[virq];
 968
 969         if (irq == -1) {
 970                 irq = xen_allocate_irq_dynamic();
 971                 if (irq < 0)
 972                         goto out;
 973
 974                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 975                                               handle_percpu_irq, "virq");
 976
 977                 bind_virq.virq = virq;
 978                 bind_virq.vcpu = cpu;
 979                 ret = HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 980                                                 &bind_virq);
 981                 if (ret == 0)
 982                         evtchn = bind_virq.port;
 983                 else {
 984                         if (ret == -EEXIST)
 985                                 ret = find_virq(virq, cpu);
 986                         BUG_ON(ret < 0);
 987                         evtchn = ret;
 988                 }
 989
 990                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
 991
 992                 bind_evtchn_to_cpu(evtchn, cpu);
 993         } else {
 994                 struct irq_info *info = info_for_irq(irq);
 995                 WARN_ON(info == NULL || info->type != IRQT_VIRQ);
 996         }
 997
 998 out:
 999         mutex_unlock(&irq_mapping_update_lock);
1000
1001         return irq;
1002 }
1003
1004 static void unbind_from_irq(unsigned int irq)
1005 {
1006         struct evtchn_close close;
1007         int evtchn = evtchn_from_irq(irq);
1008         struct irq_info *info = irq_get_handler_data(irq);
1009
1010         if (WARN_ON(!info))
1011                 return;
1012
1013         mutex_lock(&irq_mapping_update_lock);
1014
1015         if (info->refcnt > 0) {
1016                 info->refcnt--;
1017                 if (info->refcnt != 0)
1018                         goto done;
1019         }
1020
1021         if (VALID_EVTCHN(evtchn)) {
1022                 close.port = evtchn;
1023                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
1024                         BUG();
1025
1026                 switch (type_from_irq(irq)) {
1027                 case IRQT_VIRQ:
1028                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
1029                                 [virq_from_irq(irq)] = -1;
1030                         break;
1031                 case IRQT_IPI:
1032                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
1033                                 [ipi_from_irq(irq)] = -1;
1034                         break;
1035                 default:
1036                         break;
1037                 }
1038
1039                 /* Closed ports are implicitly re-bound to VCPU0. */
1040                 bind_evtchn_to_cpu(evtchn, 0);
1041
1042                 evtchn_to_irq[evtchn] = -1;
1043         }
1044
1045         BUG_ON(info_for_irq(irq)->type == IRQT_UNBOUND);
1046
1047         xen_free_irq(irq);
1048
1049  done:
1050         mutex_unlock(&irq_mapping_update_lock);
1051 }
1052
1053 int bind_evtchn_to_irqhandler(unsigned int evtchn,
1054                               irq_handler_t handler,
1055                               unsigned long irqflags,
1056                               const char *devname, void *dev_id)
1057 {
1058         int irq, retval;
1059
1060         irq = bind_evtchn_to_irq(evtchn);
1061         if (irq < 0)
1062                 return irq;
1063         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1064         if (retval != 0) {
1065                 unbind_from_irq(irq);
1066                 return retval;
1067         }
1068
1069         return irq;
1070 }
1071 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
1072
1073 int bind_interdomain_evtchn_to_irqhandler(unsigned int remote_domain,
1074                                           unsigned int remote_port,
1075                                           irq_handler_t handler,
1076                                           unsigned long irqflags,
1077                                           const char *devname,
1078                                           void *dev_id)
1079 {
1080         int irq, retval;
1081
1082         irq = bind_interdomain_evtchn_to_irq(remote_domain, remote_port);
1083         if (irq < 0)
1084                 return irq;
1085
1086         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1087         if (retval != 0) {
1088                 unbind_from_irq(irq);
1089                 return retval;
1090         }
1091
1092         return irq;
1093 }
1094 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler);
1095
1096 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
1097                             irq_handler_t handler,
1098                             unsigned long irqflags, const char *devname, void *dev_id)
1099 {
1100         int irq, retval;
1101
1102         irq = bind_virq_to_irq(virq, cpu);
1103         if (irq < 0)
1104                 return irq;
1105         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1106         if (retval != 0) {
1107                 unbind_from_irq(irq);
1108                 return retval;
1109         }
1110
1111         return irq;
1112 }
1113 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1114
1115 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1116                            unsigned int cpu,
1117                            irq_handler_t handler,
1118                            unsigned long irqflags,
1119                            const char *devname,
1120                            void *dev_id)
1121 {
1122         int irq, retval;
1123
1124         irq = bind_ipi_to_irq(ipi, cpu);
1125         if (irq < 0)
1126                 return irq;
1127
1128         irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME | IRQF_EARLY_RESUME;
1129         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1130         if (retval != 0) {
1131                 unbind_from_irq(irq);
1132                 return retval;
1133         }
1134
1135         return irq;
1136 }
1137
1138 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1139 {
1140         struct irq_info *info = irq_get_handler_data(irq);
1141
1142         if (WARN_ON(!info))
1143                 return;
1144         free_irq(irq, dev_id);
1145         unbind_from_irq(irq);
1146 }
1147 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1148
1149 int evtchn_make_refcounted(unsigned int evtchn)
1150 {
1151         int irq = evtchn_to_irq[evtchn];
1152         struct irq_info *info;
1153
1154         if (irq == -1)
1155                 return -ENOENT;
1156
1157         info = irq_get_handler_data(irq);
1158
1159         if (!info)
1160                 return -ENOENT;
1161
1162         WARN_ON(info->refcnt != -1);
1163
1164         info->refcnt = 1;
1165
1166         return 0;
1167 }
1168 EXPORT_SYMBOL_GPL(evtchn_make_refcounted);
1169
1170 int evtchn_get(unsigned int evtchn)
1171 {
1172         int irq;
1173         struct irq_info *info;
1174         int err = -ENOENT;
1175
1176         if (evtchn >= NR_EVENT_CHANNELS)
1177                 return -EINVAL;
1178
1179         mutex_lock(&irq_mapping_update_lock);
1180
1181         irq = evtchn_to_irq[evtchn];
1182         if (irq == -1)
1183                 goto done;
1184
1185         info = irq_get_handler_data(irq);
1186
1187         if (!info)
1188                 goto done;
1189
1190         err = -EINVAL;
1191         if (info->refcnt <= 0)
1192                 goto done;
1193
1194         info->refcnt++;
1195         err = 0;
1196  done:
1197         mutex_unlock(&irq_mapping_update_lock);
1198
1199         return err;
1200 }
1201 EXPORT_SYMBOL_GPL(evtchn_get);
1202
1203 void evtchn_put(unsigned int evtchn)
1204 {
1205         int irq = evtchn_to_irq[evtchn];
1206         if (WARN_ON(irq == -1))
1207                 return;
1208         unbind_from_irq(irq);
1209 }
1210 EXPORT_SYMBOL_GPL(evtchn_put);
1211
1212 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1213 {
1214         int 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                         if (word_idx == start_word_idx) {
1382                                 /* We scan the starting word in two parts */
1383                                 if (i == 0)
1384                                         /* 1st time: start in the middle */
1385                                         bit_idx = start_bit_idx;
1386                                 else
1387                                         /* 2nd time: mask bits done already */
1388                                         bit_idx &= (1UL << start_bit_idx) - 1;
1389                         }
1390
1391                         do {
1392                                 xen_ulong_t bits;
1393                                 int port;
1394
1395                                 bits = MASK_LSBS(pending_bits, bit_idx);
1396
1397                                 /* If we masked out all events, move on. */
1398                                 if (bits == 0)
1399                                         break;
1400
1401                                 bit_idx = EVTCHN_FIRST_BIT(bits);
1402
1403                                 /* Process port. */
1404                                 port = (word_idx * BITS_PER_EVTCHN_WORD) + bit_idx;
1405                                 irq = evtchn_to_irq[port];
1406
1407                                 if (irq != -1) {
1408                                         desc = irq_to_desc(irq);
1409                                         if (desc)
1410                                                 generic_handle_irq_desc(irq, desc);
1411                                 }
1412
1413                                 bit_idx = (bit_idx + 1) % BITS_PER_EVTCHN_WORD;
1414
1415                                 /* Next caller starts at last processed + 1 */
1416                                 __this_cpu_write(current_word_idx,
1417                                                  bit_idx ? word_idx :
1418                                                  (word_idx+1) % BITS_PER_EVTCHN_WORD);
1419                                 __this_cpu_write(current_bit_idx, bit_idx);
1420                         } while (bit_idx != 0);
1421
1422                         /* Scan start_l1i twice; all others once. */
1423                         if ((word_idx != start_word_idx) || (i != 0))
1424                                 pending_words &= ~(1UL << word_idx);
1425
1426                         word_idx = (word_idx + 1) % BITS_PER_EVTCHN_WORD;
1427                 }
1428
1429                 BUG_ON(!irqs_disabled());
1430
1431                 count = __this_cpu_read(xed_nesting_count);
1432                 __this_cpu_write(xed_nesting_count, 0);
1433         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
1434
1435 out:
1436
1437         put_cpu();
1438 }
1439
1440 void xen_evtchn_do_upcall(struct pt_regs *regs)
1441 {
1442         struct pt_regs *old_regs = set_irq_regs(regs);
1443
1444         irq_enter();
1445 #ifdef CONFIG_X86
1446         exit_idle();
1447 #endif
1448
1449         __xen_evtchn_do_upcall();
1450
1451         irq_exit();
1452         set_irq_regs(old_regs);
1453 }
1454
1455 void xen_hvm_evtchn_do_upcall(void)
1456 {
1457         __xen_evtchn_do_upcall();
1458 }
1459 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1460
1461 /* Rebind a new event channel to an existing irq. */
1462 void rebind_evtchn_irq(int evtchn, int irq)
1463 {
1464         struct irq_info *info = info_for_irq(irq);
1465
1466         if (WARN_ON(!info))
1467                 return;
1468
1469         /* Make sure the irq is masked, since the new event channel
1470            will also be masked. */
1471         disable_irq(irq);
1472
1473         mutex_lock(&irq_mapping_update_lock);
1474
1475         /* After resume the irq<->evtchn mappings are all cleared out */
1476         BUG_ON(evtchn_to_irq[evtchn] != -1);
1477         /* Expect irq to have been bound before,
1478            so there should be a proper type */
1479         BUG_ON(info->type == IRQT_UNBOUND);
1480
1481         xen_irq_info_evtchn_init(irq, evtchn);
1482
1483         mutex_unlock(&irq_mapping_update_lock);
1484
1485         /* new event channels are always bound to cpu 0 */
1486         irq_set_affinity(irq, cpumask_of(0));
1487
1488         /* Unmask the event channel. */
1489         enable_irq(irq);
1490 }
1491
1492 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1493 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1494 {
1495         struct shared_info *s = HYPERVISOR_shared_info;
1496         struct evtchn_bind_vcpu bind_vcpu;
1497         int evtchn = evtchn_from_irq(irq);
1498         int masked;
1499
1500         if (!VALID_EVTCHN(evtchn))
1501                 return -1;
1502
1503         /*
1504          * Events delivered via platform PCI interrupts are always
1505          * routed to vcpu 0 and hence cannot be rebound.
1506          */
1507         if (xen_hvm_domain() && !xen_have_vector_callback)
1508                 return -1;
1509
1510         /* Send future instances of this interrupt to other vcpu. */
1511         bind_vcpu.port = evtchn;
1512         bind_vcpu.vcpu = tcpu;
1513
1514         /*
1515          * Mask the event while changing the VCPU binding to prevent
1516          * it being delivered on an unexpected VCPU.
1517          */
1518         masked = sync_test_and_set_bit(evtchn, BM(s->evtchn_mask));
1519
1520         /*
1521          * If this fails, it usually just indicates that we're dealing with a
1522          * virq or IPI channel, which don't actually need to be rebound. Ignore
1523          * it, but don't do the xenlinux-level rebind in that case.
1524          */
1525         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1526                 bind_evtchn_to_cpu(evtchn, tcpu);
1527
1528         if (!masked)
1529                 unmask_evtchn(evtchn);
1530
1531         return 0;
1532 }
1533
1534 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1535                             bool force)
1536 {
1537         unsigned tcpu = cpumask_first(dest);
1538
1539         return rebind_irq_to_cpu(data->irq, tcpu);
1540 }
1541
1542 int resend_irq_on_evtchn(unsigned int irq)
1543 {
1544         int masked, evtchn = evtchn_from_irq(irq);
1545         struct shared_info *s = HYPERVISOR_shared_info;
1546
1547         if (!VALID_EVTCHN(evtchn))
1548                 return 1;
1549
1550         masked = sync_test_and_set_bit(evtchn, BM(s->evtchn_mask));
1551         sync_set_bit(evtchn, BM(s->evtchn_pending));
1552         if (!masked)
1553                 unmask_evtchn(evtchn);
1554
1555         return 1;
1556 }
1557
1558 static void enable_dynirq(struct irq_data *data)
1559 {
1560         int evtchn = evtchn_from_irq(data->irq);
1561
1562         if (VALID_EVTCHN(evtchn))
1563                 unmask_evtchn(evtchn);
1564 }
1565
1566 static void disable_dynirq(struct irq_data *data)
1567 {
1568         int evtchn = evtchn_from_irq(data->irq);
1569
1570         if (VALID_EVTCHN(evtchn))
1571                 mask_evtchn(evtchn);
1572 }
1573
1574 static void ack_dynirq(struct irq_data *data)
1575 {
1576         int evtchn = evtchn_from_irq(data->irq);
1577
1578         irq_move_irq(data);
1579
1580         if (VALID_EVTCHN(evtchn))
1581                 clear_evtchn(evtchn);
1582 }
1583
1584 static void mask_ack_dynirq(struct irq_data *data)
1585 {
1586         disable_dynirq(data);
1587         ack_dynirq(data);
1588 }
1589
1590 static int retrigger_dynirq(struct irq_data *data)
1591 {
1592         int evtchn = evtchn_from_irq(data->irq);
1593         struct shared_info *sh = HYPERVISOR_shared_info;
1594         int ret = 0;
1595
1596         if (VALID_EVTCHN(evtchn)) {
1597                 int masked;
1598
1599                 masked = sync_test_and_set_bit(evtchn, BM(sh->evtchn_mask));
1600                 sync_set_bit(evtchn, BM(sh->evtchn_pending));
1601                 if (!masked)
1602                         unmask_evtchn(evtchn);
1603                 ret = 1;
1604         }
1605
1606         return ret;
1607 }
1608
1609 static void restore_pirqs(void)
1610 {
1611         int pirq, rc, irq, gsi;
1612         struct physdev_map_pirq map_irq;
1613         struct irq_info *info;
1614
1615         list_for_each_entry(info, &xen_irq_list_head, list) {
1616                 if (info->type != IRQT_PIRQ)
1617                         continue;
1618
1619                 pirq = info->u.pirq.pirq;
1620                 gsi = info->u.pirq.gsi;
1621                 irq = info->irq;
1622
1623                 /* save/restore of PT devices doesn't work, so at this point the
1624                  * only devices present are GSI based emulated devices */
1625                 if (!gsi)
1626                         continue;
1627
1628                 map_irq.domid = DOMID_SELF;
1629                 map_irq.type = MAP_PIRQ_TYPE_GSI;
1630                 map_irq.index = gsi;
1631                 map_irq.pirq = pirq;
1632
1633                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1634                 if (rc) {
1635                         printk(KERN_WARNING "xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1636                                         gsi, irq, pirq, rc);
1637                         xen_free_irq(irq);
1638                         continue;
1639                 }
1640
1641                 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1642
1643                 __startup_pirq(irq);
1644         }
1645 }
1646
1647 static void restore_cpu_virqs(unsigned int cpu)
1648 {
1649         struct evtchn_bind_virq bind_virq;
1650         int virq, irq, evtchn;
1651
1652         for (virq = 0; virq < NR_VIRQS; virq++) {
1653                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1654                         continue;
1655
1656                 BUG_ON(virq_from_irq(irq) != virq);
1657
1658                 /* Get a new binding from Xen. */
1659                 bind_virq.virq = virq;
1660                 bind_virq.vcpu = cpu;
1661                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1662                                                 &bind_virq) != 0)
1663                         BUG();
1664                 evtchn = bind_virq.port;
1665
1666                 /* Record the new mapping. */
1667                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
1668                 bind_evtchn_to_cpu(evtchn, cpu);
1669         }
1670 }
1671
1672 static void restore_cpu_ipis(unsigned int cpu)
1673 {
1674         struct evtchn_bind_ipi bind_ipi;
1675         int ipi, irq, evtchn;
1676
1677         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1678                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1679                         continue;
1680
1681                 BUG_ON(ipi_from_irq(irq) != ipi);
1682
1683                 /* Get a new binding from Xen. */
1684                 bind_ipi.vcpu = cpu;
1685                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1686                                                 &bind_ipi) != 0)
1687                         BUG();
1688                 evtchn = bind_ipi.port;
1689
1690                 /* Record the new mapping. */
1691                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
1692                 bind_evtchn_to_cpu(evtchn, cpu);
1693         }
1694 }
1695
1696 /* Clear an irq's pending state, in preparation for polling on it */
1697 void xen_clear_irq_pending(int irq)
1698 {
1699         int evtchn = evtchn_from_irq(irq);
1700
1701         if (VALID_EVTCHN(evtchn))
1702                 clear_evtchn(evtchn);
1703 }
1704 EXPORT_SYMBOL(xen_clear_irq_pending);
1705 void xen_set_irq_pending(int irq)
1706 {
1707         int evtchn = evtchn_from_irq(irq);
1708
1709         if (VALID_EVTCHN(evtchn))
1710                 set_evtchn(evtchn);
1711 }
1712
1713 bool xen_test_irq_pending(int irq)
1714 {
1715         int evtchn = evtchn_from_irq(irq);
1716         bool ret = false;
1717
1718         if (VALID_EVTCHN(evtchn))
1719                 ret = test_evtchn(evtchn);
1720
1721         return ret;
1722 }
1723
1724 /* Poll waiting for an irq to become pending with timeout.  In the usual case,
1725  * the irq will be disabled so it won't deliver an interrupt. */
1726 void xen_poll_irq_timeout(int irq, u64 timeout)
1727 {
1728         evtchn_port_t evtchn = evtchn_from_irq(irq);
1729
1730         if (VALID_EVTCHN(evtchn)) {
1731                 struct sched_poll poll;
1732
1733                 poll.nr_ports = 1;
1734                 poll.timeout = timeout;
1735                 set_xen_guest_handle(poll.ports, &evtchn);
1736
1737                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1738                         BUG();
1739         }
1740 }
1741 EXPORT_SYMBOL(xen_poll_irq_timeout);
1742 /* Poll waiting for an irq to become pending.  In the usual case, the
1743  * irq will be disabled so it won't deliver an interrupt. */
1744 void xen_poll_irq(int irq)
1745 {
1746         xen_poll_irq_timeout(irq, 0 /* no timeout */);
1747 }
1748
1749 /* Check whether the IRQ line is shared with other guests. */
1750 int xen_test_irq_shared(int irq)
1751 {
1752         struct irq_info *info = info_for_irq(irq);
1753         struct physdev_irq_status_query irq_status;
1754
1755         if (WARN_ON(!info))
1756                 return -ENOENT;
1757
1758         irq_status.irq = info->u.pirq.pirq;
1759
1760         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
1761                 return 0;
1762         return !(irq_status.flags & XENIRQSTAT_shared);
1763 }
1764 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
1765
1766 void xen_irq_resume(void)
1767 {
1768         unsigned int cpu, evtchn;
1769         struct irq_info *info;
1770
1771         init_evtchn_cpu_bindings();
1772
1773         /* New event-channel space is not 'live' yet. */
1774         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1775                 mask_evtchn(evtchn);
1776
1777         /* No IRQ <-> event-channel mappings. */
1778         list_for_each_entry(info, &xen_irq_list_head, list)
1779                 info->evtchn = 0; /* zap event-channel binding */
1780
1781         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1782                 evtchn_to_irq[evtchn] = -1;
1783
1784         for_each_possible_cpu(cpu) {
1785                 restore_cpu_virqs(cpu);
1786                 restore_cpu_ipis(cpu);
1787         }
1788
1789         restore_pirqs();
1790 }
1791
1792 static struct irq_chip xen_dynamic_chip __read_mostly = {
1793         .name                   = "xen-dyn",
1794
1795         .irq_disable            = disable_dynirq,
1796         .irq_mask               = disable_dynirq,
1797         .irq_unmask             = enable_dynirq,
1798
1799         .irq_ack                = ack_dynirq,
1800         .irq_mask_ack           = mask_ack_dynirq,
1801
1802         .irq_set_affinity       = set_affinity_irq,
1803         .irq_retrigger          = retrigger_dynirq,
1804 };
1805
1806 static struct irq_chip xen_pirq_chip __read_mostly = {
1807         .name                   = "xen-pirq",
1808
1809         .irq_startup            = startup_pirq,
1810         .irq_shutdown           = shutdown_pirq,
1811         .irq_enable             = enable_pirq,
1812         .irq_disable            = disable_pirq,
1813
1814         .irq_mask               = disable_dynirq,
1815         .irq_unmask             = enable_dynirq,
1816
1817         .irq_ack                = eoi_pirq,
1818         .irq_eoi                = eoi_pirq,
1819         .irq_mask_ack           = mask_ack_pirq,
1820
1821         .irq_set_affinity       = set_affinity_irq,
1822
1823         .irq_retrigger          = retrigger_dynirq,
1824 };
1825
1826 static struct irq_chip xen_percpu_chip __read_mostly = {
1827         .name                   = "xen-percpu",
1828
1829         .irq_disable            = disable_dynirq,
1830         .irq_mask               = disable_dynirq,
1831         .irq_unmask             = enable_dynirq,
1832
1833         .irq_ack                = ack_dynirq,
1834 };
1835
1836 int xen_set_callback_via(uint64_t via)
1837 {
1838         struct xen_hvm_param a;
1839         a.domid = DOMID_SELF;
1840         a.index = HVM_PARAM_CALLBACK_IRQ;
1841         a.value = via;
1842         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1843 }
1844 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1845
1846 #ifdef CONFIG_XEN_PVHVM
1847 /* Vector callbacks are better than PCI interrupts to receive event
1848  * channel notifications because we can receive vector callbacks on any
1849  * vcpu and we don't need PCI support or APIC interactions. */
1850 void xen_callback_vector(void)
1851 {
1852         int rc;
1853         uint64_t callback_via;
1854         if (xen_have_vector_callback) {
1855                 callback_via = HVM_CALLBACK_VECTOR(HYPERVISOR_CALLBACK_VECTOR);
1856                 rc = xen_set_callback_via(callback_via);
1857                 if (rc) {
1858                         printk(KERN_ERR "Request for Xen HVM callback vector"
1859                                         " failed.\n");
1860                         xen_have_vector_callback = 0;
1861                         return;
1862                 }
1863                 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1864                                 "enabled\n");
1865                 /* in the restore case the vector has already been allocated */
1866                 if (!test_bit(HYPERVISOR_CALLBACK_VECTOR, used_vectors))
1867                         alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR,
1868                                         xen_hvm_callback_vector);
1869         }
1870 }
1871 #else
1872 void xen_callback_vector(void) {}
1873 #endif
1874
1875 void __init xen_init_IRQ(void)
1876 {
1877         int i;
1878
1879         evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1880                                     GFP_KERNEL);
1881         BUG_ON(!evtchn_to_irq);
1882         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1883                 evtchn_to_irq[i] = -1;
1884
1885         init_evtchn_cpu_bindings();
1886
1887         /* No event channels are 'live' right now. */
1888         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1889                 mask_evtchn(i);
1890
1891         pirq_needs_eoi = pirq_needs_eoi_flag;
1892
1893 #ifdef CONFIG_X86
1894         if (xen_hvm_domain()) {
1895                 xen_callback_vector();
1896                 native_init_IRQ();
1897                 /* pci_xen_hvm_init must be called after native_init_IRQ so that
1898                  * __acpi_register_gsi can point at the right function */
1899                 pci_xen_hvm_init();
1900         } else {
1901                 int rc;
1902                 struct physdev_pirq_eoi_gmfn eoi_gmfn;
1903
1904                 irq_ctx_init(smp_processor_id());
1905                 if (xen_initial_domain())
1906                         pci_xen_initial_domain();
1907
1908                 pirq_eoi_map = (void *)__get_free_page(GFP_KERNEL|__GFP_ZERO);
1909                 eoi_gmfn.gmfn = virt_to_mfn(pirq_eoi_map);
1910                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_pirq_eoi_gmfn_v2, &eoi_gmfn);
1911                 if (rc != 0) {
1912                         free_page((unsigned long) pirq_eoi_map);
1913                         pirq_eoi_map = NULL;
1914                 } else
1915                         pirq_needs_eoi = pirq_check_eoi_map;
1916         }
1917 #endif
1918 }