drivers/xen/events.c

   1 /*
   2  * Xen event channels
   3  *
   4  * Xen models interrupts with abstract event channels.  Because each
   5  * domain gets 1024 event channels, but NR_IRQ is not that large, we
   6  * must dynamically map irqs<->event channels.  The event channels
   7  * interface with the rest of the kernel by defining a xen interrupt
   8  * chip.  When an event is received, it is mapped to an irq and sent
   9  * through the normal interrupt processing path.
  10  *
  11  * There are four kinds of events which can be mapped to an event
  12  * channel:
  13  *
  14  * 1. Inter-domain notifications.  This includes all the virtual
  15  *    device events, since they're driven by front-ends in another domain
  16  *    (typically dom0).
  17  * 2. VIRQs, typically used for timers.  These are per-cpu events.
  18  * 3. IPIs.
  19  * 4. PIRQs - Hardware interrupts.
  20  *
  21  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
  22  */
  23
  24 #include <linux/linkage.h>
  25 #include <linux/interrupt.h>
  26 #include <linux/irq.h>
  27 #include <linux/module.h>
  28 #include <linux/string.h>
  29 #include <linux/bootmem.h>
  30 #include <linux/slab.h>
  31 #include <linux/irqnr.h>
  32 #include <linux/pci.h>
  33
  34 #ifdef CONFIG_X86
  35 #include <asm/desc.h>
  36 #include <asm/ptrace.h>
  37 #include <asm/irq.h>
  38 #include <asm/idle.h>
  39 #include <asm/io_apic.h>
  40 #include <asm/xen/page.h>
  41 #include <asm/xen/pci.h>
  42 #endif
  43 #include <asm/sync_bitops.h>
  44 #include <asm/xen/hypercall.h>
  45 #include <asm/xen/hypervisor.h>
  46
  47 #include <xen/xen.h>
  48 #include <xen/hvm.h>
  49 #include <xen/xen-ops.h>
  50 #include <xen/events.h>
  51 #include <xen/interface/xen.h>
  52 #include <xen/interface/event_channel.h>
  53 #include <xen/interface/hvm/hvm_op.h>
  54 #include <xen/interface/hvm/params.h>
  55 #include <xen/interface/physdev.h>
  56 #include <xen/interface/sched.h>
  57 #include <asm/hw_irq.h>
  58
  59 /*
  60  * This lock protects updates to the following mapping and reference-count
  61  * arrays. The lock does not need to be acquired to read the mapping tables.
  62  */
  63 static DEFINE_MUTEX(irq_mapping_update_lock);
  64
  65 static LIST_HEAD(xen_irq_list_head);
  66
  67 /* IRQ <-> VIRQ mapping. */
  68 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
  69
  70 /* IRQ <-> IPI mapping */
  71 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
  72
  73 /* Interrupt types. */
  74 enum xen_irq_type {
  75         IRQT_UNBOUND = 0,
  76         IRQT_PIRQ,
  77         IRQT_VIRQ,
  78         IRQT_IPI,
  79         IRQT_EVTCHN
  80 };
  81
  82 /*
  83  * Packed IRQ information:
  84  * type - enum xen_irq_type
  85  * event channel - irq->event channel mapping
  86  * cpu - cpu this event channel is bound to
  87  * index - type-specific information:
  88  *    PIRQ - vector, with MSB being "needs EIO", or physical IRQ of the HVM
  89  *           guest, or GSI (real passthrough IRQ) of the device.
  90  *    VIRQ - virq number
  91  *    IPI - IPI vector
  92  *    EVTCHN -
  93  */
  94 struct irq_info {
  95         struct list_head list;
  96         int refcnt;
  97         enum xen_irq_type type; /* type */
  98         unsigned irq;
  99         unsigned short evtchn;  /* event channel */
 100         unsigned short cpu;     /* cpu bound */
 101
 102         union {
 103                 unsigned short virq;
 104                 enum ipi_vector ipi;
 105                 struct {
 106                         unsigned short pirq;
 107                         unsigned short gsi;
 108                         unsigned char vector;
 109                         unsigned char flags;
 110                         uint16_t domid;
 111                 } pirq;
 112         } u;
 113 };
 114 #define PIRQ_NEEDS_EOI  (1 << 0)
 115 #define PIRQ_SHAREABLE  (1 << 1)
 116
 117 static int *evtchn_to_irq;
 118 #ifdef CONFIG_X86
 119 static unsigned long *pirq_eoi_map;
 120 #endif
 121 static bool (*pirq_needs_eoi)(unsigned irq);
 122
 123 /*
 124  * Note sizeof(xen_ulong_t) can be more than sizeof(unsigned long). Be
 125  * careful to only use bitops which allow for this (e.g
 126  * test_bit/find_first_bit and friends but not __ffs) and to pass
 127  * BITS_PER_EVTCHN_WORD as the bitmask length.
 128  */
 129 #define BITS_PER_EVTCHN_WORD (sizeof(xen_ulong_t)*8)
 130 /*
 131  * Make a bitmask (i.e. unsigned long *) of a xen_ulong_t
 132  * array. Primarily to avoid long lines (hence the terse name).
 133  */
 134 #define BM(x) (unsigned long *)(x)
 135 /* Find the first set bit in a evtchn mask */
 136 #define EVTCHN_FIRST_BIT(w) find_first_bit(BM(&(w)), BITS_PER_EVTCHN_WORD)
 137
 138 static DEFINE_PER_CPU(xen_ulong_t [NR_EVENT_CHANNELS/BITS_PER_EVTCHN_WORD],
 139                       cpu_evtchn_mask);
 140
 141 /* Xen will never allocate port zero for any purpose. */
 142 #define VALID_EVTCHN(chn)       ((chn) != 0)
 143
 144 static struct irq_chip xen_dynamic_chip;
 145 static struct irq_chip xen_percpu_chip;
 146 static struct irq_chip xen_pirq_chip;
 147 static void enable_dynirq(struct irq_data *data);
 148 static void disable_dynirq(struct irq_data *data);
 149
 150 /* Get info for IRQ */
 151 static struct irq_info *info_for_irq(unsigned irq)
 152 {
 153         return irq_get_handler_data(irq);
 154 }
 155
 156 /* Constructors for packed IRQ information. */
 157 static void xen_irq_info_common_init(struct irq_info *info,
 158                                      unsigned irq,
 159                                      enum xen_irq_type type,
 160                                      unsigned short evtchn,
 161                                      unsigned short cpu)
 162 {
 163
 164         BUG_ON(info->type != IRQT_UNBOUND && info->type != type);
 165
 166         info->type = type;
 167         info->irq = irq;
 168         info->evtchn = evtchn;
 169         info->cpu = cpu;
 170
 171         evtchn_to_irq[evtchn] = irq;
 172 }
 173
 174 static void xen_irq_info_evtchn_init(unsigned irq,
 175                                      unsigned short evtchn)
 176 {
 177         struct irq_info *info = info_for_irq(irq);
 178
 179         xen_irq_info_common_init(info, irq, IRQT_EVTCHN, evtchn, 0);
 180 }
 181
 182 static void xen_irq_info_ipi_init(unsigned cpu,
 183                                   unsigned irq,
 184                                   unsigned short evtchn,
 185                                   enum ipi_vector ipi)
 186 {
 187         struct irq_info *info = info_for_irq(irq);
 188
 189         xen_irq_info_common_init(info, irq, IRQT_IPI, evtchn, 0);
 190
 191         info->u.ipi = ipi;
 192
 193         per_cpu(ipi_to_irq, cpu)[ipi] = irq;
 194 }
 195
 196 static void xen_irq_info_virq_init(unsigned cpu,
 197                                    unsigned irq,
 198                                    unsigned short evtchn,
 199                                    unsigned short virq)
 200 {
 201         struct irq_info *info = info_for_irq(irq);
 202
 203         xen_irq_info_common_init(info, irq, IRQT_VIRQ, evtchn, 0);
 204
 205         info->u.virq = virq;
 206
 207         per_cpu(virq_to_irq, cpu)[virq] = irq;
 208 }
 209
 210 static void xen_irq_info_pirq_init(unsigned irq,
 211                                    unsigned short evtchn,
 212                                    unsigned short pirq,
 213                                    unsigned short gsi,
 214                                    unsigned short vector,
 215                                    uint16_t domid,
 216                                    unsigned char flags)
 217 {
 218         struct irq_info *info = info_for_irq(irq);
 219
 220         xen_irq_info_common_init(info, irq, IRQT_PIRQ, evtchn, 0);
 221
 222         info->u.pirq.pirq = pirq;
 223         info->u.pirq.gsi = gsi;
 224         info->u.pirq.vector = vector;
 225         info->u.pirq.domid = domid;
 226         info->u.pirq.flags = flags;
 227 }
 228
 229 /*
 230  * Accessors for packed IRQ information.
 231  */
 232 static unsigned int evtchn_from_irq(unsigned irq)
 233 {
 234         if (unlikely(WARN(irq < 0 || irq >= nr_irqs, "Invalid irq %d!\n", irq)))
 235                 return 0;
 236
 237         return info_for_irq(irq)->evtchn;
 238 }
 239
 240 unsigned irq_from_evtchn(unsigned int evtchn)
 241 {
 242         return evtchn_to_irq[evtchn];
 243 }
 244 EXPORT_SYMBOL_GPL(irq_from_evtchn);
 245
 246 static enum ipi_vector ipi_from_irq(unsigned irq)
 247 {
 248         struct irq_info *info = info_for_irq(irq);
 249
 250         BUG_ON(info == NULL);
 251         BUG_ON(info->type != IRQT_IPI);
 252
 253         return info->u.ipi;
 254 }
 255
 256 static unsigned virq_from_irq(unsigned irq)
 257 {
 258         struct irq_info *info = info_for_irq(irq);
 259
 260         BUG_ON(info == NULL);
 261         BUG_ON(info->type != IRQT_VIRQ);
 262
 263         return info->u.virq;
 264 }
 265
 266 static unsigned pirq_from_irq(unsigned irq)
 267 {
 268         struct irq_info *info = info_for_irq(irq);
 269
 270         BUG_ON(info == NULL);
 271         BUG_ON(info->type != IRQT_PIRQ);
 272
 273         return info->u.pirq.pirq;
 274 }
 275
 276 static enum xen_irq_type type_from_irq(unsigned irq)
 277 {
 278         return info_for_irq(irq)->type;
 279 }
 280
 281 static unsigned cpu_from_irq(unsigned irq)
 282 {
 283         return info_for_irq(irq)->cpu;
 284 }
 285
 286 static unsigned int cpu_from_evtchn(unsigned int evtchn)
 287 {
 288         int irq = evtchn_to_irq[evtchn];
 289         unsigned ret = 0;
 290
 291         if (irq != -1)
 292                 ret = cpu_from_irq(irq);
 293
 294         return ret;
 295 }
 296
 297 #ifdef CONFIG_X86
 298 static bool pirq_check_eoi_map(unsigned irq)
 299 {
 300         return test_bit(pirq_from_irq(irq), pirq_eoi_map);
 301 }
 302 #endif
 303
 304 static bool pirq_needs_eoi_flag(unsigned irq)
 305 {
 306         struct irq_info *info = info_for_irq(irq);
 307         BUG_ON(info->type != IRQT_PIRQ);
 308
 309         return info->u.pirq.flags & PIRQ_NEEDS_EOI;
 310 }
 311
 312 static inline xen_ulong_t active_evtchns(unsigned int cpu,
 313                                          struct shared_info *sh,
 314                                          unsigned int idx)
 315 {
 316         return sh->evtchn_pending[idx] &
 317                 per_cpu(cpu_evtchn_mask, cpu)[idx] &
 318                 ~sh->evtchn_mask[idx];
 319 }
 320
 321 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
 322 {
 323         int irq = evtchn_to_irq[chn];
 324
 325         BUG_ON(irq == -1);
 326 #ifdef CONFIG_SMP
 327         cpumask_copy(irq_to_desc(irq)->irq_data.affinity, cpumask_of(cpu));
 328 #endif
 329
 330         clear_bit(chn, BM(per_cpu(cpu_evtchn_mask, cpu_from_irq(irq))));
 331         set_bit(chn, BM(per_cpu(cpu_evtchn_mask, cpu)));
 332
 333         info_for_irq(irq)->cpu = cpu;
 334 }
 335
 336 static void init_evtchn_cpu_bindings(void)
 337 {
 338         int i;
 339 #ifdef CONFIG_SMP
 340         struct irq_info *info;
 341
 342         /* By default all event channels notify CPU#0. */
 343         list_for_each_entry(info, &xen_irq_list_head, list) {
 344                 struct irq_desc *desc = irq_to_desc(info->irq);
 345                 cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 346         }
 347 #endif
 348
 349         for_each_possible_cpu(i)
 350                 memset(per_cpu(cpu_evtchn_mask, i),
 351                        (i == 0) ? ~0 : 0, sizeof(*per_cpu(cpu_evtchn_mask, i)));
 352 }
 353
 354 static inline void clear_evtchn(int port)
 355 {
 356         struct shared_info *s = HYPERVISOR_shared_info;
 357         sync_clear_bit(port, BM(&s->evtchn_pending[0]));
 358 }
 359
 360 static inline void set_evtchn(int port)
 361 {
 362         struct shared_info *s = HYPERVISOR_shared_info;
 363         sync_set_bit(port, BM(&s->evtchn_pending[0]));
 364 }
 365
 366 static inline int test_evtchn(int port)
 367 {
 368         struct shared_info *s = HYPERVISOR_shared_info;
 369         return sync_test_bit(port, BM(&s->evtchn_pending[0]));
 370 }
 371
 372
 373 /**
 374  * notify_remote_via_irq - send event to remote end of event channel via irq
 375  * @irq: irq of event channel to send event to
 376  *
 377  * Unlike notify_remote_via_evtchn(), this is safe to use across
 378  * save/restore. Notifications on a broken connection are silently
 379  * dropped.
 380  */
 381 void notify_remote_via_irq(int irq)
 382 {
 383         int evtchn = evtchn_from_irq(irq);
 384
 385         if (VALID_EVTCHN(evtchn))
 386                 notify_remote_via_evtchn(evtchn);
 387 }
 388 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
 389
 390 static void mask_evtchn(int port)
 391 {
 392         struct shared_info *s = HYPERVISOR_shared_info;
 393         sync_set_bit(port, BM(&s->evtchn_mask[0]));
 394 }
 395
 396 static void unmask_evtchn(int port)
 397 {
 398         struct shared_info *s = HYPERVISOR_shared_info;
 399         unsigned int cpu = get_cpu();
 400         int do_hypercall = 0, evtchn_pending = 0;
 401
 402         BUG_ON(!irqs_disabled());
 403
 404         if (unlikely((cpu != cpu_from_evtchn(port))))
 405                 do_hypercall = 1;
 406         else {
 407                 /*
 408                  * Need to clear the mask before checking pending to
 409                  * avoid a race with an event becoming pending.
 410                  *
 411                  * EVTCHNOP_unmask will only trigger an upcall if the
 412                  * mask bit was set, so if a hypercall is needed
 413                  * remask the event.
 414                  */
 415                 sync_clear_bit(port, BM(&s->evtchn_mask[0]));
 416                 evtchn_pending = sync_test_bit(port, BM(&s->evtchn_pending[0]));
 417
 418                 if (unlikely(evtchn_pending && xen_hvm_domain())) {
 419                         sync_set_bit(port, BM(&s->evtchn_mask[0]));
 420                         do_hypercall = 1;
 421                 }
 422         }
 423
 424         /* Slow path (hypercall) if this is a non-local port or if this is
 425          * an hvm domain and an event is pending (hvm domains don't have
 426          * their own implementation of irq_enable). */
 427         if (do_hypercall) {
 428                 struct evtchn_unmask unmask = { .port = port };
 429                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 430         } else {
 431                 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
 432
 433                 /*
 434                  * The following is basically the equivalent of
 435                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
 436                  * the interrupt edge' if the channel is masked.
 437                  */
 438                 if (evtchn_pending &&
 439                     !sync_test_and_set_bit(port / BITS_PER_EVTCHN_WORD,
 440                                            BM(&vcpu_info->evtchn_pending_sel)))
 441                         vcpu_info->evtchn_upcall_pending = 1;
 442         }
 443
 444         put_cpu();
 445 }
 446
 447 static void xen_irq_init(unsigned irq)
 448 {
 449         struct irq_info *info;
 450 #ifdef CONFIG_SMP
 451         struct irq_desc *desc = irq_to_desc(irq);
 452
 453         /* By default all event channels notify CPU#0. */
 454         cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 455 #endif
 456
 457         info = kzalloc(sizeof(*info), GFP_KERNEL);
 458         if (info == NULL)
 459                 panic("Unable to allocate metadata for IRQ%d\n", irq);
 460
 461         info->type = IRQT_UNBOUND;
 462         info->refcnt = -1;
 463
 464         irq_set_handler_data(irq, info);
 465
 466         list_add_tail(&info->list, &xen_irq_list_head);
 467 }
 468
 469 static int __must_check xen_allocate_irq_dynamic(void)
 470 {
 471         int first = 0;
 472         int irq;
 473
 474 #ifdef CONFIG_X86_IO_APIC
 475         /*
 476          * For an HVM guest or domain 0 which see "real" (emulated or
 477          * actual respectively) GSIs we allocate dynamic IRQs
 478          * e.g. those corresponding to event channels or MSIs
 479          * etc. from the range above those "real" GSIs to avoid
 480          * collisions.
 481          */
 482         if (xen_initial_domain() || xen_hvm_domain())
 483                 first = get_nr_irqs_gsi();
 484 #endif
 485
 486         irq = irq_alloc_desc_from(first, -1);
 487
 488         if (irq >= 0)
 489                 xen_irq_init(irq);
 490
 491         return irq;
 492 }
 493
 494 static int __must_check xen_allocate_irq_gsi(unsigned gsi)
 495 {
 496         int irq;
 497
 498         /*
 499          * A PV guest has no concept of a GSI (since it has no ACPI
 500          * nor access to/knowledge of the physical APICs). Therefore
 501          * all IRQs are dynamically allocated from the entire IRQ
 502          * space.
 503          */
 504         if (xen_pv_domain() && !xen_initial_domain())
 505                 return xen_allocate_irq_dynamic();
 506
 507         /* Legacy IRQ descriptors are already allocated by the arch. */
 508         if (gsi < NR_IRQS_LEGACY)
 509                 irq = gsi;
 510         else
 511                 irq = irq_alloc_desc_at(gsi, -1);
 512
 513         xen_irq_init(irq);
 514
 515         return irq;
 516 }
 517
 518 static void xen_free_irq(unsigned irq)
 519 {
 520         struct irq_info *info = irq_get_handler_data(irq);
 521
 522         list_del(&info->list);
 523
 524         irq_set_handler_data(irq, NULL);
 525
 526         WARN_ON(info->refcnt > 0);
 527
 528         kfree(info);
 529
 530         /* Legacy IRQ descriptors are managed by the arch. */
 531         if (irq < NR_IRQS_LEGACY)
 532                 return;
 533
 534         irq_free_desc(irq);
 535 }
 536
 537 static void pirq_query_unmask(int irq)
 538 {
 539         struct physdev_irq_status_query irq_status;
 540         struct irq_info *info = info_for_irq(irq);
 541
 542         BUG_ON(info->type != IRQT_PIRQ);
 543
 544         irq_status.irq = pirq_from_irq(irq);
 545         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
 546                 irq_status.flags = 0;
 547
 548         info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
 549         if (irq_status.flags & XENIRQSTAT_needs_eoi)
 550                 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
 551 }
 552
 553 static bool probing_irq(int irq)
 554 {
 555         struct irq_desc *desc = irq_to_desc(irq);
 556
 557         return desc && desc->action == NULL;
 558 }
 559
 560 static void eoi_pirq(struct irq_data *data)
 561 {
 562         int evtchn = evtchn_from_irq(data->irq);
 563         struct physdev_eoi eoi = { .irq = pirq_from_irq(data->irq) };
 564         int rc = 0;
 565
 566         irq_move_irq(data);
 567
 568         if (VALID_EVTCHN(evtchn))
 569                 clear_evtchn(evtchn);
 570
 571         if (pirq_needs_eoi(data->irq)) {
 572                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
 573                 WARN_ON(rc);
 574         }
 575 }
 576
 577 static void mask_ack_pirq(struct irq_data *data)
 578 {
 579         disable_dynirq(data);
 580         eoi_pirq(data);
 581 }
 582
 583 static unsigned int __startup_pirq(unsigned int irq)
 584 {
 585         struct evtchn_bind_pirq bind_pirq;
 586         struct irq_info *info = info_for_irq(irq);
 587         int evtchn = evtchn_from_irq(irq);
 588         int rc;
 589
 590         BUG_ON(info->type != IRQT_PIRQ);
 591
 592         if (VALID_EVTCHN(evtchn))
 593                 goto out;
 594
 595         bind_pirq.pirq = pirq_from_irq(irq);
 596         /* NB. We are happy to share unless we are probing. */
 597         bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
 598                                         BIND_PIRQ__WILL_SHARE : 0;
 599         rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
 600         if (rc != 0) {
 601                 if (!probing_irq(irq))
 602                         printk(KERN_INFO "Failed to obtain physical IRQ %d\n",
 603                                irq);
 604                 return 0;
 605         }
 606         evtchn = bind_pirq.port;
 607
 608         pirq_query_unmask(irq);
 609
 610         evtchn_to_irq[evtchn] = irq;
 611         bind_evtchn_to_cpu(evtchn, 0);
 612         info->evtchn = evtchn;
 613
 614 out:
 615         unmask_evtchn(evtchn);
 616         eoi_pirq(irq_get_irq_data(irq));
 617
 618         return 0;
 619 }
 620
 621 static unsigned int startup_pirq(struct irq_data *data)
 622 {
 623         return __startup_pirq(data->irq);
 624 }
 625
 626 static void shutdown_pirq(struct irq_data *data)
 627 {
 628         struct evtchn_close close;
 629         unsigned int irq = data->irq;
 630         struct irq_info *info = info_for_irq(irq);
 631         int evtchn = evtchn_from_irq(irq);
 632
 633         BUG_ON(info->type != IRQT_PIRQ);
 634
 635         if (!VALID_EVTCHN(evtchn))
 636                 return;
 637
 638         mask_evtchn(evtchn);
 639
 640         close.port = evtchn;
 641         if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 642                 BUG();
 643
 644         bind_evtchn_to_cpu(evtchn, 0);
 645         evtchn_to_irq[evtchn] = -1;
 646         info->evtchn = 0;
 647 }
 648
 649 static void enable_pirq(struct irq_data *data)
 650 {
 651         startup_pirq(data);
 652 }
 653
 654 static void disable_pirq(struct irq_data *data)
 655 {
 656         disable_dynirq(data);
 657 }
 658
 659 int xen_irq_from_gsi(unsigned gsi)
 660 {
 661         struct irq_info *info;
 662
 663         list_for_each_entry(info, &xen_irq_list_head, list) {
 664                 if (info->type != IRQT_PIRQ)
 665                         continue;
 666
 667                 if (info->u.pirq.gsi == gsi)
 668                         return info->irq;
 669         }
 670
 671         return -1;
 672 }
 673 EXPORT_SYMBOL_GPL(xen_irq_from_gsi);
 674
 675 /*
 676  * Do not make any assumptions regarding the relationship between the
 677  * IRQ number returned here and the Xen pirq argument.
 678  *
 679  * Note: We don't assign an event channel until the irq actually started
 680  * up.  Return an existing irq if we've already got one for the gsi.
 681  *
 682  * Shareable implies level triggered, not shareable implies edge
 683  * triggered here.
 684  */
 685 int xen_bind_pirq_gsi_to_irq(unsigned gsi,
 686                              unsigned pirq, int shareable, char *name)
 687 {
 688         int irq = -1;
 689         struct physdev_irq irq_op;
 690
 691         mutex_lock(&irq_mapping_update_lock);
 692
 693         irq = xen_irq_from_gsi(gsi);
 694         if (irq != -1) {
 695                 printk(KERN_INFO "xen_map_pirq_gsi: returning irq %d for gsi %u\n",
 696                        irq, gsi);
 697                 goto out;
 698         }
 699
 700         irq = xen_allocate_irq_gsi(gsi);
 701         if (irq < 0)
 702                 goto out;
 703
 704         irq_op.irq = irq;
 705         irq_op.vector = 0;
 706
 707         /* Only the privileged domain can do this. For non-priv, the pcifront
 708          * driver provides a PCI bus that does the call to do exactly
 709          * this in the priv domain. */
 710         if (xen_initial_domain() &&
 711             HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
 712                 xen_free_irq(irq);
 713                 irq = -ENOSPC;
 714                 goto out;
 715         }
 716
 717         xen_irq_info_pirq_init(irq, 0, pirq, gsi, irq_op.vector, DOMID_SELF,
 718                                shareable ? PIRQ_SHAREABLE : 0);
 719
 720         pirq_query_unmask(irq);
 721         /* We try to use the handler with the appropriate semantic for the
 722          * type of interrupt: if the interrupt is an edge triggered
 723          * interrupt we use handle_edge_irq.
 724          *
 725          * On the other hand if the interrupt is level triggered we use
 726          * handle_fasteoi_irq like the native code does for this kind of
 727          * interrupts.
 728          *
 729          * Depending on the Xen version, pirq_needs_eoi might return true
 730          * not only for level triggered interrupts but for edge triggered
 731          * interrupts too. In any case Xen always honors the eoi mechanism,
 732          * not injecting any more pirqs of the same kind if the first one
 733          * hasn't received an eoi yet. Therefore using the fasteoi handler
 734          * is the right choice either way.
 735          */
 736         if (shareable)
 737                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 738                                 handle_fasteoi_irq, name);
 739         else
 740                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 741                                 handle_edge_irq, name);
 742
 743 out:
 744         mutex_unlock(&irq_mapping_update_lock);
 745
 746         return irq;
 747 }
 748
 749 #ifdef CONFIG_PCI_MSI
 750 int xen_allocate_pirq_msi(struct pci_dev *dev, struct msi_desc *msidesc)
 751 {
 752         int rc;
 753         struct physdev_get_free_pirq op_get_free_pirq;
 754
 755         op_get_free_pirq.type = MAP_PIRQ_TYPE_MSI;
 756         rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
 757
 758         WARN_ONCE(rc == -ENOSYS,
 759                   "hypervisor does not support the PHYSDEVOP_get_free_pirq interface\n");
 760
 761         return rc ? -1 : op_get_free_pirq.pirq;
 762 }
 763
 764 int xen_bind_pirq_msi_to_irq(struct pci_dev *dev, struct msi_desc *msidesc,
 765                              int pirq, int vector, const char *name,
 766                              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, vector, 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         irq_clear_status_flags(irq, IRQ_NOREQUEST|IRQ_NOAUTOEN);
 880
 881 out:
 882         mutex_unlock(&irq_mapping_update_lock);
 883
 884         return irq;
 885 }
 886 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
 887
 888 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
 889 {
 890         struct evtchn_bind_ipi bind_ipi;
 891         int evtchn, irq;
 892
 893         mutex_lock(&irq_mapping_update_lock);
 894
 895         irq = per_cpu(ipi_to_irq, cpu)[ipi];
 896
 897         if (irq == -1) {
 898                 irq = xen_allocate_irq_dynamic();
 899                 if (irq < 0)
 900                         goto out;
 901
 902                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 903                                               handle_percpu_irq, "ipi");
 904
 905                 bind_ipi.vcpu = cpu;
 906                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
 907                                                 &bind_ipi) != 0)
 908                         BUG();
 909                 evtchn = bind_ipi.port;
 910
 911                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
 912
 913                 bind_evtchn_to_cpu(evtchn, cpu);
 914         } else {
 915                 struct irq_info *info = info_for_irq(irq);
 916                 WARN_ON(info == NULL || info->type != IRQT_IPI);
 917         }
 918
 919  out:
 920         mutex_unlock(&irq_mapping_update_lock);
 921         return irq;
 922 }
 923
 924 static int bind_interdomain_evtchn_to_irq(unsigned int remote_domain,
 925                                           unsigned int remote_port)
 926 {
 927         struct evtchn_bind_interdomain bind_interdomain;
 928         int err;
 929
 930         bind_interdomain.remote_dom  = remote_domain;
 931         bind_interdomain.remote_port = remote_port;
 932
 933         err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
 934                                           &bind_interdomain);
 935
 936         return err ? : bind_evtchn_to_irq(bind_interdomain.local_port);
 937 }
 938
 939 static int find_virq(unsigned int virq, unsigned int cpu)
 940 {
 941         struct evtchn_status status;
 942         int port, rc = -ENOENT;
 943
 944         memset(&status, 0, sizeof(status));
 945         for (port = 0; port <= NR_EVENT_CHANNELS; port++) {
 946                 status.dom = DOMID_SELF;
 947                 status.port = port;
 948                 rc = HYPERVISOR_event_channel_op(EVTCHNOP_status, &status);
 949                 if (rc < 0)
 950                         continue;
 951                 if (status.status != EVTCHNSTAT_virq)
 952                         continue;
 953                 if (status.u.virq == virq && status.vcpu == cpu) {
 954                         rc = port;
 955                         break;
 956                 }
 957         }
 958         return rc;
 959 }
 960
 961 int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
 962 {
 963         struct evtchn_bind_virq bind_virq;
 964         int evtchn, irq, ret;
 965
 966         mutex_lock(&irq_mapping_update_lock);
 967
 968         irq = per_cpu(virq_to_irq, cpu)[virq];
 969
 970         if (irq == -1) {
 971                 irq = xen_allocate_irq_dynamic();
 972                 if (irq < 0)
 973                         goto out;
 974
 975                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 976                                               handle_percpu_irq, "virq");
 977
 978                 bind_virq.virq = virq;
 979                 bind_virq.vcpu = cpu;
 980                 ret = HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 981                                                 &bind_virq);
 982                 if (ret == 0)
 983                         evtchn = bind_virq.port;
 984                 else {
 985                         if (ret == -EEXIST)
 986                                 ret = find_virq(virq, cpu);
 987                         BUG_ON(ret < 0);
 988                         evtchn = ret;
 989                 }
 990
 991                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
 992
 993                 bind_evtchn_to_cpu(evtchn, cpu);
 994         } else {
 995                 struct irq_info *info = info_for_irq(irq);
 996                 WARN_ON(info == NULL || info->type != IRQT_VIRQ);
 997         }
 998
 999 out:
1000         mutex_unlock(&irq_mapping_update_lock);
1001
1002         return irq;
1003 }
1004
1005 static void unbind_from_irq(unsigned int irq)
1006 {
1007         struct evtchn_close close;
1008         int evtchn = evtchn_from_irq(irq);
1009         struct irq_info *info = irq_get_handler_data(irq);
1010
1011         mutex_lock(&irq_mapping_update_lock);
1012
1013         if (info->refcnt > 0) {
1014                 info->refcnt--;
1015                 if (info->refcnt != 0)
1016                         goto done;
1017         }
1018
1019         if (VALID_EVTCHN(evtchn)) {
1020                 close.port = evtchn;
1021                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
1022                         BUG();
1023
1024                 switch (type_from_irq(irq)) {
1025                 case IRQT_VIRQ:
1026                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
1027                                 [virq_from_irq(irq)] = -1;
1028                         break;
1029                 case IRQT_IPI:
1030                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
1031                                 [ipi_from_irq(irq)] = -1;
1032                         break;
1033                 default:
1034                         break;
1035                 }
1036
1037                 /* Closed ports are implicitly re-bound to VCPU0. */
1038                 bind_evtchn_to_cpu(evtchn, 0);
1039
1040                 evtchn_to_irq[evtchn] = -1;
1041         }
1042
1043         BUG_ON(info_for_irq(irq)->type == IRQT_UNBOUND);
1044
1045         xen_free_irq(irq);
1046
1047  done:
1048         mutex_unlock(&irq_mapping_update_lock);
1049 }
1050
1051 int bind_evtchn_to_irqhandler(unsigned int evtchn,
1052                               irq_handler_t handler,
1053                               unsigned long irqflags,
1054                               const char *devname, void *dev_id)
1055 {
1056         int irq, retval;
1057
1058         irq = bind_evtchn_to_irq(evtchn);
1059         if (irq < 0)
1060                 return irq;
1061         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1062         if (retval != 0) {
1063                 unbind_from_irq(irq);
1064                 return retval;
1065         }
1066
1067         return irq;
1068 }
1069 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
1070
1071 int bind_interdomain_evtchn_to_irqhandler(unsigned int remote_domain,
1072                                           unsigned int remote_port,
1073                                           irq_handler_t handler,
1074                                           unsigned long irqflags,
1075                                           const char *devname,
1076                                           void *dev_id)
1077 {
1078         int irq, retval;
1079
1080         irq = bind_interdomain_evtchn_to_irq(remote_domain, remote_port);
1081         if (irq < 0)
1082                 return irq;
1083
1084         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1085         if (retval != 0) {
1086                 unbind_from_irq(irq);
1087                 return retval;
1088         }
1089
1090         return irq;
1091 }
1092 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler);
1093
1094 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
1095                             irq_handler_t handler,
1096                             unsigned long irqflags, const char *devname, void *dev_id)
1097 {
1098         int irq, retval;
1099
1100         irq = bind_virq_to_irq(virq, cpu);
1101         if (irq < 0)
1102                 return irq;
1103         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1104         if (retval != 0) {
1105                 unbind_from_irq(irq);
1106                 return retval;
1107         }
1108
1109         return irq;
1110 }
1111 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1112
1113 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1114                            unsigned int cpu,
1115                            irq_handler_t handler,
1116                            unsigned long irqflags,
1117                            const char *devname,
1118                            void *dev_id)
1119 {
1120         int irq, retval;
1121
1122         irq = bind_ipi_to_irq(ipi, cpu);
1123         if (irq < 0)
1124                 return irq;
1125
1126         irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME | IRQF_EARLY_RESUME;
1127         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1128         if (retval != 0) {
1129                 unbind_from_irq(irq);
1130                 return retval;
1131         }
1132
1133         return irq;
1134 }
1135
1136 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1137 {
1138         free_irq(irq, dev_id);
1139         unbind_from_irq(irq);
1140 }
1141 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1142
1143 int evtchn_make_refcounted(unsigned int evtchn)
1144 {
1145         int irq = evtchn_to_irq[evtchn];
1146         struct irq_info *info;
1147
1148         if (irq == -1)
1149                 return -ENOENT;
1150
1151         info = irq_get_handler_data(irq);
1152
1153         if (!info)
1154                 return -ENOENT;
1155
1156         WARN_ON(info->refcnt != -1);
1157
1158         info->refcnt = 1;
1159
1160         return 0;
1161 }
1162 EXPORT_SYMBOL_GPL(evtchn_make_refcounted);
1163
1164 int evtchn_get(unsigned int evtchn)
1165 {
1166         int irq;
1167         struct irq_info *info;
1168         int err = -ENOENT;
1169
1170         if (evtchn >= NR_EVENT_CHANNELS)
1171                 return -EINVAL;
1172
1173         mutex_lock(&irq_mapping_update_lock);
1174
1175         irq = evtchn_to_irq[evtchn];
1176         if (irq == -1)
1177                 goto done;
1178
1179         info = irq_get_handler_data(irq);
1180
1181         if (!info)
1182                 goto done;
1183
1184         err = -EINVAL;
1185         if (info->refcnt <= 0)
1186                 goto done;
1187
1188         info->refcnt++;
1189         err = 0;
1190  done:
1191         mutex_unlock(&irq_mapping_update_lock);
1192
1193         return err;
1194 }
1195 EXPORT_SYMBOL_GPL(evtchn_get);
1196
1197 void evtchn_put(unsigned int evtchn)
1198 {
1199         int irq = evtchn_to_irq[evtchn];
1200         if (WARN_ON(irq == -1))
1201                 return;
1202         unbind_from_irq(irq);
1203 }
1204 EXPORT_SYMBOL_GPL(evtchn_put);
1205
1206 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1207 {
1208         int irq = per_cpu(ipi_to_irq, cpu)[vector];
1209         BUG_ON(irq < 0);
1210         notify_remote_via_irq(irq);
1211 }
1212
1213 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
1214 {
1215         struct shared_info *sh = HYPERVISOR_shared_info;
1216         int cpu = smp_processor_id();
1217         xen_ulong_t *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu);
1218         int i;
1219         unsigned long flags;
1220         static DEFINE_SPINLOCK(debug_lock);
1221         struct vcpu_info *v;
1222
1223         spin_lock_irqsave(&debug_lock, flags);
1224
1225         printk("\nvcpu %d\n  ", cpu);
1226
1227         for_each_online_cpu(i) {
1228                 int pending;
1229                 v = per_cpu(xen_vcpu, i);
1230                 pending = (get_irq_regs() && i == cpu)
1231                         ? xen_irqs_disabled(get_irq_regs())
1232                         : v->evtchn_upcall_mask;
1233                 printk("%d: masked=%d pending=%d event_sel %0*"PRI_xen_ulong"\n  ", i,
1234                        pending, v->evtchn_upcall_pending,
1235                        (int)(sizeof(v->evtchn_pending_sel)*2),
1236                        v->evtchn_pending_sel);
1237         }
1238         v = per_cpu(xen_vcpu, cpu);
1239
1240         printk("\npending:\n   ");
1241         for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
1242                 printk("%0*"PRI_xen_ulong"%s",
1243                        (int)sizeof(sh->evtchn_pending[0])*2,
1244                        sh->evtchn_pending[i],
1245                        i % 8 == 0 ? "\n   " : " ");
1246         printk("\nglobal mask:\n   ");
1247         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1248                 printk("%0*"PRI_xen_ulong"%s",
1249                        (int)(sizeof(sh->evtchn_mask[0])*2),
1250                        sh->evtchn_mask[i],
1251                        i % 8 == 0 ? "\n   " : " ");
1252
1253         printk("\nglobally unmasked:\n   ");
1254         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1255                 printk("%0*"PRI_xen_ulong"%s",
1256                        (int)(sizeof(sh->evtchn_mask[0])*2),
1257                        sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
1258                        i % 8 == 0 ? "\n   " : " ");
1259
1260         printk("\nlocal cpu%d mask:\n   ", cpu);
1261         for (i = (NR_EVENT_CHANNELS/BITS_PER_EVTCHN_WORD)-1; i >= 0; i--)
1262                 printk("%0*"PRI_xen_ulong"%s", (int)(sizeof(cpu_evtchn[0])*2),
1263                        cpu_evtchn[i],
1264                        i % 8 == 0 ? "\n   " : " ");
1265
1266         printk("\nlocally unmasked:\n   ");
1267         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
1268                 xen_ulong_t pending = sh->evtchn_pending[i]
1269                         & ~sh->evtchn_mask[i]
1270                         & cpu_evtchn[i];
1271                 printk("%0*"PRI_xen_ulong"%s",
1272                        (int)(sizeof(sh->evtchn_mask[0])*2),
1273                        pending, i % 8 == 0 ? "\n   " : " ");
1274         }
1275
1276         printk("\npending list:\n");
1277         for (i = 0; i < NR_EVENT_CHANNELS; i++) {
1278                 if (sync_test_bit(i, BM(sh->evtchn_pending))) {
1279                         int word_idx = i / BITS_PER_EVTCHN_WORD;
1280                         printk("  %d: event %d -> irq %d%s%s%s\n",
1281                                cpu_from_evtchn(i), i,
1282                                evtchn_to_irq[i],
1283                                sync_test_bit(word_idx, BM(&v->evtchn_pending_sel))
1284                                              ? "" : " l2-clear",
1285                                !sync_test_bit(i, BM(sh->evtchn_mask))
1286                                              ? "" : " globally-masked",
1287                                sync_test_bit(i, BM(cpu_evtchn))
1288                                              ? "" : " locally-masked");
1289                 }
1290         }
1291
1292         spin_unlock_irqrestore(&debug_lock, flags);
1293
1294         return IRQ_HANDLED;
1295 }
1296
1297 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
1298 static DEFINE_PER_CPU(unsigned int, current_word_idx);
1299 static DEFINE_PER_CPU(unsigned int, current_bit_idx);
1300
1301 /*
1302  * Mask out the i least significant bits of w
1303  */
1304 #define MASK_LSBS(w, i) (w & ((~((xen_ulong_t)0UL)) << i))
1305
1306 /*
1307  * Search the CPUs pending events bitmasks.  For each one found, map
1308  * the event number to an irq, and feed it into do_IRQ() for
1309  * handling.
1310  *
1311  * Xen uses a two-level bitmap to speed searching.  The first level is
1312  * a bitset of words which contain pending event bits.  The second
1313  * level is a bitset of pending events themselves.
1314  */
1315 static void __xen_evtchn_do_upcall(void)
1316 {
1317         int start_word_idx, start_bit_idx;
1318         int word_idx, bit_idx;
1319         int i;
1320         int cpu = get_cpu();
1321         struct shared_info *s = HYPERVISOR_shared_info;
1322         struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1323         unsigned count;
1324
1325         do {
1326                 xen_ulong_t pending_words;
1327
1328                 vcpu_info->evtchn_upcall_pending = 0;
1329
1330                 if (__this_cpu_inc_return(xed_nesting_count) - 1)
1331                         goto out;
1332
1333                 /*
1334                  * Master flag must be cleared /before/ clearing
1335                  * selector flag. xchg_xen_ulong must contain an
1336                  * appropriate barrier.
1337                  */
1338                 pending_words = xchg_xen_ulong(&vcpu_info->evtchn_pending_sel, 0);
1339
1340                 start_word_idx = __this_cpu_read(current_word_idx);
1341                 start_bit_idx = __this_cpu_read(current_bit_idx);
1342
1343                 word_idx = start_word_idx;
1344
1345                 for (i = 0; pending_words != 0; i++) {
1346                         xen_ulong_t pending_bits;
1347                         xen_ulong_t words;
1348
1349                         words = MASK_LSBS(pending_words, word_idx);
1350
1351                         /*
1352                          * If we masked out all events, wrap to beginning.
1353                          */
1354                         if (words == 0) {
1355                                 word_idx = 0;
1356                                 bit_idx = 0;
1357                                 continue;
1358                         }
1359                         word_idx = EVTCHN_FIRST_BIT(words);
1360
1361                         pending_bits = active_evtchns(cpu, s, word_idx);
1362                         bit_idx = 0; /* usually scan entire word from start */
1363                         if (word_idx == start_word_idx) {
1364                                 /* We scan the starting word in two parts */
1365                                 if (i == 0)
1366                                         /* 1st time: start in the middle */
1367                                         bit_idx = start_bit_idx;
1368                                 else
1369                                         /* 2nd time: mask bits done already */
1370                                         bit_idx &= (1UL << start_bit_idx) - 1;
1371                         }
1372
1373                         do {
1374                                 xen_ulong_t bits;
1375                                 int port, irq;
1376                                 struct irq_desc *desc;
1377
1378                                 bits = MASK_LSBS(pending_bits, bit_idx);
1379
1380                                 /* If we masked out all events, move on. */
1381                                 if (bits == 0)
1382                                         break;
1383
1384                                 bit_idx = EVTCHN_FIRST_BIT(bits);
1385
1386                                 /* Process port. */
1387                                 port = (word_idx * BITS_PER_EVTCHN_WORD) + bit_idx;
1388                                 irq = evtchn_to_irq[port];
1389
1390                                 if (irq != -1) {
1391                                         desc = irq_to_desc(irq);
1392                                         if (desc)
1393                                                 generic_handle_irq_desc(irq, desc);
1394                                 }
1395
1396                                 bit_idx = (bit_idx + 1) % BITS_PER_EVTCHN_WORD;
1397
1398                                 /* Next caller starts at last processed + 1 */
1399                                 __this_cpu_write(current_word_idx,
1400                                                  bit_idx ? word_idx :
1401                                                  (word_idx+1) % BITS_PER_EVTCHN_WORD);
1402                                 __this_cpu_write(current_bit_idx, bit_idx);
1403                         } while (bit_idx != 0);
1404
1405                         /* Scan start_l1i twice; all others once. */
1406                         if ((word_idx != start_word_idx) || (i != 0))
1407                                 pending_words &= ~(1UL << word_idx);
1408
1409                         word_idx = (word_idx + 1) % BITS_PER_EVTCHN_WORD;
1410                 }
1411
1412                 BUG_ON(!irqs_disabled());
1413
1414                 count = __this_cpu_read(xed_nesting_count);
1415                 __this_cpu_write(xed_nesting_count, 0);
1416         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
1417
1418 out:
1419
1420         put_cpu();
1421 }
1422
1423 void xen_evtchn_do_upcall(struct pt_regs *regs)
1424 {
1425         struct pt_regs *old_regs = set_irq_regs(regs);
1426
1427         irq_enter();
1428 #ifdef CONFIG_X86
1429         exit_idle();
1430 #endif
1431
1432         __xen_evtchn_do_upcall();
1433
1434         irq_exit();
1435         set_irq_regs(old_regs);
1436 }
1437
1438 void xen_hvm_evtchn_do_upcall(void)
1439 {
1440         __xen_evtchn_do_upcall();
1441 }
1442 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1443
1444 /* Rebind a new event channel to an existing irq. */
1445 void rebind_evtchn_irq(int evtchn, int irq)
1446 {
1447         struct irq_info *info = info_for_irq(irq);
1448
1449         /* Make sure the irq is masked, since the new event channel
1450            will also be masked. */
1451         disable_irq(irq);
1452
1453         mutex_lock(&irq_mapping_update_lock);
1454
1455         /* After resume the irq<->evtchn mappings are all cleared out */
1456         BUG_ON(evtchn_to_irq[evtchn] != -1);
1457         /* Expect irq to have been bound before,
1458            so there should be a proper type */
1459         BUG_ON(info->type == IRQT_UNBOUND);
1460
1461         xen_irq_info_evtchn_init(irq, evtchn);
1462
1463         mutex_unlock(&irq_mapping_update_lock);
1464
1465         /* new event channels are always bound to cpu 0 */
1466         irq_set_affinity(irq, cpumask_of(0));
1467
1468         /* Unmask the event channel. */
1469         enable_irq(irq);
1470 }
1471
1472 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1473 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1474 {
1475         struct evtchn_bind_vcpu bind_vcpu;
1476         int evtchn = evtchn_from_irq(irq);
1477
1478         if (!VALID_EVTCHN(evtchn))
1479                 return -1;
1480
1481         /*
1482          * Events delivered via platform PCI interrupts are always
1483          * routed to vcpu 0 and hence cannot be rebound.
1484          */
1485         if (xen_hvm_domain() && !xen_have_vector_callback)
1486                 return -1;
1487
1488         /* Send future instances of this interrupt to other vcpu. */
1489         bind_vcpu.port = evtchn;
1490         bind_vcpu.vcpu = tcpu;
1491
1492         /*
1493          * If this fails, it usually just indicates that we're dealing with a
1494          * virq or IPI channel, which don't actually need to be rebound. Ignore
1495          * it, but don't do the xenlinux-level rebind in that case.
1496          */
1497         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1498                 bind_evtchn_to_cpu(evtchn, tcpu);
1499
1500         return 0;
1501 }
1502
1503 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1504                             bool force)
1505 {
1506         unsigned tcpu = cpumask_first(dest);
1507
1508         return rebind_irq_to_cpu(data->irq, tcpu);
1509 }
1510
1511 int resend_irq_on_evtchn(unsigned int irq)
1512 {
1513         int masked, evtchn = evtchn_from_irq(irq);
1514         struct shared_info *s = HYPERVISOR_shared_info;
1515
1516         if (!VALID_EVTCHN(evtchn))
1517                 return 1;
1518
1519         masked = sync_test_and_set_bit(evtchn, BM(s->evtchn_mask));
1520         sync_set_bit(evtchn, BM(s->evtchn_pending));
1521         if (!masked)
1522                 unmask_evtchn(evtchn);
1523
1524         return 1;
1525 }
1526
1527 static void enable_dynirq(struct irq_data *data)
1528 {
1529         int evtchn = evtchn_from_irq(data->irq);
1530
1531         if (VALID_EVTCHN(evtchn))
1532                 unmask_evtchn(evtchn);
1533 }
1534
1535 static void disable_dynirq(struct irq_data *data)
1536 {
1537         int evtchn = evtchn_from_irq(data->irq);
1538
1539         if (VALID_EVTCHN(evtchn))
1540                 mask_evtchn(evtchn);
1541 }
1542
1543 static void ack_dynirq(struct irq_data *data)
1544 {
1545         int evtchn = evtchn_from_irq(data->irq);
1546
1547         irq_move_irq(data);
1548
1549         if (VALID_EVTCHN(evtchn))
1550                 clear_evtchn(evtchn);
1551 }
1552
1553 static void mask_ack_dynirq(struct irq_data *data)
1554 {
1555         disable_dynirq(data);
1556         ack_dynirq(data);
1557 }
1558
1559 static int retrigger_dynirq(struct irq_data *data)
1560 {
1561         int evtchn = evtchn_from_irq(data->irq);
1562         struct shared_info *sh = HYPERVISOR_shared_info;
1563         int ret = 0;
1564
1565         if (VALID_EVTCHN(evtchn)) {
1566                 int masked;
1567
1568                 masked = sync_test_and_set_bit(evtchn, BM(sh->evtchn_mask));
1569                 sync_set_bit(evtchn, BM(sh->evtchn_pending));
1570                 if (!masked)
1571                         unmask_evtchn(evtchn);
1572                 ret = 1;
1573         }
1574
1575         return ret;
1576 }
1577
1578 static void restore_pirqs(void)
1579 {
1580         int pirq, rc, irq, gsi;
1581         struct physdev_map_pirq map_irq;
1582         struct irq_info *info;
1583
1584         list_for_each_entry(info, &xen_irq_list_head, list) {
1585                 if (info->type != IRQT_PIRQ)
1586                         continue;
1587
1588                 pirq = info->u.pirq.pirq;
1589                 gsi = info->u.pirq.gsi;
1590                 irq = info->irq;
1591
1592                 /* save/restore of PT devices doesn't work, so at this point the
1593                  * only devices present are GSI based emulated devices */
1594                 if (!gsi)
1595                         continue;
1596
1597                 map_irq.domid = DOMID_SELF;
1598                 map_irq.type = MAP_PIRQ_TYPE_GSI;
1599                 map_irq.index = gsi;
1600                 map_irq.pirq = pirq;
1601
1602                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1603                 if (rc) {
1604                         printk(KERN_WARNING "xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1605                                         gsi, irq, pirq, rc);
1606                         xen_free_irq(irq);
1607                         continue;
1608                 }
1609
1610                 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1611
1612                 __startup_pirq(irq);
1613         }
1614 }
1615
1616 static void restore_cpu_virqs(unsigned int cpu)
1617 {
1618         struct evtchn_bind_virq bind_virq;
1619         int virq, irq, evtchn;
1620
1621         for (virq = 0; virq < NR_VIRQS; virq++) {
1622                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1623                         continue;
1624
1625                 BUG_ON(virq_from_irq(irq) != virq);
1626
1627                 /* Get a new binding from Xen. */
1628                 bind_virq.virq = virq;
1629                 bind_virq.vcpu = cpu;
1630                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1631                                                 &bind_virq) != 0)
1632                         BUG();
1633                 evtchn = bind_virq.port;
1634
1635                 /* Record the new mapping. */
1636                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
1637                 bind_evtchn_to_cpu(evtchn, cpu);
1638         }
1639 }
1640
1641 static void restore_cpu_ipis(unsigned int cpu)
1642 {
1643         struct evtchn_bind_ipi bind_ipi;
1644         int ipi, irq, evtchn;
1645
1646         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1647                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1648                         continue;
1649
1650                 BUG_ON(ipi_from_irq(irq) != ipi);
1651
1652                 /* Get a new binding from Xen. */
1653                 bind_ipi.vcpu = cpu;
1654                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1655                                                 &bind_ipi) != 0)
1656                         BUG();
1657                 evtchn = bind_ipi.port;
1658
1659                 /* Record the new mapping. */
1660                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
1661                 bind_evtchn_to_cpu(evtchn, cpu);
1662         }
1663 }
1664
1665 /* Clear an irq's pending state, in preparation for polling on it */
1666 void xen_clear_irq_pending(int irq)
1667 {
1668         int evtchn = evtchn_from_irq(irq);
1669
1670         if (VALID_EVTCHN(evtchn))
1671                 clear_evtchn(evtchn);
1672 }
1673 EXPORT_SYMBOL(xen_clear_irq_pending);
1674 void xen_set_irq_pending(int irq)
1675 {
1676         int evtchn = evtchn_from_irq(irq);
1677
1678         if (VALID_EVTCHN(evtchn))
1679                 set_evtchn(evtchn);
1680 }
1681
1682 bool xen_test_irq_pending(int irq)
1683 {
1684         int evtchn = evtchn_from_irq(irq);
1685         bool ret = false;
1686
1687         if (VALID_EVTCHN(evtchn))
1688                 ret = test_evtchn(evtchn);
1689
1690         return ret;
1691 }
1692
1693 /* Poll waiting for an irq to become pending with timeout.  In the usual case,
1694  * the irq will be disabled so it won't deliver an interrupt. */
1695 void xen_poll_irq_timeout(int irq, u64 timeout)
1696 {
1697         evtchn_port_t evtchn = evtchn_from_irq(irq);
1698
1699         if (VALID_EVTCHN(evtchn)) {
1700                 struct sched_poll poll;
1701
1702                 poll.nr_ports = 1;
1703                 poll.timeout = timeout;
1704                 set_xen_guest_handle(poll.ports, &evtchn);
1705
1706                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1707                         BUG();
1708         }
1709 }
1710 EXPORT_SYMBOL(xen_poll_irq_timeout);
1711 /* Poll waiting for an irq to become pending.  In the usual case, the
1712  * irq will be disabled so it won't deliver an interrupt. */
1713 void xen_poll_irq(int irq)
1714 {
1715         xen_poll_irq_timeout(irq, 0 /* no timeout */);
1716 }
1717
1718 /* Check whether the IRQ line is shared with other guests. */
1719 int xen_test_irq_shared(int irq)
1720 {
1721         struct irq_info *info = info_for_irq(irq);
1722         struct physdev_irq_status_query irq_status = { .irq = info->u.pirq.pirq };
1723
1724         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
1725                 return 0;
1726         return !(irq_status.flags & XENIRQSTAT_shared);
1727 }
1728 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
1729
1730 void xen_irq_resume(void)
1731 {
1732         unsigned int cpu, evtchn;
1733         struct irq_info *info;
1734
1735         init_evtchn_cpu_bindings();
1736
1737         /* New event-channel space is not 'live' yet. */
1738         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1739                 mask_evtchn(evtchn);
1740
1741         /* No IRQ <-> event-channel mappings. */
1742         list_for_each_entry(info, &xen_irq_list_head, list)
1743                 info->evtchn = 0; /* zap event-channel binding */
1744
1745         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1746                 evtchn_to_irq[evtchn] = -1;
1747
1748         for_each_possible_cpu(cpu) {
1749                 restore_cpu_virqs(cpu);
1750                 restore_cpu_ipis(cpu);
1751         }
1752
1753         restore_pirqs();
1754 }
1755
1756 static struct irq_chip xen_dynamic_chip __read_mostly = {
1757         .name                   = "xen-dyn",
1758
1759         .irq_disable            = disable_dynirq,
1760         .irq_mask               = disable_dynirq,
1761         .irq_unmask             = enable_dynirq,
1762
1763         .irq_ack                = ack_dynirq,
1764         .irq_mask_ack           = mask_ack_dynirq,
1765
1766         .irq_set_affinity       = set_affinity_irq,
1767         .irq_retrigger          = retrigger_dynirq,
1768 };
1769
1770 static struct irq_chip xen_pirq_chip __read_mostly = {
1771         .name                   = "xen-pirq",
1772
1773         .irq_startup            = startup_pirq,
1774         .irq_shutdown           = shutdown_pirq,
1775         .irq_enable             = enable_pirq,
1776         .irq_disable            = disable_pirq,
1777
1778         .irq_mask               = disable_dynirq,
1779         .irq_unmask             = enable_dynirq,
1780
1781         .irq_ack                = eoi_pirq,
1782         .irq_eoi                = eoi_pirq,
1783         .irq_mask_ack           = mask_ack_pirq,
1784
1785         .irq_set_affinity       = set_affinity_irq,
1786
1787         .irq_retrigger          = retrigger_dynirq,
1788 };
1789
1790 static struct irq_chip xen_percpu_chip __read_mostly = {
1791         .name                   = "xen-percpu",
1792
1793         .irq_disable            = disable_dynirq,
1794         .irq_mask               = disable_dynirq,
1795         .irq_unmask             = enable_dynirq,
1796
1797         .irq_ack                = ack_dynirq,
1798 };
1799
1800 int xen_set_callback_via(uint64_t via)
1801 {
1802         struct xen_hvm_param a;
1803         a.domid = DOMID_SELF;
1804         a.index = HVM_PARAM_CALLBACK_IRQ;
1805         a.value = via;
1806         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1807 }
1808 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1809
1810 #ifdef CONFIG_XEN_PVHVM
1811 /* Vector callbacks are better than PCI interrupts to receive event
1812  * channel notifications because we can receive vector callbacks on any
1813  * vcpu and we don't need PCI support or APIC interactions. */
1814 void xen_callback_vector(void)
1815 {
1816         int rc;
1817         uint64_t callback_via;
1818         if (xen_have_vector_callback) {
1819                 callback_via = HVM_CALLBACK_VECTOR(HYPERVISOR_CALLBACK_VECTOR);
1820                 rc = xen_set_callback_via(callback_via);
1821                 if (rc) {
1822                         printk(KERN_ERR "Request for Xen HVM callback vector"
1823                                         " failed.\n");
1824                         xen_have_vector_callback = 0;
1825                         return;
1826                 }
1827                 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1828                                 "enabled\n");
1829                 /* in the restore case the vector has already been allocated */
1830                 if (!test_bit(HYPERVISOR_CALLBACK_VECTOR, used_vectors))
1831                         alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR,
1832                                         xen_hvm_callback_vector);
1833         }
1834 }
1835 #else
1836 void xen_callback_vector(void) {}
1837 #endif
1838
1839 void __init xen_init_IRQ(void)
1840 {
1841         int i;
1842
1843         evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1844                                     GFP_KERNEL);
1845         BUG_ON(!evtchn_to_irq);
1846         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1847                 evtchn_to_irq[i] = -1;
1848
1849         init_evtchn_cpu_bindings();
1850
1851         /* No event channels are 'live' right now. */
1852         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1853                 mask_evtchn(i);
1854
1855         pirq_needs_eoi = pirq_needs_eoi_flag;
1856
1857 #ifdef CONFIG_X86
1858         if (xen_hvm_domain()) {
1859                 xen_callback_vector();
1860                 native_init_IRQ();
1861                 /* pci_xen_hvm_init must be called after native_init_IRQ so that
1862                  * __acpi_register_gsi can point at the right function */
1863                 pci_xen_hvm_init();
1864         } else {
1865                 int rc;
1866                 struct physdev_pirq_eoi_gmfn eoi_gmfn;
1867
1868                 irq_ctx_init(smp_processor_id());
1869                 if (xen_initial_domain())
1870                         pci_xen_initial_domain();
1871
1872                 pirq_eoi_map = (void *)__get_free_page(GFP_KERNEL|__GFP_ZERO);
1873                 eoi_gmfn.gmfn = virt_to_mfn(pirq_eoi_map);
1874                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_pirq_eoi_gmfn_v2, &eoi_gmfn);
1875                 if (rc != 0) {
1876                         free_page((unsigned long) pirq_eoi_map);
1877                         pirq_eoi_map = NULL;
1878                 } else
1879                         pirq_needs_eoi = pirq_check_eoi_map;
1880         }
1881 #endif
1882 }