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 #include <asm/desc.h>
  35 #include <asm/ptrace.h>
  36 #include <asm/irq.h>
  37 #include <asm/idle.h>
  38 #include <asm/io_apic.h>
  39 #include <asm/sync_bitops.h>
  40 #include <asm/xen/pci.h>
  41 #include <asm/xen/hypercall.h>
  42 #include <asm/xen/hypervisor.h>
  43
  44 #include <xen/xen.h>
  45 #include <xen/hvm.h>
  46 #include <xen/xen-ops.h>
  47 #include <xen/events.h>
  48 #include <xen/interface/xen.h>
  49 #include <xen/interface/event_channel.h>
  50 #include <xen/interface/hvm/hvm_op.h>
  51 #include <xen/interface/hvm/params.h>
  52
  53 /*
  54  * This lock protects updates to the following mapping and reference-count
  55  * arrays. The lock does not need to be acquired to read the mapping tables.
  56  */
  57 static DEFINE_MUTEX(irq_mapping_update_lock);
  58
  59 static LIST_HEAD(xen_irq_list_head);
  60
  61 /* IRQ <-> VIRQ mapping. */
  62 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
  63
  64 /* IRQ <-> IPI mapping */
  65 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
  66
  67 /* Interrupt types. */
  68 enum xen_irq_type {
  69         IRQT_UNBOUND = 0,
  70         IRQT_PIRQ,
  71         IRQT_VIRQ,
  72         IRQT_IPI,
  73         IRQT_EVTCHN
  74 };
  75
  76 /*
  77  * Packed IRQ information:
  78  * type - enum xen_irq_type
  79  * event channel - irq->event channel mapping
  80  * cpu - cpu this event channel is bound to
  81  * index - type-specific information:
  82  *    PIRQ - vector, with MSB being "needs EIO", or physical IRQ of the HVM
  83  *           guest, or GSI (real passthrough IRQ) of the device.
  84  *    VIRQ - virq number
  85  *    IPI - IPI vector
  86  *    EVTCHN -
  87  */
  88 struct irq_info {
  89         struct list_head list;
  90         enum xen_irq_type type; /* type */
  91         unsigned irq;
  92         unsigned short evtchn;  /* event channel */
  93         unsigned short cpu;     /* cpu bound */
  94
  95         union {
  96                 unsigned short virq;
  97                 enum ipi_vector ipi;
  98                 struct {
  99                         unsigned short pirq;
 100                         unsigned short gsi;
 101                         unsigned char vector;
 102                         unsigned char flags;
 103                         uint16_t domid;
 104                 } pirq;
 105         } u;
 106 };
 107 #define PIRQ_NEEDS_EOI  (1 << 0)
 108 #define PIRQ_SHAREABLE  (1 << 1)
 109
 110 static int *evtchn_to_irq;
 111
 112 static DEFINE_PER_CPU(unsigned long [NR_EVENT_CHANNELS/BITS_PER_LONG],
 113                       cpu_evtchn_mask);
 114
 115 /* Xen will never allocate port zero for any purpose. */
 116 #define VALID_EVTCHN(chn)       ((chn) != 0)
 117
 118 static struct irq_chip xen_dynamic_chip;
 119 static struct irq_chip xen_percpu_chip;
 120 static struct irq_chip xen_pirq_chip;
 121 static void enable_dynirq(struct irq_data *data);
 122 static void disable_dynirq(struct irq_data *data);
 123
 124 /* Get info for IRQ */
 125 static struct irq_info *info_for_irq(unsigned irq)
 126 {
 127         return irq_get_handler_data(irq);
 128 }
 129
 130 /* Constructors for packed IRQ information. */
 131 static void xen_irq_info_common_init(struct irq_info *info,
 132                                      unsigned irq,
 133                                      enum xen_irq_type type,
 134                                      unsigned short evtchn,
 135                                      unsigned short cpu)
 136 {
 137
 138         BUG_ON(info->type != IRQT_UNBOUND && info->type != type);
 139
 140         info->type = type;
 141         info->irq = irq;
 142         info->evtchn = evtchn;
 143         info->cpu = cpu;
 144
 145         evtchn_to_irq[evtchn] = irq;
 146 }
 147
 148 static void xen_irq_info_evtchn_init(unsigned irq,
 149                                      unsigned short evtchn)
 150 {
 151         struct irq_info *info = info_for_irq(irq);
 152
 153         xen_irq_info_common_init(info, irq, IRQT_EVTCHN, evtchn, 0);
 154 }
 155
 156 static void xen_irq_info_ipi_init(unsigned cpu,
 157                                   unsigned irq,
 158                                   unsigned short evtchn,
 159                                   enum ipi_vector ipi)
 160 {
 161         struct irq_info *info = info_for_irq(irq);
 162
 163         xen_irq_info_common_init(info, irq, IRQT_IPI, evtchn, 0);
 164
 165         info->u.ipi = ipi;
 166
 167         per_cpu(ipi_to_irq, cpu)[ipi] = irq;
 168 }
 169
 170 static void xen_irq_info_virq_init(unsigned cpu,
 171                                    unsigned irq,
 172                                    unsigned short evtchn,
 173                                    unsigned short virq)
 174 {
 175         struct irq_info *info = info_for_irq(irq);
 176
 177         xen_irq_info_common_init(info, irq, IRQT_VIRQ, evtchn, 0);
 178
 179         info->u.virq = virq;
 180
 181         per_cpu(virq_to_irq, cpu)[virq] = irq;
 182 }
 183
 184 static void xen_irq_info_pirq_init(unsigned irq,
 185                                    unsigned short evtchn,
 186                                    unsigned short pirq,
 187                                    unsigned short gsi,
 188                                    unsigned short vector,
 189                                    uint16_t domid,
 190                                    unsigned char flags)
 191 {
 192         struct irq_info *info = info_for_irq(irq);
 193
 194         xen_irq_info_common_init(info, irq, IRQT_PIRQ, evtchn, 0);
 195
 196         info->u.pirq.pirq = pirq;
 197         info->u.pirq.gsi = gsi;
 198         info->u.pirq.vector = vector;
 199         info->u.pirq.domid = domid;
 200         info->u.pirq.flags = flags;
 201 }
 202
 203 /*
 204  * Accessors for packed IRQ information.
 205  */
 206 static unsigned int evtchn_from_irq(unsigned irq)
 207 {
 208         if (unlikely(WARN(irq < 0 || irq >= nr_irqs, "Invalid irq %d!\n", irq)))
 209                 return 0;
 210
 211         return info_for_irq(irq)->evtchn;
 212 }
 213
 214 unsigned irq_from_evtchn(unsigned int evtchn)
 215 {
 216         return evtchn_to_irq[evtchn];
 217 }
 218 EXPORT_SYMBOL_GPL(irq_from_evtchn);
 219
 220 static enum ipi_vector ipi_from_irq(unsigned irq)
 221 {
 222         struct irq_info *info = info_for_irq(irq);
 223
 224         BUG_ON(info == NULL);
 225         BUG_ON(info->type != IRQT_IPI);
 226
 227         return info->u.ipi;
 228 }
 229
 230 static unsigned virq_from_irq(unsigned irq)
 231 {
 232         struct irq_info *info = info_for_irq(irq);
 233
 234         BUG_ON(info == NULL);
 235         BUG_ON(info->type != IRQT_VIRQ);
 236
 237         return info->u.virq;
 238 }
 239
 240 static unsigned pirq_from_irq(unsigned irq)
 241 {
 242         struct irq_info *info = info_for_irq(irq);
 243
 244         BUG_ON(info == NULL);
 245         BUG_ON(info->type != IRQT_PIRQ);
 246
 247         return info->u.pirq.pirq;
 248 }
 249
 250 static enum xen_irq_type type_from_irq(unsigned irq)
 251 {
 252         return info_for_irq(irq)->type;
 253 }
 254
 255 static unsigned cpu_from_irq(unsigned irq)
 256 {
 257         return info_for_irq(irq)->cpu;
 258 }
 259
 260 static unsigned int cpu_from_evtchn(unsigned int evtchn)
 261 {
 262         int irq = evtchn_to_irq[evtchn];
 263         unsigned ret = 0;
 264
 265         if (irq != -1)
 266                 ret = cpu_from_irq(irq);
 267
 268         return ret;
 269 }
 270
 271 static bool pirq_needs_eoi(unsigned irq)
 272 {
 273         struct irq_info *info = info_for_irq(irq);
 274
 275         BUG_ON(info->type != IRQT_PIRQ);
 276
 277         return info->u.pirq.flags & PIRQ_NEEDS_EOI;
 278 }
 279
 280 static inline unsigned long active_evtchns(unsigned int cpu,
 281                                            struct shared_info *sh,
 282                                            unsigned int idx)
 283 {
 284         return sh->evtchn_pending[idx] &
 285                 per_cpu(cpu_evtchn_mask, cpu)[idx] &
 286                 ~sh->evtchn_mask[idx];
 287 }
 288
 289 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
 290 {
 291         int irq = evtchn_to_irq[chn];
 292
 293         BUG_ON(irq == -1);
 294 #ifdef CONFIG_SMP
 295         cpumask_copy(irq_to_desc(irq)->irq_data.affinity, cpumask_of(cpu));
 296 #endif
 297
 298         clear_bit(chn, per_cpu(cpu_evtchn_mask, cpu_from_irq(irq)));
 299         set_bit(chn, per_cpu(cpu_evtchn_mask, cpu));
 300
 301         info_for_irq(irq)->cpu = cpu;
 302 }
 303
 304 static void init_evtchn_cpu_bindings(void)
 305 {
 306         int i;
 307 #ifdef CONFIG_SMP
 308         struct irq_info *info;
 309
 310         /* By default all event channels notify CPU#0. */
 311         list_for_each_entry(info, &xen_irq_list_head, list) {
 312                 struct irq_desc *desc = irq_to_desc(info->irq);
 313                 cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 314         }
 315 #endif
 316
 317         for_each_possible_cpu(i)
 318                 memset(per_cpu(cpu_evtchn_mask, i),
 319                        (i == 0) ? ~0 : 0, NR_EVENT_CHANNELS/8);
 320 }
 321
 322 static inline void clear_evtchn(int port)
 323 {
 324         struct shared_info *s = HYPERVISOR_shared_info;
 325         sync_clear_bit(port, &s->evtchn_pending[0]);
 326 }
 327
 328 static inline void set_evtchn(int port)
 329 {
 330         struct shared_info *s = HYPERVISOR_shared_info;
 331         sync_set_bit(port, &s->evtchn_pending[0]);
 332 }
 333
 334 static inline int test_evtchn(int port)
 335 {
 336         struct shared_info *s = HYPERVISOR_shared_info;
 337         return sync_test_bit(port, &s->evtchn_pending[0]);
 338 }
 339
 340 static inline int test_and_set_mask(int port)
 341 {
 342         struct shared_info *s = HYPERVISOR_shared_info;
 343         return sync_test_and_set_bit(port, &s->evtchn_mask[0]);
 344 }
 345
 346
 347 /**
 348  * notify_remote_via_irq - send event to remote end of event channel via irq
 349  * @irq: irq of event channel to send event to
 350  *
 351  * Unlike notify_remote_via_evtchn(), this is safe to use across
 352  * save/restore. Notifications on a broken connection are silently
 353  * dropped.
 354  */
 355 void notify_remote_via_irq(int irq)
 356 {
 357         int evtchn = evtchn_from_irq(irq);
 358
 359         if (VALID_EVTCHN(evtchn))
 360                 notify_remote_via_evtchn(evtchn);
 361 }
 362 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
 363
 364 static void mask_evtchn(int port)
 365 {
 366         struct shared_info *s = HYPERVISOR_shared_info;
 367         sync_set_bit(port, &s->evtchn_mask[0]);
 368 }
 369
 370 static void unmask_evtchn(int port)
 371 {
 372         struct shared_info *s = HYPERVISOR_shared_info;
 373         unsigned int cpu = get_cpu();
 374
 375         BUG_ON(!irqs_disabled());
 376
 377         /* Slow path (hypercall) if this is a non-local port. */
 378         if (unlikely(cpu != cpu_from_evtchn(port))) {
 379                 struct evtchn_unmask unmask = { .port = port };
 380                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 381         } else {
 382                 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
 383
 384                 sync_clear_bit(port, &s->evtchn_mask[0]);
 385
 386                 /*
 387                  * The following is basically the equivalent of
 388                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
 389                  * the interrupt edge' if the channel is masked.
 390                  */
 391                 if (sync_test_bit(port, &s->evtchn_pending[0]) &&
 392                     !sync_test_and_set_bit(port / BITS_PER_LONG,
 393                                            &vcpu_info->evtchn_pending_sel))
 394                         vcpu_info->evtchn_upcall_pending = 1;
 395         }
 396
 397         put_cpu();
 398 }
 399
 400 static void xen_irq_init(unsigned irq)
 401 {
 402         struct irq_info *info;
 403 #ifdef CONFIG_SMP
 404         struct irq_desc *desc = irq_to_desc(irq);
 405
 406         /* By default all event channels notify CPU#0. */
 407         cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 408 #endif
 409
 410         info = kzalloc(sizeof(*info), GFP_KERNEL);
 411         if (info == NULL)
 412                 panic("Unable to allocate metadata for IRQ%d\n", irq);
 413
 414         info->type = IRQT_UNBOUND;
 415
 416         irq_set_handler_data(irq, info);
 417
 418         list_add_tail(&info->list, &xen_irq_list_head);
 419 }
 420
 421 static int __must_check xen_allocate_irq_dynamic(void)
 422 {
 423         int first = 0;
 424         int irq;
 425
 426 #ifdef CONFIG_X86_IO_APIC
 427         /*
 428          * For an HVM guest or domain 0 which see "real" (emulated or
 429          * actual respectively) GSIs we allocate dynamic IRQs
 430          * e.g. those corresponding to event channels or MSIs
 431          * etc. from the range above those "real" GSIs to avoid
 432          * collisions.
 433          */
 434         if (xen_initial_domain() || xen_hvm_domain())
 435                 first = get_nr_irqs_gsi();
 436 #endif
 437
 438         irq = irq_alloc_desc_from(first, -1);
 439
 440         if (irq >= 0)
 441                 xen_irq_init(irq);
 442
 443         return irq;
 444 }
 445
 446 static int __must_check xen_allocate_irq_gsi(unsigned gsi)
 447 {
 448         int irq;
 449
 450         /*
 451          * A PV guest has no concept of a GSI (since it has no ACPI
 452          * nor access to/knowledge of the physical APICs). Therefore
 453          * all IRQs are dynamically allocated from the entire IRQ
 454          * space.
 455          */
 456         if (xen_pv_domain() && !xen_initial_domain())
 457                 return xen_allocate_irq_dynamic();
 458
 459         /* Legacy IRQ descriptors are already allocated by the arch. */
 460         if (gsi < NR_IRQS_LEGACY)
 461                 irq = gsi;
 462         else
 463                 irq = irq_alloc_desc_at(gsi, -1);
 464
 465         xen_irq_init(irq);
 466
 467         return irq;
 468 }
 469
 470 static void xen_free_irq(unsigned irq)
 471 {
 472         struct irq_info *info = irq_get_handler_data(irq);
 473
 474         list_del(&info->list);
 475
 476         irq_set_handler_data(irq, NULL);
 477
 478         kfree(info);
 479
 480         /* Legacy IRQ descriptors are managed by the arch. */
 481         if (irq < NR_IRQS_LEGACY)
 482                 return;
 483
 484         irq_free_desc(irq);
 485 }
 486
 487 static void pirq_query_unmask(int irq)
 488 {
 489         struct physdev_irq_status_query irq_status;
 490         struct irq_info *info = info_for_irq(irq);
 491
 492         BUG_ON(info->type != IRQT_PIRQ);
 493
 494         irq_status.irq = pirq_from_irq(irq);
 495         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
 496                 irq_status.flags = 0;
 497
 498         info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
 499         if (irq_status.flags & XENIRQSTAT_needs_eoi)
 500                 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
 501 }
 502
 503 static bool probing_irq(int irq)
 504 {
 505         struct irq_desc *desc = irq_to_desc(irq);
 506
 507         return desc && desc->action == NULL;
 508 }
 509
 510 static void eoi_pirq(struct irq_data *data)
 511 {
 512         int evtchn = evtchn_from_irq(data->irq);
 513         struct physdev_eoi eoi = { .irq = pirq_from_irq(data->irq) };
 514         int rc = 0;
 515
 516         if (!VALID_EVTCHN(evtchn))
 517                 return;
 518
 519         if (unlikely(irqd_is_setaffinity_pending(data)) &&
 520             likely(!irqd_irq_disabled(data))) {
 521                 int masked = test_and_set_mask(evtchn);
 522
 523                 clear_evtchn(evtchn);
 524
 525                 irq_move_masked_irq(data);
 526
 527                 if (!masked)
 528                         unmask_evtchn(evtchn);
 529         } else
 530                 clear_evtchn(evtchn);
 531
 532         if (pirq_needs_eoi(data->irq)) {
 533                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
 534                 WARN_ON(rc);
 535         }
 536 }
 537
 538 static void mask_ack_pirq(struct irq_data *data)
 539 {
 540         disable_dynirq(data);
 541         eoi_pirq(data);
 542 }
 543
 544 static unsigned int __startup_pirq(unsigned int irq)
 545 {
 546         struct evtchn_bind_pirq bind_pirq;
 547         struct irq_info *info = info_for_irq(irq);
 548         int evtchn = evtchn_from_irq(irq);
 549         int rc;
 550
 551         BUG_ON(info->type != IRQT_PIRQ);
 552
 553         if (VALID_EVTCHN(evtchn))
 554                 goto out;
 555
 556         bind_pirq.pirq = pirq_from_irq(irq);
 557         /* NB. We are happy to share unless we are probing. */
 558         bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
 559                                         BIND_PIRQ__WILL_SHARE : 0;
 560         rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
 561         if (rc != 0) {
 562                 if (!probing_irq(irq))
 563                         printk(KERN_INFO "Failed to obtain physical IRQ %d\n",
 564                                irq);
 565                 return 0;
 566         }
 567         evtchn = bind_pirq.port;
 568
 569         pirq_query_unmask(irq);
 570
 571         evtchn_to_irq[evtchn] = irq;
 572         bind_evtchn_to_cpu(evtchn, 0);
 573         info->evtchn = evtchn;
 574
 575 out:
 576         unmask_evtchn(evtchn);
 577         eoi_pirq(irq_get_irq_data(irq));
 578
 579         return 0;
 580 }
 581
 582 static unsigned int startup_pirq(struct irq_data *data)
 583 {
 584         return __startup_pirq(data->irq);
 585 }
 586
 587 static void shutdown_pirq(struct irq_data *data)
 588 {
 589         struct evtchn_close close;
 590         unsigned int irq = data->irq;
 591         struct irq_info *info = info_for_irq(irq);
 592         int evtchn = evtchn_from_irq(irq);
 593
 594         BUG_ON(info->type != IRQT_PIRQ);
 595
 596         if (!VALID_EVTCHN(evtchn))
 597                 return;
 598
 599         mask_evtchn(evtchn);
 600
 601         close.port = evtchn;
 602         if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 603                 BUG();
 604
 605         bind_evtchn_to_cpu(evtchn, 0);
 606         evtchn_to_irq[evtchn] = -1;
 607         info->evtchn = 0;
 608 }
 609
 610 static void enable_pirq(struct irq_data *data)
 611 {
 612         startup_pirq(data);
 613 }
 614
 615 static void disable_pirq(struct irq_data *data)
 616 {
 617         disable_dynirq(data);
 618 }
 619
 620 int xen_irq_from_gsi(unsigned gsi)
 621 {
 622         struct irq_info *info;
 623
 624         list_for_each_entry(info, &xen_irq_list_head, list) {
 625                 if (info->type != IRQT_PIRQ)
 626                         continue;
 627
 628                 if (info->u.pirq.gsi == gsi)
 629                         return info->irq;
 630         }
 631
 632         return -1;
 633 }
 634 EXPORT_SYMBOL_GPL(xen_irq_from_gsi);
 635
 636 /*
 637  * Do not make any assumptions regarding the relationship between the
 638  * IRQ number returned here and the Xen pirq argument.
 639  *
 640  * Note: We don't assign an event channel until the irq actually started
 641  * up.  Return an existing irq if we've already got one for the gsi.
 642  *
 643  * Shareable implies level triggered, not shareable implies edge
 644  * triggered here.
 645  */
 646 int xen_bind_pirq_gsi_to_irq(unsigned gsi,
 647                              unsigned pirq, int shareable, char *name)
 648 {
 649         int irq = -1;
 650         struct physdev_irq irq_op;
 651
 652         mutex_lock(&irq_mapping_update_lock);
 653
 654         irq = xen_irq_from_gsi(gsi);
 655         if (irq != -1) {
 656                 printk(KERN_INFO "xen_map_pirq_gsi: returning irq %d for gsi %u\n",
 657                        irq, gsi);
 658                 goto out;       /* XXX need refcount? */
 659         }
 660
 661         irq = xen_allocate_irq_gsi(gsi);
 662         if (irq < 0)
 663                 goto out;
 664
 665         irq_op.irq = irq;
 666         irq_op.vector = 0;
 667
 668         /* Only the privileged domain can do this. For non-priv, the pcifront
 669          * driver provides a PCI bus that does the call to do exactly
 670          * this in the priv domain. */
 671         if (xen_initial_domain() &&
 672             HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
 673                 xen_free_irq(irq);
 674                 irq = -ENOSPC;
 675                 goto out;
 676         }
 677
 678         xen_irq_info_pirq_init(irq, 0, pirq, gsi, irq_op.vector, DOMID_SELF,
 679                                shareable ? PIRQ_SHAREABLE : 0);
 680
 681         pirq_query_unmask(irq);
 682         /* We try to use the handler with the appropriate semantic for the
 683          * type of interrupt: if the interrupt is an edge triggered
 684          * interrupt we use handle_edge_irq.
 685          *
 686          * On the other hand if the interrupt is level triggered we use
 687          * handle_fasteoi_irq like the native code does for this kind of
 688          * interrupts.
 689          *
 690          * Depending on the Xen version, pirq_needs_eoi might return true
 691          * not only for level triggered interrupts but for edge triggered
 692          * interrupts too. In any case Xen always honors the eoi mechanism,
 693          * not injecting any more pirqs of the same kind if the first one
 694          * hasn't received an eoi yet. Therefore using the fasteoi handler
 695          * is the right choice either way.
 696          */
 697         if (shareable)
 698                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 699                                 handle_fasteoi_irq, name);
 700         else
 701                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 702                                 handle_edge_irq, name);
 703
 704 out:
 705         mutex_unlock(&irq_mapping_update_lock);
 706
 707         return irq;
 708 }
 709
 710 #ifdef CONFIG_PCI_MSI
 711 int xen_allocate_pirq_msi(struct pci_dev *dev, struct msi_desc *msidesc)
 712 {
 713         int rc;
 714         struct physdev_get_free_pirq op_get_free_pirq;
 715
 716         op_get_free_pirq.type = MAP_PIRQ_TYPE_MSI;
 717         rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
 718
 719         WARN_ONCE(rc == -ENOSYS,
 720                   "hypervisor does not support the PHYSDEVOP_get_free_pirq interface\n");
 721
 722         return rc ? -1 : op_get_free_pirq.pirq;
 723 }
 724
 725 int xen_bind_pirq_msi_to_irq(struct pci_dev *dev, struct msi_desc *msidesc,
 726                              int pirq, int vector, const char *name,
 727                              domid_t domid)
 728 {
 729         int irq, ret;
 730
 731         mutex_lock(&irq_mapping_update_lock);
 732
 733         irq = xen_allocate_irq_dynamic();
 734         if (irq < 0)
 735                 goto out;
 736
 737         irq_set_chip_and_handler_name(irq, &xen_pirq_chip, handle_edge_irq,
 738                         name);
 739
 740         xen_irq_info_pirq_init(irq, 0, pirq, 0, vector, domid, 0);
 741         ret = irq_set_msi_desc(irq, msidesc);
 742         if (ret < 0)
 743                 goto error_irq;
 744 out:
 745         mutex_unlock(&irq_mapping_update_lock);
 746         return irq;
 747 error_irq:
 748         mutex_unlock(&irq_mapping_update_lock);
 749         xen_free_irq(irq);
 750         return ret;
 751 }
 752 #endif
 753
 754 int xen_destroy_irq(int irq)
 755 {
 756         struct irq_desc *desc;
 757         struct physdev_unmap_pirq unmap_irq;
 758         struct irq_info *info = info_for_irq(irq);
 759         int rc = -ENOENT;
 760
 761         mutex_lock(&irq_mapping_update_lock);
 762
 763         desc = irq_to_desc(irq);
 764         if (!desc)
 765                 goto out;
 766
 767         if (xen_initial_domain()) {
 768                 unmap_irq.pirq = info->u.pirq.pirq;
 769                 unmap_irq.domid = info->u.pirq.domid;
 770                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
 771                 /* If another domain quits without making the pci_disable_msix
 772                  * call, the Xen hypervisor takes care of freeing the PIRQs
 773                  * (free_domain_pirqs).
 774                  */
 775                 if ((rc == -ESRCH && info->u.pirq.domid != DOMID_SELF))
 776                         printk(KERN_INFO "domain %d does not have %d anymore\n",
 777                                 info->u.pirq.domid, info->u.pirq.pirq);
 778                 else if (rc) {
 779                         printk(KERN_WARNING "unmap irq failed %d\n", rc);
 780                         goto out;
 781                 }
 782         }
 783
 784         xen_free_irq(irq);
 785
 786 out:
 787         mutex_unlock(&irq_mapping_update_lock);
 788         return rc;
 789 }
 790
 791 int xen_irq_from_pirq(unsigned pirq)
 792 {
 793         int irq;
 794
 795         struct irq_info *info;
 796
 797         mutex_lock(&irq_mapping_update_lock);
 798
 799         list_for_each_entry(info, &xen_irq_list_head, list) {
 800                 if (info->type != IRQT_PIRQ)
 801                         continue;
 802                 irq = info->irq;
 803                 if (info->u.pirq.pirq == pirq)
 804                         goto out;
 805         }
 806         irq = -1;
 807 out:
 808         mutex_unlock(&irq_mapping_update_lock);
 809
 810         return irq;
 811 }
 812
 813
 814 int xen_pirq_from_irq(unsigned irq)
 815 {
 816         return pirq_from_irq(irq);
 817 }
 818 EXPORT_SYMBOL_GPL(xen_pirq_from_irq);
 819 int bind_evtchn_to_irq(unsigned int evtchn)
 820 {
 821         int irq;
 822
 823         mutex_lock(&irq_mapping_update_lock);
 824
 825         irq = evtchn_to_irq[evtchn];
 826
 827         if (irq == -1) {
 828                 irq = xen_allocate_irq_dynamic();
 829                 if (irq == -1)
 830                         goto out;
 831
 832                 irq_set_chip_and_handler_name(irq, &xen_dynamic_chip,
 833                                               handle_edge_irq, "event");
 834
 835                 xen_irq_info_evtchn_init(irq, evtchn);
 836         }
 837
 838 out:
 839         mutex_unlock(&irq_mapping_update_lock);
 840
 841         return irq;
 842 }
 843 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
 844
 845 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
 846 {
 847         struct evtchn_bind_ipi bind_ipi;
 848         int evtchn, irq;
 849
 850         mutex_lock(&irq_mapping_update_lock);
 851
 852         irq = per_cpu(ipi_to_irq, cpu)[ipi];
 853
 854         if (irq == -1) {
 855                 irq = xen_allocate_irq_dynamic();
 856                 if (irq < 0)
 857                         goto out;
 858
 859                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 860                                               handle_percpu_irq, "ipi");
 861
 862                 bind_ipi.vcpu = cpu;
 863                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
 864                                                 &bind_ipi) != 0)
 865                         BUG();
 866                 evtchn = bind_ipi.port;
 867
 868                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
 869
 870                 bind_evtchn_to_cpu(evtchn, cpu);
 871         }
 872
 873  out:
 874         mutex_unlock(&irq_mapping_update_lock);
 875         return irq;
 876 }
 877
 878 static int bind_interdomain_evtchn_to_irq(unsigned int remote_domain,
 879                                           unsigned int remote_port)
 880 {
 881         struct evtchn_bind_interdomain bind_interdomain;
 882         int err;
 883
 884         bind_interdomain.remote_dom  = remote_domain;
 885         bind_interdomain.remote_port = remote_port;
 886
 887         err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
 888                                           &bind_interdomain);
 889
 890         return err ? : bind_evtchn_to_irq(bind_interdomain.local_port);
 891 }
 892
 893 static int find_virq(unsigned int virq, unsigned int cpu)
 894 {
 895         struct evtchn_status status;
 896         int port, rc = -ENOENT;
 897
 898         memset(&status, 0, sizeof(status));
 899         for (port = 0; port <= NR_EVENT_CHANNELS; port++) {
 900                 status.dom = DOMID_SELF;
 901                 status.port = port;
 902                 rc = HYPERVISOR_event_channel_op(EVTCHNOP_status, &status);
 903                 if (rc < 0)
 904                         continue;
 905                 if (status.status != EVTCHNSTAT_virq)
 906                         continue;
 907                 if (status.u.virq == virq && status.vcpu == cpu) {
 908                         rc = port;
 909                         break;
 910                 }
 911         }
 912         return rc;
 913 }
 914
 915 int bind_virq_to_irq(unsigned int virq, unsigned int cpu, bool percpu)
 916 {
 917         struct evtchn_bind_virq bind_virq;
 918         int evtchn, irq, ret;
 919
 920         mutex_lock(&irq_mapping_update_lock);
 921
 922         irq = per_cpu(virq_to_irq, cpu)[virq];
 923
 924         if (irq == -1) {
 925                 irq = xen_allocate_irq_dynamic();
 926                 if (irq == -1)
 927                         goto out;
 928
 929                 if (percpu)
 930                         irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 931                                                       handle_percpu_irq, "virq");
 932                 else
 933                         irq_set_chip_and_handler_name(irq, &xen_dynamic_chip,
 934                                                       handle_edge_irq, "virq");
 935
 936                 bind_virq.virq = virq;
 937                 bind_virq.vcpu = cpu;
 938                 ret = HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 939                                                 &bind_virq);
 940                 if (ret == 0)
 941                         evtchn = bind_virq.port;
 942                 else {
 943                         if (ret == -EEXIST)
 944                                 ret = find_virq(virq, cpu);
 945                         BUG_ON(ret < 0);
 946                         evtchn = ret;
 947                 }
 948
 949                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
 950
 951                 bind_evtchn_to_cpu(evtchn, cpu);
 952         }
 953
 954 out:
 955         mutex_unlock(&irq_mapping_update_lock);
 956
 957         return irq;
 958 }
 959
 960 static void unbind_from_irq(unsigned int irq)
 961 {
 962         struct evtchn_close close;
 963         int evtchn = evtchn_from_irq(irq);
 964
 965         mutex_lock(&irq_mapping_update_lock);
 966
 967         if (VALID_EVTCHN(evtchn)) {
 968                 close.port = evtchn;
 969                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 970                         BUG();
 971
 972                 switch (type_from_irq(irq)) {
 973                 case IRQT_VIRQ:
 974                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
 975                                 [virq_from_irq(irq)] = -1;
 976                         break;
 977                 case IRQT_IPI:
 978                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
 979                                 [ipi_from_irq(irq)] = -1;
 980                         break;
 981                 default:
 982                         break;
 983                 }
 984
 985                 /* Closed ports are implicitly re-bound to VCPU0. */
 986                 bind_evtchn_to_cpu(evtchn, 0);
 987
 988                 evtchn_to_irq[evtchn] = -1;
 989         }
 990
 991         BUG_ON(info_for_irq(irq)->type == IRQT_UNBOUND);
 992
 993         xen_free_irq(irq);
 994
 995         mutex_unlock(&irq_mapping_update_lock);
 996 }
 997
 998 int bind_evtchn_to_irqhandler(unsigned int evtchn,
 999                               irq_handler_t handler,
1000                               unsigned long irqflags,
1001                               const char *devname, void *dev_id)
1002 {
1003         int irq, retval;
1004
1005         irq = bind_evtchn_to_irq(evtchn);
1006         if (irq < 0)
1007                 return irq;
1008         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1009         if (retval != 0) {
1010                 unbind_from_irq(irq);
1011                 return retval;
1012         }
1013
1014         return irq;
1015 }
1016 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
1017
1018 int bind_interdomain_evtchn_to_irqhandler(unsigned int remote_domain,
1019                                           unsigned int remote_port,
1020                                           irq_handler_t handler,
1021                                           unsigned long irqflags,
1022                                           const char *devname,
1023                                           void *dev_id)
1024 {
1025         int irq, retval;
1026
1027         irq = bind_interdomain_evtchn_to_irq(remote_domain, remote_port);
1028         if (irq < 0)
1029                 return irq;
1030
1031         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1032         if (retval != 0) {
1033                 unbind_from_irq(irq);
1034                 return retval;
1035         }
1036
1037         return irq;
1038 }
1039 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler);
1040
1041 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
1042                             irq_handler_t handler,
1043                             unsigned long irqflags, const char *devname, void *dev_id)
1044 {
1045         int irq, retval;
1046
1047         irq = bind_virq_to_irq(virq, cpu, irqflags & IRQF_PERCPU);
1048         if (irq < 0)
1049                 return irq;
1050         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1051         if (retval != 0) {
1052                 unbind_from_irq(irq);
1053                 return retval;
1054         }
1055
1056         return irq;
1057 }
1058 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1059
1060 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1061                            unsigned int cpu,
1062                            irq_handler_t handler,
1063                            unsigned long irqflags,
1064                            const char *devname,
1065                            void *dev_id)
1066 {
1067         int irq, retval;
1068
1069         irq = bind_ipi_to_irq(ipi, cpu);
1070         if (irq < 0)
1071                 return irq;
1072
1073         irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME | IRQF_EARLY_RESUME;
1074         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1075         if (retval != 0) {
1076                 unbind_from_irq(irq);
1077                 return retval;
1078         }
1079
1080         return irq;
1081 }
1082
1083 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1084 {
1085         free_irq(irq, dev_id);
1086         unbind_from_irq(irq);
1087 }
1088 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1089
1090 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1091 {
1092         int irq = per_cpu(ipi_to_irq, cpu)[vector];
1093         BUG_ON(irq < 0);
1094         notify_remote_via_irq(irq);
1095 }
1096
1097 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
1098 {
1099         struct shared_info *sh = HYPERVISOR_shared_info;
1100         int cpu = smp_processor_id();
1101         unsigned long *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu);
1102         int i;
1103         unsigned long flags;
1104         static DEFINE_SPINLOCK(debug_lock);
1105         struct vcpu_info *v;
1106
1107         spin_lock_irqsave(&debug_lock, flags);
1108
1109         printk("\nvcpu %d\n  ", cpu);
1110
1111         for_each_online_cpu(i) {
1112                 int pending;
1113                 v = per_cpu(xen_vcpu, i);
1114                 pending = (get_irq_regs() && i == cpu)
1115                         ? xen_irqs_disabled(get_irq_regs())
1116                         : v->evtchn_upcall_mask;
1117                 printk("%d: masked=%d pending=%d event_sel %0*lx\n  ", i,
1118                        pending, v->evtchn_upcall_pending,
1119                        (int)(sizeof(v->evtchn_pending_sel)*2),
1120                        v->evtchn_pending_sel);
1121         }
1122         v = per_cpu(xen_vcpu, cpu);
1123
1124         printk("\npending:\n   ");
1125         for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
1126                 printk("%0*lx%s", (int)sizeof(sh->evtchn_pending[0])*2,
1127                        sh->evtchn_pending[i],
1128                        i % 8 == 0 ? "\n   " : " ");
1129         printk("\nglobal mask:\n   ");
1130         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1131                 printk("%0*lx%s",
1132                        (int)(sizeof(sh->evtchn_mask[0])*2),
1133                        sh->evtchn_mask[i],
1134                        i % 8 == 0 ? "\n   " : " ");
1135
1136         printk("\nglobally unmasked:\n   ");
1137         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1138                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1139                        sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
1140                        i % 8 == 0 ? "\n   " : " ");
1141
1142         printk("\nlocal cpu%d mask:\n   ", cpu);
1143         for (i = (NR_EVENT_CHANNELS/BITS_PER_LONG)-1; i >= 0; i--)
1144                 printk("%0*lx%s", (int)(sizeof(cpu_evtchn[0])*2),
1145                        cpu_evtchn[i],
1146                        i % 8 == 0 ? "\n   " : " ");
1147
1148         printk("\nlocally unmasked:\n   ");
1149         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
1150                 unsigned long pending = sh->evtchn_pending[i]
1151                         & ~sh->evtchn_mask[i]
1152                         & cpu_evtchn[i];
1153                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1154                        pending, i % 8 == 0 ? "\n   " : " ");
1155         }
1156
1157         printk("\npending list:\n");
1158         for (i = 0; i < NR_EVENT_CHANNELS; i++) {
1159                 if (sync_test_bit(i, sh->evtchn_pending)) {
1160                         int word_idx = i / BITS_PER_LONG;
1161                         printk("  %d: event %d -> irq %d%s%s%s\n",
1162                                cpu_from_evtchn(i), i,
1163                                evtchn_to_irq[i],
1164                                sync_test_bit(word_idx, &v->evtchn_pending_sel)
1165                                              ? "" : " l2-clear",
1166                                !sync_test_bit(i, sh->evtchn_mask)
1167                                              ? "" : " globally-masked",
1168                                sync_test_bit(i, cpu_evtchn)
1169                                              ? "" : " locally-masked");
1170                 }
1171         }
1172
1173         spin_unlock_irqrestore(&debug_lock, flags);
1174
1175         return IRQ_HANDLED;
1176 }
1177
1178 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
1179 static DEFINE_PER_CPU(unsigned int, current_word_idx);
1180 static DEFINE_PER_CPU(unsigned int, current_bit_idx);
1181
1182 /*
1183  * Mask out the i least significant bits of w
1184  */
1185 #define MASK_LSBS(w, i) (w & ((~0UL) << i))
1186
1187 /*
1188  * Search the CPUs pending events bitmasks.  For each one found, map
1189  * the event number to an irq, and feed it into do_IRQ() for
1190  * handling.
1191  *
1192  * Xen uses a two-level bitmap to speed searching.  The first level is
1193  * a bitset of words which contain pending event bits.  The second
1194  * level is a bitset of pending events themselves.
1195  */
1196 static void __xen_evtchn_do_upcall(void)
1197 {
1198         int start_word_idx, start_bit_idx;
1199         int word_idx, bit_idx;
1200         int i, irq;
1201         int cpu = get_cpu();
1202         struct shared_info *s = HYPERVISOR_shared_info;
1203         struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1204         unsigned count;
1205
1206         do {
1207                 unsigned long pending_words;
1208                 unsigned long pending_bits;
1209                 struct irq_desc *desc;
1210
1211                 vcpu_info->evtchn_upcall_pending = 0;
1212
1213                 if (__this_cpu_inc_return(xed_nesting_count) - 1)
1214                         goto out;
1215
1216 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
1217                 /* Clear master flag /before/ clearing selector flag. */
1218                 wmb();
1219 #endif
1220                 if ((irq = per_cpu(virq_to_irq, cpu)[VIRQ_TIMER]) != -1) {
1221                         int evtchn = evtchn_from_irq(irq);
1222                         word_idx = evtchn / BITS_PER_LONG;
1223                         pending_bits = evtchn % BITS_PER_LONG;
1224                         if (active_evtchns(cpu, s, word_idx) & (1ULL << pending_bits)) {
1225                                 desc = irq_to_desc(irq);
1226                                 if (desc)
1227                                         generic_handle_irq_desc(irq, desc);
1228                         }
1229                 }
1230
1231                 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
1232
1233                 start_word_idx = __this_cpu_read(current_word_idx);
1234                 start_bit_idx = __this_cpu_read(current_bit_idx);
1235
1236                 word_idx = start_word_idx;
1237
1238                 for (i = 0; pending_words != 0; i++) {
1239                         unsigned long words;
1240
1241                         words = MASK_LSBS(pending_words, word_idx);
1242
1243                         /*
1244                          * If we masked out all events, wrap to beginning.
1245                          */
1246                         if (words == 0) {
1247                                 word_idx = 0;
1248                                 bit_idx = 0;
1249                                 continue;
1250                         }
1251                         word_idx = __ffs(words);
1252
1253                         pending_bits = active_evtchns(cpu, s, word_idx);
1254                         bit_idx = 0; /* usually scan entire word from start */
1255                         if (word_idx == start_word_idx) {
1256                                 /* We scan the starting word in two parts */
1257                                 if (i == 0)
1258                                         /* 1st time: start in the middle */
1259                                         bit_idx = start_bit_idx;
1260                                 else
1261                                         /* 2nd time: mask bits done already */
1262                                         bit_idx &= (1UL << start_bit_idx) - 1;
1263                         }
1264
1265                         do {
1266                                 unsigned long bits;
1267                                 int port;
1268
1269                                 bits = MASK_LSBS(pending_bits, bit_idx);
1270
1271                                 /* If we masked out all events, move on. */
1272                                 if (bits == 0)
1273                                         break;
1274
1275                                 bit_idx = __ffs(bits);
1276
1277                                 /* Process port. */
1278                                 port = (word_idx * BITS_PER_LONG) + bit_idx;
1279                                 irq = evtchn_to_irq[port];
1280
1281                                 if (irq != -1) {
1282                                         desc = irq_to_desc(irq);
1283                                         if (desc)
1284                                                 generic_handle_irq_desc(irq, desc);
1285                                 }
1286
1287                                 bit_idx = (bit_idx + 1) % BITS_PER_LONG;
1288
1289                                 /* Next caller starts at last processed + 1 */
1290                                 __this_cpu_write(current_word_idx,
1291                                                  bit_idx ? word_idx :
1292                                                  (word_idx+1) % BITS_PER_LONG);
1293                                 __this_cpu_write(current_bit_idx, bit_idx);
1294                         } while (bit_idx != 0);
1295
1296                         /* Scan start_l1i twice; all others once. */
1297                         if ((word_idx != start_word_idx) || (i != 0))
1298                                 pending_words &= ~(1UL << word_idx);
1299
1300                         word_idx = (word_idx + 1) % BITS_PER_LONG;
1301                 }
1302
1303                 BUG_ON(!irqs_disabled());
1304
1305                 count = __this_cpu_read(xed_nesting_count);
1306                 __this_cpu_write(xed_nesting_count, 0);
1307         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
1308
1309 out:
1310
1311         put_cpu();
1312 }
1313
1314 void xen_evtchn_do_upcall(struct pt_regs *regs)
1315 {
1316         struct pt_regs *old_regs = set_irq_regs(regs);
1317
1318         exit_idle();
1319         irq_enter();
1320
1321         __xen_evtchn_do_upcall();
1322
1323         irq_exit();
1324         set_irq_regs(old_regs);
1325 }
1326
1327 void xen_hvm_evtchn_do_upcall(void)
1328 {
1329         __xen_evtchn_do_upcall();
1330 }
1331 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1332
1333 /* Rebind a new event channel to an existing irq. */
1334 void rebind_evtchn_irq(int evtchn, int irq)
1335 {
1336         struct irq_info *info = info_for_irq(irq);
1337
1338         /* Make sure the irq is masked, since the new event channel
1339            will also be masked. */
1340         disable_irq(irq);
1341
1342         mutex_lock(&irq_mapping_update_lock);
1343
1344         /* After resume the irq<->evtchn mappings are all cleared out */
1345         BUG_ON(evtchn_to_irq[evtchn] != -1);
1346         /* Expect irq to have been bound before,
1347            so there should be a proper type */
1348         BUG_ON(info->type == IRQT_UNBOUND);
1349
1350         xen_irq_info_evtchn_init(irq, evtchn);
1351
1352         mutex_unlock(&irq_mapping_update_lock);
1353
1354         /* new event channels are always bound to cpu 0 */
1355         irq_set_affinity(irq, cpumask_of(0));
1356
1357         /* Unmask the event channel. */
1358         enable_irq(irq);
1359 }
1360
1361 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1362 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1363 {
1364         struct shared_info *s = HYPERVISOR_shared_info;
1365         struct evtchn_bind_vcpu bind_vcpu;
1366         int evtchn = evtchn_from_irq(irq);
1367         int masked;
1368
1369         if (!VALID_EVTCHN(evtchn))
1370                 return -1;
1371
1372         /*
1373          * Events delivered via platform PCI interrupts are always
1374          * routed to vcpu 0 and hence cannot be rebound.
1375          */
1376         if (xen_hvm_domain() && !xen_have_vector_callback)
1377                 return -1;
1378
1379         /* Send future instances of this interrupt to other vcpu. */
1380         bind_vcpu.port = evtchn;
1381         bind_vcpu.vcpu = tcpu;
1382
1383         /*
1384          * Mask the event while changing the VCPU binding to prevent
1385          * it being delivered on an unexpected VCPU.
1386          */
1387         masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
1388
1389         /*
1390          * If this fails, it usually just indicates that we're dealing with a
1391          * virq or IPI channel, which don't actually need to be rebound. Ignore
1392          * it, but don't do the xenlinux-level rebind in that case.
1393          */
1394         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1395                 bind_evtchn_to_cpu(evtchn, tcpu);
1396
1397         if (!masked)
1398                 unmask_evtchn(evtchn);
1399
1400         return 0;
1401 }
1402
1403 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1404                             bool force)
1405 {
1406         unsigned tcpu = cpumask_first(dest);
1407
1408         return rebind_irq_to_cpu(data->irq, tcpu);
1409 }
1410
1411 int resend_irq_on_evtchn(unsigned int irq)
1412 {
1413         int masked, evtchn = evtchn_from_irq(irq);
1414         struct shared_info *s = HYPERVISOR_shared_info;
1415
1416         if (!VALID_EVTCHN(evtchn))
1417                 return 1;
1418
1419         masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
1420         sync_set_bit(evtchn, s->evtchn_pending);
1421         if (!masked)
1422                 unmask_evtchn(evtchn);
1423
1424         return 1;
1425 }
1426
1427 static void enable_dynirq(struct irq_data *data)
1428 {
1429         int evtchn = evtchn_from_irq(data->irq);
1430
1431         if (VALID_EVTCHN(evtchn))
1432                 unmask_evtchn(evtchn);
1433 }
1434
1435 static void disable_dynirq(struct irq_data *data)
1436 {
1437         int evtchn = evtchn_from_irq(data->irq);
1438
1439         if (VALID_EVTCHN(evtchn))
1440                 mask_evtchn(evtchn);
1441 }
1442
1443 static void ack_dynirq(struct irq_data *data)
1444 {
1445         int evtchn = evtchn_from_irq(data->irq);
1446
1447         if (!VALID_EVTCHN(evtchn))
1448                 return;
1449
1450         if (unlikely(irqd_is_setaffinity_pending(data)) &&
1451             likely(!irqd_irq_disabled(data))) {
1452                 int masked = test_and_set_mask(evtchn);
1453
1454                 clear_evtchn(evtchn);
1455
1456                 irq_move_masked_irq(data);
1457
1458                 if (!masked)
1459                         unmask_evtchn(evtchn);
1460         } else
1461                 clear_evtchn(evtchn);
1462 }
1463
1464 static void mask_ack_dynirq(struct irq_data *data)
1465 {
1466         disable_dynirq(data);
1467         ack_dynirq(data);
1468 }
1469
1470 static int retrigger_dynirq(struct irq_data *data)
1471 {
1472         int evtchn = evtchn_from_irq(data->irq);
1473         struct shared_info *sh = HYPERVISOR_shared_info;
1474         int ret = 0;
1475
1476         if (VALID_EVTCHN(evtchn)) {
1477                 int masked;
1478
1479                 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
1480                 sync_set_bit(evtchn, sh->evtchn_pending);
1481                 if (!masked)
1482                         unmask_evtchn(evtchn);
1483                 ret = 1;
1484         }
1485
1486         return ret;
1487 }
1488
1489 static void restore_pirqs(void)
1490 {
1491         int pirq, rc, irq, gsi;
1492         struct physdev_map_pirq map_irq;
1493         struct irq_info *info;
1494
1495         list_for_each_entry(info, &xen_irq_list_head, list) {
1496                 if (info->type != IRQT_PIRQ)
1497                         continue;
1498
1499                 pirq = info->u.pirq.pirq;
1500                 gsi = info->u.pirq.gsi;
1501                 irq = info->irq;
1502
1503                 /* save/restore of PT devices doesn't work, so at this point the
1504                  * only devices present are GSI based emulated devices */
1505                 if (!gsi)
1506                         continue;
1507
1508                 map_irq.domid = DOMID_SELF;
1509                 map_irq.type = MAP_PIRQ_TYPE_GSI;
1510                 map_irq.index = gsi;
1511                 map_irq.pirq = pirq;
1512
1513                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1514                 if (rc) {
1515                         printk(KERN_WARNING "xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1516                                         gsi, irq, pirq, rc);
1517                         xen_free_irq(irq);
1518                         continue;
1519                 }
1520
1521                 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1522
1523                 __startup_pirq(irq);
1524         }
1525 }
1526
1527 static void restore_cpu_virqs(unsigned int cpu)
1528 {
1529         struct evtchn_bind_virq bind_virq;
1530         int virq, irq, evtchn;
1531
1532         for (virq = 0; virq < NR_VIRQS; virq++) {
1533                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1534                         continue;
1535
1536                 BUG_ON(virq_from_irq(irq) != virq);
1537
1538                 /* Get a new binding from Xen. */
1539                 bind_virq.virq = virq;
1540                 bind_virq.vcpu = cpu;
1541                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1542                                                 &bind_virq) != 0)
1543                         BUG();
1544                 evtchn = bind_virq.port;
1545
1546                 /* Record the new mapping. */
1547                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
1548                 bind_evtchn_to_cpu(evtchn, cpu);
1549         }
1550 }
1551
1552 static void restore_cpu_ipis(unsigned int cpu)
1553 {
1554         struct evtchn_bind_ipi bind_ipi;
1555         int ipi, irq, evtchn;
1556
1557         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1558                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1559                         continue;
1560
1561                 BUG_ON(ipi_from_irq(irq) != ipi);
1562
1563                 /* Get a new binding from Xen. */
1564                 bind_ipi.vcpu = cpu;
1565                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1566                                                 &bind_ipi) != 0)
1567                         BUG();
1568                 evtchn = bind_ipi.port;
1569
1570                 /* Record the new mapping. */
1571                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
1572                 bind_evtchn_to_cpu(evtchn, cpu);
1573         }
1574 }
1575
1576 /* Clear an irq's pending state, in preparation for polling on it */
1577 void xen_clear_irq_pending(int irq)
1578 {
1579         int evtchn = evtchn_from_irq(irq);
1580
1581         if (VALID_EVTCHN(evtchn))
1582                 clear_evtchn(evtchn);
1583 }
1584 EXPORT_SYMBOL(xen_clear_irq_pending);
1585 void xen_set_irq_pending(int irq)
1586 {
1587         int evtchn = evtchn_from_irq(irq);
1588
1589         if (VALID_EVTCHN(evtchn))
1590                 set_evtchn(evtchn);
1591 }
1592
1593 bool xen_test_irq_pending(int irq)
1594 {
1595         int evtchn = evtchn_from_irq(irq);
1596         bool ret = false;
1597
1598         if (VALID_EVTCHN(evtchn))
1599                 ret = test_evtchn(evtchn);
1600
1601         return ret;
1602 }
1603
1604 /* Poll waiting for an irq to become pending with timeout.  In the usual case,
1605  * the irq will be disabled so it won't deliver an interrupt. */
1606 void xen_poll_irq_timeout(int irq, u64 timeout)
1607 {
1608         evtchn_port_t evtchn = evtchn_from_irq(irq);
1609
1610         if (VALID_EVTCHN(evtchn)) {
1611                 struct sched_poll poll;
1612
1613                 poll.nr_ports = 1;
1614                 poll.timeout = timeout;
1615                 set_xen_guest_handle(poll.ports, &evtchn);
1616
1617                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1618                         BUG();
1619         }
1620 }
1621 EXPORT_SYMBOL(xen_poll_irq_timeout);
1622 /* Poll waiting for an irq to become pending.  In the usual case, the
1623  * irq will be disabled so it won't deliver an interrupt. */
1624 void xen_poll_irq(int irq)
1625 {
1626         xen_poll_irq_timeout(irq, 0 /* no timeout */);
1627 }
1628
1629 /* Check whether the IRQ line is shared with other guests. */
1630 int xen_test_irq_shared(int irq)
1631 {
1632         struct irq_info *info = info_for_irq(irq);
1633         struct physdev_irq_status_query irq_status = { .irq = info->u.pirq.pirq };
1634
1635         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
1636                 return 0;
1637         return !(irq_status.flags & XENIRQSTAT_shared);
1638 }
1639 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
1640
1641 void xen_irq_resume(void)
1642 {
1643         unsigned int cpu, evtchn;
1644         struct irq_info *info;
1645
1646         init_evtchn_cpu_bindings();
1647
1648         /* New event-channel space is not 'live' yet. */
1649         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1650                 mask_evtchn(evtchn);
1651
1652         /* No IRQ <-> event-channel mappings. */
1653         list_for_each_entry(info, &xen_irq_list_head, list)
1654                 info->evtchn = 0; /* zap event-channel binding */
1655
1656         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1657                 evtchn_to_irq[evtchn] = -1;
1658
1659         for_each_possible_cpu(cpu) {
1660                 restore_cpu_virqs(cpu);
1661                 restore_cpu_ipis(cpu);
1662         }
1663
1664         restore_pirqs();
1665 }
1666
1667 static struct irq_chip xen_dynamic_chip __read_mostly = {
1668         .name                   = "xen-dyn",
1669
1670         .irq_disable            = disable_dynirq,
1671         .irq_mask               = disable_dynirq,
1672         .irq_unmask             = enable_dynirq,
1673
1674         .irq_ack                = ack_dynirq,
1675         .irq_mask_ack           = mask_ack_dynirq,
1676
1677         .irq_set_affinity       = set_affinity_irq,
1678         .irq_retrigger          = retrigger_dynirq,
1679 };
1680
1681 static struct irq_chip xen_pirq_chip __read_mostly = {
1682         .name                   = "xen-pirq",
1683
1684         .irq_startup            = startup_pirq,
1685         .irq_shutdown           = shutdown_pirq,
1686         .irq_enable             = enable_pirq,
1687         .irq_disable            = disable_pirq,
1688
1689         .irq_mask               = disable_dynirq,
1690         .irq_unmask             = enable_dynirq,
1691
1692         .irq_ack                = eoi_pirq,
1693         .irq_eoi                = eoi_pirq,
1694         .irq_mask_ack           = mask_ack_pirq,
1695
1696         .irq_set_affinity       = set_affinity_irq,
1697
1698         .irq_retrigger          = retrigger_dynirq,
1699 };
1700
1701 static struct irq_chip xen_percpu_chip __read_mostly = {
1702         .name                   = "xen-percpu",
1703
1704         .irq_disable            = disable_dynirq,
1705         .irq_mask               = disable_dynirq,
1706         .irq_unmask             = enable_dynirq,
1707
1708         .irq_ack                = ack_dynirq,
1709 };
1710
1711 int xen_set_callback_via(uint64_t via)
1712 {
1713         struct xen_hvm_param a;
1714         a.domid = DOMID_SELF;
1715         a.index = HVM_PARAM_CALLBACK_IRQ;
1716         a.value = via;
1717         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1718 }
1719 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1720
1721 #ifdef CONFIG_XEN_PVHVM
1722 /* Vector callbacks are better than PCI interrupts to receive event
1723  * channel notifications because we can receive vector callbacks on any
1724  * vcpu and we don't need PCI support or APIC interactions. */
1725 void xen_callback_vector(void)
1726 {
1727         int rc;
1728         uint64_t callback_via;
1729         if (xen_have_vector_callback) {
1730                 callback_via = HVM_CALLBACK_VECTOR(XEN_HVM_EVTCHN_CALLBACK);
1731                 rc = xen_set_callback_via(callback_via);
1732                 if (rc) {
1733                         printk(KERN_ERR "Request for Xen HVM callback vector"
1734                                         " failed.\n");
1735                         xen_have_vector_callback = 0;
1736                         return;
1737                 }
1738                 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1739                                 "enabled\n");
1740                 /* in the restore case the vector has already been allocated */
1741                 if (!test_bit(XEN_HVM_EVTCHN_CALLBACK, used_vectors))
1742                         alloc_intr_gate(XEN_HVM_EVTCHN_CALLBACK, xen_hvm_callback_vector);
1743         }
1744 }
1745 #else
1746 void xen_callback_vector(void) {}
1747 #endif
1748
1749 void __init xen_init_IRQ(void)
1750 {
1751         int i;
1752
1753         evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1754                                     GFP_KERNEL);
1755         BUG_ON(!evtchn_to_irq);
1756         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1757                 evtchn_to_irq[i] = -1;
1758
1759         init_evtchn_cpu_bindings();
1760
1761         /* No event channels are 'live' right now. */
1762         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1763                 mask_evtchn(i);
1764
1765         if (xen_hvm_domain()) {
1766                 xen_callback_vector();
1767                 native_init_IRQ();
1768                 /* pci_xen_hvm_init must be called after native_init_IRQ so that
1769                  * __acpi_register_gsi can point at the right function */
1770                 pci_xen_hvm_init();
1771         } else {
1772                 irq_ctx_init(smp_processor_id());
1773                 if (xen_initial_domain())
1774                         pci_xen_initial_domain();
1775         }
1776 }