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 static unsigned long *pirq_eoi_map;
 119 static bool (*pirq_needs_eoi)(unsigned irq);
 120
 121 static DEFINE_PER_CPU(unsigned long [NR_EVENT_CHANNELS/BITS_PER_LONG],
 122                       cpu_evtchn_mask);
 123
 124 /* Xen will never allocate port zero for any purpose. */
 125 #define VALID_EVTCHN(chn)       ((chn) != 0)
 126
 127 static struct irq_chip xen_dynamic_chip;
 128 static struct irq_chip xen_percpu_chip;
 129 static struct irq_chip xen_pirq_chip;
 130 static void enable_dynirq(struct irq_data *data);
 131 static void disable_dynirq(struct irq_data *data);
 132
 133 /* Get info for IRQ */
 134 static struct irq_info *info_for_irq(unsigned irq)
 135 {
 136         return irq_get_handler_data(irq);
 137 }
 138
 139 /* Constructors for packed IRQ information. */
 140 static void xen_irq_info_common_init(struct irq_info *info,
 141                                      unsigned irq,
 142                                      enum xen_irq_type type,
 143                                      unsigned short evtchn,
 144                                      unsigned short cpu)
 145 {
 146
 147         BUG_ON(info->type != IRQT_UNBOUND && info->type != type);
 148
 149         info->type = type;
 150         info->irq = irq;
 151         info->evtchn = evtchn;
 152         info->cpu = cpu;
 153
 154         evtchn_to_irq[evtchn] = irq;
 155 }
 156
 157 static void xen_irq_info_evtchn_init(unsigned irq,
 158                                      unsigned short evtchn)
 159 {
 160         struct irq_info *info = info_for_irq(irq);
 161
 162         xen_irq_info_common_init(info, irq, IRQT_EVTCHN, evtchn, 0);
 163 }
 164
 165 static void xen_irq_info_ipi_init(unsigned cpu,
 166                                   unsigned irq,
 167                                   unsigned short evtchn,
 168                                   enum ipi_vector ipi)
 169 {
 170         struct irq_info *info = info_for_irq(irq);
 171
 172         xen_irq_info_common_init(info, irq, IRQT_IPI, evtchn, 0);
 173
 174         info->u.ipi = ipi;
 175
 176         per_cpu(ipi_to_irq, cpu)[ipi] = irq;
 177 }
 178
 179 static void xen_irq_info_virq_init(unsigned cpu,
 180                                    unsigned irq,
 181                                    unsigned short evtchn,
 182                                    unsigned short virq)
 183 {
 184         struct irq_info *info = info_for_irq(irq);
 185
 186         xen_irq_info_common_init(info, irq, IRQT_VIRQ, evtchn, 0);
 187
 188         info->u.virq = virq;
 189
 190         per_cpu(virq_to_irq, cpu)[virq] = irq;
 191 }
 192
 193 static void xen_irq_info_pirq_init(unsigned irq,
 194                                    unsigned short evtchn,
 195                                    unsigned short pirq,
 196                                    unsigned short gsi,
 197                                    unsigned short vector,
 198                                    uint16_t domid,
 199                                    unsigned char flags)
 200 {
 201         struct irq_info *info = info_for_irq(irq);
 202
 203         xen_irq_info_common_init(info, irq, IRQT_PIRQ, evtchn, 0);
 204
 205         info->u.pirq.pirq = pirq;
 206         info->u.pirq.gsi = gsi;
 207         info->u.pirq.vector = vector;
 208         info->u.pirq.domid = domid;
 209         info->u.pirq.flags = flags;
 210 }
 211
 212 /*
 213  * Accessors for packed IRQ information.
 214  */
 215 static unsigned int evtchn_from_irq(unsigned irq)
 216 {
 217         if (unlikely(WARN(irq < 0 || irq >= nr_irqs, "Invalid irq %d!\n", irq)))
 218                 return 0;
 219
 220         return info_for_irq(irq)->evtchn;
 221 }
 222
 223 unsigned irq_from_evtchn(unsigned int evtchn)
 224 {
 225         return evtchn_to_irq[evtchn];
 226 }
 227 EXPORT_SYMBOL_GPL(irq_from_evtchn);
 228
 229 static enum ipi_vector ipi_from_irq(unsigned irq)
 230 {
 231         struct irq_info *info = info_for_irq(irq);
 232
 233         BUG_ON(info == NULL);
 234         BUG_ON(info->type != IRQT_IPI);
 235
 236         return info->u.ipi;
 237 }
 238
 239 static unsigned virq_from_irq(unsigned irq)
 240 {
 241         struct irq_info *info = info_for_irq(irq);
 242
 243         BUG_ON(info == NULL);
 244         BUG_ON(info->type != IRQT_VIRQ);
 245
 246         return info->u.virq;
 247 }
 248
 249 static unsigned pirq_from_irq(unsigned irq)
 250 {
 251         struct irq_info *info = info_for_irq(irq);
 252
 253         BUG_ON(info == NULL);
 254         BUG_ON(info->type != IRQT_PIRQ);
 255
 256         return info->u.pirq.pirq;
 257 }
 258
 259 static enum xen_irq_type type_from_irq(unsigned irq)
 260 {
 261         return info_for_irq(irq)->type;
 262 }
 263
 264 static unsigned cpu_from_irq(unsigned irq)
 265 {
 266         return info_for_irq(irq)->cpu;
 267 }
 268
 269 static unsigned int cpu_from_evtchn(unsigned int evtchn)
 270 {
 271         int irq = evtchn_to_irq[evtchn];
 272         unsigned ret = 0;
 273
 274         if (irq != -1)
 275                 ret = cpu_from_irq(irq);
 276
 277         return ret;
 278 }
 279
 280 static bool pirq_check_eoi_map(unsigned irq)
 281 {
 282         return test_bit(pirq_from_irq(irq), pirq_eoi_map);
 283 }
 284
 285 static bool pirq_needs_eoi_flag(unsigned irq)
 286 {
 287         struct irq_info *info = info_for_irq(irq);
 288         BUG_ON(info->type != IRQT_PIRQ);
 289
 290         return info->u.pirq.flags & PIRQ_NEEDS_EOI;
 291 }
 292
 293 static inline unsigned long active_evtchns(unsigned int cpu,
 294                                            struct shared_info *sh,
 295                                            unsigned int idx)
 296 {
 297         return sh->evtchn_pending[idx] &
 298                 per_cpu(cpu_evtchn_mask, cpu)[idx] &
 299                 ~sh->evtchn_mask[idx];
 300 }
 301
 302 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
 303 {
 304         int irq = evtchn_to_irq[chn];
 305
 306         BUG_ON(irq == -1);
 307 #ifdef CONFIG_SMP
 308         cpumask_copy(irq_to_desc(irq)->irq_data.affinity, cpumask_of(cpu));
 309 #endif
 310
 311         clear_bit(chn, per_cpu(cpu_evtchn_mask, cpu_from_irq(irq)));
 312         set_bit(chn, per_cpu(cpu_evtchn_mask, cpu));
 313
 314         info_for_irq(irq)->cpu = cpu;
 315 }
 316
 317 static void init_evtchn_cpu_bindings(void)
 318 {
 319         int i;
 320 #ifdef CONFIG_SMP
 321         struct irq_info *info;
 322
 323         /* By default all event channels notify CPU#0. */
 324         list_for_each_entry(info, &xen_irq_list_head, list) {
 325                 struct irq_desc *desc = irq_to_desc(info->irq);
 326                 cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 327         }
 328 #endif
 329
 330         for_each_possible_cpu(i)
 331                 memset(per_cpu(cpu_evtchn_mask, i),
 332                        (i == 0) ? ~0 : 0, sizeof(*per_cpu(cpu_evtchn_mask, i)));
 333 }
 334
 335 static inline void clear_evtchn(int port)
 336 {
 337         struct shared_info *s = HYPERVISOR_shared_info;
 338         sync_clear_bit(port, &s->evtchn_pending[0]);
 339 }
 340
 341 static inline void set_evtchn(int port)
 342 {
 343         struct shared_info *s = HYPERVISOR_shared_info;
 344         sync_set_bit(port, &s->evtchn_pending[0]);
 345 }
 346
 347 static inline int test_evtchn(int port)
 348 {
 349         struct shared_info *s = HYPERVISOR_shared_info;
 350         return sync_test_bit(port, &s->evtchn_pending[0]);
 351 }
 352
 353
 354 /**
 355  * notify_remote_via_irq - send event to remote end of event channel via irq
 356  * @irq: irq of event channel to send event to
 357  *
 358  * Unlike notify_remote_via_evtchn(), this is safe to use across
 359  * save/restore. Notifications on a broken connection are silently
 360  * dropped.
 361  */
 362 void notify_remote_via_irq(int irq)
 363 {
 364         int evtchn = evtchn_from_irq(irq);
 365
 366         if (VALID_EVTCHN(evtchn))
 367                 notify_remote_via_evtchn(evtchn);
 368 }
 369 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
 370
 371 static void mask_evtchn(int port)
 372 {
 373         struct shared_info *s = HYPERVISOR_shared_info;
 374         sync_set_bit(port, &s->evtchn_mask[0]);
 375 }
 376
 377 static void unmask_evtchn(int port)
 378 {
 379         struct shared_info *s = HYPERVISOR_shared_info;
 380         unsigned int cpu = get_cpu();
 381         int do_hypercall = 0, evtchn_pending = 0;
 382
 383         BUG_ON(!irqs_disabled());
 384
 385         if (unlikely((cpu != cpu_from_evtchn(port))))
 386                 do_hypercall = 1;
 387         else
 388                 evtchn_pending = sync_test_bit(port, &s->evtchn_pending[0]);
 389
 390         if (unlikely(evtchn_pending && xen_hvm_domain()))
 391                 do_hypercall = 1;
 392
 393         /* Slow path (hypercall) if this is a non-local port or if this is
 394          * an hvm domain and an event is pending (hvm domains don't have
 395          * their own implementation of irq_enable). */
 396         if (do_hypercall) {
 397                 struct evtchn_unmask unmask = { .port = port };
 398                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 399         } else {
 400                 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
 401
 402                 sync_clear_bit(port, &s->evtchn_mask[0]);
 403
 404                 /*
 405                  * The following is basically the equivalent of
 406                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
 407                  * the interrupt edge' if the channel is masked.
 408                  */
 409                 if (evtchn_pending &&
 410                     !sync_test_and_set_bit(port / BITS_PER_LONG,
 411                                            &vcpu_info->evtchn_pending_sel))
 412                         vcpu_info->evtchn_upcall_pending = 1;
 413         }
 414
 415         put_cpu();
 416 }
 417
 418 static void xen_irq_init(unsigned irq)
 419 {
 420         struct irq_info *info;
 421 #ifdef CONFIG_SMP
 422         struct irq_desc *desc = irq_to_desc(irq);
 423
 424         /* By default all event channels notify CPU#0. */
 425         cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 426 #endif
 427
 428         info = kzalloc(sizeof(*info), GFP_KERNEL);
 429         if (info == NULL)
 430                 panic("Unable to allocate metadata for IRQ%d\n", irq);
 431
 432         info->type = IRQT_UNBOUND;
 433         info->refcnt = -1;
 434
 435         irq_set_handler_data(irq, info);
 436
 437         list_add_tail(&info->list, &xen_irq_list_head);
 438 }
 439
 440 static int __must_check xen_allocate_irq_dynamic(void)
 441 {
 442         int first = 0;
 443         int irq;
 444
 445 #ifdef CONFIG_X86_IO_APIC
 446         /*
 447          * For an HVM guest or domain 0 which see "real" (emulated or
 448          * actual respectively) GSIs we allocate dynamic IRQs
 449          * e.g. those corresponding to event channels or MSIs
 450          * etc. from the range above those "real" GSIs to avoid
 451          * collisions.
 452          */
 453         if (xen_initial_domain() || xen_hvm_domain())
 454                 first = get_nr_irqs_gsi();
 455 #endif
 456
 457         irq = irq_alloc_desc_from(first, -1);
 458
 459         if (irq >= 0)
 460                 xen_irq_init(irq);
 461
 462         return irq;
 463 }
 464
 465 static int __must_check xen_allocate_irq_gsi(unsigned gsi)
 466 {
 467         int irq;
 468
 469         /*
 470          * A PV guest has no concept of a GSI (since it has no ACPI
 471          * nor access to/knowledge of the physical APICs). Therefore
 472          * all IRQs are dynamically allocated from the entire IRQ
 473          * space.
 474          */
 475         if (xen_pv_domain() && !xen_initial_domain())
 476                 return xen_allocate_irq_dynamic();
 477
 478         /* Legacy IRQ descriptors are already allocated by the arch. */
 479         if (gsi < NR_IRQS_LEGACY)
 480                 irq = gsi;
 481         else
 482                 irq = irq_alloc_desc_at(gsi, -1);
 483
 484         xen_irq_init(irq);
 485
 486         return irq;
 487 }
 488
 489 static void xen_free_irq(unsigned irq)
 490 {
 491         struct irq_info *info = irq_get_handler_data(irq);
 492
 493         list_del(&info->list);
 494
 495         irq_set_handler_data(irq, NULL);
 496
 497         WARN_ON(info->refcnt > 0);
 498
 499         kfree(info);
 500
 501         /* Legacy IRQ descriptors are managed by the arch. */
 502         if (irq < NR_IRQS_LEGACY)
 503                 return;
 504
 505         irq_free_desc(irq);
 506 }
 507
 508 static void pirq_query_unmask(int irq)
 509 {
 510         struct physdev_irq_status_query irq_status;
 511         struct irq_info *info = info_for_irq(irq);
 512
 513         BUG_ON(info->type != IRQT_PIRQ);
 514
 515         irq_status.irq = pirq_from_irq(irq);
 516         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
 517                 irq_status.flags = 0;
 518
 519         info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
 520         if (irq_status.flags & XENIRQSTAT_needs_eoi)
 521                 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
 522 }
 523
 524 static bool probing_irq(int irq)
 525 {
 526         struct irq_desc *desc = irq_to_desc(irq);
 527
 528         return desc && desc->action == NULL;
 529 }
 530
 531 static void eoi_pirq(struct irq_data *data)
 532 {
 533         int evtchn = evtchn_from_irq(data->irq);
 534         struct physdev_eoi eoi = { .irq = pirq_from_irq(data->irq) };
 535         int rc = 0;
 536
 537         irq_move_irq(data);
 538
 539         if (VALID_EVTCHN(evtchn))
 540                 clear_evtchn(evtchn);
 541
 542         if (pirq_needs_eoi(data->irq)) {
 543                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
 544                 WARN_ON(rc);
 545         }
 546 }
 547
 548 static void mask_ack_pirq(struct irq_data *data)
 549 {
 550         disable_dynirq(data);
 551         eoi_pirq(data);
 552 }
 553
 554 static unsigned int __startup_pirq(unsigned int irq)
 555 {
 556         struct evtchn_bind_pirq bind_pirq;
 557         struct irq_info *info = info_for_irq(irq);
 558         int evtchn = evtchn_from_irq(irq);
 559         int rc;
 560
 561         BUG_ON(info->type != IRQT_PIRQ);
 562
 563         if (VALID_EVTCHN(evtchn))
 564                 goto out;
 565
 566         bind_pirq.pirq = pirq_from_irq(irq);
 567         /* NB. We are happy to share unless we are probing. */
 568         bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
 569                                         BIND_PIRQ__WILL_SHARE : 0;
 570         rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
 571         if (rc != 0) {
 572                 if (!probing_irq(irq))
 573                         printk(KERN_INFO "Failed to obtain physical IRQ %d\n",
 574                                irq);
 575                 return 0;
 576         }
 577         evtchn = bind_pirq.port;
 578
 579         pirq_query_unmask(irq);
 580
 581         evtchn_to_irq[evtchn] = irq;
 582         bind_evtchn_to_cpu(evtchn, 0);
 583         info->evtchn = evtchn;
 584
 585 out:
 586         unmask_evtchn(evtchn);
 587         eoi_pirq(irq_get_irq_data(irq));
 588
 589         return 0;
 590 }
 591
 592 static unsigned int startup_pirq(struct irq_data *data)
 593 {
 594         return __startup_pirq(data->irq);
 595 }
 596
 597 static void shutdown_pirq(struct irq_data *data)
 598 {
 599         struct evtchn_close close;
 600         unsigned int irq = data->irq;
 601         struct irq_info *info = info_for_irq(irq);
 602         int evtchn = evtchn_from_irq(irq);
 603
 604         BUG_ON(info->type != IRQT_PIRQ);
 605
 606         if (!VALID_EVTCHN(evtchn))
 607                 return;
 608
 609         mask_evtchn(evtchn);
 610
 611         close.port = evtchn;
 612         if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 613                 BUG();
 614
 615         bind_evtchn_to_cpu(evtchn, 0);
 616         evtchn_to_irq[evtchn] = -1;
 617         info->evtchn = 0;
 618 }
 619
 620 static void enable_pirq(struct irq_data *data)
 621 {
 622         startup_pirq(data);
 623 }
 624
 625 static void disable_pirq(struct irq_data *data)
 626 {
 627         disable_dynirq(data);
 628 }
 629
 630 int xen_irq_from_gsi(unsigned gsi)
 631 {
 632         struct irq_info *info;
 633
 634         list_for_each_entry(info, &xen_irq_list_head, list) {
 635                 if (info->type != IRQT_PIRQ)
 636                         continue;
 637
 638                 if (info->u.pirq.gsi == gsi)
 639                         return info->irq;
 640         }
 641
 642         return -1;
 643 }
 644 EXPORT_SYMBOL_GPL(xen_irq_from_gsi);
 645
 646 /*
 647  * Do not make any assumptions regarding the relationship between the
 648  * IRQ number returned here and the Xen pirq argument.
 649  *
 650  * Note: We don't assign an event channel until the irq actually started
 651  * up.  Return an existing irq if we've already got one for the gsi.
 652  *
 653  * Shareable implies level triggered, not shareable implies edge
 654  * triggered here.
 655  */
 656 int xen_bind_pirq_gsi_to_irq(unsigned gsi,
 657                              unsigned pirq, int shareable, char *name)
 658 {
 659         int irq = -1;
 660         struct physdev_irq irq_op;
 661
 662         mutex_lock(&irq_mapping_update_lock);
 663
 664         irq = xen_irq_from_gsi(gsi);
 665         if (irq != -1) {
 666                 printk(KERN_INFO "xen_map_pirq_gsi: returning irq %d for gsi %u\n",
 667                        irq, gsi);
 668                 goto out;
 669         }
 670
 671         irq = xen_allocate_irq_gsi(gsi);
 672         if (irq < 0)
 673                 goto out;
 674
 675         irq_op.irq = irq;
 676         irq_op.vector = 0;
 677
 678         /* Only the privileged domain can do this. For non-priv, the pcifront
 679          * driver provides a PCI bus that does the call to do exactly
 680          * this in the priv domain. */
 681         if (xen_initial_domain() &&
 682             HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
 683                 xen_free_irq(irq);
 684                 irq = -ENOSPC;
 685                 goto out;
 686         }
 687
 688         xen_irq_info_pirq_init(irq, 0, pirq, gsi, irq_op.vector, DOMID_SELF,
 689                                shareable ? PIRQ_SHAREABLE : 0);
 690
 691         pirq_query_unmask(irq);
 692         /* We try to use the handler with the appropriate semantic for the
 693          * type of interrupt: if the interrupt is an edge triggered
 694          * interrupt we use handle_edge_irq.
 695          *
 696          * On the other hand if the interrupt is level triggered we use
 697          * handle_fasteoi_irq like the native code does for this kind of
 698          * interrupts.
 699          *
 700          * Depending on the Xen version, pirq_needs_eoi might return true
 701          * not only for level triggered interrupts but for edge triggered
 702          * interrupts too. In any case Xen always honors the eoi mechanism,
 703          * not injecting any more pirqs of the same kind if the first one
 704          * hasn't received an eoi yet. Therefore using the fasteoi handler
 705          * is the right choice either way.
 706          */
 707         if (shareable)
 708                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 709                                 handle_fasteoi_irq, name);
 710         else
 711                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 712                                 handle_edge_irq, name);
 713
 714 out:
 715         mutex_unlock(&irq_mapping_update_lock);
 716
 717         return irq;
 718 }
 719
 720 #ifdef CONFIG_PCI_MSI
 721 int xen_allocate_pirq_msi(struct pci_dev *dev, struct msi_desc *msidesc)
 722 {
 723         int rc;
 724         struct physdev_get_free_pirq op_get_free_pirq;
 725
 726         op_get_free_pirq.type = MAP_PIRQ_TYPE_MSI;
 727         rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
 728
 729         WARN_ONCE(rc == -ENOSYS,
 730                   "hypervisor does not support the PHYSDEVOP_get_free_pirq interface\n");
 731
 732         return rc ? -1 : op_get_free_pirq.pirq;
 733 }
 734
 735 int xen_bind_pirq_msi_to_irq(struct pci_dev *dev, struct msi_desc *msidesc,
 736                              int pirq, int vector, const char *name,
 737                              domid_t domid)
 738 {
 739         int irq, ret;
 740
 741         mutex_lock(&irq_mapping_update_lock);
 742
 743         irq = xen_allocate_irq_dynamic();
 744         if (irq < 0)
 745                 goto out;
 746
 747         irq_set_chip_and_handler_name(irq, &xen_pirq_chip, handle_edge_irq,
 748                         name);
 749
 750         xen_irq_info_pirq_init(irq, 0, pirq, 0, vector, domid, 0);
 751         ret = irq_set_msi_desc(irq, msidesc);
 752         if (ret < 0)
 753                 goto error_irq;
 754 out:
 755         mutex_unlock(&irq_mapping_update_lock);
 756         return irq;
 757 error_irq:
 758         mutex_unlock(&irq_mapping_update_lock);
 759         xen_free_irq(irq);
 760         return ret;
 761 }
 762 #endif
 763
 764 int xen_destroy_irq(int irq)
 765 {
 766         struct irq_desc *desc;
 767         struct physdev_unmap_pirq unmap_irq;
 768         struct irq_info *info = info_for_irq(irq);
 769         int rc = -ENOENT;
 770
 771         mutex_lock(&irq_mapping_update_lock);
 772
 773         desc = irq_to_desc(irq);
 774         if (!desc)
 775                 goto out;
 776
 777         if (xen_initial_domain()) {
 778                 unmap_irq.pirq = info->u.pirq.pirq;
 779                 unmap_irq.domid = info->u.pirq.domid;
 780                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
 781                 /* If another domain quits without making the pci_disable_msix
 782                  * call, the Xen hypervisor takes care of freeing the PIRQs
 783                  * (free_domain_pirqs).
 784                  */
 785                 if ((rc == -ESRCH && info->u.pirq.domid != DOMID_SELF))
 786                         printk(KERN_INFO "domain %d does not have %d anymore\n",
 787                                 info->u.pirq.domid, info->u.pirq.pirq);
 788                 else if (rc) {
 789                         printk(KERN_WARNING "unmap irq failed %d\n", rc);
 790                         goto out;
 791                 }
 792         }
 793
 794         xen_free_irq(irq);
 795
 796 out:
 797         mutex_unlock(&irq_mapping_update_lock);
 798         return rc;
 799 }
 800
 801 int xen_irq_from_pirq(unsigned pirq)
 802 {
 803         int irq;
 804
 805         struct irq_info *info;
 806
 807         mutex_lock(&irq_mapping_update_lock);
 808
 809         list_for_each_entry(info, &xen_irq_list_head, list) {
 810                 if (info->type != IRQT_PIRQ)
 811                         continue;
 812                 irq = info->irq;
 813                 if (info->u.pirq.pirq == pirq)
 814                         goto out;
 815         }
 816         irq = -1;
 817 out:
 818         mutex_unlock(&irq_mapping_update_lock);
 819
 820         return irq;
 821 }
 822
 823
 824 int xen_pirq_from_irq(unsigned irq)
 825 {
 826         return pirq_from_irq(irq);
 827 }
 828 EXPORT_SYMBOL_GPL(xen_pirq_from_irq);
 829 int bind_evtchn_to_irq(unsigned int evtchn)
 830 {
 831         int irq;
 832
 833         mutex_lock(&irq_mapping_update_lock);
 834
 835         irq = evtchn_to_irq[evtchn];
 836
 837         if (irq == -1) {
 838                 irq = xen_allocate_irq_dynamic();
 839                 if (irq == -1)
 840                         goto out;
 841
 842                 irq_set_chip_and_handler_name(irq, &xen_dynamic_chip,
 843                                               handle_edge_irq, "event");
 844
 845                 xen_irq_info_evtchn_init(irq, evtchn);
 846         } else {
 847                 struct irq_info *info = info_for_irq(irq);
 848                 WARN_ON(info == NULL || info->type != IRQT_EVTCHN);
 849         }
 850         irq_clear_status_flags(irq, IRQ_NOREQUEST|IRQ_NOAUTOEN);
 851
 852 out:
 853         mutex_unlock(&irq_mapping_update_lock);
 854
 855         return irq;
 856 }
 857 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
 858
 859 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
 860 {
 861         struct evtchn_bind_ipi bind_ipi;
 862         int evtchn, irq;
 863
 864         mutex_lock(&irq_mapping_update_lock);
 865
 866         irq = per_cpu(ipi_to_irq, cpu)[ipi];
 867
 868         if (irq == -1) {
 869                 irq = xen_allocate_irq_dynamic();
 870                 if (irq < 0)
 871                         goto out;
 872
 873                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 874                                               handle_percpu_irq, "ipi");
 875
 876                 bind_ipi.vcpu = cpu;
 877                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
 878                                                 &bind_ipi) != 0)
 879                         BUG();
 880                 evtchn = bind_ipi.port;
 881
 882                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
 883
 884                 bind_evtchn_to_cpu(evtchn, cpu);
 885         } else {
 886                 struct irq_info *info = info_for_irq(irq);
 887                 WARN_ON(info == NULL || info->type != IRQT_IPI);
 888         }
 889
 890  out:
 891         mutex_unlock(&irq_mapping_update_lock);
 892         return irq;
 893 }
 894
 895 static int bind_interdomain_evtchn_to_irq(unsigned int remote_domain,
 896                                           unsigned int remote_port)
 897 {
 898         struct evtchn_bind_interdomain bind_interdomain;
 899         int err;
 900
 901         bind_interdomain.remote_dom  = remote_domain;
 902         bind_interdomain.remote_port = remote_port;
 903
 904         err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
 905                                           &bind_interdomain);
 906
 907         return err ? : bind_evtchn_to_irq(bind_interdomain.local_port);
 908 }
 909
 910 static int find_virq(unsigned int virq, unsigned int cpu)
 911 {
 912         struct evtchn_status status;
 913         int port, rc = -ENOENT;
 914
 915         memset(&status, 0, sizeof(status));
 916         for (port = 0; port <= NR_EVENT_CHANNELS; port++) {
 917                 status.dom = DOMID_SELF;
 918                 status.port = port;
 919                 rc = HYPERVISOR_event_channel_op(EVTCHNOP_status, &status);
 920                 if (rc < 0)
 921                         continue;
 922                 if (status.status != EVTCHNSTAT_virq)
 923                         continue;
 924                 if (status.u.virq == virq && status.vcpu == cpu) {
 925                         rc = port;
 926                         break;
 927                 }
 928         }
 929         return rc;
 930 }
 931
 932 int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
 933 {
 934         struct evtchn_bind_virq bind_virq;
 935         int evtchn, irq, ret;
 936
 937         mutex_lock(&irq_mapping_update_lock);
 938
 939         irq = per_cpu(virq_to_irq, cpu)[virq];
 940
 941         if (irq == -1) {
 942                 irq = xen_allocate_irq_dynamic();
 943                 if (irq == -1)
 944                         goto out;
 945
 946                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 947                                               handle_percpu_irq, "virq");
 948
 949                 bind_virq.virq = virq;
 950                 bind_virq.vcpu = cpu;
 951                 ret = HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 952                                                 &bind_virq);
 953                 if (ret == 0)
 954                         evtchn = bind_virq.port;
 955                 else {
 956                         if (ret == -EEXIST)
 957                                 ret = find_virq(virq, cpu);
 958                         BUG_ON(ret < 0);
 959                         evtchn = ret;
 960                 }
 961
 962                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
 963
 964                 bind_evtchn_to_cpu(evtchn, cpu);
 965         } else {
 966                 struct irq_info *info = info_for_irq(irq);
 967                 WARN_ON(info == NULL || info->type != IRQT_VIRQ);
 968         }
 969
 970 out:
 971         mutex_unlock(&irq_mapping_update_lock);
 972
 973         return irq;
 974 }
 975
 976 static void unbind_from_irq(unsigned int irq)
 977 {
 978         struct evtchn_close close;
 979         int evtchn = evtchn_from_irq(irq);
 980         struct irq_info *info = irq_get_handler_data(irq);
 981
 982         mutex_lock(&irq_mapping_update_lock);
 983
 984         if (info->refcnt > 0) {
 985                 info->refcnt--;
 986                 if (info->refcnt != 0)
 987                         goto done;
 988         }
 989
 990         if (VALID_EVTCHN(evtchn)) {
 991                 close.port = evtchn;
 992                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 993                         BUG();
 994
 995                 switch (type_from_irq(irq)) {
 996                 case IRQT_VIRQ:
 997                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
 998                                 [virq_from_irq(irq)] = -1;
 999                         break;
1000                 case IRQT_IPI:
1001                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
1002                                 [ipi_from_irq(irq)] = -1;
1003                         break;
1004                 default:
1005                         break;
1006                 }
1007
1008                 /* Closed ports are implicitly re-bound to VCPU0. */
1009                 bind_evtchn_to_cpu(evtchn, 0);
1010
1011                 evtchn_to_irq[evtchn] = -1;
1012         }
1013
1014         BUG_ON(info_for_irq(irq)->type == IRQT_UNBOUND);
1015
1016         xen_free_irq(irq);
1017
1018  done:
1019         mutex_unlock(&irq_mapping_update_lock);
1020 }
1021
1022 int bind_evtchn_to_irqhandler(unsigned int evtchn,
1023                               irq_handler_t handler,
1024                               unsigned long irqflags,
1025                               const char *devname, void *dev_id)
1026 {
1027         int irq, retval;
1028
1029         irq = bind_evtchn_to_irq(evtchn);
1030         if (irq < 0)
1031                 return irq;
1032         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1033         if (retval != 0) {
1034                 unbind_from_irq(irq);
1035                 return retval;
1036         }
1037
1038         return irq;
1039 }
1040 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
1041
1042 int bind_interdomain_evtchn_to_irqhandler(unsigned int remote_domain,
1043                                           unsigned int remote_port,
1044                                           irq_handler_t handler,
1045                                           unsigned long irqflags,
1046                                           const char *devname,
1047                                           void *dev_id)
1048 {
1049         int irq, retval;
1050
1051         irq = bind_interdomain_evtchn_to_irq(remote_domain, remote_port);
1052         if (irq < 0)
1053                 return irq;
1054
1055         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1056         if (retval != 0) {
1057                 unbind_from_irq(irq);
1058                 return retval;
1059         }
1060
1061         return irq;
1062 }
1063 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler);
1064
1065 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
1066                             irq_handler_t handler,
1067                             unsigned long irqflags, const char *devname, void *dev_id)
1068 {
1069         int irq, retval;
1070
1071         irq = bind_virq_to_irq(virq, cpu);
1072         if (irq < 0)
1073                 return irq;
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 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1083
1084 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1085                            unsigned int cpu,
1086                            irq_handler_t handler,
1087                            unsigned long irqflags,
1088                            const char *devname,
1089                            void *dev_id)
1090 {
1091         int irq, retval;
1092
1093         irq = bind_ipi_to_irq(ipi, cpu);
1094         if (irq < 0)
1095                 return irq;
1096
1097         irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME | IRQF_EARLY_RESUME;
1098         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1099         if (retval != 0) {
1100                 unbind_from_irq(irq);
1101                 return retval;
1102         }
1103
1104         return irq;
1105 }
1106
1107 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1108 {
1109         free_irq(irq, dev_id);
1110         unbind_from_irq(irq);
1111 }
1112 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1113
1114 int evtchn_make_refcounted(unsigned int evtchn)
1115 {
1116         int irq = evtchn_to_irq[evtchn];
1117         struct irq_info *info;
1118
1119         if (irq == -1)
1120                 return -ENOENT;
1121
1122         info = irq_get_handler_data(irq);
1123
1124         if (!info)
1125                 return -ENOENT;
1126
1127         WARN_ON(info->refcnt != -1);
1128
1129         info->refcnt = 1;
1130
1131         return 0;
1132 }
1133 EXPORT_SYMBOL_GPL(evtchn_make_refcounted);
1134
1135 int evtchn_get(unsigned int evtchn)
1136 {
1137         int irq;
1138         struct irq_info *info;
1139         int err = -ENOENT;
1140
1141         if (evtchn >= NR_EVENT_CHANNELS)
1142                 return -EINVAL;
1143
1144         mutex_lock(&irq_mapping_update_lock);
1145
1146         irq = evtchn_to_irq[evtchn];
1147         if (irq == -1)
1148                 goto done;
1149
1150         info = irq_get_handler_data(irq);
1151
1152         if (!info)
1153                 goto done;
1154
1155         err = -EINVAL;
1156         if (info->refcnt <= 0)
1157                 goto done;
1158
1159         info->refcnt++;
1160         err = 0;
1161  done:
1162         mutex_unlock(&irq_mapping_update_lock);
1163
1164         return err;
1165 }
1166 EXPORT_SYMBOL_GPL(evtchn_get);
1167
1168 void evtchn_put(unsigned int evtchn)
1169 {
1170         int irq = evtchn_to_irq[evtchn];
1171         if (WARN_ON(irq == -1))
1172                 return;
1173         unbind_from_irq(irq);
1174 }
1175 EXPORT_SYMBOL_GPL(evtchn_put);
1176
1177 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1178 {
1179         int irq = per_cpu(ipi_to_irq, cpu)[vector];
1180         BUG_ON(irq < 0);
1181         notify_remote_via_irq(irq);
1182 }
1183
1184 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
1185 {
1186         struct shared_info *sh = HYPERVISOR_shared_info;
1187         int cpu = smp_processor_id();
1188         unsigned long *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu);
1189         int i;
1190         unsigned long flags;
1191         static DEFINE_SPINLOCK(debug_lock);
1192         struct vcpu_info *v;
1193
1194         spin_lock_irqsave(&debug_lock, flags);
1195
1196         printk("\nvcpu %d\n  ", cpu);
1197
1198         for_each_online_cpu(i) {
1199                 int pending;
1200                 v = per_cpu(xen_vcpu, i);
1201                 pending = (get_irq_regs() && i == cpu)
1202                         ? xen_irqs_disabled(get_irq_regs())
1203                         : v->evtchn_upcall_mask;
1204                 printk("%d: masked=%d pending=%d event_sel %0*lx\n  ", i,
1205                        pending, v->evtchn_upcall_pending,
1206                        (int)(sizeof(v->evtchn_pending_sel)*2),
1207                        v->evtchn_pending_sel);
1208         }
1209         v = per_cpu(xen_vcpu, cpu);
1210
1211         printk("\npending:\n   ");
1212         for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
1213                 printk("%0*lx%s", (int)sizeof(sh->evtchn_pending[0])*2,
1214                        sh->evtchn_pending[i],
1215                        i % 8 == 0 ? "\n   " : " ");
1216         printk("\nglobal mask:\n   ");
1217         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1218                 printk("%0*lx%s",
1219                        (int)(sizeof(sh->evtchn_mask[0])*2),
1220                        sh->evtchn_mask[i],
1221                        i % 8 == 0 ? "\n   " : " ");
1222
1223         printk("\nglobally unmasked:\n   ");
1224         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1225                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1226                        sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
1227                        i % 8 == 0 ? "\n   " : " ");
1228
1229         printk("\nlocal cpu%d mask:\n   ", cpu);
1230         for (i = (NR_EVENT_CHANNELS/BITS_PER_LONG)-1; i >= 0; i--)
1231                 printk("%0*lx%s", (int)(sizeof(cpu_evtchn[0])*2),
1232                        cpu_evtchn[i],
1233                        i % 8 == 0 ? "\n   " : " ");
1234
1235         printk("\nlocally unmasked:\n   ");
1236         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
1237                 unsigned long pending = sh->evtchn_pending[i]
1238                         & ~sh->evtchn_mask[i]
1239                         & cpu_evtchn[i];
1240                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1241                        pending, i % 8 == 0 ? "\n   " : " ");
1242         }
1243
1244         printk("\npending list:\n");
1245         for (i = 0; i < NR_EVENT_CHANNELS; i++) {
1246                 if (sync_test_bit(i, sh->evtchn_pending)) {
1247                         int word_idx = i / BITS_PER_LONG;
1248                         printk("  %d: event %d -> irq %d%s%s%s\n",
1249                                cpu_from_evtchn(i), i,
1250                                evtchn_to_irq[i],
1251                                sync_test_bit(word_idx, &v->evtchn_pending_sel)
1252                                              ? "" : " l2-clear",
1253                                !sync_test_bit(i, sh->evtchn_mask)
1254                                              ? "" : " globally-masked",
1255                                sync_test_bit(i, cpu_evtchn)
1256                                              ? "" : " locally-masked");
1257                 }
1258         }
1259
1260         spin_unlock_irqrestore(&debug_lock, flags);
1261
1262         return IRQ_HANDLED;
1263 }
1264
1265 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
1266 static DEFINE_PER_CPU(unsigned int, current_word_idx);
1267 static DEFINE_PER_CPU(unsigned int, current_bit_idx);
1268
1269 /*
1270  * Mask out the i least significant bits of w
1271  */
1272 #define MASK_LSBS(w, i) (w & ((~0UL) << i))
1273
1274 /*
1275  * Search the CPUs pending events bitmasks.  For each one found, map
1276  * the event number to an irq, and feed it into do_IRQ() for
1277  * handling.
1278  *
1279  * Xen uses a two-level bitmap to speed searching.  The first level is
1280  * a bitset of words which contain pending event bits.  The second
1281  * level is a bitset of pending events themselves.
1282  */
1283 static void __xen_evtchn_do_upcall(void)
1284 {
1285         int start_word_idx, start_bit_idx;
1286         int word_idx, bit_idx;
1287         int i;
1288         int cpu = get_cpu();
1289         struct shared_info *s = HYPERVISOR_shared_info;
1290         struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1291         unsigned count;
1292
1293         do {
1294                 unsigned long pending_words;
1295
1296                 vcpu_info->evtchn_upcall_pending = 0;
1297
1298                 if (__this_cpu_inc_return(xed_nesting_count) - 1)
1299                         goto out;
1300
1301 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
1302                 /* Clear master flag /before/ clearing selector flag. */
1303                 wmb();
1304 #endif
1305                 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
1306
1307                 start_word_idx = __this_cpu_read(current_word_idx);
1308                 start_bit_idx = __this_cpu_read(current_bit_idx);
1309
1310                 word_idx = start_word_idx;
1311
1312                 for (i = 0; pending_words != 0; i++) {
1313                         unsigned long pending_bits;
1314                         unsigned long words;
1315
1316                         words = MASK_LSBS(pending_words, word_idx);
1317
1318                         /*
1319                          * If we masked out all events, wrap to beginning.
1320                          */
1321                         if (words == 0) {
1322                                 word_idx = 0;
1323                                 bit_idx = 0;
1324                                 continue;
1325                         }
1326                         word_idx = __ffs(words);
1327
1328                         pending_bits = active_evtchns(cpu, s, word_idx);
1329                         bit_idx = 0; /* usually scan entire word from start */
1330                         if (word_idx == start_word_idx) {
1331                                 /* We scan the starting word in two parts */
1332                                 if (i == 0)
1333                                         /* 1st time: start in the middle */
1334                                         bit_idx = start_bit_idx;
1335                                 else
1336                                         /* 2nd time: mask bits done already */
1337                                         bit_idx &= (1UL << start_bit_idx) - 1;
1338                         }
1339
1340                         do {
1341                                 unsigned long bits;
1342                                 int port, irq;
1343                                 struct irq_desc *desc;
1344
1345                                 bits = MASK_LSBS(pending_bits, bit_idx);
1346
1347                                 /* If we masked out all events, move on. */
1348                                 if (bits == 0)
1349                                         break;
1350
1351                                 bit_idx = __ffs(bits);
1352
1353                                 /* Process port. */
1354                                 port = (word_idx * BITS_PER_LONG) + bit_idx;
1355                                 irq = evtchn_to_irq[port];
1356
1357                                 if (irq != -1) {
1358                                         desc = irq_to_desc(irq);
1359                                         if (desc)
1360                                                 generic_handle_irq_desc(irq, desc);
1361                                 }
1362
1363                                 bit_idx = (bit_idx + 1) % BITS_PER_LONG;
1364
1365                                 /* Next caller starts at last processed + 1 */
1366                                 __this_cpu_write(current_word_idx,
1367                                                  bit_idx ? word_idx :
1368                                                  (word_idx+1) % BITS_PER_LONG);
1369                                 __this_cpu_write(current_bit_idx, bit_idx);
1370                         } while (bit_idx != 0);
1371
1372                         /* Scan start_l1i twice; all others once. */
1373                         if ((word_idx != start_word_idx) || (i != 0))
1374                                 pending_words &= ~(1UL << word_idx);
1375
1376                         word_idx = (word_idx + 1) % BITS_PER_LONG;
1377                 }
1378
1379                 BUG_ON(!irqs_disabled());
1380
1381                 count = __this_cpu_read(xed_nesting_count);
1382                 __this_cpu_write(xed_nesting_count, 0);
1383         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
1384
1385 out:
1386
1387         put_cpu();
1388 }
1389
1390 void xen_evtchn_do_upcall(struct pt_regs *regs)
1391 {
1392         struct pt_regs *old_regs = set_irq_regs(regs);
1393
1394 #ifdef CONFIG_X86
1395         exit_idle();
1396 #endif
1397         irq_enter();
1398
1399         __xen_evtchn_do_upcall();
1400
1401         irq_exit();
1402         set_irq_regs(old_regs);
1403 }
1404
1405 void xen_hvm_evtchn_do_upcall(void)
1406 {
1407         __xen_evtchn_do_upcall();
1408 }
1409 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1410
1411 /* Rebind a new event channel to an existing irq. */
1412 void rebind_evtchn_irq(int evtchn, int irq)
1413 {
1414         struct irq_info *info = info_for_irq(irq);
1415
1416         /* Make sure the irq is masked, since the new event channel
1417            will also be masked. */
1418         disable_irq(irq);
1419
1420         mutex_lock(&irq_mapping_update_lock);
1421
1422         /* After resume the irq<->evtchn mappings are all cleared out */
1423         BUG_ON(evtchn_to_irq[evtchn] != -1);
1424         /* Expect irq to have been bound before,
1425            so there should be a proper type */
1426         BUG_ON(info->type == IRQT_UNBOUND);
1427
1428         xen_irq_info_evtchn_init(irq, evtchn);
1429
1430         mutex_unlock(&irq_mapping_update_lock);
1431
1432         /* new event channels are always bound to cpu 0 */
1433         irq_set_affinity(irq, cpumask_of(0));
1434
1435         /* Unmask the event channel. */
1436         enable_irq(irq);
1437 }
1438
1439 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1440 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1441 {
1442         struct evtchn_bind_vcpu bind_vcpu;
1443         int evtchn = evtchn_from_irq(irq);
1444
1445         if (!VALID_EVTCHN(evtchn))
1446                 return -1;
1447
1448         /*
1449          * Events delivered via platform PCI interrupts are always
1450          * routed to vcpu 0 and hence cannot be rebound.
1451          */
1452         if (xen_hvm_domain() && !xen_have_vector_callback)
1453                 return -1;
1454
1455         /* Send future instances of this interrupt to other vcpu. */
1456         bind_vcpu.port = evtchn;
1457         bind_vcpu.vcpu = tcpu;
1458
1459         /*
1460          * If this fails, it usually just indicates that we're dealing with a
1461          * virq or IPI channel, which don't actually need to be rebound. Ignore
1462          * it, but don't do the xenlinux-level rebind in that case.
1463          */
1464         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1465                 bind_evtchn_to_cpu(evtchn, tcpu);
1466
1467         return 0;
1468 }
1469
1470 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1471                             bool force)
1472 {
1473         unsigned tcpu = cpumask_first(dest);
1474
1475         return rebind_irq_to_cpu(data->irq, tcpu);
1476 }
1477
1478 int resend_irq_on_evtchn(unsigned int irq)
1479 {
1480         int masked, evtchn = evtchn_from_irq(irq);
1481         struct shared_info *s = HYPERVISOR_shared_info;
1482
1483         if (!VALID_EVTCHN(evtchn))
1484                 return 1;
1485
1486         masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
1487         sync_set_bit(evtchn, s->evtchn_pending);
1488         if (!masked)
1489                 unmask_evtchn(evtchn);
1490
1491         return 1;
1492 }
1493
1494 static void enable_dynirq(struct irq_data *data)
1495 {
1496         int evtchn = evtchn_from_irq(data->irq);
1497
1498         if (VALID_EVTCHN(evtchn))
1499                 unmask_evtchn(evtchn);
1500 }
1501
1502 static void disable_dynirq(struct irq_data *data)
1503 {
1504         int evtchn = evtchn_from_irq(data->irq);
1505
1506         if (VALID_EVTCHN(evtchn))
1507                 mask_evtchn(evtchn);
1508 }
1509
1510 static void ack_dynirq(struct irq_data *data)
1511 {
1512         int evtchn = evtchn_from_irq(data->irq);
1513
1514         irq_move_irq(data);
1515
1516         if (VALID_EVTCHN(evtchn))
1517                 clear_evtchn(evtchn);
1518 }
1519
1520 static void mask_ack_dynirq(struct irq_data *data)
1521 {
1522         disable_dynirq(data);
1523         ack_dynirq(data);
1524 }
1525
1526 static int retrigger_dynirq(struct irq_data *data)
1527 {
1528         int evtchn = evtchn_from_irq(data->irq);
1529         struct shared_info *sh = HYPERVISOR_shared_info;
1530         int ret = 0;
1531
1532         if (VALID_EVTCHN(evtchn)) {
1533                 int masked;
1534
1535                 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
1536                 sync_set_bit(evtchn, sh->evtchn_pending);
1537                 if (!masked)
1538                         unmask_evtchn(evtchn);
1539                 ret = 1;
1540         }
1541
1542         return ret;
1543 }
1544
1545 static void restore_pirqs(void)
1546 {
1547         int pirq, rc, irq, gsi;
1548         struct physdev_map_pirq map_irq;
1549         struct irq_info *info;
1550
1551         list_for_each_entry(info, &xen_irq_list_head, list) {
1552                 if (info->type != IRQT_PIRQ)
1553                         continue;
1554
1555                 pirq = info->u.pirq.pirq;
1556                 gsi = info->u.pirq.gsi;
1557                 irq = info->irq;
1558
1559                 /* save/restore of PT devices doesn't work, so at this point the
1560                  * only devices present are GSI based emulated devices */
1561                 if (!gsi)
1562                         continue;
1563
1564                 map_irq.domid = DOMID_SELF;
1565                 map_irq.type = MAP_PIRQ_TYPE_GSI;
1566                 map_irq.index = gsi;
1567                 map_irq.pirq = pirq;
1568
1569                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1570                 if (rc) {
1571                         printk(KERN_WARNING "xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1572                                         gsi, irq, pirq, rc);
1573                         xen_free_irq(irq);
1574                         continue;
1575                 }
1576
1577                 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1578
1579                 __startup_pirq(irq);
1580         }
1581 }
1582
1583 static void restore_cpu_virqs(unsigned int cpu)
1584 {
1585         struct evtchn_bind_virq bind_virq;
1586         int virq, irq, evtchn;
1587
1588         for (virq = 0; virq < NR_VIRQS; virq++) {
1589                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1590                         continue;
1591
1592                 BUG_ON(virq_from_irq(irq) != virq);
1593
1594                 /* Get a new binding from Xen. */
1595                 bind_virq.virq = virq;
1596                 bind_virq.vcpu = cpu;
1597                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1598                                                 &bind_virq) != 0)
1599                         BUG();
1600                 evtchn = bind_virq.port;
1601
1602                 /* Record the new mapping. */
1603                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
1604                 bind_evtchn_to_cpu(evtchn, cpu);
1605         }
1606 }
1607
1608 static void restore_cpu_ipis(unsigned int cpu)
1609 {
1610         struct evtchn_bind_ipi bind_ipi;
1611         int ipi, irq, evtchn;
1612
1613         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1614                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1615                         continue;
1616
1617                 BUG_ON(ipi_from_irq(irq) != ipi);
1618
1619                 /* Get a new binding from Xen. */
1620                 bind_ipi.vcpu = cpu;
1621                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1622                                                 &bind_ipi) != 0)
1623                         BUG();
1624                 evtchn = bind_ipi.port;
1625
1626                 /* Record the new mapping. */
1627                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
1628                 bind_evtchn_to_cpu(evtchn, cpu);
1629         }
1630 }
1631
1632 /* Clear an irq's pending state, in preparation for polling on it */
1633 void xen_clear_irq_pending(int irq)
1634 {
1635         int evtchn = evtchn_from_irq(irq);
1636
1637         if (VALID_EVTCHN(evtchn))
1638                 clear_evtchn(evtchn);
1639 }
1640 EXPORT_SYMBOL(xen_clear_irq_pending);
1641 void xen_set_irq_pending(int irq)
1642 {
1643         int evtchn = evtchn_from_irq(irq);
1644
1645         if (VALID_EVTCHN(evtchn))
1646                 set_evtchn(evtchn);
1647 }
1648
1649 bool xen_test_irq_pending(int irq)
1650 {
1651         int evtchn = evtchn_from_irq(irq);
1652         bool ret = false;
1653
1654         if (VALID_EVTCHN(evtchn))
1655                 ret = test_evtchn(evtchn);
1656
1657         return ret;
1658 }
1659
1660 /* Poll waiting for an irq to become pending with timeout.  In the usual case,
1661  * the irq will be disabled so it won't deliver an interrupt. */
1662 void xen_poll_irq_timeout(int irq, u64 timeout)
1663 {
1664         evtchn_port_t evtchn = evtchn_from_irq(irq);
1665
1666         if (VALID_EVTCHN(evtchn)) {
1667                 struct sched_poll poll;
1668
1669                 poll.nr_ports = 1;
1670                 poll.timeout = timeout;
1671                 set_xen_guest_handle(poll.ports, &evtchn);
1672
1673                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1674                         BUG();
1675         }
1676 }
1677 EXPORT_SYMBOL(xen_poll_irq_timeout);
1678 /* Poll waiting for an irq to become pending.  In the usual case, the
1679  * irq will be disabled so it won't deliver an interrupt. */
1680 void xen_poll_irq(int irq)
1681 {
1682         xen_poll_irq_timeout(irq, 0 /* no timeout */);
1683 }
1684
1685 /* Check whether the IRQ line is shared with other guests. */
1686 int xen_test_irq_shared(int irq)
1687 {
1688         struct irq_info *info = info_for_irq(irq);
1689         struct physdev_irq_status_query irq_status = { .irq = info->u.pirq.pirq };
1690
1691         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
1692                 return 0;
1693         return !(irq_status.flags & XENIRQSTAT_shared);
1694 }
1695 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
1696
1697 void xen_irq_resume(void)
1698 {
1699         unsigned int cpu, evtchn;
1700         struct irq_info *info;
1701
1702         init_evtchn_cpu_bindings();
1703
1704         /* New event-channel space is not 'live' yet. */
1705         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1706                 mask_evtchn(evtchn);
1707
1708         /* No IRQ <-> event-channel mappings. */
1709         list_for_each_entry(info, &xen_irq_list_head, list)
1710                 info->evtchn = 0; /* zap event-channel binding */
1711
1712         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1713                 evtchn_to_irq[evtchn] = -1;
1714
1715         for_each_possible_cpu(cpu) {
1716                 restore_cpu_virqs(cpu);
1717                 restore_cpu_ipis(cpu);
1718         }
1719
1720         restore_pirqs();
1721 }
1722
1723 static struct irq_chip xen_dynamic_chip __read_mostly = {
1724         .name                   = "xen-dyn",
1725
1726         .irq_disable            = disable_dynirq,
1727         .irq_mask               = disable_dynirq,
1728         .irq_unmask             = enable_dynirq,
1729
1730         .irq_ack                = ack_dynirq,
1731         .irq_mask_ack           = mask_ack_dynirq,
1732
1733         .irq_set_affinity       = set_affinity_irq,
1734         .irq_retrigger          = retrigger_dynirq,
1735 };
1736
1737 static struct irq_chip xen_pirq_chip __read_mostly = {
1738         .name                   = "xen-pirq",
1739
1740         .irq_startup            = startup_pirq,
1741         .irq_shutdown           = shutdown_pirq,
1742         .irq_enable             = enable_pirq,
1743         .irq_disable            = disable_pirq,
1744
1745         .irq_mask               = disable_dynirq,
1746         .irq_unmask             = enable_dynirq,
1747
1748         .irq_ack                = eoi_pirq,
1749         .irq_eoi                = eoi_pirq,
1750         .irq_mask_ack           = mask_ack_pirq,
1751
1752         .irq_set_affinity       = set_affinity_irq,
1753
1754         .irq_retrigger          = retrigger_dynirq,
1755 };
1756
1757 static struct irq_chip xen_percpu_chip __read_mostly = {
1758         .name                   = "xen-percpu",
1759
1760         .irq_disable            = disable_dynirq,
1761         .irq_mask               = disable_dynirq,
1762         .irq_unmask             = enable_dynirq,
1763
1764         .irq_ack                = ack_dynirq,
1765 };
1766
1767 int xen_set_callback_via(uint64_t via)
1768 {
1769         struct xen_hvm_param a;
1770         a.domid = DOMID_SELF;
1771         a.index = HVM_PARAM_CALLBACK_IRQ;
1772         a.value = via;
1773         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1774 }
1775 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1776
1777 #ifdef CONFIG_XEN_PVHVM
1778 /* Vector callbacks are better than PCI interrupts to receive event
1779  * channel notifications because we can receive vector callbacks on any
1780  * vcpu and we don't need PCI support or APIC interactions. */
1781 void xen_callback_vector(void)
1782 {
1783         int rc;
1784         uint64_t callback_via;
1785         if (xen_have_vector_callback) {
1786                 callback_via = HVM_CALLBACK_VECTOR(XEN_HVM_EVTCHN_CALLBACK);
1787                 rc = xen_set_callback_via(callback_via);
1788                 if (rc) {
1789                         printk(KERN_ERR "Request for Xen HVM callback vector"
1790                                         " failed.\n");
1791                         xen_have_vector_callback = 0;
1792                         return;
1793                 }
1794                 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1795                                 "enabled\n");
1796                 /* in the restore case the vector has already been allocated */
1797                 if (!test_bit(XEN_HVM_EVTCHN_CALLBACK, used_vectors))
1798                         alloc_intr_gate(XEN_HVM_EVTCHN_CALLBACK, xen_hvm_callback_vector);
1799         }
1800 }
1801 #else
1802 void xen_callback_vector(void) {}
1803 #endif
1804
1805 void __init xen_init_IRQ(void)
1806 {
1807         int i;
1808
1809         evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1810                                     GFP_KERNEL);
1811         BUG_ON(!evtchn_to_irq);
1812         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1813                 evtchn_to_irq[i] = -1;
1814
1815         init_evtchn_cpu_bindings();
1816
1817         /* No event channels are 'live' right now. */
1818         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1819                 mask_evtchn(i);
1820
1821         pirq_needs_eoi = pirq_needs_eoi_flag;
1822
1823 #ifdef CONFIG_X86
1824         if (xen_hvm_domain()) {
1825                 xen_callback_vector();
1826                 native_init_IRQ();
1827                 /* pci_xen_hvm_init must be called after native_init_IRQ so that
1828                  * __acpi_register_gsi can point at the right function */
1829                 pci_xen_hvm_init();
1830         } else {
1831                 int rc;
1832                 struct physdev_pirq_eoi_gmfn eoi_gmfn;
1833
1834                 irq_ctx_init(smp_processor_id());
1835                 if (xen_initial_domain())
1836                         pci_xen_initial_domain();
1837
1838                 pirq_eoi_map = (void *)__get_free_page(GFP_KERNEL|__GFP_ZERO);
1839                 eoi_gmfn.gmfn = virt_to_mfn(pirq_eoi_map);
1840                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_pirq_eoi_gmfn_v2, &eoi_gmfn);
1841                 if (rc != 0) {
1842                         free_page((unsigned long) pirq_eoi_map);
1843                         pirq_eoi_map = NULL;
1844                 } else
1845                         pirq_needs_eoi = pirq_check_eoi_map;
1846         }
1847 #endif
1848 }