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