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 recieved, 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. Hardware interrupts. Not supported at present.
  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
  32 #include <asm/desc.h>
  33 #include <asm/ptrace.h>
  34 #include <asm/irq.h>
  35 #include <asm/idle.h>
  36 #include <asm/sync_bitops.h>
  37 #include <asm/xen/hypercall.h>
  38 #include <asm/xen/hypervisor.h>
  39
  40 #include <xen/xen.h>
  41 #include <xen/hvm.h>
  42 #include <xen/xen-ops.h>
  43 #include <xen/events.h>
  44 #include <xen/interface/xen.h>
  45 #include <xen/interface/event_channel.h>
  46 #include <xen/interface/hvm/hvm_op.h>
  47 #include <xen/interface/hvm/params.h>
  48
  49 /*
  50  * This lock protects updates to the following mapping and reference-count
  51  * arrays. The lock does not need to be acquired to read the mapping tables.
  52  */
  53 static DEFINE_SPINLOCK(irq_mapping_update_lock);
  54
  55 /* IRQ <-> VIRQ mapping. */
  56 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
  57
  58 /* IRQ <-> IPI mapping */
  59 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
  60
  61 /* Interrupt types. */
  62 enum xen_irq_type {
  63         IRQT_UNBOUND = 0,
  64         IRQT_PIRQ,
  65         IRQT_VIRQ,
  66         IRQT_IPI,
  67         IRQT_EVTCHN
  68 };
  69
  70 /*
  71  * Packed IRQ information:
  72  * type - enum xen_irq_type
  73  * event channel - irq->event channel mapping
  74  * cpu - cpu this event channel is bound to
  75  * index - type-specific information:
  76  *    PIRQ - vector, with MSB being "needs EIO"
  77  *    VIRQ - virq number
  78  *    IPI - IPI vector
  79  *    EVTCHN -
  80  */
  81 struct irq_info
  82 {
  83         enum xen_irq_type type; /* type */
  84         unsigned short evtchn;  /* event channel */
  85         unsigned short cpu;     /* cpu bound */
  86
  87         union {
  88                 unsigned short virq;
  89                 enum ipi_vector ipi;
  90                 struct {
  91                         unsigned short gsi;
  92                         unsigned short vector;
  93                 } pirq;
  94         } u;
  95 };
  96
  97 static struct irq_info irq_info[NR_IRQS];
  98
  99 static int evtchn_to_irq[NR_EVENT_CHANNELS] = {
 100         [0 ... NR_EVENT_CHANNELS-1] = -1
 101 };
 102 struct cpu_evtchn_s {
 103         unsigned long bits[NR_EVENT_CHANNELS/BITS_PER_LONG];
 104 };
 105 static struct cpu_evtchn_s *cpu_evtchn_mask_p;
 106 static inline unsigned long *cpu_evtchn_mask(int cpu)
 107 {
 108         return cpu_evtchn_mask_p[cpu].bits;
 109 }
 110
 111 /* Xen will never allocate port zero for any purpose. */
 112 #define VALID_EVTCHN(chn)       ((chn) != 0)
 113
 114 static struct irq_chip xen_dynamic_chip;
 115 static struct irq_chip xen_percpu_chip;
 116
 117 /* Constructor for packed IRQ information. */
 118 static struct irq_info mk_unbound_info(void)
 119 {
 120         return (struct irq_info) { .type = IRQT_UNBOUND };
 121 }
 122
 123 static struct irq_info mk_evtchn_info(unsigned short evtchn)
 124 {
 125         return (struct irq_info) { .type = IRQT_EVTCHN, .evtchn = evtchn,
 126                         .cpu = 0 };
 127 }
 128
 129 static struct irq_info mk_ipi_info(unsigned short evtchn, enum ipi_vector ipi)
 130 {
 131         return (struct irq_info) { .type = IRQT_IPI, .evtchn = evtchn,
 132                         .cpu = 0, .u.ipi = ipi };
 133 }
 134
 135 static struct irq_info mk_virq_info(unsigned short evtchn, unsigned short virq)
 136 {
 137         return (struct irq_info) { .type = IRQT_VIRQ, .evtchn = evtchn,
 138                         .cpu = 0, .u.virq = virq };
 139 }
 140
 141 static struct irq_info mk_pirq_info(unsigned short evtchn,
 142                                     unsigned short gsi, unsigned short vector)
 143 {
 144         return (struct irq_info) { .type = IRQT_PIRQ, .evtchn = evtchn,
 145                         .cpu = 0, .u.pirq = { .gsi = gsi, .vector = vector } };
 146 }
 147
 148 /*
 149  * Accessors for packed IRQ information.
 150  */
 151 static struct irq_info *info_for_irq(unsigned irq)
 152 {
 153         return &irq_info[irq];
 154 }
 155
 156 static unsigned int evtchn_from_irq(unsigned irq)
 157 {
 158         return info_for_irq(irq)->evtchn;
 159 }
 160
 161 unsigned irq_from_evtchn(unsigned int evtchn)
 162 {
 163         return evtchn_to_irq[evtchn];
 164 }
 165 EXPORT_SYMBOL_GPL(irq_from_evtchn);
 166
 167 static enum ipi_vector ipi_from_irq(unsigned irq)
 168 {
 169         struct irq_info *info = info_for_irq(irq);
 170
 171         BUG_ON(info == NULL);
 172         BUG_ON(info->type != IRQT_IPI);
 173
 174         return info->u.ipi;
 175 }
 176
 177 static unsigned virq_from_irq(unsigned irq)
 178 {
 179         struct irq_info *info = info_for_irq(irq);
 180
 181         BUG_ON(info == NULL);
 182         BUG_ON(info->type != IRQT_VIRQ);
 183
 184         return info->u.virq;
 185 }
 186
 187 static unsigned gsi_from_irq(unsigned irq)
 188 {
 189         struct irq_info *info = info_for_irq(irq);
 190
 191         BUG_ON(info == NULL);
 192         BUG_ON(info->type != IRQT_PIRQ);
 193
 194         return info->u.pirq.gsi;
 195 }
 196
 197 static unsigned vector_from_irq(unsigned irq)
 198 {
 199         struct irq_info *info = info_for_irq(irq);
 200
 201         BUG_ON(info == NULL);
 202         BUG_ON(info->type != IRQT_PIRQ);
 203
 204         return info->u.pirq.vector;
 205 }
 206
 207 static enum xen_irq_type type_from_irq(unsigned irq)
 208 {
 209         return info_for_irq(irq)->type;
 210 }
 211
 212 static unsigned cpu_from_irq(unsigned irq)
 213 {
 214         return info_for_irq(irq)->cpu;
 215 }
 216
 217 static unsigned int cpu_from_evtchn(unsigned int evtchn)
 218 {
 219         int irq = evtchn_to_irq[evtchn];
 220         unsigned ret = 0;
 221
 222         if (irq != -1)
 223                 ret = cpu_from_irq(irq);
 224
 225         return ret;
 226 }
 227
 228 static inline unsigned long active_evtchns(unsigned int cpu,
 229                                            struct shared_info *sh,
 230                                            unsigned int idx)
 231 {
 232         return (sh->evtchn_pending[idx] &
 233                 cpu_evtchn_mask(cpu)[idx] &
 234                 ~sh->evtchn_mask[idx]);
 235 }
 236
 237 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
 238 {
 239         int irq = evtchn_to_irq[chn];
 240
 241         BUG_ON(irq == -1);
 242 #ifdef CONFIG_SMP
 243         cpumask_copy(irq_to_desc(irq)->affinity, cpumask_of(cpu));
 244 #endif
 245
 246         __clear_bit(chn, cpu_evtchn_mask(cpu_from_irq(irq)));
 247         __set_bit(chn, cpu_evtchn_mask(cpu));
 248
 249         irq_info[irq].cpu = cpu;
 250 }
 251
 252 static void init_evtchn_cpu_bindings(void)
 253 {
 254 #ifdef CONFIG_SMP
 255         struct irq_desc *desc;
 256         int i;
 257
 258         /* By default all event channels notify CPU#0. */
 259         for_each_irq_desc(i, desc) {
 260                 cpumask_copy(desc->affinity, cpumask_of(0));
 261         }
 262 #endif
 263
 264         memset(cpu_evtchn_mask(0), ~0, sizeof(struct cpu_evtchn_s));
 265 }
 266
 267 static inline void clear_evtchn(int port)
 268 {
 269         struct shared_info *s = HYPERVISOR_shared_info;
 270         sync_clear_bit(port, &s->evtchn_pending[0]);
 271 }
 272
 273 static inline void set_evtchn(int port)
 274 {
 275         struct shared_info *s = HYPERVISOR_shared_info;
 276         sync_set_bit(port, &s->evtchn_pending[0]);
 277 }
 278
 279 static inline int test_evtchn(int port)
 280 {
 281         struct shared_info *s = HYPERVISOR_shared_info;
 282         return sync_test_bit(port, &s->evtchn_pending[0]);
 283 }
 284
 285
 286 /**
 287  * notify_remote_via_irq - send event to remote end of event channel via irq
 288  * @irq: irq of event channel to send event to
 289  *
 290  * Unlike notify_remote_via_evtchn(), this is safe to use across
 291  * save/restore. Notifications on a broken connection are silently
 292  * dropped.
 293  */
 294 void notify_remote_via_irq(int irq)
 295 {
 296         int evtchn = evtchn_from_irq(irq);
 297
 298         if (VALID_EVTCHN(evtchn))
 299                 notify_remote_via_evtchn(evtchn);
 300 }
 301 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
 302
 303 static void mask_evtchn(int port)
 304 {
 305         struct shared_info *s = HYPERVISOR_shared_info;
 306         sync_set_bit(port, &s->evtchn_mask[0]);
 307 }
 308
 309 static void unmask_evtchn(int port)
 310 {
 311         struct shared_info *s = HYPERVISOR_shared_info;
 312         unsigned int cpu = get_cpu();
 313
 314         BUG_ON(!irqs_disabled());
 315
 316         /* Slow path (hypercall) if this is a non-local port. */
 317         if (unlikely(cpu != cpu_from_evtchn(port))) {
 318                 struct evtchn_unmask unmask = { .port = port };
 319                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 320         } else {
 321                 struct vcpu_info *vcpu_info = __get_cpu_var(xen_vcpu);
 322
 323                 sync_clear_bit(port, &s->evtchn_mask[0]);
 324
 325                 /*
 326                  * The following is basically the equivalent of
 327                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
 328                  * the interrupt edge' if the channel is masked.
 329                  */
 330                 if (sync_test_bit(port, &s->evtchn_pending[0]) &&
 331                     !sync_test_and_set_bit(port / BITS_PER_LONG,
 332                                            &vcpu_info->evtchn_pending_sel))
 333                         vcpu_info->evtchn_upcall_pending = 1;
 334         }
 335
 336         put_cpu();
 337 }
 338
 339 static int find_unbound_irq(void)
 340 {
 341         int irq;
 342         struct irq_desc *desc;
 343
 344         for (irq = 0; irq < nr_irqs; irq++) {
 345                 desc = irq_to_desc(irq);
 346                 /* only 0->15 have init'd desc; handle irq > 16 */
 347                 if (desc == NULL)
 348                         break;
 349                 if (desc->chip == &no_irq_chip)
 350                         break;
 351                 if (desc->chip != &xen_dynamic_chip)
 352                         continue;
 353                 if (irq_info[irq].type == IRQT_UNBOUND)
 354                         break;
 355         }
 356
 357         if (irq == nr_irqs)
 358                 panic("No available IRQ to bind to: increase nr_irqs!\n");
 359
 360         desc = irq_to_desc_alloc_node(irq, 0);
 361         if (WARN_ON(desc == NULL))
 362                 return -1;
 363
 364         dynamic_irq_init_keep_chip_data(irq);
 365
 366         return irq;
 367 }
 368
 369 int bind_evtchn_to_irq(unsigned int evtchn)
 370 {
 371         int irq;
 372
 373         spin_lock(&irq_mapping_update_lock);
 374
 375         irq = evtchn_to_irq[evtchn];
 376
 377         if (irq == -1) {
 378                 irq = find_unbound_irq();
 379
 380                 set_irq_chip_and_handler_name(irq, &xen_dynamic_chip,
 381                                               handle_edge_irq, "event");
 382
 383                 evtchn_to_irq[evtchn] = irq;
 384                 irq_info[irq] = mk_evtchn_info(evtchn);
 385         }
 386
 387         spin_unlock(&irq_mapping_update_lock);
 388
 389         return irq;
 390 }
 391 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
 392
 393 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
 394 {
 395         struct evtchn_bind_ipi bind_ipi;
 396         int evtchn, irq;
 397
 398         spin_lock(&irq_mapping_update_lock);
 399
 400         irq = per_cpu(ipi_to_irq, cpu)[ipi];
 401
 402         if (irq == -1) {
 403                 irq = find_unbound_irq();
 404                 if (irq < 0)
 405                         goto out;
 406
 407                 set_irq_chip_and_handler_name(irq, &xen_percpu_chip,
 408                                               handle_percpu_irq, "ipi");
 409
 410                 bind_ipi.vcpu = cpu;
 411                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
 412                                                 &bind_ipi) != 0)
 413                         BUG();
 414                 evtchn = bind_ipi.port;
 415
 416                 evtchn_to_irq[evtchn] = irq;
 417                 irq_info[irq] = mk_ipi_info(evtchn, ipi);
 418                 per_cpu(ipi_to_irq, cpu)[ipi] = irq;
 419
 420                 bind_evtchn_to_cpu(evtchn, cpu);
 421         }
 422
 423  out:
 424         spin_unlock(&irq_mapping_update_lock);
 425         return irq;
 426 }
 427
 428
 429 static int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
 430 {
 431         struct evtchn_bind_virq bind_virq;
 432         int evtchn, irq;
 433
 434         spin_lock(&irq_mapping_update_lock);
 435
 436         irq = per_cpu(virq_to_irq, cpu)[virq];
 437
 438         if (irq == -1) {
 439                 bind_virq.virq = virq;
 440                 bind_virq.vcpu = cpu;
 441                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 442                                                 &bind_virq) != 0)
 443                         BUG();
 444                 evtchn = bind_virq.port;
 445
 446                 irq = find_unbound_irq();
 447
 448                 set_irq_chip_and_handler_name(irq, &xen_percpu_chip,
 449                                               handle_percpu_irq, "virq");
 450
 451                 evtchn_to_irq[evtchn] = irq;
 452                 irq_info[irq] = mk_virq_info(evtchn, virq);
 453
 454                 per_cpu(virq_to_irq, cpu)[virq] = irq;
 455
 456                 bind_evtchn_to_cpu(evtchn, cpu);
 457         }
 458
 459         spin_unlock(&irq_mapping_update_lock);
 460
 461         return irq;
 462 }
 463
 464 static void unbind_from_irq(unsigned int irq)
 465 {
 466         struct evtchn_close close;
 467         int evtchn = evtchn_from_irq(irq);
 468
 469         spin_lock(&irq_mapping_update_lock);
 470
 471         if (VALID_EVTCHN(evtchn)) {
 472                 close.port = evtchn;
 473                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 474                         BUG();
 475
 476                 switch (type_from_irq(irq)) {
 477                 case IRQT_VIRQ:
 478                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
 479                                 [virq_from_irq(irq)] = -1;
 480                         break;
 481                 case IRQT_IPI:
 482                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
 483                                 [ipi_from_irq(irq)] = -1;
 484                         break;
 485                 default:
 486                         break;
 487                 }
 488
 489                 /* Closed ports are implicitly re-bound to VCPU0. */
 490                 bind_evtchn_to_cpu(evtchn, 0);
 491
 492                 evtchn_to_irq[evtchn] = -1;
 493         }
 494
 495         if (irq_info[irq].type != IRQT_UNBOUND) {
 496                 irq_info[irq] = mk_unbound_info();
 497
 498                 dynamic_irq_cleanup(irq);
 499         }
 500
 501         spin_unlock(&irq_mapping_update_lock);
 502 }
 503
 504 int bind_evtchn_to_irqhandler(unsigned int evtchn,
 505                               irq_handler_t handler,
 506                               unsigned long irqflags,
 507                               const char *devname, void *dev_id)
 508 {
 509         unsigned int irq;
 510         int retval;
 511
 512         irq = bind_evtchn_to_irq(evtchn);
 513         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 514         if (retval != 0) {
 515                 unbind_from_irq(irq);
 516                 return retval;
 517         }
 518
 519         return irq;
 520 }
 521 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
 522
 523 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
 524                             irq_handler_t handler,
 525                             unsigned long irqflags, const char *devname, void *dev_id)
 526 {
 527         unsigned int irq;
 528         int retval;
 529
 530         irq = bind_virq_to_irq(virq, cpu);
 531         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 532         if (retval != 0) {
 533                 unbind_from_irq(irq);
 534                 return retval;
 535         }
 536
 537         return irq;
 538 }
 539 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
 540
 541 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
 542                            unsigned int cpu,
 543                            irq_handler_t handler,
 544                            unsigned long irqflags,
 545                            const char *devname,
 546                            void *dev_id)
 547 {
 548         int irq, retval;
 549
 550         irq = bind_ipi_to_irq(ipi, cpu);
 551         if (irq < 0)
 552                 return irq;
 553
 554         irqflags |= IRQF_NO_SUSPEND;
 555         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 556         if (retval != 0) {
 557                 unbind_from_irq(irq);
 558                 return retval;
 559         }
 560
 561         return irq;
 562 }
 563
 564 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
 565 {
 566         free_irq(irq, dev_id);
 567         unbind_from_irq(irq);
 568 }
 569 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
 570
 571 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
 572 {
 573         int irq = per_cpu(ipi_to_irq, cpu)[vector];
 574         BUG_ON(irq < 0);
 575         notify_remote_via_irq(irq);
 576 }
 577
 578 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
 579 {
 580         struct shared_info *sh = HYPERVISOR_shared_info;
 581         int cpu = smp_processor_id();
 582         int i;
 583         unsigned long flags;
 584         static DEFINE_SPINLOCK(debug_lock);
 585
 586         spin_lock_irqsave(&debug_lock, flags);
 587
 588         printk("vcpu %d\n  ", cpu);
 589
 590         for_each_online_cpu(i) {
 591                 struct vcpu_info *v = per_cpu(xen_vcpu, i);
 592                 printk("%d: masked=%d pending=%d event_sel %08lx\n  ", i,
 593                         (get_irq_regs() && i == cpu) ? xen_irqs_disabled(get_irq_regs()) : v->evtchn_upcall_mask,
 594                         v->evtchn_upcall_pending,
 595                         v->evtchn_pending_sel);
 596         }
 597         printk("pending:\n   ");
 598         for(i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
 599                 printk("%08lx%s", sh->evtchn_pending[i],
 600                         i % 8 == 0 ? "\n   " : " ");
 601         printk("\nmasks:\n   ");
 602         for(i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
 603                 printk("%08lx%s", sh->evtchn_mask[i],
 604                         i % 8 == 0 ? "\n   " : " ");
 605
 606         printk("\nunmasked:\n   ");
 607         for(i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
 608                 printk("%08lx%s", sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
 609                         i % 8 == 0 ? "\n   " : " ");
 610
 611         printk("\npending list:\n");
 612         for(i = 0; i < NR_EVENT_CHANNELS; i++) {
 613                 if (sync_test_bit(i, sh->evtchn_pending)) {
 614                         printk("  %d: event %d -> irq %d\n",
 615                                cpu_from_evtchn(i), i,
 616                                evtchn_to_irq[i]);
 617                 }
 618         }
 619
 620         spin_unlock_irqrestore(&debug_lock, flags);
 621
 622         return IRQ_HANDLED;
 623 }
 624
 625 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
 626
 627 /*
 628  * Search the CPUs pending events bitmasks.  For each one found, map
 629  * the event number to an irq, and feed it into do_IRQ() for
 630  * handling.
 631  *
 632  * Xen uses a two-level bitmap to speed searching.  The first level is
 633  * a bitset of words which contain pending event bits.  The second
 634  * level is a bitset of pending events themselves.
 635  */
 636 static void __xen_evtchn_do_upcall(void)
 637 {
 638         int cpu = get_cpu();
 639         struct shared_info *s = HYPERVISOR_shared_info;
 640         struct vcpu_info *vcpu_info = __get_cpu_var(xen_vcpu);
 641         unsigned count;
 642
 643         do {
 644                 unsigned long pending_words;
 645
 646                 vcpu_info->evtchn_upcall_pending = 0;
 647
 648                 if (__get_cpu_var(xed_nesting_count)++)
 649                         goto out;
 650
 651 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
 652                 /* Clear master flag /before/ clearing selector flag. */
 653                 wmb();
 654 #endif
 655                 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
 656                 while (pending_words != 0) {
 657                         unsigned long pending_bits;
 658                         int word_idx = __ffs(pending_words);
 659                         pending_words &= ~(1UL << word_idx);
 660
 661                         while ((pending_bits = active_evtchns(cpu, s, word_idx)) != 0) {
 662                                 int bit_idx = __ffs(pending_bits);
 663                                 int port = (word_idx * BITS_PER_LONG) + bit_idx;
 664                                 int irq = evtchn_to_irq[port];
 665                                 struct irq_desc *desc;
 666
 667                                 if (irq != -1) {
 668                                         desc = irq_to_desc(irq);
 669                                         if (desc)
 670                                                 generic_handle_irq_desc(irq, desc);
 671                                 }
 672                         }
 673                 }
 674
 675                 BUG_ON(!irqs_disabled());
 676
 677                 count = __get_cpu_var(xed_nesting_count);
 678                 __get_cpu_var(xed_nesting_count) = 0;
 679         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
 680
 681 out:
 682
 683         put_cpu();
 684 }
 685
 686 void xen_evtchn_do_upcall(struct pt_regs *regs)
 687 {
 688         struct pt_regs *old_regs = set_irq_regs(regs);
 689
 690         exit_idle();
 691         irq_enter();
 692
 693         __xen_evtchn_do_upcall();
 694
 695         irq_exit();
 696         set_irq_regs(old_regs);
 697 }
 698
 699 void xen_hvm_evtchn_do_upcall(void)
 700 {
 701         __xen_evtchn_do_upcall();
 702 }
 703 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
 704
 705 /* Rebind a new event channel to an existing irq. */
 706 void rebind_evtchn_irq(int evtchn, int irq)
 707 {
 708         struct irq_info *info = info_for_irq(irq);
 709
 710         /* Make sure the irq is masked, since the new event channel
 711            will also be masked. */
 712         disable_irq(irq);
 713
 714         spin_lock(&irq_mapping_update_lock);
 715
 716         /* After resume the irq<->evtchn mappings are all cleared out */
 717         BUG_ON(evtchn_to_irq[evtchn] != -1);
 718         /* Expect irq to have been bound before,
 719            so there should be a proper type */
 720         BUG_ON(info->type == IRQT_UNBOUND);
 721
 722         evtchn_to_irq[evtchn] = irq;
 723         irq_info[irq] = mk_evtchn_info(evtchn);
 724
 725         spin_unlock(&irq_mapping_update_lock);
 726
 727         /* new event channels are always bound to cpu 0 */
 728         irq_set_affinity(irq, cpumask_of(0));
 729
 730         /* Unmask the event channel. */
 731         enable_irq(irq);
 732 }
 733
 734 /* Rebind an evtchn so that it gets delivered to a specific cpu */
 735 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
 736 {
 737         struct evtchn_bind_vcpu bind_vcpu;
 738         int evtchn = evtchn_from_irq(irq);
 739
 740         /* events delivered via platform PCI interrupts are always
 741          * routed to vcpu 0 */
 742         if (!VALID_EVTCHN(evtchn) ||
 743                 (xen_hvm_domain() && !xen_have_vector_callback))
 744                 return -1;
 745
 746         /* Send future instances of this interrupt to other vcpu. */
 747         bind_vcpu.port = evtchn;
 748         bind_vcpu.vcpu = tcpu;
 749
 750         /*
 751          * If this fails, it usually just indicates that we're dealing with a
 752          * virq or IPI channel, which don't actually need to be rebound. Ignore
 753          * it, but don't do the xenlinux-level rebind in that case.
 754          */
 755         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
 756                 bind_evtchn_to_cpu(evtchn, tcpu);
 757
 758         return 0;
 759 }
 760
 761 static int set_affinity_irq(unsigned irq, const struct cpumask *dest)
 762 {
 763         unsigned tcpu = cpumask_first(dest);
 764
 765         return rebind_irq_to_cpu(irq, tcpu);
 766 }
 767
 768 int resend_irq_on_evtchn(unsigned int irq)
 769 {
 770         int masked, evtchn = evtchn_from_irq(irq);
 771         struct shared_info *s = HYPERVISOR_shared_info;
 772
 773         if (!VALID_EVTCHN(evtchn))
 774                 return 1;
 775
 776         masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
 777         sync_set_bit(evtchn, s->evtchn_pending);
 778         if (!masked)
 779                 unmask_evtchn(evtchn);
 780
 781         return 1;
 782 }
 783
 784 static void enable_dynirq(unsigned int irq)
 785 {
 786         int evtchn = evtchn_from_irq(irq);
 787
 788         if (VALID_EVTCHN(evtchn))
 789                 unmask_evtchn(evtchn);
 790 }
 791
 792 static void disable_dynirq(unsigned int irq)
 793 {
 794         int evtchn = evtchn_from_irq(irq);
 795
 796         if (VALID_EVTCHN(evtchn))
 797                 mask_evtchn(evtchn);
 798 }
 799
 800 static void ack_dynirq(unsigned int irq)
 801 {
 802         int evtchn = evtchn_from_irq(irq);
 803
 804         move_native_irq(irq);
 805
 806         if (VALID_EVTCHN(evtchn))
 807                 clear_evtchn(evtchn);
 808 }
 809
 810 static int retrigger_dynirq(unsigned int irq)
 811 {
 812         int evtchn = evtchn_from_irq(irq);
 813         struct shared_info *sh = HYPERVISOR_shared_info;
 814         int ret = 0;
 815
 816         if (VALID_EVTCHN(evtchn)) {
 817                 int masked;
 818
 819                 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
 820                 sync_set_bit(evtchn, sh->evtchn_pending);
 821                 if (!masked)
 822                         unmask_evtchn(evtchn);
 823                 ret = 1;
 824         }
 825
 826         return ret;
 827 }
 828
 829 static void restore_cpu_virqs(unsigned int cpu)
 830 {
 831         struct evtchn_bind_virq bind_virq;
 832         int virq, irq, evtchn;
 833
 834         for (virq = 0; virq < NR_VIRQS; virq++) {
 835                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
 836                         continue;
 837
 838                 BUG_ON(virq_from_irq(irq) != virq);
 839
 840                 /* Get a new binding from Xen. */
 841                 bind_virq.virq = virq;
 842                 bind_virq.vcpu = cpu;
 843                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 844                                                 &bind_virq) != 0)
 845                         BUG();
 846                 evtchn = bind_virq.port;
 847
 848                 /* Record the new mapping. */
 849                 evtchn_to_irq[evtchn] = irq;
 850                 irq_info[irq] = mk_virq_info(evtchn, virq);
 851                 bind_evtchn_to_cpu(evtchn, cpu);
 852
 853                 /* Ready for use. */
 854                 unmask_evtchn(evtchn);
 855         }
 856 }
 857
 858 static void restore_cpu_ipis(unsigned int cpu)
 859 {
 860         struct evtchn_bind_ipi bind_ipi;
 861         int ipi, irq, evtchn;
 862
 863         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
 864                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
 865                         continue;
 866
 867                 BUG_ON(ipi_from_irq(irq) != ipi);
 868
 869                 /* Get a new binding from Xen. */
 870                 bind_ipi.vcpu = cpu;
 871                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
 872                                                 &bind_ipi) != 0)
 873                         BUG();
 874                 evtchn = bind_ipi.port;
 875
 876                 /* Record the new mapping. */
 877                 evtchn_to_irq[evtchn] = irq;
 878                 irq_info[irq] = mk_ipi_info(evtchn, ipi);
 879                 bind_evtchn_to_cpu(evtchn, cpu);
 880
 881                 /* Ready for use. */
 882                 unmask_evtchn(evtchn);
 883
 884         }
 885 }
 886
 887 /* Clear an irq's pending state, in preparation for polling on it */
 888 void xen_clear_irq_pending(int irq)
 889 {
 890         int evtchn = evtchn_from_irq(irq);
 891
 892         if (VALID_EVTCHN(evtchn))
 893                 clear_evtchn(evtchn);
 894 }
 895
 896 void xen_set_irq_pending(int irq)
 897 {
 898         int evtchn = evtchn_from_irq(irq);
 899
 900         if (VALID_EVTCHN(evtchn))
 901                 set_evtchn(evtchn);
 902 }
 903
 904 bool xen_test_irq_pending(int irq)
 905 {
 906         int evtchn = evtchn_from_irq(irq);
 907         bool ret = false;
 908
 909         if (VALID_EVTCHN(evtchn))
 910                 ret = test_evtchn(evtchn);
 911
 912         return ret;
 913 }
 914
 915 /* Poll waiting for an irq to become pending.  In the usual case, the
 916    irq will be disabled so it won't deliver an interrupt. */
 917 void xen_poll_irq(int irq)
 918 {
 919         evtchn_port_t evtchn = evtchn_from_irq(irq);
 920
 921         if (VALID_EVTCHN(evtchn)) {
 922                 struct sched_poll poll;
 923
 924                 poll.nr_ports = 1;
 925                 poll.timeout = 0;
 926                 set_xen_guest_handle(poll.ports, &evtchn);
 927
 928                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
 929                         BUG();
 930         }
 931 }
 932
 933 void xen_irq_resume(void)
 934 {
 935         unsigned int cpu, irq, evtchn;
 936
 937         init_evtchn_cpu_bindings();
 938
 939         /* New event-channel space is not 'live' yet. */
 940         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
 941                 mask_evtchn(evtchn);
 942
 943         /* No IRQ <-> event-channel mappings. */
 944         for (irq = 0; irq < nr_irqs; irq++)
 945                 irq_info[irq].evtchn = 0; /* zap event-channel binding */
 946
 947         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
 948                 evtchn_to_irq[evtchn] = -1;
 949
 950         for_each_possible_cpu(cpu) {
 951                 restore_cpu_virqs(cpu);
 952                 restore_cpu_ipis(cpu);
 953         }
 954 }
 955
 956 static struct irq_chip xen_dynamic_chip __read_mostly = {
 957         .name           = "xen-dyn",
 958
 959         .disable        = disable_dynirq,
 960         .mask           = disable_dynirq,
 961         .unmask         = enable_dynirq,
 962
 963         .ack            = ack_dynirq,
 964         .set_affinity   = set_affinity_irq,
 965         .retrigger      = retrigger_dynirq,
 966 };
 967
 968 static struct irq_chip xen_percpu_chip __read_mostly = {
 969         .name           = "xen-percpu",
 970
 971         .disable        = disable_dynirq,
 972         .mask           = disable_dynirq,
 973         .unmask         = enable_dynirq,
 974
 975         .ack            = ack_dynirq,
 976 };
 977
 978 int xen_set_callback_via(uint64_t via)
 979 {
 980         struct xen_hvm_param a;
 981         a.domid = DOMID_SELF;
 982         a.index = HVM_PARAM_CALLBACK_IRQ;
 983         a.value = via;
 984         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
 985 }
 986 EXPORT_SYMBOL_GPL(xen_set_callback_via);
 987
 988 #ifdef CONFIG_XEN_PVHVM
 989 /* Vector callbacks are better than PCI interrupts to receive event
 990  * channel notifications because we can receive vector callbacks on any
 991  * vcpu and we don't need PCI support or APIC interactions. */
 992 void xen_callback_vector(void)
 993 {
 994         int rc;
 995         uint64_t callback_via;
 996         if (xen_have_vector_callback) {
 997                 callback_via = HVM_CALLBACK_VECTOR(XEN_HVM_EVTCHN_CALLBACK);
 998                 rc = xen_set_callback_via(callback_via);
 999                 if (rc) {
1000                         printk(KERN_ERR "Request for Xen HVM callback vector"
1001                                         " failed.\n");
1002                         xen_have_vector_callback = 0;
1003                         return;
1004                 }
1005                 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1006                                 "enabled\n");
1007                 /* in the restore case the vector has already been allocated */
1008                 if (!test_bit(XEN_HVM_EVTCHN_CALLBACK, used_vectors))
1009                         alloc_intr_gate(XEN_HVM_EVTCHN_CALLBACK, xen_hvm_callback_vector);
1010         }
1011 }
1012 #else
1013 void xen_callback_vector(void) {}
1014 #endif
1015
1016 void __init xen_init_IRQ(void)
1017 {
1018         int i;
1019
1020         cpu_evtchn_mask_p = kcalloc(nr_cpu_ids, sizeof(struct cpu_evtchn_s),
1021                                     GFP_KERNEL);
1022         BUG_ON(cpu_evtchn_mask_p == NULL);
1023
1024         init_evtchn_cpu_bindings();
1025
1026         /* No event channels are 'live' right now. */
1027         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1028                 mask_evtchn(i);
1029
1030         if (xen_hvm_domain()) {
1031                 xen_callback_vector();
1032                 native_init_IRQ();
1033         } else {
1034                 irq_ctx_init(smp_processor_id());
1035         }
1036 }