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