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         struct irq_data *data;
 342         int irq, res;
 343
 344         for (irq = 0; irq < nr_irqs; irq++) {
 345                 data = irq_get_irq_data(irq);
 346                 /* only 0->15 have init'd desc; handle irq > 16 */
 347                 if (!data)
 348                         break;
 349                 if (data->chip == &no_irq_chip)
 350                         break;
 351                 if (data->chip != &xen_dynamic_chip)
 352                         continue;
 353                 if (irq_info[irq].type == IRQT_UNBOUND)
 354                         return irq;
 355         }
 356
 357         if (irq == nr_irqs)
 358                 panic("No available IRQ to bind to: increase nr_irqs!\n");
 359
 360         res = irq_alloc_desc_at(irq, 0);
 361
 362         if (WARN_ON(res != irq))
 363                 return -1;
 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_fasteoi_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_percpu_chip,
 407                                               handle_percpu_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                 irq = find_unbound_irq();
 439
 440                 set_irq_chip_and_handler_name(irq, &xen_percpu_chip,
 441                                               handle_percpu_irq, "virq");
 442
 443                 bind_virq.virq = virq;
 444                 bind_virq.vcpu = cpu;
 445                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 446                                                 &bind_virq) != 0)
 447                         BUG();
 448                 evtchn = bind_virq.port;
 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                 irq_free_desc(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         unsigned long *cpu_evtchn = cpu_evtchn_mask(cpu);
 582         int i;
 583         unsigned long flags;
 584         static DEFINE_SPINLOCK(debug_lock);
 585         struct vcpu_info *v;
 586
 587         spin_lock_irqsave(&debug_lock, flags);
 588
 589         printk("\nvcpu %d\n  ", cpu);
 590
 591         for_each_online_cpu(i) {
 592                 int pending;
 593                 v = per_cpu(xen_vcpu, i);
 594                 pending = (get_irq_regs() && i == cpu)
 595                         ? xen_irqs_disabled(get_irq_regs())
 596                         : v->evtchn_upcall_mask;
 597                 printk("%d: masked=%d pending=%d event_sel %0*lx\n  ", i,
 598                        pending, v->evtchn_upcall_pending,
 599                        (int)(sizeof(v->evtchn_pending_sel)*2),
 600                        v->evtchn_pending_sel);
 601         }
 602         v = per_cpu(xen_vcpu, cpu);
 603
 604         printk("\npending:\n   ");
 605         for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
 606                 printk("%0*lx%s", (int)sizeof(sh->evtchn_pending[0])*2,
 607                        sh->evtchn_pending[i],
 608                        i % 8 == 0 ? "\n   " : " ");
 609         printk("\nglobal mask:\n   ");
 610         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
 611                 printk("%0*lx%s",
 612                        (int)(sizeof(sh->evtchn_mask[0])*2),
 613                        sh->evtchn_mask[i],
 614                        i % 8 == 0 ? "\n   " : " ");
 615
 616         printk("\nglobally unmasked:\n   ");
 617         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
 618                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
 619                        sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
 620                        i % 8 == 0 ? "\n   " : " ");
 621
 622         printk("\nlocal cpu%d mask:\n   ", cpu);
 623         for (i = (NR_EVENT_CHANNELS/BITS_PER_LONG)-1; i >= 0; i--)
 624                 printk("%0*lx%s", (int)(sizeof(cpu_evtchn[0])*2),
 625                        cpu_evtchn[i],
 626                        i % 8 == 0 ? "\n   " : " ");
 627
 628         printk("\nlocally unmasked:\n   ");
 629         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
 630                 unsigned long pending = sh->evtchn_pending[i]
 631                         & ~sh->evtchn_mask[i]
 632                         & cpu_evtchn[i];
 633                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
 634                        pending, i % 8 == 0 ? "\n   " : " ");
 635         }
 636
 637         printk("\npending list:\n");
 638         for (i = 0; i < NR_EVENT_CHANNELS; i++) {
 639                 if (sync_test_bit(i, sh->evtchn_pending)) {
 640                         int word_idx = i / BITS_PER_LONG;
 641                         printk("  %d: event %d -> irq %d%s%s%s\n",
 642                                cpu_from_evtchn(i), i,
 643                                evtchn_to_irq[i],
 644                                sync_test_bit(word_idx, &v->evtchn_pending_sel)
 645                                              ? "" : " l2-clear",
 646                                !sync_test_bit(i, sh->evtchn_mask)
 647                                              ? "" : " globally-masked",
 648                                sync_test_bit(i, cpu_evtchn)
 649                                              ? "" : " locally-masked");
 650                 }
 651         }
 652
 653         spin_unlock_irqrestore(&debug_lock, flags);
 654
 655         return IRQ_HANDLED;
 656 }
 657
 658 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
 659
 660 /*
 661  * Search the CPUs pending events bitmasks.  For each one found, map
 662  * the event number to an irq, and feed it into do_IRQ() for
 663  * handling.
 664  *
 665  * Xen uses a two-level bitmap to speed searching.  The first level is
 666  * a bitset of words which contain pending event bits.  The second
 667  * level is a bitset of pending events themselves.
 668  */
 669 static void __xen_evtchn_do_upcall(void)
 670 {
 671         int cpu = get_cpu();
 672         struct shared_info *s = HYPERVISOR_shared_info;
 673         struct vcpu_info *vcpu_info = __get_cpu_var(xen_vcpu);
 674         unsigned count;
 675
 676         do {
 677                 unsigned long pending_words;
 678
 679                 vcpu_info->evtchn_upcall_pending = 0;
 680
 681                 if (__get_cpu_var(xed_nesting_count)++)
 682                         goto out;
 683
 684 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
 685                 /* Clear master flag /before/ clearing selector flag. */
 686                 wmb();
 687 #endif
 688                 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
 689                 while (pending_words != 0) {
 690                         unsigned long pending_bits;
 691                         int word_idx = __ffs(pending_words);
 692                         pending_words &= ~(1UL << word_idx);
 693
 694                         while ((pending_bits = active_evtchns(cpu, s, word_idx)) != 0) {
 695                                 int bit_idx = __ffs(pending_bits);
 696                                 int port = (word_idx * BITS_PER_LONG) + bit_idx;
 697                                 int irq = evtchn_to_irq[port];
 698                                 struct irq_desc *desc;
 699
 700                                 mask_evtchn(port);
 701                                 clear_evtchn(port);
 702
 703                                 if (irq != -1) {
 704                                         desc = irq_to_desc(irq);
 705                                         if (desc)
 706                                                 generic_handle_irq_desc(irq, desc);
 707                                 }
 708                         }
 709                 }
 710
 711                 BUG_ON(!irqs_disabled());
 712
 713                 count = __get_cpu_var(xed_nesting_count);
 714                 __get_cpu_var(xed_nesting_count) = 0;
 715         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
 716
 717 out:
 718
 719         put_cpu();
 720 }
 721
 722 void xen_evtchn_do_upcall(struct pt_regs *regs)
 723 {
 724         struct pt_regs *old_regs = set_irq_regs(regs);
 725
 726         exit_idle();
 727         irq_enter();
 728
 729         __xen_evtchn_do_upcall();
 730
 731         irq_exit();
 732         set_irq_regs(old_regs);
 733 }
 734
 735 void xen_hvm_evtchn_do_upcall(void)
 736 {
 737         __xen_evtchn_do_upcall();
 738 }
 739 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
 740
 741 /* Rebind a new event channel to an existing irq. */
 742 void rebind_evtchn_irq(int evtchn, int irq)
 743 {
 744         struct irq_info *info = info_for_irq(irq);
 745
 746         /* Make sure the irq is masked, since the new event channel
 747            will also be masked. */
 748         disable_irq(irq);
 749
 750         spin_lock(&irq_mapping_update_lock);
 751
 752         /* After resume the irq<->evtchn mappings are all cleared out */
 753         BUG_ON(evtchn_to_irq[evtchn] != -1);
 754         /* Expect irq to have been bound before,
 755            so there should be a proper type */
 756         BUG_ON(info->type == IRQT_UNBOUND);
 757
 758         evtchn_to_irq[evtchn] = irq;
 759         irq_info[irq] = mk_evtchn_info(evtchn);
 760
 761         spin_unlock(&irq_mapping_update_lock);
 762
 763         /* new event channels are always bound to cpu 0 */
 764         irq_set_affinity(irq, cpumask_of(0));
 765
 766         /* Unmask the event channel. */
 767         enable_irq(irq);
 768 }
 769
 770 /* Rebind an evtchn so that it gets delivered to a specific cpu */
 771 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
 772 {
 773         struct evtchn_bind_vcpu bind_vcpu;
 774         int evtchn = evtchn_from_irq(irq);
 775
 776         /* events delivered via platform PCI interrupts are always
 777          * routed to vcpu 0 */
 778         if (!VALID_EVTCHN(evtchn) ||
 779                 (xen_hvm_domain() && !xen_have_vector_callback))
 780                 return -1;
 781
 782         /* Send future instances of this interrupt to other vcpu. */
 783         bind_vcpu.port = evtchn;
 784         bind_vcpu.vcpu = tcpu;
 785
 786         /*
 787          * If this fails, it usually just indicates that we're dealing with a
 788          * virq or IPI channel, which don't actually need to be rebound. Ignore
 789          * it, but don't do the xenlinux-level rebind in that case.
 790          */
 791         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
 792                 bind_evtchn_to_cpu(evtchn, tcpu);
 793
 794         return 0;
 795 }
 796
 797 static int set_affinity_irq(unsigned irq, const struct cpumask *dest)
 798 {
 799         unsigned tcpu = cpumask_first(dest);
 800
 801         return rebind_irq_to_cpu(irq, tcpu);
 802 }
 803
 804 int resend_irq_on_evtchn(unsigned int irq)
 805 {
 806         int masked, evtchn = evtchn_from_irq(irq);
 807         struct shared_info *s = HYPERVISOR_shared_info;
 808
 809         if (!VALID_EVTCHN(evtchn))
 810                 return 1;
 811
 812         masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
 813         sync_set_bit(evtchn, s->evtchn_pending);
 814         if (!masked)
 815                 unmask_evtchn(evtchn);
 816
 817         return 1;
 818 }
 819
 820 static void enable_dynirq(unsigned int irq)
 821 {
 822         int evtchn = evtchn_from_irq(irq);
 823
 824         if (VALID_EVTCHN(evtchn))
 825                 unmask_evtchn(evtchn);
 826 }
 827
 828 static void disable_dynirq(unsigned int irq)
 829 {
 830         int evtchn = evtchn_from_irq(irq);
 831
 832         if (VALID_EVTCHN(evtchn))
 833                 mask_evtchn(evtchn);
 834 }
 835
 836 static void ack_dynirq(unsigned int irq)
 837 {
 838         int evtchn = evtchn_from_irq(irq);
 839
 840         move_masked_irq(irq);
 841
 842         if (VALID_EVTCHN(evtchn))
 843                 unmask_evtchn(evtchn);
 844 }
 845
 846 static int retrigger_dynirq(unsigned int irq)
 847 {
 848         int evtchn = evtchn_from_irq(irq);
 849         struct shared_info *sh = HYPERVISOR_shared_info;
 850         int ret = 0;
 851
 852         if (VALID_EVTCHN(evtchn)) {
 853                 int masked;
 854
 855                 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
 856                 sync_set_bit(evtchn, sh->evtchn_pending);
 857                 if (!masked)
 858                         unmask_evtchn(evtchn);
 859                 ret = 1;
 860         }
 861
 862         return ret;
 863 }
 864
 865 static void restore_cpu_virqs(unsigned int cpu)
 866 {
 867         struct evtchn_bind_virq bind_virq;
 868         int virq, irq, evtchn;
 869
 870         for (virq = 0; virq < NR_VIRQS; virq++) {
 871                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
 872                         continue;
 873
 874                 BUG_ON(virq_from_irq(irq) != virq);
 875
 876                 /* Get a new binding from Xen. */
 877                 bind_virq.virq = virq;
 878                 bind_virq.vcpu = cpu;
 879                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 880                                                 &bind_virq) != 0)
 881                         BUG();
 882                 evtchn = bind_virq.port;
 883
 884                 /* Record the new mapping. */
 885                 evtchn_to_irq[evtchn] = irq;
 886                 irq_info[irq] = mk_virq_info(evtchn, virq);
 887                 bind_evtchn_to_cpu(evtchn, cpu);
 888
 889                 /* Ready for use. */
 890                 unmask_evtchn(evtchn);
 891         }
 892 }
 893
 894 static void restore_cpu_ipis(unsigned int cpu)
 895 {
 896         struct evtchn_bind_ipi bind_ipi;
 897         int ipi, irq, evtchn;
 898
 899         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
 900                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
 901                         continue;
 902
 903                 BUG_ON(ipi_from_irq(irq) != ipi);
 904
 905                 /* Get a new binding from Xen. */
 906                 bind_ipi.vcpu = cpu;
 907                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
 908                                                 &bind_ipi) != 0)
 909                         BUG();
 910                 evtchn = bind_ipi.port;
 911
 912                 /* Record the new mapping. */
 913                 evtchn_to_irq[evtchn] = irq;
 914                 irq_info[irq] = mk_ipi_info(evtchn, ipi);
 915                 bind_evtchn_to_cpu(evtchn, cpu);
 916
 917                 /* Ready for use. */
 918                 unmask_evtchn(evtchn);
 919
 920         }
 921 }
 922
 923 /* Clear an irq's pending state, in preparation for polling on it */
 924 void xen_clear_irq_pending(int irq)
 925 {
 926         int evtchn = evtchn_from_irq(irq);
 927
 928         if (VALID_EVTCHN(evtchn))
 929                 clear_evtchn(evtchn);
 930 }
 931
 932 void xen_set_irq_pending(int irq)
 933 {
 934         int evtchn = evtchn_from_irq(irq);
 935
 936         if (VALID_EVTCHN(evtchn))
 937                 set_evtchn(evtchn);
 938 }
 939
 940 bool xen_test_irq_pending(int irq)
 941 {
 942         int evtchn = evtchn_from_irq(irq);
 943         bool ret = false;
 944
 945         if (VALID_EVTCHN(evtchn))
 946                 ret = test_evtchn(evtchn);
 947
 948         return ret;
 949 }
 950
 951 /* Poll waiting for an irq to become pending.  In the usual case, the
 952    irq will be disabled so it won't deliver an interrupt. */
 953 void xen_poll_irq(int irq)
 954 {
 955         evtchn_port_t evtchn = evtchn_from_irq(irq);
 956
 957         if (VALID_EVTCHN(evtchn)) {
 958                 struct sched_poll poll;
 959
 960                 poll.nr_ports = 1;
 961                 poll.timeout = 0;
 962                 set_xen_guest_handle(poll.ports, &evtchn);
 963
 964                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
 965                         BUG();
 966         }
 967 }
 968
 969 void xen_irq_resume(void)
 970 {
 971         unsigned int cpu, irq, evtchn;
 972
 973         init_evtchn_cpu_bindings();
 974
 975         /* New event-channel space is not 'live' yet. */
 976         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
 977                 mask_evtchn(evtchn);
 978
 979         /* No IRQ <-> event-channel mappings. */
 980         for (irq = 0; irq < nr_irqs; irq++)
 981                 irq_info[irq].evtchn = 0; /* zap event-channel binding */
 982
 983         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
 984                 evtchn_to_irq[evtchn] = -1;
 985
 986         for_each_possible_cpu(cpu) {
 987                 restore_cpu_virqs(cpu);
 988                 restore_cpu_ipis(cpu);
 989         }
 990 }
 991
 992 static struct irq_chip xen_dynamic_chip __read_mostly = {
 993         .name           = "xen-dyn",
 994
 995         .disable        = disable_dynirq,
 996         .mask           = disable_dynirq,
 997         .unmask         = enable_dynirq,
 998
 999         .eoi            = ack_dynirq,
1000         .set_affinity   = set_affinity_irq,
1001         .retrigger      = retrigger_dynirq,
1002 };
1003
1004 static struct irq_chip xen_percpu_chip __read_mostly = {
1005         .name           = "xen-percpu",
1006
1007         .disable        = disable_dynirq,
1008         .mask           = disable_dynirq,
1009         .unmask         = enable_dynirq,
1010
1011         .ack            = ack_dynirq,
1012 };
1013
1014 int xen_set_callback_via(uint64_t via)
1015 {
1016         struct xen_hvm_param a;
1017         a.domid = DOMID_SELF;
1018         a.index = HVM_PARAM_CALLBACK_IRQ;
1019         a.value = via;
1020         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1021 }
1022 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1023
1024 #ifdef CONFIG_XEN_PVHVM
1025 /* Vector callbacks are better than PCI interrupts to receive event
1026  * channel notifications because we can receive vector callbacks on any
1027  * vcpu and we don't need PCI support or APIC interactions. */
1028 void xen_callback_vector(void)
1029 {
1030         int rc;
1031         uint64_t callback_via;
1032         if (xen_have_vector_callback) {
1033                 callback_via = HVM_CALLBACK_VECTOR(XEN_HVM_EVTCHN_CALLBACK);
1034                 rc = xen_set_callback_via(callback_via);
1035                 if (rc) {
1036                         printk(KERN_ERR "Request for Xen HVM callback vector"
1037                                         " failed.\n");
1038                         xen_have_vector_callback = 0;
1039                         return;
1040                 }
1041                 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1042                                 "enabled\n");
1043                 /* in the restore case the vector has already been allocated */
1044                 if (!test_bit(XEN_HVM_EVTCHN_CALLBACK, used_vectors))
1045                         alloc_intr_gate(XEN_HVM_EVTCHN_CALLBACK, xen_hvm_callback_vector);
1046         }
1047 }
1048 #else
1049 void xen_callback_vector(void) {}
1050 #endif
1051
1052 void __init xen_init_IRQ(void)
1053 {
1054         int i;
1055
1056         cpu_evtchn_mask_p = kcalloc(nr_cpu_ids, sizeof(struct cpu_evtchn_s),
1057                                     GFP_KERNEL);
1058         BUG_ON(cpu_evtchn_mask_p == NULL);
1059
1060         init_evtchn_cpu_bindings();
1061
1062         /* No event channels are 'live' right now. */
1063         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1064                 mask_evtchn(i);
1065
1066         if (xen_hvm_domain()) {
1067                 xen_callback_vector();
1068                 native_init_IRQ();
1069         } else {
1070                 irq_ctx_init(smp_processor_id());
1071         }
1072 }