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. PIRQs - Hardware interrupts.
  20  *
  21  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
  22  */
  23
  24 #include <linux/linkage.h>
  25 #include <linux/interrupt.h>
  26 #include <linux/irq.h>
  27 #include <linux/module.h>
  28 #include <linux/string.h>
  29 #include <linux/bootmem.h>
  30 #include <linux/slab.h>
  31 #include <linux/irqnr.h>
  32
  33 #include <asm/desc.h>
  34 #include <asm/ptrace.h>
  35 #include <asm/irq.h>
  36 #include <asm/idle.h>
  37 #include <asm/io_apic.h>
  38 #include <asm/sync_bitops.h>
  39 #include <asm/xen/hypercall.h>
  40 #include <asm/xen/hypervisor.h>
  41
  42 #include <xen/xen.h>
  43 #include <xen/hvm.h>
  44 #include <xen/xen-ops.h>
  45 #include <xen/events.h>
  46 #include <xen/interface/xen.h>
  47 #include <xen/interface/event_channel.h>
  48 #include <xen/interface/hvm/hvm_op.h>
  49 #include <xen/interface/hvm/params.h>
  50
  51 /*
  52  * This lock protects updates to the following mapping and reference-count
  53  * arrays. The lock does not need to be acquired to read the mapping tables.
  54  */
  55 static DEFINE_SPINLOCK(irq_mapping_update_lock);
  56
  57 /* IRQ <-> VIRQ mapping. */
  58 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
  59
  60 /* IRQ <-> IPI mapping */
  61 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
  62
  63 /* Interrupt types. */
  64 enum xen_irq_type {
  65         IRQT_UNBOUND = 0,
  66         IRQT_PIRQ,
  67         IRQT_VIRQ,
  68         IRQT_IPI,
  69         IRQT_EVTCHN
  70 };
  71
  72 /*
  73  * Packed IRQ information:
  74  * type - enum xen_irq_type
  75  * event channel - irq->event channel mapping
  76  * cpu - cpu this event channel is bound to
  77  * index - type-specific information:
  78  *    PIRQ - vector, with MSB being "needs EIO"
  79  *    VIRQ - virq number
  80  *    IPI - IPI vector
  81  *    EVTCHN -
  82  */
  83 struct irq_info
  84 {
  85         enum xen_irq_type type; /* type */
  86         unsigned short evtchn;  /* event channel */
  87         unsigned short cpu;     /* cpu bound */
  88
  89         union {
  90                 unsigned short virq;
  91                 enum ipi_vector ipi;
  92                 struct {
  93                         unsigned short gsi;
  94                         unsigned char vector;
  95                         unsigned char flags;
  96                 } pirq;
  97         } u;
  98 };
  99 #define PIRQ_NEEDS_EOI  (1 << 0)
 100 #define PIRQ_SHAREABLE  (1 << 1)
 101
 102 static struct irq_info *irq_info;
 103
 104 static int *evtchn_to_irq;
 105 struct cpu_evtchn_s {
 106         unsigned long bits[NR_EVENT_CHANNELS/BITS_PER_LONG];
 107 };
 108
 109 static __initdata struct cpu_evtchn_s init_evtchn_mask = {
 110         .bits[0 ... (NR_EVENT_CHANNELS/BITS_PER_LONG)-1] = ~0ul,
 111 };
 112 static struct cpu_evtchn_s *cpu_evtchn_mask_p = &init_evtchn_mask;
 113
 114 static inline unsigned long *cpu_evtchn_mask(int cpu)
 115 {
 116         return cpu_evtchn_mask_p[cpu].bits;
 117 }
 118
 119 /* Xen will never allocate port zero for any purpose. */
 120 #define VALID_EVTCHN(chn)       ((chn) != 0)
 121
 122 static struct irq_chip xen_dynamic_chip;
 123 static struct irq_chip xen_percpu_chip;
 124 static struct irq_chip xen_pirq_chip;
 125
 126 /* Constructor for packed IRQ information. */
 127 static struct irq_info mk_unbound_info(void)
 128 {
 129         return (struct irq_info) { .type = IRQT_UNBOUND };
 130 }
 131
 132 static struct irq_info mk_evtchn_info(unsigned short evtchn)
 133 {
 134         return (struct irq_info) { .type = IRQT_EVTCHN, .evtchn = evtchn,
 135                         .cpu = 0 };
 136 }
 137
 138 static struct irq_info mk_ipi_info(unsigned short evtchn, enum ipi_vector ipi)
 139 {
 140         return (struct irq_info) { .type = IRQT_IPI, .evtchn = evtchn,
 141                         .cpu = 0, .u.ipi = ipi };
 142 }
 143
 144 static struct irq_info mk_virq_info(unsigned short evtchn, unsigned short virq)
 145 {
 146         return (struct irq_info) { .type = IRQT_VIRQ, .evtchn = evtchn,
 147                         .cpu = 0, .u.virq = virq };
 148 }
 149
 150 static struct irq_info mk_pirq_info(unsigned short evtchn,
 151                                     unsigned short gsi, unsigned short vector)
 152 {
 153         return (struct irq_info) { .type = IRQT_PIRQ, .evtchn = evtchn,
 154                         .cpu = 0, .u.pirq = { .gsi = gsi, .vector = vector } };
 155 }
 156
 157 /*
 158  * Accessors for packed IRQ information.
 159  */
 160 static struct irq_info *info_for_irq(unsigned irq)
 161 {
 162         return &irq_info[irq];
 163 }
 164
 165 static unsigned int evtchn_from_irq(unsigned irq)
 166 {
 167         return info_for_irq(irq)->evtchn;
 168 }
 169
 170 unsigned irq_from_evtchn(unsigned int evtchn)
 171 {
 172         return evtchn_to_irq[evtchn];
 173 }
 174 EXPORT_SYMBOL_GPL(irq_from_evtchn);
 175
 176 static enum ipi_vector ipi_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_IPI);
 182
 183         return info->u.ipi;
 184 }
 185
 186 static unsigned virq_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_VIRQ);
 192
 193         return info->u.virq;
 194 }
 195
 196 static unsigned gsi_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.gsi;
 204 }
 205
 206 static unsigned vector_from_irq(unsigned irq)
 207 {
 208         struct irq_info *info = info_for_irq(irq);
 209
 210         BUG_ON(info == NULL);
 211         BUG_ON(info->type != IRQT_PIRQ);
 212
 213         return info->u.pirq.vector;
 214 }
 215
 216 static enum xen_irq_type type_from_irq(unsigned irq)
 217 {
 218         return info_for_irq(irq)->type;
 219 }
 220
 221 static unsigned cpu_from_irq(unsigned irq)
 222 {
 223         return info_for_irq(irq)->cpu;
 224 }
 225
 226 static unsigned int cpu_from_evtchn(unsigned int evtchn)
 227 {
 228         int irq = evtchn_to_irq[evtchn];
 229         unsigned ret = 0;
 230
 231         if (irq != -1)
 232                 ret = cpu_from_irq(irq);
 233
 234         return ret;
 235 }
 236
 237 static bool pirq_needs_eoi(unsigned irq)
 238 {
 239         struct irq_info *info = info_for_irq(irq);
 240
 241         BUG_ON(info->type != IRQT_PIRQ);
 242
 243         return info->u.pirq.flags & PIRQ_NEEDS_EOI;
 244 }
 245
 246 static inline unsigned long active_evtchns(unsigned int cpu,
 247                                            struct shared_info *sh,
 248                                            unsigned int idx)
 249 {
 250         return (sh->evtchn_pending[idx] &
 251                 cpu_evtchn_mask(cpu)[idx] &
 252                 ~sh->evtchn_mask[idx]);
 253 }
 254
 255 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
 256 {
 257         int irq = evtchn_to_irq[chn];
 258
 259         BUG_ON(irq == -1);
 260 #ifdef CONFIG_SMP
 261         cpumask_copy(irq_to_desc(irq)->affinity, cpumask_of(cpu));
 262 #endif
 263
 264         __clear_bit(chn, cpu_evtchn_mask(cpu_from_irq(irq)));
 265         __set_bit(chn, cpu_evtchn_mask(cpu));
 266
 267         irq_info[irq].cpu = cpu;
 268 }
 269
 270 static void init_evtchn_cpu_bindings(void)
 271 {
 272 #ifdef CONFIG_SMP
 273         struct irq_desc *desc;
 274         int i;
 275
 276         /* By default all event channels notify CPU#0. */
 277         for_each_irq_desc(i, desc) {
 278                 cpumask_copy(desc->affinity, cpumask_of(0));
 279         }
 280 #endif
 281
 282         memset(cpu_evtchn_mask(0), ~0, sizeof(cpu_evtchn_mask(0)));
 283 }
 284
 285 static inline void clear_evtchn(int port)
 286 {
 287         struct shared_info *s = HYPERVISOR_shared_info;
 288         sync_clear_bit(port, &s->evtchn_pending[0]);
 289 }
 290
 291 static inline void set_evtchn(int port)
 292 {
 293         struct shared_info *s = HYPERVISOR_shared_info;
 294         sync_set_bit(port, &s->evtchn_pending[0]);
 295 }
 296
 297 static inline int test_evtchn(int port)
 298 {
 299         struct shared_info *s = HYPERVISOR_shared_info;
 300         return sync_test_bit(port, &s->evtchn_pending[0]);
 301 }
 302
 303
 304 /**
 305  * notify_remote_via_irq - send event to remote end of event channel via irq
 306  * @irq: irq of event channel to send event to
 307  *
 308  * Unlike notify_remote_via_evtchn(), this is safe to use across
 309  * save/restore. Notifications on a broken connection are silently
 310  * dropped.
 311  */
 312 void notify_remote_via_irq(int irq)
 313 {
 314         int evtchn = evtchn_from_irq(irq);
 315
 316         if (VALID_EVTCHN(evtchn))
 317                 notify_remote_via_evtchn(evtchn);
 318 }
 319 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
 320
 321 static void mask_evtchn(int port)
 322 {
 323         struct shared_info *s = HYPERVISOR_shared_info;
 324         sync_set_bit(port, &s->evtchn_mask[0]);
 325 }
 326
 327 static void unmask_evtchn(int port)
 328 {
 329         struct shared_info *s = HYPERVISOR_shared_info;
 330         unsigned int cpu = get_cpu();
 331
 332         BUG_ON(!irqs_disabled());
 333
 334         /* Slow path (hypercall) if this is a non-local port. */
 335         if (unlikely(cpu != cpu_from_evtchn(port))) {
 336                 struct evtchn_unmask unmask = { .port = port };
 337                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 338         } else {
 339                 struct vcpu_info *vcpu_info = __get_cpu_var(xen_vcpu);
 340
 341                 sync_clear_bit(port, &s->evtchn_mask[0]);
 342
 343                 /*
 344                  * The following is basically the equivalent of
 345                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
 346                  * the interrupt edge' if the channel is masked.
 347                  */
 348                 if (sync_test_bit(port, &s->evtchn_pending[0]) &&
 349                     !sync_test_and_set_bit(port / BITS_PER_LONG,
 350                                            &vcpu_info->evtchn_pending_sel))
 351                         vcpu_info->evtchn_upcall_pending = 1;
 352         }
 353
 354         put_cpu();
 355 }
 356
 357 static int get_nr_hw_irqs(void)
 358 {
 359         int ret = 1;
 360
 361 #ifdef CONFIG_X86_IO_APIC
 362         ret = get_nr_irqs_gsi();
 363 #endif
 364
 365         return ret;
 366 }
 367
 368 static int find_unbound_irq(void)
 369 {
 370         struct irq_data *data;
 371         int irq, res;
 372         int start = get_nr_hw_irqs();
 373
 374         if (start == nr_irqs)
 375                 goto no_irqs;
 376
 377         /* nr_irqs is a magic value. Must not use it.*/
 378         for (irq = nr_irqs-1; irq > start; irq--) {
 379                 data = irq_get_irq_data(irq);
 380                 /* only 0->15 have init'd desc; handle irq > 16 */
 381                 if (!data)
 382                         break;
 383                 if (data->chip == &no_irq_chip)
 384                         break;
 385                 if (data->chip != &xen_dynamic_chip)
 386                         continue;
 387                 if (irq_info[irq].type == IRQT_UNBOUND)
 388                         return irq;
 389         }
 390
 391         if (irq == start)
 392                 goto no_irqs;
 393
 394         res = irq_alloc_desc_at(irq, 0);
 395
 396         if (WARN_ON(res != irq))
 397                 return -1;
 398
 399         return irq;
 400
 401 no_irqs:
 402         panic("No available IRQ to bind to: increase nr_irqs!\n");
 403 }
 404
 405 static bool identity_mapped_irq(unsigned irq)
 406 {
 407         /* identity map all the hardware irqs */
 408         return irq < get_nr_hw_irqs();
 409 }
 410
 411 static void pirq_unmask_notify(int irq)
 412 {
 413         struct physdev_eoi eoi = { .irq = irq };
 414
 415         if (unlikely(pirq_needs_eoi(irq))) {
 416                 int rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
 417                 WARN_ON(rc);
 418         }
 419 }
 420
 421 static void pirq_query_unmask(int irq)
 422 {
 423         struct physdev_irq_status_query irq_status;
 424         struct irq_info *info = info_for_irq(irq);
 425
 426         BUG_ON(info->type != IRQT_PIRQ);
 427
 428         irq_status.irq = irq;
 429         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
 430                 irq_status.flags = 0;
 431
 432         info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
 433         if (irq_status.flags & XENIRQSTAT_needs_eoi)
 434                 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
 435 }
 436
 437 static bool probing_irq(int irq)
 438 {
 439         struct irq_desc *desc = irq_to_desc(irq);
 440
 441         return desc && desc->action == NULL;
 442 }
 443
 444 static unsigned int startup_pirq(unsigned int irq)
 445 {
 446         struct evtchn_bind_pirq bind_pirq;
 447         struct irq_info *info = info_for_irq(irq);
 448         int evtchn = evtchn_from_irq(irq);
 449         int rc;
 450
 451         BUG_ON(info->type != IRQT_PIRQ);
 452
 453         if (VALID_EVTCHN(evtchn))
 454                 goto out;
 455
 456         bind_pirq.pirq = irq;
 457         /* NB. We are happy to share unless we are probing. */
 458         bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
 459                                         BIND_PIRQ__WILL_SHARE : 0;
 460         rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
 461         if (rc != 0) {
 462                 if (!probing_irq(irq))
 463                         printk(KERN_INFO "Failed to obtain physical IRQ %d\n",
 464                                irq);
 465                 return 0;
 466         }
 467         evtchn = bind_pirq.port;
 468
 469         pirq_query_unmask(irq);
 470
 471         evtchn_to_irq[evtchn] = irq;
 472         bind_evtchn_to_cpu(evtchn, 0);
 473         info->evtchn = evtchn;
 474
 475 out:
 476         unmask_evtchn(evtchn);
 477         pirq_unmask_notify(irq);
 478
 479         return 0;
 480 }
 481
 482 static void shutdown_pirq(unsigned int irq)
 483 {
 484         struct evtchn_close close;
 485         struct irq_info *info = info_for_irq(irq);
 486         int evtchn = evtchn_from_irq(irq);
 487
 488         BUG_ON(info->type != IRQT_PIRQ);
 489
 490         if (!VALID_EVTCHN(evtchn))
 491                 return;
 492
 493         mask_evtchn(evtchn);
 494
 495         close.port = evtchn;
 496         if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 497                 BUG();
 498
 499         bind_evtchn_to_cpu(evtchn, 0);
 500         evtchn_to_irq[evtchn] = -1;
 501         info->evtchn = 0;
 502 }
 503
 504 static void enable_pirq(unsigned int irq)
 505 {
 506         startup_pirq(irq);
 507 }
 508
 509 static void disable_pirq(unsigned int irq)
 510 {
 511 }
 512
 513 static void ack_pirq(unsigned int irq)
 514 {
 515         int evtchn = evtchn_from_irq(irq);
 516
 517         move_native_irq(irq);
 518
 519         if (VALID_EVTCHN(evtchn)) {
 520                 mask_evtchn(evtchn);
 521                 clear_evtchn(evtchn);
 522         }
 523 }
 524
 525 static void end_pirq(unsigned int irq)
 526 {
 527         int evtchn = evtchn_from_irq(irq);
 528         struct irq_desc *desc = irq_to_desc(irq);
 529
 530         if (WARN_ON(!desc))
 531                 return;
 532
 533         if ((desc->status & (IRQ_DISABLED|IRQ_PENDING)) ==
 534             (IRQ_DISABLED|IRQ_PENDING)) {
 535                 shutdown_pirq(irq);
 536         } else if (VALID_EVTCHN(evtchn)) {
 537                 unmask_evtchn(evtchn);
 538                 pirq_unmask_notify(irq);
 539         }
 540 }
 541
 542 static int find_irq_by_gsi(unsigned gsi)
 543 {
 544         int irq;
 545
 546         for (irq = 0; irq < nr_irqs; irq++) {
 547                 struct irq_info *info = info_for_irq(irq);
 548
 549                 if (info == NULL || info->type != IRQT_PIRQ)
 550                         continue;
 551
 552                 if (gsi_from_irq(irq) == gsi)
 553                         return irq;
 554         }
 555
 556         return -1;
 557 }
 558
 559 /* xen_allocate_irq might allocate irqs from the top down, as a
 560  * consequence don't assume that the irq number returned has a low value
 561  * or can be used as a pirq number unless you know otherwise.
 562  *
 563  * One notable exception is when xen_allocate_irq is called passing an
 564  * hardware gsi as argument, in that case the irq number returned
 565  * matches the gsi number passed as first argument.
 566
 567  * Note: We don't assign an
 568  * event channel until the irq actually started up.  Return an
 569  * existing irq if we've already got one for the gsi.
 570  */
 571 int xen_allocate_pirq(unsigned gsi, int shareable, char *name)
 572 {
 573         int irq;
 574         struct physdev_irq irq_op;
 575
 576         spin_lock(&irq_mapping_update_lock);
 577
 578         irq = find_irq_by_gsi(gsi);
 579         if (irq != -1) {
 580                 printk(KERN_INFO "xen_allocate_pirq: returning irq %d for gsi %u\n",
 581                        irq, gsi);
 582                 goto out;       /* XXX need refcount? */
 583         }
 584
 585         /* If we are a PV guest, we don't have GSIs (no ACPI passed). Therefore
 586          * we are using the !xen_initial_domain() to drop in the function.*/
 587         if (identity_mapped_irq(gsi) || !xen_initial_domain()) {
 588                 irq = gsi;
 589                 irq_alloc_desc_at(irq, 0);
 590         } else
 591                 irq = find_unbound_irq();
 592
 593         set_irq_chip_and_handler_name(irq, &xen_pirq_chip,
 594                                       handle_level_irq, name);
 595
 596         irq_op.irq = irq;
 597         irq_op.vector = 0;
 598
 599         /* Only the privileged domain can do this. For non-priv, the pcifront
 600          * driver provides a PCI bus that does the call to do exactly
 601          * this in the priv domain. */
 602         if (xen_initial_domain() &&
 603             HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
 604                 irq_free_desc(irq);
 605                 irq = -ENOSPC;
 606                 goto out;
 607         }
 608
 609         irq_info[irq] = mk_pirq_info(0, gsi, irq_op.vector);
 610         irq_info[irq].u.pirq.flags |= shareable ? PIRQ_SHAREABLE : 0;
 611
 612 out:
 613         spin_unlock(&irq_mapping_update_lock);
 614
 615         return irq;
 616 }
 617
 618 int xen_destroy_irq(int irq)
 619 {
 620         struct irq_desc *desc;
 621         int rc = -ENOENT;
 622
 623         spin_lock(&irq_mapping_update_lock);
 624
 625         desc = irq_to_desc(irq);
 626         if (!desc)
 627                 goto out;
 628
 629         irq_info[irq] = mk_unbound_info();
 630
 631         irq_free_desc(irq);
 632
 633 out:
 634         spin_unlock(&irq_mapping_update_lock);
 635         return rc;
 636 }
 637
 638 int xen_vector_from_irq(unsigned irq)
 639 {
 640         return vector_from_irq(irq);
 641 }
 642
 643 int xen_gsi_from_irq(unsigned irq)
 644 {
 645         return gsi_from_irq(irq);
 646 }
 647
 648 int bind_evtchn_to_irq(unsigned int evtchn)
 649 {
 650         int irq;
 651
 652         spin_lock(&irq_mapping_update_lock);
 653
 654         irq = evtchn_to_irq[evtchn];
 655
 656         if (irq == -1) {
 657                 irq = find_unbound_irq();
 658
 659                 set_irq_chip_and_handler_name(irq, &xen_dynamic_chip,
 660                                               handle_edge_irq, "event");
 661
 662                 evtchn_to_irq[evtchn] = irq;
 663                 irq_info[irq] = mk_evtchn_info(evtchn);
 664         }
 665
 666         spin_unlock(&irq_mapping_update_lock);
 667
 668         return irq;
 669 }
 670 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
 671
 672 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
 673 {
 674         struct evtchn_bind_ipi bind_ipi;
 675         int evtchn, irq;
 676
 677         spin_lock(&irq_mapping_update_lock);
 678
 679         irq = per_cpu(ipi_to_irq, cpu)[ipi];
 680
 681         if (irq == -1) {
 682                 irq = find_unbound_irq();
 683                 if (irq < 0)
 684                         goto out;
 685
 686                 set_irq_chip_and_handler_name(irq, &xen_percpu_chip,
 687                                               handle_percpu_irq, "ipi");
 688
 689                 bind_ipi.vcpu = cpu;
 690                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
 691                                                 &bind_ipi) != 0)
 692                         BUG();
 693                 evtchn = bind_ipi.port;
 694
 695                 evtchn_to_irq[evtchn] = irq;
 696                 irq_info[irq] = mk_ipi_info(evtchn, ipi);
 697                 per_cpu(ipi_to_irq, cpu)[ipi] = irq;
 698
 699                 bind_evtchn_to_cpu(evtchn, cpu);
 700         }
 701
 702  out:
 703         spin_unlock(&irq_mapping_update_lock);
 704         return irq;
 705 }
 706
 707
 708 static int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
 709 {
 710         struct evtchn_bind_virq bind_virq;
 711         int evtchn, irq;
 712
 713         spin_lock(&irq_mapping_update_lock);
 714
 715         irq = per_cpu(virq_to_irq, cpu)[virq];
 716
 717         if (irq == -1) {
 718                 bind_virq.virq = virq;
 719                 bind_virq.vcpu = cpu;
 720                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 721                                                 &bind_virq) != 0)
 722                         BUG();
 723                 evtchn = bind_virq.port;
 724
 725                 irq = find_unbound_irq();
 726
 727                 set_irq_chip_and_handler_name(irq, &xen_percpu_chip,
 728                                               handle_percpu_irq, "virq");
 729
 730                 evtchn_to_irq[evtchn] = irq;
 731                 irq_info[irq] = mk_virq_info(evtchn, virq);
 732
 733                 per_cpu(virq_to_irq, cpu)[virq] = irq;
 734
 735                 bind_evtchn_to_cpu(evtchn, cpu);
 736         }
 737
 738         spin_unlock(&irq_mapping_update_lock);
 739
 740         return irq;
 741 }
 742
 743 static void unbind_from_irq(unsigned int irq)
 744 {
 745         struct evtchn_close close;
 746         int evtchn = evtchn_from_irq(irq);
 747
 748         spin_lock(&irq_mapping_update_lock);
 749
 750         if (VALID_EVTCHN(evtchn)) {
 751                 close.port = evtchn;
 752                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 753                         BUG();
 754
 755                 switch (type_from_irq(irq)) {
 756                 case IRQT_VIRQ:
 757                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
 758                                 [virq_from_irq(irq)] = -1;
 759                         break;
 760                 case IRQT_IPI:
 761                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
 762                                 [ipi_from_irq(irq)] = -1;
 763                         break;
 764                 default:
 765                         break;
 766                 }
 767
 768                 /* Closed ports are implicitly re-bound to VCPU0. */
 769                 bind_evtchn_to_cpu(evtchn, 0);
 770
 771                 evtchn_to_irq[evtchn] = -1;
 772         }
 773
 774         if (irq_info[irq].type != IRQT_UNBOUND) {
 775                 irq_info[irq] = mk_unbound_info();
 776
 777                 irq_free_desc(irq);
 778         }
 779
 780         spin_unlock(&irq_mapping_update_lock);
 781 }
 782
 783 int bind_evtchn_to_irqhandler(unsigned int evtchn,
 784                               irq_handler_t handler,
 785                               unsigned long irqflags,
 786                               const char *devname, void *dev_id)
 787 {
 788         unsigned int irq;
 789         int retval;
 790
 791         irq = bind_evtchn_to_irq(evtchn);
 792         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 793         if (retval != 0) {
 794                 unbind_from_irq(irq);
 795                 return retval;
 796         }
 797
 798         return irq;
 799 }
 800 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
 801
 802 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
 803                             irq_handler_t handler,
 804                             unsigned long irqflags, const char *devname, void *dev_id)
 805 {
 806         unsigned int irq;
 807         int retval;
 808
 809         irq = bind_virq_to_irq(virq, cpu);
 810         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 811         if (retval != 0) {
 812                 unbind_from_irq(irq);
 813                 return retval;
 814         }
 815
 816         return irq;
 817 }
 818 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
 819
 820 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
 821                            unsigned int cpu,
 822                            irq_handler_t handler,
 823                            unsigned long irqflags,
 824                            const char *devname,
 825                            void *dev_id)
 826 {
 827         int irq, retval;
 828
 829         irq = bind_ipi_to_irq(ipi, cpu);
 830         if (irq < 0)
 831                 return irq;
 832
 833         irqflags |= IRQF_NO_SUSPEND;
 834         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 835         if (retval != 0) {
 836                 unbind_from_irq(irq);
 837                 return retval;
 838         }
 839
 840         return irq;
 841 }
 842
 843 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
 844 {
 845         free_irq(irq, dev_id);
 846         unbind_from_irq(irq);
 847 }
 848 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
 849
 850 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
 851 {
 852         int irq = per_cpu(ipi_to_irq, cpu)[vector];
 853         BUG_ON(irq < 0);
 854         notify_remote_via_irq(irq);
 855 }
 856
 857 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
 858 {
 859         struct shared_info *sh = HYPERVISOR_shared_info;
 860         int cpu = smp_processor_id();
 861         int i;
 862         unsigned long flags;
 863         static DEFINE_SPINLOCK(debug_lock);
 864
 865         spin_lock_irqsave(&debug_lock, flags);
 866
 867         printk("vcpu %d\n  ", cpu);
 868
 869         for_each_online_cpu(i) {
 870                 struct vcpu_info *v = per_cpu(xen_vcpu, i);
 871                 printk("%d: masked=%d pending=%d event_sel %08lx\n  ", i,
 872                         (get_irq_regs() && i == cpu) ? xen_irqs_disabled(get_irq_regs()) : v->evtchn_upcall_mask,
 873                         v->evtchn_upcall_pending,
 874                         v->evtchn_pending_sel);
 875         }
 876         printk("pending:\n   ");
 877         for(i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
 878                 printk("%08lx%s", sh->evtchn_pending[i],
 879                         i % 8 == 0 ? "\n   " : " ");
 880         printk("\nmasks:\n   ");
 881         for(i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
 882                 printk("%08lx%s", sh->evtchn_mask[i],
 883                         i % 8 == 0 ? "\n   " : " ");
 884
 885         printk("\nunmasked:\n   ");
 886         for(i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
 887                 printk("%08lx%s", sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
 888                         i % 8 == 0 ? "\n   " : " ");
 889
 890         printk("\npending list:\n");
 891         for(i = 0; i < NR_EVENT_CHANNELS; i++) {
 892                 if (sync_test_bit(i, sh->evtchn_pending)) {
 893                         printk("  %d: event %d -> irq %d\n",
 894                                cpu_from_evtchn(i), i,
 895                                evtchn_to_irq[i]);
 896                 }
 897         }
 898
 899         spin_unlock_irqrestore(&debug_lock, flags);
 900
 901         return IRQ_HANDLED;
 902 }
 903
 904 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
 905
 906 /*
 907  * Search the CPUs pending events bitmasks.  For each one found, map
 908  * the event number to an irq, and feed it into do_IRQ() for
 909  * handling.
 910  *
 911  * Xen uses a two-level bitmap to speed searching.  The first level is
 912  * a bitset of words which contain pending event bits.  The second
 913  * level is a bitset of pending events themselves.
 914  */
 915 static void __xen_evtchn_do_upcall(void)
 916 {
 917         int cpu = get_cpu();
 918         struct shared_info *s = HYPERVISOR_shared_info;
 919         struct vcpu_info *vcpu_info = __get_cpu_var(xen_vcpu);
 920         unsigned count;
 921
 922         do {
 923                 unsigned long pending_words;
 924
 925                 vcpu_info->evtchn_upcall_pending = 0;
 926
 927                 if (__get_cpu_var(xed_nesting_count)++)
 928                         goto out;
 929
 930 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
 931                 /* Clear master flag /before/ clearing selector flag. */
 932                 wmb();
 933 #endif
 934                 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
 935                 while (pending_words != 0) {
 936                         unsigned long pending_bits;
 937                         int word_idx = __ffs(pending_words);
 938                         pending_words &= ~(1UL << word_idx);
 939
 940                         while ((pending_bits = active_evtchns(cpu, s, word_idx)) != 0) {
 941                                 int bit_idx = __ffs(pending_bits);
 942                                 int port = (word_idx * BITS_PER_LONG) + bit_idx;
 943                                 int irq = evtchn_to_irq[port];
 944                                 struct irq_desc *desc;
 945
 946                                 if (irq != -1) {
 947                                         desc = irq_to_desc(irq);
 948                                         if (desc)
 949                                                 generic_handle_irq_desc(irq, desc);
 950                                 }
 951                         }
 952                 }
 953
 954                 BUG_ON(!irqs_disabled());
 955
 956                 count = __get_cpu_var(xed_nesting_count);
 957                 __get_cpu_var(xed_nesting_count) = 0;
 958         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
 959
 960 out:
 961
 962         put_cpu();
 963 }
 964
 965 void xen_evtchn_do_upcall(struct pt_regs *regs)
 966 {
 967         struct pt_regs *old_regs = set_irq_regs(regs);
 968
 969         exit_idle();
 970         irq_enter();
 971
 972         __xen_evtchn_do_upcall();
 973
 974         irq_exit();
 975         set_irq_regs(old_regs);
 976 }
 977
 978 void xen_hvm_evtchn_do_upcall(void)
 979 {
 980         __xen_evtchn_do_upcall();
 981 }
 982 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
 983
 984 /* Rebind a new event channel to an existing irq. */
 985 void rebind_evtchn_irq(int evtchn, int irq)
 986 {
 987         struct irq_info *info = info_for_irq(irq);
 988
 989         /* Make sure the irq is masked, since the new event channel
 990            will also be masked. */
 991         disable_irq(irq);
 992
 993         spin_lock(&irq_mapping_update_lock);
 994
 995         /* After resume the irq<->evtchn mappings are all cleared out */
 996         BUG_ON(evtchn_to_irq[evtchn] != -1);
 997         /* Expect irq to have been bound before,
 998            so there should be a proper type */
 999         BUG_ON(info->type == IRQT_UNBOUND);
1000
1001         evtchn_to_irq[evtchn] = irq;
1002         irq_info[irq] = mk_evtchn_info(evtchn);
1003
1004         spin_unlock(&irq_mapping_update_lock);
1005
1006         /* new event channels are always bound to cpu 0 */
1007         irq_set_affinity(irq, cpumask_of(0));
1008
1009         /* Unmask the event channel. */
1010         enable_irq(irq);
1011 }
1012
1013 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1014 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1015 {
1016         struct evtchn_bind_vcpu bind_vcpu;
1017         int evtchn = evtchn_from_irq(irq);
1018
1019         /* events delivered via platform PCI interrupts are always
1020          * routed to vcpu 0 */
1021         if (!VALID_EVTCHN(evtchn) ||
1022                 (xen_hvm_domain() && !xen_have_vector_callback))
1023                 return -1;
1024
1025         /* Send future instances of this interrupt to other vcpu. */
1026         bind_vcpu.port = evtchn;
1027         bind_vcpu.vcpu = tcpu;
1028
1029         /*
1030          * If this fails, it usually just indicates that we're dealing with a
1031          * virq or IPI channel, which don't actually need to be rebound. Ignore
1032          * it, but don't do the xenlinux-level rebind in that case.
1033          */
1034         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1035                 bind_evtchn_to_cpu(evtchn, tcpu);
1036
1037         return 0;
1038 }
1039
1040 static int set_affinity_irq(unsigned irq, const struct cpumask *dest)
1041 {
1042         unsigned tcpu = cpumask_first(dest);
1043
1044         return rebind_irq_to_cpu(irq, tcpu);
1045 }
1046
1047 int resend_irq_on_evtchn(unsigned int irq)
1048 {
1049         int masked, evtchn = evtchn_from_irq(irq);
1050         struct shared_info *s = HYPERVISOR_shared_info;
1051
1052         if (!VALID_EVTCHN(evtchn))
1053                 return 1;
1054
1055         masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
1056         sync_set_bit(evtchn, s->evtchn_pending);
1057         if (!masked)
1058                 unmask_evtchn(evtchn);
1059
1060         return 1;
1061 }
1062
1063 static void enable_dynirq(unsigned int irq)
1064 {
1065         int evtchn = evtchn_from_irq(irq);
1066
1067         if (VALID_EVTCHN(evtchn))
1068                 unmask_evtchn(evtchn);
1069 }
1070
1071 static void disable_dynirq(unsigned int irq)
1072 {
1073         int evtchn = evtchn_from_irq(irq);
1074
1075         if (VALID_EVTCHN(evtchn))
1076                 mask_evtchn(evtchn);
1077 }
1078
1079 static void ack_dynirq(unsigned int irq)
1080 {
1081         int evtchn = evtchn_from_irq(irq);
1082
1083         move_native_irq(irq);
1084
1085         if (VALID_EVTCHN(evtchn))
1086                 clear_evtchn(evtchn);
1087 }
1088
1089 static int retrigger_dynirq(unsigned int irq)
1090 {
1091         int evtchn = evtchn_from_irq(irq);
1092         struct shared_info *sh = HYPERVISOR_shared_info;
1093         int ret = 0;
1094
1095         if (VALID_EVTCHN(evtchn)) {
1096                 int masked;
1097
1098                 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
1099                 sync_set_bit(evtchn, sh->evtchn_pending);
1100                 if (!masked)
1101                         unmask_evtchn(evtchn);
1102                 ret = 1;
1103         }
1104
1105         return ret;
1106 }
1107
1108 static void restore_cpu_virqs(unsigned int cpu)
1109 {
1110         struct evtchn_bind_virq bind_virq;
1111         int virq, irq, evtchn;
1112
1113         for (virq = 0; virq < NR_VIRQS; virq++) {
1114                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1115                         continue;
1116
1117                 BUG_ON(virq_from_irq(irq) != virq);
1118
1119                 /* Get a new binding from Xen. */
1120                 bind_virq.virq = virq;
1121                 bind_virq.vcpu = cpu;
1122                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1123                                                 &bind_virq) != 0)
1124                         BUG();
1125                 evtchn = bind_virq.port;
1126
1127                 /* Record the new mapping. */
1128                 evtchn_to_irq[evtchn] = irq;
1129                 irq_info[irq] = mk_virq_info(evtchn, virq);
1130                 bind_evtchn_to_cpu(evtchn, cpu);
1131
1132                 /* Ready for use. */
1133                 unmask_evtchn(evtchn);
1134         }
1135 }
1136
1137 static void restore_cpu_ipis(unsigned int cpu)
1138 {
1139         struct evtchn_bind_ipi bind_ipi;
1140         int ipi, irq, evtchn;
1141
1142         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1143                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1144                         continue;
1145
1146                 BUG_ON(ipi_from_irq(irq) != ipi);
1147
1148                 /* Get a new binding from Xen. */
1149                 bind_ipi.vcpu = cpu;
1150                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1151                                                 &bind_ipi) != 0)
1152                         BUG();
1153                 evtchn = bind_ipi.port;
1154
1155                 /* Record the new mapping. */
1156                 evtchn_to_irq[evtchn] = irq;
1157                 irq_info[irq] = mk_ipi_info(evtchn, ipi);
1158                 bind_evtchn_to_cpu(evtchn, cpu);
1159
1160                 /* Ready for use. */
1161                 unmask_evtchn(evtchn);
1162
1163         }
1164 }
1165
1166 /* Clear an irq's pending state, in preparation for polling on it */
1167 void xen_clear_irq_pending(int irq)
1168 {
1169         int evtchn = evtchn_from_irq(irq);
1170
1171         if (VALID_EVTCHN(evtchn))
1172                 clear_evtchn(evtchn);
1173 }
1174 EXPORT_SYMBOL(xen_clear_irq_pending);
1175 void xen_set_irq_pending(int irq)
1176 {
1177         int evtchn = evtchn_from_irq(irq);
1178
1179         if (VALID_EVTCHN(evtchn))
1180                 set_evtchn(evtchn);
1181 }
1182
1183 bool xen_test_irq_pending(int irq)
1184 {
1185         int evtchn = evtchn_from_irq(irq);
1186         bool ret = false;
1187
1188         if (VALID_EVTCHN(evtchn))
1189                 ret = test_evtchn(evtchn);
1190
1191         return ret;
1192 }
1193
1194 /* Poll waiting for an irq to become pending with timeout.  In the usual case,
1195  * the irq will be disabled so it won't deliver an interrupt. */
1196 void xen_poll_irq_timeout(int irq, u64 timeout)
1197 {
1198         evtchn_port_t evtchn = evtchn_from_irq(irq);
1199
1200         if (VALID_EVTCHN(evtchn)) {
1201                 struct sched_poll poll;
1202
1203                 poll.nr_ports = 1;
1204                 poll.timeout = timeout;
1205                 set_xen_guest_handle(poll.ports, &evtchn);
1206
1207                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1208                         BUG();
1209         }
1210 }
1211 EXPORT_SYMBOL(xen_poll_irq_timeout);
1212 /* Poll waiting for an irq to become pending.  In the usual case, the
1213  * irq will be disabled so it won't deliver an interrupt. */
1214 void xen_poll_irq(int irq)
1215 {
1216         xen_poll_irq_timeout(irq, 0 /* no timeout */);
1217 }
1218
1219 void xen_irq_resume(void)
1220 {
1221         unsigned int cpu, irq, evtchn;
1222
1223         init_evtchn_cpu_bindings();
1224
1225         /* New event-channel space is not 'live' yet. */
1226         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1227                 mask_evtchn(evtchn);
1228
1229         /* No IRQ <-> event-channel mappings. */
1230         for (irq = 0; irq < nr_irqs; irq++)
1231                 irq_info[irq].evtchn = 0; /* zap event-channel binding */
1232
1233         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1234                 evtchn_to_irq[evtchn] = -1;
1235
1236         for_each_possible_cpu(cpu) {
1237                 restore_cpu_virqs(cpu);
1238                 restore_cpu_ipis(cpu);
1239         }
1240 }
1241
1242 static struct irq_chip xen_dynamic_chip __read_mostly = {
1243         .name           = "xen-dyn",
1244
1245         .disable        = disable_dynirq,
1246         .mask           = disable_dynirq,
1247         .unmask         = enable_dynirq,
1248
1249         .ack            = ack_dynirq,
1250         .set_affinity   = set_affinity_irq,
1251         .retrigger      = retrigger_dynirq,
1252 };
1253
1254 static struct irq_chip xen_pirq_chip __read_mostly = {
1255         .name           = "xen-pirq",
1256
1257         .startup        = startup_pirq,
1258         .shutdown       = shutdown_pirq,
1259
1260         .enable         = enable_pirq,
1261         .unmask         = enable_pirq,
1262
1263         .disable        = disable_pirq,
1264         .mask           = disable_pirq,
1265
1266         .ack            = ack_pirq,
1267         .end            = end_pirq,
1268
1269         .set_affinity   = set_affinity_irq,
1270
1271         .retrigger      = retrigger_dynirq,
1272 };
1273
1274 static struct irq_chip xen_percpu_chip __read_mostly = {
1275         .name           = "xen-percpu",
1276
1277         .disable        = disable_dynirq,
1278         .mask           = disable_dynirq,
1279         .unmask         = enable_dynirq,
1280
1281         .ack            = ack_dynirq,
1282 };
1283
1284 int xen_set_callback_via(uint64_t via)
1285 {
1286         struct xen_hvm_param a;
1287         a.domid = DOMID_SELF;
1288         a.index = HVM_PARAM_CALLBACK_IRQ;
1289         a.value = via;
1290         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1291 }
1292 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1293
1294 #ifdef CONFIG_XEN_PVHVM
1295 /* Vector callbacks are better than PCI interrupts to receive event
1296  * channel notifications because we can receive vector callbacks on any
1297  * vcpu and we don't need PCI support or APIC interactions. */
1298 void xen_callback_vector(void)
1299 {
1300         int rc;
1301         uint64_t callback_via;
1302         if (xen_have_vector_callback) {
1303                 callback_via = HVM_CALLBACK_VECTOR(XEN_HVM_EVTCHN_CALLBACK);
1304                 rc = xen_set_callback_via(callback_via);
1305                 if (rc) {
1306                         printk(KERN_ERR "Request for Xen HVM callback vector"
1307                                         " failed.\n");
1308                         xen_have_vector_callback = 0;
1309                         return;
1310                 }
1311                 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1312                                 "enabled\n");
1313                 /* in the restore case the vector has already been allocated */
1314                 if (!test_bit(XEN_HVM_EVTCHN_CALLBACK, used_vectors))
1315                         alloc_intr_gate(XEN_HVM_EVTCHN_CALLBACK, xen_hvm_callback_vector);
1316         }
1317 }
1318 #else
1319 void xen_callback_vector(void) {}
1320 #endif
1321
1322 void __init xen_init_IRQ(void)
1323 {
1324         int i;
1325
1326         cpu_evtchn_mask_p = kcalloc(nr_cpu_ids, sizeof(struct cpu_evtchn_s),
1327                                     GFP_KERNEL);
1328         irq_info = kcalloc(nr_irqs, sizeof(*irq_info), GFP_KERNEL);
1329
1330         evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1331                                     GFP_KERNEL);
1332         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1333                 evtchn_to_irq[i] = -1;
1334
1335         init_evtchn_cpu_bindings();
1336
1337         /* No event channels are 'live' right now. */
1338         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1339                 mask_evtchn(i);
1340
1341         if (xen_hvm_domain()) {
1342                 xen_callback_vector();
1343                 native_init_IRQ();
1344         } else {
1345                 irq_ctx_init(smp_processor_id());
1346         }
1347 }