drivers/xen/events.c

   1 /*
   2  * Xen event channels
   3  *
   4  * Xen models interrupts with abstract event channels.  Because each
   5  * domain gets 1024 event channels, but NR_IRQ is not that large, we
   6  * must dynamically map irqs<->event channels.  The event channels
   7  * interface with the rest of the kernel by defining a xen interrupt
   8  * chip.  When an event is received, it is mapped to an irq and sent
   9  * through the normal interrupt processing path.
  10  *
  11  * There are four kinds of events which can be mapped to an event
  12  * channel:
  13  *
  14  * 1. Inter-domain notifications.  This includes all the virtual
  15  *    device events, since they're driven by front-ends in another domain
  16  *    (typically dom0).
  17  * 2. VIRQs, typically used for timers.  These are per-cpu events.
  18  * 3. IPIs.
  19  * 4. PIRQs - Hardware interrupts.
  20  *
  21  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
  22  */
  23
  24 #include <linux/linkage.h>
  25 #include <linux/interrupt.h>
  26 #include <linux/irq.h>
  27 #include <linux/module.h>
  28 #include <linux/string.h>
  29 #include <linux/bootmem.h>
  30 #include <linux/slab.h>
  31 #include <linux/irqnr.h>
  32 #include <linux/pci.h>
  33
  34 #include <asm/desc.h>
  35 #include <asm/ptrace.h>
  36 #include <asm/irq.h>
  37 #include <asm/idle.h>
  38 #include <asm/io_apic.h>
  39 #include <asm/sync_bitops.h>
  40 #include <asm/xen/pci.h>
  41 #include <asm/xen/hypercall.h>
  42 #include <asm/xen/hypervisor.h>
  43
  44 #include <xen/xen.h>
  45 #include <xen/hvm.h>
  46 #include <xen/xen-ops.h>
  47 #include <xen/events.h>
  48 #include <xen/interface/xen.h>
  49 #include <xen/interface/event_channel.h>
  50 #include <xen/interface/hvm/hvm_op.h>
  51 #include <xen/interface/hvm/params.h>
  52
  53 /*
  54  * This lock protects updates to the following mapping and reference-count
  55  * arrays. The lock does not need to be acquired to read the mapping tables.
  56  */
  57 static DEFINE_MUTEX(irq_mapping_update_lock);
  58
  59 static LIST_HEAD(xen_irq_list_head);
  60
  61 /* IRQ <-> VIRQ mapping. */
  62 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
  63
  64 /* IRQ <-> IPI mapping */
  65 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
  66
  67 /* Interrupt types. */
  68 enum xen_irq_type {
  69         IRQT_UNBOUND = 0,
  70         IRQT_PIRQ,
  71         IRQT_VIRQ,
  72         IRQT_IPI,
  73         IRQT_EVTCHN
  74 };
  75
  76 /*
  77  * Packed IRQ information:
  78  * type - enum xen_irq_type
  79  * event channel - irq->event channel mapping
  80  * cpu - cpu this event channel is bound to
  81  * index - type-specific information:
  82  *    PIRQ - vector, with MSB being "needs EIO", or physical IRQ of the HVM
  83  *           guest, or GSI (real passthrough IRQ) of the device.
  84  *    VIRQ - virq number
  85  *    IPI - IPI vector
  86  *    EVTCHN -
  87  */
  88 struct irq_info {
  89         struct list_head list;
  90         int refcnt;
  91         enum xen_irq_type type; /* type */
  92         unsigned irq;
  93         unsigned short evtchn;  /* event channel */
  94         unsigned short cpu;     /* cpu bound */
  95
  96         union {
  97                 unsigned short virq;
  98                 enum ipi_vector ipi;
  99                 struct {
 100                         unsigned short pirq;
 101                         unsigned short gsi;
 102                         unsigned char vector;
 103                         unsigned char flags;
 104                         uint16_t domid;
 105                 } pirq;
 106         } u;
 107 };
 108 #define PIRQ_NEEDS_EOI  (1 << 0)
 109 #define PIRQ_SHAREABLE  (1 << 1)
 110
 111 static int *evtchn_to_irq;
 112
 113 static DEFINE_PER_CPU(unsigned long [NR_EVENT_CHANNELS/BITS_PER_LONG],
 114                       cpu_evtchn_mask);
 115
 116 /* Xen will never allocate port zero for any purpose. */
 117 #define VALID_EVTCHN(chn)       ((chn) != 0)
 118
 119 static struct irq_chip xen_dynamic_chip;
 120 static struct irq_chip xen_percpu_chip;
 121 static struct irq_chip xen_pirq_chip;
 122 static void enable_dynirq(struct irq_data *data);
 123 static void disable_dynirq(struct irq_data *data);
 124
 125 /* Get info for IRQ */
 126 static struct irq_info *info_for_irq(unsigned irq)
 127 {
 128         return irq_get_handler_data(irq);
 129 }
 130
 131 /* Constructors for packed IRQ information. */
 132 static void xen_irq_info_common_init(struct irq_info *info,
 133                                      unsigned irq,
 134                                      enum xen_irq_type type,
 135                                      unsigned short evtchn,
 136                                      unsigned short cpu)
 137 {
 138
 139         BUG_ON(info->type != IRQT_UNBOUND && info->type != type);
 140
 141         info->type = type;
 142         info->irq = irq;
 143         info->evtchn = evtchn;
 144         info->cpu = cpu;
 145
 146         evtchn_to_irq[evtchn] = irq;
 147 }
 148
 149 static void xen_irq_info_evtchn_init(unsigned irq,
 150                                      unsigned short evtchn)
 151 {
 152         struct irq_info *info = info_for_irq(irq);
 153
 154         xen_irq_info_common_init(info, irq, IRQT_EVTCHN, evtchn, 0);
 155 }
 156
 157 static void xen_irq_info_ipi_init(unsigned cpu,
 158                                   unsigned irq,
 159                                   unsigned short evtchn,
 160                                   enum ipi_vector ipi)
 161 {
 162         struct irq_info *info = info_for_irq(irq);
 163
 164         xen_irq_info_common_init(info, irq, IRQT_IPI, evtchn, 0);
 165
 166         info->u.ipi = ipi;
 167
 168         per_cpu(ipi_to_irq, cpu)[ipi] = irq;
 169 }
 170
 171 static void xen_irq_info_virq_init(unsigned cpu,
 172                                    unsigned irq,
 173                                    unsigned short evtchn,
 174                                    unsigned short virq)
 175 {
 176         struct irq_info *info = info_for_irq(irq);
 177
 178         xen_irq_info_common_init(info, irq, IRQT_VIRQ, evtchn, 0);
 179
 180         info->u.virq = virq;
 181
 182         per_cpu(virq_to_irq, cpu)[virq] = irq;
 183 }
 184
 185 static void xen_irq_info_pirq_init(unsigned irq,
 186                                    unsigned short evtchn,
 187                                    unsigned short pirq,
 188                                    unsigned short gsi,
 189                                    unsigned short vector,
 190                                    uint16_t domid,
 191                                    unsigned char flags)
 192 {
 193         struct irq_info *info = info_for_irq(irq);
 194
 195         xen_irq_info_common_init(info, irq, IRQT_PIRQ, evtchn, 0);
 196
 197         info->u.pirq.pirq = pirq;
 198         info->u.pirq.gsi = gsi;
 199         info->u.pirq.vector = vector;
 200         info->u.pirq.domid = domid;
 201         info->u.pirq.flags = flags;
 202 }
 203
 204 /*
 205  * Accessors for packed IRQ information.
 206  */
 207 static unsigned int evtchn_from_irq(unsigned irq)
 208 {
 209         if (unlikely(WARN(irq < 0 || irq >= nr_irqs, "Invalid irq %d!\n", irq)))
 210                 return 0;
 211
 212         return info_for_irq(irq)->evtchn;
 213 }
 214
 215 unsigned irq_from_evtchn(unsigned int evtchn)
 216 {
 217         return evtchn_to_irq[evtchn];
 218 }
 219 EXPORT_SYMBOL_GPL(irq_from_evtchn);
 220
 221 static enum ipi_vector ipi_from_irq(unsigned irq)
 222 {
 223         struct irq_info *info = info_for_irq(irq);
 224
 225         BUG_ON(info == NULL);
 226         BUG_ON(info->type != IRQT_IPI);
 227
 228         return info->u.ipi;
 229 }
 230
 231 static unsigned virq_from_irq(unsigned irq)
 232 {
 233         struct irq_info *info = info_for_irq(irq);
 234
 235         BUG_ON(info == NULL);
 236         BUG_ON(info->type != IRQT_VIRQ);
 237
 238         return info->u.virq;
 239 }
 240
 241 static unsigned pirq_from_irq(unsigned irq)
 242 {
 243         struct irq_info *info = info_for_irq(irq);
 244
 245         BUG_ON(info == NULL);
 246         BUG_ON(info->type != IRQT_PIRQ);
 247
 248         return info->u.pirq.pirq;
 249 }
 250
 251 static enum xen_irq_type type_from_irq(unsigned irq)
 252 {
 253         return info_for_irq(irq)->type;
 254 }
 255
 256 static unsigned cpu_from_irq(unsigned irq)
 257 {
 258         return info_for_irq(irq)->cpu;
 259 }
 260
 261 static unsigned int cpu_from_evtchn(unsigned int evtchn)
 262 {
 263         int irq = evtchn_to_irq[evtchn];
 264         unsigned ret = 0;
 265
 266         if (irq != -1)
 267                 ret = cpu_from_irq(irq);
 268
 269         return ret;
 270 }
 271
 272 static bool pirq_needs_eoi(unsigned irq)
 273 {
 274         struct irq_info *info = info_for_irq(irq);
 275
 276         BUG_ON(info->type != IRQT_PIRQ);
 277
 278         return info->u.pirq.flags & PIRQ_NEEDS_EOI;
 279 }
 280
 281 static inline unsigned long active_evtchns(unsigned int cpu,
 282                                            struct shared_info *sh,
 283                                            unsigned int idx)
 284 {
 285         return sh->evtchn_pending[idx] &
 286                 per_cpu(cpu_evtchn_mask, cpu)[idx] &
 287                 ~sh->evtchn_mask[idx];
 288 }
 289
 290 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
 291 {
 292         int irq = evtchn_to_irq[chn];
 293
 294         BUG_ON(irq == -1);
 295 #ifdef CONFIG_SMP
 296         cpumask_copy(irq_to_desc(irq)->irq_data.affinity, cpumask_of(cpu));
 297 #endif
 298
 299         clear_bit(chn, per_cpu(cpu_evtchn_mask, cpu_from_irq(irq)));
 300         set_bit(chn, per_cpu(cpu_evtchn_mask, cpu));
 301
 302         info_for_irq(irq)->cpu = cpu;
 303 }
 304
 305 static void init_evtchn_cpu_bindings(void)
 306 {
 307         int i;
 308 #ifdef CONFIG_SMP
 309         struct irq_info *info;
 310
 311         /* By default all event channels notify CPU#0. */
 312         list_for_each_entry(info, &xen_irq_list_head, list) {
 313                 struct irq_desc *desc = irq_to_desc(info->irq);
 314                 cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 315         }
 316 #endif
 317
 318         for_each_possible_cpu(i)
 319                 memset(per_cpu(cpu_evtchn_mask, i),
 320                        (i == 0) ? ~0 : 0, sizeof(*per_cpu(cpu_evtchn_mask, i)));
 321 }
 322
 323 static inline void clear_evtchn(int port)
 324 {
 325         struct shared_info *s = HYPERVISOR_shared_info;
 326         sync_clear_bit(port, &s->evtchn_pending[0]);
 327 }
 328
 329 static inline void set_evtchn(int port)
 330 {
 331         struct shared_info *s = HYPERVISOR_shared_info;
 332         sync_set_bit(port, &s->evtchn_pending[0]);
 333 }
 334
 335 static inline int test_evtchn(int port)
 336 {
 337         struct shared_info *s = HYPERVISOR_shared_info;
 338         return sync_test_bit(port, &s->evtchn_pending[0]);
 339 }
 340
 341
 342 /**
 343  * notify_remote_via_irq - send event to remote end of event channel via irq
 344  * @irq: irq of event channel to send event to
 345  *
 346  * Unlike notify_remote_via_evtchn(), this is safe to use across
 347  * save/restore. Notifications on a broken connection are silently
 348  * dropped.
 349  */
 350 void notify_remote_via_irq(int irq)
 351 {
 352         int evtchn = evtchn_from_irq(irq);
 353
 354         if (VALID_EVTCHN(evtchn))
 355                 notify_remote_via_evtchn(evtchn);
 356 }
 357 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
 358
 359 static void mask_evtchn(int port)
 360 {
 361         struct shared_info *s = HYPERVISOR_shared_info;
 362         sync_set_bit(port, &s->evtchn_mask[0]);
 363 }
 364
 365 static void unmask_evtchn(int port)
 366 {
 367         struct shared_info *s = HYPERVISOR_shared_info;
 368         unsigned int cpu = get_cpu();
 369
 370         BUG_ON(!irqs_disabled());
 371
 372         /* Slow path (hypercall) if this is a non-local port. */
 373         if (unlikely(cpu != cpu_from_evtchn(port))) {
 374                 struct evtchn_unmask unmask = { .port = port };
 375                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 376         } else {
 377                 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
 378
 379                 sync_clear_bit(port, &s->evtchn_mask[0]);
 380
 381                 /*
 382                  * The following is basically the equivalent of
 383                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
 384                  * the interrupt edge' if the channel is masked.
 385                  */
 386                 if (sync_test_bit(port, &s->evtchn_pending[0]) &&
 387                     !sync_test_and_set_bit(port / BITS_PER_LONG,
 388                                            &vcpu_info->evtchn_pending_sel))
 389                         vcpu_info->evtchn_upcall_pending = 1;
 390         }
 391
 392         put_cpu();
 393 }
 394
 395 static void xen_irq_init(unsigned irq)
 396 {
 397         struct irq_info *info;
 398 #ifdef CONFIG_SMP
 399         struct irq_desc *desc = irq_to_desc(irq);
 400
 401         /* By default all event channels notify CPU#0. */
 402         cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 403 #endif
 404
 405         info = kzalloc(sizeof(*info), GFP_KERNEL);
 406         if (info == NULL)
 407                 panic("Unable to allocate metadata for IRQ%d\n", irq);
 408
 409         info->type = IRQT_UNBOUND;
 410         info->refcnt = -1;
 411
 412         irq_set_handler_data(irq, info);
 413
 414         list_add_tail(&info->list, &xen_irq_list_head);
 415 }
 416
 417 static int __must_check xen_allocate_irq_dynamic(void)
 418 {
 419         int first = 0;
 420         int irq;
 421
 422 #ifdef CONFIG_X86_IO_APIC
 423         /*
 424          * For an HVM guest or domain 0 which see "real" (emulated or
 425          * actual respectively) GSIs we allocate dynamic IRQs
 426          * e.g. those corresponding to event channels or MSIs
 427          * etc. from the range above those "real" GSIs to avoid
 428          * collisions.
 429          */
 430         if (xen_initial_domain() || xen_hvm_domain())
 431                 first = get_nr_irqs_gsi();
 432 #endif
 433
 434         irq = irq_alloc_desc_from(first, -1);
 435
 436         if (irq >= 0)
 437                 xen_irq_init(irq);
 438
 439         return irq;
 440 }
 441
 442 static int __must_check xen_allocate_irq_gsi(unsigned gsi)
 443 {
 444         int irq;
 445
 446         /*
 447          * A PV guest has no concept of a GSI (since it has no ACPI
 448          * nor access to/knowledge of the physical APICs). Therefore
 449          * all IRQs are dynamically allocated from the entire IRQ
 450          * space.
 451          */
 452         if (xen_pv_domain() && !xen_initial_domain())
 453                 return xen_allocate_irq_dynamic();
 454
 455         /* Legacy IRQ descriptors are already allocated by the arch. */
 456         if (gsi < NR_IRQS_LEGACY)
 457                 irq = gsi;
 458         else
 459                 irq = irq_alloc_desc_at(gsi, -1);
 460
 461         xen_irq_init(irq);
 462
 463         return irq;
 464 }
 465
 466 static void xen_free_irq(unsigned irq)
 467 {
 468         struct irq_info *info = irq_get_handler_data(irq);
 469
 470         list_del(&info->list);
 471
 472         irq_set_handler_data(irq, NULL);
 473
 474         WARN_ON(info->refcnt > 0);
 475
 476         kfree(info);
 477
 478         /* Legacy IRQ descriptors are managed by the arch. */
 479         if (irq < NR_IRQS_LEGACY)
 480                 return;
 481
 482         irq_free_desc(irq);
 483 }
 484
 485 static void pirq_query_unmask(int irq)
 486 {
 487         struct physdev_irq_status_query irq_status;
 488         struct irq_info *info = info_for_irq(irq);
 489
 490         BUG_ON(info->type != IRQT_PIRQ);
 491
 492         irq_status.irq = pirq_from_irq(irq);
 493         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
 494                 irq_status.flags = 0;
 495
 496         info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
 497         if (irq_status.flags & XENIRQSTAT_needs_eoi)
 498                 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
 499 }
 500
 501 static bool probing_irq(int irq)
 502 {
 503         struct irq_desc *desc = irq_to_desc(irq);
 504
 505         return desc && desc->action == NULL;
 506 }
 507
 508 static void eoi_pirq(struct irq_data *data)
 509 {
 510         int evtchn = evtchn_from_irq(data->irq);
 511         struct physdev_eoi eoi = { .irq = pirq_from_irq(data->irq) };
 512         int rc = 0;
 513
 514         irq_move_irq(data);
 515
 516         if (VALID_EVTCHN(evtchn))
 517                 clear_evtchn(evtchn);
 518
 519         if (pirq_needs_eoi(data->irq)) {
 520                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
 521                 WARN_ON(rc);
 522         }
 523 }
 524
 525 static void mask_ack_pirq(struct irq_data *data)
 526 {
 527         disable_dynirq(data);
 528         eoi_pirq(data);
 529 }
 530
 531 static unsigned int __startup_pirq(unsigned int irq)
 532 {
 533         struct evtchn_bind_pirq bind_pirq;
 534         struct irq_info *info = info_for_irq(irq);
 535         int evtchn = evtchn_from_irq(irq);
 536         int rc;
 537
 538         BUG_ON(info->type != IRQT_PIRQ);
 539
 540         if (VALID_EVTCHN(evtchn))
 541                 goto out;
 542
 543         bind_pirq.pirq = pirq_from_irq(irq);
 544         /* NB. We are happy to share unless we are probing. */
 545         bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
 546                                         BIND_PIRQ__WILL_SHARE : 0;
 547         rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
 548         if (rc != 0) {
 549                 if (!probing_irq(irq))
 550                         printk(KERN_INFO "Failed to obtain physical IRQ %d\n",
 551                                irq);
 552                 return 0;
 553         }
 554         evtchn = bind_pirq.port;
 555
 556         pirq_query_unmask(irq);
 557
 558         evtchn_to_irq[evtchn] = irq;
 559         bind_evtchn_to_cpu(evtchn, 0);
 560         info->evtchn = evtchn;
 561
 562 out:
 563         unmask_evtchn(evtchn);
 564         eoi_pirq(irq_get_irq_data(irq));
 565
 566         return 0;
 567 }
 568
 569 static unsigned int startup_pirq(struct irq_data *data)
 570 {
 571         return __startup_pirq(data->irq);
 572 }
 573
 574 static void shutdown_pirq(struct irq_data *data)
 575 {
 576         struct evtchn_close close;
 577         unsigned int irq = data->irq;
 578         struct irq_info *info = info_for_irq(irq);
 579         int evtchn = evtchn_from_irq(irq);
 580
 581         BUG_ON(info->type != IRQT_PIRQ);
 582
 583         if (!VALID_EVTCHN(evtchn))
 584                 return;
 585
 586         mask_evtchn(evtchn);
 587
 588         close.port = evtchn;
 589         if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 590                 BUG();
 591
 592         bind_evtchn_to_cpu(evtchn, 0);
 593         evtchn_to_irq[evtchn] = -1;
 594         info->evtchn = 0;
 595 }
 596
 597 static void enable_pirq(struct irq_data *data)
 598 {
 599         startup_pirq(data);
 600 }
 601
 602 static void disable_pirq(struct irq_data *data)
 603 {
 604         disable_dynirq(data);
 605 }
 606
 607 static int find_irq_by_gsi(unsigned gsi)
 608 {
 609         struct irq_info *info;
 610
 611         list_for_each_entry(info, &xen_irq_list_head, list) {
 612                 if (info->type != IRQT_PIRQ)
 613                         continue;
 614
 615                 if (info->u.pirq.gsi == gsi)
 616                         return info->irq;
 617         }
 618
 619         return -1;
 620 }
 621
 622 /*
 623  * Do not make any assumptions regarding the relationship between the
 624  * IRQ number returned here and the Xen pirq argument.
 625  *
 626  * Note: We don't assign an event channel until the irq actually started
 627  * up.  Return an existing irq if we've already got one for the gsi.
 628  *
 629  * Shareable implies level triggered, not shareable implies edge
 630  * triggered here.
 631  */
 632 int xen_bind_pirq_gsi_to_irq(unsigned gsi,
 633                              unsigned pirq, int shareable, char *name)
 634 {
 635         int irq = -1;
 636         struct physdev_irq irq_op;
 637
 638         mutex_lock(&irq_mapping_update_lock);
 639
 640         irq = find_irq_by_gsi(gsi);
 641         if (irq != -1) {
 642                 printk(KERN_INFO "xen_map_pirq_gsi: returning irq %d for gsi %u\n",
 643                        irq, gsi);
 644                 goto out;
 645         }
 646
 647         irq = xen_allocate_irq_gsi(gsi);
 648         if (irq < 0)
 649                 goto out;
 650
 651         irq_op.irq = irq;
 652         irq_op.vector = 0;
 653
 654         /* Only the privileged domain can do this. For non-priv, the pcifront
 655          * driver provides a PCI bus that does the call to do exactly
 656          * this in the priv domain. */
 657         if (xen_initial_domain() &&
 658             HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
 659                 xen_free_irq(irq);
 660                 irq = -ENOSPC;
 661                 goto out;
 662         }
 663
 664         xen_irq_info_pirq_init(irq, 0, pirq, gsi, irq_op.vector, DOMID_SELF,
 665                                shareable ? PIRQ_SHAREABLE : 0);
 666
 667         pirq_query_unmask(irq);
 668         /* We try to use the handler with the appropriate semantic for the
 669          * type of interrupt: if the interrupt is an edge triggered
 670          * interrupt we use handle_edge_irq.
 671          *
 672          * On the other hand if the interrupt is level triggered we use
 673          * handle_fasteoi_irq like the native code does for this kind of
 674          * interrupts.
 675          *
 676          * Depending on the Xen version, pirq_needs_eoi might return true
 677          * not only for level triggered interrupts but for edge triggered
 678          * interrupts too. In any case Xen always honors the eoi mechanism,
 679          * not injecting any more pirqs of the same kind if the first one
 680          * hasn't received an eoi yet. Therefore using the fasteoi handler
 681          * is the right choice either way.
 682          */
 683         if (shareable)
 684                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 685                                 handle_fasteoi_irq, name);
 686         else
 687                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 688                                 handle_edge_irq, name);
 689
 690 out:
 691         mutex_unlock(&irq_mapping_update_lock);
 692
 693         return irq;
 694 }
 695
 696 #ifdef CONFIG_PCI_MSI
 697 int xen_allocate_pirq_msi(struct pci_dev *dev, struct msi_desc *msidesc)
 698 {
 699         int rc;
 700         struct physdev_get_free_pirq op_get_free_pirq;
 701
 702         op_get_free_pirq.type = MAP_PIRQ_TYPE_MSI;
 703         rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
 704
 705         WARN_ONCE(rc == -ENOSYS,
 706                   "hypervisor does not support the PHYSDEVOP_get_free_pirq interface\n");
 707
 708         return rc ? -1 : op_get_free_pirq.pirq;
 709 }
 710
 711 int xen_bind_pirq_msi_to_irq(struct pci_dev *dev, struct msi_desc *msidesc,
 712                              int pirq, int vector, const char *name,
 713                              domid_t domid)
 714 {
 715         int irq, ret;
 716
 717         mutex_lock(&irq_mapping_update_lock);
 718
 719         irq = xen_allocate_irq_dynamic();
 720         if (irq < 0)
 721                 goto out;
 722
 723         irq_set_chip_and_handler_name(irq, &xen_pirq_chip, handle_edge_irq,
 724                         name);
 725
 726         xen_irq_info_pirq_init(irq, 0, pirq, 0, vector, domid, 0);
 727         ret = irq_set_msi_desc(irq, msidesc);
 728         if (ret < 0)
 729                 goto error_irq;
 730 out:
 731         mutex_unlock(&irq_mapping_update_lock);
 732         return irq;
 733 error_irq:
 734         mutex_unlock(&irq_mapping_update_lock);
 735         xen_free_irq(irq);
 736         return ret;
 737 }
 738 #endif
 739
 740 int xen_destroy_irq(int irq)
 741 {
 742         struct irq_desc *desc;
 743         struct physdev_unmap_pirq unmap_irq;
 744         struct irq_info *info = info_for_irq(irq);
 745         int rc = -ENOENT;
 746
 747         mutex_lock(&irq_mapping_update_lock);
 748
 749         desc = irq_to_desc(irq);
 750         if (!desc)
 751                 goto out;
 752
 753         if (xen_initial_domain()) {
 754                 unmap_irq.pirq = info->u.pirq.pirq;
 755                 unmap_irq.domid = info->u.pirq.domid;
 756                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
 757                 /* If another domain quits without making the pci_disable_msix
 758                  * call, the Xen hypervisor takes care of freeing the PIRQs
 759                  * (free_domain_pirqs).
 760                  */
 761                 if ((rc == -ESRCH && info->u.pirq.domid != DOMID_SELF))
 762                         printk(KERN_INFO "domain %d does not have %d anymore\n",
 763                                 info->u.pirq.domid, info->u.pirq.pirq);
 764                 else if (rc) {
 765                         printk(KERN_WARNING "unmap irq failed %d\n", rc);
 766                         goto out;
 767                 }
 768         }
 769
 770         xen_free_irq(irq);
 771
 772 out:
 773         mutex_unlock(&irq_mapping_update_lock);
 774         return rc;
 775 }
 776
 777 int xen_irq_from_pirq(unsigned pirq)
 778 {
 779         int irq;
 780
 781         struct irq_info *info;
 782
 783         mutex_lock(&irq_mapping_update_lock);
 784
 785         list_for_each_entry(info, &xen_irq_list_head, list) {
 786                 if (info->type != IRQT_PIRQ)
 787                         continue;
 788                 irq = info->irq;
 789                 if (info->u.pirq.pirq == pirq)
 790                         goto out;
 791         }
 792         irq = -1;
 793 out:
 794         mutex_unlock(&irq_mapping_update_lock);
 795
 796         return irq;
 797 }
 798
 799
 800 int xen_pirq_from_irq(unsigned irq)
 801 {
 802         return pirq_from_irq(irq);
 803 }
 804 EXPORT_SYMBOL_GPL(xen_pirq_from_irq);
 805 int bind_evtchn_to_irq(unsigned int evtchn)
 806 {
 807         int irq;
 808
 809         mutex_lock(&irq_mapping_update_lock);
 810
 811         irq = evtchn_to_irq[evtchn];
 812
 813         if (irq == -1) {
 814                 irq = xen_allocate_irq_dynamic();
 815                 if (irq == -1)
 816                         goto out;
 817
 818                 irq_set_chip_and_handler_name(irq, &xen_dynamic_chip,
 819                                               handle_edge_irq, "event");
 820
 821                 xen_irq_info_evtchn_init(irq, evtchn);
 822         }
 823
 824 out:
 825         mutex_unlock(&irq_mapping_update_lock);
 826
 827         return irq;
 828 }
 829 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
 830
 831 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
 832 {
 833         struct evtchn_bind_ipi bind_ipi;
 834         int evtchn, irq;
 835
 836         mutex_lock(&irq_mapping_update_lock);
 837
 838         irq = per_cpu(ipi_to_irq, cpu)[ipi];
 839
 840         if (irq == -1) {
 841                 irq = xen_allocate_irq_dynamic();
 842                 if (irq < 0)
 843                         goto out;
 844
 845                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 846                                               handle_percpu_irq, "ipi");
 847
 848                 bind_ipi.vcpu = cpu;
 849                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
 850                                                 &bind_ipi) != 0)
 851                         BUG();
 852                 evtchn = bind_ipi.port;
 853
 854                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
 855
 856                 bind_evtchn_to_cpu(evtchn, cpu);
 857         }
 858
 859  out:
 860         mutex_unlock(&irq_mapping_update_lock);
 861         return irq;
 862 }
 863
 864 static int bind_interdomain_evtchn_to_irq(unsigned int remote_domain,
 865                                           unsigned int remote_port)
 866 {
 867         struct evtchn_bind_interdomain bind_interdomain;
 868         int err;
 869
 870         bind_interdomain.remote_dom  = remote_domain;
 871         bind_interdomain.remote_port = remote_port;
 872
 873         err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
 874                                           &bind_interdomain);
 875
 876         return err ? : bind_evtchn_to_irq(bind_interdomain.local_port);
 877 }
 878
 879 static int find_virq(unsigned int virq, unsigned int cpu)
 880 {
 881         struct evtchn_status status;
 882         int port, rc = -ENOENT;
 883
 884         memset(&status, 0, sizeof(status));
 885         for (port = 0; port <= NR_EVENT_CHANNELS; port++) {
 886                 status.dom = DOMID_SELF;
 887                 status.port = port;
 888                 rc = HYPERVISOR_event_channel_op(EVTCHNOP_status, &status);
 889                 if (rc < 0)
 890                         continue;
 891                 if (status.status != EVTCHNSTAT_virq)
 892                         continue;
 893                 if (status.u.virq == virq && status.vcpu == cpu) {
 894                         rc = port;
 895                         break;
 896                 }
 897         }
 898         return rc;
 899 }
 900
 901 int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
 902 {
 903         struct evtchn_bind_virq bind_virq;
 904         int evtchn, irq, ret;
 905
 906         mutex_lock(&irq_mapping_update_lock);
 907
 908         irq = per_cpu(virq_to_irq, cpu)[virq];
 909
 910         if (irq == -1) {
 911                 irq = xen_allocate_irq_dynamic();
 912                 if (irq == -1)
 913                         goto out;
 914
 915                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 916                                               handle_percpu_irq, "virq");
 917
 918                 bind_virq.virq = virq;
 919                 bind_virq.vcpu = cpu;
 920                 ret = HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 921                                                 &bind_virq);
 922                 if (ret == 0)
 923                         evtchn = bind_virq.port;
 924                 else {
 925                         if (ret == -EEXIST)
 926                                 ret = find_virq(virq, cpu);
 927                         BUG_ON(ret < 0);
 928                         evtchn = ret;
 929                 }
 930
 931                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
 932
 933                 bind_evtchn_to_cpu(evtchn, cpu);
 934         }
 935
 936 out:
 937         mutex_unlock(&irq_mapping_update_lock);
 938
 939         return irq;
 940 }
 941
 942 static void unbind_from_irq(unsigned int irq)
 943 {
 944         struct evtchn_close close;
 945         int evtchn = evtchn_from_irq(irq);
 946         struct irq_info *info = irq_get_handler_data(irq);
 947
 948         mutex_lock(&irq_mapping_update_lock);
 949
 950         if (info->refcnt > 0) {
 951                 info->refcnt--;
 952                 if (info->refcnt != 0)
 953                         goto done;
 954         }
 955
 956         if (VALID_EVTCHN(evtchn)) {
 957                 close.port = evtchn;
 958                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 959                         BUG();
 960
 961                 switch (type_from_irq(irq)) {
 962                 case IRQT_VIRQ:
 963                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
 964                                 [virq_from_irq(irq)] = -1;
 965                         break;
 966                 case IRQT_IPI:
 967                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
 968                                 [ipi_from_irq(irq)] = -1;
 969                         break;
 970                 default:
 971                         break;
 972                 }
 973
 974                 /* Closed ports are implicitly re-bound to VCPU0. */
 975                 bind_evtchn_to_cpu(evtchn, 0);
 976
 977                 evtchn_to_irq[evtchn] = -1;
 978         }
 979
 980         BUG_ON(info_for_irq(irq)->type == IRQT_UNBOUND);
 981
 982         xen_free_irq(irq);
 983
 984  done:
 985         mutex_unlock(&irq_mapping_update_lock);
 986 }
 987
 988 int bind_evtchn_to_irqhandler(unsigned int evtchn,
 989                               irq_handler_t handler,
 990                               unsigned long irqflags,
 991                               const char *devname, void *dev_id)
 992 {
 993         int irq, retval;
 994
 995         irq = bind_evtchn_to_irq(evtchn);
 996         if (irq < 0)
 997                 return irq;
 998         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 999         if (retval != 0) {
1000                 unbind_from_irq(irq);
1001                 return retval;
1002         }
1003
1004         return irq;
1005 }
1006 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
1007
1008 int bind_interdomain_evtchn_to_irqhandler(unsigned int remote_domain,
1009                                           unsigned int remote_port,
1010                                           irq_handler_t handler,
1011                                           unsigned long irqflags,
1012                                           const char *devname,
1013                                           void *dev_id)
1014 {
1015         int irq, retval;
1016
1017         irq = bind_interdomain_evtchn_to_irq(remote_domain, remote_port);
1018         if (irq < 0)
1019                 return irq;
1020
1021         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1022         if (retval != 0) {
1023                 unbind_from_irq(irq);
1024                 return retval;
1025         }
1026
1027         return irq;
1028 }
1029 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler);
1030
1031 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
1032                             irq_handler_t handler,
1033                             unsigned long irqflags, const char *devname, void *dev_id)
1034 {
1035         int irq, retval;
1036
1037         irq = bind_virq_to_irq(virq, cpu);
1038         if (irq < 0)
1039                 return irq;
1040         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1041         if (retval != 0) {
1042                 unbind_from_irq(irq);
1043                 return retval;
1044         }
1045
1046         return irq;
1047 }
1048 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1049
1050 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1051                            unsigned int cpu,
1052                            irq_handler_t handler,
1053                            unsigned long irqflags,
1054                            const char *devname,
1055                            void *dev_id)
1056 {
1057         int irq, retval;
1058
1059         irq = bind_ipi_to_irq(ipi, cpu);
1060         if (irq < 0)
1061                 return irq;
1062
1063         irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME | IRQF_EARLY_RESUME;
1064         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1065         if (retval != 0) {
1066                 unbind_from_irq(irq);
1067                 return retval;
1068         }
1069
1070         return irq;
1071 }
1072
1073 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1074 {
1075         free_irq(irq, dev_id);
1076         unbind_from_irq(irq);
1077 }
1078 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1079
1080 int evtchn_make_refcounted(unsigned int evtchn)
1081 {
1082         int irq = evtchn_to_irq[evtchn];
1083         struct irq_info *info;
1084
1085         if (irq == -1)
1086                 return -ENOENT;
1087
1088         info = irq_get_handler_data(irq);
1089
1090         if (!info)
1091                 return -ENOENT;
1092
1093         WARN_ON(info->refcnt != -1);
1094
1095         info->refcnt = 1;
1096
1097         return 0;
1098 }
1099 EXPORT_SYMBOL_GPL(evtchn_make_refcounted);
1100
1101 int evtchn_get(unsigned int evtchn)
1102 {
1103         int irq;
1104         struct irq_info *info;
1105         int err = -ENOENT;
1106
1107         if (evtchn >= NR_EVENT_CHANNELS)
1108                 return -EINVAL;
1109
1110         mutex_lock(&irq_mapping_update_lock);
1111
1112         irq = evtchn_to_irq[evtchn];
1113         if (irq == -1)
1114                 goto done;
1115
1116         info = irq_get_handler_data(irq);
1117
1118         if (!info)
1119                 goto done;
1120
1121         err = -EINVAL;
1122         if (info->refcnt <= 0)
1123                 goto done;
1124
1125         info->refcnt++;
1126         err = 0;
1127  done:
1128         mutex_unlock(&irq_mapping_update_lock);
1129
1130         return err;
1131 }
1132 EXPORT_SYMBOL_GPL(evtchn_get);
1133
1134 void evtchn_put(unsigned int evtchn)
1135 {
1136         int irq = evtchn_to_irq[evtchn];
1137         if (WARN_ON(irq == -1))
1138                 return;
1139         unbind_from_irq(irq);
1140 }
1141 EXPORT_SYMBOL_GPL(evtchn_put);
1142
1143 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1144 {
1145         int irq = per_cpu(ipi_to_irq, cpu)[vector];
1146         BUG_ON(irq < 0);
1147         notify_remote_via_irq(irq);
1148 }
1149
1150 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
1151 {
1152         struct shared_info *sh = HYPERVISOR_shared_info;
1153         int cpu = smp_processor_id();
1154         unsigned long *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu);
1155         int i;
1156         unsigned long flags;
1157         static DEFINE_SPINLOCK(debug_lock);
1158         struct vcpu_info *v;
1159
1160         spin_lock_irqsave(&debug_lock, flags);
1161
1162         printk("\nvcpu %d\n  ", cpu);
1163
1164         for_each_online_cpu(i) {
1165                 int pending;
1166                 v = per_cpu(xen_vcpu, i);
1167                 pending = (get_irq_regs() && i == cpu)
1168                         ? xen_irqs_disabled(get_irq_regs())
1169                         : v->evtchn_upcall_mask;
1170                 printk("%d: masked=%d pending=%d event_sel %0*lx\n  ", i,
1171                        pending, v->evtchn_upcall_pending,
1172                        (int)(sizeof(v->evtchn_pending_sel)*2),
1173                        v->evtchn_pending_sel);
1174         }
1175         v = per_cpu(xen_vcpu, cpu);
1176
1177         printk("\npending:\n   ");
1178         for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
1179                 printk("%0*lx%s", (int)sizeof(sh->evtchn_pending[0])*2,
1180                        sh->evtchn_pending[i],
1181                        i % 8 == 0 ? "\n   " : " ");
1182         printk("\nglobal mask:\n   ");
1183         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1184                 printk("%0*lx%s",
1185                        (int)(sizeof(sh->evtchn_mask[0])*2),
1186                        sh->evtchn_mask[i],
1187                        i % 8 == 0 ? "\n   " : " ");
1188
1189         printk("\nglobally unmasked:\n   ");
1190         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1191                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1192                        sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
1193                        i % 8 == 0 ? "\n   " : " ");
1194
1195         printk("\nlocal cpu%d mask:\n   ", cpu);
1196         for (i = (NR_EVENT_CHANNELS/BITS_PER_LONG)-1; i >= 0; i--)
1197                 printk("%0*lx%s", (int)(sizeof(cpu_evtchn[0])*2),
1198                        cpu_evtchn[i],
1199                        i % 8 == 0 ? "\n   " : " ");
1200
1201         printk("\nlocally unmasked:\n   ");
1202         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
1203                 unsigned long pending = sh->evtchn_pending[i]
1204                         & ~sh->evtchn_mask[i]
1205                         & cpu_evtchn[i];
1206                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1207                        pending, i % 8 == 0 ? "\n   " : " ");
1208         }
1209
1210         printk("\npending list:\n");
1211         for (i = 0; i < NR_EVENT_CHANNELS; i++) {
1212                 if (sync_test_bit(i, sh->evtchn_pending)) {
1213                         int word_idx = i / BITS_PER_LONG;
1214                         printk("  %d: event %d -> irq %d%s%s%s\n",
1215                                cpu_from_evtchn(i), i,
1216                                evtchn_to_irq[i],
1217                                sync_test_bit(word_idx, &v->evtchn_pending_sel)
1218                                              ? "" : " l2-clear",
1219                                !sync_test_bit(i, sh->evtchn_mask)
1220                                              ? "" : " globally-masked",
1221                                sync_test_bit(i, cpu_evtchn)
1222                                              ? "" : " locally-masked");
1223                 }
1224         }
1225
1226         spin_unlock_irqrestore(&debug_lock, flags);
1227
1228         return IRQ_HANDLED;
1229 }
1230
1231 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
1232 static DEFINE_PER_CPU(unsigned int, current_word_idx);
1233 static DEFINE_PER_CPU(unsigned int, current_bit_idx);
1234
1235 /*
1236  * Mask out the i least significant bits of w
1237  */
1238 #define MASK_LSBS(w, i) (w & ((~0UL) << i))
1239
1240 /*
1241  * Search the CPUs pending events bitmasks.  For each one found, map
1242  * the event number to an irq, and feed it into do_IRQ() for
1243  * handling.
1244  *
1245  * Xen uses a two-level bitmap to speed searching.  The first level is
1246  * a bitset of words which contain pending event bits.  The second
1247  * level is a bitset of pending events themselves.
1248  */
1249 static void __xen_evtchn_do_upcall(void)
1250 {
1251         int start_word_idx, start_bit_idx;
1252         int word_idx, bit_idx;
1253         int i;
1254         int cpu = get_cpu();
1255         struct shared_info *s = HYPERVISOR_shared_info;
1256         struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1257         unsigned count;
1258
1259         do {
1260                 unsigned long pending_words;
1261
1262                 vcpu_info->evtchn_upcall_pending = 0;
1263
1264                 if (__this_cpu_inc_return(xed_nesting_count) - 1)
1265                         goto out;
1266
1267 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
1268                 /* Clear master flag /before/ clearing selector flag. */
1269                 wmb();
1270 #endif
1271                 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
1272
1273                 start_word_idx = __this_cpu_read(current_word_idx);
1274                 start_bit_idx = __this_cpu_read(current_bit_idx);
1275
1276                 word_idx = start_word_idx;
1277
1278                 for (i = 0; pending_words != 0; i++) {
1279                         unsigned long pending_bits;
1280                         unsigned long words;
1281
1282                         words = MASK_LSBS(pending_words, word_idx);
1283
1284                         /*
1285                          * If we masked out all events, wrap to beginning.
1286                          */
1287                         if (words == 0) {
1288                                 word_idx = 0;
1289                                 bit_idx = 0;
1290                                 continue;
1291                         }
1292                         word_idx = __ffs(words);
1293
1294                         pending_bits = active_evtchns(cpu, s, word_idx);
1295                         bit_idx = 0; /* usually scan entire word from start */
1296                         if (word_idx == start_word_idx) {
1297                                 /* We scan the starting word in two parts */
1298                                 if (i == 0)
1299                                         /* 1st time: start in the middle */
1300                                         bit_idx = start_bit_idx;
1301                                 else
1302                                         /* 2nd time: mask bits done already */
1303                                         bit_idx &= (1UL << start_bit_idx) - 1;
1304                         }
1305
1306                         do {
1307                                 unsigned long bits;
1308                                 int port, irq;
1309                                 struct irq_desc *desc;
1310
1311                                 bits = MASK_LSBS(pending_bits, bit_idx);
1312
1313                                 /* If we masked out all events, move on. */
1314                                 if (bits == 0)
1315                                         break;
1316
1317                                 bit_idx = __ffs(bits);
1318
1319                                 /* Process port. */
1320                                 port = (word_idx * BITS_PER_LONG) + bit_idx;
1321                                 irq = evtchn_to_irq[port];
1322
1323                                 if (irq != -1) {
1324                                         desc = irq_to_desc(irq);
1325                                         if (desc)
1326                                                 generic_handle_irq_desc(irq, desc);
1327                                 }
1328
1329                                 bit_idx = (bit_idx + 1) % BITS_PER_LONG;
1330
1331                                 /* Next caller starts at last processed + 1 */
1332                                 __this_cpu_write(current_word_idx,
1333                                                  bit_idx ? word_idx :
1334                                                  (word_idx+1) % BITS_PER_LONG);
1335                                 __this_cpu_write(current_bit_idx, bit_idx);
1336                         } while (bit_idx != 0);
1337
1338                         /* Scan start_l1i twice; all others once. */
1339                         if ((word_idx != start_word_idx) || (i != 0))
1340                                 pending_words &= ~(1UL << word_idx);
1341
1342                         word_idx = (word_idx + 1) % BITS_PER_LONG;
1343                 }
1344
1345                 BUG_ON(!irqs_disabled());
1346
1347                 count = __this_cpu_read(xed_nesting_count);
1348                 __this_cpu_write(xed_nesting_count, 0);
1349         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
1350
1351 out:
1352
1353         put_cpu();
1354 }
1355
1356 void xen_evtchn_do_upcall(struct pt_regs *regs)
1357 {
1358         struct pt_regs *old_regs = set_irq_regs(regs);
1359
1360         exit_idle();
1361         irq_enter();
1362
1363         __xen_evtchn_do_upcall();
1364
1365         irq_exit();
1366         set_irq_regs(old_regs);
1367 }
1368
1369 void xen_hvm_evtchn_do_upcall(void)
1370 {
1371         __xen_evtchn_do_upcall();
1372 }
1373 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1374
1375 /* Rebind a new event channel to an existing irq. */
1376 void rebind_evtchn_irq(int evtchn, int irq)
1377 {
1378         struct irq_info *info = info_for_irq(irq);
1379
1380         /* Make sure the irq is masked, since the new event channel
1381            will also be masked. */
1382         disable_irq(irq);
1383
1384         mutex_lock(&irq_mapping_update_lock);
1385
1386         /* After resume the irq<->evtchn mappings are all cleared out */
1387         BUG_ON(evtchn_to_irq[evtchn] != -1);
1388         /* Expect irq to have been bound before,
1389            so there should be a proper type */
1390         BUG_ON(info->type == IRQT_UNBOUND);
1391
1392         xen_irq_info_evtchn_init(irq, evtchn);
1393
1394         mutex_unlock(&irq_mapping_update_lock);
1395
1396         /* new event channels are always bound to cpu 0 */
1397         irq_set_affinity(irq, cpumask_of(0));
1398
1399         /* Unmask the event channel. */
1400         enable_irq(irq);
1401 }
1402
1403 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1404 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1405 {
1406         struct evtchn_bind_vcpu bind_vcpu;
1407         int evtchn = evtchn_from_irq(irq);
1408
1409         if (!VALID_EVTCHN(evtchn))
1410                 return -1;
1411
1412         /*
1413          * Events delivered via platform PCI interrupts are always
1414          * routed to vcpu 0 and hence cannot be rebound.
1415          */
1416         if (xen_hvm_domain() && !xen_have_vector_callback)
1417                 return -1;
1418
1419         /* Send future instances of this interrupt to other vcpu. */
1420         bind_vcpu.port = evtchn;
1421         bind_vcpu.vcpu = tcpu;
1422
1423         /*
1424          * If this fails, it usually just indicates that we're dealing with a
1425          * virq or IPI channel, which don't actually need to be rebound. Ignore
1426          * it, but don't do the xenlinux-level rebind in that case.
1427          */
1428         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1429                 bind_evtchn_to_cpu(evtchn, tcpu);
1430
1431         return 0;
1432 }
1433
1434 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1435                             bool force)
1436 {
1437         unsigned tcpu = cpumask_first(dest);
1438
1439         return rebind_irq_to_cpu(data->irq, tcpu);
1440 }
1441
1442 int resend_irq_on_evtchn(unsigned int irq)
1443 {
1444         int masked, evtchn = evtchn_from_irq(irq);
1445         struct shared_info *s = HYPERVISOR_shared_info;
1446
1447         if (!VALID_EVTCHN(evtchn))
1448                 return 1;
1449
1450         masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
1451         sync_set_bit(evtchn, s->evtchn_pending);
1452         if (!masked)
1453                 unmask_evtchn(evtchn);
1454
1455         return 1;
1456 }
1457
1458 static void enable_dynirq(struct irq_data *data)
1459 {
1460         int evtchn = evtchn_from_irq(data->irq);
1461
1462         if (VALID_EVTCHN(evtchn))
1463                 unmask_evtchn(evtchn);
1464 }
1465
1466 static void disable_dynirq(struct irq_data *data)
1467 {
1468         int evtchn = evtchn_from_irq(data->irq);
1469
1470         if (VALID_EVTCHN(evtchn))
1471                 mask_evtchn(evtchn);
1472 }
1473
1474 static void ack_dynirq(struct irq_data *data)
1475 {
1476         int evtchn = evtchn_from_irq(data->irq);
1477
1478         irq_move_irq(data);
1479
1480         if (VALID_EVTCHN(evtchn))
1481                 clear_evtchn(evtchn);
1482 }
1483
1484 static void mask_ack_dynirq(struct irq_data *data)
1485 {
1486         disable_dynirq(data);
1487         ack_dynirq(data);
1488 }
1489
1490 static int retrigger_dynirq(struct irq_data *data)
1491 {
1492         int evtchn = evtchn_from_irq(data->irq);
1493         struct shared_info *sh = HYPERVISOR_shared_info;
1494         int ret = 0;
1495
1496         if (VALID_EVTCHN(evtchn)) {
1497                 int masked;
1498
1499                 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
1500                 sync_set_bit(evtchn, sh->evtchn_pending);
1501                 if (!masked)
1502                         unmask_evtchn(evtchn);
1503                 ret = 1;
1504         }
1505
1506         return ret;
1507 }
1508
1509 static void restore_pirqs(void)
1510 {
1511         int pirq, rc, irq, gsi;
1512         struct physdev_map_pirq map_irq;
1513         struct irq_info *info;
1514
1515         list_for_each_entry(info, &xen_irq_list_head, list) {
1516                 if (info->type != IRQT_PIRQ)
1517                         continue;
1518
1519                 pirq = info->u.pirq.pirq;
1520                 gsi = info->u.pirq.gsi;
1521                 irq = info->irq;
1522
1523                 /* save/restore of PT devices doesn't work, so at this point the
1524                  * only devices present are GSI based emulated devices */
1525                 if (!gsi)
1526                         continue;
1527
1528                 map_irq.domid = DOMID_SELF;
1529                 map_irq.type = MAP_PIRQ_TYPE_GSI;
1530                 map_irq.index = gsi;
1531                 map_irq.pirq = pirq;
1532
1533                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1534                 if (rc) {
1535                         printk(KERN_WARNING "xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1536                                         gsi, irq, pirq, rc);
1537                         xen_free_irq(irq);
1538                         continue;
1539                 }
1540
1541                 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1542
1543                 __startup_pirq(irq);
1544         }
1545 }
1546
1547 static void restore_cpu_virqs(unsigned int cpu)
1548 {
1549         struct evtchn_bind_virq bind_virq;
1550         int virq, irq, evtchn;
1551
1552         for (virq = 0; virq < NR_VIRQS; virq++) {
1553                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1554                         continue;
1555
1556                 BUG_ON(virq_from_irq(irq) != virq);
1557
1558                 /* Get a new binding from Xen. */
1559                 bind_virq.virq = virq;
1560                 bind_virq.vcpu = cpu;
1561                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1562                                                 &bind_virq) != 0)
1563                         BUG();
1564                 evtchn = bind_virq.port;
1565
1566                 /* Record the new mapping. */
1567                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
1568                 bind_evtchn_to_cpu(evtchn, cpu);
1569         }
1570 }
1571
1572 static void restore_cpu_ipis(unsigned int cpu)
1573 {
1574         struct evtchn_bind_ipi bind_ipi;
1575         int ipi, irq, evtchn;
1576
1577         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1578                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1579                         continue;
1580
1581                 BUG_ON(ipi_from_irq(irq) != ipi);
1582
1583                 /* Get a new binding from Xen. */
1584                 bind_ipi.vcpu = cpu;
1585                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1586                                                 &bind_ipi) != 0)
1587                         BUG();
1588                 evtchn = bind_ipi.port;
1589
1590                 /* Record the new mapping. */
1591                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
1592                 bind_evtchn_to_cpu(evtchn, cpu);
1593         }
1594 }
1595
1596 /* Clear an irq's pending state, in preparation for polling on it */
1597 void xen_clear_irq_pending(int irq)
1598 {
1599         int evtchn = evtchn_from_irq(irq);
1600
1601         if (VALID_EVTCHN(evtchn))
1602                 clear_evtchn(evtchn);
1603 }
1604 EXPORT_SYMBOL(xen_clear_irq_pending);
1605 void xen_set_irq_pending(int irq)
1606 {
1607         int evtchn = evtchn_from_irq(irq);
1608
1609         if (VALID_EVTCHN(evtchn))
1610                 set_evtchn(evtchn);
1611 }
1612
1613 bool xen_test_irq_pending(int irq)
1614 {
1615         int evtchn = evtchn_from_irq(irq);
1616         bool ret = false;
1617
1618         if (VALID_EVTCHN(evtchn))
1619                 ret = test_evtchn(evtchn);
1620
1621         return ret;
1622 }
1623
1624 /* Poll waiting for an irq to become pending with timeout.  In the usual case,
1625  * the irq will be disabled so it won't deliver an interrupt. */
1626 void xen_poll_irq_timeout(int irq, u64 timeout)
1627 {
1628         evtchn_port_t evtchn = evtchn_from_irq(irq);
1629
1630         if (VALID_EVTCHN(evtchn)) {
1631                 struct sched_poll poll;
1632
1633                 poll.nr_ports = 1;
1634                 poll.timeout = timeout;
1635                 set_xen_guest_handle(poll.ports, &evtchn);
1636
1637                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1638                         BUG();
1639         }
1640 }
1641 EXPORT_SYMBOL(xen_poll_irq_timeout);
1642 /* Poll waiting for an irq to become pending.  In the usual case, the
1643  * irq will be disabled so it won't deliver an interrupt. */
1644 void xen_poll_irq(int irq)
1645 {
1646         xen_poll_irq_timeout(irq, 0 /* no timeout */);
1647 }
1648
1649 /* Check whether the IRQ line is shared with other guests. */
1650 int xen_test_irq_shared(int irq)
1651 {
1652         struct irq_info *info = info_for_irq(irq);
1653         struct physdev_irq_status_query irq_status = { .irq = info->u.pirq.pirq };
1654
1655         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
1656                 return 0;
1657         return !(irq_status.flags & XENIRQSTAT_shared);
1658 }
1659 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
1660
1661 void xen_irq_resume(void)
1662 {
1663         unsigned int cpu, evtchn;
1664         struct irq_info *info;
1665
1666         init_evtchn_cpu_bindings();
1667
1668         /* New event-channel space is not 'live' yet. */
1669         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1670                 mask_evtchn(evtchn);
1671
1672         /* No IRQ <-> event-channel mappings. */
1673         list_for_each_entry(info, &xen_irq_list_head, list)
1674                 info->evtchn = 0; /* zap event-channel binding */
1675
1676         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1677                 evtchn_to_irq[evtchn] = -1;
1678
1679         for_each_possible_cpu(cpu) {
1680                 restore_cpu_virqs(cpu);
1681                 restore_cpu_ipis(cpu);
1682         }
1683
1684         restore_pirqs();
1685 }
1686
1687 static struct irq_chip xen_dynamic_chip __read_mostly = {
1688         .name                   = "xen-dyn",
1689
1690         .irq_disable            = disable_dynirq,
1691         .irq_mask               = disable_dynirq,
1692         .irq_unmask             = enable_dynirq,
1693
1694         .irq_ack                = ack_dynirq,
1695         .irq_mask_ack           = mask_ack_dynirq,
1696
1697         .irq_set_affinity       = set_affinity_irq,
1698         .irq_retrigger          = retrigger_dynirq,
1699 };
1700
1701 static struct irq_chip xen_pirq_chip __read_mostly = {
1702         .name                   = "xen-pirq",
1703
1704         .irq_startup            = startup_pirq,
1705         .irq_shutdown           = shutdown_pirq,
1706         .irq_enable             = enable_pirq,
1707         .irq_disable            = disable_pirq,
1708
1709         .irq_mask               = disable_dynirq,
1710         .irq_unmask             = enable_dynirq,
1711
1712         .irq_ack                = eoi_pirq,
1713         .irq_eoi                = eoi_pirq,
1714         .irq_mask_ack           = mask_ack_pirq,
1715
1716         .irq_set_affinity       = set_affinity_irq,
1717
1718         .irq_retrigger          = retrigger_dynirq,
1719 };
1720
1721 static struct irq_chip xen_percpu_chip __read_mostly = {
1722         .name                   = "xen-percpu",
1723
1724         .irq_disable            = disable_dynirq,
1725         .irq_mask               = disable_dynirq,
1726         .irq_unmask             = enable_dynirq,
1727
1728         .irq_ack                = ack_dynirq,
1729 };
1730
1731 int xen_set_callback_via(uint64_t via)
1732 {
1733         struct xen_hvm_param a;
1734         a.domid = DOMID_SELF;
1735         a.index = HVM_PARAM_CALLBACK_IRQ;
1736         a.value = via;
1737         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1738 }
1739 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1740
1741 #ifdef CONFIG_XEN_PVHVM
1742 /* Vector callbacks are better than PCI interrupts to receive event
1743  * channel notifications because we can receive vector callbacks on any
1744  * vcpu and we don't need PCI support or APIC interactions. */
1745 void xen_callback_vector(void)
1746 {
1747         int rc;
1748         uint64_t callback_via;
1749         if (xen_have_vector_callback) {
1750                 callback_via = HVM_CALLBACK_VECTOR(XEN_HVM_EVTCHN_CALLBACK);
1751                 rc = xen_set_callback_via(callback_via);
1752                 if (rc) {
1753                         printk(KERN_ERR "Request for Xen HVM callback vector"
1754                                         " failed.\n");
1755                         xen_have_vector_callback = 0;
1756                         return;
1757                 }
1758                 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1759                                 "enabled\n");
1760                 /* in the restore case the vector has already been allocated */
1761                 if (!test_bit(XEN_HVM_EVTCHN_CALLBACK, used_vectors))
1762                         alloc_intr_gate(XEN_HVM_EVTCHN_CALLBACK, xen_hvm_callback_vector);
1763         }
1764 }
1765 #else
1766 void xen_callback_vector(void) {}
1767 #endif
1768
1769 void __init xen_init_IRQ(void)
1770 {
1771         int i;
1772
1773         evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1774                                     GFP_KERNEL);
1775         BUG_ON(!evtchn_to_irq);
1776         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1777                 evtchn_to_irq[i] = -1;
1778
1779         init_evtchn_cpu_bindings();
1780
1781         /* No event channels are 'live' right now. */
1782         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1783                 mask_evtchn(i);
1784
1785         if (xen_hvm_domain()) {
1786                 xen_callback_vector();
1787                 native_init_IRQ();
1788                 /* pci_xen_hvm_init must be called after native_init_IRQ so that
1789                  * __acpi_register_gsi can point at the right function */
1790                 pci_xen_hvm_init();
1791         } else {
1792                 irq_ctx_init(smp_processor_id());
1793                 if (xen_initial_domain())
1794                         pci_xen_initial_domain();
1795         }
1796 }