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 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
  25
  26 #include <linux/linkage.h>
  27 #include <linux/interrupt.h>
  28 #include <linux/irq.h>
  29 #include <linux/module.h>
  30 #include <linux/string.h>
  31 #include <linux/bootmem.h>
  32 #include <linux/slab.h>
  33 #include <linux/irqnr.h>
  34 #include <linux/pci.h>
  35
  36 #ifdef CONFIG_X86
  37 #include <asm/desc.h>
  38 #include <asm/ptrace.h>
  39 #include <asm/irq.h>
  40 #include <asm/idle.h>
  41 #include <asm/io_apic.h>
  42 #include <asm/xen/page.h>
  43 #include <asm/xen/pci.h>
  44 #endif
  45 #include <asm/sync_bitops.h>
  46 #include <asm/xen/hypercall.h>
  47 #include <asm/xen/hypervisor.h>
  48
  49 #include <xen/xen.h>
  50 #include <xen/hvm.h>
  51 #include <xen/xen-ops.h>
  52 #include <xen/events.h>
  53 #include <xen/interface/xen.h>
  54 #include <xen/interface/event_channel.h>
  55 #include <xen/interface/hvm/hvm_op.h>
  56 #include <xen/interface/hvm/params.h>
  57 #include <xen/interface/physdev.h>
  58 #include <xen/interface/sched.h>
  59 #include <asm/hw_irq.h>
  60
  61 /*
  62  * This lock protects updates to the following mapping and reference-count
  63  * arrays. The lock does not need to be acquired to read the mapping tables.
  64  */
  65 static DEFINE_MUTEX(irq_mapping_update_lock);
  66
  67 static LIST_HEAD(xen_irq_list_head);
  68
  69 /* IRQ <-> VIRQ mapping. */
  70 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
  71
  72 /* IRQ <-> IPI mapping */
  73 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
  74
  75 /* Interrupt types. */
  76 enum xen_irq_type {
  77         IRQT_UNBOUND = 0,
  78         IRQT_PIRQ,
  79         IRQT_VIRQ,
  80         IRQT_IPI,
  81         IRQT_EVTCHN
  82 };
  83
  84 /*
  85  * Packed IRQ information:
  86  * type - enum xen_irq_type
  87  * event channel - irq->event channel mapping
  88  * cpu - cpu this event channel is bound to
  89  * index - type-specific information:
  90  *    PIRQ - physical IRQ, GSI, flags, and owner domain
  91  *    VIRQ - virq number
  92  *    IPI - IPI vector
  93  *    EVTCHN -
  94  */
  95 struct irq_info {
  96         struct list_head list;
  97         int refcnt;
  98         enum xen_irq_type type; /* type */
  99         unsigned irq;
 100         unsigned short evtchn;  /* event channel */
 101         unsigned short cpu;     /* cpu bound */
 102
 103         union {
 104                 unsigned short virq;
 105                 enum ipi_vector ipi;
 106                 struct {
 107                         unsigned short pirq;
 108                         unsigned short gsi;
 109                         unsigned char flags;
 110                         uint16_t domid;
 111                 } pirq;
 112         } u;
 113 };
 114 #define PIRQ_NEEDS_EOI  (1 << 0)
 115 #define PIRQ_SHAREABLE  (1 << 1)
 116
 117 static int *evtchn_to_irq;
 118 #ifdef CONFIG_X86
 119 static unsigned long *pirq_eoi_map;
 120 #endif
 121 static bool (*pirq_needs_eoi)(unsigned irq);
 122
 123 /*
 124  * Note sizeof(xen_ulong_t) can be more than sizeof(unsigned long). Be
 125  * careful to only use bitops which allow for this (e.g
 126  * test_bit/find_first_bit and friends but not __ffs) and to pass
 127  * BITS_PER_EVTCHN_WORD as the bitmask length.
 128  */
 129 #define BITS_PER_EVTCHN_WORD (sizeof(xen_ulong_t)*8)
 130 /*
 131  * Make a bitmask (i.e. unsigned long *) of a xen_ulong_t
 132  * array. Primarily to avoid long lines (hence the terse name).
 133  */
 134 #define BM(x) (unsigned long *)(x)
 135 /* Find the first set bit in a evtchn mask */
 136 #define EVTCHN_FIRST_BIT(w) find_first_bit(BM(&(w)), BITS_PER_EVTCHN_WORD)
 137
 138 static DEFINE_PER_CPU(xen_ulong_t [NR_EVENT_CHANNELS/BITS_PER_EVTCHN_WORD],
 139                       cpu_evtchn_mask);
 140
 141 /* Xen will never allocate port zero for any purpose. */
 142 #define VALID_EVTCHN(chn)       ((chn) != 0)
 143
 144 static struct irq_chip xen_dynamic_chip;
 145 static struct irq_chip xen_percpu_chip;
 146 static struct irq_chip xen_pirq_chip;
 147 static void enable_dynirq(struct irq_data *data);
 148 static void disable_dynirq(struct irq_data *data);
 149
 150 /* Get info for IRQ */
 151 static struct irq_info *info_for_irq(unsigned irq)
 152 {
 153         return irq_get_handler_data(irq);
 154 }
 155
 156 /* Constructors for packed IRQ information. */
 157 static void xen_irq_info_common_init(struct irq_info *info,
 158                                      unsigned irq,
 159                                      enum xen_irq_type type,
 160                                      unsigned short evtchn,
 161                                      unsigned short cpu)
 162 {
 163
 164         BUG_ON(info->type != IRQT_UNBOUND && info->type != type);
 165
 166         info->type = type;
 167         info->irq = irq;
 168         info->evtchn = evtchn;
 169         info->cpu = cpu;
 170
 171         evtchn_to_irq[evtchn] = irq;
 172
 173         irq_clear_status_flags(irq, IRQ_NOREQUEST|IRQ_NOAUTOEN);
 174 }
 175
 176 static void xen_irq_info_evtchn_init(unsigned irq,
 177                                      unsigned short evtchn)
 178 {
 179         struct irq_info *info = info_for_irq(irq);
 180
 181         xen_irq_info_common_init(info, irq, IRQT_EVTCHN, evtchn, 0);
 182 }
 183
 184 static void xen_irq_info_ipi_init(unsigned cpu,
 185                                   unsigned irq,
 186                                   unsigned short evtchn,
 187                                   enum ipi_vector ipi)
 188 {
 189         struct irq_info *info = info_for_irq(irq);
 190
 191         xen_irq_info_common_init(info, irq, IRQT_IPI, evtchn, 0);
 192
 193         info->u.ipi = ipi;
 194
 195         per_cpu(ipi_to_irq, cpu)[ipi] = irq;
 196 }
 197
 198 static void xen_irq_info_virq_init(unsigned cpu,
 199                                    unsigned irq,
 200                                    unsigned short evtchn,
 201                                    unsigned short virq)
 202 {
 203         struct irq_info *info = info_for_irq(irq);
 204
 205         xen_irq_info_common_init(info, irq, IRQT_VIRQ, evtchn, 0);
 206
 207         info->u.virq = virq;
 208
 209         per_cpu(virq_to_irq, cpu)[virq] = irq;
 210 }
 211
 212 static void xen_irq_info_pirq_init(unsigned irq,
 213                                    unsigned short evtchn,
 214                                    unsigned short pirq,
 215                                    unsigned short gsi,
 216                                    uint16_t domid,
 217                                    unsigned char flags)
 218 {
 219         struct irq_info *info = info_for_irq(irq);
 220
 221         xen_irq_info_common_init(info, irq, IRQT_PIRQ, evtchn, 0);
 222
 223         info->u.pirq.pirq = pirq;
 224         info->u.pirq.gsi = gsi;
 225         info->u.pirq.domid = domid;
 226         info->u.pirq.flags = flags;
 227 }
 228
 229 /*
 230  * Accessors for packed IRQ information.
 231  */
 232 static unsigned int evtchn_from_irq(unsigned irq)
 233 {
 234         if (unlikely(WARN(irq < 0 || irq >= nr_irqs, "Invalid irq %d!\n", irq)))
 235                 return 0;
 236
 237         return info_for_irq(irq)->evtchn;
 238 }
 239
 240 unsigned irq_from_evtchn(unsigned int evtchn)
 241 {
 242         return evtchn_to_irq[evtchn];
 243 }
 244 EXPORT_SYMBOL_GPL(irq_from_evtchn);
 245
 246 static enum ipi_vector ipi_from_irq(unsigned irq)
 247 {
 248         struct irq_info *info = info_for_irq(irq);
 249
 250         BUG_ON(info == NULL);
 251         BUG_ON(info->type != IRQT_IPI);
 252
 253         return info->u.ipi;
 254 }
 255
 256 static unsigned virq_from_irq(unsigned irq)
 257 {
 258         struct irq_info *info = info_for_irq(irq);
 259
 260         BUG_ON(info == NULL);
 261         BUG_ON(info->type != IRQT_VIRQ);
 262
 263         return info->u.virq;
 264 }
 265
 266 static unsigned pirq_from_irq(unsigned irq)
 267 {
 268         struct irq_info *info = info_for_irq(irq);
 269
 270         BUG_ON(info == NULL);
 271         BUG_ON(info->type != IRQT_PIRQ);
 272
 273         return info->u.pirq.pirq;
 274 }
 275
 276 static enum xen_irq_type type_from_irq(unsigned irq)
 277 {
 278         return info_for_irq(irq)->type;
 279 }
 280
 281 static unsigned cpu_from_irq(unsigned irq)
 282 {
 283         return info_for_irq(irq)->cpu;
 284 }
 285
 286 static unsigned int cpu_from_evtchn(unsigned int evtchn)
 287 {
 288         int irq = evtchn_to_irq[evtchn];
 289         unsigned ret = 0;
 290
 291         if (irq != -1)
 292                 ret = cpu_from_irq(irq);
 293
 294         return ret;
 295 }
 296
 297 #ifdef CONFIG_X86
 298 static bool pirq_check_eoi_map(unsigned irq)
 299 {
 300         return test_bit(pirq_from_irq(irq), pirq_eoi_map);
 301 }
 302 #endif
 303
 304 static bool pirq_needs_eoi_flag(unsigned irq)
 305 {
 306         struct irq_info *info = info_for_irq(irq);
 307         BUG_ON(info->type != IRQT_PIRQ);
 308
 309         return info->u.pirq.flags & PIRQ_NEEDS_EOI;
 310 }
 311
 312 static inline xen_ulong_t active_evtchns(unsigned int cpu,
 313                                          struct shared_info *sh,
 314                                          unsigned int idx)
 315 {
 316         return sh->evtchn_pending[idx] &
 317                 per_cpu(cpu_evtchn_mask, cpu)[idx] &
 318                 ~sh->evtchn_mask[idx];
 319 }
 320
 321 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
 322 {
 323         int irq = evtchn_to_irq[chn];
 324
 325         BUG_ON(irq == -1);
 326 #ifdef CONFIG_SMP
 327         cpumask_copy(irq_to_desc(irq)->irq_data.affinity, cpumask_of(cpu));
 328 #endif
 329
 330         clear_bit(chn, BM(per_cpu(cpu_evtchn_mask, cpu_from_irq(irq))));
 331         set_bit(chn, BM(per_cpu(cpu_evtchn_mask, cpu)));
 332
 333         info_for_irq(irq)->cpu = cpu;
 334 }
 335
 336 static void init_evtchn_cpu_bindings(void)
 337 {
 338         int i;
 339 #ifdef CONFIG_SMP
 340         struct irq_info *info;
 341
 342         /* By default all event channels notify CPU#0. */
 343         list_for_each_entry(info, &xen_irq_list_head, list) {
 344                 struct irq_desc *desc = irq_to_desc(info->irq);
 345                 cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 346         }
 347 #endif
 348
 349         for_each_possible_cpu(i)
 350                 memset(per_cpu(cpu_evtchn_mask, i),
 351                        (i == 0) ? ~0 : 0, NR_EVENT_CHANNELS/8);
 352 }
 353
 354 static inline void clear_evtchn(int port)
 355 {
 356         struct shared_info *s = HYPERVISOR_shared_info;
 357         sync_clear_bit(port, BM(&s->evtchn_pending[0]));
 358 }
 359
 360 static inline void set_evtchn(int port)
 361 {
 362         struct shared_info *s = HYPERVISOR_shared_info;
 363         sync_set_bit(port, BM(&s->evtchn_pending[0]));
 364 }
 365
 366 static inline int test_evtchn(int port)
 367 {
 368         struct shared_info *s = HYPERVISOR_shared_info;
 369         return sync_test_bit(port, BM(&s->evtchn_pending[0]));
 370 }
 371
 372
 373 /**
 374  * notify_remote_via_irq - send event to remote end of event channel via irq
 375  * @irq: irq of event channel to send event to
 376  *
 377  * Unlike notify_remote_via_evtchn(), this is safe to use across
 378  * save/restore. Notifications on a broken connection are silently
 379  * dropped.
 380  */
 381 void notify_remote_via_irq(int irq)
 382 {
 383         int evtchn = evtchn_from_irq(irq);
 384
 385         if (VALID_EVTCHN(evtchn))
 386                 notify_remote_via_evtchn(evtchn);
 387 }
 388 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
 389
 390 static void mask_evtchn(int port)
 391 {
 392         struct shared_info *s = HYPERVISOR_shared_info;
 393         sync_set_bit(port, BM(&s->evtchn_mask[0]));
 394 }
 395
 396 static void unmask_evtchn(int port)
 397 {
 398         struct shared_info *s = HYPERVISOR_shared_info;
 399         unsigned int cpu = get_cpu();
 400         int do_hypercall = 0, evtchn_pending = 0;
 401
 402         BUG_ON(!irqs_disabled());
 403
 404         if (unlikely((cpu != cpu_from_evtchn(port))))
 405                 do_hypercall = 1;
 406         else {
 407                 /*
 408                  * Need to clear the mask before checking pending to
 409                  * avoid a race with an event becoming pending.
 410                  *
 411                  * EVTCHNOP_unmask will only trigger an upcall if the
 412                  * mask bit was set, so if a hypercall is needed
 413                  * remask the event.
 414                  */
 415                 sync_clear_bit(port, BM(&s->evtchn_mask[0]));
 416                 evtchn_pending = sync_test_bit(port, BM(&s->evtchn_pending[0]));
 417
 418                 if (unlikely(evtchn_pending && xen_hvm_domain())) {
 419                         sync_set_bit(port, BM(&s->evtchn_mask[0]));
 420                         do_hypercall = 1;
 421                 }
 422         }
 423
 424         /* Slow path (hypercall) if this is a non-local port or if this is
 425          * an hvm domain and an event is pending (hvm domains don't have
 426          * their own implementation of irq_enable). */
 427         if (do_hypercall) {
 428                 struct evtchn_unmask unmask = { .port = port };
 429                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 430         } else {
 431                 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
 432
 433                 /*
 434                  * The following is basically the equivalent of
 435                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
 436                  * the interrupt edge' if the channel is masked.
 437                  */
 438                 if (evtchn_pending &&
 439                     !sync_test_and_set_bit(port / BITS_PER_EVTCHN_WORD,
 440                                            BM(&vcpu_info->evtchn_pending_sel)))
 441                         vcpu_info->evtchn_upcall_pending = 1;
 442         }
 443
 444         put_cpu();
 445 }
 446
 447 static void xen_irq_init(unsigned irq)
 448 {
 449         struct irq_info *info;
 450 #ifdef CONFIG_SMP
 451         struct irq_desc *desc = irq_to_desc(irq);
 452
 453         /* By default all event channels notify CPU#0. */
 454         cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 455 #endif
 456
 457         info = kzalloc(sizeof(*info), GFP_KERNEL);
 458         if (info == NULL)
 459                 panic("Unable to allocate metadata for IRQ%d\n", irq);
 460
 461         info->type = IRQT_UNBOUND;
 462         info->refcnt = -1;
 463
 464         irq_set_handler_data(irq, info);
 465
 466         list_add_tail(&info->list, &xen_irq_list_head);
 467 }
 468
 469 static int __must_check xen_allocate_irq_dynamic(void)
 470 {
 471         int first = 0;
 472         int irq;
 473
 474 #ifdef CONFIG_X86_IO_APIC
 475         /*
 476          * For an HVM guest or domain 0 which see "real" (emulated or
 477          * actual respectively) GSIs we allocate dynamic IRQs
 478          * e.g. those corresponding to event channels or MSIs
 479          * etc. from the range above those "real" GSIs to avoid
 480          * collisions.
 481          */
 482         if (xen_initial_domain() || xen_hvm_domain())
 483                 first = get_nr_irqs_gsi();
 484 #endif
 485
 486         irq = irq_alloc_desc_from(first, -1);
 487
 488         if (irq >= 0)
 489                 xen_irq_init(irq);
 490
 491         return irq;
 492 }
 493
 494 static int __must_check xen_allocate_irq_gsi(unsigned gsi)
 495 {
 496         int irq;
 497
 498         /*
 499          * A PV guest has no concept of a GSI (since it has no ACPI
 500          * nor access to/knowledge of the physical APICs). Therefore
 501          * all IRQs are dynamically allocated from the entire IRQ
 502          * space.
 503          */
 504         if (xen_pv_domain() && !xen_initial_domain())
 505                 return xen_allocate_irq_dynamic();
 506
 507         /* Legacy IRQ descriptors are already allocated by the arch. */
 508         if (gsi < NR_IRQS_LEGACY)
 509                 irq = gsi;
 510         else
 511                 irq = irq_alloc_desc_at(gsi, -1);
 512
 513         xen_irq_init(irq);
 514
 515         return irq;
 516 }
 517
 518 static void xen_free_irq(unsigned irq)
 519 {
 520         struct irq_info *info = irq_get_handler_data(irq);
 521
 522         if (WARN_ON(!info))
 523                 return;
 524
 525         list_del(&info->list);
 526
 527         irq_set_handler_data(irq, NULL);
 528
 529         WARN_ON(info->refcnt > 0);
 530
 531         kfree(info);
 532
 533         /* Legacy IRQ descriptors are managed by the arch. */
 534         if (irq < NR_IRQS_LEGACY)
 535                 return;
 536
 537         irq_free_desc(irq);
 538 }
 539
 540 static void pirq_query_unmask(int irq)
 541 {
 542         struct physdev_irq_status_query irq_status;
 543         struct irq_info *info = info_for_irq(irq);
 544
 545         BUG_ON(info->type != IRQT_PIRQ);
 546
 547         irq_status.irq = pirq_from_irq(irq);
 548         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
 549                 irq_status.flags = 0;
 550
 551         info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
 552         if (irq_status.flags & XENIRQSTAT_needs_eoi)
 553                 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
 554 }
 555
 556 static bool probing_irq(int irq)
 557 {
 558         struct irq_desc *desc = irq_to_desc(irq);
 559
 560         return desc && desc->action == NULL;
 561 }
 562
 563 static void eoi_pirq(struct irq_data *data)
 564 {
 565         int evtchn = evtchn_from_irq(data->irq);
 566         struct physdev_eoi eoi = { .irq = pirq_from_irq(data->irq) };
 567         int rc = 0;
 568
 569         irq_move_irq(data);
 570
 571         if (VALID_EVTCHN(evtchn))
 572                 clear_evtchn(evtchn);
 573
 574         if (pirq_needs_eoi(data->irq)) {
 575                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
 576                 WARN_ON(rc);
 577         }
 578 }
 579
 580 static void mask_ack_pirq(struct irq_data *data)
 581 {
 582         disable_dynirq(data);
 583         eoi_pirq(data);
 584 }
 585
 586 static unsigned int __startup_pirq(unsigned int irq)
 587 {
 588         struct evtchn_bind_pirq bind_pirq;
 589         struct irq_info *info = info_for_irq(irq);
 590         int evtchn = evtchn_from_irq(irq);
 591         int rc;
 592
 593         BUG_ON(info->type != IRQT_PIRQ);
 594
 595         if (VALID_EVTCHN(evtchn))
 596                 goto out;
 597
 598         bind_pirq.pirq = pirq_from_irq(irq);
 599         /* NB. We are happy to share unless we are probing. */
 600         bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
 601                                         BIND_PIRQ__WILL_SHARE : 0;
 602         rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
 603         if (rc != 0) {
 604                 if (!probing_irq(irq))
 605                         pr_info("Failed to obtain physical IRQ %d\n", irq);
 606                 return 0;
 607         }
 608         evtchn = bind_pirq.port;
 609
 610         pirq_query_unmask(irq);
 611
 612         evtchn_to_irq[evtchn] = irq;
 613         bind_evtchn_to_cpu(evtchn, 0);
 614         info->evtchn = evtchn;
 615
 616 out:
 617         unmask_evtchn(evtchn);
 618         eoi_pirq(irq_get_irq_data(irq));
 619
 620         return 0;
 621 }
 622
 623 static unsigned int startup_pirq(struct irq_data *data)
 624 {
 625         return __startup_pirq(data->irq);
 626 }
 627
 628 static void shutdown_pirq(struct irq_data *data)
 629 {
 630         struct evtchn_close close;
 631         unsigned int irq = data->irq;
 632         struct irq_info *info = info_for_irq(irq);
 633         int evtchn = evtchn_from_irq(irq);
 634
 635         BUG_ON(info->type != IRQT_PIRQ);
 636
 637         if (!VALID_EVTCHN(evtchn))
 638                 return;
 639
 640         mask_evtchn(evtchn);
 641
 642         close.port = evtchn;
 643         if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 644                 BUG();
 645
 646         bind_evtchn_to_cpu(evtchn, 0);
 647         evtchn_to_irq[evtchn] = -1;
 648         info->evtchn = 0;
 649 }
 650
 651 static void enable_pirq(struct irq_data *data)
 652 {
 653         startup_pirq(data);
 654 }
 655
 656 static void disable_pirq(struct irq_data *data)
 657 {
 658         disable_dynirq(data);
 659 }
 660
 661 int xen_irq_from_gsi(unsigned gsi)
 662 {
 663         struct irq_info *info;
 664
 665         list_for_each_entry(info, &xen_irq_list_head, list) {
 666                 if (info->type != IRQT_PIRQ)
 667                         continue;
 668
 669                 if (info->u.pirq.gsi == gsi)
 670                         return info->irq;
 671         }
 672
 673         return -1;
 674 }
 675 EXPORT_SYMBOL_GPL(xen_irq_from_gsi);
 676
 677 /*
 678  * Do not make any assumptions regarding the relationship between the
 679  * IRQ number returned here and the Xen pirq argument.
 680  *
 681  * Note: We don't assign an event channel until the irq actually started
 682  * up.  Return an existing irq if we've already got one for the gsi.
 683  *
 684  * Shareable implies level triggered, not shareable implies edge
 685  * triggered here.
 686  */
 687 int xen_bind_pirq_gsi_to_irq(unsigned gsi,
 688                              unsigned pirq, int shareable, char *name)
 689 {
 690         int irq = -1;
 691         struct physdev_irq irq_op;
 692
 693         mutex_lock(&irq_mapping_update_lock);
 694
 695         irq = xen_irq_from_gsi(gsi);
 696         if (irq != -1) {
 697                 pr_info("%s: returning irq %d for gsi %u\n",
 698                         __func__, irq, gsi);
 699                 goto out;
 700         }
 701
 702         irq = xen_allocate_irq_gsi(gsi);
 703         if (irq < 0)
 704                 goto out;
 705
 706         irq_op.irq = irq;
 707         irq_op.vector = 0;
 708
 709         /* Only the privileged domain can do this. For non-priv, the pcifront
 710          * driver provides a PCI bus that does the call to do exactly
 711          * this in the priv domain. */
 712         if (xen_initial_domain() &&
 713             HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
 714                 xen_free_irq(irq);
 715                 irq = -ENOSPC;
 716                 goto out;
 717         }
 718
 719         xen_irq_info_pirq_init(irq, 0, pirq, gsi, DOMID_SELF,
 720                                shareable ? PIRQ_SHAREABLE : 0);
 721
 722         pirq_query_unmask(irq);
 723         /* We try to use the handler with the appropriate semantic for the
 724          * type of interrupt: if the interrupt is an edge triggered
 725          * interrupt we use handle_edge_irq.
 726          *
 727          * On the other hand if the interrupt is level triggered we use
 728          * handle_fasteoi_irq like the native code does for this kind of
 729          * interrupts.
 730          *
 731          * Depending on the Xen version, pirq_needs_eoi might return true
 732          * not only for level triggered interrupts but for edge triggered
 733          * interrupts too. In any case Xen always honors the eoi mechanism,
 734          * not injecting any more pirqs of the same kind if the first one
 735          * hasn't received an eoi yet. Therefore using the fasteoi handler
 736          * is the right choice either way.
 737          */
 738         if (shareable)
 739                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 740                                 handle_fasteoi_irq, name);
 741         else
 742                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 743                                 handle_edge_irq, name);
 744
 745 out:
 746         mutex_unlock(&irq_mapping_update_lock);
 747
 748         return irq;
 749 }
 750
 751 #ifdef CONFIG_PCI_MSI
 752 int xen_allocate_pirq_msi(struct pci_dev *dev, struct msi_desc *msidesc)
 753 {
 754         int rc;
 755         struct physdev_get_free_pirq op_get_free_pirq;
 756
 757         op_get_free_pirq.type = MAP_PIRQ_TYPE_MSI;
 758         rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
 759
 760         WARN_ONCE(rc == -ENOSYS,
 761                   "hypervisor does not support the PHYSDEVOP_get_free_pirq interface\n");
 762
 763         return rc ? -1 : op_get_free_pirq.pirq;
 764 }
 765
 766 int xen_bind_pirq_msi_to_irq(struct pci_dev *dev, struct msi_desc *msidesc,
 767                              int pirq, const char *name, domid_t domid)
 768 {
 769         int irq, ret;
 770
 771         mutex_lock(&irq_mapping_update_lock);
 772
 773         irq = xen_allocate_irq_dynamic();
 774         if (irq < 0)
 775                 goto out;
 776
 777         irq_set_chip_and_handler_name(irq, &xen_pirq_chip, handle_edge_irq,
 778                         name);
 779
 780         xen_irq_info_pirq_init(irq, 0, pirq, 0, domid, 0);
 781         ret = irq_set_msi_desc(irq, msidesc);
 782         if (ret < 0)
 783                 goto error_irq;
 784 out:
 785         mutex_unlock(&irq_mapping_update_lock);
 786         return irq;
 787 error_irq:
 788         mutex_unlock(&irq_mapping_update_lock);
 789         xen_free_irq(irq);
 790         return ret;
 791 }
 792 #endif
 793
 794 int xen_destroy_irq(int irq)
 795 {
 796         struct irq_desc *desc;
 797         struct physdev_unmap_pirq unmap_irq;
 798         struct irq_info *info = info_for_irq(irq);
 799         int rc = -ENOENT;
 800
 801         mutex_lock(&irq_mapping_update_lock);
 802
 803         desc = irq_to_desc(irq);
 804         if (!desc)
 805                 goto out;
 806
 807         if (xen_initial_domain()) {
 808                 unmap_irq.pirq = info->u.pirq.pirq;
 809                 unmap_irq.domid = info->u.pirq.domid;
 810                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
 811                 /* If another domain quits without making the pci_disable_msix
 812                  * call, the Xen hypervisor takes care of freeing the PIRQs
 813                  * (free_domain_pirqs).
 814                  */
 815                 if ((rc == -ESRCH && info->u.pirq.domid != DOMID_SELF))
 816                         pr_info("domain %d does not have %d anymore\n",
 817                                 info->u.pirq.domid, info->u.pirq.pirq);
 818                 else if (rc) {
 819                         pr_warn("unmap irq failed %d\n", rc);
 820                         goto out;
 821                 }
 822         }
 823
 824         xen_free_irq(irq);
 825
 826 out:
 827         mutex_unlock(&irq_mapping_update_lock);
 828         return rc;
 829 }
 830
 831 int xen_irq_from_pirq(unsigned pirq)
 832 {
 833         int irq;
 834
 835         struct irq_info *info;
 836
 837         mutex_lock(&irq_mapping_update_lock);
 838
 839         list_for_each_entry(info, &xen_irq_list_head, list) {
 840                 if (info->type != IRQT_PIRQ)
 841                         continue;
 842                 irq = info->irq;
 843                 if (info->u.pirq.pirq == pirq)
 844                         goto out;
 845         }
 846         irq = -1;
 847 out:
 848         mutex_unlock(&irq_mapping_update_lock);
 849
 850         return irq;
 851 }
 852
 853
 854 int xen_pirq_from_irq(unsigned irq)
 855 {
 856         return pirq_from_irq(irq);
 857 }
 858 EXPORT_SYMBOL_GPL(xen_pirq_from_irq);
 859 int bind_evtchn_to_irq(unsigned int evtchn)
 860 {
 861         int irq;
 862
 863         mutex_lock(&irq_mapping_update_lock);
 864
 865         irq = evtchn_to_irq[evtchn];
 866
 867         if (irq == -1) {
 868                 irq = xen_allocate_irq_dynamic();
 869                 if (irq < 0)
 870                         goto out;
 871
 872                 irq_set_chip_and_handler_name(irq, &xen_dynamic_chip,
 873                                               handle_edge_irq, "event");
 874
 875                 xen_irq_info_evtchn_init(irq, evtchn);
 876         } else {
 877                 struct irq_info *info = info_for_irq(irq);
 878                 WARN_ON(info == NULL || info->type != IRQT_EVTCHN);
 879         }
 880
 881 out:
 882         mutex_unlock(&irq_mapping_update_lock);
 883
 884         return irq;
 885 }
 886 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
 887
 888 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
 889 {
 890         struct evtchn_bind_ipi bind_ipi;
 891         int evtchn, irq;
 892
 893         mutex_lock(&irq_mapping_update_lock);
 894
 895         irq = per_cpu(ipi_to_irq, cpu)[ipi];
 896
 897         if (irq == -1) {
 898                 irq = xen_allocate_irq_dynamic();
 899                 if (irq < 0)
 900                         goto out;
 901
 902                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 903                                               handle_percpu_irq, "ipi");
 904
 905                 bind_ipi.vcpu = cpu;
 906                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
 907                                                 &bind_ipi) != 0)
 908                         BUG();
 909                 evtchn = bind_ipi.port;
 910
 911                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
 912
 913                 bind_evtchn_to_cpu(evtchn, cpu);
 914         } else {
 915                 struct irq_info *info = info_for_irq(irq);
 916                 WARN_ON(info == NULL || info->type != IRQT_IPI);
 917         }
 918
 919  out:
 920         mutex_unlock(&irq_mapping_update_lock);
 921         return irq;
 922 }
 923
 924 static int bind_interdomain_evtchn_to_irq(unsigned int remote_domain,
 925                                           unsigned int remote_port)
 926 {
 927         struct evtchn_bind_interdomain bind_interdomain;
 928         int err;
 929
 930         bind_interdomain.remote_dom  = remote_domain;
 931         bind_interdomain.remote_port = remote_port;
 932
 933         err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
 934                                           &bind_interdomain);
 935
 936         return err ? : bind_evtchn_to_irq(bind_interdomain.local_port);
 937 }
 938
 939 static int find_virq(unsigned int virq, unsigned int cpu)
 940 {
 941         struct evtchn_status status;
 942         int port, rc = -ENOENT;
 943
 944         memset(&status, 0, sizeof(status));
 945         for (port = 0; port <= NR_EVENT_CHANNELS; port++) {
 946                 status.dom = DOMID_SELF;
 947                 status.port = port;
 948                 rc = HYPERVISOR_event_channel_op(EVTCHNOP_status, &status);
 949                 if (rc < 0)
 950                         continue;
 951                 if (status.status != EVTCHNSTAT_virq)
 952                         continue;
 953                 if (status.u.virq == virq && status.vcpu == cpu) {
 954                         rc = port;
 955                         break;
 956                 }
 957         }
 958         return rc;
 959 }
 960
 961 int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
 962 {
 963         struct evtchn_bind_virq bind_virq;
 964         int evtchn, irq, ret;
 965
 966         mutex_lock(&irq_mapping_update_lock);
 967
 968         irq = per_cpu(virq_to_irq, cpu)[virq];
 969
 970         if (irq == -1) {
 971                 irq = xen_allocate_irq_dynamic();
 972                 if (irq < 0)
 973                         goto out;
 974
 975                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 976                                               handle_percpu_irq, "virq");
 977
 978                 bind_virq.virq = virq;
 979                 bind_virq.vcpu = cpu;
 980                 ret = HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 981                                                 &bind_virq);
 982                 if (ret == 0)
 983                         evtchn = bind_virq.port;
 984                 else {
 985                         if (ret == -EEXIST)
 986                                 ret = find_virq(virq, cpu);
 987                         BUG_ON(ret < 0);
 988                         evtchn = ret;
 989                 }
 990
 991                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
 992
 993                 bind_evtchn_to_cpu(evtchn, cpu);
 994         } else {
 995                 struct irq_info *info = info_for_irq(irq);
 996                 WARN_ON(info == NULL || info->type != IRQT_VIRQ);
 997         }
 998
 999 out:
1000         mutex_unlock(&irq_mapping_update_lock);
1001
1002         return irq;
1003 }
1004
1005 static void unbind_from_irq(unsigned int irq)
1006 {
1007         struct evtchn_close close;
1008         int evtchn = evtchn_from_irq(irq);
1009         struct irq_info *info = irq_get_handler_data(irq);
1010
1011         if (WARN_ON(!info))
1012                 return;
1013
1014         mutex_lock(&irq_mapping_update_lock);
1015
1016         if (info->refcnt > 0) {
1017                 info->refcnt--;
1018                 if (info->refcnt != 0)
1019                         goto done;
1020         }
1021
1022         if (VALID_EVTCHN(evtchn)) {
1023                 close.port = evtchn;
1024                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
1025                         BUG();
1026
1027                 switch (type_from_irq(irq)) {
1028                 case IRQT_VIRQ:
1029                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
1030                                 [virq_from_irq(irq)] = -1;
1031                         break;
1032                 case IRQT_IPI:
1033                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
1034                                 [ipi_from_irq(irq)] = -1;
1035                         break;
1036                 default:
1037                         break;
1038                 }
1039
1040                 /* Closed ports are implicitly re-bound to VCPU0. */
1041                 bind_evtchn_to_cpu(evtchn, 0);
1042
1043                 evtchn_to_irq[evtchn] = -1;
1044         }
1045
1046         BUG_ON(info_for_irq(irq)->type == IRQT_UNBOUND);
1047
1048         xen_free_irq(irq);
1049
1050  done:
1051         mutex_unlock(&irq_mapping_update_lock);
1052 }
1053
1054 int bind_evtchn_to_irqhandler(unsigned int evtchn,
1055                               irq_handler_t handler,
1056                               unsigned long irqflags,
1057                               const char *devname, void *dev_id)
1058 {
1059         int irq, retval;
1060
1061         irq = bind_evtchn_to_irq(evtchn);
1062         if (irq < 0)
1063                 return irq;
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 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
1073
1074 int bind_interdomain_evtchn_to_irqhandler(unsigned int remote_domain,
1075                                           unsigned int remote_port,
1076                                           irq_handler_t handler,
1077                                           unsigned long irqflags,
1078                                           const char *devname,
1079                                           void *dev_id)
1080 {
1081         int irq, retval;
1082
1083         irq = bind_interdomain_evtchn_to_irq(remote_domain, remote_port);
1084         if (irq < 0)
1085                 return irq;
1086
1087         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1088         if (retval != 0) {
1089                 unbind_from_irq(irq);
1090                 return retval;
1091         }
1092
1093         return irq;
1094 }
1095 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler);
1096
1097 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
1098                             irq_handler_t handler,
1099                             unsigned long irqflags, const char *devname, void *dev_id)
1100 {
1101         int irq, retval;
1102
1103         irq = bind_virq_to_irq(virq, cpu);
1104         if (irq < 0)
1105                 return irq;
1106         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1107         if (retval != 0) {
1108                 unbind_from_irq(irq);
1109                 return retval;
1110         }
1111
1112         return irq;
1113 }
1114 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1115
1116 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1117                            unsigned int cpu,
1118                            irq_handler_t handler,
1119                            unsigned long irqflags,
1120                            const char *devname,
1121                            void *dev_id)
1122 {
1123         int irq, retval;
1124
1125         irq = bind_ipi_to_irq(ipi, cpu);
1126         if (irq < 0)
1127                 return irq;
1128
1129         irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME | IRQF_EARLY_RESUME;
1130         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1131         if (retval != 0) {
1132                 unbind_from_irq(irq);
1133                 return retval;
1134         }
1135
1136         return irq;
1137 }
1138
1139 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1140 {
1141         struct irq_info *info = irq_get_handler_data(irq);
1142
1143         if (WARN_ON(!info))
1144                 return;
1145         free_irq(irq, dev_id);
1146         unbind_from_irq(irq);
1147 }
1148 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1149
1150 int evtchn_make_refcounted(unsigned int evtchn)
1151 {
1152         int irq = evtchn_to_irq[evtchn];
1153         struct irq_info *info;
1154
1155         if (irq == -1)
1156                 return -ENOENT;
1157
1158         info = irq_get_handler_data(irq);
1159
1160         if (!info)
1161                 return -ENOENT;
1162
1163         WARN_ON(info->refcnt != -1);
1164
1165         info->refcnt = 1;
1166
1167         return 0;
1168 }
1169 EXPORT_SYMBOL_GPL(evtchn_make_refcounted);
1170
1171 int evtchn_get(unsigned int evtchn)
1172 {
1173         int irq;
1174         struct irq_info *info;
1175         int err = -ENOENT;
1176
1177         if (evtchn >= NR_EVENT_CHANNELS)
1178                 return -EINVAL;
1179
1180         mutex_lock(&irq_mapping_update_lock);
1181
1182         irq = evtchn_to_irq[evtchn];
1183         if (irq == -1)
1184                 goto done;
1185
1186         info = irq_get_handler_data(irq);
1187
1188         if (!info)
1189                 goto done;
1190
1191         err = -EINVAL;
1192         if (info->refcnt <= 0)
1193                 goto done;
1194
1195         info->refcnt++;
1196         err = 0;
1197  done:
1198         mutex_unlock(&irq_mapping_update_lock);
1199
1200         return err;
1201 }
1202 EXPORT_SYMBOL_GPL(evtchn_get);
1203
1204 void evtchn_put(unsigned int evtchn)
1205 {
1206         int irq = evtchn_to_irq[evtchn];
1207         if (WARN_ON(irq == -1))
1208                 return;
1209         unbind_from_irq(irq);
1210 }
1211 EXPORT_SYMBOL_GPL(evtchn_put);
1212
1213 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1214 {
1215         int irq = per_cpu(ipi_to_irq, cpu)[vector];
1216         BUG_ON(irq < 0);
1217         notify_remote_via_irq(irq);
1218 }
1219
1220 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
1221 {
1222         struct shared_info *sh = HYPERVISOR_shared_info;
1223         int cpu = smp_processor_id();
1224         xen_ulong_t *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu);
1225         int i;
1226         unsigned long flags;
1227         static DEFINE_SPINLOCK(debug_lock);
1228         struct vcpu_info *v;
1229
1230         spin_lock_irqsave(&debug_lock, flags);
1231
1232         printk("\nvcpu %d\n  ", cpu);
1233
1234         for_each_online_cpu(i) {
1235                 int pending;
1236                 v = per_cpu(xen_vcpu, i);
1237                 pending = (get_irq_regs() && i == cpu)
1238                         ? xen_irqs_disabled(get_irq_regs())
1239                         : v->evtchn_upcall_mask;
1240                 printk("%d: masked=%d pending=%d event_sel %0*"PRI_xen_ulong"\n  ", i,
1241                        pending, v->evtchn_upcall_pending,
1242                        (int)(sizeof(v->evtchn_pending_sel)*2),
1243                        v->evtchn_pending_sel);
1244         }
1245         v = per_cpu(xen_vcpu, cpu);
1246
1247         printk("\npending:\n   ");
1248         for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
1249                 printk("%0*"PRI_xen_ulong"%s",
1250                        (int)sizeof(sh->evtchn_pending[0])*2,
1251                        sh->evtchn_pending[i],
1252                        i % 8 == 0 ? "\n   " : " ");
1253         printk("\nglobal mask:\n   ");
1254         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1255                 printk("%0*"PRI_xen_ulong"%s",
1256                        (int)(sizeof(sh->evtchn_mask[0])*2),
1257                        sh->evtchn_mask[i],
1258                        i % 8 == 0 ? "\n   " : " ");
1259
1260         printk("\nglobally unmasked:\n   ");
1261         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1262                 printk("%0*"PRI_xen_ulong"%s",
1263                        (int)(sizeof(sh->evtchn_mask[0])*2),
1264                        sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
1265                        i % 8 == 0 ? "\n   " : " ");
1266
1267         printk("\nlocal cpu%d mask:\n   ", cpu);
1268         for (i = (NR_EVENT_CHANNELS/BITS_PER_EVTCHN_WORD)-1; i >= 0; i--)
1269                 printk("%0*"PRI_xen_ulong"%s", (int)(sizeof(cpu_evtchn[0])*2),
1270                        cpu_evtchn[i],
1271                        i % 8 == 0 ? "\n   " : " ");
1272
1273         printk("\nlocally unmasked:\n   ");
1274         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
1275                 xen_ulong_t pending = sh->evtchn_pending[i]
1276                         & ~sh->evtchn_mask[i]
1277                         & cpu_evtchn[i];
1278                 printk("%0*"PRI_xen_ulong"%s",
1279                        (int)(sizeof(sh->evtchn_mask[0])*2),
1280                        pending, i % 8 == 0 ? "\n   " : " ");
1281         }
1282
1283         printk("\npending list:\n");
1284         for (i = 0; i < NR_EVENT_CHANNELS; i++) {
1285                 if (sync_test_bit(i, BM(sh->evtchn_pending))) {
1286                         int word_idx = i / BITS_PER_EVTCHN_WORD;
1287                         printk("  %d: event %d -> irq %d%s%s%s\n",
1288                                cpu_from_evtchn(i), i,
1289                                evtchn_to_irq[i],
1290                                sync_test_bit(word_idx, BM(&v->evtchn_pending_sel))
1291                                              ? "" : " l2-clear",
1292                                !sync_test_bit(i, BM(sh->evtchn_mask))
1293                                              ? "" : " globally-masked",
1294                                sync_test_bit(i, BM(cpu_evtchn))
1295                                              ? "" : " locally-masked");
1296                 }
1297         }
1298
1299         spin_unlock_irqrestore(&debug_lock, flags);
1300
1301         return IRQ_HANDLED;
1302 }
1303
1304 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
1305 static DEFINE_PER_CPU(unsigned int, current_word_idx);
1306 static DEFINE_PER_CPU(unsigned int, current_bit_idx);
1307
1308 /*
1309  * Mask out the i least significant bits of w
1310  */
1311 #define MASK_LSBS(w, i) (w & ((~((xen_ulong_t)0UL)) << i))
1312
1313 /*
1314  * Search the CPUs pending events bitmasks.  For each one found, map
1315  * the event number to an irq, and feed it into do_IRQ() for
1316  * handling.
1317  *
1318  * Xen uses a two-level bitmap to speed searching.  The first level is
1319  * a bitset of words which contain pending event bits.  The second
1320  * level is a bitset of pending events themselves.
1321  */
1322 static void __xen_evtchn_do_upcall(void)
1323 {
1324         int start_word_idx, start_bit_idx;
1325         int word_idx, bit_idx;
1326         int i, irq;
1327         int cpu = get_cpu();
1328         struct shared_info *s = HYPERVISOR_shared_info;
1329         struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1330         unsigned count;
1331
1332         do {
1333                 xen_ulong_t pending_words;
1334                 xen_ulong_t pending_bits;
1335                 struct irq_desc *desc;
1336
1337                 vcpu_info->evtchn_upcall_pending = 0;
1338
1339                 if (__this_cpu_inc_return(xed_nesting_count) - 1)
1340                         goto out;
1341
1342                 /*
1343                  * Master flag must be cleared /before/ clearing
1344                  * selector flag. xchg_xen_ulong must contain an
1345                  * appropriate barrier.
1346                  */
1347                 if ((irq = per_cpu(virq_to_irq, cpu)[VIRQ_TIMER]) != -1) {
1348                         int evtchn = evtchn_from_irq(irq);
1349                         word_idx = evtchn / BITS_PER_LONG;
1350                         pending_bits = evtchn % BITS_PER_LONG;
1351                         if (active_evtchns(cpu, s, word_idx) & (1ULL << pending_bits)) {
1352                                 desc = irq_to_desc(irq);
1353                                 if (desc)
1354                                         generic_handle_irq_desc(irq, desc);
1355                         }
1356                 }
1357
1358                 pending_words = xchg_xen_ulong(&vcpu_info->evtchn_pending_sel, 0);
1359
1360                 start_word_idx = __this_cpu_read(current_word_idx);
1361                 start_bit_idx = __this_cpu_read(current_bit_idx);
1362
1363                 word_idx = start_word_idx;
1364
1365                 for (i = 0; pending_words != 0; i++) {
1366                         xen_ulong_t words;
1367
1368                         words = MASK_LSBS(pending_words, word_idx);
1369
1370                         /*
1371                          * If we masked out all events, wrap to beginning.
1372                          */
1373                         if (words == 0) {
1374                                 word_idx = 0;
1375                                 bit_idx = 0;
1376                                 continue;
1377                         }
1378                         word_idx = EVTCHN_FIRST_BIT(words);
1379
1380                         pending_bits = active_evtchns(cpu, s, word_idx);
1381                         bit_idx = 0; /* usually scan entire word from start */
1382                         if (word_idx == start_word_idx) {
1383                                 /* We scan the starting word in two parts */
1384                                 if (i == 0)
1385                                         /* 1st time: start in the middle */
1386                                         bit_idx = start_bit_idx;
1387                                 else
1388                                         /* 2nd time: mask bits done already */
1389                                         bit_idx &= (1UL << start_bit_idx) - 1;
1390                         }
1391
1392                         do {
1393                                 xen_ulong_t bits;
1394                                 int port;
1395
1396                                 bits = MASK_LSBS(pending_bits, bit_idx);
1397
1398                                 /* If we masked out all events, move on. */
1399                                 if (bits == 0)
1400                                         break;
1401
1402                                 bit_idx = EVTCHN_FIRST_BIT(bits);
1403
1404                                 /* Process port. */
1405                                 port = (word_idx * BITS_PER_EVTCHN_WORD) + bit_idx;
1406                                 irq = evtchn_to_irq[port];
1407
1408                                 if (irq != -1) {
1409                                         desc = irq_to_desc(irq);
1410                                         if (desc)
1411                                                 generic_handle_irq_desc(irq, desc);
1412                                 }
1413
1414                                 bit_idx = (bit_idx + 1) % BITS_PER_EVTCHN_WORD;
1415
1416                                 /* Next caller starts at last processed + 1 */
1417                                 __this_cpu_write(current_word_idx,
1418                                                  bit_idx ? word_idx :
1419                                                  (word_idx+1) % BITS_PER_EVTCHN_WORD);
1420                                 __this_cpu_write(current_bit_idx, bit_idx);
1421                         } while (bit_idx != 0);
1422
1423                         /* Scan start_l1i twice; all others once. */
1424                         if ((word_idx != start_word_idx) || (i != 0))
1425                                 pending_words &= ~(1UL << word_idx);
1426
1427                         word_idx = (word_idx + 1) % BITS_PER_EVTCHN_WORD;
1428                 }
1429
1430                 BUG_ON(!irqs_disabled());
1431
1432                 count = __this_cpu_read(xed_nesting_count);
1433                 __this_cpu_write(xed_nesting_count, 0);
1434         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
1435
1436 out:
1437
1438         put_cpu();
1439 }
1440
1441 void xen_evtchn_do_upcall(struct pt_regs *regs)
1442 {
1443         struct pt_regs *old_regs = set_irq_regs(regs);
1444
1445         irq_enter();
1446 #ifdef CONFIG_X86
1447         exit_idle();
1448 #endif
1449
1450         __xen_evtchn_do_upcall();
1451
1452         irq_exit();
1453         set_irq_regs(old_regs);
1454 }
1455
1456 void xen_hvm_evtchn_do_upcall(void)
1457 {
1458         __xen_evtchn_do_upcall();
1459 }
1460 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1461
1462 /* Rebind a new event channel to an existing irq. */
1463 void rebind_evtchn_irq(int evtchn, int irq)
1464 {
1465         struct irq_info *info = info_for_irq(irq);
1466
1467         if (WARN_ON(!info))
1468                 return;
1469
1470         /* Make sure the irq is masked, since the new event channel
1471            will also be masked. */
1472         disable_irq(irq);
1473
1474         mutex_lock(&irq_mapping_update_lock);
1475
1476         /* After resume the irq<->evtchn mappings are all cleared out */
1477         BUG_ON(evtchn_to_irq[evtchn] != -1);
1478         /* Expect irq to have been bound before,
1479            so there should be a proper type */
1480         BUG_ON(info->type == IRQT_UNBOUND);
1481
1482         xen_irq_info_evtchn_init(irq, evtchn);
1483
1484         mutex_unlock(&irq_mapping_update_lock);
1485
1486         /* new event channels are always bound to cpu 0 */
1487         irq_set_affinity(irq, cpumask_of(0));
1488
1489         /* Unmask the event channel. */
1490         enable_irq(irq);
1491 }
1492
1493 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1494 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1495 {
1496         struct evtchn_bind_vcpu bind_vcpu;
1497         int evtchn = evtchn_from_irq(irq);
1498
1499         if (!VALID_EVTCHN(evtchn))
1500                 return -1;
1501
1502         /*
1503          * Events delivered via platform PCI interrupts are always
1504          * routed to vcpu 0 and hence cannot be rebound.
1505          */
1506         if (xen_hvm_domain() && !xen_have_vector_callback)
1507                 return -1;
1508
1509         /* Send future instances of this interrupt to other vcpu. */
1510         bind_vcpu.port = evtchn;
1511         bind_vcpu.vcpu = tcpu;
1512
1513         /*
1514          * If this fails, it usually just indicates that we're dealing with a
1515          * virq or IPI channel, which don't actually need to be rebound. Ignore
1516          * it, but don't do the xenlinux-level rebind in that case.
1517          */
1518         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1519                 bind_evtchn_to_cpu(evtchn, tcpu);
1520
1521         return 0;
1522 }
1523
1524 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1525                             bool force)
1526 {
1527         unsigned tcpu = cpumask_first(dest);
1528
1529         return rebind_irq_to_cpu(data->irq, tcpu);
1530 }
1531
1532 int resend_irq_on_evtchn(unsigned int irq)
1533 {
1534         int masked, evtchn = evtchn_from_irq(irq);
1535         struct shared_info *s = HYPERVISOR_shared_info;
1536
1537         if (!VALID_EVTCHN(evtchn))
1538                 return 1;
1539
1540         masked = sync_test_and_set_bit(evtchn, BM(s->evtchn_mask));
1541         sync_set_bit(evtchn, BM(s->evtchn_pending));
1542         if (!masked)
1543                 unmask_evtchn(evtchn);
1544
1545         return 1;
1546 }
1547
1548 static void enable_dynirq(struct irq_data *data)
1549 {
1550         int evtchn = evtchn_from_irq(data->irq);
1551
1552         if (VALID_EVTCHN(evtchn))
1553                 unmask_evtchn(evtchn);
1554 }
1555
1556 static void disable_dynirq(struct irq_data *data)
1557 {
1558         int evtchn = evtchn_from_irq(data->irq);
1559
1560         if (VALID_EVTCHN(evtchn))
1561                 mask_evtchn(evtchn);
1562 }
1563
1564 static void ack_dynirq(struct irq_data *data)
1565 {
1566         int evtchn = evtchn_from_irq(data->irq);
1567
1568         irq_move_irq(data);
1569
1570         if (VALID_EVTCHN(evtchn))
1571                 clear_evtchn(evtchn);
1572 }
1573
1574 static void mask_ack_dynirq(struct irq_data *data)
1575 {
1576         disable_dynirq(data);
1577         ack_dynirq(data);
1578 }
1579
1580 static int retrigger_dynirq(struct irq_data *data)
1581 {
1582         int evtchn = evtchn_from_irq(data->irq);
1583         struct shared_info *sh = HYPERVISOR_shared_info;
1584         int ret = 0;
1585
1586         if (VALID_EVTCHN(evtchn)) {
1587                 int masked;
1588
1589                 masked = sync_test_and_set_bit(evtchn, BM(sh->evtchn_mask));
1590                 sync_set_bit(evtchn, BM(sh->evtchn_pending));
1591                 if (!masked)
1592                         unmask_evtchn(evtchn);
1593                 ret = 1;
1594         }
1595
1596         return ret;
1597 }
1598
1599 static void restore_pirqs(void)
1600 {
1601         int pirq, rc, irq, gsi;
1602         struct physdev_map_pirq map_irq;
1603         struct irq_info *info;
1604
1605         list_for_each_entry(info, &xen_irq_list_head, list) {
1606                 if (info->type != IRQT_PIRQ)
1607                         continue;
1608
1609                 pirq = info->u.pirq.pirq;
1610                 gsi = info->u.pirq.gsi;
1611                 irq = info->irq;
1612
1613                 /* save/restore of PT devices doesn't work, so at this point the
1614                  * only devices present are GSI based emulated devices */
1615                 if (!gsi)
1616                         continue;
1617
1618                 map_irq.domid = DOMID_SELF;
1619                 map_irq.type = MAP_PIRQ_TYPE_GSI;
1620                 map_irq.index = gsi;
1621                 map_irq.pirq = pirq;
1622
1623                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1624                 if (rc) {
1625                         pr_warn("xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1626                                 gsi, irq, pirq, rc);
1627                         xen_free_irq(irq);
1628                         continue;
1629                 }
1630
1631                 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1632
1633                 __startup_pirq(irq);
1634         }
1635 }
1636
1637 static void restore_cpu_virqs(unsigned int cpu)
1638 {
1639         struct evtchn_bind_virq bind_virq;
1640         int virq, irq, evtchn;
1641
1642         for (virq = 0; virq < NR_VIRQS; virq++) {
1643                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1644                         continue;
1645
1646                 BUG_ON(virq_from_irq(irq) != virq);
1647
1648                 /* Get a new binding from Xen. */
1649                 bind_virq.virq = virq;
1650                 bind_virq.vcpu = cpu;
1651                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1652                                                 &bind_virq) != 0)
1653                         BUG();
1654                 evtchn = bind_virq.port;
1655
1656                 /* Record the new mapping. */
1657                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
1658                 bind_evtchn_to_cpu(evtchn, cpu);
1659         }
1660 }
1661
1662 static void restore_cpu_ipis(unsigned int cpu)
1663 {
1664         struct evtchn_bind_ipi bind_ipi;
1665         int ipi, irq, evtchn;
1666
1667         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1668                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1669                         continue;
1670
1671                 BUG_ON(ipi_from_irq(irq) != ipi);
1672
1673                 /* Get a new binding from Xen. */
1674                 bind_ipi.vcpu = cpu;
1675                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1676                                                 &bind_ipi) != 0)
1677                         BUG();
1678                 evtchn = bind_ipi.port;
1679
1680                 /* Record the new mapping. */
1681                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
1682                 bind_evtchn_to_cpu(evtchn, cpu);
1683         }
1684 }
1685
1686 /* Clear an irq's pending state, in preparation for polling on it */
1687 void xen_clear_irq_pending(int irq)
1688 {
1689         int evtchn = evtchn_from_irq(irq);
1690
1691         if (VALID_EVTCHN(evtchn))
1692                 clear_evtchn(evtchn);
1693 }
1694 EXPORT_SYMBOL(xen_clear_irq_pending);
1695 void xen_set_irq_pending(int irq)
1696 {
1697         int evtchn = evtchn_from_irq(irq);
1698
1699         if (VALID_EVTCHN(evtchn))
1700                 set_evtchn(evtchn);
1701 }
1702
1703 bool xen_test_irq_pending(int irq)
1704 {
1705         int evtchn = evtchn_from_irq(irq);
1706         bool ret = false;
1707
1708         if (VALID_EVTCHN(evtchn))
1709                 ret = test_evtchn(evtchn);
1710
1711         return ret;
1712 }
1713
1714 /* Poll waiting for an irq to become pending with timeout.  In the usual case,
1715  * the irq will be disabled so it won't deliver an interrupt. */
1716 void xen_poll_irq_timeout(int irq, u64 timeout)
1717 {
1718         evtchn_port_t evtchn = evtchn_from_irq(irq);
1719
1720         if (VALID_EVTCHN(evtchn)) {
1721                 struct sched_poll poll;
1722
1723                 poll.nr_ports = 1;
1724                 poll.timeout = timeout;
1725                 set_xen_guest_handle(poll.ports, &evtchn);
1726
1727                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1728                         BUG();
1729         }
1730 }
1731 EXPORT_SYMBOL(xen_poll_irq_timeout);
1732 /* Poll waiting for an irq to become pending.  In the usual case, the
1733  * irq will be disabled so it won't deliver an interrupt. */
1734 void xen_poll_irq(int irq)
1735 {
1736         xen_poll_irq_timeout(irq, 0 /* no timeout */);
1737 }
1738
1739 /* Check whether the IRQ line is shared with other guests. */
1740 int xen_test_irq_shared(int irq)
1741 {
1742         struct irq_info *info = info_for_irq(irq);
1743         struct physdev_irq_status_query irq_status;
1744
1745         if (WARN_ON(!info))
1746                 return -ENOENT;
1747
1748         irq_status.irq = info->u.pirq.pirq;
1749
1750         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
1751                 return 0;
1752         return !(irq_status.flags & XENIRQSTAT_shared);
1753 }
1754 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
1755
1756 void xen_irq_resume(void)
1757 {
1758         unsigned int cpu, evtchn;
1759         struct irq_info *info;
1760
1761         init_evtchn_cpu_bindings();
1762
1763         /* New event-channel space is not 'live' yet. */
1764         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1765                 mask_evtchn(evtchn);
1766
1767         /* No IRQ <-> event-channel mappings. */
1768         list_for_each_entry(info, &xen_irq_list_head, list)
1769                 info->evtchn = 0; /* zap event-channel binding */
1770
1771         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1772                 evtchn_to_irq[evtchn] = -1;
1773
1774         for_each_possible_cpu(cpu) {
1775                 restore_cpu_virqs(cpu);
1776                 restore_cpu_ipis(cpu);
1777         }
1778
1779         restore_pirqs();
1780 }
1781
1782 static struct irq_chip xen_dynamic_chip __read_mostly = {
1783         .name                   = "xen-dyn",
1784
1785         .irq_disable            = disable_dynirq,
1786         .irq_mask               = disable_dynirq,
1787         .irq_unmask             = enable_dynirq,
1788
1789         .irq_ack                = ack_dynirq,
1790         .irq_mask_ack           = mask_ack_dynirq,
1791
1792         .irq_set_affinity       = set_affinity_irq,
1793         .irq_retrigger          = retrigger_dynirq,
1794 };
1795
1796 static struct irq_chip xen_pirq_chip __read_mostly = {
1797         .name                   = "xen-pirq",
1798
1799         .irq_startup            = startup_pirq,
1800         .irq_shutdown           = shutdown_pirq,
1801         .irq_enable             = enable_pirq,
1802         .irq_disable            = disable_pirq,
1803
1804         .irq_mask               = disable_dynirq,
1805         .irq_unmask             = enable_dynirq,
1806
1807         .irq_ack                = eoi_pirq,
1808         .irq_eoi                = eoi_pirq,
1809         .irq_mask_ack           = mask_ack_pirq,
1810
1811         .irq_set_affinity       = set_affinity_irq,
1812
1813         .irq_retrigger          = retrigger_dynirq,
1814 };
1815
1816 static struct irq_chip xen_percpu_chip __read_mostly = {
1817         .name                   = "xen-percpu",
1818
1819         .irq_disable            = disable_dynirq,
1820         .irq_mask               = disable_dynirq,
1821         .irq_unmask             = enable_dynirq,
1822
1823         .irq_ack                = ack_dynirq,
1824 };
1825
1826 int xen_set_callback_via(uint64_t via)
1827 {
1828         struct xen_hvm_param a;
1829         a.domid = DOMID_SELF;
1830         a.index = HVM_PARAM_CALLBACK_IRQ;
1831         a.value = via;
1832         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1833 }
1834 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1835
1836 #ifdef CONFIG_XEN_PVHVM
1837 /* Vector callbacks are better than PCI interrupts to receive event
1838  * channel notifications because we can receive vector callbacks on any
1839  * vcpu and we don't need PCI support or APIC interactions. */
1840 void xen_callback_vector(void)
1841 {
1842         int rc;
1843         uint64_t callback_via;
1844         if (xen_have_vector_callback) {
1845                 callback_via = HVM_CALLBACK_VECTOR(HYPERVISOR_CALLBACK_VECTOR);
1846                 rc = xen_set_callback_via(callback_via);
1847                 if (rc) {
1848                         pr_err("Request for Xen HVM callback vector failed\n");
1849                         xen_have_vector_callback = 0;
1850                         return;
1851                 }
1852                 pr_info("Xen HVM callback vector for event delivery is enabled\n");
1853                 /* in the restore case the vector has already been allocated */
1854                 if (!test_bit(HYPERVISOR_CALLBACK_VECTOR, used_vectors))
1855                         alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR,
1856                                         xen_hvm_callback_vector);
1857         }
1858 }
1859 #else
1860 void xen_callback_vector(void) {}
1861 #endif
1862
1863 void __init xen_init_IRQ(void)
1864 {
1865         int i;
1866
1867         evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1868                                     GFP_KERNEL);
1869         BUG_ON(!evtchn_to_irq);
1870         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1871                 evtchn_to_irq[i] = -1;
1872
1873         init_evtchn_cpu_bindings();
1874
1875         /* No event channels are 'live' right now. */
1876         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1877                 mask_evtchn(i);
1878
1879         pirq_needs_eoi = pirq_needs_eoi_flag;
1880
1881 #ifdef CONFIG_X86
1882         if (xen_hvm_domain()) {
1883                 xen_callback_vector();
1884                 native_init_IRQ();
1885                 /* pci_xen_hvm_init must be called after native_init_IRQ so that
1886                  * __acpi_register_gsi can point at the right function */
1887                 pci_xen_hvm_init();
1888         } else {
1889                 int rc;
1890                 struct physdev_pirq_eoi_gmfn eoi_gmfn;
1891
1892                 irq_ctx_init(smp_processor_id());
1893                 if (xen_initial_domain())
1894                         pci_xen_initial_domain();
1895
1896                 pirq_eoi_map = (void *)__get_free_page(GFP_KERNEL|__GFP_ZERO);
1897                 eoi_gmfn.gmfn = virt_to_mfn(pirq_eoi_map);
1898                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_pirq_eoi_gmfn_v2, &eoi_gmfn);
1899                 if (rc != 0) {
1900                         free_page((unsigned long) pirq_eoi_map);
1901                         pirq_eoi_map = NULL;
1902                 } else
1903                         pirq_needs_eoi = pirq_check_eoi_map;
1904         }
1905 #endif
1906 }