drivers/xen/events.c

   1 /*
   2  * Xen event channels
   3  *
   4  * Xen models interrupts with abstract event channels.  Because each
   5  * domain gets 1024 event channels, but NR_IRQ is not that large, we
   6  * must dynamically map irqs<->event channels.  The event channels
   7  * interface with the rest of the kernel by defining a xen interrupt
   8  * chip.  When an event is recieved, it is mapped to an irq and sent
   9  * through the normal interrupt processing path.
  10  *
  11  * There are four kinds of events which can be mapped to an event
  12  * channel:
  13  *
  14  * 1. Inter-domain notifications.  This includes all the virtual
  15  *    device events, since they're driven by front-ends in another domain
  16  *    (typically dom0).
  17  * 2. VIRQs, typically used for timers.  These are per-cpu events.
  18  * 3. IPIs.
  19  * 4. PIRQs - Hardware interrupts.
  20  *
  21  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
  22  */
  23
  24 #include <linux/linkage.h>
  25 #include <linux/interrupt.h>
  26 #include <linux/irq.h>
  27 #include <linux/module.h>
  28 #include <linux/string.h>
  29 #include <linux/bootmem.h>
  30 #include <linux/slab.h>
  31 #include <linux/irqnr.h>
  32 #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_SPINLOCK(irq_mapping_update_lock);
  58
  59 /* IRQ <-> VIRQ mapping. */
  60 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
  61
  62 /* IRQ <-> IPI mapping */
  63 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
  64
  65 /* Interrupt types. */
  66 enum xen_irq_type {
  67         IRQT_UNBOUND = 0,
  68         IRQT_PIRQ,
  69         IRQT_VIRQ,
  70         IRQT_IPI,
  71         IRQT_EVTCHN
  72 };
  73
  74 /*
  75  * Packed IRQ information:
  76  * type - enum xen_irq_type
  77  * event channel - irq->event channel mapping
  78  * cpu - cpu this event channel is bound to
  79  * index - type-specific information:
  80  *    PIRQ - vector, with MSB being "needs EIO", or physical IRQ of the HVM
  81  *           guest, or GSI (real passthrough IRQ) of the device.
  82  *    VIRQ - virq number
  83  *    IPI - IPI vector
  84  *    EVTCHN -
  85  */
  86 struct irq_info
  87 {
  88         enum xen_irq_type type; /* type */
  89         unsigned short evtchn;  /* event channel */
  90         unsigned short cpu;     /* cpu bound */
  91
  92         union {
  93                 unsigned short virq;
  94                 enum ipi_vector ipi;
  95                 struct {
  96                         unsigned short pirq;
  97                         unsigned short gsi;
  98                         unsigned char vector;
  99                         unsigned char flags;
 100                 } pirq;
 101         } u;
 102 };
 103 #define PIRQ_NEEDS_EOI  (1 << 0)
 104 #define PIRQ_SHAREABLE  (1 << 1)
 105
 106 static struct irq_info *irq_info;
 107 static int *pirq_to_irq;
 108
 109 static int *evtchn_to_irq;
 110 struct cpu_evtchn_s {
 111         unsigned long bits[NR_EVENT_CHANNELS/BITS_PER_LONG];
 112 };
 113
 114 static __initdata struct cpu_evtchn_s init_evtchn_mask = {
 115         .bits[0 ... (NR_EVENT_CHANNELS/BITS_PER_LONG)-1] = ~0ul,
 116 };
 117 static struct cpu_evtchn_s *cpu_evtchn_mask_p = &init_evtchn_mask;
 118
 119 static inline unsigned long *cpu_evtchn_mask(int cpu)
 120 {
 121         return cpu_evtchn_mask_p[cpu].bits;
 122 }
 123
 124 /* Xen will never allocate port zero for any purpose. */
 125 #define VALID_EVTCHN(chn)       ((chn) != 0)
 126
 127 static struct irq_chip xen_dynamic_chip;
 128 static struct irq_chip xen_percpu_chip;
 129 static struct irq_chip xen_pirq_chip;
 130
 131 /* Constructor for packed IRQ information. */
 132 static struct irq_info mk_unbound_info(void)
 133 {
 134         return (struct irq_info) { .type = IRQT_UNBOUND };
 135 }
 136
 137 static struct irq_info mk_evtchn_info(unsigned short evtchn)
 138 {
 139         return (struct irq_info) { .type = IRQT_EVTCHN, .evtchn = evtchn,
 140                         .cpu = 0 };
 141 }
 142
 143 static struct irq_info mk_ipi_info(unsigned short evtchn, enum ipi_vector ipi)
 144 {
 145         return (struct irq_info) { .type = IRQT_IPI, .evtchn = evtchn,
 146                         .cpu = 0, .u.ipi = ipi };
 147 }
 148
 149 static struct irq_info mk_virq_info(unsigned short evtchn, unsigned short virq)
 150 {
 151         return (struct irq_info) { .type = IRQT_VIRQ, .evtchn = evtchn,
 152                         .cpu = 0, .u.virq = virq };
 153 }
 154
 155 static struct irq_info mk_pirq_info(unsigned short evtchn, unsigned short pirq,
 156                                     unsigned short gsi, unsigned short vector)
 157 {
 158         return (struct irq_info) { .type = IRQT_PIRQ, .evtchn = evtchn,
 159                         .cpu = 0,
 160                         .u.pirq = { .pirq = pirq, .gsi = gsi, .vector = vector } };
 161 }
 162
 163 /*
 164  * Accessors for packed IRQ information.
 165  */
 166 static struct irq_info *info_for_irq(unsigned irq)
 167 {
 168         return &irq_info[irq];
 169 }
 170
 171 static unsigned int evtchn_from_irq(unsigned irq)
 172 {
 173         if (unlikely(WARN(irq < 0 || irq >= nr_irqs, "Invalid irq %d!\n", irq)))
 174                 return 0;
 175
 176         return info_for_irq(irq)->evtchn;
 177 }
 178
 179 unsigned irq_from_evtchn(unsigned int evtchn)
 180 {
 181         return evtchn_to_irq[evtchn];
 182 }
 183 EXPORT_SYMBOL_GPL(irq_from_evtchn);
 184
 185 static enum ipi_vector ipi_from_irq(unsigned irq)
 186 {
 187         struct irq_info *info = info_for_irq(irq);
 188
 189         BUG_ON(info == NULL);
 190         BUG_ON(info->type != IRQT_IPI);
 191
 192         return info->u.ipi;
 193 }
 194
 195 static unsigned virq_from_irq(unsigned irq)
 196 {
 197         struct irq_info *info = info_for_irq(irq);
 198
 199         BUG_ON(info == NULL);
 200         BUG_ON(info->type != IRQT_VIRQ);
 201
 202         return info->u.virq;
 203 }
 204
 205 static unsigned pirq_from_irq(unsigned irq)
 206 {
 207         struct irq_info *info = info_for_irq(irq);
 208
 209         BUG_ON(info == NULL);
 210         BUG_ON(info->type != IRQT_PIRQ);
 211
 212         return info->u.pirq.pirq;
 213 }
 214
 215 static unsigned gsi_from_irq(unsigned irq)
 216 {
 217         struct irq_info *info = info_for_irq(irq);
 218
 219         BUG_ON(info == NULL);
 220         BUG_ON(info->type != IRQT_PIRQ);
 221
 222         return info->u.pirq.gsi;
 223 }
 224
 225 static unsigned vector_from_irq(unsigned irq)
 226 {
 227         struct irq_info *info = info_for_irq(irq);
 228
 229         BUG_ON(info == NULL);
 230         BUG_ON(info->type != IRQT_PIRQ);
 231
 232         return info->u.pirq.vector;
 233 }
 234
 235 static enum xen_irq_type type_from_irq(unsigned irq)
 236 {
 237         return info_for_irq(irq)->type;
 238 }
 239
 240 static unsigned cpu_from_irq(unsigned irq)
 241 {
 242         return info_for_irq(irq)->cpu;
 243 }
 244
 245 static unsigned int cpu_from_evtchn(unsigned int evtchn)
 246 {
 247         int irq = evtchn_to_irq[evtchn];
 248         unsigned ret = 0;
 249
 250         if (irq != -1)
 251                 ret = cpu_from_irq(irq);
 252
 253         return ret;
 254 }
 255
 256 static bool pirq_needs_eoi(unsigned irq)
 257 {
 258         struct irq_info *info = info_for_irq(irq);
 259
 260         BUG_ON(info->type != IRQT_PIRQ);
 261
 262         return info->u.pirq.flags & PIRQ_NEEDS_EOI;
 263 }
 264
 265 static inline unsigned long active_evtchns(unsigned int cpu,
 266                                            struct shared_info *sh,
 267                                            unsigned int idx)
 268 {
 269         return (sh->evtchn_pending[idx] &
 270                 cpu_evtchn_mask(cpu)[idx] &
 271                 ~sh->evtchn_mask[idx]);
 272 }
 273
 274 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
 275 {
 276         int irq = evtchn_to_irq[chn];
 277
 278         BUG_ON(irq == -1);
 279 #ifdef CONFIG_SMP
 280         cpumask_copy(irq_to_desc(irq)->affinity, cpumask_of(cpu));
 281 #endif
 282
 283         clear_bit(chn, cpu_evtchn_mask(cpu_from_irq(irq)));
 284         set_bit(chn, cpu_evtchn_mask(cpu));
 285
 286         irq_info[irq].cpu = cpu;
 287 }
 288
 289 static void init_evtchn_cpu_bindings(void)
 290 {
 291         int i;
 292 #ifdef CONFIG_SMP
 293         struct irq_desc *desc;
 294
 295         /* By default all event channels notify CPU#0. */
 296         for_each_irq_desc(i, desc) {
 297                 cpumask_copy(desc->affinity, cpumask_of(0));
 298         }
 299 #endif
 300
 301         for_each_possible_cpu(i)
 302                 memset(cpu_evtchn_mask(i),
 303                        (i == 0) ? ~0 : 0, sizeof(struct cpu_evtchn_s));
 304
 305 }
 306
 307 static inline void clear_evtchn(int port)
 308 {
 309         struct shared_info *s = HYPERVISOR_shared_info;
 310         sync_clear_bit(port, &s->evtchn_pending[0]);
 311 }
 312
 313 static inline void set_evtchn(int port)
 314 {
 315         struct shared_info *s = HYPERVISOR_shared_info;
 316         sync_set_bit(port, &s->evtchn_pending[0]);
 317 }
 318
 319 static inline int test_evtchn(int port)
 320 {
 321         struct shared_info *s = HYPERVISOR_shared_info;
 322         return sync_test_bit(port, &s->evtchn_pending[0]);
 323 }
 324
 325
 326 /**
 327  * notify_remote_via_irq - send event to remote end of event channel via irq
 328  * @irq: irq of event channel to send event to
 329  *
 330  * Unlike notify_remote_via_evtchn(), this is safe to use across
 331  * save/restore. Notifications on a broken connection are silently
 332  * dropped.
 333  */
 334 void notify_remote_via_irq(int irq)
 335 {
 336         int evtchn = evtchn_from_irq(irq);
 337
 338         if (VALID_EVTCHN(evtchn))
 339                 notify_remote_via_evtchn(evtchn);
 340 }
 341 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
 342
 343 static void mask_evtchn(int port)
 344 {
 345         struct shared_info *s = HYPERVISOR_shared_info;
 346         sync_set_bit(port, &s->evtchn_mask[0]);
 347 }
 348
 349 static void unmask_evtchn(int port)
 350 {
 351         struct shared_info *s = HYPERVISOR_shared_info;
 352         unsigned int cpu = get_cpu();
 353
 354         BUG_ON(!irqs_disabled());
 355
 356         /* Slow path (hypercall) if this is a non-local port. */
 357         if (unlikely(cpu != cpu_from_evtchn(port))) {
 358                 struct evtchn_unmask unmask = { .port = port };
 359                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 360         } else {
 361                 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
 362
 363                 sync_clear_bit(port, &s->evtchn_mask[0]);
 364
 365                 /*
 366                  * The following is basically the equivalent of
 367                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
 368                  * the interrupt edge' if the channel is masked.
 369                  */
 370                 if (sync_test_bit(port, &s->evtchn_pending[0]) &&
 371                     !sync_test_and_set_bit(port / BITS_PER_LONG,
 372                                            &vcpu_info->evtchn_pending_sel))
 373                         vcpu_info->evtchn_upcall_pending = 1;
 374         }
 375
 376         put_cpu();
 377 }
 378
 379 static int get_nr_hw_irqs(void)
 380 {
 381         int ret = 1;
 382
 383 #ifdef CONFIG_X86_IO_APIC
 384         ret = get_nr_irqs_gsi();
 385 #endif
 386
 387         return ret;
 388 }
 389
 390 static int find_unbound_pirq(int type)
 391 {
 392         int rc, i;
 393         struct physdev_get_free_pirq op_get_free_pirq;
 394         op_get_free_pirq.type = type;
 395
 396         rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
 397         if (!rc)
 398                 return op_get_free_pirq.pirq;
 399
 400         for (i = 0; i < nr_irqs; i++) {
 401                 if (pirq_to_irq[i] < 0)
 402                         return i;
 403         }
 404         return -1;
 405 }
 406
 407 static int find_unbound_irq(void)
 408 {
 409         struct irq_data *data;
 410         int irq, res;
 411         int bottom = get_nr_hw_irqs();
 412         int top = nr_irqs-1;
 413
 414         if (bottom == nr_irqs)
 415                 goto no_irqs;
 416
 417         /* This loop starts from the top of IRQ space and goes down.
 418          * We need this b/c if we have a PCI device in a Xen PV guest
 419          * we do not have an IO-APIC (though the backend might have them)
 420          * mapped in. To not have a collision of physical IRQs with the Xen
 421          * event channels start at the top of the IRQ space for virtual IRQs.
 422          */
 423         for (irq = top; irq > bottom; irq--) {
 424                 data = irq_get_irq_data(irq);
 425                 /* only 15->0 have init'd desc; handle irq > 16 */
 426                 if (!data)
 427                         break;
 428                 if (data->chip == &no_irq_chip)
 429                         break;
 430                 if (data->chip != &xen_dynamic_chip)
 431                         continue;
 432                 if (irq_info[irq].type == IRQT_UNBOUND)
 433                         return irq;
 434         }
 435
 436         if (irq == bottom)
 437                 goto no_irqs;
 438
 439         res = irq_alloc_desc_at(irq, -1);
 440
 441         if (WARN_ON(res != irq))
 442                 return -1;
 443
 444         return irq;
 445
 446 no_irqs:
 447         panic("No available IRQ to bind to: increase nr_irqs!\n");
 448 }
 449
 450 static bool identity_mapped_irq(unsigned irq)
 451 {
 452         /* identity map all the hardware irqs */
 453         return irq < get_nr_hw_irqs();
 454 }
 455
 456 static void pirq_unmask_notify(int irq)
 457 {
 458         struct physdev_eoi eoi = { .irq = pirq_from_irq(irq) };
 459
 460         if (unlikely(pirq_needs_eoi(irq))) {
 461                 int rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
 462                 WARN_ON(rc);
 463         }
 464 }
 465
 466 static void pirq_query_unmask(int irq)
 467 {
 468         struct physdev_irq_status_query irq_status;
 469         struct irq_info *info = info_for_irq(irq);
 470
 471         BUG_ON(info->type != IRQT_PIRQ);
 472
 473         irq_status.irq = pirq_from_irq(irq);
 474         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
 475                 irq_status.flags = 0;
 476
 477         info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
 478         if (irq_status.flags & XENIRQSTAT_needs_eoi)
 479                 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
 480 }
 481
 482 static bool probing_irq(int irq)
 483 {
 484         struct irq_desc *desc = irq_to_desc(irq);
 485
 486         return desc && desc->action == NULL;
 487 }
 488
 489 static unsigned int startup_pirq(unsigned int irq)
 490 {
 491         struct evtchn_bind_pirq bind_pirq;
 492         struct irq_info *info = info_for_irq(irq);
 493         int evtchn = evtchn_from_irq(irq);
 494         int rc;
 495
 496         BUG_ON(info->type != IRQT_PIRQ);
 497
 498         if (VALID_EVTCHN(evtchn))
 499                 goto out;
 500
 501         bind_pirq.pirq = pirq_from_irq(irq);
 502         /* NB. We are happy to share unless we are probing. */
 503         bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
 504                                         BIND_PIRQ__WILL_SHARE : 0;
 505         rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
 506         if (rc != 0) {
 507                 if (!probing_irq(irq))
 508                         printk(KERN_INFO "Failed to obtain physical IRQ %d\n",
 509                                irq);
 510                 return 0;
 511         }
 512         evtchn = bind_pirq.port;
 513
 514         pirq_query_unmask(irq);
 515
 516         evtchn_to_irq[evtchn] = irq;
 517         bind_evtchn_to_cpu(evtchn, 0);
 518         info->evtchn = evtchn;
 519
 520 out:
 521         unmask_evtchn(evtchn);
 522         pirq_unmask_notify(irq);
 523
 524         return 0;
 525 }
 526
 527 static void shutdown_pirq(unsigned int irq)
 528 {
 529         struct evtchn_close close;
 530         struct irq_info *info = info_for_irq(irq);
 531         int evtchn = evtchn_from_irq(irq);
 532
 533         BUG_ON(info->type != IRQT_PIRQ);
 534
 535         if (!VALID_EVTCHN(evtchn))
 536                 return;
 537
 538         mask_evtchn(evtchn);
 539
 540         close.port = evtchn;
 541         if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 542                 BUG();
 543
 544         bind_evtchn_to_cpu(evtchn, 0);
 545         evtchn_to_irq[evtchn] = -1;
 546         info->evtchn = 0;
 547 }
 548
 549 static void enable_pirq(unsigned int irq)
 550 {
 551         startup_pirq(irq);
 552 }
 553
 554 static void disable_pirq(unsigned int irq)
 555 {
 556 }
 557
 558 static void ack_pirq(unsigned int irq)
 559 {
 560         int evtchn = evtchn_from_irq(irq);
 561
 562         move_native_irq(irq);
 563
 564         if (VALID_EVTCHN(evtchn)) {
 565                 mask_evtchn(evtchn);
 566                 clear_evtchn(evtchn);
 567         }
 568 }
 569
 570 static void end_pirq(unsigned int irq)
 571 {
 572         int evtchn = evtchn_from_irq(irq);
 573         struct irq_desc *desc = irq_to_desc(irq);
 574
 575         if (WARN_ON(!desc))
 576                 return;
 577
 578         if ((desc->status & (IRQ_DISABLED|IRQ_PENDING)) ==
 579             (IRQ_DISABLED|IRQ_PENDING)) {
 580                 shutdown_pirq(irq);
 581         } else if (VALID_EVTCHN(evtchn)) {
 582                 unmask_evtchn(evtchn);
 583                 pirq_unmask_notify(irq);
 584         }
 585 }
 586
 587 static int find_irq_by_gsi(unsigned gsi)
 588 {
 589         int irq;
 590
 591         for (irq = 0; irq < nr_irqs; irq++) {
 592                 struct irq_info *info = info_for_irq(irq);
 593
 594                 if (info == NULL || info->type != IRQT_PIRQ)
 595                         continue;
 596
 597                 if (gsi_from_irq(irq) == gsi)
 598                         return irq;
 599         }
 600
 601         return -1;
 602 }
 603
 604 int xen_allocate_pirq(unsigned gsi, int shareable, char *name)
 605 {
 606         return xen_map_pirq_gsi(gsi, gsi, shareable, name);
 607 }
 608
 609 /* xen_map_pirq_gsi might allocate irqs from the top down, as a
 610  * consequence don't assume that the irq number returned has a low value
 611  * or can be used as a pirq number unless you know otherwise.
 612  *
 613  * One notable exception is when xen_map_pirq_gsi is called passing an
 614  * hardware gsi as argument, in that case the irq number returned
 615  * matches the gsi number passed as second argument.
 616  *
 617  * Note: We don't assign an event channel until the irq actually started
 618  * up.  Return an existing irq if we've already got one for the gsi.
 619  */
 620 int xen_map_pirq_gsi(unsigned pirq, unsigned gsi, int shareable, char *name)
 621 {
 622         int irq = 0;
 623         struct physdev_irq irq_op;
 624
 625         spin_lock(&irq_mapping_update_lock);
 626
 627         if ((pirq > nr_irqs) || (gsi > nr_irqs)) {
 628                 printk(KERN_WARNING "xen_map_pirq_gsi: %s %s is incorrect!\n",
 629                         pirq > nr_irqs ? "pirq" :"",
 630                         gsi > nr_irqs ? "gsi" : "");
 631                 goto out;
 632         }
 633
 634         irq = find_irq_by_gsi(gsi);
 635         if (irq != -1) {
 636                 printk(KERN_INFO "xen_map_pirq_gsi: returning irq %d for gsi %u\n",
 637                        irq, gsi);
 638                 goto out;       /* XXX need refcount? */
 639         }
 640
 641         /* If we are a PV guest, we don't have GSIs (no ACPI passed). Therefore
 642          * we are using the !xen_initial_domain() to drop in the function.*/
 643         if (identity_mapped_irq(gsi) || (!xen_initial_domain() &&
 644                                 xen_pv_domain())) {
 645                 irq = gsi;
 646                 irq_alloc_desc_at(irq, -1);
 647         } else
 648                 irq = find_unbound_irq();
 649
 650         set_irq_chip_and_handler_name(irq, &xen_pirq_chip,
 651                                       handle_level_irq, name);
 652
 653         irq_op.irq = irq;
 654         irq_op.vector = 0;
 655
 656         /* Only the privileged domain can do this. For non-priv, the pcifront
 657          * driver provides a PCI bus that does the call to do exactly
 658          * this in the priv domain. */
 659         if (xen_initial_domain() &&
 660             HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
 661                 irq_free_desc(irq);
 662                 irq = -ENOSPC;
 663                 goto out;
 664         }
 665
 666         irq_info[irq] = mk_pirq_info(0, pirq, gsi, irq_op.vector);
 667         irq_info[irq].u.pirq.flags |= shareable ? PIRQ_SHAREABLE : 0;
 668         pirq_to_irq[pirq] = irq;
 669
 670 out:
 671         spin_unlock(&irq_mapping_update_lock);
 672
 673         return irq;
 674 }
 675
 676 #ifdef CONFIG_PCI_MSI
 677 #include <linux/msi.h>
 678 #include "../pci/msi.h"
 679
 680 void xen_allocate_pirq_msi(char *name, int *irq, int *pirq, int alloc)
 681 {
 682         spin_lock(&irq_mapping_update_lock);
 683
 684         if (alloc & XEN_ALLOC_IRQ) {
 685                 *irq = find_unbound_irq();
 686                 if (*irq == -1)
 687                         goto out;
 688         }
 689
 690         if (alloc & XEN_ALLOC_PIRQ) {
 691                 *pirq = find_unbound_pirq(MAP_PIRQ_TYPE_MSI);
 692                 if (*pirq == -1)
 693                         goto out;
 694         }
 695
 696         set_irq_chip_and_handler_name(*irq, &xen_pirq_chip,
 697                                       handle_level_irq, name);
 698
 699         irq_info[*irq] = mk_pirq_info(0, *pirq, 0, 0);
 700         pirq_to_irq[*pirq] = *irq;
 701
 702 out:
 703         spin_unlock(&irq_mapping_update_lock);
 704 }
 705
 706 int xen_create_msi_irq(struct pci_dev *dev, struct msi_desc *msidesc, int type)
 707 {
 708         int irq = -1;
 709         struct physdev_map_pirq map_irq;
 710         int rc;
 711         int pos;
 712         u32 table_offset, bir;
 713
 714         memset(&map_irq, 0, sizeof(map_irq));
 715         map_irq.domid = DOMID_SELF;
 716         map_irq.type = MAP_PIRQ_TYPE_MSI;
 717         map_irq.index = -1;
 718         map_irq.pirq = -1;
 719         map_irq.bus = dev->bus->number;
 720         map_irq.devfn = dev->devfn;
 721
 722         if (type == PCI_CAP_ID_MSIX) {
 723                 pos = pci_find_capability(dev, PCI_CAP_ID_MSIX);
 724
 725                 pci_read_config_dword(dev, msix_table_offset_reg(pos),
 726                                         &table_offset);
 727                 bir = (u8)(table_offset & PCI_MSIX_FLAGS_BIRMASK);
 728
 729                 map_irq.table_base = pci_resource_start(dev, bir);
 730                 map_irq.entry_nr = msidesc->msi_attrib.entry_nr;
 731         }
 732
 733         spin_lock(&irq_mapping_update_lock);
 734
 735         irq = find_unbound_irq();
 736
 737         if (irq == -1)
 738                 goto out;
 739
 740         rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
 741         if (rc) {
 742                 printk(KERN_WARNING "xen map irq failed %d\n", rc);
 743
 744                 irq_free_desc(irq);
 745
 746                 irq = -1;
 747                 goto out;
 748         }
 749         irq_info[irq] = mk_pirq_info(0, map_irq.pirq, 0, map_irq.index);
 750
 751         set_irq_chip_and_handler_name(irq, &xen_pirq_chip,
 752                         handle_level_irq,
 753                         (type == PCI_CAP_ID_MSIX) ? "msi-x":"msi");
 754
 755 out:
 756         spin_unlock(&irq_mapping_update_lock);
 757         return irq;
 758 }
 759 #endif
 760
 761 int xen_destroy_irq(int irq)
 762 {
 763         struct irq_desc *desc;
 764         struct physdev_unmap_pirq unmap_irq;
 765         struct irq_info *info = info_for_irq(irq);
 766         int rc = -ENOENT;
 767
 768         spin_lock(&irq_mapping_update_lock);
 769
 770         desc = irq_to_desc(irq);
 771         if (!desc)
 772                 goto out;
 773
 774         if (xen_initial_domain()) {
 775                 unmap_irq.pirq = info->u.pirq.pirq;
 776                 unmap_irq.domid = DOMID_SELF;
 777                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
 778                 if (rc) {
 779                         printk(KERN_WARNING "unmap irq failed %d\n", rc);
 780                         goto out;
 781                 }
 782                 pirq_to_irq[info->u.pirq.pirq] = -1;
 783         }
 784         irq_info[irq] = mk_unbound_info();
 785
 786         irq_free_desc(irq);
 787
 788 out:
 789         spin_unlock(&irq_mapping_update_lock);
 790         return rc;
 791 }
 792
 793 int xen_vector_from_irq(unsigned irq)
 794 {
 795         return vector_from_irq(irq);
 796 }
 797
 798 int xen_gsi_from_irq(unsigned irq)
 799 {
 800         return gsi_from_irq(irq);
 801 }
 802
 803 int xen_irq_from_pirq(unsigned pirq)
 804 {
 805         return pirq_to_irq[pirq];
 806 }
 807
 808 int bind_evtchn_to_irq(unsigned int evtchn)
 809 {
 810         int irq;
 811
 812         spin_lock(&irq_mapping_update_lock);
 813
 814         irq = evtchn_to_irq[evtchn];
 815
 816         if (irq == -1) {
 817                 irq = find_unbound_irq();
 818
 819                 set_irq_chip_and_handler_name(irq, &xen_dynamic_chip,
 820                                               handle_fasteoi_irq, "event");
 821
 822                 evtchn_to_irq[evtchn] = irq;
 823                 irq_info[irq] = mk_evtchn_info(evtchn);
 824         }
 825
 826         spin_unlock(&irq_mapping_update_lock);
 827
 828         return irq;
 829 }
 830 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
 831
 832 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
 833 {
 834         struct evtchn_bind_ipi bind_ipi;
 835         int evtchn, irq;
 836
 837         spin_lock(&irq_mapping_update_lock);
 838
 839         irq = per_cpu(ipi_to_irq, cpu)[ipi];
 840
 841         if (irq == -1) {
 842                 irq = find_unbound_irq();
 843                 if (irq < 0)
 844                         goto out;
 845
 846                 set_irq_chip_and_handler_name(irq, &xen_percpu_chip,
 847                                               handle_percpu_irq, "ipi");
 848
 849                 bind_ipi.vcpu = cpu;
 850                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
 851                                                 &bind_ipi) != 0)
 852                         BUG();
 853                 evtchn = bind_ipi.port;
 854
 855                 evtchn_to_irq[evtchn] = irq;
 856                 irq_info[irq] = mk_ipi_info(evtchn, ipi);
 857                 per_cpu(ipi_to_irq, cpu)[ipi] = irq;
 858
 859                 bind_evtchn_to_cpu(evtchn, cpu);
 860         }
 861
 862  out:
 863         spin_unlock(&irq_mapping_update_lock);
 864         return irq;
 865 }
 866
 867
 868 int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
 869 {
 870         struct evtchn_bind_virq bind_virq;
 871         int evtchn, irq;
 872
 873         spin_lock(&irq_mapping_update_lock);
 874
 875         irq = per_cpu(virq_to_irq, cpu)[virq];
 876
 877         if (irq == -1) {
 878                 irq = find_unbound_irq();
 879
 880                 set_irq_chip_and_handler_name(irq, &xen_percpu_chip,
 881                                               handle_percpu_irq, "virq");
 882
 883                 bind_virq.virq = virq;
 884                 bind_virq.vcpu = cpu;
 885                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 886                                                 &bind_virq) != 0)
 887                         BUG();
 888                 evtchn = bind_virq.port;
 889
 890                 evtchn_to_irq[evtchn] = irq;
 891                 irq_info[irq] = mk_virq_info(evtchn, virq);
 892
 893                 per_cpu(virq_to_irq, cpu)[virq] = irq;
 894
 895                 bind_evtchn_to_cpu(evtchn, cpu);
 896         }
 897
 898         spin_unlock(&irq_mapping_update_lock);
 899
 900         return irq;
 901 }
 902
 903 static void unbind_from_irq(unsigned int irq)
 904 {
 905         struct evtchn_close close;
 906         int evtchn = evtchn_from_irq(irq);
 907
 908         spin_lock(&irq_mapping_update_lock);
 909
 910         if (VALID_EVTCHN(evtchn)) {
 911                 close.port = evtchn;
 912                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 913                         BUG();
 914
 915                 switch (type_from_irq(irq)) {
 916                 case IRQT_VIRQ:
 917                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
 918                                 [virq_from_irq(irq)] = -1;
 919                         break;
 920                 case IRQT_IPI:
 921                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
 922                                 [ipi_from_irq(irq)] = -1;
 923                         break;
 924                 default:
 925                         break;
 926                 }
 927
 928                 /* Closed ports are implicitly re-bound to VCPU0. */
 929                 bind_evtchn_to_cpu(evtchn, 0);
 930
 931                 evtchn_to_irq[evtchn] = -1;
 932         }
 933
 934         if (irq_info[irq].type != IRQT_UNBOUND) {
 935                 irq_info[irq] = mk_unbound_info();
 936
 937                 irq_free_desc(irq);
 938         }
 939
 940         spin_unlock(&irq_mapping_update_lock);
 941 }
 942
 943 int bind_evtchn_to_irqhandler(unsigned int evtchn,
 944                               irq_handler_t handler,
 945                               unsigned long irqflags,
 946                               const char *devname, void *dev_id)
 947 {
 948         unsigned int irq;
 949         int retval;
 950
 951         irq = bind_evtchn_to_irq(evtchn);
 952         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 953         if (retval != 0) {
 954                 unbind_from_irq(irq);
 955                 return retval;
 956         }
 957
 958         return irq;
 959 }
 960 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
 961
 962 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
 963                             irq_handler_t handler,
 964                             unsigned long irqflags, const char *devname, void *dev_id)
 965 {
 966         unsigned int irq;
 967         int retval;
 968
 969         irq = bind_virq_to_irq(virq, cpu);
 970         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 971         if (retval != 0) {
 972                 unbind_from_irq(irq);
 973                 return retval;
 974         }
 975
 976         return irq;
 977 }
 978 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
 979
 980 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
 981                            unsigned int cpu,
 982                            irq_handler_t handler,
 983                            unsigned long irqflags,
 984                            const char *devname,
 985                            void *dev_id)
 986 {
 987         int irq, retval;
 988
 989         irq = bind_ipi_to_irq(ipi, cpu);
 990         if (irq < 0)
 991                 return irq;
 992
 993         irqflags |= IRQF_NO_SUSPEND;
 994         retval = request_irq(irq, handler, irqflags, devname, dev_id);
 995         if (retval != 0) {
 996                 unbind_from_irq(irq);
 997                 return retval;
 998         }
 999
1000         return irq;
1001 }
1002
1003 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1004 {
1005         free_irq(irq, dev_id);
1006         unbind_from_irq(irq);
1007 }
1008 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1009
1010 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1011 {
1012         int irq = per_cpu(ipi_to_irq, cpu)[vector];
1013         BUG_ON(irq < 0);
1014         notify_remote_via_irq(irq);
1015 }
1016
1017 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
1018 {
1019         struct shared_info *sh = HYPERVISOR_shared_info;
1020         int cpu = smp_processor_id();
1021         unsigned long *cpu_evtchn = cpu_evtchn_mask(cpu);
1022         int i;
1023         unsigned long flags;
1024         static DEFINE_SPINLOCK(debug_lock);
1025         struct vcpu_info *v;
1026
1027         spin_lock_irqsave(&debug_lock, flags);
1028
1029         printk("\nvcpu %d\n  ", cpu);
1030
1031         for_each_online_cpu(i) {
1032                 int pending;
1033                 v = per_cpu(xen_vcpu, i);
1034                 pending = (get_irq_regs() && i == cpu)
1035                         ? xen_irqs_disabled(get_irq_regs())
1036                         : v->evtchn_upcall_mask;
1037                 printk("%d: masked=%d pending=%d event_sel %0*lx\n  ", i,
1038                        pending, v->evtchn_upcall_pending,
1039                        (int)(sizeof(v->evtchn_pending_sel)*2),
1040                        v->evtchn_pending_sel);
1041         }
1042         v = per_cpu(xen_vcpu, cpu);
1043
1044         printk("\npending:\n   ");
1045         for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
1046                 printk("%0*lx%s", (int)sizeof(sh->evtchn_pending[0])*2,
1047                        sh->evtchn_pending[i],
1048                        i % 8 == 0 ? "\n   " : " ");
1049         printk("\nglobal mask:\n   ");
1050         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1051                 printk("%0*lx%s",
1052                        (int)(sizeof(sh->evtchn_mask[0])*2),
1053                        sh->evtchn_mask[i],
1054                        i % 8 == 0 ? "\n   " : " ");
1055
1056         printk("\nglobally unmasked:\n   ");
1057         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1058                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1059                        sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
1060                        i % 8 == 0 ? "\n   " : " ");
1061
1062         printk("\nlocal cpu%d mask:\n   ", cpu);
1063         for (i = (NR_EVENT_CHANNELS/BITS_PER_LONG)-1; i >= 0; i--)
1064                 printk("%0*lx%s", (int)(sizeof(cpu_evtchn[0])*2),
1065                        cpu_evtchn[i],
1066                        i % 8 == 0 ? "\n   " : " ");
1067
1068         printk("\nlocally unmasked:\n   ");
1069         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
1070                 unsigned long pending = sh->evtchn_pending[i]
1071                         & ~sh->evtchn_mask[i]
1072                         & cpu_evtchn[i];
1073                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1074                        pending, i % 8 == 0 ? "\n   " : " ");
1075         }
1076
1077         printk("\npending list:\n");
1078         for (i = 0; i < NR_EVENT_CHANNELS; i++) {
1079                 if (sync_test_bit(i, sh->evtchn_pending)) {
1080                         int word_idx = i / BITS_PER_LONG;
1081                         printk("  %d: event %d -> irq %d%s%s%s\n",
1082                                cpu_from_evtchn(i), i,
1083                                evtchn_to_irq[i],
1084                                sync_test_bit(word_idx, &v->evtchn_pending_sel)
1085                                              ? "" : " l2-clear",
1086                                !sync_test_bit(i, sh->evtchn_mask)
1087                                              ? "" : " globally-masked",
1088                                sync_test_bit(i, cpu_evtchn)
1089                                              ? "" : " locally-masked");
1090                 }
1091         }
1092
1093         spin_unlock_irqrestore(&debug_lock, flags);
1094
1095         return IRQ_HANDLED;
1096 }
1097
1098 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
1099
1100 /*
1101  * Search the CPUs pending events bitmasks.  For each one found, map
1102  * the event number to an irq, and feed it into do_IRQ() for
1103  * handling.
1104  *
1105  * Xen uses a two-level bitmap to speed searching.  The first level is
1106  * a bitset of words which contain pending event bits.  The second
1107  * level is a bitset of pending events themselves.
1108  */
1109 static void __xen_evtchn_do_upcall(void)
1110 {
1111         int cpu = get_cpu();
1112         struct shared_info *s = HYPERVISOR_shared_info;
1113         struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1114         unsigned count;
1115
1116         do {
1117                 unsigned long pending_words;
1118
1119                 vcpu_info->evtchn_upcall_pending = 0;
1120
1121                 if (__this_cpu_inc_return(xed_nesting_count) - 1)
1122                         goto out;
1123
1124 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
1125                 /* Clear master flag /before/ clearing selector flag. */
1126                 wmb();
1127 #endif
1128                 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
1129                 while (pending_words != 0) {
1130                         unsigned long pending_bits;
1131                         int word_idx = __ffs(pending_words);
1132                         pending_words &= ~(1UL << word_idx);
1133
1134                         while ((pending_bits = active_evtchns(cpu, s, word_idx)) != 0) {
1135                                 int bit_idx = __ffs(pending_bits);
1136                                 int port = (word_idx * BITS_PER_LONG) + bit_idx;
1137                                 int irq = evtchn_to_irq[port];
1138                                 struct irq_desc *desc;
1139
1140                                 mask_evtchn(port);
1141                                 clear_evtchn(port);
1142
1143                                 if (irq != -1) {
1144                                         desc = irq_to_desc(irq);
1145                                         if (desc)
1146                                                 generic_handle_irq_desc(irq, desc);
1147                                 }
1148                         }
1149                 }
1150
1151                 BUG_ON(!irqs_disabled());
1152
1153                 count = __this_cpu_read(xed_nesting_count);
1154                 __this_cpu_write(xed_nesting_count, 0);
1155         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
1156
1157 out:
1158
1159         put_cpu();
1160 }
1161
1162 void xen_evtchn_do_upcall(struct pt_regs *regs)
1163 {
1164         struct pt_regs *old_regs = set_irq_regs(regs);
1165
1166         exit_idle();
1167         irq_enter();
1168
1169         __xen_evtchn_do_upcall();
1170
1171         irq_exit();
1172         set_irq_regs(old_regs);
1173 }
1174
1175 void xen_hvm_evtchn_do_upcall(void)
1176 {
1177         __xen_evtchn_do_upcall();
1178 }
1179 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1180
1181 /* Rebind a new event channel to an existing irq. */
1182 void rebind_evtchn_irq(int evtchn, int irq)
1183 {
1184         struct irq_info *info = info_for_irq(irq);
1185
1186         /* Make sure the irq is masked, since the new event channel
1187            will also be masked. */
1188         disable_irq(irq);
1189
1190         spin_lock(&irq_mapping_update_lock);
1191
1192         /* After resume the irq<->evtchn mappings are all cleared out */
1193         BUG_ON(evtchn_to_irq[evtchn] != -1);
1194         /* Expect irq to have been bound before,
1195            so there should be a proper type */
1196         BUG_ON(info->type == IRQT_UNBOUND);
1197
1198         evtchn_to_irq[evtchn] = irq;
1199         irq_info[irq] = mk_evtchn_info(evtchn);
1200
1201         spin_unlock(&irq_mapping_update_lock);
1202
1203         /* new event channels are always bound to cpu 0 */
1204         irq_set_affinity(irq, cpumask_of(0));
1205
1206         /* Unmask the event channel. */
1207         enable_irq(irq);
1208 }
1209
1210 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1211 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1212 {
1213         struct evtchn_bind_vcpu bind_vcpu;
1214         int evtchn = evtchn_from_irq(irq);
1215
1216         /* events delivered via platform PCI interrupts are always
1217          * routed to vcpu 0 */
1218         if (!VALID_EVTCHN(evtchn) ||
1219                 (xen_hvm_domain() && !xen_have_vector_callback))
1220                 return -1;
1221
1222         /* Send future instances of this interrupt to other vcpu. */
1223         bind_vcpu.port = evtchn;
1224         bind_vcpu.vcpu = tcpu;
1225
1226         /*
1227          * If this fails, it usually just indicates that we're dealing with a
1228          * virq or IPI channel, which don't actually need to be rebound. Ignore
1229          * it, but don't do the xenlinux-level rebind in that case.
1230          */
1231         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1232                 bind_evtchn_to_cpu(evtchn, tcpu);
1233
1234         return 0;
1235 }
1236
1237 static int set_affinity_irq(unsigned irq, const struct cpumask *dest)
1238 {
1239         unsigned tcpu = cpumask_first(dest);
1240
1241         return rebind_irq_to_cpu(irq, tcpu);
1242 }
1243
1244 int resend_irq_on_evtchn(unsigned int irq)
1245 {
1246         int masked, evtchn = evtchn_from_irq(irq);
1247         struct shared_info *s = HYPERVISOR_shared_info;
1248
1249         if (!VALID_EVTCHN(evtchn))
1250                 return 1;
1251
1252         masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
1253         sync_set_bit(evtchn, s->evtchn_pending);
1254         if (!masked)
1255                 unmask_evtchn(evtchn);
1256
1257         return 1;
1258 }
1259
1260 static void enable_dynirq(unsigned int irq)
1261 {
1262         int evtchn = evtchn_from_irq(irq);
1263
1264         if (VALID_EVTCHN(evtchn))
1265                 unmask_evtchn(evtchn);
1266 }
1267
1268 static void disable_dynirq(unsigned int irq)
1269 {
1270         int evtchn = evtchn_from_irq(irq);
1271
1272         if (VALID_EVTCHN(evtchn))
1273                 mask_evtchn(evtchn);
1274 }
1275
1276 static void ack_dynirq(unsigned int irq)
1277 {
1278         int evtchn = evtchn_from_irq(irq);
1279
1280         move_masked_irq(irq);
1281
1282         if (VALID_EVTCHN(evtchn))
1283                 unmask_evtchn(evtchn);
1284 }
1285
1286 static int retrigger_dynirq(unsigned int irq)
1287 {
1288         int evtchn = evtchn_from_irq(irq);
1289         struct shared_info *sh = HYPERVISOR_shared_info;
1290         int ret = 0;
1291
1292         if (VALID_EVTCHN(evtchn)) {
1293                 int masked;
1294
1295                 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
1296                 sync_set_bit(evtchn, sh->evtchn_pending);
1297                 if (!masked)
1298                         unmask_evtchn(evtchn);
1299                 ret = 1;
1300         }
1301
1302         return ret;
1303 }
1304
1305 static void restore_cpu_pirqs(void)
1306 {
1307         int pirq, rc, irq, gsi;
1308         struct physdev_map_pirq map_irq;
1309
1310         for (pirq = 0; pirq < nr_irqs; pirq++) {
1311                 irq = pirq_to_irq[pirq];
1312                 if (irq == -1)
1313                         continue;
1314
1315                 /* save/restore of PT devices doesn't work, so at this point the
1316                  * only devices present are GSI based emulated devices */
1317                 gsi = gsi_from_irq(irq);
1318                 if (!gsi)
1319                         continue;
1320
1321                 map_irq.domid = DOMID_SELF;
1322                 map_irq.type = MAP_PIRQ_TYPE_GSI;
1323                 map_irq.index = gsi;
1324                 map_irq.pirq = pirq;
1325
1326                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1327                 if (rc) {
1328                         printk(KERN_WARNING "xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1329                                         gsi, irq, pirq, rc);
1330                         irq_info[irq] = mk_unbound_info();
1331                         pirq_to_irq[pirq] = -1;
1332                         continue;
1333                 }
1334
1335                 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1336
1337                 startup_pirq(irq);
1338         }
1339 }
1340
1341 static void restore_cpu_virqs(unsigned int cpu)
1342 {
1343         struct evtchn_bind_virq bind_virq;
1344         int virq, irq, evtchn;
1345
1346         for (virq = 0; virq < NR_VIRQS; virq++) {
1347                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1348                         continue;
1349
1350                 BUG_ON(virq_from_irq(irq) != virq);
1351
1352                 /* Get a new binding from Xen. */
1353                 bind_virq.virq = virq;
1354                 bind_virq.vcpu = cpu;
1355                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1356                                                 &bind_virq) != 0)
1357                         BUG();
1358                 evtchn = bind_virq.port;
1359
1360                 /* Record the new mapping. */
1361                 evtchn_to_irq[evtchn] = irq;
1362                 irq_info[irq] = mk_virq_info(evtchn, virq);
1363                 bind_evtchn_to_cpu(evtchn, cpu);
1364         }
1365 }
1366
1367 static void restore_cpu_ipis(unsigned int cpu)
1368 {
1369         struct evtchn_bind_ipi bind_ipi;
1370         int ipi, irq, evtchn;
1371
1372         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1373                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1374                         continue;
1375
1376                 BUG_ON(ipi_from_irq(irq) != ipi);
1377
1378                 /* Get a new binding from Xen. */
1379                 bind_ipi.vcpu = cpu;
1380                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1381                                                 &bind_ipi) != 0)
1382                         BUG();
1383                 evtchn = bind_ipi.port;
1384
1385                 /* Record the new mapping. */
1386                 evtchn_to_irq[evtchn] = irq;
1387                 irq_info[irq] = mk_ipi_info(evtchn, ipi);
1388                 bind_evtchn_to_cpu(evtchn, cpu);
1389         }
1390 }
1391
1392 /* Clear an irq's pending state, in preparation for polling on it */
1393 void xen_clear_irq_pending(int irq)
1394 {
1395         int evtchn = evtchn_from_irq(irq);
1396
1397         if (VALID_EVTCHN(evtchn))
1398                 clear_evtchn(evtchn);
1399 }
1400 EXPORT_SYMBOL(xen_clear_irq_pending);
1401 void xen_set_irq_pending(int irq)
1402 {
1403         int evtchn = evtchn_from_irq(irq);
1404
1405         if (VALID_EVTCHN(evtchn))
1406                 set_evtchn(evtchn);
1407 }
1408
1409 bool xen_test_irq_pending(int irq)
1410 {
1411         int evtchn = evtchn_from_irq(irq);
1412         bool ret = false;
1413
1414         if (VALID_EVTCHN(evtchn))
1415                 ret = test_evtchn(evtchn);
1416
1417         return ret;
1418 }
1419
1420 /* Poll waiting for an irq to become pending with timeout.  In the usual case,
1421  * the irq will be disabled so it won't deliver an interrupt. */
1422 void xen_poll_irq_timeout(int irq, u64 timeout)
1423 {
1424         evtchn_port_t evtchn = evtchn_from_irq(irq);
1425
1426         if (VALID_EVTCHN(evtchn)) {
1427                 struct sched_poll poll;
1428
1429                 poll.nr_ports = 1;
1430                 poll.timeout = timeout;
1431                 set_xen_guest_handle(poll.ports, &evtchn);
1432
1433                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1434                         BUG();
1435         }
1436 }
1437 EXPORT_SYMBOL(xen_poll_irq_timeout);
1438 /* Poll waiting for an irq to become pending.  In the usual case, the
1439  * irq will be disabled so it won't deliver an interrupt. */
1440 void xen_poll_irq(int irq)
1441 {
1442         xen_poll_irq_timeout(irq, 0 /* no timeout */);
1443 }
1444
1445 void xen_irq_resume(void)
1446 {
1447         unsigned int cpu, irq, evtchn;
1448         struct irq_desc *desc;
1449
1450         init_evtchn_cpu_bindings();
1451
1452         /* New event-channel space is not 'live' yet. */
1453         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1454                 mask_evtchn(evtchn);
1455
1456         /* No IRQ <-> event-channel mappings. */
1457         for (irq = 0; irq < nr_irqs; irq++)
1458                 irq_info[irq].evtchn = 0; /* zap event-channel binding */
1459
1460         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1461                 evtchn_to_irq[evtchn] = -1;
1462
1463         for_each_possible_cpu(cpu) {
1464                 restore_cpu_virqs(cpu);
1465                 restore_cpu_ipis(cpu);
1466         }
1467
1468         /*
1469          * Unmask any IRQF_NO_SUSPEND IRQs which are enabled. These
1470          * are not handled by the IRQ core.
1471          */
1472         for_each_irq_desc(irq, desc) {
1473                 if (!desc->action || !(desc->action->flags & IRQF_NO_SUSPEND))
1474                         continue;
1475                 if (desc->status & IRQ_DISABLED)
1476                         continue;
1477
1478                 evtchn = evtchn_from_irq(irq);
1479                 if (evtchn == -1)
1480                         continue;
1481
1482                 unmask_evtchn(evtchn);
1483         }
1484
1485         restore_cpu_pirqs();
1486 }
1487
1488 static struct irq_chip xen_dynamic_chip __read_mostly = {
1489         .name           = "xen-dyn",
1490
1491         .disable        = disable_dynirq,
1492         .mask           = disable_dynirq,
1493         .unmask         = enable_dynirq,
1494
1495         .eoi            = ack_dynirq,
1496         .set_affinity   = set_affinity_irq,
1497         .retrigger      = retrigger_dynirq,
1498 };
1499
1500 static struct irq_chip xen_pirq_chip __read_mostly = {
1501         .name           = "xen-pirq",
1502
1503         .startup        = startup_pirq,
1504         .shutdown       = shutdown_pirq,
1505
1506         .enable         = enable_pirq,
1507         .unmask         = enable_pirq,
1508
1509         .disable        = disable_pirq,
1510         .mask           = disable_pirq,
1511
1512         .ack            = ack_pirq,
1513         .end            = end_pirq,
1514
1515         .set_affinity   = set_affinity_irq,
1516
1517         .retrigger      = retrigger_dynirq,
1518 };
1519
1520 static struct irq_chip xen_percpu_chip __read_mostly = {
1521         .name           = "xen-percpu",
1522
1523         .disable        = disable_dynirq,
1524         .mask           = disable_dynirq,
1525         .unmask         = enable_dynirq,
1526
1527         .ack            = ack_dynirq,
1528 };
1529
1530 int xen_set_callback_via(uint64_t via)
1531 {
1532         struct xen_hvm_param a;
1533         a.domid = DOMID_SELF;
1534         a.index = HVM_PARAM_CALLBACK_IRQ;
1535         a.value = via;
1536         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1537 }
1538 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1539
1540 #ifdef CONFIG_XEN_PVHVM
1541 /* Vector callbacks are better than PCI interrupts to receive event
1542  * channel notifications because we can receive vector callbacks on any
1543  * vcpu and we don't need PCI support or APIC interactions. */
1544 void xen_callback_vector(void)
1545 {
1546         int rc;
1547         uint64_t callback_via;
1548         if (xen_have_vector_callback) {
1549                 callback_via = HVM_CALLBACK_VECTOR(XEN_HVM_EVTCHN_CALLBACK);
1550                 rc = xen_set_callback_via(callback_via);
1551                 if (rc) {
1552                         printk(KERN_ERR "Request for Xen HVM callback vector"
1553                                         " failed.\n");
1554                         xen_have_vector_callback = 0;
1555                         return;
1556                 }
1557                 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1558                                 "enabled\n");
1559                 /* in the restore case the vector has already been allocated */
1560                 if (!test_bit(XEN_HVM_EVTCHN_CALLBACK, used_vectors))
1561                         alloc_intr_gate(XEN_HVM_EVTCHN_CALLBACK, xen_hvm_callback_vector);
1562         }
1563 }
1564 #else
1565 void xen_callback_vector(void) {}
1566 #endif
1567
1568 void __init xen_init_IRQ(void)
1569 {
1570         int i;
1571
1572         cpu_evtchn_mask_p = kcalloc(nr_cpu_ids, sizeof(struct cpu_evtchn_s),
1573                                     GFP_KERNEL);
1574         irq_info = kcalloc(nr_irqs, sizeof(*irq_info), GFP_KERNEL);
1575
1576         /* We are using nr_irqs as the maximum number of pirq available but
1577          * that number is actually chosen by Xen and we don't know exactly
1578          * what it is. Be careful choosing high pirq numbers. */
1579         pirq_to_irq = kcalloc(nr_irqs, sizeof(*pirq_to_irq), GFP_KERNEL);
1580         for (i = 0; i < nr_irqs; i++)
1581                 pirq_to_irq[i] = -1;
1582
1583         evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1584                                     GFP_KERNEL);
1585         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1586                 evtchn_to_irq[i] = -1;
1587
1588         init_evtchn_cpu_bindings();
1589
1590         /* No event channels are 'live' right now. */
1591         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1592                 mask_evtchn(i);
1593
1594         if (xen_hvm_domain()) {
1595                 xen_callback_vector();
1596                 native_init_IRQ();
1597                 /* pci_xen_hvm_init must be called after native_init_IRQ so that
1598                  * __acpi_register_gsi can point at the right function */
1599                 pci_xen_hvm_init();
1600         } else {
1601                 irq_ctx_init(smp_processor_id());
1602                 if (xen_initial_domain())
1603                         xen_setup_pirqs();
1604         }
1605 }