drivers/xen/events/events_base.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Xen event channels
   4  *
   5  * Xen models interrupts with abstract event channels.  Because each
   6  * domain gets 1024 event channels, but NR_IRQ is not that large, we
   7  * must dynamically map irqs<->event channels.  The event channels
   8  * interface with the rest of the kernel by defining a xen interrupt
   9  * chip.  When an event is received, it is mapped to an irq and sent
  10  * through the normal interrupt processing path.
  11  *
  12  * There are four kinds of events which can be mapped to an event
  13  * channel:
  14  *
  15  * 1. Inter-domain notifications.  This includes all the virtual
  16  *    device events, since they're driven by front-ends in another domain
  17  *    (typically dom0).
  18  * 2. VIRQs, typically used for timers.  These are per-cpu events.
  19  * 3. IPIs.
  20  * 4. PIRQs - Hardware interrupts.
  21  *
  22  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
  23  */
  24
  25 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
  26
  27 #include <linux/linkage.h>
  28 #include <linux/interrupt.h>
  29 #include <linux/irq.h>
  30 #include <linux/moduleparam.h>
  31 #include <linux/string.h>
  32 #include <linux/memblock.h>
  33 #include <linux/slab.h>
  34 #include <linux/irqnr.h>
  35 #include <linux/pci.h>
  36 #include <linux/rcupdate.h>
  37 #include <linux/spinlock.h>
  38 #include <linux/cpuhotplug.h>
  39 #include <linux/atomic.h>
  40 #include <linux/ktime.h>
  41
  42 #ifdef CONFIG_X86
  43 #include <asm/desc.h>
  44 #include <asm/ptrace.h>
  45 #include <asm/idtentry.h>
  46 #include <asm/irq.h>
  47 #include <asm/io_apic.h>
  48 #include <asm/i8259.h>
  49 #include <asm/xen/cpuid.h>
  50 #include <asm/xen/pci.h>
  51 #endif
  52 #include <asm/sync_bitops.h>
  53 #include <asm/xen/hypercall.h>
  54 #include <asm/xen/hypervisor.h>
  55 #include <xen/page.h>
  56
  57 #include <xen/xen.h>
  58 #include <xen/hvm.h>
  59 #include <xen/xen-ops.h>
  60 #include <xen/events.h>
  61 #include <xen/interface/xen.h>
  62 #include <xen/interface/event_channel.h>
  63 #include <xen/interface/hvm/hvm_op.h>
  64 #include <xen/interface/hvm/params.h>
  65 #include <xen/interface/physdev.h>
  66 #include <xen/interface/sched.h>
  67 #include <xen/interface/vcpu.h>
  68 #include <xen/xenbus.h>
  69 #include <asm/hw_irq.h>
  70
  71 #include "events_internal.h"
  72
  73 #undef MODULE_PARAM_PREFIX
  74 #define MODULE_PARAM_PREFIX "xen."
  75
  76 /* Interrupt types. */
  77 enum xen_irq_type {
  78         IRQT_UNBOUND = 0,
  79         IRQT_PIRQ,
  80         IRQT_VIRQ,
  81         IRQT_IPI,
  82         IRQT_EVTCHN
  83 };
  84
  85 /*
  86  * Packed IRQ information:
  87  * type - enum xen_irq_type
  88  * event channel - irq->event channel mapping
  89  * cpu - cpu this event channel is bound to
  90  * index - type-specific information:
  91  *    PIRQ - vector, with MSB being "needs EIO", or physical IRQ of the HVM
  92  *           guest, or GSI (real passthrough IRQ) of the device.
  93  *    VIRQ - virq number
  94  *    IPI - IPI vector
  95  *    EVTCHN -
  96  */
  97 struct irq_info {
  98         struct list_head list;
  99         struct list_head eoi_list;
 100         struct rcu_work rwork;
 101         short refcnt;
 102         u8 spurious_cnt;
 103         u8 is_accounted;
 104         short type;             /* type: IRQT_* */
 105         u8 mask_reason;         /* Why is event channel masked */
 106 #define EVT_MASK_REASON_EXPLICIT        0x01
 107 #define EVT_MASK_REASON_TEMPORARY       0x02
 108 #define EVT_MASK_REASON_EOI_PENDING     0x04
 109         u8 is_active;           /* Is event just being handled? */
 110         unsigned irq;
 111         evtchn_port_t evtchn;   /* event channel */
 112         unsigned short cpu;     /* cpu bound */
 113         unsigned short eoi_cpu; /* EOI must happen on this cpu-1 */
 114         unsigned int irq_epoch; /* If eoi_cpu valid: irq_epoch of event */
 115         u64 eoi_time;           /* Time in jiffies when to EOI. */
 116         raw_spinlock_t lock;
 117         bool is_static;           /* Is event channel static */
 118
 119         union {
 120                 unsigned short virq;
 121                 enum ipi_vector ipi;
 122                 struct {
 123                         unsigned short pirq;
 124                         unsigned short gsi;
 125                         unsigned char vector;
 126                         unsigned char flags;
 127                         uint16_t domid;
 128                 } pirq;
 129                 struct xenbus_device *interdomain;
 130         } u;
 131 };
 132
 133 #define PIRQ_NEEDS_EOI  (1 << 0)
 134 #define PIRQ_SHAREABLE  (1 << 1)
 135 #define PIRQ_MSI_GROUP  (1 << 2)
 136
 137 static uint __read_mostly event_loop_timeout = 2;
 138 module_param(event_loop_timeout, uint, 0644);
 139
 140 static uint __read_mostly event_eoi_delay = 10;
 141 module_param(event_eoi_delay, uint, 0644);
 142
 143 const struct evtchn_ops *evtchn_ops;
 144
 145 /*
 146  * This lock protects updates to the following mapping and reference-count
 147  * arrays. The lock does not need to be acquired to read the mapping tables.
 148  */
 149 static DEFINE_MUTEX(irq_mapping_update_lock);
 150
 151 /*
 152  * Lock hierarchy:
 153  *
 154  * irq_mapping_update_lock
 155  *   IRQ-desc lock
 156  *     percpu eoi_list_lock
 157  *       irq_info->lock
 158  */
 159
 160 static LIST_HEAD(xen_irq_list_head);
 161
 162 /* IRQ <-> VIRQ mapping. */
 163 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
 164
 165 /* IRQ <-> IPI mapping */
 166 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
 167 /* Cache for IPI event channels - needed for hot cpu unplug (avoid RCU usage). */
 168 static DEFINE_PER_CPU(evtchn_port_t [XEN_NR_IPIS], ipi_to_evtchn) = {[0 ... XEN_NR_IPIS-1] = 0};
 169
 170 /* Event channel distribution data */
 171 static atomic_t channels_on_cpu[NR_CPUS];
 172
 173 static int **evtchn_to_irq;
 174 #ifdef CONFIG_X86
 175 static unsigned long *pirq_eoi_map;
 176 #endif
 177 static bool (*pirq_needs_eoi)(struct irq_info *info);
 178
 179 #define EVTCHN_ROW(e)  (e / (PAGE_SIZE/sizeof(**evtchn_to_irq)))
 180 #define EVTCHN_COL(e)  (e % (PAGE_SIZE/sizeof(**evtchn_to_irq)))
 181 #define EVTCHN_PER_ROW (PAGE_SIZE / sizeof(**evtchn_to_irq))
 182
 183 /* Xen will never allocate port zero for any purpose. */
 184 #define VALID_EVTCHN(chn)       ((chn) != 0)
 185
 186 static struct irq_info *legacy_info_ptrs[NR_IRQS_LEGACY];
 187
 188 static struct irq_chip xen_dynamic_chip;
 189 static struct irq_chip xen_lateeoi_chip;
 190 static struct irq_chip xen_percpu_chip;
 191 static struct irq_chip xen_pirq_chip;
 192 static void enable_dynirq(struct irq_data *data);
 193
 194 static DEFINE_PER_CPU(unsigned int, irq_epoch);
 195
 196 static void clear_evtchn_to_irq_row(int *evtchn_row)
 197 {
 198         unsigned col;
 199
 200         for (col = 0; col < EVTCHN_PER_ROW; col++)
 201                 WRITE_ONCE(evtchn_row[col], -1);
 202 }
 203
 204 static void clear_evtchn_to_irq_all(void)
 205 {
 206         unsigned row;
 207
 208         for (row = 0; row < EVTCHN_ROW(xen_evtchn_max_channels()); row++) {
 209                 if (evtchn_to_irq[row] == NULL)
 210                         continue;
 211                 clear_evtchn_to_irq_row(evtchn_to_irq[row]);
 212         }
 213 }
 214
 215 static int set_evtchn_to_irq(evtchn_port_t evtchn, unsigned int irq)
 216 {
 217         unsigned row;
 218         unsigned col;
 219         int *evtchn_row;
 220
 221         if (evtchn >= xen_evtchn_max_channels())
 222                 return -EINVAL;
 223
 224         row = EVTCHN_ROW(evtchn);
 225         col = EVTCHN_COL(evtchn);
 226
 227         if (evtchn_to_irq[row] == NULL) {
 228                 /* Unallocated irq entries return -1 anyway */
 229                 if (irq == -1)
 230                         return 0;
 231
 232                 evtchn_row = (int *) __get_free_pages(GFP_KERNEL, 0);
 233                 if (evtchn_row == NULL)
 234                         return -ENOMEM;
 235
 236                 clear_evtchn_to_irq_row(evtchn_row);
 237
 238                 /*
 239                  * We've prepared an empty row for the mapping. If a different
 240                  * thread was faster inserting it, we can drop ours.
 241                  */
 242                 if (cmpxchg(&evtchn_to_irq[row], NULL, evtchn_row) != NULL)
 243                         free_page((unsigned long) evtchn_row);
 244         }
 245
 246         WRITE_ONCE(evtchn_to_irq[row][col], irq);
 247         return 0;
 248 }
 249
 250 /* Get info for IRQ */
 251 static struct irq_info *info_for_irq(unsigned irq)
 252 {
 253         if (irq < nr_legacy_irqs())
 254                 return legacy_info_ptrs[irq];
 255         else
 256                 return irq_get_chip_data(irq);
 257 }
 258
 259 static void set_info_for_irq(unsigned int irq, struct irq_info *info)
 260 {
 261         if (irq < nr_legacy_irqs())
 262                 legacy_info_ptrs[irq] = info;
 263         else
 264                 irq_set_chip_data(irq, info);
 265 }
 266
 267 static struct irq_info *evtchn_to_info(evtchn_port_t evtchn)
 268 {
 269         int irq;
 270
 271         if (evtchn >= xen_evtchn_max_channels())
 272                 return NULL;
 273         if (evtchn_to_irq[EVTCHN_ROW(evtchn)] == NULL)
 274                 return NULL;
 275         irq = READ_ONCE(evtchn_to_irq[EVTCHN_ROW(evtchn)][EVTCHN_COL(evtchn)]);
 276
 277         return (irq < 0) ? NULL : info_for_irq(irq);
 278 }
 279
 280 /* Per CPU channel accounting */
 281 static void channels_on_cpu_dec(struct irq_info *info)
 282 {
 283         if (!info->is_accounted)
 284                 return;
 285
 286         info->is_accounted = 0;
 287
 288         if (WARN_ON_ONCE(info->cpu >= nr_cpu_ids))
 289                 return;
 290
 291         WARN_ON_ONCE(!atomic_add_unless(&channels_on_cpu[info->cpu], -1 , 0));
 292 }
 293
 294 static void channels_on_cpu_inc(struct irq_info *info)
 295 {
 296         if (WARN_ON_ONCE(info->cpu >= nr_cpu_ids))
 297                 return;
 298
 299         if (WARN_ON_ONCE(!atomic_add_unless(&channels_on_cpu[info->cpu], 1,
 300                                             INT_MAX)))
 301                 return;
 302
 303         info->is_accounted = 1;
 304 }
 305
 306 static void xen_irq_free_desc(unsigned int irq)
 307 {
 308         /* Legacy IRQ descriptors are managed by the arch. */
 309         if (irq >= nr_legacy_irqs())
 310                 irq_free_desc(irq);
 311 }
 312
 313 static void delayed_free_irq(struct work_struct *work)
 314 {
 315         struct irq_info *info = container_of(to_rcu_work(work), struct irq_info,
 316                                              rwork);
 317         unsigned int irq = info->irq;
 318
 319         /* Remove the info pointer only now, with no potential users left. */
 320         set_info_for_irq(irq, NULL);
 321
 322         kfree(info);
 323
 324         xen_irq_free_desc(irq);
 325 }
 326
 327 /* Constructors for packed IRQ information. */
 328 static int xen_irq_info_common_setup(struct irq_info *info,
 329                                      enum xen_irq_type type,
 330                                      evtchn_port_t evtchn,
 331                                      unsigned short cpu)
 332 {
 333         int ret;
 334
 335         BUG_ON(info->type != IRQT_UNBOUND && info->type != type);
 336
 337         info->type = type;
 338         info->evtchn = evtchn;
 339         info->cpu = cpu;
 340         info->mask_reason = EVT_MASK_REASON_EXPLICIT;
 341         raw_spin_lock_init(&info->lock);
 342
 343         ret = set_evtchn_to_irq(evtchn, info->irq);
 344         if (ret < 0)
 345                 return ret;
 346
 347         irq_clear_status_flags(info->irq, IRQ_NOREQUEST | IRQ_NOAUTOEN);
 348
 349         return xen_evtchn_port_setup(evtchn);
 350 }
 351
 352 static int xen_irq_info_evtchn_setup(struct irq_info *info,
 353                                      evtchn_port_t evtchn,
 354                                      struct xenbus_device *dev)
 355 {
 356         int ret;
 357
 358         ret = xen_irq_info_common_setup(info, IRQT_EVTCHN, evtchn, 0);
 359         info->u.interdomain = dev;
 360         if (dev)
 361                 atomic_inc(&dev->event_channels);
 362
 363         return ret;
 364 }
 365
 366 static int xen_irq_info_ipi_setup(struct irq_info *info, unsigned int cpu,
 367                                   evtchn_port_t evtchn, enum ipi_vector ipi)
 368 {
 369         info->u.ipi = ipi;
 370
 371         per_cpu(ipi_to_irq, cpu)[ipi] = info->irq;
 372         per_cpu(ipi_to_evtchn, cpu)[ipi] = evtchn;
 373
 374         return xen_irq_info_common_setup(info, IRQT_IPI, evtchn, 0);
 375 }
 376
 377 static int xen_irq_info_virq_setup(struct irq_info *info, unsigned int cpu,
 378                                    evtchn_port_t evtchn, unsigned int virq)
 379 {
 380         info->u.virq = virq;
 381
 382         per_cpu(virq_to_irq, cpu)[virq] = info->irq;
 383
 384         return xen_irq_info_common_setup(info, IRQT_VIRQ, evtchn, 0);
 385 }
 386
 387 static int xen_irq_info_pirq_setup(struct irq_info *info, evtchn_port_t evtchn,
 388                                    unsigned int pirq, unsigned int gsi,
 389                                    uint16_t domid, unsigned char flags)
 390 {
 391         info->u.pirq.pirq = pirq;
 392         info->u.pirq.gsi = gsi;
 393         info->u.pirq.domid = domid;
 394         info->u.pirq.flags = flags;
 395
 396         return xen_irq_info_common_setup(info, IRQT_PIRQ, evtchn, 0);
 397 }
 398
 399 static void xen_irq_info_cleanup(struct irq_info *info)
 400 {
 401         set_evtchn_to_irq(info->evtchn, -1);
 402         xen_evtchn_port_remove(info->evtchn, info->cpu);
 403         info->evtchn = 0;
 404         channels_on_cpu_dec(info);
 405 }
 406
 407 /*
 408  * Accessors for packed IRQ information.
 409  */
 410 static evtchn_port_t evtchn_from_irq(unsigned int irq)
 411 {
 412         const struct irq_info *info = NULL;
 413
 414         if (likely(irq < nr_irqs))
 415                 info = info_for_irq(irq);
 416         if (!info)
 417                 return 0;
 418
 419         return info->evtchn;
 420 }
 421
 422 unsigned int irq_from_evtchn(evtchn_port_t evtchn)
 423 {
 424         struct irq_info *info = evtchn_to_info(evtchn);
 425
 426         return info ? info->irq : -1;
 427 }
 428 EXPORT_SYMBOL_GPL(irq_from_evtchn);
 429
 430 int irq_evtchn_from_virq(unsigned int cpu, unsigned int virq,
 431                          evtchn_port_t *evtchn)
 432 {
 433         int irq = per_cpu(virq_to_irq, cpu)[virq];
 434
 435         *evtchn = evtchn_from_irq(irq);
 436
 437         return irq;
 438 }
 439
 440 static enum ipi_vector ipi_from_irq(struct irq_info *info)
 441 {
 442         BUG_ON(info == NULL);
 443         BUG_ON(info->type != IRQT_IPI);
 444
 445         return info->u.ipi;
 446 }
 447
 448 static unsigned int virq_from_irq(struct irq_info *info)
 449 {
 450         BUG_ON(info == NULL);
 451         BUG_ON(info->type != IRQT_VIRQ);
 452
 453         return info->u.virq;
 454 }
 455
 456 static unsigned int pirq_from_irq(struct irq_info *info)
 457 {
 458         BUG_ON(info == NULL);
 459         BUG_ON(info->type != IRQT_PIRQ);
 460
 461         return info->u.pirq.pirq;
 462 }
 463
 464 unsigned int cpu_from_evtchn(evtchn_port_t evtchn)
 465 {
 466         struct irq_info *info = evtchn_to_info(evtchn);
 467
 468         return info ? info->cpu : 0;
 469 }
 470
 471 static void do_mask(struct irq_info *info, u8 reason)
 472 {
 473         unsigned long flags;
 474
 475         raw_spin_lock_irqsave(&info->lock, flags);
 476
 477         if (!info->mask_reason)
 478                 mask_evtchn(info->evtchn);
 479
 480         info->mask_reason |= reason;
 481
 482         raw_spin_unlock_irqrestore(&info->lock, flags);
 483 }
 484
 485 static void do_unmask(struct irq_info *info, u8 reason)
 486 {
 487         unsigned long flags;
 488
 489         raw_spin_lock_irqsave(&info->lock, flags);
 490
 491         info->mask_reason &= ~reason;
 492
 493         if (!info->mask_reason)
 494                 unmask_evtchn(info->evtchn);
 495
 496         raw_spin_unlock_irqrestore(&info->lock, flags);
 497 }
 498
 499 #ifdef CONFIG_X86
 500 static bool pirq_check_eoi_map(struct irq_info *info)
 501 {
 502         return test_bit(pirq_from_irq(info), pirq_eoi_map);
 503 }
 504 #endif
 505
 506 static bool pirq_needs_eoi_flag(struct irq_info *info)
 507 {
 508         BUG_ON(info->type != IRQT_PIRQ);
 509
 510         return info->u.pirq.flags & PIRQ_NEEDS_EOI;
 511 }
 512
 513 static void bind_evtchn_to_cpu(struct irq_info *info, unsigned int cpu,
 514                                bool force_affinity)
 515 {
 516         if (IS_ENABLED(CONFIG_SMP) && force_affinity) {
 517                 struct irq_data *data = irq_get_irq_data(info->irq);
 518
 519                 irq_data_update_affinity(data, cpumask_of(cpu));
 520                 irq_data_update_effective_affinity(data, cpumask_of(cpu));
 521         }
 522
 523         xen_evtchn_port_bind_to_cpu(info->evtchn, cpu, info->cpu);
 524
 525         channels_on_cpu_dec(info);
 526         info->cpu = cpu;
 527         channels_on_cpu_inc(info);
 528 }
 529
 530 /**
 531  * notify_remote_via_irq - send event to remote end of event channel via irq
 532  * @irq: irq of event channel to send event to
 533  *
 534  * Unlike notify_remote_via_evtchn(), this is safe to use across
 535  * save/restore. Notifications on a broken connection are silently
 536  * dropped.
 537  */
 538 void notify_remote_via_irq(int irq)
 539 {
 540         evtchn_port_t evtchn = evtchn_from_irq(irq);
 541
 542         if (VALID_EVTCHN(evtchn))
 543                 notify_remote_via_evtchn(evtchn);
 544 }
 545 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
 546
 547 struct lateeoi_work {
 548         struct delayed_work delayed;
 549         spinlock_t eoi_list_lock;
 550         struct list_head eoi_list;
 551 };
 552
 553 static DEFINE_PER_CPU(struct lateeoi_work, lateeoi);
 554
 555 static void lateeoi_list_del(struct irq_info *info)
 556 {
 557         struct lateeoi_work *eoi = &per_cpu(lateeoi, info->eoi_cpu);
 558         unsigned long flags;
 559
 560         spin_lock_irqsave(&eoi->eoi_list_lock, flags);
 561         list_del_init(&info->eoi_list);
 562         spin_unlock_irqrestore(&eoi->eoi_list_lock, flags);
 563 }
 564
 565 static void lateeoi_list_add(struct irq_info *info)
 566 {
 567         struct lateeoi_work *eoi = &per_cpu(lateeoi, info->eoi_cpu);
 568         struct irq_info *elem;
 569         u64 now = get_jiffies_64();
 570         unsigned long delay;
 571         unsigned long flags;
 572
 573         if (now < info->eoi_time)
 574                 delay = info->eoi_time - now;
 575         else
 576                 delay = 1;
 577
 578         spin_lock_irqsave(&eoi->eoi_list_lock, flags);
 579
 580         elem = list_first_entry_or_null(&eoi->eoi_list, struct irq_info,
 581                                         eoi_list);
 582         if (!elem || info->eoi_time < elem->eoi_time) {
 583                 list_add(&info->eoi_list, &eoi->eoi_list);
 584                 mod_delayed_work_on(info->eoi_cpu, system_wq,
 585                                     &eoi->delayed, delay);
 586         } else {
 587                 list_for_each_entry_reverse(elem, &eoi->eoi_list, eoi_list) {
 588                         if (elem->eoi_time <= info->eoi_time)
 589                                 break;
 590                 }
 591                 list_add(&info->eoi_list, &elem->eoi_list);
 592         }
 593
 594         spin_unlock_irqrestore(&eoi->eoi_list_lock, flags);
 595 }
 596
 597 static void xen_irq_lateeoi_locked(struct irq_info *info, bool spurious)
 598 {
 599         evtchn_port_t evtchn;
 600         unsigned int cpu;
 601         unsigned int delay = 0;
 602
 603         evtchn = info->evtchn;
 604         if (!VALID_EVTCHN(evtchn) || !list_empty(&info->eoi_list))
 605                 return;
 606
 607         if (spurious) {
 608                 struct xenbus_device *dev = info->u.interdomain;
 609                 unsigned int threshold = 1;
 610
 611                 if (dev && dev->spurious_threshold)
 612                         threshold = dev->spurious_threshold;
 613
 614                 if ((1 << info->spurious_cnt) < (HZ << 2)) {
 615                         if (info->spurious_cnt != 0xFF)
 616                                 info->spurious_cnt++;
 617                 }
 618                 if (info->spurious_cnt > threshold) {
 619                         delay = 1 << (info->spurious_cnt - 1 - threshold);
 620                         if (delay > HZ)
 621                                 delay = HZ;
 622                         if (!info->eoi_time)
 623                                 info->eoi_cpu = smp_processor_id();
 624                         info->eoi_time = get_jiffies_64() + delay;
 625                         if (dev)
 626                                 atomic_add(delay, &dev->jiffies_eoi_delayed);
 627                 }
 628                 if (dev)
 629                         atomic_inc(&dev->spurious_events);
 630         } else {
 631                 info->spurious_cnt = 0;
 632         }
 633
 634         cpu = info->eoi_cpu;
 635         if (info->eoi_time &&
 636             (info->irq_epoch == per_cpu(irq_epoch, cpu) || delay)) {
 637                 lateeoi_list_add(info);
 638                 return;
 639         }
 640
 641         info->eoi_time = 0;
 642
 643         /* is_active hasn't been reset yet, do it now. */
 644         smp_store_release(&info->is_active, 0);
 645         do_unmask(info, EVT_MASK_REASON_EOI_PENDING);
 646 }
 647
 648 static void xen_irq_lateeoi_worker(struct work_struct *work)
 649 {
 650         struct lateeoi_work *eoi;
 651         struct irq_info *info;
 652         u64 now = get_jiffies_64();
 653         unsigned long flags;
 654
 655         eoi = container_of(to_delayed_work(work), struct lateeoi_work, delayed);
 656
 657         rcu_read_lock();
 658
 659         while (true) {
 660                 spin_lock_irqsave(&eoi->eoi_list_lock, flags);
 661
 662                 info = list_first_entry_or_null(&eoi->eoi_list, struct irq_info,
 663                                                 eoi_list);
 664
 665                 if (info == NULL)
 666                         break;
 667
 668                 if (now < info->eoi_time) {
 669                         mod_delayed_work_on(info->eoi_cpu, system_wq,
 670                                             &eoi->delayed,
 671                                             info->eoi_time - now);
 672                         break;
 673                 }
 674
 675                 list_del_init(&info->eoi_list);
 676
 677                 spin_unlock_irqrestore(&eoi->eoi_list_lock, flags);
 678
 679                 info->eoi_time = 0;
 680
 681                 xen_irq_lateeoi_locked(info, false);
 682         }
 683
 684         spin_unlock_irqrestore(&eoi->eoi_list_lock, flags);
 685
 686         rcu_read_unlock();
 687 }
 688
 689 static void xen_cpu_init_eoi(unsigned int cpu)
 690 {
 691         struct lateeoi_work *eoi = &per_cpu(lateeoi, cpu);
 692
 693         INIT_DELAYED_WORK(&eoi->delayed, xen_irq_lateeoi_worker);
 694         spin_lock_init(&eoi->eoi_list_lock);
 695         INIT_LIST_HEAD(&eoi->eoi_list);
 696 }
 697
 698 void xen_irq_lateeoi(unsigned int irq, unsigned int eoi_flags)
 699 {
 700         struct irq_info *info;
 701
 702         rcu_read_lock();
 703
 704         info = info_for_irq(irq);
 705
 706         if (info)
 707                 xen_irq_lateeoi_locked(info, eoi_flags & XEN_EOI_FLAG_SPURIOUS);
 708
 709         rcu_read_unlock();
 710 }
 711 EXPORT_SYMBOL_GPL(xen_irq_lateeoi);
 712
 713 static struct irq_info *xen_irq_init(unsigned int irq)
 714 {
 715         struct irq_info *info;
 716
 717         info = kzalloc(sizeof(*info), GFP_KERNEL);
 718         if (info) {
 719                 info->irq = irq;
 720                 info->type = IRQT_UNBOUND;
 721                 info->refcnt = -1;
 722                 INIT_RCU_WORK(&info->rwork, delayed_free_irq);
 723
 724                 set_info_for_irq(irq, info);
 725                 /*
 726                  * Interrupt affinity setting can be immediate. No point
 727                  * in delaying it until an interrupt is handled.
 728                  */
 729                 irq_set_status_flags(irq, IRQ_MOVE_PCNTXT);
 730
 731                 INIT_LIST_HEAD(&info->eoi_list);
 732                 list_add_tail(&info->list, &xen_irq_list_head);
 733         }
 734
 735         return info;
 736 }
 737
 738 static struct irq_info *xen_allocate_irq_dynamic(void)
 739 {
 740         int irq = irq_alloc_desc_from(0, -1);
 741         struct irq_info *info = NULL;
 742
 743         if (irq >= 0) {
 744                 info = xen_irq_init(irq);
 745                 if (!info)
 746                         xen_irq_free_desc(irq);
 747         }
 748
 749         return info;
 750 }
 751
 752 static struct irq_info *xen_allocate_irq_gsi(unsigned int gsi)
 753 {
 754         int irq;
 755         struct irq_info *info;
 756
 757         /*
 758          * A PV guest has no concept of a GSI (since it has no ACPI
 759          * nor access to/knowledge of the physical APICs). Therefore
 760          * all IRQs are dynamically allocated from the entire IRQ
 761          * space.
 762          */
 763         if (xen_pv_domain() && !xen_initial_domain())
 764                 return xen_allocate_irq_dynamic();
 765
 766         /* Legacy IRQ descriptors are already allocated by the arch. */
 767         if (gsi < nr_legacy_irqs())
 768                 irq = gsi;
 769         else
 770                 irq = irq_alloc_desc_at(gsi, -1);
 771
 772         info = xen_irq_init(irq);
 773         if (!info)
 774                 xen_irq_free_desc(irq);
 775
 776         return info;
 777 }
 778
 779 static void xen_free_irq(struct irq_info *info)
 780 {
 781         if (WARN_ON(!info))
 782                 return;
 783
 784         if (!list_empty(&info->eoi_list))
 785                 lateeoi_list_del(info);
 786
 787         list_del(&info->list);
 788
 789         WARN_ON(info->refcnt > 0);
 790
 791         queue_rcu_work(system_wq, &info->rwork);
 792 }
 793
 794 /* Not called for lateeoi events. */
 795 static void event_handler_exit(struct irq_info *info)
 796 {
 797         smp_store_release(&info->is_active, 0);
 798         clear_evtchn(info->evtchn);
 799 }
 800
 801 static void pirq_query_unmask(struct irq_info *info)
 802 {
 803         struct physdev_irq_status_query irq_status;
 804
 805         irq_status.irq = pirq_from_irq(info);
 806         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
 807                 irq_status.flags = 0;
 808
 809         info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
 810         if (irq_status.flags & XENIRQSTAT_needs_eoi)
 811                 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
 812 }
 813
 814 static void do_eoi_pirq(struct irq_info *info)
 815 {
 816         struct physdev_eoi eoi = { .irq = pirq_from_irq(info) };
 817         int rc = 0;
 818
 819         if (!VALID_EVTCHN(info->evtchn))
 820                 return;
 821
 822         event_handler_exit(info);
 823
 824         if (pirq_needs_eoi(info)) {
 825                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
 826                 WARN_ON(rc);
 827         }
 828 }
 829
 830 static void eoi_pirq(struct irq_data *data)
 831 {
 832         struct irq_info *info = info_for_irq(data->irq);
 833
 834         do_eoi_pirq(info);
 835 }
 836
 837 static void do_disable_dynirq(struct irq_info *info)
 838 {
 839         if (VALID_EVTCHN(info->evtchn))
 840                 do_mask(info, EVT_MASK_REASON_EXPLICIT);
 841 }
 842
 843 static void disable_dynirq(struct irq_data *data)
 844 {
 845         struct irq_info *info = info_for_irq(data->irq);
 846
 847         if (info)
 848                 do_disable_dynirq(info);
 849 }
 850
 851 static void mask_ack_pirq(struct irq_data *data)
 852 {
 853         struct irq_info *info = info_for_irq(data->irq);
 854
 855         if (info) {
 856                 do_disable_dynirq(info);
 857                 do_eoi_pirq(info);
 858         }
 859 }
 860
 861 static unsigned int __startup_pirq(struct irq_info *info)
 862 {
 863         struct evtchn_bind_pirq bind_pirq;
 864         evtchn_port_t evtchn = info->evtchn;
 865         int rc;
 866
 867         if (VALID_EVTCHN(evtchn))
 868                 goto out;
 869
 870         bind_pirq.pirq = pirq_from_irq(info);
 871         /* NB. We are happy to share unless we are probing. */
 872         bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
 873                                         BIND_PIRQ__WILL_SHARE : 0;
 874         rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
 875         if (rc != 0) {
 876                 pr_warn("Failed to obtain physical IRQ %d\n", info->irq);
 877                 return 0;
 878         }
 879         evtchn = bind_pirq.port;
 880
 881         pirq_query_unmask(info);
 882
 883         rc = set_evtchn_to_irq(evtchn, info->irq);
 884         if (rc)
 885                 goto err;
 886
 887         info->evtchn = evtchn;
 888         bind_evtchn_to_cpu(info, 0, false);
 889
 890         rc = xen_evtchn_port_setup(evtchn);
 891         if (rc)
 892                 goto err;
 893
 894 out:
 895         do_unmask(info, EVT_MASK_REASON_EXPLICIT);
 896
 897         do_eoi_pirq(info);
 898
 899         return 0;
 900
 901 err:
 902         pr_err("irq%d: Failed to set port to irq mapping (%d)\n", info->irq,
 903                rc);
 904         xen_evtchn_close(evtchn);
 905         return 0;
 906 }
 907
 908 static unsigned int startup_pirq(struct irq_data *data)
 909 {
 910         struct irq_info *info = info_for_irq(data->irq);
 911
 912         return __startup_pirq(info);
 913 }
 914
 915 static void shutdown_pirq(struct irq_data *data)
 916 {
 917         struct irq_info *info = info_for_irq(data->irq);
 918         evtchn_port_t evtchn = info->evtchn;
 919
 920         BUG_ON(info->type != IRQT_PIRQ);
 921
 922         if (!VALID_EVTCHN(evtchn))
 923                 return;
 924
 925         do_mask(info, EVT_MASK_REASON_EXPLICIT);
 926         xen_irq_info_cleanup(info);
 927         xen_evtchn_close(evtchn);
 928 }
 929
 930 static void enable_pirq(struct irq_data *data)
 931 {
 932         enable_dynirq(data);
 933 }
 934
 935 static void disable_pirq(struct irq_data *data)
 936 {
 937         disable_dynirq(data);
 938 }
 939
 940 int xen_irq_from_gsi(unsigned gsi)
 941 {
 942         struct irq_info *info;
 943
 944         list_for_each_entry(info, &xen_irq_list_head, list) {
 945                 if (info->type != IRQT_PIRQ)
 946                         continue;
 947
 948                 if (info->u.pirq.gsi == gsi)
 949                         return info->irq;
 950         }
 951
 952         return -1;
 953 }
 954 EXPORT_SYMBOL_GPL(xen_irq_from_gsi);
 955
 956 static void __unbind_from_irq(struct irq_info *info, unsigned int irq)
 957 {
 958         evtchn_port_t evtchn;
 959         bool close_evtchn = false;
 960
 961         if (!info) {
 962                 xen_irq_free_desc(irq);
 963                 return;
 964         }
 965
 966         if (info->refcnt > 0) {
 967                 info->refcnt--;
 968                 if (info->refcnt != 0)
 969                         return;
 970         }
 971
 972         evtchn = info->evtchn;
 973
 974         if (VALID_EVTCHN(evtchn)) {
 975                 unsigned int cpu = info->cpu;
 976                 struct xenbus_device *dev;
 977
 978                 if (!info->is_static)
 979                         close_evtchn = true;
 980
 981                 switch (info->type) {
 982                 case IRQT_VIRQ:
 983                         per_cpu(virq_to_irq, cpu)[virq_from_irq(info)] = -1;
 984                         break;
 985                 case IRQT_IPI:
 986                         per_cpu(ipi_to_irq, cpu)[ipi_from_irq(info)] = -1;
 987                         per_cpu(ipi_to_evtchn, cpu)[ipi_from_irq(info)] = 0;
 988                         break;
 989                 case IRQT_EVTCHN:
 990                         dev = info->u.interdomain;
 991                         if (dev)
 992                                 atomic_dec(&dev->event_channels);
 993                         break;
 994                 default:
 995                         break;
 996                 }
 997
 998                 xen_irq_info_cleanup(info);
 999
1000                 if (close_evtchn)
1001                         xen_evtchn_close(evtchn);
1002         }
1003
1004         xen_free_irq(info);
1005 }
1006
1007 /*
1008  * Do not make any assumptions regarding the relationship between the
1009  * IRQ number returned here and the Xen pirq argument.
1010  *
1011  * Note: We don't assign an event channel until the irq actually started
1012  * up.  Return an existing irq if we've already got one for the gsi.
1013  *
1014  * Shareable implies level triggered, not shareable implies edge
1015  * triggered here.
1016  */
1017 int xen_bind_pirq_gsi_to_irq(unsigned gsi,
1018                              unsigned pirq, int shareable, char *name)
1019 {
1020         struct irq_info *info;
1021         struct physdev_irq irq_op;
1022         int ret;
1023
1024         mutex_lock(&irq_mapping_update_lock);
1025
1026         ret = xen_irq_from_gsi(gsi);
1027         if (ret != -1) {
1028                 pr_info("%s: returning irq %d for gsi %u\n",
1029                         __func__, ret, gsi);
1030                 goto out;
1031         }
1032
1033         info = xen_allocate_irq_gsi(gsi);
1034         if (!info)
1035                 goto out;
1036
1037         irq_op.irq = info->irq;
1038         irq_op.vector = 0;
1039
1040         /* Only the privileged domain can do this. For non-priv, the pcifront
1041          * driver provides a PCI bus that does the call to do exactly
1042          * this in the priv domain. */
1043         if (xen_initial_domain() &&
1044             HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
1045                 xen_free_irq(info);
1046                 ret = -ENOSPC;
1047                 goto out;
1048         }
1049
1050         ret = xen_irq_info_pirq_setup(info, 0, pirq, gsi, DOMID_SELF,
1051                                shareable ? PIRQ_SHAREABLE : 0);
1052         if (ret < 0) {
1053                 __unbind_from_irq(info, info->irq);
1054                 goto out;
1055         }
1056
1057         pirq_query_unmask(info);
1058         /* We try to use the handler with the appropriate semantic for the
1059          * type of interrupt: if the interrupt is an edge triggered
1060          * interrupt we use handle_edge_irq.
1061          *
1062          * On the other hand if the interrupt is level triggered we use
1063          * handle_fasteoi_irq like the native code does for this kind of
1064          * interrupts.
1065          *
1066          * Depending on the Xen version, pirq_needs_eoi might return true
1067          * not only for level triggered interrupts but for edge triggered
1068          * interrupts too. In any case Xen always honors the eoi mechanism,
1069          * not injecting any more pirqs of the same kind if the first one
1070          * hasn't received an eoi yet. Therefore using the fasteoi handler
1071          * is the right choice either way.
1072          */
1073         if (shareable)
1074                 irq_set_chip_and_handler_name(info->irq, &xen_pirq_chip,
1075                                 handle_fasteoi_irq, name);
1076         else
1077                 irq_set_chip_and_handler_name(info->irq, &xen_pirq_chip,
1078                                 handle_edge_irq, name);
1079
1080         ret = info->irq;
1081
1082 out:
1083         mutex_unlock(&irq_mapping_update_lock);
1084
1085         return ret;
1086 }
1087
1088 #ifdef CONFIG_PCI_MSI
1089 int xen_allocate_pirq_msi(struct pci_dev *dev, struct msi_desc *msidesc)
1090 {
1091         int rc;
1092         struct physdev_get_free_pirq op_get_free_pirq;
1093
1094         op_get_free_pirq.type = MAP_PIRQ_TYPE_MSI;
1095         rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
1096
1097         WARN_ONCE(rc == -ENOSYS,
1098                   "hypervisor does not support the PHYSDEVOP_get_free_pirq interface\n");
1099
1100         return rc ? -1 : op_get_free_pirq.pirq;
1101 }
1102
1103 int xen_bind_pirq_msi_to_irq(struct pci_dev *dev, struct msi_desc *msidesc,
1104                              int pirq, int nvec, const char *name, domid_t domid)
1105 {
1106         int i, irq, ret;
1107         struct irq_info *info;
1108
1109         mutex_lock(&irq_mapping_update_lock);
1110
1111         irq = irq_alloc_descs(-1, 0, nvec, -1);
1112         if (irq < 0)
1113                 goto out;
1114
1115         for (i = 0; i < nvec; i++) {
1116                 info = xen_irq_init(irq + i);
1117                 if (!info) {
1118                         ret = -ENOMEM;
1119                         goto error_irq;
1120                 }
1121
1122                 irq_set_chip_and_handler_name(irq + i, &xen_pirq_chip, handle_edge_irq, name);
1123
1124                 ret = xen_irq_info_pirq_setup(info, 0, pirq + i, 0, domid,
1125                                               i == 0 ? 0 : PIRQ_MSI_GROUP);
1126                 if (ret < 0)
1127                         goto error_irq;
1128         }
1129
1130         ret = irq_set_msi_desc(irq, msidesc);
1131         if (ret < 0)
1132                 goto error_irq;
1133 out:
1134         mutex_unlock(&irq_mapping_update_lock);
1135         return irq;
1136
1137 error_irq:
1138         while (nvec--) {
1139                 info = info_for_irq(irq + nvec);
1140                 __unbind_from_irq(info, irq + nvec);
1141         }
1142         mutex_unlock(&irq_mapping_update_lock);
1143         return ret;
1144 }
1145 #endif
1146
1147 int xen_destroy_irq(int irq)
1148 {
1149         struct physdev_unmap_pirq unmap_irq;
1150         struct irq_info *info = info_for_irq(irq);
1151         int rc = -ENOENT;
1152
1153         mutex_lock(&irq_mapping_update_lock);
1154
1155         /*
1156          * If trying to remove a vector in a MSI group different
1157          * than the first one skip the PIRQ unmap unless this vector
1158          * is the first one in the group.
1159          */
1160         if (xen_initial_domain() && !(info->u.pirq.flags & PIRQ_MSI_GROUP)) {
1161                 unmap_irq.pirq = info->u.pirq.pirq;
1162                 unmap_irq.domid = info->u.pirq.domid;
1163                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
1164                 /* If another domain quits without making the pci_disable_msix
1165                  * call, the Xen hypervisor takes care of freeing the PIRQs
1166                  * (free_domain_pirqs).
1167                  */
1168                 if ((rc == -ESRCH && info->u.pirq.domid != DOMID_SELF))
1169                         pr_info("domain %d does not have %d anymore\n",
1170                                 info->u.pirq.domid, info->u.pirq.pirq);
1171                 else if (rc) {
1172                         pr_warn("unmap irq failed %d\n", rc);
1173                         goto out;
1174                 }
1175         }
1176
1177         xen_free_irq(info);
1178
1179 out:
1180         mutex_unlock(&irq_mapping_update_lock);
1181         return rc;
1182 }
1183
1184 int xen_pirq_from_irq(unsigned irq)
1185 {
1186         struct irq_info *info = info_for_irq(irq);
1187
1188         return pirq_from_irq(info);
1189 }
1190 EXPORT_SYMBOL_GPL(xen_pirq_from_irq);
1191
1192 static int bind_evtchn_to_irq_chip(evtchn_port_t evtchn, struct irq_chip *chip,
1193                                    struct xenbus_device *dev, bool shared)
1194 {
1195         int ret = -ENOMEM;
1196         struct irq_info *info;
1197
1198         if (evtchn >= xen_evtchn_max_channels())
1199                 return -ENOMEM;
1200
1201         mutex_lock(&irq_mapping_update_lock);
1202
1203         info = evtchn_to_info(evtchn);
1204
1205         if (!info) {
1206                 info = xen_allocate_irq_dynamic();
1207                 if (!info)
1208                         goto out;
1209
1210                 irq_set_chip_and_handler_name(info->irq, chip,
1211                                               handle_edge_irq, "event");
1212
1213                 ret = xen_irq_info_evtchn_setup(info, evtchn, dev);
1214                 if (ret < 0) {
1215                         __unbind_from_irq(info, info->irq);
1216                         goto out;
1217                 }
1218                 /*
1219                  * New interdomain events are initially bound to vCPU0 This
1220                  * is required to setup the event channel in the first
1221                  * place and also important for UP guests because the
1222                  * affinity setting is not invoked on them so nothing would
1223                  * bind the channel.
1224                  */
1225                 bind_evtchn_to_cpu(info, 0, false);
1226         } else if (!WARN_ON(info->type != IRQT_EVTCHN)) {
1227                 if (shared && !WARN_ON(info->refcnt < 0))
1228                         info->refcnt++;
1229         }
1230
1231         ret = info->irq;
1232
1233 out:
1234         mutex_unlock(&irq_mapping_update_lock);
1235
1236         return ret;
1237 }
1238
1239 int bind_evtchn_to_irq(evtchn_port_t evtchn)
1240 {
1241         return bind_evtchn_to_irq_chip(evtchn, &xen_dynamic_chip, NULL, false);
1242 }
1243 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
1244
1245 int bind_evtchn_to_irq_lateeoi(evtchn_port_t evtchn)
1246 {
1247         return bind_evtchn_to_irq_chip(evtchn, &xen_lateeoi_chip, NULL, false);
1248 }
1249 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq_lateeoi);
1250
1251 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
1252 {
1253         struct evtchn_bind_ipi bind_ipi;
1254         evtchn_port_t evtchn;
1255         struct irq_info *info;
1256         int ret;
1257
1258         mutex_lock(&irq_mapping_update_lock);
1259
1260         ret = per_cpu(ipi_to_irq, cpu)[ipi];
1261
1262         if (ret == -1) {
1263                 info = xen_allocate_irq_dynamic();
1264                 if (!info)
1265                         goto out;
1266
1267                 irq_set_chip_and_handler_name(info->irq, &xen_percpu_chip,
1268                                               handle_percpu_irq, "ipi");
1269
1270                 bind_ipi.vcpu = xen_vcpu_nr(cpu);
1271                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1272                                                 &bind_ipi) != 0)
1273                         BUG();
1274                 evtchn = bind_ipi.port;
1275
1276                 ret = xen_irq_info_ipi_setup(info, cpu, evtchn, ipi);
1277                 if (ret < 0) {
1278                         __unbind_from_irq(info, info->irq);
1279                         goto out;
1280                 }
1281                 /*
1282                  * Force the affinity mask to the target CPU so proc shows
1283                  * the correct target.
1284                  */
1285                 bind_evtchn_to_cpu(info, cpu, true);
1286                 ret = info->irq;
1287         } else {
1288                 info = info_for_irq(ret);
1289                 WARN_ON(info == NULL || info->type != IRQT_IPI);
1290         }
1291
1292  out:
1293         mutex_unlock(&irq_mapping_update_lock);
1294         return ret;
1295 }
1296
1297 static int bind_interdomain_evtchn_to_irq_chip(struct xenbus_device *dev,
1298                                                evtchn_port_t remote_port,
1299                                                struct irq_chip *chip,
1300                                                bool shared)
1301 {
1302         struct evtchn_bind_interdomain bind_interdomain;
1303         int err;
1304
1305         bind_interdomain.remote_dom  = dev->otherend_id;
1306         bind_interdomain.remote_port = remote_port;
1307
1308         err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
1309                                           &bind_interdomain);
1310
1311         return err ? : bind_evtchn_to_irq_chip(bind_interdomain.local_port,
1312                                                chip, dev, shared);
1313 }
1314
1315 int bind_interdomain_evtchn_to_irq_lateeoi(struct xenbus_device *dev,
1316                                            evtchn_port_t remote_port)
1317 {
1318         return bind_interdomain_evtchn_to_irq_chip(dev, remote_port,
1319                                                    &xen_lateeoi_chip, false);
1320 }
1321 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irq_lateeoi);
1322
1323 static int find_virq(unsigned int virq, unsigned int cpu, evtchn_port_t *evtchn)
1324 {
1325         struct evtchn_status status;
1326         evtchn_port_t port;
1327         int rc = -ENOENT;
1328
1329         memset(&status, 0, sizeof(status));
1330         for (port = 0; port < xen_evtchn_max_channels(); port++) {
1331                 status.dom = DOMID_SELF;
1332                 status.port = port;
1333                 rc = HYPERVISOR_event_channel_op(EVTCHNOP_status, &status);
1334                 if (rc < 0)
1335                         continue;
1336                 if (status.status != EVTCHNSTAT_virq)
1337                         continue;
1338                 if (status.u.virq == virq && status.vcpu == xen_vcpu_nr(cpu)) {
1339                         *evtchn = port;
1340                         break;
1341                 }
1342         }
1343         return rc;
1344 }
1345
1346 /**
1347  * xen_evtchn_nr_channels - number of usable event channel ports
1348  *
1349  * This may be less than the maximum supported by the current
1350  * hypervisor ABI. Use xen_evtchn_max_channels() for the maximum
1351  * supported.
1352  */
1353 unsigned xen_evtchn_nr_channels(void)
1354 {
1355         return evtchn_ops->nr_channels();
1356 }
1357 EXPORT_SYMBOL_GPL(xen_evtchn_nr_channels);
1358
1359 int bind_virq_to_irq(unsigned int virq, unsigned int cpu, bool percpu)
1360 {
1361         struct evtchn_bind_virq bind_virq;
1362         evtchn_port_t evtchn = 0;
1363         struct irq_info *info;
1364         int ret;
1365
1366         mutex_lock(&irq_mapping_update_lock);
1367
1368         ret = per_cpu(virq_to_irq, cpu)[virq];
1369
1370         if (ret == -1) {
1371                 info = xen_allocate_irq_dynamic();
1372                 if (!info)
1373                         goto out;
1374
1375                 if (percpu)
1376                         irq_set_chip_and_handler_name(info->irq, &xen_percpu_chip,
1377                                                       handle_percpu_irq, "virq");
1378                 else
1379                         irq_set_chip_and_handler_name(info->irq, &xen_dynamic_chip,
1380                                                       handle_edge_irq, "virq");
1381
1382                 bind_virq.virq = virq;
1383                 bind_virq.vcpu = xen_vcpu_nr(cpu);
1384                 ret = HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1385                                                 &bind_virq);
1386                 if (ret == 0)
1387                         evtchn = bind_virq.port;
1388                 else {
1389                         if (ret == -EEXIST)
1390                                 ret = find_virq(virq, cpu, &evtchn);
1391                         BUG_ON(ret < 0);
1392                 }
1393
1394                 ret = xen_irq_info_virq_setup(info, cpu, evtchn, virq);
1395                 if (ret < 0) {
1396                         __unbind_from_irq(info, info->irq);
1397                         goto out;
1398                 }
1399
1400                 /*
1401                  * Force the affinity mask for percpu interrupts so proc
1402                  * shows the correct target.
1403                  */
1404                 bind_evtchn_to_cpu(info, cpu, percpu);
1405                 ret = info->irq;
1406         } else {
1407                 info = info_for_irq(ret);
1408                 WARN_ON(info == NULL || info->type != IRQT_VIRQ);
1409         }
1410
1411 out:
1412         mutex_unlock(&irq_mapping_update_lock);
1413
1414         return ret;
1415 }
1416
1417 static void unbind_from_irq(unsigned int irq)
1418 {
1419         struct irq_info *info;
1420
1421         mutex_lock(&irq_mapping_update_lock);
1422         info = info_for_irq(irq);
1423         __unbind_from_irq(info, irq);
1424         mutex_unlock(&irq_mapping_update_lock);
1425 }
1426
1427 static int bind_evtchn_to_irqhandler_chip(evtchn_port_t evtchn,
1428                                           irq_handler_t handler,
1429                                           unsigned long irqflags,
1430                                           const char *devname, void *dev_id,
1431                                           struct irq_chip *chip)
1432 {
1433         int irq, retval;
1434
1435         irq = bind_evtchn_to_irq_chip(evtchn, chip, NULL,
1436                                       irqflags & IRQF_SHARED);
1437         if (irq < 0)
1438                 return irq;
1439         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1440         if (retval != 0) {
1441                 unbind_from_irq(irq);
1442                 return retval;
1443         }
1444
1445         return irq;
1446 }
1447
1448 int bind_evtchn_to_irqhandler(evtchn_port_t evtchn,
1449                               irq_handler_t handler,
1450                               unsigned long irqflags,
1451                               const char *devname, void *dev_id)
1452 {
1453         return bind_evtchn_to_irqhandler_chip(evtchn, handler, irqflags,
1454                                               devname, dev_id,
1455                                               &xen_dynamic_chip);
1456 }
1457 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
1458
1459 int bind_evtchn_to_irqhandler_lateeoi(evtchn_port_t evtchn,
1460                                       irq_handler_t handler,
1461                                       unsigned long irqflags,
1462                                       const char *devname, void *dev_id)
1463 {
1464         return bind_evtchn_to_irqhandler_chip(evtchn, handler, irqflags,
1465                                               devname, dev_id,
1466                                               &xen_lateeoi_chip);
1467 }
1468 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler_lateeoi);
1469
1470 static int bind_interdomain_evtchn_to_irqhandler_chip(
1471                 struct xenbus_device *dev, evtchn_port_t remote_port,
1472                 irq_handler_t handler, unsigned long irqflags,
1473                 const char *devname, void *dev_id, struct irq_chip *chip)
1474 {
1475         int irq, retval;
1476
1477         irq = bind_interdomain_evtchn_to_irq_chip(dev, remote_port, chip,
1478                                                   irqflags & IRQF_SHARED);
1479         if (irq < 0)
1480                 return irq;
1481
1482         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1483         if (retval != 0) {
1484                 unbind_from_irq(irq);
1485                 return retval;
1486         }
1487
1488         return irq;
1489 }
1490
1491 int bind_interdomain_evtchn_to_irqhandler_lateeoi(struct xenbus_device *dev,
1492                                                   evtchn_port_t remote_port,
1493                                                   irq_handler_t handler,
1494                                                   unsigned long irqflags,
1495                                                   const char *devname,
1496                                                   void *dev_id)
1497 {
1498         return bind_interdomain_evtchn_to_irqhandler_chip(dev,
1499                                 remote_port, handler, irqflags, devname,
1500                                 dev_id, &xen_lateeoi_chip);
1501 }
1502 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler_lateeoi);
1503
1504 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
1505                             irq_handler_t handler,
1506                             unsigned long irqflags, const char *devname, void *dev_id)
1507 {
1508         int irq, retval;
1509
1510         irq = bind_virq_to_irq(virq, cpu, irqflags & IRQF_PERCPU);
1511         if (irq < 0)
1512                 return irq;
1513         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1514         if (retval != 0) {
1515                 unbind_from_irq(irq);
1516                 return retval;
1517         }
1518
1519         return irq;
1520 }
1521 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1522
1523 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1524                            unsigned int cpu,
1525                            irq_handler_t handler,
1526                            unsigned long irqflags,
1527                            const char *devname,
1528                            void *dev_id)
1529 {
1530         int irq, retval;
1531
1532         irq = bind_ipi_to_irq(ipi, cpu);
1533         if (irq < 0)
1534                 return irq;
1535
1536         irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME | IRQF_EARLY_RESUME;
1537         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1538         if (retval != 0) {
1539                 unbind_from_irq(irq);
1540                 return retval;
1541         }
1542
1543         return irq;
1544 }
1545
1546 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1547 {
1548         struct irq_info *info = info_for_irq(irq);
1549
1550         if (WARN_ON(!info))
1551                 return;
1552         free_irq(irq, dev_id);
1553         unbind_from_irq(irq);
1554 }
1555 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1556
1557 /**
1558  * xen_set_irq_priority() - set an event channel priority.
1559  * @irq:irq bound to an event channel.
1560  * @priority: priority between XEN_IRQ_PRIORITY_MAX and XEN_IRQ_PRIORITY_MIN.
1561  */
1562 int xen_set_irq_priority(unsigned irq, unsigned priority)
1563 {
1564         struct evtchn_set_priority set_priority;
1565
1566         set_priority.port = evtchn_from_irq(irq);
1567         set_priority.priority = priority;
1568
1569         return HYPERVISOR_event_channel_op(EVTCHNOP_set_priority,
1570                                            &set_priority);
1571 }
1572 EXPORT_SYMBOL_GPL(xen_set_irq_priority);
1573
1574 int evtchn_make_refcounted(evtchn_port_t evtchn, bool is_static)
1575 {
1576         struct irq_info *info = evtchn_to_info(evtchn);
1577
1578         if (!info)
1579                 return -ENOENT;
1580
1581         WARN_ON(info->refcnt != -1);
1582
1583         info->refcnt = 1;
1584         info->is_static = is_static;
1585
1586         return 0;
1587 }
1588 EXPORT_SYMBOL_GPL(evtchn_make_refcounted);
1589
1590 int evtchn_get(evtchn_port_t evtchn)
1591 {
1592         struct irq_info *info;
1593         int err = -ENOENT;
1594
1595         if (evtchn >= xen_evtchn_max_channels())
1596                 return -EINVAL;
1597
1598         mutex_lock(&irq_mapping_update_lock);
1599
1600         info = evtchn_to_info(evtchn);
1601
1602         if (!info)
1603                 goto done;
1604
1605         err = -EINVAL;
1606         if (info->refcnt <= 0 || info->refcnt == SHRT_MAX)
1607                 goto done;
1608
1609         info->refcnt++;
1610         err = 0;
1611  done:
1612         mutex_unlock(&irq_mapping_update_lock);
1613
1614         return err;
1615 }
1616 EXPORT_SYMBOL_GPL(evtchn_get);
1617
1618 void evtchn_put(evtchn_port_t evtchn)
1619 {
1620         struct irq_info *info = evtchn_to_info(evtchn);
1621
1622         if (WARN_ON(!info))
1623                 return;
1624         unbind_from_irq(info->irq);
1625 }
1626 EXPORT_SYMBOL_GPL(evtchn_put);
1627
1628 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1629 {
1630         evtchn_port_t evtchn;
1631
1632 #ifdef CONFIG_X86
1633         if (unlikely(vector == XEN_NMI_VECTOR)) {
1634                 int rc =  HYPERVISOR_vcpu_op(VCPUOP_send_nmi, xen_vcpu_nr(cpu),
1635                                              NULL);
1636                 if (rc < 0)
1637                         printk(KERN_WARNING "Sending nmi to CPU%d failed (rc:%d)\n", cpu, rc);
1638                 return;
1639         }
1640 #endif
1641         evtchn = per_cpu(ipi_to_evtchn, cpu)[vector];
1642         BUG_ON(evtchn == 0);
1643         notify_remote_via_evtchn(evtchn);
1644 }
1645
1646 struct evtchn_loop_ctrl {
1647         ktime_t timeout;
1648         unsigned count;
1649         bool defer_eoi;
1650 };
1651
1652 void handle_irq_for_port(evtchn_port_t port, struct evtchn_loop_ctrl *ctrl)
1653 {
1654         struct irq_info *info = evtchn_to_info(port);
1655         struct xenbus_device *dev;
1656
1657         if (!info)
1658                 return;
1659
1660         /*
1661          * Check for timeout every 256 events.
1662          * We are setting the timeout value only after the first 256
1663          * events in order to not hurt the common case of few loop
1664          * iterations. The 256 is basically an arbitrary value.
1665          *
1666          * In case we are hitting the timeout we need to defer all further
1667          * EOIs in order to ensure to leave the event handling loop rather
1668          * sooner than later.
1669          */
1670         if (!ctrl->defer_eoi && !(++ctrl->count & 0xff)) {
1671                 ktime_t kt = ktime_get();
1672
1673                 if (!ctrl->timeout) {
1674                         kt = ktime_add_ms(kt,
1675                                           jiffies_to_msecs(event_loop_timeout));
1676                         ctrl->timeout = kt;
1677                 } else if (kt > ctrl->timeout) {
1678                         ctrl->defer_eoi = true;
1679                 }
1680         }
1681
1682         if (xchg_acquire(&info->is_active, 1))
1683                 return;
1684
1685         dev = (info->type == IRQT_EVTCHN) ? info->u.interdomain : NULL;
1686         if (dev)
1687                 atomic_inc(&dev->events);
1688
1689         if (ctrl->defer_eoi) {
1690                 info->eoi_cpu = smp_processor_id();
1691                 info->irq_epoch = __this_cpu_read(irq_epoch);
1692                 info->eoi_time = get_jiffies_64() + event_eoi_delay;
1693         }
1694
1695         generic_handle_irq(info->irq);
1696 }
1697
1698 int xen_evtchn_do_upcall(void)
1699 {
1700         struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1701         int ret = vcpu_info->evtchn_upcall_pending ? IRQ_HANDLED : IRQ_NONE;
1702         int cpu = smp_processor_id();
1703         struct evtchn_loop_ctrl ctrl = { 0 };
1704
1705         /*
1706          * When closing an event channel the associated IRQ must not be freed
1707          * until all cpus have left the event handling loop. This is ensured
1708          * by taking the rcu_read_lock() while handling events, as freeing of
1709          * the IRQ is handled via queue_rcu_work() _after_ closing the event
1710          * channel.
1711          */
1712         rcu_read_lock();
1713
1714         do {
1715                 vcpu_info->evtchn_upcall_pending = 0;
1716
1717                 xen_evtchn_handle_events(cpu, &ctrl);
1718
1719                 BUG_ON(!irqs_disabled());
1720
1721                 virt_rmb(); /* Hypervisor can set upcall pending. */
1722
1723         } while (vcpu_info->evtchn_upcall_pending);
1724
1725         rcu_read_unlock();
1726
1727         /*
1728          * Increment irq_epoch only now to defer EOIs only for
1729          * xen_irq_lateeoi() invocations occurring from inside the loop
1730          * above.
1731          */
1732         __this_cpu_inc(irq_epoch);
1733
1734         return ret;
1735 }
1736 EXPORT_SYMBOL_GPL(xen_evtchn_do_upcall);
1737
1738 /* Rebind a new event channel to an existing irq. */
1739 void rebind_evtchn_irq(evtchn_port_t evtchn, int irq)
1740 {
1741         struct irq_info *info = info_for_irq(irq);
1742
1743         if (WARN_ON(!info))
1744                 return;
1745
1746         /* Make sure the irq is masked, since the new event channel
1747            will also be masked. */
1748         disable_irq(irq);
1749
1750         mutex_lock(&irq_mapping_update_lock);
1751
1752         /* After resume the irq<->evtchn mappings are all cleared out */
1753         BUG_ON(evtchn_to_info(evtchn));
1754         /* Expect irq to have been bound before,
1755            so there should be a proper type */
1756         BUG_ON(info->type == IRQT_UNBOUND);
1757
1758         info->irq = irq;
1759         (void)xen_irq_info_evtchn_setup(info, evtchn, NULL);
1760
1761         mutex_unlock(&irq_mapping_update_lock);
1762
1763         bind_evtchn_to_cpu(info, info->cpu, false);
1764
1765         /* Unmask the event channel. */
1766         enable_irq(irq);
1767 }
1768
1769 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1770 static int xen_rebind_evtchn_to_cpu(struct irq_info *info, unsigned int tcpu)
1771 {
1772         struct evtchn_bind_vcpu bind_vcpu;
1773         evtchn_port_t evtchn = info ? info->evtchn : 0;
1774
1775         if (!VALID_EVTCHN(evtchn))
1776                 return -1;
1777
1778         if (!xen_support_evtchn_rebind())
1779                 return -1;
1780
1781         /* Send future instances of this interrupt to other vcpu. */
1782         bind_vcpu.port = evtchn;
1783         bind_vcpu.vcpu = xen_vcpu_nr(tcpu);
1784
1785         /*
1786          * Mask the event while changing the VCPU binding to prevent
1787          * it being delivered on an unexpected VCPU.
1788          */
1789         do_mask(info, EVT_MASK_REASON_TEMPORARY);
1790
1791         /*
1792          * If this fails, it usually just indicates that we're dealing with a
1793          * virq or IPI channel, which don't actually need to be rebound. Ignore
1794          * it, but don't do the xenlinux-level rebind in that case.
1795          */
1796         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1797                 bind_evtchn_to_cpu(info, tcpu, false);
1798
1799         do_unmask(info, EVT_MASK_REASON_TEMPORARY);
1800
1801         return 0;
1802 }
1803
1804 /*
1805  * Find the CPU within @dest mask which has the least number of channels
1806  * assigned. This is not precise as the per cpu counts can be modified
1807  * concurrently.
1808  */
1809 static unsigned int select_target_cpu(const struct cpumask *dest)
1810 {
1811         unsigned int cpu, best_cpu = UINT_MAX, minch = UINT_MAX;
1812
1813         for_each_cpu_and(cpu, dest, cpu_online_mask) {
1814                 unsigned int curch = atomic_read(&channels_on_cpu[cpu]);
1815
1816                 if (curch < minch) {
1817                         minch = curch;
1818                         best_cpu = cpu;
1819                 }
1820         }
1821
1822         /*
1823          * Catch the unlikely case that dest contains no online CPUs. Can't
1824          * recurse.
1825          */
1826         if (best_cpu == UINT_MAX)
1827                 return select_target_cpu(cpu_online_mask);
1828
1829         return best_cpu;
1830 }
1831
1832 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1833                             bool force)
1834 {
1835         unsigned int tcpu = select_target_cpu(dest);
1836         int ret;
1837
1838         ret = xen_rebind_evtchn_to_cpu(info_for_irq(data->irq), tcpu);
1839         if (!ret)
1840                 irq_data_update_effective_affinity(data, cpumask_of(tcpu));
1841
1842         return ret;
1843 }
1844
1845 static void enable_dynirq(struct irq_data *data)
1846 {
1847         struct irq_info *info = info_for_irq(data->irq);
1848         evtchn_port_t evtchn = info ? info->evtchn : 0;
1849
1850         if (VALID_EVTCHN(evtchn))
1851                 do_unmask(info, EVT_MASK_REASON_EXPLICIT);
1852 }
1853
1854 static void do_ack_dynirq(struct irq_info *info)
1855 {
1856         evtchn_port_t evtchn = info->evtchn;
1857
1858         if (VALID_EVTCHN(evtchn))
1859                 event_handler_exit(info);
1860 }
1861
1862 static void ack_dynirq(struct irq_data *data)
1863 {
1864         struct irq_info *info = info_for_irq(data->irq);
1865
1866         if (info)
1867                 do_ack_dynirq(info);
1868 }
1869
1870 static void mask_ack_dynirq(struct irq_data *data)
1871 {
1872         struct irq_info *info = info_for_irq(data->irq);
1873
1874         if (info) {
1875                 do_disable_dynirq(info);
1876                 do_ack_dynirq(info);
1877         }
1878 }
1879
1880 static void lateeoi_ack_dynirq(struct irq_data *data)
1881 {
1882         struct irq_info *info = info_for_irq(data->irq);
1883         evtchn_port_t evtchn = info ? info->evtchn : 0;
1884
1885         if (VALID_EVTCHN(evtchn)) {
1886                 do_mask(info, EVT_MASK_REASON_EOI_PENDING);
1887                 /*
1888                  * Don't call event_handler_exit().
1889                  * Need to keep is_active non-zero in order to ignore re-raised
1890                  * events after cpu affinity changes while a lateeoi is pending.
1891                  */
1892                 clear_evtchn(evtchn);
1893         }
1894 }
1895
1896 static void lateeoi_mask_ack_dynirq(struct irq_data *data)
1897 {
1898         struct irq_info *info = info_for_irq(data->irq);
1899         evtchn_port_t evtchn = info ? info->evtchn : 0;
1900
1901         if (VALID_EVTCHN(evtchn)) {
1902                 do_mask(info, EVT_MASK_REASON_EXPLICIT);
1903                 event_handler_exit(info);
1904         }
1905 }
1906
1907 static int retrigger_dynirq(struct irq_data *data)
1908 {
1909         struct irq_info *info = info_for_irq(data->irq);
1910         evtchn_port_t evtchn = info ? info->evtchn : 0;
1911
1912         if (!VALID_EVTCHN(evtchn))
1913                 return 0;
1914
1915         do_mask(info, EVT_MASK_REASON_TEMPORARY);
1916         set_evtchn(evtchn);
1917         do_unmask(info, EVT_MASK_REASON_TEMPORARY);
1918
1919         return 1;
1920 }
1921
1922 static void restore_pirqs(void)
1923 {
1924         int pirq, rc, irq, gsi;
1925         struct physdev_map_pirq map_irq;
1926         struct irq_info *info;
1927
1928         list_for_each_entry(info, &xen_irq_list_head, list) {
1929                 if (info->type != IRQT_PIRQ)
1930                         continue;
1931
1932                 pirq = info->u.pirq.pirq;
1933                 gsi = info->u.pirq.gsi;
1934                 irq = info->irq;
1935
1936                 /* save/restore of PT devices doesn't work, so at this point the
1937                  * only devices present are GSI based emulated devices */
1938                 if (!gsi)
1939                         continue;
1940
1941                 map_irq.domid = DOMID_SELF;
1942                 map_irq.type = MAP_PIRQ_TYPE_GSI;
1943                 map_irq.index = gsi;
1944                 map_irq.pirq = pirq;
1945
1946                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1947                 if (rc) {
1948                         pr_warn("xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1949                                 gsi, irq, pirq, rc);
1950                         xen_free_irq(info);
1951                         continue;
1952                 }
1953
1954                 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1955
1956                 __startup_pirq(info);
1957         }
1958 }
1959
1960 static void restore_cpu_virqs(unsigned int cpu)
1961 {
1962         struct evtchn_bind_virq bind_virq;
1963         evtchn_port_t evtchn;
1964         struct irq_info *info;
1965         int virq, irq;
1966
1967         for (virq = 0; virq < NR_VIRQS; virq++) {
1968                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1969                         continue;
1970                 info = info_for_irq(irq);
1971
1972                 BUG_ON(virq_from_irq(info) != virq);
1973
1974                 /* Get a new binding from Xen. */
1975                 bind_virq.virq = virq;
1976                 bind_virq.vcpu = xen_vcpu_nr(cpu);
1977                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1978                                                 &bind_virq) != 0)
1979                         BUG();
1980                 evtchn = bind_virq.port;
1981
1982                 /* Record the new mapping. */
1983                 xen_irq_info_virq_setup(info, cpu, evtchn, virq);
1984                 /* The affinity mask is still valid */
1985                 bind_evtchn_to_cpu(info, cpu, false);
1986         }
1987 }
1988
1989 static void restore_cpu_ipis(unsigned int cpu)
1990 {
1991         struct evtchn_bind_ipi bind_ipi;
1992         evtchn_port_t evtchn;
1993         struct irq_info *info;
1994         int ipi, irq;
1995
1996         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1997                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1998                         continue;
1999                 info = info_for_irq(irq);
2000
2001                 BUG_ON(ipi_from_irq(info) != ipi);
2002
2003                 /* Get a new binding from Xen. */
2004                 bind_ipi.vcpu = xen_vcpu_nr(cpu);
2005                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
2006                                                 &bind_ipi) != 0)
2007                         BUG();
2008                 evtchn = bind_ipi.port;
2009
2010                 /* Record the new mapping. */
2011                 xen_irq_info_ipi_setup(info, cpu, evtchn, ipi);
2012                 /* The affinity mask is still valid */
2013                 bind_evtchn_to_cpu(info, cpu, false);
2014         }
2015 }
2016
2017 /* Clear an irq's pending state, in preparation for polling on it */
2018 void xen_clear_irq_pending(int irq)
2019 {
2020         struct irq_info *info = info_for_irq(irq);
2021         evtchn_port_t evtchn = info ? info->evtchn : 0;
2022
2023         if (VALID_EVTCHN(evtchn))
2024                 event_handler_exit(info);
2025 }
2026 EXPORT_SYMBOL(xen_clear_irq_pending);
2027
2028 bool xen_test_irq_pending(int irq)
2029 {
2030         evtchn_port_t evtchn = evtchn_from_irq(irq);
2031         bool ret = false;
2032
2033         if (VALID_EVTCHN(evtchn))
2034                 ret = test_evtchn(evtchn);
2035
2036         return ret;
2037 }
2038
2039 /* Poll waiting for an irq to become pending with timeout.  In the usual case,
2040  * the irq will be disabled so it won't deliver an interrupt. */
2041 void xen_poll_irq_timeout(int irq, u64 timeout)
2042 {
2043         evtchn_port_t evtchn = evtchn_from_irq(irq);
2044
2045         if (VALID_EVTCHN(evtchn)) {
2046                 struct sched_poll poll;
2047
2048                 poll.nr_ports = 1;
2049                 poll.timeout = timeout;
2050                 set_xen_guest_handle(poll.ports, &evtchn);
2051
2052                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
2053                         BUG();
2054         }
2055 }
2056 EXPORT_SYMBOL(xen_poll_irq_timeout);
2057 /* Poll waiting for an irq to become pending.  In the usual case, the
2058  * irq will be disabled so it won't deliver an interrupt. */
2059 void xen_poll_irq(int irq)
2060 {
2061         xen_poll_irq_timeout(irq, 0 /* no timeout */);
2062 }
2063
2064 /* Check whether the IRQ line is shared with other guests. */
2065 int xen_test_irq_shared(int irq)
2066 {
2067         struct irq_info *info = info_for_irq(irq);
2068         struct physdev_irq_status_query irq_status;
2069
2070         if (WARN_ON(!info))
2071                 return -ENOENT;
2072
2073         irq_status.irq = info->u.pirq.pirq;
2074
2075         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
2076                 return 0;
2077         return !(irq_status.flags & XENIRQSTAT_shared);
2078 }
2079 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
2080
2081 void xen_irq_resume(void)
2082 {
2083         unsigned int cpu;
2084         struct irq_info *info;
2085
2086         /* New event-channel space is not 'live' yet. */
2087         xen_evtchn_resume();
2088
2089         /* No IRQ <-> event-channel mappings. */
2090         list_for_each_entry(info, &xen_irq_list_head, list) {
2091                 /* Zap event-channel binding */
2092                 info->evtchn = 0;
2093                 /* Adjust accounting */
2094                 channels_on_cpu_dec(info);
2095         }
2096
2097         clear_evtchn_to_irq_all();
2098
2099         for_each_possible_cpu(cpu) {
2100                 restore_cpu_virqs(cpu);
2101                 restore_cpu_ipis(cpu);
2102         }
2103
2104         restore_pirqs();
2105 }
2106
2107 static struct irq_chip xen_dynamic_chip __read_mostly = {
2108         .name                   = "xen-dyn",
2109
2110         .irq_disable            = disable_dynirq,
2111         .irq_mask               = disable_dynirq,
2112         .irq_unmask             = enable_dynirq,
2113
2114         .irq_ack                = ack_dynirq,
2115         .irq_mask_ack           = mask_ack_dynirq,
2116
2117         .irq_set_affinity       = set_affinity_irq,
2118         .irq_retrigger          = retrigger_dynirq,
2119 };
2120
2121 static struct irq_chip xen_lateeoi_chip __read_mostly = {
2122         /* The chip name needs to contain "xen-dyn" for irqbalance to work. */
2123         .name                   = "xen-dyn-lateeoi",
2124
2125         .irq_disable            = disable_dynirq,
2126         .irq_mask               = disable_dynirq,
2127         .irq_unmask             = enable_dynirq,
2128
2129         .irq_ack                = lateeoi_ack_dynirq,
2130         .irq_mask_ack           = lateeoi_mask_ack_dynirq,
2131
2132         .irq_set_affinity       = set_affinity_irq,
2133         .irq_retrigger          = retrigger_dynirq,
2134 };
2135
2136 static struct irq_chip xen_pirq_chip __read_mostly = {
2137         .name                   = "xen-pirq",
2138
2139         .irq_startup            = startup_pirq,
2140         .irq_shutdown           = shutdown_pirq,
2141         .irq_enable             = enable_pirq,
2142         .irq_disable            = disable_pirq,
2143
2144         .irq_mask               = disable_dynirq,
2145         .irq_unmask             = enable_dynirq,
2146
2147         .irq_ack                = eoi_pirq,
2148         .irq_eoi                = eoi_pirq,
2149         .irq_mask_ack           = mask_ack_pirq,
2150
2151         .irq_set_affinity       = set_affinity_irq,
2152
2153         .irq_retrigger          = retrigger_dynirq,
2154 };
2155
2156 static struct irq_chip xen_percpu_chip __read_mostly = {
2157         .name                   = "xen-percpu",
2158
2159         .irq_disable            = disable_dynirq,
2160         .irq_mask               = disable_dynirq,
2161         .irq_unmask             = enable_dynirq,
2162
2163         .irq_ack                = ack_dynirq,
2164 };
2165
2166 #ifdef CONFIG_X86
2167 #ifdef CONFIG_XEN_PVHVM
2168 /* Vector callbacks are better than PCI interrupts to receive event
2169  * channel notifications because we can receive vector callbacks on any
2170  * vcpu and we don't need PCI support or APIC interactions. */
2171 void xen_setup_callback_vector(void)
2172 {
2173         uint64_t callback_via;
2174
2175         if (xen_have_vector_callback) {
2176                 callback_via = HVM_CALLBACK_VECTOR(HYPERVISOR_CALLBACK_VECTOR);
2177                 if (xen_set_callback_via(callback_via)) {
2178                         pr_err("Request for Xen HVM callback vector failed\n");
2179                         xen_have_vector_callback = false;
2180                 }
2181         }
2182 }
2183
2184 /*
2185  * Setup per-vCPU vector-type callbacks. If this setup is unavailable,
2186  * fallback to the global vector-type callback.
2187  */
2188 static __init void xen_init_setup_upcall_vector(void)
2189 {
2190         if (!xen_have_vector_callback)
2191                 return;
2192
2193         if ((cpuid_eax(xen_cpuid_base() + 4) & XEN_HVM_CPUID_UPCALL_VECTOR) &&
2194             !xen_set_upcall_vector(0))
2195                 xen_percpu_upcall = true;
2196         else if (xen_feature(XENFEAT_hvm_callback_vector))
2197                 xen_setup_callback_vector();
2198         else
2199                 xen_have_vector_callback = false;
2200 }
2201
2202 int xen_set_upcall_vector(unsigned int cpu)
2203 {
2204         int rc;
2205         xen_hvm_evtchn_upcall_vector_t op = {
2206                 .vector = HYPERVISOR_CALLBACK_VECTOR,
2207                 .vcpu = per_cpu(xen_vcpu_id, cpu),
2208         };
2209
2210         rc = HYPERVISOR_hvm_op(HVMOP_set_evtchn_upcall_vector, &op);
2211         if (rc)
2212                 return rc;
2213
2214         /* Trick toolstack to think we are enlightened. */
2215         if (!cpu)
2216                 rc = xen_set_callback_via(1);
2217
2218         return rc;
2219 }
2220
2221 static __init void xen_alloc_callback_vector(void)
2222 {
2223         if (!xen_have_vector_callback)
2224                 return;
2225
2226         pr_info("Xen HVM callback vector for event delivery is enabled\n");
2227         sysvec_install(HYPERVISOR_CALLBACK_VECTOR, sysvec_xen_hvm_callback);
2228 }
2229 #else
2230 void xen_setup_callback_vector(void) {}
2231 static inline void xen_init_setup_upcall_vector(void) {}
2232 int xen_set_upcall_vector(unsigned int cpu) {}
2233 static inline void xen_alloc_callback_vector(void) {}
2234 #endif /* CONFIG_XEN_PVHVM */
2235 #endif /* CONFIG_X86 */
2236
2237 bool xen_fifo_events = true;
2238 module_param_named(fifo_events, xen_fifo_events, bool, 0);
2239
2240 static int xen_evtchn_cpu_prepare(unsigned int cpu)
2241 {
2242         int ret = 0;
2243
2244         xen_cpu_init_eoi(cpu);
2245
2246         if (evtchn_ops->percpu_init)
2247                 ret = evtchn_ops->percpu_init(cpu);
2248
2249         return ret;
2250 }
2251
2252 static int xen_evtchn_cpu_dead(unsigned int cpu)
2253 {
2254         int ret = 0;
2255
2256         if (evtchn_ops->percpu_deinit)
2257                 ret = evtchn_ops->percpu_deinit(cpu);
2258
2259         return ret;
2260 }
2261
2262 void __init xen_init_IRQ(void)
2263 {
2264         int ret = -EINVAL;
2265         evtchn_port_t evtchn;
2266
2267         if (xen_fifo_events)
2268                 ret = xen_evtchn_fifo_init();
2269         if (ret < 0) {
2270                 xen_evtchn_2l_init();
2271                 xen_fifo_events = false;
2272         }
2273
2274         xen_cpu_init_eoi(smp_processor_id());
2275
2276         cpuhp_setup_state_nocalls(CPUHP_XEN_EVTCHN_PREPARE,
2277                                   "xen/evtchn:prepare",
2278                                   xen_evtchn_cpu_prepare, xen_evtchn_cpu_dead);
2279
2280         evtchn_to_irq = kcalloc(EVTCHN_ROW(xen_evtchn_max_channels()),
2281                                 sizeof(*evtchn_to_irq), GFP_KERNEL);
2282         BUG_ON(!evtchn_to_irq);
2283
2284         /* No event channels are 'live' right now. */
2285         for (evtchn = 0; evtchn < xen_evtchn_nr_channels(); evtchn++)
2286                 mask_evtchn(evtchn);
2287
2288         pirq_needs_eoi = pirq_needs_eoi_flag;
2289
2290 #ifdef CONFIG_X86
2291         if (xen_pv_domain()) {
2292                 if (xen_initial_domain())
2293                         pci_xen_initial_domain();
2294         }
2295         xen_init_setup_upcall_vector();
2296         xen_alloc_callback_vector();
2297
2298
2299         if (xen_hvm_domain()) {
2300                 native_init_IRQ();
2301                 /* pci_xen_hvm_init must be called after native_init_IRQ so that
2302                  * __acpi_register_gsi can point at the right function */
2303                 pci_xen_hvm_init();
2304         } else {
2305                 int rc;
2306                 struct physdev_pirq_eoi_gmfn eoi_gmfn;
2307
2308                 pirq_eoi_map = (void *)__get_free_page(GFP_KERNEL|__GFP_ZERO);
2309                 eoi_gmfn.gmfn = virt_to_gfn(pirq_eoi_map);
2310                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_pirq_eoi_gmfn_v2, &eoi_gmfn);
2311                 if (rc != 0) {
2312                         free_page((unsigned long) pirq_eoi_map);
2313                         pirq_eoi_map = NULL;
2314                 } else
2315                         pirq_needs_eoi = pirq_check_eoi_map;
2316         }
2317 #endif
2318 }