virt/kvm/eventfd.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * kvm eventfd support - use eventfd objects to signal various KVM events
   4  *
   5  * Copyright 2009 Novell.  All Rights Reserved.
   6  * Copyright 2010 Red Hat, Inc. and/or its affiliates.
   7  *
   8  * Author:
   9  *      Gregory Haskins <[email protected]>
  10  */
  11
  12 #include <linux/kvm_host.h>
  13 #include <linux/kvm.h>
  14 #include <linux/kvm_irqfd.h>
  15 #include <linux/workqueue.h>
  16 #include <linux/syscalls.h>
  17 #include <linux/wait.h>
  18 #include <linux/poll.h>
  19 #include <linux/file.h>
  20 #include <linux/list.h>
  21 #include <linux/eventfd.h>
  22 #include <linux/kernel.h>
  23 #include <linux/srcu.h>
  24 #include <linux/slab.h>
  25 #include <linux/seqlock.h>
  26 #include <linux/irqbypass.h>
  27 #include <trace/events/kvm.h>
  28
  29 #include <kvm/iodev.h>
  30
  31 #ifdef CONFIG_HAVE_KVM_IRQFD
  32
  33 static struct workqueue_struct *irqfd_cleanup_wq;
  34
  35 bool __attribute__((weak))
  36 kvm_arch_irqfd_allowed(struct kvm *kvm, struct kvm_irqfd *args)
  37 {
  38         return true;
  39 }
  40
  41 static void
  42 irqfd_inject(struct work_struct *work)
  43 {
  44         struct kvm_kernel_irqfd *irqfd =
  45                 container_of(work, struct kvm_kernel_irqfd, inject);
  46         struct kvm *kvm = irqfd->kvm;
  47
  48         if (!irqfd->resampler) {
  49                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1,
  50                                 false);
  51                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0,
  52                                 false);
  53         } else
  54                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
  55                             irqfd->gsi, 1, false);
  56 }
  57
  58 /*
  59  * Since resampler irqfds share an IRQ source ID, we de-assert once
  60  * then notify all of the resampler irqfds using this GSI.  We can't
  61  * do multiple de-asserts or we risk racing with incoming re-asserts.
  62  */
  63 static void
  64 irqfd_resampler_ack(struct kvm_irq_ack_notifier *kian)
  65 {
  66         struct kvm_kernel_irqfd_resampler *resampler;
  67         struct kvm *kvm;
  68         struct kvm_kernel_irqfd *irqfd;
  69         int idx;
  70
  71         resampler = container_of(kian,
  72                         struct kvm_kernel_irqfd_resampler, notifier);
  73         kvm = resampler->kvm;
  74
  75         kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
  76                     resampler->notifier.gsi, 0, false);
  77
  78         idx = srcu_read_lock(&kvm->irq_srcu);
  79
  80         list_for_each_entry_rcu(irqfd, &resampler->list, resampler_link)
  81                 eventfd_signal(irqfd->resamplefd, 1);
  82
  83         srcu_read_unlock(&kvm->irq_srcu, idx);
  84 }
  85
  86 static void
  87 irqfd_resampler_shutdown(struct kvm_kernel_irqfd *irqfd)
  88 {
  89         struct kvm_kernel_irqfd_resampler *resampler = irqfd->resampler;
  90         struct kvm *kvm = resampler->kvm;
  91
  92         mutex_lock(&kvm->irqfds.resampler_lock);
  93
  94         list_del_rcu(&irqfd->resampler_link);
  95         synchronize_srcu(&kvm->irq_srcu);
  96
  97         if (list_empty(&resampler->list)) {
  98                 list_del(&resampler->link);
  99                 kvm_unregister_irq_ack_notifier(kvm, &resampler->notifier);
 100                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 101                             resampler->notifier.gsi, 0, false);
 102                 kfree(resampler);
 103         }
 104
 105         mutex_unlock(&kvm->irqfds.resampler_lock);
 106 }
 107
 108 /*
 109  * Race-free decouple logic (ordering is critical)
 110  */
 111 static void
 112 irqfd_shutdown(struct work_struct *work)
 113 {
 114         struct kvm_kernel_irqfd *irqfd =
 115                 container_of(work, struct kvm_kernel_irqfd, shutdown);
 116         struct kvm *kvm = irqfd->kvm;
 117         u64 cnt;
 118
 119         /* Make sure irqfd has been initialized in assign path. */
 120         synchronize_srcu(&kvm->irq_srcu);
 121
 122         /*
 123          * Synchronize with the wait-queue and unhook ourselves to prevent
 124          * further events.
 125          */
 126         eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
 127
 128         /*
 129          * We know no new events will be scheduled at this point, so block
 130          * until all previously outstanding events have completed
 131          */
 132         flush_work(&irqfd->inject);
 133
 134         if (irqfd->resampler) {
 135                 irqfd_resampler_shutdown(irqfd);
 136                 eventfd_ctx_put(irqfd->resamplefd);
 137         }
 138
 139         /*
 140          * It is now safe to release the object's resources
 141          */
 142 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 143         irq_bypass_unregister_consumer(&irqfd->consumer);
 144 #endif
 145         eventfd_ctx_put(irqfd->eventfd);
 146         kfree(irqfd);
 147 }
 148
 149
 150 /* assumes kvm->irqfds.lock is held */
 151 static bool
 152 irqfd_is_active(struct kvm_kernel_irqfd *irqfd)
 153 {
 154         return list_empty(&irqfd->list) ? false : true;
 155 }
 156
 157 /*
 158  * Mark the irqfd as inactive and schedule it for removal
 159  *
 160  * assumes kvm->irqfds.lock is held
 161  */
 162 static void
 163 irqfd_deactivate(struct kvm_kernel_irqfd *irqfd)
 164 {
 165         BUG_ON(!irqfd_is_active(irqfd));
 166
 167         list_del_init(&irqfd->list);
 168
 169         queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
 170 }
 171
 172 int __attribute__((weak)) kvm_arch_set_irq_inatomic(
 173                                 struct kvm_kernel_irq_routing_entry *irq,
 174                                 struct kvm *kvm, int irq_source_id,
 175                                 int level,
 176                                 bool line_status)
 177 {
 178         return -EWOULDBLOCK;
 179 }
 180
 181 /*
 182  * Called with wqh->lock held and interrupts disabled
 183  */
 184 static int
 185 irqfd_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
 186 {
 187         struct kvm_kernel_irqfd *irqfd =
 188                 container_of(wait, struct kvm_kernel_irqfd, wait);
 189         __poll_t flags = key_to_poll(key);
 190         struct kvm_kernel_irq_routing_entry irq;
 191         struct kvm *kvm = irqfd->kvm;
 192         unsigned seq;
 193         int idx;
 194         int ret = 0;
 195
 196         if (flags & EPOLLIN) {
 197                 u64 cnt;
 198                 eventfd_ctx_do_read(irqfd->eventfd, &cnt);
 199
 200                 idx = srcu_read_lock(&kvm->irq_srcu);
 201                 do {
 202                         seq = read_seqcount_begin(&irqfd->irq_entry_sc);
 203                         irq = irqfd->irq_entry;
 204                 } while (read_seqcount_retry(&irqfd->irq_entry_sc, seq));
 205                 /* An event has been signaled, inject an interrupt */
 206                 if (kvm_arch_set_irq_inatomic(&irq, kvm,
 207                                               KVM_USERSPACE_IRQ_SOURCE_ID, 1,
 208                                               false) == -EWOULDBLOCK)
 209                         schedule_work(&irqfd->inject);
 210                 srcu_read_unlock(&kvm->irq_srcu, idx);
 211                 ret = 1;
 212         }
 213
 214         if (flags & EPOLLHUP) {
 215                 /* The eventfd is closing, detach from KVM */
 216                 unsigned long iflags;
 217
 218                 spin_lock_irqsave(&kvm->irqfds.lock, iflags);
 219
 220                 /*
 221                  * We must check if someone deactivated the irqfd before
 222                  * we could acquire the irqfds.lock since the item is
 223                  * deactivated from the KVM side before it is unhooked from
 224                  * the wait-queue.  If it is already deactivated, we can
 225                  * simply return knowing the other side will cleanup for us.
 226                  * We cannot race against the irqfd going away since the
 227                  * other side is required to acquire wqh->lock, which we hold
 228                  */
 229                 if (irqfd_is_active(irqfd))
 230                         irqfd_deactivate(irqfd);
 231
 232                 spin_unlock_irqrestore(&kvm->irqfds.lock, iflags);
 233         }
 234
 235         return ret;
 236 }
 237
 238 static void
 239 irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
 240                         poll_table *pt)
 241 {
 242         struct kvm_kernel_irqfd *irqfd =
 243                 container_of(pt, struct kvm_kernel_irqfd, pt);
 244         add_wait_queue_priority(wqh, &irqfd->wait);
 245 }
 246
 247 /* Must be called under irqfds.lock */
 248 static void irqfd_update(struct kvm *kvm, struct kvm_kernel_irqfd *irqfd)
 249 {
 250         struct kvm_kernel_irq_routing_entry *e;
 251         struct kvm_kernel_irq_routing_entry entries[KVM_NR_IRQCHIPS];
 252         int n_entries;
 253
 254         n_entries = kvm_irq_map_gsi(kvm, entries, irqfd->gsi);
 255
 256         write_seqcount_begin(&irqfd->irq_entry_sc);
 257
 258         e = entries;
 259         if (n_entries == 1)
 260                 irqfd->irq_entry = *e;
 261         else
 262                 irqfd->irq_entry.type = 0;
 263
 264         write_seqcount_end(&irqfd->irq_entry_sc);
 265 }
 266
 267 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 268 void __attribute__((weak)) kvm_arch_irq_bypass_stop(
 269                                 struct irq_bypass_consumer *cons)
 270 {
 271 }
 272
 273 void __attribute__((weak)) kvm_arch_irq_bypass_start(
 274                                 struct irq_bypass_consumer *cons)
 275 {
 276 }
 277
 278 int  __attribute__((weak)) kvm_arch_update_irqfd_routing(
 279                                 struct kvm *kvm, unsigned int host_irq,
 280                                 uint32_t guest_irq, bool set)
 281 {
 282         return 0;
 283 }
 284
 285 bool __attribute__((weak)) kvm_arch_irqfd_route_changed(
 286                                 struct kvm_kernel_irq_routing_entry *old,
 287                                 struct kvm_kernel_irq_routing_entry *new)
 288 {
 289         return true;
 290 }
 291 #endif
 292
 293 static int
 294 kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
 295 {
 296         struct kvm_kernel_irqfd *irqfd, *tmp;
 297         struct fd f;
 298         struct eventfd_ctx *eventfd = NULL, *resamplefd = NULL;
 299         int ret;
 300         __poll_t events;
 301         int idx;
 302
 303         if (!kvm_arch_intc_initialized(kvm))
 304                 return -EAGAIN;
 305
 306         if (!kvm_arch_irqfd_allowed(kvm, args))
 307                 return -EINVAL;
 308
 309         irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL_ACCOUNT);
 310         if (!irqfd)
 311                 return -ENOMEM;
 312
 313         irqfd->kvm = kvm;
 314         irqfd->gsi = args->gsi;
 315         INIT_LIST_HEAD(&irqfd->list);
 316         INIT_WORK(&irqfd->inject, irqfd_inject);
 317         INIT_WORK(&irqfd->shutdown, irqfd_shutdown);
 318         seqcount_spinlock_init(&irqfd->irq_entry_sc, &kvm->irqfds.lock);
 319
 320         f = fdget(args->fd);
 321         if (!f.file) {
 322                 ret = -EBADF;
 323                 goto out;
 324         }
 325
 326         eventfd = eventfd_ctx_fileget(f.file);
 327         if (IS_ERR(eventfd)) {
 328                 ret = PTR_ERR(eventfd);
 329                 goto fail;
 330         }
 331
 332         irqfd->eventfd = eventfd;
 333
 334         if (args->flags & KVM_IRQFD_FLAG_RESAMPLE) {
 335                 struct kvm_kernel_irqfd_resampler *resampler;
 336
 337                 resamplefd = eventfd_ctx_fdget(args->resamplefd);
 338                 if (IS_ERR(resamplefd)) {
 339                         ret = PTR_ERR(resamplefd);
 340                         goto fail;
 341                 }
 342
 343                 irqfd->resamplefd = resamplefd;
 344                 INIT_LIST_HEAD(&irqfd->resampler_link);
 345
 346                 mutex_lock(&kvm->irqfds.resampler_lock);
 347
 348                 list_for_each_entry(resampler,
 349                                     &kvm->irqfds.resampler_list, link) {
 350                         if (resampler->notifier.gsi == irqfd->gsi) {
 351                                 irqfd->resampler = resampler;
 352                                 break;
 353                         }
 354                 }
 355
 356                 if (!irqfd->resampler) {
 357                         resampler = kzalloc(sizeof(*resampler),
 358                                             GFP_KERNEL_ACCOUNT);
 359                         if (!resampler) {
 360                                 ret = -ENOMEM;
 361                                 mutex_unlock(&kvm->irqfds.resampler_lock);
 362                                 goto fail;
 363                         }
 364
 365                         resampler->kvm = kvm;
 366                         INIT_LIST_HEAD(&resampler->list);
 367                         resampler->notifier.gsi = irqfd->gsi;
 368                         resampler->notifier.irq_acked = irqfd_resampler_ack;
 369                         INIT_LIST_HEAD(&resampler->link);
 370
 371                         list_add(&resampler->link, &kvm->irqfds.resampler_list);
 372                         kvm_register_irq_ack_notifier(kvm,
 373                                                       &resampler->notifier);
 374                         irqfd->resampler = resampler;
 375                 }
 376
 377                 list_add_rcu(&irqfd->resampler_link, &irqfd->resampler->list);
 378                 synchronize_srcu(&kvm->irq_srcu);
 379
 380                 mutex_unlock(&kvm->irqfds.resampler_lock);
 381         }
 382
 383         /*
 384          * Install our own custom wake-up handling so we are notified via
 385          * a callback whenever someone signals the underlying eventfd
 386          */
 387         init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
 388         init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
 389
 390         spin_lock_irq(&kvm->irqfds.lock);
 391
 392         ret = 0;
 393         list_for_each_entry(tmp, &kvm->irqfds.items, list) {
 394                 if (irqfd->eventfd != tmp->eventfd)
 395                         continue;
 396                 /* This fd is used for another irq already. */
 397                 ret = -EBUSY;
 398                 spin_unlock_irq(&kvm->irqfds.lock);
 399                 goto fail;
 400         }
 401
 402         idx = srcu_read_lock(&kvm->irq_srcu);
 403         irqfd_update(kvm, irqfd);
 404
 405         list_add_tail(&irqfd->list, &kvm->irqfds.items);
 406
 407         spin_unlock_irq(&kvm->irqfds.lock);
 408
 409         /*
 410          * Check if there was an event already pending on the eventfd
 411          * before we registered, and trigger it as if we didn't miss it.
 412          */
 413         events = vfs_poll(f.file, &irqfd->pt);
 414
 415         if (events & EPOLLIN)
 416                 schedule_work(&irqfd->inject);
 417
 418 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 419         if (kvm_arch_has_irq_bypass()) {
 420                 irqfd->consumer.token = (void *)irqfd->eventfd;
 421                 irqfd->consumer.add_producer = kvm_arch_irq_bypass_add_producer;
 422                 irqfd->consumer.del_producer = kvm_arch_irq_bypass_del_producer;
 423                 irqfd->consumer.stop = kvm_arch_irq_bypass_stop;
 424                 irqfd->consumer.start = kvm_arch_irq_bypass_start;
 425                 ret = irq_bypass_register_consumer(&irqfd->consumer);
 426                 if (ret)
 427                         pr_info("irq bypass consumer (token %p) registration fails: %d\n",
 428                                 irqfd->consumer.token, ret);
 429         }
 430 #endif
 431
 432         srcu_read_unlock(&kvm->irq_srcu, idx);
 433
 434         /*
 435          * do not drop the file until the irqfd is fully initialized, otherwise
 436          * we might race against the EPOLLHUP
 437          */
 438         fdput(f);
 439         return 0;
 440
 441 fail:
 442         if (irqfd->resampler)
 443                 irqfd_resampler_shutdown(irqfd);
 444
 445         if (resamplefd && !IS_ERR(resamplefd))
 446                 eventfd_ctx_put(resamplefd);
 447
 448         if (eventfd && !IS_ERR(eventfd))
 449                 eventfd_ctx_put(eventfd);
 450
 451         fdput(f);
 452
 453 out:
 454         kfree(irqfd);
 455         return ret;
 456 }
 457
 458 bool kvm_irq_has_notifier(struct kvm *kvm, unsigned irqchip, unsigned pin)
 459 {
 460         struct kvm_irq_ack_notifier *kian;
 461         int gsi, idx;
 462
 463         idx = srcu_read_lock(&kvm->irq_srcu);
 464         gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
 465         if (gsi != -1)
 466                 hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
 467                                          link)
 468                         if (kian->gsi == gsi) {
 469                                 srcu_read_unlock(&kvm->irq_srcu, idx);
 470                                 return true;
 471                         }
 472
 473         srcu_read_unlock(&kvm->irq_srcu, idx);
 474
 475         return false;
 476 }
 477 EXPORT_SYMBOL_GPL(kvm_irq_has_notifier);
 478
 479 void kvm_notify_acked_gsi(struct kvm *kvm, int gsi)
 480 {
 481         struct kvm_irq_ack_notifier *kian;
 482
 483         hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
 484                                  link)
 485                 if (kian->gsi == gsi)
 486                         kian->irq_acked(kian);
 487 }
 488
 489 void kvm_notify_acked_irq(struct kvm *kvm, unsigned irqchip, unsigned pin)
 490 {
 491         int gsi, idx;
 492
 493         trace_kvm_ack_irq(irqchip, pin);
 494
 495         idx = srcu_read_lock(&kvm->irq_srcu);
 496         gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
 497         if (gsi != -1)
 498                 kvm_notify_acked_gsi(kvm, gsi);
 499         srcu_read_unlock(&kvm->irq_srcu, idx);
 500 }
 501
 502 void kvm_register_irq_ack_notifier(struct kvm *kvm,
 503                                    struct kvm_irq_ack_notifier *kian)
 504 {
 505         mutex_lock(&kvm->irq_lock);
 506         hlist_add_head_rcu(&kian->link, &kvm->irq_ack_notifier_list);
 507         mutex_unlock(&kvm->irq_lock);
 508         kvm_arch_post_irq_ack_notifier_list_update(kvm);
 509 }
 510
 511 void kvm_unregister_irq_ack_notifier(struct kvm *kvm,
 512                                     struct kvm_irq_ack_notifier *kian)
 513 {
 514         mutex_lock(&kvm->irq_lock);
 515         hlist_del_init_rcu(&kian->link);
 516         mutex_unlock(&kvm->irq_lock);
 517         synchronize_srcu(&kvm->irq_srcu);
 518         kvm_arch_post_irq_ack_notifier_list_update(kvm);
 519 }
 520 #endif
 521
 522 void
 523 kvm_eventfd_init(struct kvm *kvm)
 524 {
 525 #ifdef CONFIG_HAVE_KVM_IRQFD
 526         spin_lock_init(&kvm->irqfds.lock);
 527         INIT_LIST_HEAD(&kvm->irqfds.items);
 528         INIT_LIST_HEAD(&kvm->irqfds.resampler_list);
 529         mutex_init(&kvm->irqfds.resampler_lock);
 530 #endif
 531         INIT_LIST_HEAD(&kvm->ioeventfds);
 532 }
 533
 534 #ifdef CONFIG_HAVE_KVM_IRQFD
 535 /*
 536  * shutdown any irqfd's that match fd+gsi
 537  */
 538 static int
 539 kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
 540 {
 541         struct kvm_kernel_irqfd *irqfd, *tmp;
 542         struct eventfd_ctx *eventfd;
 543
 544         eventfd = eventfd_ctx_fdget(args->fd);
 545         if (IS_ERR(eventfd))
 546                 return PTR_ERR(eventfd);
 547
 548         spin_lock_irq(&kvm->irqfds.lock);
 549
 550         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
 551                 if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
 552                         /*
 553                          * This clearing of irq_entry.type is needed for when
 554                          * another thread calls kvm_irq_routing_update before
 555                          * we flush workqueue below (we synchronize with
 556                          * kvm_irq_routing_update using irqfds.lock).
 557                          */
 558                         write_seqcount_begin(&irqfd->irq_entry_sc);
 559                         irqfd->irq_entry.type = 0;
 560                         write_seqcount_end(&irqfd->irq_entry_sc);
 561                         irqfd_deactivate(irqfd);
 562                 }
 563         }
 564
 565         spin_unlock_irq(&kvm->irqfds.lock);
 566         eventfd_ctx_put(eventfd);
 567
 568         /*
 569          * Block until we know all outstanding shutdown jobs have completed
 570          * so that we guarantee there will not be any more interrupts on this
 571          * gsi once this deassign function returns.
 572          */
 573         flush_workqueue(irqfd_cleanup_wq);
 574
 575         return 0;
 576 }
 577
 578 int
 579 kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
 580 {
 581         if (args->flags & ~(KVM_IRQFD_FLAG_DEASSIGN | KVM_IRQFD_FLAG_RESAMPLE))
 582                 return -EINVAL;
 583
 584         if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
 585                 return kvm_irqfd_deassign(kvm, args);
 586
 587         return kvm_irqfd_assign(kvm, args);
 588 }
 589
 590 /*
 591  * This function is called as the kvm VM fd is being released. Shutdown all
 592  * irqfds that still remain open
 593  */
 594 void
 595 kvm_irqfd_release(struct kvm *kvm)
 596 {
 597         struct kvm_kernel_irqfd *irqfd, *tmp;
 598
 599         spin_lock_irq(&kvm->irqfds.lock);
 600
 601         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
 602                 irqfd_deactivate(irqfd);
 603
 604         spin_unlock_irq(&kvm->irqfds.lock);
 605
 606         /*
 607          * Block until we know all outstanding shutdown jobs have completed
 608          * since we do not take a kvm* reference.
 609          */
 610         flush_workqueue(irqfd_cleanup_wq);
 611
 612 }
 613
 614 /*
 615  * Take note of a change in irq routing.
 616  * Caller must invoke synchronize_srcu(&kvm->irq_srcu) afterwards.
 617  */
 618 void kvm_irq_routing_update(struct kvm *kvm)
 619 {
 620         struct kvm_kernel_irqfd *irqfd;
 621
 622         spin_lock_irq(&kvm->irqfds.lock);
 623
 624         list_for_each_entry(irqfd, &kvm->irqfds.items, list) {
 625 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 626                 /* Under irqfds.lock, so can read irq_entry safely */
 627                 struct kvm_kernel_irq_routing_entry old = irqfd->irq_entry;
 628 #endif
 629
 630                 irqfd_update(kvm, irqfd);
 631
 632 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 633                 if (irqfd->producer &&
 634                     kvm_arch_irqfd_route_changed(&old, &irqfd->irq_entry)) {
 635                         int ret = kvm_arch_update_irqfd_routing(
 636                                         irqfd->kvm, irqfd->producer->irq,
 637                                         irqfd->gsi, 1);
 638                         WARN_ON(ret);
 639                 }
 640 #endif
 641         }
 642
 643         spin_unlock_irq(&kvm->irqfds.lock);
 644 }
 645
 646 /*
 647  * create a host-wide workqueue for issuing deferred shutdown requests
 648  * aggregated from all vm* instances. We need our own isolated
 649  * queue to ease flushing work items when a VM exits.
 650  */
 651 int kvm_irqfd_init(void)
 652 {
 653         irqfd_cleanup_wq = alloc_workqueue("kvm-irqfd-cleanup", 0, 0);
 654         if (!irqfd_cleanup_wq)
 655                 return -ENOMEM;
 656
 657         return 0;
 658 }
 659
 660 void kvm_irqfd_exit(void)
 661 {
 662         destroy_workqueue(irqfd_cleanup_wq);
 663 }
 664 #endif
 665
 666 /*
 667  * --------------------------------------------------------------------
 668  * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
 669  *
 670  * userspace can register a PIO/MMIO address with an eventfd for receiving
 671  * notification when the memory has been touched.
 672  * --------------------------------------------------------------------
 673  */
 674
 675 struct _ioeventfd {
 676         struct list_head     list;
 677         u64                  addr;
 678         int                  length;
 679         struct eventfd_ctx  *eventfd;
 680         u64                  datamatch;
 681         struct kvm_io_device dev;
 682         u8                   bus_idx;
 683         bool                 wildcard;
 684 };
 685
 686 static inline struct _ioeventfd *
 687 to_ioeventfd(struct kvm_io_device *dev)
 688 {
 689         return container_of(dev, struct _ioeventfd, dev);
 690 }
 691
 692 static void
 693 ioeventfd_release(struct _ioeventfd *p)
 694 {
 695         eventfd_ctx_put(p->eventfd);
 696         list_del(&p->list);
 697         kfree(p);
 698 }
 699
 700 static bool
 701 ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
 702 {
 703         u64 _val;
 704
 705         if (addr != p->addr)
 706                 /* address must be precise for a hit */
 707                 return false;
 708
 709         if (!p->length)
 710                 /* length = 0 means only look at the address, so always a hit */
 711                 return true;
 712
 713         if (len != p->length)
 714                 /* address-range must be precise for a hit */
 715                 return false;
 716
 717         if (p->wildcard)
 718                 /* all else equal, wildcard is always a hit */
 719                 return true;
 720
 721         /* otherwise, we have to actually compare the data */
 722
 723         BUG_ON(!IS_ALIGNED((unsigned long)val, len));
 724
 725         switch (len) {
 726         case 1:
 727                 _val = *(u8 *)val;
 728                 break;
 729         case 2:
 730                 _val = *(u16 *)val;
 731                 break;
 732         case 4:
 733                 _val = *(u32 *)val;
 734                 break;
 735         case 8:
 736                 _val = *(u64 *)val;
 737                 break;
 738         default:
 739                 return false;
 740         }
 741
 742         return _val == p->datamatch;
 743 }
 744
 745 /* MMIO/PIO writes trigger an event if the addr/val match */
 746 static int
 747 ioeventfd_write(struct kvm_vcpu *vcpu, struct kvm_io_device *this, gpa_t addr,
 748                 int len, const void *val)
 749 {
 750         struct _ioeventfd *p = to_ioeventfd(this);
 751
 752         if (!ioeventfd_in_range(p, addr, len, val))
 753                 return -EOPNOTSUPP;
 754
 755         eventfd_signal(p->eventfd, 1);
 756         return 0;
 757 }
 758
 759 /*
 760  * This function is called as KVM is completely shutting down.  We do not
 761  * need to worry about locking just nuke anything we have as quickly as possible
 762  */
 763 static void
 764 ioeventfd_destructor(struct kvm_io_device *this)
 765 {
 766         struct _ioeventfd *p = to_ioeventfd(this);
 767
 768         ioeventfd_release(p);
 769 }
 770
 771 static const struct kvm_io_device_ops ioeventfd_ops = {
 772         .write      = ioeventfd_write,
 773         .destructor = ioeventfd_destructor,
 774 };
 775
 776 /* assumes kvm->slots_lock held */
 777 static bool
 778 ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
 779 {
 780         struct _ioeventfd *_p;
 781
 782         list_for_each_entry(_p, &kvm->ioeventfds, list)
 783                 if (_p->bus_idx == p->bus_idx &&
 784                     _p->addr == p->addr &&
 785                     (!_p->length || !p->length ||
 786                      (_p->length == p->length &&
 787                       (_p->wildcard || p->wildcard ||
 788                        _p->datamatch == p->datamatch))))
 789                         return true;
 790
 791         return false;
 792 }
 793
 794 static enum kvm_bus ioeventfd_bus_from_flags(__u32 flags)
 795 {
 796         if (flags & KVM_IOEVENTFD_FLAG_PIO)
 797                 return KVM_PIO_BUS;
 798         if (flags & KVM_IOEVENTFD_FLAG_VIRTIO_CCW_NOTIFY)
 799                 return KVM_VIRTIO_CCW_NOTIFY_BUS;
 800         return KVM_MMIO_BUS;
 801 }
 802
 803 static int kvm_assign_ioeventfd_idx(struct kvm *kvm,
 804                                 enum kvm_bus bus_idx,
 805                                 struct kvm_ioeventfd *args)
 806 {
 807
 808         struct eventfd_ctx *eventfd;
 809         struct _ioeventfd *p;
 810         int ret;
 811
 812         eventfd = eventfd_ctx_fdget(args->fd);
 813         if (IS_ERR(eventfd))
 814                 return PTR_ERR(eventfd);
 815
 816         p = kzalloc(sizeof(*p), GFP_KERNEL_ACCOUNT);
 817         if (!p) {
 818                 ret = -ENOMEM;
 819                 goto fail;
 820         }
 821
 822         INIT_LIST_HEAD(&p->list);
 823         p->addr    = args->addr;
 824         p->bus_idx = bus_idx;
 825         p->length  = args->len;
 826         p->eventfd = eventfd;
 827
 828         /* The datamatch feature is optional, otherwise this is a wildcard */
 829         if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
 830                 p->datamatch = args->datamatch;
 831         else
 832                 p->wildcard = true;
 833
 834         mutex_lock(&kvm->slots_lock);
 835
 836         /* Verify that there isn't a match already */
 837         if (ioeventfd_check_collision(kvm, p)) {
 838                 ret = -EEXIST;
 839                 goto unlock_fail;
 840         }
 841
 842         kvm_iodevice_init(&p->dev, &ioeventfd_ops);
 843
 844         ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
 845                                       &p->dev);
 846         if (ret < 0)
 847                 goto unlock_fail;
 848
 849         kvm_get_bus(kvm, bus_idx)->ioeventfd_count++;
 850         list_add_tail(&p->list, &kvm->ioeventfds);
 851
 852         mutex_unlock(&kvm->slots_lock);
 853
 854         return 0;
 855
 856 unlock_fail:
 857         mutex_unlock(&kvm->slots_lock);
 858
 859 fail:
 860         kfree(p);
 861         eventfd_ctx_put(eventfd);
 862
 863         return ret;
 864 }
 865
 866 static int
 867 kvm_deassign_ioeventfd_idx(struct kvm *kvm, enum kvm_bus bus_idx,
 868                            struct kvm_ioeventfd *args)
 869 {
 870         struct _ioeventfd        *p, *tmp;
 871         struct eventfd_ctx       *eventfd;
 872         struct kvm_io_bus        *bus;
 873         int                       ret = -ENOENT;
 874         bool                      wildcard;
 875
 876         eventfd = eventfd_ctx_fdget(args->fd);
 877         if (IS_ERR(eventfd))
 878                 return PTR_ERR(eventfd);
 879
 880         wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
 881
 882         mutex_lock(&kvm->slots_lock);
 883
 884         list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
 885
 886                 if (p->bus_idx != bus_idx ||
 887                     p->eventfd != eventfd  ||
 888                     p->addr != args->addr  ||
 889                     p->length != args->len ||
 890                     p->wildcard != wildcard)
 891                         continue;
 892
 893                 if (!p->wildcard && p->datamatch != args->datamatch)
 894                         continue;
 895
 896                 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
 897                 bus = kvm_get_bus(kvm, bus_idx);
 898                 if (bus)
 899                         bus->ioeventfd_count--;
 900                 ioeventfd_release(p);
 901                 ret = 0;
 902                 break;
 903         }
 904
 905         mutex_unlock(&kvm->slots_lock);
 906
 907         eventfd_ctx_put(eventfd);
 908
 909         return ret;
 910 }
 911
 912 static int kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 913 {
 914         enum kvm_bus bus_idx = ioeventfd_bus_from_flags(args->flags);
 915         int ret = kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
 916
 917         if (!args->len && bus_idx == KVM_MMIO_BUS)
 918                 kvm_deassign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
 919
 920         return ret;
 921 }
 922
 923 static int
 924 kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 925 {
 926         enum kvm_bus              bus_idx;
 927         int ret;
 928
 929         bus_idx = ioeventfd_bus_from_flags(args->flags);
 930         /* must be natural-word sized, or 0 to ignore length */
 931         switch (args->len) {
 932         case 0:
 933         case 1:
 934         case 2:
 935         case 4:
 936         case 8:
 937                 break;
 938         default:
 939                 return -EINVAL;
 940         }
 941
 942         /* check for range overflow */
 943         if (args->addr + args->len < args->addr)
 944                 return -EINVAL;
 945
 946         /* check for extra flags that we don't understand */
 947         if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
 948                 return -EINVAL;
 949
 950         /* ioeventfd with no length can't be combined with DATAMATCH */
 951         if (!args->len && (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH))
 952                 return -EINVAL;
 953
 954         ret = kvm_assign_ioeventfd_idx(kvm, bus_idx, args);
 955         if (ret)
 956                 goto fail;
 957
 958         /* When length is ignored, MMIO is also put on a separate bus, for
 959          * faster lookups.
 960          */
 961         if (!args->len && bus_idx == KVM_MMIO_BUS) {
 962                 ret = kvm_assign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
 963                 if (ret < 0)
 964                         goto fast_fail;
 965         }
 966
 967         return 0;
 968
 969 fast_fail:
 970         kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
 971 fail:
 972         return ret;
 973 }
 974
 975 int
 976 kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 977 {
 978         if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
 979                 return kvm_deassign_ioeventfd(kvm, args);
 980
 981         return kvm_assign_ioeventfd(kvm, args);
 982 }