virt/kvm/eventfd.c

   1 /*
   2  * kvm eventfd support - use eventfd objects to signal various KVM events
   3  *
   4  * Copyright 2009 Novell.  All Rights Reserved.
   5  * Copyright 2010 Red Hat, Inc. and/or its affiliates.
   6  *
   7  * Author:
   8  *      Gregory Haskins <ghaskins@novell.com>
   9  *
  10  * This file is free software; you can redistribute it and/or modify
  11  * it under the terms of version 2 of the GNU General Public License
  12  * as published by the Free Software Foundation.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU General Public License
  20  * along with this program; if not, write to the Free Software Foundation,
  21  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA.
  22  */
  23
  24 #include <linux/kvm_host.h>
  25 #include <linux/kvm.h>
  26 #include <linux/kvm_irqfd.h>
  27 #include <linux/workqueue.h>
  28 #include <linux/syscalls.h>
  29 #include <linux/wait.h>
  30 #include <linux/poll.h>
  31 #include <linux/file.h>
  32 #include <linux/list.h>
  33 #include <linux/eventfd.h>
  34 #include <linux/kernel.h>
  35 #include <linux/srcu.h>
  36 #include <linux/slab.h>
  37 #include <linux/seqlock.h>
  38 #include <linux/irqbypass.h>
  39 #include <trace/events/kvm.h>
  40
  41 #include <kvm/iodev.h>
  42
  43 #ifdef CONFIG_HAVE_KVM_IRQFD
  44
  45 static struct workqueue_struct *irqfd_cleanup_wq;
  46
  47 bool __attribute__((weak))
  48 kvm_arch_irqfd_allowed(struct kvm *kvm, struct kvm_irqfd *args)
  49 {
  50         return true;
  51 }
  52
  53 static void
  54 irqfd_inject(struct work_struct *work)
  55 {
  56         struct kvm_kernel_irqfd *irqfd =
  57                 container_of(work, struct kvm_kernel_irqfd, inject);
  58         struct kvm *kvm = irqfd->kvm;
  59
  60         if (!irqfd->resampler) {
  61                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1,
  62                                 false);
  63                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0,
  64                                 false);
  65         } else
  66                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
  67                             irqfd->gsi, 1, false);
  68 }
  69
  70 /*
  71  * Since resampler irqfds share an IRQ source ID, we de-assert once
  72  * then notify all of the resampler irqfds using this GSI.  We can't
  73  * do multiple de-asserts or we risk racing with incoming re-asserts.
  74  */
  75 static void
  76 irqfd_resampler_ack(struct kvm_irq_ack_notifier *kian)
  77 {
  78         struct kvm_kernel_irqfd_resampler *resampler;
  79         struct kvm *kvm;
  80         struct kvm_kernel_irqfd *irqfd;
  81         int idx;
  82
  83         resampler = container_of(kian,
  84                         struct kvm_kernel_irqfd_resampler, notifier);
  85         kvm = resampler->kvm;
  86
  87         kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
  88                     resampler->notifier.gsi, 0, false);
  89
  90         idx = srcu_read_lock(&kvm->irq_srcu);
  91
  92         list_for_each_entry_rcu(irqfd, &resampler->list, resampler_link)
  93                 eventfd_signal(irqfd->resamplefd, 1);
  94
  95         srcu_read_unlock(&kvm->irq_srcu, idx);
  96 }
  97
  98 static void
  99 irqfd_resampler_shutdown(struct kvm_kernel_irqfd *irqfd)
 100 {
 101         struct kvm_kernel_irqfd_resampler *resampler = irqfd->resampler;
 102         struct kvm *kvm = resampler->kvm;
 103
 104         mutex_lock(&kvm->irqfds.resampler_lock);
 105
 106         list_del_rcu(&irqfd->resampler_link);
 107         synchronize_srcu(&kvm->irq_srcu);
 108
 109         if (list_empty(&resampler->list)) {
 110                 list_del(&resampler->link);
 111                 kvm_unregister_irq_ack_notifier(kvm, &resampler->notifier);
 112                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 113                             resampler->notifier.gsi, 0, false);
 114                 kfree(resampler);
 115         }
 116
 117         mutex_unlock(&kvm->irqfds.resampler_lock);
 118 }
 119
 120 /*
 121  * Race-free decouple logic (ordering is critical)
 122  */
 123 static void
 124 irqfd_shutdown(struct work_struct *work)
 125 {
 126         struct kvm_kernel_irqfd *irqfd =
 127                 container_of(work, struct kvm_kernel_irqfd, shutdown);
 128         struct kvm *kvm = irqfd->kvm;
 129         u64 cnt;
 130
 131         /* Make sure irqfd has been initalized in assign path. */
 132         synchronize_srcu(&kvm->irq_srcu);
 133
 134         /*
 135          * Synchronize with the wait-queue and unhook ourselves to prevent
 136          * further events.
 137          */
 138         eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
 139
 140         /*
 141          * We know no new events will be scheduled at this point, so block
 142          * until all previously outstanding events have completed
 143          */
 144         flush_work(&irqfd->inject);
 145
 146         if (irqfd->resampler) {
 147                 irqfd_resampler_shutdown(irqfd);
 148                 eventfd_ctx_put(irqfd->resamplefd);
 149         }
 150
 151         /*
 152          * It is now safe to release the object's resources
 153          */
 154 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 155         irq_bypass_unregister_consumer(&irqfd->consumer);
 156 #endif
 157         eventfd_ctx_put(irqfd->eventfd);
 158         kfree(irqfd);
 159 }
 160
 161
 162 /* assumes kvm->irqfds.lock is held */
 163 static bool
 164 irqfd_is_active(struct kvm_kernel_irqfd *irqfd)
 165 {
 166         return list_empty(&irqfd->list) ? false : true;
 167 }
 168
 169 /*
 170  * Mark the irqfd as inactive and schedule it for removal
 171  *
 172  * assumes kvm->irqfds.lock is held
 173  */
 174 static void
 175 irqfd_deactivate(struct kvm_kernel_irqfd *irqfd)
 176 {
 177         BUG_ON(!irqfd_is_active(irqfd));
 178
 179         list_del_init(&irqfd->list);
 180
 181         queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
 182 }
 183
 184 int __attribute__((weak)) kvm_arch_set_irq_inatomic(
 185                                 struct kvm_kernel_irq_routing_entry *irq,
 186                                 struct kvm *kvm, int irq_source_id,
 187                                 int level,
 188                                 bool line_status)
 189 {
 190         return -EWOULDBLOCK;
 191 }
 192
 193 /*
 194  * Called with wqh->lock held and interrupts disabled
 195  */
 196 static int
 197 irqfd_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
 198 {
 199         struct kvm_kernel_irqfd *irqfd =
 200                 container_of(wait, struct kvm_kernel_irqfd, wait);
 201         __poll_t flags = key_to_poll(key);
 202         struct kvm_kernel_irq_routing_entry irq;
 203         struct kvm *kvm = irqfd->kvm;
 204         unsigned seq;
 205         int idx;
 206
 207         if (flags & EPOLLIN) {
 208                 idx = srcu_read_lock(&kvm->irq_srcu);
 209                 do {
 210                         seq = read_seqcount_begin(&irqfd->irq_entry_sc);
 211                         irq = irqfd->irq_entry;
 212                 } while (read_seqcount_retry(&irqfd->irq_entry_sc, seq));
 213                 /* An event has been signaled, inject an interrupt */
 214                 if (kvm_arch_set_irq_inatomic(&irq, kvm,
 215                                               KVM_USERSPACE_IRQ_SOURCE_ID, 1,
 216                                               false) == -EWOULDBLOCK)
 217                         schedule_work(&irqfd->inject);
 218                 srcu_read_unlock(&kvm->irq_srcu, idx);
 219         }
 220
 221         if (flags & EPOLLHUP) {
 222                 /* The eventfd is closing, detach from KVM */
 223                 unsigned long flags;
 224
 225                 spin_lock_irqsave(&kvm->irqfds.lock, flags);
 226
 227                 /*
 228                  * We must check if someone deactivated the irqfd before
 229                  * we could acquire the irqfds.lock since the item is
 230                  * deactivated from the KVM side before it is unhooked from
 231                  * the wait-queue.  If it is already deactivated, we can
 232                  * simply return knowing the other side will cleanup for us.
 233                  * We cannot race against the irqfd going away since the
 234                  * other side is required to acquire wqh->lock, which we hold
 235                  */
 236                 if (irqfd_is_active(irqfd))
 237                         irqfd_deactivate(irqfd);
 238
 239                 spin_unlock_irqrestore(&kvm->irqfds.lock, flags);
 240         }
 241
 242         return 0;
 243 }
 244
 245 static void
 246 irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
 247                         poll_table *pt)
 248 {
 249         struct kvm_kernel_irqfd *irqfd =
 250                 container_of(pt, struct kvm_kernel_irqfd, pt);
 251         add_wait_queue(wqh, &irqfd->wait);
 252 }
 253
 254 /* Must be called under irqfds.lock */
 255 static void irqfd_update(struct kvm *kvm, struct kvm_kernel_irqfd *irqfd)
 256 {
 257         struct kvm_kernel_irq_routing_entry *e;
 258         struct kvm_kernel_irq_routing_entry entries[KVM_NR_IRQCHIPS];
 259         int n_entries;
 260
 261         n_entries = kvm_irq_map_gsi(kvm, entries, irqfd->gsi);
 262
 263         write_seqcount_begin(&irqfd->irq_entry_sc);
 264
 265         e = entries;
 266         if (n_entries == 1)
 267                 irqfd->irq_entry = *e;
 268         else
 269                 irqfd->irq_entry.type = 0;
 270
 271         write_seqcount_end(&irqfd->irq_entry_sc);
 272 }
 273
 274 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 275 void __attribute__((weak)) kvm_arch_irq_bypass_stop(
 276                                 struct irq_bypass_consumer *cons)
 277 {
 278 }
 279
 280 void __attribute__((weak)) kvm_arch_irq_bypass_start(
 281                                 struct irq_bypass_consumer *cons)
 282 {
 283 }
 284
 285 int  __attribute__((weak)) kvm_arch_update_irqfd_routing(
 286                                 struct kvm *kvm, unsigned int host_irq,
 287                                 uint32_t guest_irq, bool set)
 288 {
 289         return 0;
 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);
 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_init(&irqfd->irq_entry_sc);
 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), GFP_KERNEL);
 358                         if (!resampler) {
 359                                 ret = -ENOMEM;
 360                                 mutex_unlock(&kvm->irqfds.resampler_lock);
 361                                 goto fail;
 362                         }
 363
 364                         resampler->kvm = kvm;
 365                         INIT_LIST_HEAD(&resampler->list);
 366                         resampler->notifier.gsi = irqfd->gsi;
 367                         resampler->notifier.irq_acked = irqfd_resampler_ack;
 368                         INIT_LIST_HEAD(&resampler->link);
 369
 370                         list_add(&resampler->link, &kvm->irqfds.resampler_list);
 371                         kvm_register_irq_ack_notifier(kvm,
 372                                                       &resampler->notifier);
 373                         irqfd->resampler = resampler;
 374                 }
 375
 376                 list_add_rcu(&irqfd->resampler_link, &irqfd->resampler->list);
 377                 synchronize_srcu(&kvm->irq_srcu);
 378
 379                 mutex_unlock(&kvm->irqfds.resampler_lock);
 380         }
 381
 382         /*
 383          * Install our own custom wake-up handling so we are notified via
 384          * a callback whenever someone signals the underlying eventfd
 385          */
 386         init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
 387         init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
 388
 389         spin_lock_irq(&kvm->irqfds.lock);
 390
 391         ret = 0;
 392         list_for_each_entry(tmp, &kvm->irqfds.items, list) {
 393                 if (irqfd->eventfd != tmp->eventfd)
 394                         continue;
 395                 /* This fd is used for another irq already. */
 396                 ret = -EBUSY;
 397                 spin_unlock_irq(&kvm->irqfds.lock);
 398                 goto fail;
 399         }
 400
 401         idx = srcu_read_lock(&kvm->irq_srcu);
 402         irqfd_update(kvm, irqfd);
 403
 404         list_add_tail(&irqfd->list, &kvm->irqfds.items);
 405
 406         spin_unlock_irq(&kvm->irqfds.lock);
 407
 408         /*
 409          * Check if there was an event already pending on the eventfd
 410          * before we registered, and trigger it as if we didn't miss it.
 411          */
 412         events = vfs_poll(f.file, &irqfd->pt);
 413
 414         if (events & EPOLLIN)
 415                 schedule_work(&irqfd->inject);
 416
 417 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 418         if (kvm_arch_has_irq_bypass()) {
 419                 irqfd->consumer.token = (void *)irqfd->eventfd;
 420                 irqfd->consumer.add_producer = kvm_arch_irq_bypass_add_producer;
 421                 irqfd->consumer.del_producer = kvm_arch_irq_bypass_del_producer;
 422                 irqfd->consumer.stop = kvm_arch_irq_bypass_stop;
 423                 irqfd->consumer.start = kvm_arch_irq_bypass_start;
 424                 ret = irq_bypass_register_consumer(&irqfd->consumer);
 425                 if (ret)
 426                         pr_info("irq bypass consumer (token %p) registration fails: %d\n",
 427                                 irqfd->consumer.token, ret);
 428         }
 429 #endif
 430
 431         srcu_read_unlock(&kvm->irq_srcu, idx);
 432
 433         /*
 434          * do not drop the file until the irqfd is fully initialized, otherwise
 435          * we might race against the EPOLLHUP
 436          */
 437         fdput(f);
 438         return 0;
 439
 440 fail:
 441         if (irqfd->resampler)
 442                 irqfd_resampler_shutdown(irqfd);
 443
 444         if (resamplefd && !IS_ERR(resamplefd))
 445                 eventfd_ctx_put(resamplefd);
 446
 447         if (eventfd && !IS_ERR(eventfd))
 448                 eventfd_ctx_put(eventfd);
 449
 450         fdput(f);
 451
 452 out:
 453         kfree(irqfd);
 454         return ret;
 455 }
 456
 457 bool kvm_irq_has_notifier(struct kvm *kvm, unsigned irqchip, unsigned pin)
 458 {
 459         struct kvm_irq_ack_notifier *kian;
 460         int gsi, idx;
 461
 462         idx = srcu_read_lock(&kvm->irq_srcu);
 463         gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
 464         if (gsi != -1)
 465                 hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
 466                                          link)
 467                         if (kian->gsi == gsi) {
 468                                 srcu_read_unlock(&kvm->irq_srcu, idx);
 469                                 return true;
 470                         }
 471
 472         srcu_read_unlock(&kvm->irq_srcu, idx);
 473
 474         return false;
 475 }
 476 EXPORT_SYMBOL_GPL(kvm_irq_has_notifier);
 477
 478 void kvm_notify_acked_gsi(struct kvm *kvm, int gsi)
 479 {
 480         struct kvm_irq_ack_notifier *kian;
 481
 482         hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
 483                                  link)
 484                 if (kian->gsi == gsi)
 485                         kian->irq_acked(kian);
 486 }
 487
 488 void kvm_notify_acked_irq(struct kvm *kvm, unsigned irqchip, unsigned pin)
 489 {
 490         int gsi, idx;
 491
 492         trace_kvm_ack_irq(irqchip, pin);
 493
 494         idx = srcu_read_lock(&kvm->irq_srcu);
 495         gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
 496         if (gsi != -1)
 497                 kvm_notify_acked_gsi(kvm, gsi);
 498         srcu_read_unlock(&kvm->irq_srcu, idx);
 499 }
 500
 501 void kvm_register_irq_ack_notifier(struct kvm *kvm,
 502                                    struct kvm_irq_ack_notifier *kian)
 503 {
 504         mutex_lock(&kvm->irq_lock);
 505         hlist_add_head_rcu(&kian->link, &kvm->irq_ack_notifier_list);
 506         mutex_unlock(&kvm->irq_lock);
 507         kvm_arch_post_irq_ack_notifier_list_update(kvm);
 508 }
 509
 510 void kvm_unregister_irq_ack_notifier(struct kvm *kvm,
 511                                     struct kvm_irq_ack_notifier *kian)
 512 {
 513         mutex_lock(&kvm->irq_lock);
 514         hlist_del_init_rcu(&kian->link);
 515         mutex_unlock(&kvm->irq_lock);
 516         synchronize_srcu(&kvm->irq_srcu);
 517         kvm_arch_post_irq_ack_notifier_list_update(kvm);
 518 }
 519 #endif
 520
 521 void
 522 kvm_eventfd_init(struct kvm *kvm)
 523 {
 524 #ifdef CONFIG_HAVE_KVM_IRQFD
 525         spin_lock_init(&kvm->irqfds.lock);
 526         INIT_LIST_HEAD(&kvm->irqfds.items);
 527         INIT_LIST_HEAD(&kvm->irqfds.resampler_list);
 528         mutex_init(&kvm->irqfds.resampler_lock);
 529 #endif
 530         INIT_LIST_HEAD(&kvm->ioeventfds);
 531 }
 532
 533 #ifdef CONFIG_HAVE_KVM_IRQFD
 534 /*
 535  * shutdown any irqfd's that match fd+gsi
 536  */
 537 static int
 538 kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
 539 {
 540         struct kvm_kernel_irqfd *irqfd, *tmp;
 541         struct eventfd_ctx *eventfd;
 542
 543         eventfd = eventfd_ctx_fdget(args->fd);
 544         if (IS_ERR(eventfd))
 545                 return PTR_ERR(eventfd);
 546
 547         spin_lock_irq(&kvm->irqfds.lock);
 548
 549         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
 550                 if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
 551                         /*
 552                          * This clearing of irq_entry.type is needed for when
 553                          * another thread calls kvm_irq_routing_update before
 554                          * we flush workqueue below (we synchronize with
 555                          * kvm_irq_routing_update using irqfds.lock).
 556                          */
 557                         write_seqcount_begin(&irqfd->irq_entry_sc);
 558                         irqfd->irq_entry.type = 0;
 559                         write_seqcount_end(&irqfd->irq_entry_sc);
 560                         irqfd_deactivate(irqfd);
 561                 }
 562         }
 563
 564         spin_unlock_irq(&kvm->irqfds.lock);
 565         eventfd_ctx_put(eventfd);
 566
 567         /*
 568          * Block until we know all outstanding shutdown jobs have completed
 569          * so that we guarantee there will not be any more interrupts on this
 570          * gsi once this deassign function returns.
 571          */
 572         flush_workqueue(irqfd_cleanup_wq);
 573
 574         return 0;
 575 }
 576
 577 int
 578 kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
 579 {
 580         if (args->flags & ~(KVM_IRQFD_FLAG_DEASSIGN | KVM_IRQFD_FLAG_RESAMPLE))
 581                 return -EINVAL;
 582
 583         if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
 584                 return kvm_irqfd_deassign(kvm, args);
 585
 586         return kvm_irqfd_assign(kvm, args);
 587 }
 588
 589 /*
 590  * This function is called as the kvm VM fd is being released. Shutdown all
 591  * irqfds that still remain open
 592  */
 593 void
 594 kvm_irqfd_release(struct kvm *kvm)
 595 {
 596         struct kvm_kernel_irqfd *irqfd, *tmp;
 597
 598         spin_lock_irq(&kvm->irqfds.lock);
 599
 600         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
 601                 irqfd_deactivate(irqfd);
 602
 603         spin_unlock_irq(&kvm->irqfds.lock);
 604
 605         /*
 606          * Block until we know all outstanding shutdown jobs have completed
 607          * since we do not take a kvm* reference.
 608          */
 609         flush_workqueue(irqfd_cleanup_wq);
 610
 611 }
 612
 613 /*
 614  * Take note of a change in irq routing.
 615  * Caller must invoke synchronize_srcu(&kvm->irq_srcu) afterwards.
 616  */
 617 void kvm_irq_routing_update(struct kvm *kvm)
 618 {
 619         struct kvm_kernel_irqfd *irqfd;
 620
 621         spin_lock_irq(&kvm->irqfds.lock);
 622
 623         list_for_each_entry(irqfd, &kvm->irqfds.items, list) {
 624                 irqfd_update(kvm, irqfd);
 625
 626 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 627                 if (irqfd->producer) {
 628                         int ret = kvm_arch_update_irqfd_routing(
 629                                         irqfd->kvm, irqfd->producer->irq,
 630                                         irqfd->gsi, 1);
 631                         WARN_ON(ret);
 632                 }
 633 #endif
 634         }
 635
 636         spin_unlock_irq(&kvm->irqfds.lock);
 637 }
 638
 639 /*
 640  * create a host-wide workqueue for issuing deferred shutdown requests
 641  * aggregated from all vm* instances. We need our own isolated
 642  * queue to ease flushing work items when a VM exits.
 643  */
 644 int kvm_irqfd_init(void)
 645 {
 646         irqfd_cleanup_wq = alloc_workqueue("kvm-irqfd-cleanup", 0, 0);
 647         if (!irqfd_cleanup_wq)
 648                 return -ENOMEM;
 649
 650         return 0;
 651 }
 652
 653 void kvm_irqfd_exit(void)
 654 {
 655         destroy_workqueue(irqfd_cleanup_wq);
 656 }
 657 #endif
 658
 659 /*
 660  * --------------------------------------------------------------------
 661  * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
 662  *
 663  * userspace can register a PIO/MMIO address with an eventfd for receiving
 664  * notification when the memory has been touched.
 665  * --------------------------------------------------------------------
 666  */
 667
 668 struct _ioeventfd {
 669         struct list_head     list;
 670         u64                  addr;
 671         int                  length;
 672         struct eventfd_ctx  *eventfd;
 673         u64                  datamatch;
 674         struct kvm_io_device dev;
 675         u8                   bus_idx;
 676         bool                 wildcard;
 677 };
 678
 679 static inline struct _ioeventfd *
 680 to_ioeventfd(struct kvm_io_device *dev)
 681 {
 682         return container_of(dev, struct _ioeventfd, dev);
 683 }
 684
 685 static void
 686 ioeventfd_release(struct _ioeventfd *p)
 687 {
 688         eventfd_ctx_put(p->eventfd);
 689         list_del(&p->list);
 690         kfree(p);
 691 }
 692
 693 static bool
 694 ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
 695 {
 696         u64 _val;
 697
 698         if (addr != p->addr)
 699                 /* address must be precise for a hit */
 700                 return false;
 701
 702         if (!p->length)
 703                 /* length = 0 means only look at the address, so always a hit */
 704                 return true;
 705
 706         if (len != p->length)
 707                 /* address-range must be precise for a hit */
 708                 return false;
 709
 710         if (p->wildcard)
 711                 /* all else equal, wildcard is always a hit */
 712                 return true;
 713
 714         /* otherwise, we have to actually compare the data */
 715
 716         BUG_ON(!IS_ALIGNED((unsigned long)val, len));
 717
 718         switch (len) {
 719         case 1:
 720                 _val = *(u8 *)val;
 721                 break;
 722         case 2:
 723                 _val = *(u16 *)val;
 724                 break;
 725         case 4:
 726                 _val = *(u32 *)val;
 727                 break;
 728         case 8:
 729                 _val = *(u64 *)val;
 730                 break;
 731         default:
 732                 return false;
 733         }
 734
 735         return _val == p->datamatch ? true : false;
 736 }
 737
 738 /* MMIO/PIO writes trigger an event if the addr/val match */
 739 static int
 740 ioeventfd_write(struct kvm_vcpu *vcpu, struct kvm_io_device *this, gpa_t addr,
 741                 int len, const void *val)
 742 {
 743         struct _ioeventfd *p = to_ioeventfd(this);
 744
 745         if (!ioeventfd_in_range(p, addr, len, val))
 746                 return -EOPNOTSUPP;
 747
 748         eventfd_signal(p->eventfd, 1);
 749         return 0;
 750 }
 751
 752 /*
 753  * This function is called as KVM is completely shutting down.  We do not
 754  * need to worry about locking just nuke anything we have as quickly as possible
 755  */
 756 static void
 757 ioeventfd_destructor(struct kvm_io_device *this)
 758 {
 759         struct _ioeventfd *p = to_ioeventfd(this);
 760
 761         ioeventfd_release(p);
 762 }
 763
 764 static const struct kvm_io_device_ops ioeventfd_ops = {
 765         .write      = ioeventfd_write,
 766         .destructor = ioeventfd_destructor,
 767 };
 768
 769 /* assumes kvm->slots_lock held */
 770 static bool
 771 ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
 772 {
 773         struct _ioeventfd *_p;
 774
 775         list_for_each_entry(_p, &kvm->ioeventfds, list)
 776                 if (_p->bus_idx == p->bus_idx &&
 777                     _p->addr == p->addr &&
 778                     (!_p->length || !p->length ||
 779                      (_p->length == p->length &&
 780                       (_p->wildcard || p->wildcard ||
 781                        _p->datamatch == p->datamatch))))
 782                         return true;
 783
 784         return false;
 785 }
 786
 787 static enum kvm_bus ioeventfd_bus_from_flags(__u32 flags)
 788 {
 789         if (flags & KVM_IOEVENTFD_FLAG_PIO)
 790                 return KVM_PIO_BUS;
 791         if (flags & KVM_IOEVENTFD_FLAG_VIRTIO_CCW_NOTIFY)
 792                 return KVM_VIRTIO_CCW_NOTIFY_BUS;
 793         return KVM_MMIO_BUS;
 794 }
 795
 796 static int kvm_assign_ioeventfd_idx(struct kvm *kvm,
 797                                 enum kvm_bus bus_idx,
 798                                 struct kvm_ioeventfd *args)
 799 {
 800
 801         struct eventfd_ctx *eventfd;
 802         struct _ioeventfd *p;
 803         int ret;
 804
 805         eventfd = eventfd_ctx_fdget(args->fd);
 806         if (IS_ERR(eventfd))
 807                 return PTR_ERR(eventfd);
 808
 809         p = kzalloc(sizeof(*p), GFP_KERNEL);
 810         if (!p) {
 811                 ret = -ENOMEM;
 812                 goto fail;
 813         }
 814
 815         INIT_LIST_HEAD(&p->list);
 816         p->addr    = args->addr;
 817         p->bus_idx = bus_idx;
 818         p->length  = args->len;
 819         p->eventfd = eventfd;
 820
 821         /* The datamatch feature is optional, otherwise this is a wildcard */
 822         if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
 823                 p->datamatch = args->datamatch;
 824         else
 825                 p->wildcard = true;
 826
 827         mutex_lock(&kvm->slots_lock);
 828
 829         /* Verify that there isn't a match already */
 830         if (ioeventfd_check_collision(kvm, p)) {
 831                 ret = -EEXIST;
 832                 goto unlock_fail;
 833         }
 834
 835         kvm_iodevice_init(&p->dev, &ioeventfd_ops);
 836
 837         ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
 838                                       &p->dev);
 839         if (ret < 0)
 840                 goto unlock_fail;
 841
 842         kvm_get_bus(kvm, bus_idx)->ioeventfd_count++;
 843         list_add_tail(&p->list, &kvm->ioeventfds);
 844
 845         mutex_unlock(&kvm->slots_lock);
 846
 847         return 0;
 848
 849 unlock_fail:
 850         mutex_unlock(&kvm->slots_lock);
 851
 852 fail:
 853         kfree(p);
 854         eventfd_ctx_put(eventfd);
 855
 856         return ret;
 857 }
 858
 859 static int
 860 kvm_deassign_ioeventfd_idx(struct kvm *kvm, enum kvm_bus bus_idx,
 861                            struct kvm_ioeventfd *args)
 862 {
 863         struct _ioeventfd        *p, *tmp;
 864         struct eventfd_ctx       *eventfd;
 865         struct kvm_io_bus        *bus;
 866         int                       ret = -ENOENT;
 867
 868         eventfd = eventfd_ctx_fdget(args->fd);
 869         if (IS_ERR(eventfd))
 870                 return PTR_ERR(eventfd);
 871
 872         mutex_lock(&kvm->slots_lock);
 873
 874         list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
 875                 bool wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
 876
 877                 if (p->bus_idx != bus_idx ||
 878                     p->eventfd != eventfd  ||
 879                     p->addr != args->addr  ||
 880                     p->length != args->len ||
 881                     p->wildcard != wildcard)
 882                         continue;
 883
 884                 if (!p->wildcard && p->datamatch != args->datamatch)
 885                         continue;
 886
 887                 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
 888                 bus = kvm_get_bus(kvm, bus_idx);
 889                 if (bus)
 890                         bus->ioeventfd_count--;
 891                 ioeventfd_release(p);
 892                 ret = 0;
 893                 break;
 894         }
 895
 896         mutex_unlock(&kvm->slots_lock);
 897
 898         eventfd_ctx_put(eventfd);
 899
 900         return ret;
 901 }
 902
 903 static int kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 904 {
 905         enum kvm_bus bus_idx = ioeventfd_bus_from_flags(args->flags);
 906         int ret = kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
 907
 908         if (!args->len && bus_idx == KVM_MMIO_BUS)
 909                 kvm_deassign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
 910
 911         return ret;
 912 }
 913
 914 static int
 915 kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 916 {
 917         enum kvm_bus              bus_idx;
 918         int ret;
 919
 920         bus_idx = ioeventfd_bus_from_flags(args->flags);
 921         /* must be natural-word sized, or 0 to ignore length */
 922         switch (args->len) {
 923         case 0:
 924         case 1:
 925         case 2:
 926         case 4:
 927         case 8:
 928                 break;
 929         default:
 930                 return -EINVAL;
 931         }
 932
 933         /* check for range overflow */
 934         if (args->addr + args->len < args->addr)
 935                 return -EINVAL;
 936
 937         /* check for extra flags that we don't understand */
 938         if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
 939                 return -EINVAL;
 940
 941         /* ioeventfd with no length can't be combined with DATAMATCH */
 942         if (!args->len && (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH))
 943                 return -EINVAL;
 944
 945         ret = kvm_assign_ioeventfd_idx(kvm, bus_idx, args);
 946         if (ret)
 947                 goto fail;
 948
 949         /* When length is ignored, MMIO is also put on a separate bus, for
 950          * faster lookups.
 951          */
 952         if (!args->len && bus_idx == KVM_MMIO_BUS) {
 953                 ret = kvm_assign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
 954                 if (ret < 0)
 955                         goto fast_fail;
 956         }
 957
 958         return 0;
 959
 960 fast_fail:
 961         kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
 962 fail:
 963         return ret;
 964 }
 965
 966 int
 967 kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 968 {
 969         if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
 970                 return kvm_deassign_ioeventfd(kvm, args);
 971
 972         return kvm_assign_ioeventfd(kvm, args);
 973 }