Pileus Git - ~andy/linux/blob - 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/workqueue.h>
  27 #include <linux/syscalls.h>
  28 #include <linux/wait.h>
  29 #include <linux/poll.h>
  30 #include <linux/file.h>
  31 #include <linux/list.h>
  32 #include <linux/eventfd.h>
  33 #include <linux/kernel.h>
  34 #include <linux/slab.h>
  35
  36 #include "iodev.h"
  37
  38 #ifdef __KVM_HAVE_IOAPIC
  39 /*
  40  * --------------------------------------------------------------------
  41  * irqfd: Allows an fd to be used to inject an interrupt to the guest
  42  *
  43  * Credit goes to Avi Kivity for the original idea.
  44  * --------------------------------------------------------------------
  45  */
  46
  47 /*
  48  * Resampling irqfds are a special variety of irqfds used to emulate
  49  * level triggered interrupts.  The interrupt is asserted on eventfd
  50  * trigger.  On acknowledgement through the irq ack notifier, the
  51  * interrupt is de-asserted and userspace is notified through the
  52  * resamplefd.  All resamplers on the same gsi are de-asserted
  53  * together, so we don't need to track the state of each individual
  54  * user.  We can also therefore share the same irq source ID.
  55  */
  56 struct _irqfd_resampler {
  57         struct kvm *kvm;
  58         /*
  59          * List of resampling struct _irqfd objects sharing this gsi.
  60          * RCU list modified under kvm->irqfds.resampler_lock
  61          */
  62         struct list_head list;
  63         struct kvm_irq_ack_notifier notifier;
  64         /*
  65          * Entry in list of kvm->irqfd.resampler_list.  Use for sharing
  66          * resamplers among irqfds on the same gsi.
  67          * Accessed and modified under kvm->irqfds.resampler_lock
  68          */
  69         struct list_head link;
  70 };
  71
  72 struct _irqfd {
  73         /* Used for MSI fast-path */
  74         struct kvm *kvm;
  75         wait_queue_t wait;
  76         /* Update side is protected by irqfds.lock */
  77         struct kvm_kernel_irq_routing_entry __rcu *irq_entry;
  78         /* Used for level IRQ fast-path */
  79         int gsi;
  80         struct work_struct inject;
  81         /* The resampler used by this irqfd (resampler-only) */
  82         struct _irqfd_resampler *resampler;
  83         /* Eventfd notified on resample (resampler-only) */
  84         struct eventfd_ctx *resamplefd;
  85         /* Entry in list of irqfds for a resampler (resampler-only) */
  86         struct list_head resampler_link;
  87         /* Used for setup/shutdown */
  88         struct eventfd_ctx *eventfd;
  89         struct list_head list;
  90         poll_table pt;
  91         struct work_struct shutdown;
  92 };
  93
  94 static struct workqueue_struct *irqfd_cleanup_wq;
  95
  96 static void
  97 irqfd_inject(struct work_struct *work)
  98 {
  99         struct _irqfd *irqfd = container_of(work, struct _irqfd, inject);
 100         struct kvm *kvm = irqfd->kvm;
 101
 102         if (!irqfd->resampler) {
 103                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1);
 104                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0);
 105         } else
 106                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 107                             irqfd->gsi, 1);
 108 }
 109
 110 /*
 111  * Since resampler irqfds share an IRQ source ID, we de-assert once
 112  * then notify all of the resampler irqfds using this GSI.  We can't
 113  * do multiple de-asserts or we risk racing with incoming re-asserts.
 114  */
 115 static void
 116 irqfd_resampler_ack(struct kvm_irq_ack_notifier *kian)
 117 {
 118         struct _irqfd_resampler *resampler;
 119         struct _irqfd *irqfd;
 120
 121         resampler = container_of(kian, struct _irqfd_resampler, notifier);
 122
 123         kvm_set_irq(resampler->kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 124                     resampler->notifier.gsi, 0);
 125
 126         rcu_read_lock();
 127
 128         list_for_each_entry_rcu(irqfd, &resampler->list, resampler_link)
 129                 eventfd_signal(irqfd->resamplefd, 1);
 130
 131         rcu_read_unlock();
 132 }
 133
 134 static void
 135 irqfd_resampler_shutdown(struct _irqfd *irqfd)
 136 {
 137         struct _irqfd_resampler *resampler = irqfd->resampler;
 138         struct kvm *kvm = resampler->kvm;
 139
 140         mutex_lock(&kvm->irqfds.resampler_lock);
 141
 142         list_del_rcu(&irqfd->resampler_link);
 143         synchronize_rcu();
 144
 145         if (list_empty(&resampler->list)) {
 146                 list_del(&resampler->link);
 147                 kvm_unregister_irq_ack_notifier(kvm, &resampler->notifier);
 148                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 149                             resampler->notifier.gsi, 0);
 150                 kfree(resampler);
 151         }
 152
 153         mutex_unlock(&kvm->irqfds.resampler_lock);
 154 }
 155
 156 /*
 157  * Race-free decouple logic (ordering is critical)
 158  */
 159 static void
 160 irqfd_shutdown(struct work_struct *work)
 161 {
 162         struct _irqfd *irqfd = container_of(work, struct _irqfd, shutdown);
 163         u64 cnt;
 164
 165         /*
 166          * Synchronize with the wait-queue and unhook ourselves to prevent
 167          * further events.
 168          */
 169         eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
 170
 171         /*
 172          * We know no new events will be scheduled at this point, so block
 173          * until all previously outstanding events have completed
 174          */
 175         flush_work(&irqfd->inject);
 176
 177         if (irqfd->resampler) {
 178                 irqfd_resampler_shutdown(irqfd);
 179                 eventfd_ctx_put(irqfd->resamplefd);
 180         }
 181
 182         /*
 183          * It is now safe to release the object's resources
 184          */
 185         eventfd_ctx_put(irqfd->eventfd);
 186         kfree(irqfd);
 187 }
 188
 189
 190 /* assumes kvm->irqfds.lock is held */
 191 static bool
 192 irqfd_is_active(struct _irqfd *irqfd)
 193 {
 194         return list_empty(&irqfd->list) ? false : true;
 195 }
 196
 197 /*
 198  * Mark the irqfd as inactive and schedule it for removal
 199  *
 200  * assumes kvm->irqfds.lock is held
 201  */
 202 static void
 203 irqfd_deactivate(struct _irqfd *irqfd)
 204 {
 205         BUG_ON(!irqfd_is_active(irqfd));
 206
 207         list_del_init(&irqfd->list);
 208
 209         queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
 210 }
 211
 212 /*
 213  * Called with wqh->lock held and interrupts disabled
 214  */
 215 static int
 216 irqfd_wakeup(wait_queue_t *wait, unsigned mode, int sync, void *key)
 217 {
 218         struct _irqfd *irqfd = container_of(wait, struct _irqfd, wait);
 219         unsigned long flags = (unsigned long)key;
 220         struct kvm_kernel_irq_routing_entry *irq;
 221         struct kvm *kvm = irqfd->kvm;
 222
 223         if (flags & POLLIN) {
 224                 rcu_read_lock();
 225                 irq = rcu_dereference(irqfd->irq_entry);
 226                 /* An event has been signaled, inject an interrupt */
 227                 if (irq)
 228                         kvm_set_msi(irq, kvm, KVM_USERSPACE_IRQ_SOURCE_ID, 1);
 229                 else
 230                         schedule_work(&irqfd->inject);
 231                 rcu_read_unlock();
 232         }
 233
 234         if (flags & POLLHUP) {
 235                 /* The eventfd is closing, detach from KVM */
 236                 unsigned long flags;
 237
 238                 spin_lock_irqsave(&kvm->irqfds.lock, flags);
 239
 240                 /*
 241                  * We must check if someone deactivated the irqfd before
 242                  * we could acquire the irqfds.lock since the item is
 243                  * deactivated from the KVM side before it is unhooked from
 244                  * the wait-queue.  If it is already deactivated, we can
 245                  * simply return knowing the other side will cleanup for us.
 246                  * We cannot race against the irqfd going away since the
 247                  * other side is required to acquire wqh->lock, which we hold
 248                  */
 249                 if (irqfd_is_active(irqfd))
 250                         irqfd_deactivate(irqfd);
 251
 252                 spin_unlock_irqrestore(&kvm->irqfds.lock, flags);
 253         }
 254
 255         return 0;
 256 }
 257
 258 static void
 259 irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
 260                         poll_table *pt)
 261 {
 262         struct _irqfd *irqfd = container_of(pt, struct _irqfd, pt);
 263         add_wait_queue(wqh, &irqfd->wait);
 264 }
 265
 266 /* Must be called under irqfds.lock */
 267 static void irqfd_update(struct kvm *kvm, struct _irqfd *irqfd,
 268                          struct kvm_irq_routing_table *irq_rt)
 269 {
 270         struct kvm_kernel_irq_routing_entry *e;
 271
 272         if (irqfd->gsi >= irq_rt->nr_rt_entries) {
 273                 rcu_assign_pointer(irqfd->irq_entry, NULL);
 274                 return;
 275         }
 276
 277         hlist_for_each_entry(e, &irq_rt->map[irqfd->gsi], link) {
 278                 /* Only fast-path MSI. */
 279                 if (e->type == KVM_IRQ_ROUTING_MSI)
 280                         rcu_assign_pointer(irqfd->irq_entry, e);
 281                 else
 282                         rcu_assign_pointer(irqfd->irq_entry, NULL);
 283         }
 284 }
 285
 286 static int
 287 kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
 288 {
 289         struct kvm_irq_routing_table *irq_rt;
 290         struct _irqfd *irqfd, *tmp;
 291         struct file *file = NULL;
 292         struct eventfd_ctx *eventfd = NULL, *resamplefd = NULL;
 293         int ret;
 294         unsigned int events;
 295
 296         irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL);
 297         if (!irqfd)
 298                 return -ENOMEM;
 299
 300         irqfd->kvm = kvm;
 301         irqfd->gsi = args->gsi;
 302         INIT_LIST_HEAD(&irqfd->list);
 303         INIT_WORK(&irqfd->inject, irqfd_inject);
 304         INIT_WORK(&irqfd->shutdown, irqfd_shutdown);
 305
 306         file = eventfd_fget(args->fd);
 307         if (IS_ERR(file)) {
 308                 ret = PTR_ERR(file);
 309                 goto fail;
 310         }
 311
 312         eventfd = eventfd_ctx_fileget(file);
 313         if (IS_ERR(eventfd)) {
 314                 ret = PTR_ERR(eventfd);
 315                 goto fail;
 316         }
 317
 318         irqfd->eventfd = eventfd;
 319
 320         if (args->flags & KVM_IRQFD_FLAG_RESAMPLE) {
 321                 struct _irqfd_resampler *resampler;
 322
 323                 resamplefd = eventfd_ctx_fdget(args->resamplefd);
 324                 if (IS_ERR(resamplefd)) {
 325                         ret = PTR_ERR(resamplefd);
 326                         goto fail;
 327                 }
 328
 329                 irqfd->resamplefd = resamplefd;
 330                 INIT_LIST_HEAD(&irqfd->resampler_link);
 331
 332                 mutex_lock(&kvm->irqfds.resampler_lock);
 333
 334                 list_for_each_entry(resampler,
 335                                     &kvm->irqfds.resampler_list, link) {
 336                         if (resampler->notifier.gsi == irqfd->gsi) {
 337                                 irqfd->resampler = resampler;
 338                                 break;
 339                         }
 340                 }
 341
 342                 if (!irqfd->resampler) {
 343                         resampler = kzalloc(sizeof(*resampler), GFP_KERNEL);
 344                         if (!resampler) {
 345                                 ret = -ENOMEM;
 346                                 mutex_unlock(&kvm->irqfds.resampler_lock);
 347                                 goto fail;
 348                         }
 349
 350                         resampler->kvm = kvm;
 351                         INIT_LIST_HEAD(&resampler->list);
 352                         resampler->notifier.gsi = irqfd->gsi;
 353                         resampler->notifier.irq_acked = irqfd_resampler_ack;
 354                         INIT_LIST_HEAD(&resampler->link);
 355
 356                         list_add(&resampler->link, &kvm->irqfds.resampler_list);
 357                         kvm_register_irq_ack_notifier(kvm,
 358                                                       &resampler->notifier);
 359                         irqfd->resampler = resampler;
 360                 }
 361
 362                 list_add_rcu(&irqfd->resampler_link, &irqfd->resampler->list);
 363                 synchronize_rcu();
 364
 365                 mutex_unlock(&kvm->irqfds.resampler_lock);
 366         }
 367
 368         /*
 369          * Install our own custom wake-up handling so we are notified via
 370          * a callback whenever someone signals the underlying eventfd
 371          */
 372         init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
 373         init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
 374
 375         spin_lock_irq(&kvm->irqfds.lock);
 376
 377         ret = 0;
 378         list_for_each_entry(tmp, &kvm->irqfds.items, list) {
 379                 if (irqfd->eventfd != tmp->eventfd)
 380                         continue;
 381                 /* This fd is used for another irq already. */
 382                 ret = -EBUSY;
 383                 spin_unlock_irq(&kvm->irqfds.lock);
 384                 goto fail;
 385         }
 386
 387         irq_rt = rcu_dereference_protected(kvm->irq_routing,
 388                                            lockdep_is_held(&kvm->irqfds.lock));
 389         irqfd_update(kvm, irqfd, irq_rt);
 390
 391         events = file->f_op->poll(file, &irqfd->pt);
 392
 393         list_add_tail(&irqfd->list, &kvm->irqfds.items);
 394
 395         /*
 396          * Check if there was an event already pending on the eventfd
 397          * before we registered, and trigger it as if we didn't miss it.
 398          */
 399         if (events & POLLIN)
 400                 schedule_work(&irqfd->inject);
 401
 402         spin_unlock_irq(&kvm->irqfds.lock);
 403
 404         /*
 405          * do not drop the file until the irqfd is fully initialized, otherwise
 406          * we might race against the POLLHUP
 407          */
 408         fput(file);
 409
 410         return 0;
 411
 412 fail:
 413         if (irqfd->resampler)
 414                 irqfd_resampler_shutdown(irqfd);
 415
 416         if (resamplefd && !IS_ERR(resamplefd))
 417                 eventfd_ctx_put(resamplefd);
 418
 419         if (eventfd && !IS_ERR(eventfd))
 420                 eventfd_ctx_put(eventfd);
 421
 422         if (!IS_ERR(file))
 423                 fput(file);
 424
 425         kfree(irqfd);
 426         return ret;
 427 }
 428 #endif
 429
 430 void
 431 kvm_eventfd_init(struct kvm *kvm)
 432 {
 433 #ifdef __KVM_HAVE_IOAPIC
 434         spin_lock_init(&kvm->irqfds.lock);
 435         INIT_LIST_HEAD(&kvm->irqfds.items);
 436         INIT_LIST_HEAD(&kvm->irqfds.resampler_list);
 437         mutex_init(&kvm->irqfds.resampler_lock);
 438 #endif
 439         INIT_LIST_HEAD(&kvm->ioeventfds);
 440 }
 441
 442 #ifdef __KVM_HAVE_IOAPIC
 443 /*
 444  * shutdown any irqfd's that match fd+gsi
 445  */
 446 static int
 447 kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
 448 {
 449         struct _irqfd *irqfd, *tmp;
 450         struct eventfd_ctx *eventfd;
 451
 452         eventfd = eventfd_ctx_fdget(args->fd);
 453         if (IS_ERR(eventfd))
 454                 return PTR_ERR(eventfd);
 455
 456         spin_lock_irq(&kvm->irqfds.lock);
 457
 458         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
 459                 if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
 460                         /*
 461                          * This rcu_assign_pointer is needed for when
 462                          * another thread calls kvm_irq_routing_update before
 463                          * we flush workqueue below (we synchronize with
 464                          * kvm_irq_routing_update using irqfds.lock).
 465                          * It is paired with synchronize_rcu done by caller
 466                          * of that function.
 467                          */
 468                         rcu_assign_pointer(irqfd->irq_entry, NULL);
 469                         irqfd_deactivate(irqfd);
 470                 }
 471         }
 472
 473         spin_unlock_irq(&kvm->irqfds.lock);
 474         eventfd_ctx_put(eventfd);
 475
 476         /*
 477          * Block until we know all outstanding shutdown jobs have completed
 478          * so that we guarantee there will not be any more interrupts on this
 479          * gsi once this deassign function returns.
 480          */
 481         flush_workqueue(irqfd_cleanup_wq);
 482
 483         return 0;
 484 }
 485
 486 int
 487 kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
 488 {
 489         if (args->flags & ~(KVM_IRQFD_FLAG_DEASSIGN | KVM_IRQFD_FLAG_RESAMPLE))
 490                 return -EINVAL;
 491
 492         if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
 493                 return kvm_irqfd_deassign(kvm, args);
 494
 495         return kvm_irqfd_assign(kvm, args);
 496 }
 497
 498 /*
 499  * This function is called as the kvm VM fd is being released. Shutdown all
 500  * irqfds that still remain open
 501  */
 502 void
 503 kvm_irqfd_release(struct kvm *kvm)
 504 {
 505         struct _irqfd *irqfd, *tmp;
 506
 507         spin_lock_irq(&kvm->irqfds.lock);
 508
 509         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
 510                 irqfd_deactivate(irqfd);
 511
 512         spin_unlock_irq(&kvm->irqfds.lock);
 513
 514         /*
 515          * Block until we know all outstanding shutdown jobs have completed
 516          * since we do not take a kvm* reference.
 517          */
 518         flush_workqueue(irqfd_cleanup_wq);
 519
 520 }
 521
 522 /*
 523  * Change irq_routing and irqfd.
 524  * Caller must invoke synchronize_rcu afterwards.
 525  */
 526 void kvm_irq_routing_update(struct kvm *kvm,
 527                             struct kvm_irq_routing_table *irq_rt)
 528 {
 529         struct _irqfd *irqfd;
 530
 531         spin_lock_irq(&kvm->irqfds.lock);
 532
 533         rcu_assign_pointer(kvm->irq_routing, irq_rt);
 534
 535         list_for_each_entry(irqfd, &kvm->irqfds.items, list)
 536                 irqfd_update(kvm, irqfd, irq_rt);
 537
 538         spin_unlock_irq(&kvm->irqfds.lock);
 539 }
 540
 541 /*
 542  * create a host-wide workqueue for issuing deferred shutdown requests
 543  * aggregated from all vm* instances. We need our own isolated single-thread
 544  * queue to prevent deadlock against flushing the normal work-queue.
 545  */
 546 int kvm_irqfd_init(void)
 547 {
 548         irqfd_cleanup_wq = create_singlethread_workqueue("kvm-irqfd-cleanup");
 549         if (!irqfd_cleanup_wq)
 550                 return -ENOMEM;
 551
 552         return 0;
 553 }
 554
 555 void kvm_irqfd_exit(void)
 556 {
 557         destroy_workqueue(irqfd_cleanup_wq);
 558 }
 559 #endif
 560
 561 /*
 562  * --------------------------------------------------------------------
 563  * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
 564  *
 565  * userspace can register a PIO/MMIO address with an eventfd for receiving
 566  * notification when the memory has been touched.
 567  * --------------------------------------------------------------------
 568  */
 569
 570 struct _ioeventfd {
 571         struct list_head     list;
 572         u64                  addr;
 573         int                  length;
 574         struct eventfd_ctx  *eventfd;
 575         u64                  datamatch;
 576         struct kvm_io_device dev;
 577         bool                 wildcard;
 578 };
 579
 580 static inline struct _ioeventfd *
 581 to_ioeventfd(struct kvm_io_device *dev)
 582 {
 583         return container_of(dev, struct _ioeventfd, dev);
 584 }
 585
 586 static void
 587 ioeventfd_release(struct _ioeventfd *p)
 588 {
 589         eventfd_ctx_put(p->eventfd);
 590         list_del(&p->list);
 591         kfree(p);
 592 }
 593
 594 static bool
 595 ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
 596 {
 597         u64 _val;
 598
 599         if (!(addr == p->addr && len == p->length))
 600                 /* address-range must be precise for a hit */
 601                 return false;
 602
 603         if (p->wildcard)
 604                 /* all else equal, wildcard is always a hit */
 605                 return true;
 606
 607         /* otherwise, we have to actually compare the data */
 608
 609         BUG_ON(!IS_ALIGNED((unsigned long)val, len));
 610
 611         switch (len) {
 612         case 1:
 613                 _val = *(u8 *)val;
 614                 break;
 615         case 2:
 616                 _val = *(u16 *)val;
 617                 break;
 618         case 4:
 619                 _val = *(u32 *)val;
 620                 break;
 621         case 8:
 622                 _val = *(u64 *)val;
 623                 break;
 624         default:
 625                 return false;
 626         }
 627
 628         return _val == p->datamatch ? true : false;
 629 }
 630
 631 /* MMIO/PIO writes trigger an event if the addr/val match */
 632 static int
 633 ioeventfd_write(struct kvm_io_device *this, gpa_t addr, int len,
 634                 const void *val)
 635 {
 636         struct _ioeventfd *p = to_ioeventfd(this);
 637
 638         if (!ioeventfd_in_range(p, addr, len, val))
 639                 return -EOPNOTSUPP;
 640
 641         eventfd_signal(p->eventfd, 1);
 642         return 0;
 643 }
 644
 645 /*
 646  * This function is called as KVM is completely shutting down.  We do not
 647  * need to worry about locking just nuke anything we have as quickly as possible
 648  */
 649 static void
 650 ioeventfd_destructor(struct kvm_io_device *this)
 651 {
 652         struct _ioeventfd *p = to_ioeventfd(this);
 653
 654         ioeventfd_release(p);
 655 }
 656
 657 static const struct kvm_io_device_ops ioeventfd_ops = {
 658         .write      = ioeventfd_write,
 659         .destructor = ioeventfd_destructor,
 660 };
 661
 662 /* assumes kvm->slots_lock held */
 663 static bool
 664 ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
 665 {
 666         struct _ioeventfd *_p;
 667
 668         list_for_each_entry(_p, &kvm->ioeventfds, list)
 669                 if (_p->addr == p->addr && _p->length == p->length &&
 670                     (_p->wildcard || p->wildcard ||
 671                      _p->datamatch == p->datamatch))
 672                         return true;
 673
 674         return false;
 675 }
 676
 677 static enum kvm_bus ioeventfd_bus_from_flags(__u32 flags)
 678 {
 679         if (flags & KVM_IOEVENTFD_FLAG_PIO)
 680                 return KVM_PIO_BUS;
 681         if (flags & KVM_IOEVENTFD_FLAG_VIRTIO_CCW_NOTIFY)
 682                 return KVM_VIRTIO_CCW_NOTIFY_BUS;
 683         return KVM_MMIO_BUS;
 684 }
 685
 686 static int
 687 kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 688 {
 689         enum kvm_bus              bus_idx;
 690         struct _ioeventfd        *p;
 691         struct eventfd_ctx       *eventfd;
 692         int                       ret;
 693
 694         bus_idx = ioeventfd_bus_from_flags(args->flags);
 695         /* must be natural-word sized */
 696         switch (args->len) {
 697         case 1:
 698         case 2:
 699         case 4:
 700         case 8:
 701                 break;
 702         default:
 703                 return -EINVAL;
 704         }
 705
 706         /* check for range overflow */
 707         if (args->addr + args->len < args->addr)
 708                 return -EINVAL;
 709
 710         /* check for extra flags that we don't understand */
 711         if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
 712                 return -EINVAL;
 713
 714         eventfd = eventfd_ctx_fdget(args->fd);
 715         if (IS_ERR(eventfd))
 716                 return PTR_ERR(eventfd);
 717
 718         p = kzalloc(sizeof(*p), GFP_KERNEL);
 719         if (!p) {
 720                 ret = -ENOMEM;
 721                 goto fail;
 722         }
 723
 724         INIT_LIST_HEAD(&p->list);
 725         p->addr    = args->addr;
 726         p->length  = args->len;
 727         p->eventfd = eventfd;
 728
 729         /* The datamatch feature is optional, otherwise this is a wildcard */
 730         if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
 731                 p->datamatch = args->datamatch;
 732         else
 733                 p->wildcard = true;
 734
 735         mutex_lock(&kvm->slots_lock);
 736
 737         /* Verify that there isn't a match already */
 738         if (ioeventfd_check_collision(kvm, p)) {
 739                 ret = -EEXIST;
 740                 goto unlock_fail;
 741         }
 742
 743         kvm_iodevice_init(&p->dev, &ioeventfd_ops);
 744
 745         ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
 746                                       &p->dev);
 747         if (ret < 0)
 748                 goto unlock_fail;
 749
 750         list_add_tail(&p->list, &kvm->ioeventfds);
 751
 752         mutex_unlock(&kvm->slots_lock);
 753
 754         return 0;
 755
 756 unlock_fail:
 757         mutex_unlock(&kvm->slots_lock);
 758
 759 fail:
 760         kfree(p);
 761         eventfd_ctx_put(eventfd);
 762
 763         return ret;
 764 }
 765
 766 static int
 767 kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 768 {
 769         enum kvm_bus              bus_idx;
 770         struct _ioeventfd        *p, *tmp;
 771         struct eventfd_ctx       *eventfd;
 772         int                       ret = -ENOENT;
 773
 774         bus_idx = ioeventfd_bus_from_flags(args->flags);
 775         eventfd = eventfd_ctx_fdget(args->fd);
 776         if (IS_ERR(eventfd))
 777                 return PTR_ERR(eventfd);
 778
 779         mutex_lock(&kvm->slots_lock);
 780
 781         list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
 782                 bool wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
 783
 784                 if (p->eventfd != eventfd  ||
 785                     p->addr != args->addr  ||
 786                     p->length != args->len ||
 787                     p->wildcard != wildcard)
 788                         continue;
 789
 790                 if (!p->wildcard && p->datamatch != args->datamatch)
 791                         continue;
 792
 793                 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
 794                 ioeventfd_release(p);
 795                 ret = 0;
 796                 break;
 797         }
 798
 799         mutex_unlock(&kvm->slots_lock);
 800
 801         eventfd_ctx_put(eventfd);
 802
 803         return ret;
 804 }
 805
 806 int
 807 kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 808 {
 809         if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
 810                 return kvm_deassign_ioeventfd(kvm, args);
 811
 812         return kvm_assign_ioeventfd(kvm, args);
 813 }