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/srcu.h>
  35 #include <linux/slab.h>
  36 #include <linux/seqlock.h>
  37
  38 #include "iodev.h"
  39
  40 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
  41 /*
  42  * --------------------------------------------------------------------
  43  * irqfd: Allows an fd to be used to inject an interrupt to the guest
  44  *
  45  * Credit goes to Avi Kivity for the original idea.
  46  * --------------------------------------------------------------------
  47  */
  48
  49 /*
  50  * Resampling irqfds are a special variety of irqfds used to emulate
  51  * level triggered interrupts.  The interrupt is asserted on eventfd
  52  * trigger.  On acknowledgement through the irq ack notifier, the
  53  * interrupt is de-asserted and userspace is notified through the
  54  * resamplefd.  All resamplers on the same gsi are de-asserted
  55  * together, so we don't need to track the state of each individual
  56  * user.  We can also therefore share the same irq source ID.
  57  */
  58 struct _irqfd_resampler {
  59         struct kvm *kvm;
  60         /*
  61          * List of resampling struct _irqfd objects sharing this gsi.
  62          * RCU list modified under kvm->irqfds.resampler_lock
  63          */
  64         struct list_head list;
  65         struct kvm_irq_ack_notifier notifier;
  66         /*
  67          * Entry in list of kvm->irqfd.resampler_list.  Use for sharing
  68          * resamplers among irqfds on the same gsi.
  69          * Accessed and modified under kvm->irqfds.resampler_lock
  70          */
  71         struct list_head link;
  72 };
  73
  74 struct _irqfd {
  75         /* Used for MSI fast-path */
  76         struct kvm *kvm;
  77         wait_queue_t wait;
  78         /* Update side is protected by irqfds.lock */
  79         struct kvm_kernel_irq_routing_entry irq_entry;
  80         seqcount_t irq_entry_sc;
  81         /* Used for level IRQ fast-path */
  82         int gsi;
  83         struct work_struct inject;
  84         /* The resampler used by this irqfd (resampler-only) */
  85         struct _irqfd_resampler *resampler;
  86         /* Eventfd notified on resample (resampler-only) */
  87         struct eventfd_ctx *resamplefd;
  88         /* Entry in list of irqfds for a resampler (resampler-only) */
  89         struct list_head resampler_link;
  90         /* Used for setup/shutdown */
  91         struct eventfd_ctx *eventfd;
  92         struct list_head list;
  93         poll_table pt;
  94         struct work_struct shutdown;
  95 };
  96
  97 static struct workqueue_struct *irqfd_cleanup_wq;
  98
  99 static void
 100 irqfd_inject(struct work_struct *work)
 101 {
 102         struct _irqfd *irqfd = container_of(work, struct _irqfd, inject);
 103         struct kvm *kvm = irqfd->kvm;
 104
 105         if (!irqfd->resampler) {
 106                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1,
 107                                 false);
 108                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0,
 109                                 false);
 110         } else
 111                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 112                             irqfd->gsi, 1, false);
 113 }
 114
 115 /*
 116  * Since resampler irqfds share an IRQ source ID, we de-assert once
 117  * then notify all of the resampler irqfds using this GSI.  We can't
 118  * do multiple de-asserts or we risk racing with incoming re-asserts.
 119  */
 120 static void
 121 irqfd_resampler_ack(struct kvm_irq_ack_notifier *kian)
 122 {
 123         struct _irqfd_resampler *resampler;
 124         struct kvm *kvm;
 125         struct _irqfd *irqfd;
 126         int idx;
 127
 128         resampler = container_of(kian, struct _irqfd_resampler, notifier);
 129         kvm = resampler->kvm;
 130
 131         kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 132                     resampler->notifier.gsi, 0, false);
 133
 134         idx = srcu_read_lock(&kvm->irq_srcu);
 135
 136         list_for_each_entry_rcu(irqfd, &resampler->list, resampler_link)
 137                 eventfd_signal(irqfd->resamplefd, 1);
 138
 139         srcu_read_unlock(&kvm->irq_srcu, idx);
 140 }
 141
 142 static void
 143 irqfd_resampler_shutdown(struct _irqfd *irqfd)
 144 {
 145         struct _irqfd_resampler *resampler = irqfd->resampler;
 146         struct kvm *kvm = resampler->kvm;
 147
 148         mutex_lock(&kvm->irqfds.resampler_lock);
 149
 150         list_del_rcu(&irqfd->resampler_link);
 151         synchronize_srcu(&kvm->irq_srcu);
 152
 153         if (list_empty(&resampler->list)) {
 154                 list_del(&resampler->link);
 155                 kvm_unregister_irq_ack_notifier(kvm, &resampler->notifier);
 156                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 157                             resampler->notifier.gsi, 0, false);
 158                 kfree(resampler);
 159         }
 160
 161         mutex_unlock(&kvm->irqfds.resampler_lock);
 162 }
 163
 164 /*
 165  * Race-free decouple logic (ordering is critical)
 166  */
 167 static void
 168 irqfd_shutdown(struct work_struct *work)
 169 {
 170         struct _irqfd *irqfd = container_of(work, struct _irqfd, shutdown);
 171         u64 cnt;
 172
 173         /*
 174          * Synchronize with the wait-queue and unhook ourselves to prevent
 175          * further events.
 176          */
 177         eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
 178
 179         /*
 180          * We know no new events will be scheduled at this point, so block
 181          * until all previously outstanding events have completed
 182          */
 183         flush_work(&irqfd->inject);
 184
 185         if (irqfd->resampler) {
 186                 irqfd_resampler_shutdown(irqfd);
 187                 eventfd_ctx_put(irqfd->resamplefd);
 188         }
 189
 190         /*
 191          * It is now safe to release the object's resources
 192          */
 193         eventfd_ctx_put(irqfd->eventfd);
 194         kfree(irqfd);
 195 }
 196
 197
 198 /* assumes kvm->irqfds.lock is held */
 199 static bool
 200 irqfd_is_active(struct _irqfd *irqfd)
 201 {
 202         return list_empty(&irqfd->list) ? false : true;
 203 }
 204
 205 /*
 206  * Mark the irqfd as inactive and schedule it for removal
 207  *
 208  * assumes kvm->irqfds.lock is held
 209  */
 210 static void
 211 irqfd_deactivate(struct _irqfd *irqfd)
 212 {
 213         BUG_ON(!irqfd_is_active(irqfd));
 214
 215         list_del_init(&irqfd->list);
 216
 217         queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
 218 }
 219
 220 /*
 221  * Called with wqh->lock held and interrupts disabled
 222  */
 223 static int
 224 irqfd_wakeup(wait_queue_t *wait, unsigned mode, int sync, void *key)
 225 {
 226         struct _irqfd *irqfd = container_of(wait, struct _irqfd, wait);
 227         unsigned long flags = (unsigned long)key;
 228         struct kvm_kernel_irq_routing_entry irq;
 229         struct kvm *kvm = irqfd->kvm;
 230         unsigned seq;
 231         int idx;
 232
 233         if (flags & POLLIN) {
 234                 idx = srcu_read_lock(&kvm->irq_srcu);
 235                 do {
 236                         seq = read_seqcount_begin(&irqfd->irq_entry_sc);
 237                         irq = irqfd->irq_entry;
 238                 } while (read_seqcount_retry(&irqfd->irq_entry_sc, seq));
 239                 /* An event has been signaled, inject an interrupt */
 240                 if (irq.type == KVM_IRQ_ROUTING_MSI)
 241                         kvm_set_msi(&irq, kvm, KVM_USERSPACE_IRQ_SOURCE_ID, 1,
 242                                         false);
 243                 else
 244                         schedule_work(&irqfd->inject);
 245                 srcu_read_unlock(&kvm->irq_srcu, idx);
 246         }
 247
 248         if (flags & POLLHUP) {
 249                 /* The eventfd is closing, detach from KVM */
 250                 unsigned long flags;
 251
 252                 spin_lock_irqsave(&kvm->irqfds.lock, flags);
 253
 254                 /*
 255                  * We must check if someone deactivated the irqfd before
 256                  * we could acquire the irqfds.lock since the item is
 257                  * deactivated from the KVM side before it is unhooked from
 258                  * the wait-queue.  If it is already deactivated, we can
 259                  * simply return knowing the other side will cleanup for us.
 260                  * We cannot race against the irqfd going away since the
 261                  * other side is required to acquire wqh->lock, which we hold
 262                  */
 263                 if (irqfd_is_active(irqfd))
 264                         irqfd_deactivate(irqfd);
 265
 266                 spin_unlock_irqrestore(&kvm->irqfds.lock, flags);
 267         }
 268
 269         return 0;
 270 }
 271
 272 static void
 273 irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
 274                         poll_table *pt)
 275 {
 276         struct _irqfd *irqfd = container_of(pt, struct _irqfd, pt);
 277         add_wait_queue(wqh, &irqfd->wait);
 278 }
 279
 280 /* Must be called under irqfds.lock */
 281 static void irqfd_update(struct kvm *kvm, struct _irqfd *irqfd,
 282                          struct kvm_irq_routing_table *irq_rt)
 283 {
 284         struct kvm_kernel_irq_routing_entry *e;
 285
 286         write_seqcount_begin(&irqfd->irq_entry_sc);
 287
 288         irqfd->irq_entry.type = 0;
 289         if (irqfd->gsi >= irq_rt->nr_rt_entries)
 290                 goto out;
 291
 292         hlist_for_each_entry(e, &irq_rt->map[irqfd->gsi], link) {
 293                 /* Only fast-path MSI. */
 294                 if (e->type == KVM_IRQ_ROUTING_MSI)
 295                         irqfd->irq_entry = *e;
 296         }
 297
 298  out:
 299         write_seqcount_end(&irqfd->irq_entry_sc);
 300 }
 301
 302 static int
 303 kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
 304 {
 305         struct kvm_irq_routing_table *irq_rt;
 306         struct _irqfd *irqfd, *tmp;
 307         struct fd f;
 308         struct eventfd_ctx *eventfd = NULL, *resamplefd = NULL;
 309         int ret;
 310         unsigned int events;
 311
 312         irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL);
 313         if (!irqfd)
 314                 return -ENOMEM;
 315
 316         irqfd->kvm = kvm;
 317         irqfd->gsi = args->gsi;
 318         INIT_LIST_HEAD(&irqfd->list);
 319         INIT_WORK(&irqfd->inject, irqfd_inject);
 320         INIT_WORK(&irqfd->shutdown, irqfd_shutdown);
 321         seqcount_init(&irqfd->irq_entry_sc);
 322
 323         f = fdget(args->fd);
 324         if (!f.file) {
 325                 ret = -EBADF;
 326                 goto out;
 327         }
 328
 329         eventfd = eventfd_ctx_fileget(f.file);
 330         if (IS_ERR(eventfd)) {
 331                 ret = PTR_ERR(eventfd);
 332                 goto fail;
 333         }
 334
 335         irqfd->eventfd = eventfd;
 336
 337         if (args->flags & KVM_IRQFD_FLAG_RESAMPLE) {
 338                 struct _irqfd_resampler *resampler;
 339
 340                 resamplefd = eventfd_ctx_fdget(args->resamplefd);
 341                 if (IS_ERR(resamplefd)) {
 342                         ret = PTR_ERR(resamplefd);
 343                         goto fail;
 344                 }
 345
 346                 irqfd->resamplefd = resamplefd;
 347                 INIT_LIST_HEAD(&irqfd->resampler_link);
 348
 349                 mutex_lock(&kvm->irqfds.resampler_lock);
 350
 351                 list_for_each_entry(resampler,
 352                                     &kvm->irqfds.resampler_list, link) {
 353                         if (resampler->notifier.gsi == irqfd->gsi) {
 354                                 irqfd->resampler = resampler;
 355                                 break;
 356                         }
 357                 }
 358
 359                 if (!irqfd->resampler) {
 360                         resampler = kzalloc(sizeof(*resampler), GFP_KERNEL);
 361                         if (!resampler) {
 362                                 ret = -ENOMEM;
 363                                 mutex_unlock(&kvm->irqfds.resampler_lock);
 364                                 goto fail;
 365                         }
 366
 367                         resampler->kvm = kvm;
 368                         INIT_LIST_HEAD(&resampler->list);
 369                         resampler->notifier.gsi = irqfd->gsi;
 370                         resampler->notifier.irq_acked = irqfd_resampler_ack;
 371                         INIT_LIST_HEAD(&resampler->link);
 372
 373                         list_add(&resampler->link, &kvm->irqfds.resampler_list);
 374                         kvm_register_irq_ack_notifier(kvm,
 375                                                       &resampler->notifier);
 376                         irqfd->resampler = resampler;
 377                 }
 378
 379                 list_add_rcu(&irqfd->resampler_link, &irqfd->resampler->list);
 380                 synchronize_srcu(&kvm->irq_srcu);
 381
 382                 mutex_unlock(&kvm->irqfds.resampler_lock);
 383         }
 384
 385         /*
 386          * Install our own custom wake-up handling so we are notified via
 387          * a callback whenever someone signals the underlying eventfd
 388          */
 389         init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
 390         init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
 391
 392         spin_lock_irq(&kvm->irqfds.lock);
 393
 394         ret = 0;
 395         list_for_each_entry(tmp, &kvm->irqfds.items, list) {
 396                 if (irqfd->eventfd != tmp->eventfd)
 397                         continue;
 398                 /* This fd is used for another irq already. */
 399                 ret = -EBUSY;
 400                 spin_unlock_irq(&kvm->irqfds.lock);
 401                 goto fail;
 402         }
 403
 404         irq_rt = rcu_dereference_protected(kvm->irq_routing,
 405                                            lockdep_is_held(&kvm->irqfds.lock));
 406         irqfd_update(kvm, irqfd, irq_rt);
 407
 408         list_add_tail(&irqfd->list, &kvm->irqfds.items);
 409
 410         spin_unlock_irq(&kvm->irqfds.lock);
 411
 412         /*
 413          * Check if there was an event already pending on the eventfd
 414          * before we registered, and trigger it as if we didn't miss it.
 415          */
 416         events = f.file->f_op->poll(f.file, &irqfd->pt);
 417
 418         if (events & POLLIN)
 419                 schedule_work(&irqfd->inject);
 420
 421         /*
 422          * do not drop the file until the irqfd is fully initialized, otherwise
 423          * we might race against the POLLHUP
 424          */
 425         fdput(f);
 426
 427         return 0;
 428
 429 fail:
 430         if (irqfd->resampler)
 431                 irqfd_resampler_shutdown(irqfd);
 432
 433         if (resamplefd && !IS_ERR(resamplefd))
 434                 eventfd_ctx_put(resamplefd);
 435
 436         if (eventfd && !IS_ERR(eventfd))
 437                 eventfd_ctx_put(eventfd);
 438
 439         fdput(f);
 440
 441 out:
 442         kfree(irqfd);
 443         return ret;
 444 }
 445 #endif
 446
 447 void
 448 kvm_eventfd_init(struct kvm *kvm)
 449 {
 450 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
 451         spin_lock_init(&kvm->irqfds.lock);
 452         INIT_LIST_HEAD(&kvm->irqfds.items);
 453         INIT_LIST_HEAD(&kvm->irqfds.resampler_list);
 454         mutex_init(&kvm->irqfds.resampler_lock);
 455 #endif
 456         INIT_LIST_HEAD(&kvm->ioeventfds);
 457 }
 458
 459 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
 460 /*
 461  * shutdown any irqfd's that match fd+gsi
 462  */
 463 static int
 464 kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
 465 {
 466         struct _irqfd *irqfd, *tmp;
 467         struct eventfd_ctx *eventfd;
 468
 469         eventfd = eventfd_ctx_fdget(args->fd);
 470         if (IS_ERR(eventfd))
 471                 return PTR_ERR(eventfd);
 472
 473         spin_lock_irq(&kvm->irqfds.lock);
 474
 475         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
 476                 if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
 477                         /*
 478                          * This clearing of irq_entry.type is needed for when
 479                          * another thread calls kvm_irq_routing_update before
 480                          * we flush workqueue below (we synchronize with
 481                          * kvm_irq_routing_update using irqfds.lock).
 482                          */
 483                         write_seqcount_begin(&irqfd->irq_entry_sc);
 484                         irqfd->irq_entry.type = 0;
 485                         write_seqcount_end(&irqfd->irq_entry_sc);
 486                         irqfd_deactivate(irqfd);
 487                 }
 488         }
 489
 490         spin_unlock_irq(&kvm->irqfds.lock);
 491         eventfd_ctx_put(eventfd);
 492
 493         /*
 494          * Block until we know all outstanding shutdown jobs have completed
 495          * so that we guarantee there will not be any more interrupts on this
 496          * gsi once this deassign function returns.
 497          */
 498         flush_workqueue(irqfd_cleanup_wq);
 499
 500         return 0;
 501 }
 502
 503 int
 504 kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
 505 {
 506         if (args->flags & ~(KVM_IRQFD_FLAG_DEASSIGN | KVM_IRQFD_FLAG_RESAMPLE))
 507                 return -EINVAL;
 508
 509         if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
 510                 return kvm_irqfd_deassign(kvm, args);
 511
 512         return kvm_irqfd_assign(kvm, args);
 513 }
 514
 515 /*
 516  * This function is called as the kvm VM fd is being released. Shutdown all
 517  * irqfds that still remain open
 518  */
 519 void
 520 kvm_irqfd_release(struct kvm *kvm)
 521 {
 522         struct _irqfd *irqfd, *tmp;
 523
 524         spin_lock_irq(&kvm->irqfds.lock);
 525
 526         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
 527                 irqfd_deactivate(irqfd);
 528
 529         spin_unlock_irq(&kvm->irqfds.lock);
 530
 531         /*
 532          * Block until we know all outstanding shutdown jobs have completed
 533          * since we do not take a kvm* reference.
 534          */
 535         flush_workqueue(irqfd_cleanup_wq);
 536
 537 }
 538
 539 /*
 540  * Change irq_routing and irqfd.
 541  * Caller must invoke synchronize_srcu(&kvm->irq_srcu) afterwards.
 542  */
 543 void kvm_irq_routing_update(struct kvm *kvm,
 544                             struct kvm_irq_routing_table *irq_rt)
 545 {
 546         struct _irqfd *irqfd;
 547
 548         spin_lock_irq(&kvm->irqfds.lock);
 549
 550         rcu_assign_pointer(kvm->irq_routing, irq_rt);
 551
 552         list_for_each_entry(irqfd, &kvm->irqfds.items, list)
 553                 irqfd_update(kvm, irqfd, irq_rt);
 554
 555         spin_unlock_irq(&kvm->irqfds.lock);
 556 }
 557
 558 /*
 559  * create a host-wide workqueue for issuing deferred shutdown requests
 560  * aggregated from all vm* instances. We need our own isolated single-thread
 561  * queue to prevent deadlock against flushing the normal work-queue.
 562  */
 563 int kvm_irqfd_init(void)
 564 {
 565         irqfd_cleanup_wq = create_singlethread_workqueue("kvm-irqfd-cleanup");
 566         if (!irqfd_cleanup_wq)
 567                 return -ENOMEM;
 568
 569         return 0;
 570 }
 571
 572 void kvm_irqfd_exit(void)
 573 {
 574         destroy_workqueue(irqfd_cleanup_wq);
 575 }
 576 #endif
 577
 578 /*
 579  * --------------------------------------------------------------------
 580  * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
 581  *
 582  * userspace can register a PIO/MMIO address with an eventfd for receiving
 583  * notification when the memory has been touched.
 584  * --------------------------------------------------------------------
 585  */
 586
 587 struct _ioeventfd {
 588         struct list_head     list;
 589         u64                  addr;
 590         int                  length;
 591         struct eventfd_ctx  *eventfd;
 592         u64                  datamatch;
 593         struct kvm_io_device dev;
 594         u8                   bus_idx;
 595         bool                 wildcard;
 596 };
 597
 598 static inline struct _ioeventfd *
 599 to_ioeventfd(struct kvm_io_device *dev)
 600 {
 601         return container_of(dev, struct _ioeventfd, dev);
 602 }
 603
 604 static void
 605 ioeventfd_release(struct _ioeventfd *p)
 606 {
 607         eventfd_ctx_put(p->eventfd);
 608         list_del(&p->list);
 609         kfree(p);
 610 }
 611
 612 static bool
 613 ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
 614 {
 615         u64 _val;
 616
 617         if (addr != p->addr)
 618                 /* address must be precise for a hit */
 619                 return false;
 620
 621         if (!p->length)
 622                 /* length = 0 means only look at the address, so always a hit */
 623                 return true;
 624
 625         if (len != p->length)
 626                 /* address-range must be precise for a hit */
 627                 return false;
 628
 629         if (p->wildcard)
 630                 /* all else equal, wildcard is always a hit */
 631                 return true;
 632
 633         /* otherwise, we have to actually compare the data */
 634
 635         BUG_ON(!IS_ALIGNED((unsigned long)val, len));
 636
 637         switch (len) {
 638         case 1:
 639                 _val = *(u8 *)val;
 640                 break;
 641         case 2:
 642                 _val = *(u16 *)val;
 643                 break;
 644         case 4:
 645                 _val = *(u32 *)val;
 646                 break;
 647         case 8:
 648                 _val = *(u64 *)val;
 649                 break;
 650         default:
 651                 return false;
 652         }
 653
 654         return _val == p->datamatch ? true : false;
 655 }
 656
 657 /* MMIO/PIO writes trigger an event if the addr/val match */
 658 static int
 659 ioeventfd_write(struct kvm_io_device *this, gpa_t addr, int len,
 660                 const void *val)
 661 {
 662         struct _ioeventfd *p = to_ioeventfd(this);
 663
 664         if (!ioeventfd_in_range(p, addr, len, val))
 665                 return -EOPNOTSUPP;
 666
 667         eventfd_signal(p->eventfd, 1);
 668         return 0;
 669 }
 670
 671 /*
 672  * This function is called as KVM is completely shutting down.  We do not
 673  * need to worry about locking just nuke anything we have as quickly as possible
 674  */
 675 static void
 676 ioeventfd_destructor(struct kvm_io_device *this)
 677 {
 678         struct _ioeventfd *p = to_ioeventfd(this);
 679
 680         ioeventfd_release(p);
 681 }
 682
 683 static const struct kvm_io_device_ops ioeventfd_ops = {
 684         .write      = ioeventfd_write,
 685         .destructor = ioeventfd_destructor,
 686 };
 687
 688 /* assumes kvm->slots_lock held */
 689 static bool
 690 ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
 691 {
 692         struct _ioeventfd *_p;
 693
 694         list_for_each_entry(_p, &kvm->ioeventfds, list)
 695                 if (_p->bus_idx == p->bus_idx &&
 696                     _p->addr == p->addr &&
 697                     (!_p->length || !p->length ||
 698                      (_p->length == p->length &&
 699                       (_p->wildcard || p->wildcard ||
 700                        _p->datamatch == p->datamatch))))
 701                         return true;
 702
 703         return false;
 704 }
 705
 706 static enum kvm_bus ioeventfd_bus_from_flags(__u32 flags)
 707 {
 708         if (flags & KVM_IOEVENTFD_FLAG_PIO)
 709                 return KVM_PIO_BUS;
 710         if (flags & KVM_IOEVENTFD_FLAG_VIRTIO_CCW_NOTIFY)
 711                 return KVM_VIRTIO_CCW_NOTIFY_BUS;
 712         return KVM_MMIO_BUS;
 713 }
 714
 715 static int
 716 kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 717 {
 718         enum kvm_bus              bus_idx;
 719         struct _ioeventfd        *p;
 720         struct eventfd_ctx       *eventfd;
 721         int                       ret;
 722
 723         bus_idx = ioeventfd_bus_from_flags(args->flags);
 724         /* must be natural-word sized, or 0 to ignore length */
 725         switch (args->len) {
 726         case 0:
 727         case 1:
 728         case 2:
 729         case 4:
 730         case 8:
 731                 break;
 732         default:
 733                 return -EINVAL;
 734         }
 735
 736         /* check for range overflow */
 737         if (args->addr + args->len < args->addr)
 738                 return -EINVAL;
 739
 740         /* check for extra flags that we don't understand */
 741         if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
 742                 return -EINVAL;
 743
 744         /* ioeventfd with no length can't be combined with DATAMATCH */
 745         if (!args->len &&
 746             args->flags & (KVM_IOEVENTFD_FLAG_PIO |
 747                            KVM_IOEVENTFD_FLAG_DATAMATCH))
 748                 return -EINVAL;
 749
 750         eventfd = eventfd_ctx_fdget(args->fd);
 751         if (IS_ERR(eventfd))
 752                 return PTR_ERR(eventfd);
 753
 754         p = kzalloc(sizeof(*p), GFP_KERNEL);
 755         if (!p) {
 756                 ret = -ENOMEM;
 757                 goto fail;
 758         }
 759
 760         INIT_LIST_HEAD(&p->list);
 761         p->addr    = args->addr;
 762         p->bus_idx = bus_idx;
 763         p->length  = args->len;
 764         p->eventfd = eventfd;
 765
 766         /* The datamatch feature is optional, otherwise this is a wildcard */
 767         if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
 768                 p->datamatch = args->datamatch;
 769         else
 770                 p->wildcard = true;
 771
 772         mutex_lock(&kvm->slots_lock);
 773
 774         /* Verify that there isn't a match already */
 775         if (ioeventfd_check_collision(kvm, p)) {
 776                 ret = -EEXIST;
 777                 goto unlock_fail;
 778         }
 779
 780         kvm_iodevice_init(&p->dev, &ioeventfd_ops);
 781
 782         ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
 783                                       &p->dev);
 784         if (ret < 0)
 785                 goto unlock_fail;
 786
 787         /* When length is ignored, MMIO is also put on a separate bus, for
 788          * faster lookups.
 789          */
 790         if (!args->len && !(args->flags & KVM_IOEVENTFD_FLAG_PIO)) {
 791                 ret = kvm_io_bus_register_dev(kvm, KVM_FAST_MMIO_BUS,
 792                                               p->addr, 0, &p->dev);
 793                 if (ret < 0)
 794                         goto register_fail;
 795         }
 796
 797         kvm->buses[bus_idx]->ioeventfd_count++;
 798         list_add_tail(&p->list, &kvm->ioeventfds);
 799
 800         mutex_unlock(&kvm->slots_lock);
 801
 802         return 0;
 803
 804 register_fail:
 805         kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
 806 unlock_fail:
 807         mutex_unlock(&kvm->slots_lock);
 808
 809 fail:
 810         kfree(p);
 811         eventfd_ctx_put(eventfd);
 812
 813         return ret;
 814 }
 815
 816 static int
 817 kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 818 {
 819         enum kvm_bus              bus_idx;
 820         struct _ioeventfd        *p, *tmp;
 821         struct eventfd_ctx       *eventfd;
 822         int                       ret = -ENOENT;
 823
 824         bus_idx = ioeventfd_bus_from_flags(args->flags);
 825         eventfd = eventfd_ctx_fdget(args->fd);
 826         if (IS_ERR(eventfd))
 827                 return PTR_ERR(eventfd);
 828
 829         mutex_lock(&kvm->slots_lock);
 830
 831         list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
 832                 bool wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
 833
 834                 if (p->bus_idx != bus_idx ||
 835                     p->eventfd != eventfd  ||
 836                     p->addr != args->addr  ||
 837                     p->length != args->len ||
 838                     p->wildcard != wildcard)
 839                         continue;
 840
 841                 if (!p->wildcard && p->datamatch != args->datamatch)
 842                         continue;
 843
 844                 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
 845                 if (!p->length) {
 846                         kvm_io_bus_unregister_dev(kvm, KVM_FAST_MMIO_BUS,
 847                                                   &p->dev);
 848                 }
 849                 kvm->buses[bus_idx]->ioeventfd_count--;
 850                 ioeventfd_release(p);
 851                 ret = 0;
 852                 break;
 853         }
 854
 855         mutex_unlock(&kvm->slots_lock);
 856
 857         eventfd_ctx_put(eventfd);
 858
 859         return ret;
 860 }
 861
 862 int
 863 kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 864 {
 865         if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
 866                 return kvm_deassign_ioeventfd(kvm, args);
 867
 868         return kvm_assign_ioeventfd(kvm, args);
 869 }