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 {
 283         struct kvm_kernel_irq_routing_entry *e;
 284         struct kvm_kernel_irq_routing_entry entries[KVM_NR_IRQCHIPS];
 285         int i, n_entries;
 286
 287         n_entries = kvm_irq_map_gsi(kvm, entries, irqfd->gsi);
 288
 289         write_seqcount_begin(&irqfd->irq_entry_sc);
 290
 291         irqfd->irq_entry.type = 0;
 292
 293         e = entries;
 294         for (i = 0; i < n_entries; ++i, ++e) {
 295                 /* Only fast-path MSI. */
 296                 if (e->type == KVM_IRQ_ROUTING_MSI)
 297                         irqfd->irq_entry = *e;
 298         }
 299
 300         write_seqcount_end(&irqfd->irq_entry_sc);
 301 }
 302
 303 static int
 304 kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
 305 {
 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         int idx;
 312
 313         irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL);
 314         if (!irqfd)
 315                 return -ENOMEM;
 316
 317         irqfd->kvm = kvm;
 318         irqfd->gsi = args->gsi;
 319         INIT_LIST_HEAD(&irqfd->list);
 320         INIT_WORK(&irqfd->inject, irqfd_inject);
 321         INIT_WORK(&irqfd->shutdown, irqfd_shutdown);
 322         seqcount_init(&irqfd->irq_entry_sc);
 323
 324         f = fdget(args->fd);
 325         if (!f.file) {
 326                 ret = -EBADF;
 327                 goto out;
 328         }
 329
 330         eventfd = eventfd_ctx_fileget(f.file);
 331         if (IS_ERR(eventfd)) {
 332                 ret = PTR_ERR(eventfd);
 333                 goto fail;
 334         }
 335
 336         irqfd->eventfd = eventfd;
 337
 338         if (args->flags & KVM_IRQFD_FLAG_RESAMPLE) {
 339                 struct _irqfd_resampler *resampler;
 340
 341                 resamplefd = eventfd_ctx_fdget(args->resamplefd);
 342                 if (IS_ERR(resamplefd)) {
 343                         ret = PTR_ERR(resamplefd);
 344                         goto fail;
 345                 }
 346
 347                 irqfd->resamplefd = resamplefd;
 348                 INIT_LIST_HEAD(&irqfd->resampler_link);
 349
 350                 mutex_lock(&kvm->irqfds.resampler_lock);
 351
 352                 list_for_each_entry(resampler,
 353                                     &kvm->irqfds.resampler_list, link) {
 354                         if (resampler->notifier.gsi == irqfd->gsi) {
 355                                 irqfd->resampler = resampler;
 356                                 break;
 357                         }
 358                 }
 359
 360                 if (!irqfd->resampler) {
 361                         resampler = kzalloc(sizeof(*resampler), GFP_KERNEL);
 362                         if (!resampler) {
 363                                 ret = -ENOMEM;
 364                                 mutex_unlock(&kvm->irqfds.resampler_lock);
 365                                 goto fail;
 366                         }
 367
 368                         resampler->kvm = kvm;
 369                         INIT_LIST_HEAD(&resampler->list);
 370                         resampler->notifier.gsi = irqfd->gsi;
 371                         resampler->notifier.irq_acked = irqfd_resampler_ack;
 372                         INIT_LIST_HEAD(&resampler->link);
 373
 374                         list_add(&resampler->link, &kvm->irqfds.resampler_list);
 375                         kvm_register_irq_ack_notifier(kvm,
 376                                                       &resampler->notifier);
 377                         irqfd->resampler = resampler;
 378                 }
 379
 380                 list_add_rcu(&irqfd->resampler_link, &irqfd->resampler->list);
 381                 synchronize_srcu(&kvm->irq_srcu);
 382
 383                 mutex_unlock(&kvm->irqfds.resampler_lock);
 384         }
 385
 386         /*
 387          * Install our own custom wake-up handling so we are notified via
 388          * a callback whenever someone signals the underlying eventfd
 389          */
 390         init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
 391         init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
 392
 393         spin_lock_irq(&kvm->irqfds.lock);
 394
 395         ret = 0;
 396         list_for_each_entry(tmp, &kvm->irqfds.items, list) {
 397                 if (irqfd->eventfd != tmp->eventfd)
 398                         continue;
 399                 /* This fd is used for another irq already. */
 400                 ret = -EBUSY;
 401                 spin_unlock_irq(&kvm->irqfds.lock);
 402                 goto fail;
 403         }
 404
 405         idx = srcu_read_lock(&kvm->irq_srcu);
 406         irqfd_update(kvm, irqfd);
 407         srcu_read_unlock(&kvm->irq_srcu, idx);
 408
 409         list_add_tail(&irqfd->list, &kvm->irqfds.items);
 410
 411         spin_unlock_irq(&kvm->irqfds.lock);
 412
 413         /*
 414          * Check if there was an event already pending on the eventfd
 415          * before we registered, and trigger it as if we didn't miss it.
 416          */
 417         events = f.file->f_op->poll(f.file, &irqfd->pt);
 418
 419         if (events & POLLIN)
 420                 schedule_work(&irqfd->inject);
 421
 422         /*
 423          * do not drop the file until the irqfd is fully initialized, otherwise
 424          * we might race against the POLLHUP
 425          */
 426         fdput(f);
 427
 428         return 0;
 429
 430 fail:
 431         if (irqfd->resampler)
 432                 irqfd_resampler_shutdown(irqfd);
 433
 434         if (resamplefd && !IS_ERR(resamplefd))
 435                 eventfd_ctx_put(resamplefd);
 436
 437         if (eventfd && !IS_ERR(eventfd))
 438                 eventfd_ctx_put(eventfd);
 439
 440         fdput(f);
 441
 442 out:
 443         kfree(irqfd);
 444         return ret;
 445 }
 446 #endif
 447
 448 void
 449 kvm_eventfd_init(struct kvm *kvm)
 450 {
 451 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
 452         spin_lock_init(&kvm->irqfds.lock);
 453         INIT_LIST_HEAD(&kvm->irqfds.items);
 454         INIT_LIST_HEAD(&kvm->irqfds.resampler_list);
 455         mutex_init(&kvm->irqfds.resampler_lock);
 456 #endif
 457         INIT_LIST_HEAD(&kvm->ioeventfds);
 458 }
 459
 460 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
 461 /*
 462  * shutdown any irqfd's that match fd+gsi
 463  */
 464 static int
 465 kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
 466 {
 467         struct _irqfd *irqfd, *tmp;
 468         struct eventfd_ctx *eventfd;
 469
 470         eventfd = eventfd_ctx_fdget(args->fd);
 471         if (IS_ERR(eventfd))
 472                 return PTR_ERR(eventfd);
 473
 474         spin_lock_irq(&kvm->irqfds.lock);
 475
 476         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
 477                 if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
 478                         /*
 479                          * This clearing of irq_entry.type is needed for when
 480                          * another thread calls kvm_irq_routing_update before
 481                          * we flush workqueue below (we synchronize with
 482                          * kvm_irq_routing_update using irqfds.lock).
 483                          */
 484                         write_seqcount_begin(&irqfd->irq_entry_sc);
 485                         irqfd->irq_entry.type = 0;
 486                         write_seqcount_end(&irqfd->irq_entry_sc);
 487                         irqfd_deactivate(irqfd);
 488                 }
 489         }
 490
 491         spin_unlock_irq(&kvm->irqfds.lock);
 492         eventfd_ctx_put(eventfd);
 493
 494         /*
 495          * Block until we know all outstanding shutdown jobs have completed
 496          * so that we guarantee there will not be any more interrupts on this
 497          * gsi once this deassign function returns.
 498          */
 499         flush_workqueue(irqfd_cleanup_wq);
 500
 501         return 0;
 502 }
 503
 504 int
 505 kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
 506 {
 507         if (args->flags & ~(KVM_IRQFD_FLAG_DEASSIGN | KVM_IRQFD_FLAG_RESAMPLE))
 508                 return -EINVAL;
 509
 510         if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
 511                 return kvm_irqfd_deassign(kvm, args);
 512
 513         return kvm_irqfd_assign(kvm, args);
 514 }
 515
 516 /*
 517  * This function is called as the kvm VM fd is being released. Shutdown all
 518  * irqfds that still remain open
 519  */
 520 void
 521 kvm_irqfd_release(struct kvm *kvm)
 522 {
 523         struct _irqfd *irqfd, *tmp;
 524
 525         spin_lock_irq(&kvm->irqfds.lock);
 526
 527         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
 528                 irqfd_deactivate(irqfd);
 529
 530         spin_unlock_irq(&kvm->irqfds.lock);
 531
 532         /*
 533          * Block until we know all outstanding shutdown jobs have completed
 534          * since we do not take a kvm* reference.
 535          */
 536         flush_workqueue(irqfd_cleanup_wq);
 537
 538 }
 539
 540 /*
 541  * Take note of a change in irq routing.
 542  * Caller must invoke synchronize_srcu(&kvm->irq_srcu) afterwards.
 543  */
 544 void kvm_irq_routing_update(struct kvm *kvm)
 545 {
 546         struct _irqfd *irqfd;
 547
 548         spin_lock_irq(&kvm->irqfds.lock);
 549
 550         list_for_each_entry(irqfd, &kvm->irqfds.items, list)
 551                 irqfd_update(kvm, irqfd);
 552
 553         spin_unlock_irq(&kvm->irqfds.lock);
 554 }
 555
 556 /*
 557  * create a host-wide workqueue for issuing deferred shutdown requests
 558  * aggregated from all vm* instances. We need our own isolated single-thread
 559  * queue to prevent deadlock against flushing the normal work-queue.
 560  */
 561 int kvm_irqfd_init(void)
 562 {
 563         irqfd_cleanup_wq = create_singlethread_workqueue("kvm-irqfd-cleanup");
 564         if (!irqfd_cleanup_wq)
 565                 return -ENOMEM;
 566
 567         return 0;
 568 }
 569
 570 void kvm_irqfd_exit(void)
 571 {
 572         destroy_workqueue(irqfd_cleanup_wq);
 573 }
 574 #endif
 575
 576 /*
 577  * --------------------------------------------------------------------
 578  * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
 579  *
 580  * userspace can register a PIO/MMIO address with an eventfd for receiving
 581  * notification when the memory has been touched.
 582  * --------------------------------------------------------------------
 583  */
 584
 585 struct _ioeventfd {
 586         struct list_head     list;
 587         u64                  addr;
 588         int                  length;
 589         struct eventfd_ctx  *eventfd;
 590         u64                  datamatch;
 591         struct kvm_io_device dev;
 592         u8                   bus_idx;
 593         bool                 wildcard;
 594 };
 595
 596 static inline struct _ioeventfd *
 597 to_ioeventfd(struct kvm_io_device *dev)
 598 {
 599         return container_of(dev, struct _ioeventfd, dev);
 600 }
 601
 602 static void
 603 ioeventfd_release(struct _ioeventfd *p)
 604 {
 605         eventfd_ctx_put(p->eventfd);
 606         list_del(&p->list);
 607         kfree(p);
 608 }
 609
 610 static bool
 611 ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
 612 {
 613         u64 _val;
 614
 615         if (addr != p->addr)
 616                 /* address must be precise for a hit */
 617                 return false;
 618
 619         if (!p->length)
 620                 /* length = 0 means only look at the address, so always a hit */
 621                 return true;
 622
 623         if (len != p->length)
 624                 /* address-range must be precise for a hit */
 625                 return false;
 626
 627         if (p->wildcard)
 628                 /* all else equal, wildcard is always a hit */
 629                 return true;
 630
 631         /* otherwise, we have to actually compare the data */
 632
 633         BUG_ON(!IS_ALIGNED((unsigned long)val, len));
 634
 635         switch (len) {
 636         case 1:
 637                 _val = *(u8 *)val;
 638                 break;
 639         case 2:
 640                 _val = *(u16 *)val;
 641                 break;
 642         case 4:
 643                 _val = *(u32 *)val;
 644                 break;
 645         case 8:
 646                 _val = *(u64 *)val;
 647                 break;
 648         default:
 649                 return false;
 650         }
 651
 652         return _val == p->datamatch ? true : false;
 653 }
 654
 655 /* MMIO/PIO writes trigger an event if the addr/val match */
 656 static int
 657 ioeventfd_write(struct kvm_io_device *this, gpa_t addr, int len,
 658                 const void *val)
 659 {
 660         struct _ioeventfd *p = to_ioeventfd(this);
 661
 662         if (!ioeventfd_in_range(p, addr, len, val))
 663                 return -EOPNOTSUPP;
 664
 665         eventfd_signal(p->eventfd, 1);
 666         return 0;
 667 }
 668
 669 /*
 670  * This function is called as KVM is completely shutting down.  We do not
 671  * need to worry about locking just nuke anything we have as quickly as possible
 672  */
 673 static void
 674 ioeventfd_destructor(struct kvm_io_device *this)
 675 {
 676         struct _ioeventfd *p = to_ioeventfd(this);
 677
 678         ioeventfd_release(p);
 679 }
 680
 681 static const struct kvm_io_device_ops ioeventfd_ops = {
 682         .write      = ioeventfd_write,
 683         .destructor = ioeventfd_destructor,
 684 };
 685
 686 /* assumes kvm->slots_lock held */
 687 static bool
 688 ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
 689 {
 690         struct _ioeventfd *_p;
 691
 692         list_for_each_entry(_p, &kvm->ioeventfds, list)
 693                 if (_p->bus_idx == p->bus_idx &&
 694                     _p->addr == p->addr &&
 695                     (!_p->length || !p->length ||
 696                      (_p->length == p->length &&
 697                       (_p->wildcard || p->wildcard ||
 698                        _p->datamatch == p->datamatch))))
 699                         return true;
 700
 701         return false;
 702 }
 703
 704 static enum kvm_bus ioeventfd_bus_from_flags(__u32 flags)
 705 {
 706         if (flags & KVM_IOEVENTFD_FLAG_PIO)
 707                 return KVM_PIO_BUS;
 708         if (flags & KVM_IOEVENTFD_FLAG_VIRTIO_CCW_NOTIFY)
 709                 return KVM_VIRTIO_CCW_NOTIFY_BUS;
 710         return KVM_MMIO_BUS;
 711 }
 712
 713 static int
 714 kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 715 {
 716         enum kvm_bus              bus_idx;
 717         struct _ioeventfd        *p;
 718         struct eventfd_ctx       *eventfd;
 719         int                       ret;
 720
 721         bus_idx = ioeventfd_bus_from_flags(args->flags);
 722         /* must be natural-word sized, or 0 to ignore length */
 723         switch (args->len) {
 724         case 0:
 725         case 1:
 726         case 2:
 727         case 4:
 728         case 8:
 729                 break;
 730         default:
 731                 return -EINVAL;
 732         }
 733
 734         /* check for range overflow */
 735         if (args->addr + args->len < args->addr)
 736                 return -EINVAL;
 737
 738         /* check for extra flags that we don't understand */
 739         if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
 740                 return -EINVAL;
 741
 742         /* ioeventfd with no length can't be combined with DATAMATCH */
 743         if (!args->len &&
 744             args->flags & (KVM_IOEVENTFD_FLAG_PIO |
 745                            KVM_IOEVENTFD_FLAG_DATAMATCH))
 746                 return -EINVAL;
 747
 748         eventfd = eventfd_ctx_fdget(args->fd);
 749         if (IS_ERR(eventfd))
 750                 return PTR_ERR(eventfd);
 751
 752         p = kzalloc(sizeof(*p), GFP_KERNEL);
 753         if (!p) {
 754                 ret = -ENOMEM;
 755                 goto fail;
 756         }
 757
 758         INIT_LIST_HEAD(&p->list);
 759         p->addr    = args->addr;
 760         p->bus_idx = bus_idx;
 761         p->length  = args->len;
 762         p->eventfd = eventfd;
 763
 764         /* The datamatch feature is optional, otherwise this is a wildcard */
 765         if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
 766                 p->datamatch = args->datamatch;
 767         else
 768                 p->wildcard = true;
 769
 770         mutex_lock(&kvm->slots_lock);
 771
 772         /* Verify that there isn't a match already */
 773         if (ioeventfd_check_collision(kvm, p)) {
 774                 ret = -EEXIST;
 775                 goto unlock_fail;
 776         }
 777
 778         kvm_iodevice_init(&p->dev, &ioeventfd_ops);
 779
 780         ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
 781                                       &p->dev);
 782         if (ret < 0)
 783                 goto unlock_fail;
 784
 785         /* When length is ignored, MMIO is also put on a separate bus, for
 786          * faster lookups.
 787          */
 788         if (!args->len && !(args->flags & KVM_IOEVENTFD_FLAG_PIO)) {
 789                 ret = kvm_io_bus_register_dev(kvm, KVM_FAST_MMIO_BUS,
 790                                               p->addr, 0, &p->dev);
 791                 if (ret < 0)
 792                         goto register_fail;
 793         }
 794
 795         kvm->buses[bus_idx]->ioeventfd_count++;
 796         list_add_tail(&p->list, &kvm->ioeventfds);
 797
 798         mutex_unlock(&kvm->slots_lock);
 799
 800         return 0;
 801
 802 register_fail:
 803         kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
 804 unlock_fail:
 805         mutex_unlock(&kvm->slots_lock);
 806
 807 fail:
 808         kfree(p);
 809         eventfd_ctx_put(eventfd);
 810
 811         return ret;
 812 }
 813
 814 static int
 815 kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 816 {
 817         enum kvm_bus              bus_idx;
 818         struct _ioeventfd        *p, *tmp;
 819         struct eventfd_ctx       *eventfd;
 820         int                       ret = -ENOENT;
 821
 822         bus_idx = ioeventfd_bus_from_flags(args->flags);
 823         eventfd = eventfd_ctx_fdget(args->fd);
 824         if (IS_ERR(eventfd))
 825                 return PTR_ERR(eventfd);
 826
 827         mutex_lock(&kvm->slots_lock);
 828
 829         list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
 830                 bool wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
 831
 832                 if (p->bus_idx != bus_idx ||
 833                     p->eventfd != eventfd  ||
 834                     p->addr != args->addr  ||
 835                     p->length != args->len ||
 836                     p->wildcard != wildcard)
 837                         continue;
 838
 839                 if (!p->wildcard && p->datamatch != args->datamatch)
 840                         continue;
 841
 842                 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
 843                 if (!p->length) {
 844                         kvm_io_bus_unregister_dev(kvm, KVM_FAST_MMIO_BUS,
 845                                                   &p->dev);
 846                 }
 847                 kvm->buses[bus_idx]->ioeventfd_count--;
 848                 ioeventfd_release(p);
 849                 ret = 0;
 850                 break;
 851         }
 852
 853         mutex_unlock(&kvm->slots_lock);
 854
 855         eventfd_ctx_put(eventfd);
 856
 857         return ret;
 858 }
 859
 860 int
 861 kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 862 {
 863         if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
 864                 return kvm_deassign_ioeventfd(kvm, args);
 865
 866         return kvm_assign_ioeventfd(kvm, args);
 867 }