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