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 #include <trace/events/kvm.h>
  38
  39 #include <kvm/iodev.h>
  40
  41 #ifdef CONFIG_HAVE_KVM_IRQFD
  42 /*
  43  * --------------------------------------------------------------------
  44  * irqfd: Allows an fd to be used to inject an interrupt to the guest
  45  *
  46  * Credit goes to Avi Kivity for the original idea.
  47  * --------------------------------------------------------------------
  48  */
  49
  50 /*
  51  * Resampling irqfds are a special variety of irqfds used to emulate
  52  * level triggered interrupts.  The interrupt is asserted on eventfd
  53  * trigger.  On acknowledgement through the irq ack notifier, the
  54  * interrupt is de-asserted and userspace is notified through the
  55  * resamplefd.  All resamplers on the same gsi are de-asserted
  56  * together, so we don't need to track the state of each individual
  57  * user.  We can also therefore share the same irq source ID.
  58  */
  59 struct _irqfd_resampler {
  60         struct kvm *kvm;
  61         /*
  62          * List of resampling struct _irqfd objects sharing this gsi.
  63          * RCU list modified under kvm->irqfds.resampler_lock
  64          */
  65         struct list_head list;
  66         struct kvm_irq_ack_notifier notifier;
  67         /*
  68          * Entry in list of kvm->irqfd.resampler_list.  Use for sharing
  69          * resamplers among irqfds on the same gsi.
  70          * Accessed and modified under kvm->irqfds.resampler_lock
  71          */
  72         struct list_head link;
  73 };
  74
  75 struct _irqfd {
  76         /* Used for MSI fast-path */
  77         struct kvm *kvm;
  78         wait_queue_t wait;
  79         /* Update side is protected by irqfds.lock */
  80         struct kvm_kernel_irq_routing_entry irq_entry;
  81         seqcount_t irq_entry_sc;
  82         /* Used for level IRQ fast-path */
  83         int gsi;
  84         struct work_struct inject;
  85         /* The resampler used by this irqfd (resampler-only) */
  86         struct _irqfd_resampler *resampler;
  87         /* Eventfd notified on resample (resampler-only) */
  88         struct eventfd_ctx *resamplefd;
  89         /* Entry in list of irqfds for a resampler (resampler-only) */
  90         struct list_head resampler_link;
  91         /* Used for setup/shutdown */
  92         struct eventfd_ctx *eventfd;
  93         struct list_head list;
  94         poll_table pt;
  95         struct work_struct shutdown;
  96 };
  97
  98 static struct workqueue_struct *irqfd_cleanup_wq;
  99
 100 static void
 101 irqfd_inject(struct work_struct *work)
 102 {
 103         struct _irqfd *irqfd = container_of(work, struct _irqfd, inject);
 104         struct kvm *kvm = irqfd->kvm;
 105
 106         if (!irqfd->resampler) {
 107                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1,
 108                                 false);
 109                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0,
 110                                 false);
 111         } else
 112                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 113                             irqfd->gsi, 1, false);
 114 }
 115
 116 /*
 117  * Since resampler irqfds share an IRQ source ID, we de-assert once
 118  * then notify all of the resampler irqfds using this GSI.  We can't
 119  * do multiple de-asserts or we risk racing with incoming re-asserts.
 120  */
 121 static void
 122 irqfd_resampler_ack(struct kvm_irq_ack_notifier *kian)
 123 {
 124         struct _irqfd_resampler *resampler;
 125         struct kvm *kvm;
 126         struct _irqfd *irqfd;
 127         int idx;
 128
 129         resampler = container_of(kian, struct _irqfd_resampler, notifier);
 130         kvm = resampler->kvm;
 131
 132         kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 133                     resampler->notifier.gsi, 0, false);
 134
 135         idx = srcu_read_lock(&kvm->irq_srcu);
 136
 137         list_for_each_entry_rcu(irqfd, &resampler->list, resampler_link)
 138                 eventfd_signal(irqfd->resamplefd, 1);
 139
 140         srcu_read_unlock(&kvm->irq_srcu, idx);
 141 }
 142
 143 static void
 144 irqfd_resampler_shutdown(struct _irqfd *irqfd)
 145 {
 146         struct _irqfd_resampler *resampler = irqfd->resampler;
 147         struct kvm *kvm = resampler->kvm;
 148
 149         mutex_lock(&kvm->irqfds.resampler_lock);
 150
 151         list_del_rcu(&irqfd->resampler_link);
 152         synchronize_srcu(&kvm->irq_srcu);
 153
 154         if (list_empty(&resampler->list)) {
 155                 list_del(&resampler->link);
 156                 kvm_unregister_irq_ack_notifier(kvm, &resampler->notifier);
 157                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 158                             resampler->notifier.gsi, 0, false);
 159                 kfree(resampler);
 160         }
 161
 162         mutex_unlock(&kvm->irqfds.resampler_lock);
 163 }
 164
 165 /*
 166  * Race-free decouple logic (ordering is critical)
 167  */
 168 static void
 169 irqfd_shutdown(struct work_struct *work)
 170 {
 171         struct _irqfd *irqfd = container_of(work, struct _irqfd, shutdown);
 172         u64 cnt;
 173
 174         /*
 175          * Synchronize with the wait-queue and unhook ourselves to prevent
 176          * further events.
 177          */
 178         eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
 179
 180         /*
 181          * We know no new events will be scheduled at this point, so block
 182          * until all previously outstanding events have completed
 183          */
 184         flush_work(&irqfd->inject);
 185
 186         if (irqfd->resampler) {
 187                 irqfd_resampler_shutdown(irqfd);
 188                 eventfd_ctx_put(irqfd->resamplefd);
 189         }
 190
 191         /*
 192          * It is now safe to release the object's resources
 193          */
 194         eventfd_ctx_put(irqfd->eventfd);
 195         kfree(irqfd);
 196 }
 197
 198
 199 /* assumes kvm->irqfds.lock is held */
 200 static bool
 201 irqfd_is_active(struct _irqfd *irqfd)
 202 {
 203         return list_empty(&irqfd->list) ? false : true;
 204 }
 205
 206 /*
 207  * Mark the irqfd as inactive and schedule it for removal
 208  *
 209  * assumes kvm->irqfds.lock is held
 210  */
 211 static void
 212 irqfd_deactivate(struct _irqfd *irqfd)
 213 {
 214         BUG_ON(!irqfd_is_active(irqfd));
 215
 216         list_del_init(&irqfd->list);
 217
 218         queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
 219 }
 220
 221 /*
 222  * Called with wqh->lock held and interrupts disabled
 223  */
 224 static int
 225 irqfd_wakeup(wait_queue_t *wait, unsigned mode, int sync, void *key)
 226 {
 227         struct _irqfd *irqfd = container_of(wait, struct _irqfd, wait);
 228         unsigned long flags = (unsigned long)key;
 229         struct kvm_kernel_irq_routing_entry irq;
 230         struct kvm *kvm = irqfd->kvm;
 231         unsigned seq;
 232         int idx;
 233
 234         if (flags & POLLIN) {
 235                 idx = srcu_read_lock(&kvm->irq_srcu);
 236                 do {
 237                         seq = read_seqcount_begin(&irqfd->irq_entry_sc);
 238                         irq = irqfd->irq_entry;
 239                 } while (read_seqcount_retry(&irqfd->irq_entry_sc, seq));
 240                 /* An event has been signaled, inject an interrupt */
 241                 if (irq.type == KVM_IRQ_ROUTING_MSI)
 242                         kvm_set_msi(&irq, kvm, KVM_USERSPACE_IRQ_SOURCE_ID, 1,
 243                                         false);
 244                 else
 245                         schedule_work(&irqfd->inject);
 246                 srcu_read_unlock(&kvm->irq_srcu, idx);
 247         }
 248
 249         if (flags & POLLHUP) {
 250                 /* The eventfd is closing, detach from KVM */
 251                 unsigned long flags;
 252
 253                 spin_lock_irqsave(&kvm->irqfds.lock, flags);
 254
 255                 /*
 256                  * We must check if someone deactivated the irqfd before
 257                  * we could acquire the irqfds.lock since the item is
 258                  * deactivated from the KVM side before it is unhooked from
 259                  * the wait-queue.  If it is already deactivated, we can
 260                  * simply return knowing the other side will cleanup for us.
 261                  * We cannot race against the irqfd going away since the
 262                  * other side is required to acquire wqh->lock, which we hold
 263                  */
 264                 if (irqfd_is_active(irqfd))
 265                         irqfd_deactivate(irqfd);
 266
 267                 spin_unlock_irqrestore(&kvm->irqfds.lock, flags);
 268         }
 269
 270         return 0;
 271 }
 272
 273 static void
 274 irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
 275                         poll_table *pt)
 276 {
 277         struct _irqfd *irqfd = container_of(pt, struct _irqfd, pt);
 278         add_wait_queue(wqh, &irqfd->wait);
 279 }
 280
 281 /* Must be called under irqfds.lock */
 282 static void irqfd_update(struct kvm *kvm, struct _irqfd *irqfd)
 283 {
 284         struct kvm_kernel_irq_routing_entry *e;
 285         struct kvm_kernel_irq_routing_entry entries[KVM_NR_IRQCHIPS];
 286         int i, n_entries;
 287
 288         n_entries = kvm_irq_map_gsi(kvm, entries, irqfd->gsi);
 289
 290         write_seqcount_begin(&irqfd->irq_entry_sc);
 291
 292         irqfd->irq_entry.type = 0;
 293
 294         e = entries;
 295         for (i = 0; i < n_entries; ++i, ++e) {
 296                 /* Only fast-path MSI. */
 297                 if (e->type == KVM_IRQ_ROUTING_MSI)
 298                         irqfd->irq_entry = *e;
 299         }
 300
 301         write_seqcount_end(&irqfd->irq_entry_sc);
 302 }
 303
 304 static int
 305 kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
 306 {
 307         struct _irqfd *irqfd, *tmp;
 308         struct fd f;
 309         struct eventfd_ctx *eventfd = NULL, *resamplefd = NULL;
 310         int ret;
 311         unsigned int events;
 312         int idx;
 313
 314         if (!kvm_arch_intc_initialized(kvm))
 315                 return -EAGAIN;
 316
 317         irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL);
 318         if (!irqfd)
 319                 return -ENOMEM;
 320
 321         irqfd->kvm = kvm;
 322         irqfd->gsi = args->gsi;
 323         INIT_LIST_HEAD(&irqfd->list);
 324         INIT_WORK(&irqfd->inject, irqfd_inject);
 325         INIT_WORK(&irqfd->shutdown, irqfd_shutdown);
 326         seqcount_init(&irqfd->irq_entry_sc);
 327
 328         f = fdget(args->fd);
 329         if (!f.file) {
 330                 ret = -EBADF;
 331                 goto out;
 332         }
 333
 334         eventfd = eventfd_ctx_fileget(f.file);
 335         if (IS_ERR(eventfd)) {
 336                 ret = PTR_ERR(eventfd);
 337                 goto fail;
 338         }
 339
 340         irqfd->eventfd = eventfd;
 341
 342         if (args->flags & KVM_IRQFD_FLAG_RESAMPLE) {
 343                 struct _irqfd_resampler *resampler;
 344
 345                 resamplefd = eventfd_ctx_fdget(args->resamplefd);
 346                 if (IS_ERR(resamplefd)) {
 347                         ret = PTR_ERR(resamplefd);
 348                         goto fail;
 349                 }
 350
 351                 irqfd->resamplefd = resamplefd;
 352                 INIT_LIST_HEAD(&irqfd->resampler_link);
 353
 354                 mutex_lock(&kvm->irqfds.resampler_lock);
 355
 356                 list_for_each_entry(resampler,
 357                                     &kvm->irqfds.resampler_list, link) {
 358                         if (resampler->notifier.gsi == irqfd->gsi) {
 359                                 irqfd->resampler = resampler;
 360                                 break;
 361                         }
 362                 }
 363
 364                 if (!irqfd->resampler) {
 365                         resampler = kzalloc(sizeof(*resampler), GFP_KERNEL);
 366                         if (!resampler) {
 367                                 ret = -ENOMEM;
 368                                 mutex_unlock(&kvm->irqfds.resampler_lock);
 369                                 goto fail;
 370                         }
 371
 372                         resampler->kvm = kvm;
 373                         INIT_LIST_HEAD(&resampler->list);
 374                         resampler->notifier.gsi = irqfd->gsi;
 375                         resampler->notifier.irq_acked = irqfd_resampler_ack;
 376                         INIT_LIST_HEAD(&resampler->link);
 377
 378                         list_add(&resampler->link, &kvm->irqfds.resampler_list);
 379                         kvm_register_irq_ack_notifier(kvm,
 380                                                       &resampler->notifier);
 381                         irqfd->resampler = resampler;
 382                 }
 383
 384                 list_add_rcu(&irqfd->resampler_link, &irqfd->resampler->list);
 385                 synchronize_srcu(&kvm->irq_srcu);
 386
 387                 mutex_unlock(&kvm->irqfds.resampler_lock);
 388         }
 389
 390         /*
 391          * Install our own custom wake-up handling so we are notified via
 392          * a callback whenever someone signals the underlying eventfd
 393          */
 394         init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
 395         init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
 396
 397         spin_lock_irq(&kvm->irqfds.lock);
 398
 399         ret = 0;
 400         list_for_each_entry(tmp, &kvm->irqfds.items, list) {
 401                 if (irqfd->eventfd != tmp->eventfd)
 402                         continue;
 403                 /* This fd is used for another irq already. */
 404                 ret = -EBUSY;
 405                 spin_unlock_irq(&kvm->irqfds.lock);
 406                 goto fail;
 407         }
 408
 409         idx = srcu_read_lock(&kvm->irq_srcu);
 410         irqfd_update(kvm, irqfd);
 411         srcu_read_unlock(&kvm->irq_srcu, idx);
 412
 413         list_add_tail(&irqfd->list, &kvm->irqfds.items);
 414
 415         spin_unlock_irq(&kvm->irqfds.lock);
 416
 417         /*
 418          * Check if there was an event already pending on the eventfd
 419          * before we registered, and trigger it as if we didn't miss it.
 420          */
 421         events = f.file->f_op->poll(f.file, &irqfd->pt);
 422
 423         if (events & POLLIN)
 424                 schedule_work(&irqfd->inject);
 425
 426         /*
 427          * do not drop the file until the irqfd is fully initialized, otherwise
 428          * we might race against the POLLHUP
 429          */
 430         fdput(f);
 431
 432         return 0;
 433
 434 fail:
 435         if (irqfd->resampler)
 436                 irqfd_resampler_shutdown(irqfd);
 437
 438         if (resamplefd && !IS_ERR(resamplefd))
 439                 eventfd_ctx_put(resamplefd);
 440
 441         if (eventfd && !IS_ERR(eventfd))
 442                 eventfd_ctx_put(eventfd);
 443
 444         fdput(f);
 445
 446 out:
 447         kfree(irqfd);
 448         return ret;
 449 }
 450
 451 bool kvm_irq_has_notifier(struct kvm *kvm, unsigned irqchip, unsigned pin)
 452 {
 453         struct kvm_irq_ack_notifier *kian;
 454         int gsi, idx;
 455
 456         idx = srcu_read_lock(&kvm->irq_srcu);
 457         gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
 458         if (gsi != -1)
 459                 hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
 460                                          link)
 461                         if (kian->gsi == gsi) {
 462                                 srcu_read_unlock(&kvm->irq_srcu, idx);
 463                                 return true;
 464                         }
 465
 466         srcu_read_unlock(&kvm->irq_srcu, idx);
 467
 468         return false;
 469 }
 470 EXPORT_SYMBOL_GPL(kvm_irq_has_notifier);
 471
 472 void kvm_notify_acked_irq(struct kvm *kvm, unsigned irqchip, unsigned pin)
 473 {
 474         struct kvm_irq_ack_notifier *kian;
 475         int gsi, idx;
 476
 477         trace_kvm_ack_irq(irqchip, pin);
 478
 479         idx = srcu_read_lock(&kvm->irq_srcu);
 480         gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
 481         if (gsi != -1)
 482                 hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
 483                                          link)
 484                         if (kian->gsi == gsi)
 485                                 kian->irq_acked(kian);
 486         srcu_read_unlock(&kvm->irq_srcu, idx);
 487 }
 488
 489 void kvm_register_irq_ack_notifier(struct kvm *kvm,
 490                                    struct kvm_irq_ack_notifier *kian)
 491 {
 492         mutex_lock(&kvm->irq_lock);
 493         hlist_add_head_rcu(&kian->link, &kvm->irq_ack_notifier_list);
 494         mutex_unlock(&kvm->irq_lock);
 495         kvm_vcpu_request_scan_ioapic(kvm);
 496 }
 497
 498 void kvm_unregister_irq_ack_notifier(struct kvm *kvm,
 499                                     struct kvm_irq_ack_notifier *kian)
 500 {
 501         mutex_lock(&kvm->irq_lock);
 502         hlist_del_init_rcu(&kian->link);
 503         mutex_unlock(&kvm->irq_lock);
 504         synchronize_srcu(&kvm->irq_srcu);
 505         kvm_vcpu_request_scan_ioapic(kvm);
 506 }
 507 #endif
 508
 509 void
 510 kvm_eventfd_init(struct kvm *kvm)
 511 {
 512 #ifdef CONFIG_HAVE_KVM_IRQFD
 513         spin_lock_init(&kvm->irqfds.lock);
 514         INIT_LIST_HEAD(&kvm->irqfds.items);
 515         INIT_LIST_HEAD(&kvm->irqfds.resampler_list);
 516         mutex_init(&kvm->irqfds.resampler_lock);
 517 #endif
 518         INIT_LIST_HEAD(&kvm->ioeventfds);
 519 }
 520
 521 #ifdef CONFIG_HAVE_KVM_IRQFD
 522 /*
 523  * shutdown any irqfd's that match fd+gsi
 524  */
 525 static int
 526 kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
 527 {
 528         struct _irqfd *irqfd, *tmp;
 529         struct eventfd_ctx *eventfd;
 530
 531         eventfd = eventfd_ctx_fdget(args->fd);
 532         if (IS_ERR(eventfd))
 533                 return PTR_ERR(eventfd);
 534
 535         spin_lock_irq(&kvm->irqfds.lock);
 536
 537         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
 538                 if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
 539                         /*
 540                          * This clearing of irq_entry.type is needed for when
 541                          * another thread calls kvm_irq_routing_update before
 542                          * we flush workqueue below (we synchronize with
 543                          * kvm_irq_routing_update using irqfds.lock).
 544                          */
 545                         write_seqcount_begin(&irqfd->irq_entry_sc);
 546                         irqfd->irq_entry.type = 0;
 547                         write_seqcount_end(&irqfd->irq_entry_sc);
 548                         irqfd_deactivate(irqfd);
 549                 }
 550         }
 551
 552         spin_unlock_irq(&kvm->irqfds.lock);
 553         eventfd_ctx_put(eventfd);
 554
 555         /*
 556          * Block until we know all outstanding shutdown jobs have completed
 557          * so that we guarantee there will not be any more interrupts on this
 558          * gsi once this deassign function returns.
 559          */
 560         flush_workqueue(irqfd_cleanup_wq);
 561
 562         return 0;
 563 }
 564
 565 int
 566 kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
 567 {
 568         if (args->flags & ~(KVM_IRQFD_FLAG_DEASSIGN | KVM_IRQFD_FLAG_RESAMPLE))
 569                 return -EINVAL;
 570
 571         if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
 572                 return kvm_irqfd_deassign(kvm, args);
 573
 574         return kvm_irqfd_assign(kvm, args);
 575 }
 576
 577 /*
 578  * This function is called as the kvm VM fd is being released. Shutdown all
 579  * irqfds that still remain open
 580  */
 581 void
 582 kvm_irqfd_release(struct kvm *kvm)
 583 {
 584         struct _irqfd *irqfd, *tmp;
 585
 586         spin_lock_irq(&kvm->irqfds.lock);
 587
 588         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
 589                 irqfd_deactivate(irqfd);
 590
 591         spin_unlock_irq(&kvm->irqfds.lock);
 592
 593         /*
 594          * Block until we know all outstanding shutdown jobs have completed
 595          * since we do not take a kvm* reference.
 596          */
 597         flush_workqueue(irqfd_cleanup_wq);
 598
 599 }
 600
 601 /*
 602  * Take note of a change in irq routing.
 603  * Caller must invoke synchronize_srcu(&kvm->irq_srcu) afterwards.
 604  */
 605 void kvm_irq_routing_update(struct kvm *kvm)
 606 {
 607         struct _irqfd *irqfd;
 608
 609         spin_lock_irq(&kvm->irqfds.lock);
 610
 611         list_for_each_entry(irqfd, &kvm->irqfds.items, list)
 612                 irqfd_update(kvm, irqfd);
 613
 614         spin_unlock_irq(&kvm->irqfds.lock);
 615 }
 616
 617 /*
 618  * create a host-wide workqueue for issuing deferred shutdown requests
 619  * aggregated from all vm* instances. We need our own isolated single-thread
 620  * queue to prevent deadlock against flushing the normal work-queue.
 621  */
 622 int kvm_irqfd_init(void)
 623 {
 624         irqfd_cleanup_wq = create_singlethread_workqueue("kvm-irqfd-cleanup");
 625         if (!irqfd_cleanup_wq)
 626                 return -ENOMEM;
 627
 628         return 0;
 629 }
 630
 631 void kvm_irqfd_exit(void)
 632 {
 633         destroy_workqueue(irqfd_cleanup_wq);
 634 }
 635 #endif
 636
 637 /*
 638  * --------------------------------------------------------------------
 639  * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
 640  *
 641  * userspace can register a PIO/MMIO address with an eventfd for receiving
 642  * notification when the memory has been touched.
 643  * --------------------------------------------------------------------
 644  */
 645
 646 struct _ioeventfd {
 647         struct list_head     list;
 648         u64                  addr;
 649         int                  length;
 650         struct eventfd_ctx  *eventfd;
 651         u64                  datamatch;
 652         struct kvm_io_device dev;
 653         u8                   bus_idx;
 654         bool                 wildcard;
 655 };
 656
 657 static inline struct _ioeventfd *
 658 to_ioeventfd(struct kvm_io_device *dev)
 659 {
 660         return container_of(dev, struct _ioeventfd, dev);
 661 }
 662
 663 static void
 664 ioeventfd_release(struct _ioeventfd *p)
 665 {
 666         eventfd_ctx_put(p->eventfd);
 667         list_del(&p->list);
 668         kfree(p);
 669 }
 670
 671 static bool
 672 ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
 673 {
 674         u64 _val;
 675
 676         if (addr != p->addr)
 677                 /* address must be precise for a hit */
 678                 return false;
 679
 680         if (!p->length)
 681                 /* length = 0 means only look at the address, so always a hit */
 682                 return true;
 683
 684         if (len != p->length)
 685                 /* address-range must be precise for a hit */
 686                 return false;
 687
 688         if (p->wildcard)
 689                 /* all else equal, wildcard is always a hit */
 690                 return true;
 691
 692         /* otherwise, we have to actually compare the data */
 693
 694         BUG_ON(!IS_ALIGNED((unsigned long)val, len));
 695
 696         switch (len) {
 697         case 1:
 698                 _val = *(u8 *)val;
 699                 break;
 700         case 2:
 701                 _val = *(u16 *)val;
 702                 break;
 703         case 4:
 704                 _val = *(u32 *)val;
 705                 break;
 706         case 8:
 707                 _val = *(u64 *)val;
 708                 break;
 709         default:
 710                 return false;
 711         }
 712
 713         return _val == p->datamatch ? true : false;
 714 }
 715
 716 /* MMIO/PIO writes trigger an event if the addr/val match */
 717 static int
 718 ioeventfd_write(struct kvm_vcpu *vcpu, struct kvm_io_device *this, gpa_t addr,
 719                 int len, const void *val)
 720 {
 721         struct _ioeventfd *p = to_ioeventfd(this);
 722
 723         if (!ioeventfd_in_range(p, addr, len, val))
 724                 return -EOPNOTSUPP;
 725
 726         eventfd_signal(p->eventfd, 1);
 727         return 0;
 728 }
 729
 730 /*
 731  * This function is called as KVM is completely shutting down.  We do not
 732  * need to worry about locking just nuke anything we have as quickly as possible
 733  */
 734 static void
 735 ioeventfd_destructor(struct kvm_io_device *this)
 736 {
 737         struct _ioeventfd *p = to_ioeventfd(this);
 738
 739         ioeventfd_release(p);
 740 }
 741
 742 static const struct kvm_io_device_ops ioeventfd_ops = {
 743         .write      = ioeventfd_write,
 744         .destructor = ioeventfd_destructor,
 745 };
 746
 747 /* assumes kvm->slots_lock held */
 748 static bool
 749 ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
 750 {
 751         struct _ioeventfd *_p;
 752
 753         list_for_each_entry(_p, &kvm->ioeventfds, list)
 754                 if (_p->bus_idx == p->bus_idx &&
 755                     _p->addr == p->addr &&
 756                     (!_p->length || !p->length ||
 757                      (_p->length == p->length &&
 758                       (_p->wildcard || p->wildcard ||
 759                        _p->datamatch == p->datamatch))))
 760                         return true;
 761
 762         return false;
 763 }
 764
 765 static enum kvm_bus ioeventfd_bus_from_flags(__u32 flags)
 766 {
 767         if (flags & KVM_IOEVENTFD_FLAG_PIO)
 768                 return KVM_PIO_BUS;
 769         if (flags & KVM_IOEVENTFD_FLAG_VIRTIO_CCW_NOTIFY)
 770                 return KVM_VIRTIO_CCW_NOTIFY_BUS;
 771         return KVM_MMIO_BUS;
 772 }
 773
 774 static int kvm_assign_ioeventfd_idx(struct kvm *kvm,
 775                                 enum kvm_bus bus_idx,
 776                                 struct kvm_ioeventfd *args)
 777 {
 778
 779         struct eventfd_ctx *eventfd;
 780         struct _ioeventfd *p;
 781         int ret;
 782
 783         eventfd = eventfd_ctx_fdget(args->fd);
 784         if (IS_ERR(eventfd))
 785                 return PTR_ERR(eventfd);
 786
 787         p = kzalloc(sizeof(*p), GFP_KERNEL);
 788         if (!p) {
 789                 ret = -ENOMEM;
 790                 goto fail;
 791         }
 792
 793         INIT_LIST_HEAD(&p->list);
 794         p->addr    = args->addr;
 795         p->bus_idx = bus_idx;
 796         p->length  = args->len;
 797         p->eventfd = eventfd;
 798
 799         /* The datamatch feature is optional, otherwise this is a wildcard */
 800         if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
 801                 p->datamatch = args->datamatch;
 802         else
 803                 p->wildcard = true;
 804
 805         mutex_lock(&kvm->slots_lock);
 806
 807         /* Verify that there isn't a match already */
 808         if (ioeventfd_check_collision(kvm, p)) {
 809                 ret = -EEXIST;
 810                 goto unlock_fail;
 811         }
 812
 813         kvm_iodevice_init(&p->dev, &ioeventfd_ops);
 814
 815         ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
 816                                       &p->dev);
 817         if (ret < 0)
 818                 goto unlock_fail;
 819
 820         kvm->buses[bus_idx]->ioeventfd_count++;
 821         list_add_tail(&p->list, &kvm->ioeventfds);
 822
 823         mutex_unlock(&kvm->slots_lock);
 824
 825         return 0;
 826
 827 unlock_fail:
 828         mutex_unlock(&kvm->slots_lock);
 829
 830 fail:
 831         kfree(p);
 832         eventfd_ctx_put(eventfd);
 833
 834         return ret;
 835 }
 836
 837 static int
 838 kvm_deassign_ioeventfd_idx(struct kvm *kvm, enum kvm_bus bus_idx,
 839                            struct kvm_ioeventfd *args)
 840 {
 841         struct _ioeventfd        *p, *tmp;
 842         struct eventfd_ctx       *eventfd;
 843         int                       ret = -ENOENT;
 844
 845         eventfd = eventfd_ctx_fdget(args->fd);
 846         if (IS_ERR(eventfd))
 847                 return PTR_ERR(eventfd);
 848
 849         mutex_lock(&kvm->slots_lock);
 850
 851         list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
 852                 bool wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
 853
 854                 if (p->bus_idx != bus_idx ||
 855                     p->eventfd != eventfd  ||
 856                     p->addr != args->addr  ||
 857                     p->length != args->len ||
 858                     p->wildcard != wildcard)
 859                         continue;
 860
 861                 if (!p->wildcard && p->datamatch != args->datamatch)
 862                         continue;
 863
 864                 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
 865                 kvm->buses[bus_idx]->ioeventfd_count--;
 866                 ioeventfd_release(p);
 867                 ret = 0;
 868                 break;
 869         }
 870
 871         mutex_unlock(&kvm->slots_lock);
 872
 873         eventfd_ctx_put(eventfd);
 874
 875         return ret;
 876 }
 877
 878 static int kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 879 {
 880         enum kvm_bus bus_idx = ioeventfd_bus_from_flags(args->flags);
 881         int ret = kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
 882
 883         if (!args->len && bus_idx == KVM_MMIO_BUS)
 884                 kvm_deassign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
 885
 886         return ret;
 887 }
 888
 889 static int
 890 kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 891 {
 892         enum kvm_bus              bus_idx;
 893         int ret;
 894
 895         bus_idx = ioeventfd_bus_from_flags(args->flags);
 896         /* must be natural-word sized, or 0 to ignore length */
 897         switch (args->len) {
 898         case 0:
 899         case 1:
 900         case 2:
 901         case 4:
 902         case 8:
 903                 break;
 904         default:
 905                 return -EINVAL;
 906         }
 907
 908         /* check for range overflow */
 909         if (args->addr + args->len < args->addr)
 910                 return -EINVAL;
 911
 912         /* check for extra flags that we don't understand */
 913         if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
 914                 return -EINVAL;
 915
 916         /* ioeventfd with no length can't be combined with DATAMATCH */
 917         if (!args->len &&
 918             args->flags & (KVM_IOEVENTFD_FLAG_PIO |
 919                            KVM_IOEVENTFD_FLAG_DATAMATCH))
 920                 return -EINVAL;
 921
 922         ret = kvm_assign_ioeventfd_idx(kvm, bus_idx, args);
 923         if (ret)
 924                 goto fail;
 925
 926         /* When length is ignored, MMIO is also put on a separate bus, for
 927          * faster lookups.
 928          */
 929         if (!args->len && bus_idx == KVM_MMIO_BUS) {
 930                 ret = kvm_assign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
 931                 if (ret < 0)
 932                         goto fast_fail;
 933         }
 934
 935         return 0;
 936
 937 fast_fail:
 938         kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
 939 fail:
 940         return ret;
 941 }
 942
 943 int
 944 kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 945 {
 946         if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
 947                 return kvm_deassign_ioeventfd(kvm, args);
 948
 949         return kvm_assign_ioeventfd(kvm, args);
 950 }