virt/kvm/eventfd.c

   1 /*
   2  * kvm eventfd support - use eventfd objects to signal various KVM events
   3  *
   4  * Copyright 2009 Novell.  All Rights Reserved.
   5  * Copyright 2010 Red Hat, Inc. and/or its affiliates.
   6  *
   7  * Author:
   8  *      Gregory Haskins <ghaskins@novell.com>
   9  *
  10  * This file is free software; you can redistribute it and/or modify
  11  * it under the terms of version 2 of the GNU General Public License
  12  * as published by the Free Software Foundation.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU General Public License
  20  * along with this program; if not, write to the Free Software Foundation,
  21  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA.
  22  */
  23
  24 #include <linux/kvm_host.h>
  25 #include <linux/kvm.h>
  26 #include <linux/workqueue.h>
  27 #include <linux/syscalls.h>
  28 #include <linux/wait.h>
  29 #include <linux/poll.h>
  30 #include <linux/file.h>
  31 #include <linux/list.h>
  32 #include <linux/eventfd.h>
  33 #include <linux/kernel.h>
  34 #include <linux/srcu.h>
  35 #include <linux/slab.h>
  36 #include <linux/seqlock.h>
  37
  38 #include "iodev.h"
  39
  40 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
  41 /*
  42  * --------------------------------------------------------------------
  43  * irqfd: Allows an fd to be used to inject an interrupt to the guest
  44  *
  45  * Credit goes to Avi Kivity for the original idea.
  46  * --------------------------------------------------------------------
  47  */
  48
  49 /*
  50  * Resampling irqfds are a special variety of irqfds used to emulate
  51  * level triggered interrupts.  The interrupt is asserted on eventfd
  52  * trigger.  On acknowledgement through the irq ack notifier, the
  53  * interrupt is de-asserted and userspace is notified through the
  54  * resamplefd.  All resamplers on the same gsi are de-asserted
  55  * together, so we don't need to track the state of each individual
  56  * user.  We can also therefore share the same irq source ID.
  57  */
  58 struct _irqfd_resampler {
  59         struct kvm *kvm;
  60         /*
  61          * List of resampling struct _irqfd objects sharing this gsi.
  62          * RCU list modified under kvm->irqfds.resampler_lock
  63          */
  64         struct list_head list;
  65         struct kvm_irq_ack_notifier notifier;
  66         /*
  67          * Entry in list of kvm->irqfd.resampler_list.  Use for sharing
  68          * resamplers among irqfds on the same gsi.
  69          * Accessed and modified under kvm->irqfds.resampler_lock
  70          */
  71         struct list_head link;
  72 };
  73
  74 struct _irqfd {
  75         /* Used for MSI fast-path */
  76         struct kvm *kvm;
  77         wait_queue_t wait;
  78         /* Update side is protected by irqfds.lock */
  79         struct kvm_kernel_irq_routing_entry irq_entry;
  80         seqcount_t irq_entry_sc;
  81         /* Used for level IRQ fast-path */
  82         int gsi;
  83         struct work_struct inject;
  84         /* The resampler used by this irqfd (resampler-only) */
  85         struct _irqfd_resampler *resampler;
  86         /* Eventfd notified on resample (resampler-only) */
  87         struct eventfd_ctx *resamplefd;
  88         /* Entry in list of irqfds for a resampler (resampler-only) */
  89         struct list_head resampler_link;
  90         /* Used for setup/shutdown */
  91         struct eventfd_ctx *eventfd;
  92         struct list_head list;
  93         poll_table pt;
  94         struct work_struct shutdown;
  95 };
  96
  97 static struct workqueue_struct *irqfd_cleanup_wq;
  98
  99 static void
 100 irqfd_inject(struct work_struct *work)
 101 {
 102         struct _irqfd *irqfd = container_of(work, struct _irqfd, inject);
 103         struct kvm *kvm = irqfd->kvm;
 104
 105         if (!irqfd->resampler) {
 106                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1,
 107                                 false);
 108                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0,
 109                                 false);
 110         } else
 111                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 112                             irqfd->gsi, 1, false);
 113 }
 114
 115 /*
 116  * Since resampler irqfds share an IRQ source ID, we de-assert once
 117  * then notify all of the resampler irqfds using this GSI.  We can't
 118  * do multiple de-asserts or we risk racing with incoming re-asserts.
 119  */
 120 static void
 121 irqfd_resampler_ack(struct kvm_irq_ack_notifier *kian)
 122 {
 123         struct _irqfd_resampler *resampler;
 124         struct kvm *kvm;
 125         struct _irqfd *irqfd;
 126         int idx;
 127
 128         resampler = container_of(kian, struct _irqfd_resampler, notifier);
 129         kvm = resampler->kvm;
 130
 131         kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 132                     resampler->notifier.gsi, 0, false);
 133
 134         idx = srcu_read_lock(&kvm->irq_srcu);
 135
 136         list_for_each_entry_rcu(irqfd, &resampler->list, resampler_link)
 137                 eventfd_signal(irqfd->resamplefd, 1);
 138
 139         srcu_read_unlock(&kvm->irq_srcu, idx);
 140 }
 141
 142 static void
 143 irqfd_resampler_shutdown(struct _irqfd *irqfd)
 144 {
 145         struct _irqfd_resampler *resampler = irqfd->resampler;
 146         struct kvm *kvm = resampler->kvm;
 147
 148         mutex_lock(&kvm->irqfds.resampler_lock);
 149
 150         list_del_rcu(&irqfd->resampler_link);
 151         synchronize_srcu(&kvm->irq_srcu);
 152
 153         if (list_empty(&resampler->list)) {
 154                 list_del(&resampler->link);
 155                 kvm_unregister_irq_ack_notifier(kvm, &resampler->notifier);
 156                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 157                             resampler->notifier.gsi, 0, false);
 158                 kfree(resampler);
 159         }
 160
 161         mutex_unlock(&kvm->irqfds.resampler_lock);
 162 }
 163
 164 /*
 165  * Race-free decouple logic (ordering is critical)
 166  */
 167 static void
 168 irqfd_shutdown(struct work_struct *work)
 169 {
 170         struct _irqfd *irqfd = container_of(work, struct _irqfd, shutdown);
 171         u64 cnt;
 172
 173         /*
 174          * Synchronize with the wait-queue and unhook ourselves to prevent
 175          * further events.
 176          */
 177         eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
 178
 179         /*
 180          * We know no new events will be scheduled at this point, so block
 181          * until all previously outstanding events have completed
 182          */
 183         flush_work(&irqfd->inject);
 184
 185         if (irqfd->resampler) {
 186                 irqfd_resampler_shutdown(irqfd);
 187                 eventfd_ctx_put(irqfd->resamplefd);
 188         }
 189
 190         /*
 191          * It is now safe to release the object's resources
 192          */
 193         eventfd_ctx_put(irqfd->eventfd);
 194         kfree(irqfd);
 195 }
 196
 197
 198 /* assumes kvm->irqfds.lock is held */
 199 static bool
 200 irqfd_is_active(struct _irqfd *irqfd)
 201 {
 202         return list_empty(&irqfd->list) ? false : true;
 203 }
 204
 205 /*
 206  * Mark the irqfd as inactive and schedule it for removal
 207  *
 208  * assumes kvm->irqfds.lock is held
 209  */
 210 static void
 211 irqfd_deactivate(struct _irqfd *irqfd)
 212 {
 213         BUG_ON(!irqfd_is_active(irqfd));
 214
 215         list_del_init(&irqfd->list);
 216
 217         queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
 218 }
 219
 220 /*
 221  * Called with wqh->lock held and interrupts disabled
 222  */
 223 static int
 224 irqfd_wakeup(wait_queue_t *wait, unsigned mode, int sync, void *key)
 225 {
 226         struct _irqfd *irqfd = container_of(wait, struct _irqfd, wait);
 227         unsigned long flags = (unsigned long)key;
 228         struct kvm_kernel_irq_routing_entry irq;
 229         struct kvm *kvm = irqfd->kvm;
 230         unsigned seq;
 231         int idx;
 232
 233         if (flags & POLLIN) {
 234                 idx = srcu_read_lock(&kvm->irq_srcu);
 235                 do {
 236                         seq = read_seqcount_begin(&irqfd->irq_entry_sc);
 237                         irq = irqfd->irq_entry;
 238                 } while (read_seqcount_retry(&irqfd->irq_entry_sc, seq));
 239                 /* An event has been signaled, inject an interrupt */
 240                 if (irq.type == KVM_IRQ_ROUTING_MSI)
 241                         kvm_set_msi(&irq, kvm, KVM_USERSPACE_IRQ_SOURCE_ID, 1,
 242                                         false);
 243                 else
 244                         schedule_work(&irqfd->inject);
 245                 srcu_read_unlock(&kvm->irq_srcu, idx);
 246         }
 247
 248         if (flags & POLLHUP) {
 249                 /* The eventfd is closing, detach from KVM */
 250                 unsigned long flags;
 251
 252                 spin_lock_irqsave(&kvm->irqfds.lock, flags);
 253
 254                 /*
 255                  * We must check if someone deactivated the irqfd before
 256                  * we could acquire the irqfds.lock since the item is
 257                  * deactivated from the KVM side before it is unhooked from
 258                  * the wait-queue.  If it is already deactivated, we can
 259                  * simply return knowing the other side will cleanup for us.
 260                  * We cannot race against the irqfd going away since the
 261                  * other side is required to acquire wqh->lock, which we hold
 262                  */
 263                 if (irqfd_is_active(irqfd))
 264                         irqfd_deactivate(irqfd);
 265
 266                 spin_unlock_irqrestore(&kvm->irqfds.lock, flags);
 267         }
 268
 269         return 0;
 270 }
 271
 272 static void
 273 irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
 274                         poll_table *pt)
 275 {
 276         struct _irqfd *irqfd = container_of(pt, struct _irqfd, pt);
 277         add_wait_queue(wqh, &irqfd->wait);
 278 }
 279
 280 /* Must be called under irqfds.lock */
 281 static void irqfd_update(struct kvm *kvm, struct _irqfd *irqfd,
 282                          struct kvm_irq_routing_table *irq_rt)
 283 {
 284         struct kvm_kernel_irq_routing_entry *e;
 285         struct kvm_kernel_irq_routing_entry entries[KVM_NR_IRQCHIPS];
 286         int i, n_entries;
 287
 288         n_entries = kvm_irq_map_gsi(entries, irq_rt, 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 kvm_irq_routing_table *irq_rt;
 308         struct _irqfd *irqfd, *tmp;
 309         struct fd f;
 310         struct eventfd_ctx *eventfd = NULL, *resamplefd = NULL;
 311         int ret;
 312         unsigned int events;
 313
 314         irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL);
 315         if (!irqfd)
 316                 return -ENOMEM;
 317
 318         irqfd->kvm = kvm;
 319         irqfd->gsi = args->gsi;
 320         INIT_LIST_HEAD(&irqfd->list);
 321         INIT_WORK(&irqfd->inject, irqfd_inject);
 322         INIT_WORK(&irqfd->shutdown, irqfd_shutdown);
 323         seqcount_init(&irqfd->irq_entry_sc);
 324
 325         f = fdget(args->fd);
 326         if (!f.file) {
 327                 ret = -EBADF;
 328                 goto out;
 329         }
 330
 331         eventfd = eventfd_ctx_fileget(f.file);
 332         if (IS_ERR(eventfd)) {
 333                 ret = PTR_ERR(eventfd);
 334                 goto fail;
 335         }
 336
 337         irqfd->eventfd = eventfd;
 338
 339         if (args->flags & KVM_IRQFD_FLAG_RESAMPLE) {
 340                 struct _irqfd_resampler *resampler;
 341
 342                 resamplefd = eventfd_ctx_fdget(args->resamplefd);
 343                 if (IS_ERR(resamplefd)) {
 344                         ret = PTR_ERR(resamplefd);
 345                         goto fail;
 346                 }
 347
 348                 irqfd->resamplefd = resamplefd;
 349                 INIT_LIST_HEAD(&irqfd->resampler_link);
 350
 351                 mutex_lock(&kvm->irqfds.resampler_lock);
 352
 353                 list_for_each_entry(resampler,
 354                                     &kvm->irqfds.resampler_list, link) {
 355                         if (resampler->notifier.gsi == irqfd->gsi) {
 356                                 irqfd->resampler = resampler;
 357                                 break;
 358                         }
 359                 }
 360
 361                 if (!irqfd->resampler) {
 362                         resampler = kzalloc(sizeof(*resampler), GFP_KERNEL);
 363                         if (!resampler) {
 364                                 ret = -ENOMEM;
 365                                 mutex_unlock(&kvm->irqfds.resampler_lock);
 366                                 goto fail;
 367                         }
 368
 369                         resampler->kvm = kvm;
 370                         INIT_LIST_HEAD(&resampler->list);
 371                         resampler->notifier.gsi = irqfd->gsi;
 372                         resampler->notifier.irq_acked = irqfd_resampler_ack;
 373                         INIT_LIST_HEAD(&resampler->link);
 374
 375                         list_add(&resampler->link, &kvm->irqfds.resampler_list);
 376                         kvm_register_irq_ack_notifier(kvm,
 377                                                       &resampler->notifier);
 378                         irqfd->resampler = resampler;
 379                 }
 380
 381                 list_add_rcu(&irqfd->resampler_link, &irqfd->resampler->list);
 382                 synchronize_srcu(&kvm->irq_srcu);
 383
 384                 mutex_unlock(&kvm->irqfds.resampler_lock);
 385         }
 386
 387         /*
 388          * Install our own custom wake-up handling so we are notified via
 389          * a callback whenever someone signals the underlying eventfd
 390          */
 391         init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
 392         init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
 393
 394         spin_lock_irq(&kvm->irqfds.lock);
 395
 396         ret = 0;
 397         list_for_each_entry(tmp, &kvm->irqfds.items, list) {
 398                 if (irqfd->eventfd != tmp->eventfd)
 399                         continue;
 400                 /* This fd is used for another irq already. */
 401                 ret = -EBUSY;
 402                 spin_unlock_irq(&kvm->irqfds.lock);
 403                 goto fail;
 404         }
 405
 406         irq_rt = rcu_dereference_protected(kvm->irq_routing,
 407                                            lockdep_is_held(&kvm->irqfds.lock));
 408         irqfd_update(kvm, irqfd, irq_rt);
 409
 410         list_add_tail(&irqfd->list, &kvm->irqfds.items);
 411
 412         spin_unlock_irq(&kvm->irqfds.lock);
 413
 414         /*
 415          * Check if there was an event already pending on the eventfd
 416          * before we registered, and trigger it as if we didn't miss it.
 417          */
 418         events = f.file->f_op->poll(f.file, &irqfd->pt);
 419
 420         if (events & POLLIN)
 421                 schedule_work(&irqfd->inject);
 422
 423         /*
 424          * do not drop the file until the irqfd is fully initialized, otherwise
 425          * we might race against the POLLHUP
 426          */
 427         fdput(f);
 428
 429         return 0;
 430
 431 fail:
 432         if (irqfd->resampler)
 433                 irqfd_resampler_shutdown(irqfd);
 434
 435         if (resamplefd && !IS_ERR(resamplefd))
 436                 eventfd_ctx_put(resamplefd);
 437
 438         if (eventfd && !IS_ERR(eventfd))
 439                 eventfd_ctx_put(eventfd);
 440
 441         fdput(f);
 442
 443 out:
 444         kfree(irqfd);
 445         return ret;
 446 }
 447 #endif
 448
 449 void
 450 kvm_eventfd_init(struct kvm *kvm)
 451 {
 452 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
 453         spin_lock_init(&kvm->irqfds.lock);
 454         INIT_LIST_HEAD(&kvm->irqfds.items);
 455         INIT_LIST_HEAD(&kvm->irqfds.resampler_list);
 456         mutex_init(&kvm->irqfds.resampler_lock);
 457 #endif
 458         INIT_LIST_HEAD(&kvm->ioeventfds);
 459 }
 460
 461 #ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
 462 /*
 463  * shutdown any irqfd's that match fd+gsi
 464  */
 465 static int
 466 kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
 467 {
 468         struct _irqfd *irqfd, *tmp;
 469         struct eventfd_ctx *eventfd;
 470
 471         eventfd = eventfd_ctx_fdget(args->fd);
 472         if (IS_ERR(eventfd))
 473                 return PTR_ERR(eventfd);
 474
 475         spin_lock_irq(&kvm->irqfds.lock);
 476
 477         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
 478                 if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
 479                         /*
 480                          * This clearing of irq_entry.type is needed for when
 481                          * another thread calls kvm_irq_routing_update before
 482                          * we flush workqueue below (we synchronize with
 483                          * kvm_irq_routing_update using irqfds.lock).
 484                          */
 485                         write_seqcount_begin(&irqfd->irq_entry_sc);
 486                         irqfd->irq_entry.type = 0;
 487                         write_seqcount_end(&irqfd->irq_entry_sc);
 488                         irqfd_deactivate(irqfd);
 489                 }
 490         }
 491
 492         spin_unlock_irq(&kvm->irqfds.lock);
 493         eventfd_ctx_put(eventfd);
 494
 495         /*
 496          * Block until we know all outstanding shutdown jobs have completed
 497          * so that we guarantee there will not be any more interrupts on this
 498          * gsi once this deassign function returns.
 499          */
 500         flush_workqueue(irqfd_cleanup_wq);
 501
 502         return 0;
 503 }
 504
 505 int
 506 kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
 507 {
 508         if (args->flags & ~(KVM_IRQFD_FLAG_DEASSIGN | KVM_IRQFD_FLAG_RESAMPLE))
 509                 return -EINVAL;
 510
 511         if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
 512                 return kvm_irqfd_deassign(kvm, args);
 513
 514         return kvm_irqfd_assign(kvm, args);
 515 }
 516
 517 /*
 518  * This function is called as the kvm VM fd is being released. Shutdown all
 519  * irqfds that still remain open
 520  */
 521 void
 522 kvm_irqfd_release(struct kvm *kvm)
 523 {
 524         struct _irqfd *irqfd, *tmp;
 525
 526         spin_lock_irq(&kvm->irqfds.lock);
 527
 528         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
 529                 irqfd_deactivate(irqfd);
 530
 531         spin_unlock_irq(&kvm->irqfds.lock);
 532
 533         /*
 534          * Block until we know all outstanding shutdown jobs have completed
 535          * since we do not take a kvm* reference.
 536          */
 537         flush_workqueue(irqfd_cleanup_wq);
 538
 539 }
 540
 541 /*
 542  * Change irq_routing and irqfd.
 543  * Caller must invoke synchronize_srcu(&kvm->irq_srcu) afterwards.
 544  */
 545 void kvm_irq_routing_update(struct kvm *kvm,
 546                             struct kvm_irq_routing_table *irq_rt)
 547 {
 548         struct _irqfd *irqfd;
 549
 550         spin_lock_irq(&kvm->irqfds.lock);
 551
 552         rcu_assign_pointer(kvm->irq_routing, irq_rt);
 553
 554         list_for_each_entry(irqfd, &kvm->irqfds.items, list)
 555                 irqfd_update(kvm, irqfd, irq_rt);
 556
 557         spin_unlock_irq(&kvm->irqfds.lock);
 558 }
 559
 560 /*
 561  * create a host-wide workqueue for issuing deferred shutdown requests
 562  * aggregated from all vm* instances. We need our own isolated single-thread
 563  * queue to prevent deadlock against flushing the normal work-queue.
 564  */
 565 int kvm_irqfd_init(void)
 566 {
 567         irqfd_cleanup_wq = create_singlethread_workqueue("kvm-irqfd-cleanup");
 568         if (!irqfd_cleanup_wq)
 569                 return -ENOMEM;
 570
 571         return 0;
 572 }
 573
 574 void kvm_irqfd_exit(void)
 575 {
 576         destroy_workqueue(irqfd_cleanup_wq);
 577 }
 578 #endif
 579
 580 /*
 581  * --------------------------------------------------------------------
 582  * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
 583  *
 584  * userspace can register a PIO/MMIO address with an eventfd for receiving
 585  * notification when the memory has been touched.
 586  * --------------------------------------------------------------------
 587  */
 588
 589 struct _ioeventfd {
 590         struct list_head     list;
 591         u64                  addr;
 592         int                  length;
 593         struct eventfd_ctx  *eventfd;
 594         u64                  datamatch;
 595         struct kvm_io_device dev;
 596         u8                   bus_idx;
 597         bool                 wildcard;
 598 };
 599
 600 static inline struct _ioeventfd *
 601 to_ioeventfd(struct kvm_io_device *dev)
 602 {
 603         return container_of(dev, struct _ioeventfd, dev);
 604 }
 605
 606 static void
 607 ioeventfd_release(struct _ioeventfd *p)
 608 {
 609         eventfd_ctx_put(p->eventfd);
 610         list_del(&p->list);
 611         kfree(p);
 612 }
 613
 614 static bool
 615 ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
 616 {
 617         u64 _val;
 618
 619         if (addr != p->addr)
 620                 /* address must be precise for a hit */
 621                 return false;
 622
 623         if (!p->length)
 624                 /* length = 0 means only look at the address, so always a hit */
 625                 return true;
 626
 627         if (len != p->length)
 628                 /* address-range must be precise for a hit */
 629                 return false;
 630
 631         if (p->wildcard)
 632                 /* all else equal, wildcard is always a hit */
 633                 return true;
 634
 635         /* otherwise, we have to actually compare the data */
 636
 637         BUG_ON(!IS_ALIGNED((unsigned long)val, len));
 638
 639         switch (len) {
 640         case 1:
 641                 _val = *(u8 *)val;
 642                 break;
 643         case 2:
 644                 _val = *(u16 *)val;
 645                 break;
 646         case 4:
 647                 _val = *(u32 *)val;
 648                 break;
 649         case 8:
 650                 _val = *(u64 *)val;
 651                 break;
 652         default:
 653                 return false;
 654         }
 655
 656         return _val == p->datamatch ? true : false;
 657 }
 658
 659 /* MMIO/PIO writes trigger an event if the addr/val match */
 660 static int
 661 ioeventfd_write(struct kvm_io_device *this, gpa_t addr, int len,
 662                 const void *val)
 663 {
 664         struct _ioeventfd *p = to_ioeventfd(this);
 665
 666         if (!ioeventfd_in_range(p, addr, len, val))
 667                 return -EOPNOTSUPP;
 668
 669         eventfd_signal(p->eventfd, 1);
 670         return 0;
 671 }
 672
 673 /*
 674  * This function is called as KVM is completely shutting down.  We do not
 675  * need to worry about locking just nuke anything we have as quickly as possible
 676  */
 677 static void
 678 ioeventfd_destructor(struct kvm_io_device *this)
 679 {
 680         struct _ioeventfd *p = to_ioeventfd(this);
 681
 682         ioeventfd_release(p);
 683 }
 684
 685 static const struct kvm_io_device_ops ioeventfd_ops = {
 686         .write      = ioeventfd_write,
 687         .destructor = ioeventfd_destructor,
 688 };
 689
 690 /* assumes kvm->slots_lock held */
 691 static bool
 692 ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
 693 {
 694         struct _ioeventfd *_p;
 695
 696         list_for_each_entry(_p, &kvm->ioeventfds, list)
 697                 if (_p->bus_idx == p->bus_idx &&
 698                     _p->addr == p->addr &&
 699                     (!_p->length || !p->length ||
 700                      (_p->length == p->length &&
 701                       (_p->wildcard || p->wildcard ||
 702                        _p->datamatch == p->datamatch))))
 703                         return true;
 704
 705         return false;
 706 }
 707
 708 static enum kvm_bus ioeventfd_bus_from_flags(__u32 flags)
 709 {
 710         if (flags & KVM_IOEVENTFD_FLAG_PIO)
 711                 return KVM_PIO_BUS;
 712         if (flags & KVM_IOEVENTFD_FLAG_VIRTIO_CCW_NOTIFY)
 713                 return KVM_VIRTIO_CCW_NOTIFY_BUS;
 714         return KVM_MMIO_BUS;
 715 }
 716
 717 static int
 718 kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 719 {
 720         enum kvm_bus              bus_idx;
 721         struct _ioeventfd        *p;
 722         struct eventfd_ctx       *eventfd;
 723         int                       ret;
 724
 725         bus_idx = ioeventfd_bus_from_flags(args->flags);
 726         /* must be natural-word sized, or 0 to ignore length */
 727         switch (args->len) {
 728         case 0:
 729         case 1:
 730         case 2:
 731         case 4:
 732         case 8:
 733                 break;
 734         default:
 735                 return -EINVAL;
 736         }
 737
 738         /* check for range overflow */
 739         if (args->addr + args->len < args->addr)
 740                 return -EINVAL;
 741
 742         /* check for extra flags that we don't understand */
 743         if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
 744                 return -EINVAL;
 745
 746         /* ioeventfd with no length can't be combined with DATAMATCH */
 747         if (!args->len &&
 748             args->flags & (KVM_IOEVENTFD_FLAG_PIO |
 749                            KVM_IOEVENTFD_FLAG_DATAMATCH))
 750                 return -EINVAL;
 751
 752         eventfd = eventfd_ctx_fdget(args->fd);
 753         if (IS_ERR(eventfd))
 754                 return PTR_ERR(eventfd);
 755
 756         p = kzalloc(sizeof(*p), GFP_KERNEL);
 757         if (!p) {
 758                 ret = -ENOMEM;
 759                 goto fail;
 760         }
 761
 762         INIT_LIST_HEAD(&p->list);
 763         p->addr    = args->addr;
 764         p->bus_idx = bus_idx;
 765         p->length  = args->len;
 766         p->eventfd = eventfd;
 767
 768         /* The datamatch feature is optional, otherwise this is a wildcard */
 769         if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
 770                 p->datamatch = args->datamatch;
 771         else
 772                 p->wildcard = true;
 773
 774         mutex_lock(&kvm->slots_lock);
 775
 776         /* Verify that there isn't a match already */
 777         if (ioeventfd_check_collision(kvm, p)) {
 778                 ret = -EEXIST;
 779                 goto unlock_fail;
 780         }
 781
 782         kvm_iodevice_init(&p->dev, &ioeventfd_ops);
 783
 784         ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
 785                                       &p->dev);
 786         if (ret < 0)
 787                 goto unlock_fail;
 788
 789         /* When length is ignored, MMIO is also put on a separate bus, for
 790          * faster lookups.
 791          */
 792         if (!args->len && !(args->flags & KVM_IOEVENTFD_FLAG_PIO)) {
 793                 ret = kvm_io_bus_register_dev(kvm, KVM_FAST_MMIO_BUS,
 794                                               p->addr, 0, &p->dev);
 795                 if (ret < 0)
 796                         goto register_fail;
 797         }
 798
 799         kvm->buses[bus_idx]->ioeventfd_count++;
 800         list_add_tail(&p->list, &kvm->ioeventfds);
 801
 802         mutex_unlock(&kvm->slots_lock);
 803
 804         return 0;
 805
 806 register_fail:
 807         kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
 808 unlock_fail:
 809         mutex_unlock(&kvm->slots_lock);
 810
 811 fail:
 812         kfree(p);
 813         eventfd_ctx_put(eventfd);
 814
 815         return ret;
 816 }
 817
 818 static int
 819 kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 820 {
 821         enum kvm_bus              bus_idx;
 822         struct _ioeventfd        *p, *tmp;
 823         struct eventfd_ctx       *eventfd;
 824         int                       ret = -ENOENT;
 825
 826         bus_idx = ioeventfd_bus_from_flags(args->flags);
 827         eventfd = eventfd_ctx_fdget(args->fd);
 828         if (IS_ERR(eventfd))
 829                 return PTR_ERR(eventfd);
 830
 831         mutex_lock(&kvm->slots_lock);
 832
 833         list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
 834                 bool wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
 835
 836                 if (p->bus_idx != bus_idx ||
 837                     p->eventfd != eventfd  ||
 838                     p->addr != args->addr  ||
 839                     p->length != args->len ||
 840                     p->wildcard != wildcard)
 841                         continue;
 842
 843                 if (!p->wildcard && p->datamatch != args->datamatch)
 844                         continue;
 845
 846                 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
 847                 if (!p->length) {
 848                         kvm_io_bus_unregister_dev(kvm, KVM_FAST_MMIO_BUS,
 849                                                   &p->dev);
 850                 }
 851                 kvm->buses[bus_idx]->ioeventfd_count--;
 852                 ioeventfd_release(p);
 853                 ret = 0;
 854                 break;
 855         }
 856
 857         mutex_unlock(&kvm->slots_lock);
 858
 859         eventfd_ctx_put(eventfd);
 860
 861         return ret;
 862 }
 863
 864 int
 865 kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 866 {
 867         if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
 868                 return kvm_deassign_ioeventfd(kvm, args);
 869
 870         return kvm_assign_ioeventfd(kvm, args);
 871 }