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 <[email protected]>
   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/kvm_irqfd.h>
  27 #include <linux/workqueue.h>
  28 #include <linux/syscalls.h>
  29 #include <linux/wait.h>
  30 #include <linux/poll.h>
  31 #include <linux/file.h>
  32 #include <linux/list.h>
  33 #include <linux/eventfd.h>
  34 #include <linux/kernel.h>
  35 #include <linux/srcu.h>
  36 #include <linux/slab.h>
  37 #include <linux/seqlock.h>
  38 #include <linux/irqbypass.h>
  39 #include <trace/events/kvm.h>
  40
  41 #include <kvm/iodev.h>
  42
  43 #ifdef CONFIG_HAVE_KVM_IRQFD
  44
  45 static struct workqueue_struct *irqfd_cleanup_wq;
  46
  47 static void
  48 irqfd_inject(struct work_struct *work)
  49 {
  50         struct kvm_kernel_irqfd *irqfd =
  51                 container_of(work, struct kvm_kernel_irqfd, inject);
  52         struct kvm *kvm = irqfd->kvm;
  53
  54         if (!irqfd->resampler) {
  55                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1,
  56                                 false);
  57                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0,
  58                                 false);
  59         } else
  60                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
  61                             irqfd->gsi, 1, false);
  62 }
  63
  64 /*
  65  * Since resampler irqfds share an IRQ source ID, we de-assert once
  66  * then notify all of the resampler irqfds using this GSI.  We can't
  67  * do multiple de-asserts or we risk racing with incoming re-asserts.
  68  */
  69 static void
  70 irqfd_resampler_ack(struct kvm_irq_ack_notifier *kian)
  71 {
  72         struct kvm_kernel_irqfd_resampler *resampler;
  73         struct kvm *kvm;
  74         struct kvm_kernel_irqfd *irqfd;
  75         int idx;
  76
  77         resampler = container_of(kian,
  78                         struct kvm_kernel_irqfd_resampler, notifier);
  79         kvm = resampler->kvm;
  80
  81         kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
  82                     resampler->notifier.gsi, 0, false);
  83
  84         idx = srcu_read_lock(&kvm->irq_srcu);
  85
  86         list_for_each_entry_rcu(irqfd, &resampler->list, resampler_link)
  87                 eventfd_signal(irqfd->resamplefd, 1);
  88
  89         srcu_read_unlock(&kvm->irq_srcu, idx);
  90 }
  91
  92 static void
  93 irqfd_resampler_shutdown(struct kvm_kernel_irqfd *irqfd)
  94 {
  95         struct kvm_kernel_irqfd_resampler *resampler = irqfd->resampler;
  96         struct kvm *kvm = resampler->kvm;
  97
  98         mutex_lock(&kvm->irqfds.resampler_lock);
  99
 100         list_del_rcu(&irqfd->resampler_link);
 101         synchronize_srcu(&kvm->irq_srcu);
 102
 103         if (list_empty(&resampler->list)) {
 104                 list_del(&resampler->link);
 105                 kvm_unregister_irq_ack_notifier(kvm, &resampler->notifier);
 106                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 107                             resampler->notifier.gsi, 0, false);
 108                 kfree(resampler);
 109         }
 110
 111         mutex_unlock(&kvm->irqfds.resampler_lock);
 112 }
 113
 114 /*
 115  * Race-free decouple logic (ordering is critical)
 116  */
 117 static void
 118 irqfd_shutdown(struct work_struct *work)
 119 {
 120         struct kvm_kernel_irqfd *irqfd =
 121                 container_of(work, struct kvm_kernel_irqfd, shutdown);
 122         u64 cnt;
 123
 124         /*
 125          * Synchronize with the wait-queue and unhook ourselves to prevent
 126          * further events.
 127          */
 128         eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
 129
 130         /*
 131          * We know no new events will be scheduled at this point, so block
 132          * until all previously outstanding events have completed
 133          */
 134         flush_work(&irqfd->inject);
 135
 136         if (irqfd->resampler) {
 137                 irqfd_resampler_shutdown(irqfd);
 138                 eventfd_ctx_put(irqfd->resamplefd);
 139         }
 140
 141         /*
 142          * It is now safe to release the object's resources
 143          */
 144 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 145         irq_bypass_unregister_consumer(&irqfd->consumer);
 146 #endif
 147         eventfd_ctx_put(irqfd->eventfd);
 148         kfree(irqfd);
 149 }
 150
 151
 152 /* assumes kvm->irqfds.lock is held */
 153 static bool
 154 irqfd_is_active(struct kvm_kernel_irqfd *irqfd)
 155 {
 156         return list_empty(&irqfd->list) ? false : true;
 157 }
 158
 159 /*
 160  * Mark the irqfd as inactive and schedule it for removal
 161  *
 162  * assumes kvm->irqfds.lock is held
 163  */
 164 static void
 165 irqfd_deactivate(struct kvm_kernel_irqfd *irqfd)
 166 {
 167         BUG_ON(!irqfd_is_active(irqfd));
 168
 169         list_del_init(&irqfd->list);
 170
 171         queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
 172 }
 173
 174 int __attribute__((weak)) kvm_arch_set_irq_inatomic(
 175                                 struct kvm_kernel_irq_routing_entry *irq,
 176                                 struct kvm *kvm, int irq_source_id,
 177                                 int level,
 178                                 bool line_status)
 179 {
 180         return -EWOULDBLOCK;
 181 }
 182
 183 /*
 184  * Called with wqh->lock held and interrupts disabled
 185  */
 186 static int
 187 irqfd_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
 188 {
 189         struct kvm_kernel_irqfd *irqfd =
 190                 container_of(wait, struct kvm_kernel_irqfd, wait);
 191         __poll_t flags = key_to_poll(key);
 192         struct kvm_kernel_irq_routing_entry irq;
 193         struct kvm *kvm = irqfd->kvm;
 194         unsigned seq;
 195         int idx;
 196
 197         if (flags & EPOLLIN) {
 198                 idx = srcu_read_lock(&kvm->irq_srcu);
 199                 do {
 200                         seq = read_seqcount_begin(&irqfd->irq_entry_sc);
 201                         irq = irqfd->irq_entry;
 202                 } while (read_seqcount_retry(&irqfd->irq_entry_sc, seq));
 203                 /* An event has been signaled, inject an interrupt */
 204                 if (kvm_arch_set_irq_inatomic(&irq, kvm,
 205                                               KVM_USERSPACE_IRQ_SOURCE_ID, 1,
 206                                               false) == -EWOULDBLOCK)
 207                         schedule_work(&irqfd->inject);
 208                 srcu_read_unlock(&kvm->irq_srcu, idx);
 209         }
 210
 211         if (flags & EPOLLHUP) {
 212                 /* The eventfd is closing, detach from KVM */
 213                 unsigned long flags;
 214
 215                 spin_lock_irqsave(&kvm->irqfds.lock, flags);
 216
 217                 /*
 218                  * We must check if someone deactivated the irqfd before
 219                  * we could acquire the irqfds.lock since the item is
 220                  * deactivated from the KVM side before it is unhooked from
 221                  * the wait-queue.  If it is already deactivated, we can
 222                  * simply return knowing the other side will cleanup for us.
 223                  * We cannot race against the irqfd going away since the
 224                  * other side is required to acquire wqh->lock, which we hold
 225                  */
 226                 if (irqfd_is_active(irqfd))
 227                         irqfd_deactivate(irqfd);
 228
 229                 spin_unlock_irqrestore(&kvm->irqfds.lock, flags);
 230         }
 231
 232         return 0;
 233 }
 234
 235 static void
 236 irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
 237                         poll_table *pt)
 238 {
 239         struct kvm_kernel_irqfd *irqfd =
 240                 container_of(pt, struct kvm_kernel_irqfd, pt);
 241         add_wait_queue(wqh, &irqfd->wait);
 242 }
 243
 244 /* Must be called under irqfds.lock */
 245 static void irqfd_update(struct kvm *kvm, struct kvm_kernel_irqfd *irqfd)
 246 {
 247         struct kvm_kernel_irq_routing_entry *e;
 248         struct kvm_kernel_irq_routing_entry entries[KVM_NR_IRQCHIPS];
 249         int n_entries;
 250
 251         n_entries = kvm_irq_map_gsi(kvm, entries, irqfd->gsi);
 252
 253         write_seqcount_begin(&irqfd->irq_entry_sc);
 254
 255         e = entries;
 256         if (n_entries == 1)
 257                 irqfd->irq_entry = *e;
 258         else
 259                 irqfd->irq_entry.type = 0;
 260
 261         write_seqcount_end(&irqfd->irq_entry_sc);
 262 }
 263
 264 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 265 void __attribute__((weak)) kvm_arch_irq_bypass_stop(
 266                                 struct irq_bypass_consumer *cons)
 267 {
 268 }
 269
 270 void __attribute__((weak)) kvm_arch_irq_bypass_start(
 271                                 struct irq_bypass_consumer *cons)
 272 {
 273 }
 274
 275 int  __attribute__((weak)) kvm_arch_update_irqfd_routing(
 276                                 struct kvm *kvm, unsigned int host_irq,
 277                                 uint32_t guest_irq, bool set)
 278 {
 279         return 0;
 280 }
 281 #endif
 282
 283 static int
 284 kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
 285 {
 286         struct kvm_kernel_irqfd *irqfd, *tmp;
 287         struct fd f;
 288         struct eventfd_ctx *eventfd = NULL, *resamplefd = NULL;
 289         int ret;
 290         __poll_t events;
 291         int idx;
 292
 293         if (!kvm_arch_intc_initialized(kvm))
 294                 return -EAGAIN;
 295
 296         irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL);
 297         if (!irqfd)
 298                 return -ENOMEM;
 299
 300         irqfd->kvm = kvm;
 301         irqfd->gsi = args->gsi;
 302         INIT_LIST_HEAD(&irqfd->list);
 303         INIT_WORK(&irqfd->inject, irqfd_inject);
 304         INIT_WORK(&irqfd->shutdown, irqfd_shutdown);
 305         seqcount_init(&irqfd->irq_entry_sc);
 306
 307         f = fdget(args->fd);
 308         if (!f.file) {
 309                 ret = -EBADF;
 310                 goto out;
 311         }
 312
 313         eventfd = eventfd_ctx_fileget(f.file);
 314         if (IS_ERR(eventfd)) {
 315                 ret = PTR_ERR(eventfd);
 316                 goto fail;
 317         }
 318
 319         irqfd->eventfd = eventfd;
 320
 321         if (args->flags & KVM_IRQFD_FLAG_RESAMPLE) {
 322                 struct kvm_kernel_irqfd_resampler *resampler;
 323
 324                 resamplefd = eventfd_ctx_fdget(args->resamplefd);
 325                 if (IS_ERR(resamplefd)) {
 326                         ret = PTR_ERR(resamplefd);
 327                         goto fail;
 328                 }
 329
 330                 irqfd->resamplefd = resamplefd;
 331                 INIT_LIST_HEAD(&irqfd->resampler_link);
 332
 333                 mutex_lock(&kvm->irqfds.resampler_lock);
 334
 335                 list_for_each_entry(resampler,
 336                                     &kvm->irqfds.resampler_list, link) {
 337                         if (resampler->notifier.gsi == irqfd->gsi) {
 338                                 irqfd->resampler = resampler;
 339                                 break;
 340                         }
 341                 }
 342
 343                 if (!irqfd->resampler) {
 344                         resampler = kzalloc(sizeof(*resampler), GFP_KERNEL);
 345                         if (!resampler) {
 346                                 ret = -ENOMEM;
 347                                 mutex_unlock(&kvm->irqfds.resampler_lock);
 348                                 goto fail;
 349                         }
 350
 351                         resampler->kvm = kvm;
 352                         INIT_LIST_HEAD(&resampler->list);
 353                         resampler->notifier.gsi = irqfd->gsi;
 354                         resampler->notifier.irq_acked = irqfd_resampler_ack;
 355                         INIT_LIST_HEAD(&resampler->link);
 356
 357                         list_add(&resampler->link, &kvm->irqfds.resampler_list);
 358                         kvm_register_irq_ack_notifier(kvm,
 359                                                       &resampler->notifier);
 360                         irqfd->resampler = resampler;
 361                 }
 362
 363                 list_add_rcu(&irqfd->resampler_link, &irqfd->resampler->list);
 364                 synchronize_srcu(&kvm->irq_srcu);
 365
 366                 mutex_unlock(&kvm->irqfds.resampler_lock);
 367         }
 368
 369         /*
 370          * Install our own custom wake-up handling so we are notified via
 371          * a callback whenever someone signals the underlying eventfd
 372          */
 373         init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
 374         init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
 375
 376         spin_lock_irq(&kvm->irqfds.lock);
 377
 378         ret = 0;
 379         list_for_each_entry(tmp, &kvm->irqfds.items, list) {
 380                 if (irqfd->eventfd != tmp->eventfd)
 381                         continue;
 382                 /* This fd is used for another irq already. */
 383                 ret = -EBUSY;
 384                 spin_unlock_irq(&kvm->irqfds.lock);
 385                 goto fail;
 386         }
 387
 388         idx = srcu_read_lock(&kvm->irq_srcu);
 389         irqfd_update(kvm, irqfd);
 390         srcu_read_unlock(&kvm->irq_srcu, idx);
 391
 392         list_add_tail(&irqfd->list, &kvm->irqfds.items);
 393
 394         spin_unlock_irq(&kvm->irqfds.lock);
 395
 396         /*
 397          * Check if there was an event already pending on the eventfd
 398          * before we registered, and trigger it as if we didn't miss it.
 399          */
 400         events = f.file->f_op->poll(f.file, &irqfd->pt);
 401
 402         if (events & EPOLLIN)
 403                 schedule_work(&irqfd->inject);
 404
 405         /*
 406          * do not drop the file until the irqfd is fully initialized, otherwise
 407          * we might race against the EPOLLHUP
 408          */
 409         fdput(f);
 410 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 411         if (kvm_arch_has_irq_bypass()) {
 412                 irqfd->consumer.token = (void *)irqfd->eventfd;
 413                 irqfd->consumer.add_producer = kvm_arch_irq_bypass_add_producer;
 414                 irqfd->consumer.del_producer = kvm_arch_irq_bypass_del_producer;
 415                 irqfd->consumer.stop = kvm_arch_irq_bypass_stop;
 416                 irqfd->consumer.start = kvm_arch_irq_bypass_start;
 417                 ret = irq_bypass_register_consumer(&irqfd->consumer);
 418                 if (ret)
 419                         pr_info("irq bypass consumer (token %p) registration fails: %d\n",
 420                                 irqfd->consumer.token, ret);
 421         }
 422 #endif
 423
 424         return 0;
 425
 426 fail:
 427         if (irqfd->resampler)
 428                 irqfd_resampler_shutdown(irqfd);
 429
 430         if (resamplefd && !IS_ERR(resamplefd))
 431                 eventfd_ctx_put(resamplefd);
 432
 433         if (eventfd && !IS_ERR(eventfd))
 434                 eventfd_ctx_put(eventfd);
 435
 436         fdput(f);
 437
 438 out:
 439         kfree(irqfd);
 440         return ret;
 441 }
 442
 443 bool kvm_irq_has_notifier(struct kvm *kvm, unsigned irqchip, unsigned pin)
 444 {
 445         struct kvm_irq_ack_notifier *kian;
 446         int gsi, idx;
 447
 448         idx = srcu_read_lock(&kvm->irq_srcu);
 449         gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
 450         if (gsi != -1)
 451                 hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
 452                                          link)
 453                         if (kian->gsi == gsi) {
 454                                 srcu_read_unlock(&kvm->irq_srcu, idx);
 455                                 return true;
 456                         }
 457
 458         srcu_read_unlock(&kvm->irq_srcu, idx);
 459
 460         return false;
 461 }
 462 EXPORT_SYMBOL_GPL(kvm_irq_has_notifier);
 463
 464 void kvm_notify_acked_gsi(struct kvm *kvm, int gsi)
 465 {
 466         struct kvm_irq_ack_notifier *kian;
 467
 468         hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
 469                                  link)
 470                 if (kian->gsi == gsi)
 471                         kian->irq_acked(kian);
 472 }
 473
 474 void kvm_notify_acked_irq(struct kvm *kvm, unsigned irqchip, unsigned pin)
 475 {
 476         int gsi, idx;
 477
 478         trace_kvm_ack_irq(irqchip, pin);
 479
 480         idx = srcu_read_lock(&kvm->irq_srcu);
 481         gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
 482         if (gsi != -1)
 483                 kvm_notify_acked_gsi(kvm, gsi);
 484         srcu_read_unlock(&kvm->irq_srcu, idx);
 485 }
 486
 487 void kvm_register_irq_ack_notifier(struct kvm *kvm,
 488                                    struct kvm_irq_ack_notifier *kian)
 489 {
 490         mutex_lock(&kvm->irq_lock);
 491         hlist_add_head_rcu(&kian->link, &kvm->irq_ack_notifier_list);
 492         mutex_unlock(&kvm->irq_lock);
 493         kvm_arch_post_irq_ack_notifier_list_update(kvm);
 494 }
 495
 496 void kvm_unregister_irq_ack_notifier(struct kvm *kvm,
 497                                     struct kvm_irq_ack_notifier *kian)
 498 {
 499         mutex_lock(&kvm->irq_lock);
 500         hlist_del_init_rcu(&kian->link);
 501         mutex_unlock(&kvm->irq_lock);
 502         synchronize_srcu(&kvm->irq_srcu);
 503         kvm_arch_post_irq_ack_notifier_list_update(kvm);
 504 }
 505 #endif
 506
 507 void
 508 kvm_eventfd_init(struct kvm *kvm)
 509 {
 510 #ifdef CONFIG_HAVE_KVM_IRQFD
 511         spin_lock_init(&kvm->irqfds.lock);
 512         INIT_LIST_HEAD(&kvm->irqfds.items);
 513         INIT_LIST_HEAD(&kvm->irqfds.resampler_list);
 514         mutex_init(&kvm->irqfds.resampler_lock);
 515 #endif
 516         INIT_LIST_HEAD(&kvm->ioeventfds);
 517 }
 518
 519 #ifdef CONFIG_HAVE_KVM_IRQFD
 520 /*
 521  * shutdown any irqfd's that match fd+gsi
 522  */
 523 static int
 524 kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
 525 {
 526         struct kvm_kernel_irqfd *irqfd, *tmp;
 527         struct eventfd_ctx *eventfd;
 528
 529         eventfd = eventfd_ctx_fdget(args->fd);
 530         if (IS_ERR(eventfd))
 531                 return PTR_ERR(eventfd);
 532
 533         spin_lock_irq(&kvm->irqfds.lock);
 534
 535         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
 536                 if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
 537                         /*
 538                          * This clearing of irq_entry.type is needed for when
 539                          * another thread calls kvm_irq_routing_update before
 540                          * we flush workqueue below (we synchronize with
 541                          * kvm_irq_routing_update using irqfds.lock).
 542                          */
 543                         write_seqcount_begin(&irqfd->irq_entry_sc);
 544                         irqfd->irq_entry.type = 0;
 545                         write_seqcount_end(&irqfd->irq_entry_sc);
 546                         irqfd_deactivate(irqfd);
 547                 }
 548         }
 549
 550         spin_unlock_irq(&kvm->irqfds.lock);
 551         eventfd_ctx_put(eventfd);
 552
 553         /*
 554          * Block until we know all outstanding shutdown jobs have completed
 555          * so that we guarantee there will not be any more interrupts on this
 556          * gsi once this deassign function returns.
 557          */
 558         flush_workqueue(irqfd_cleanup_wq);
 559
 560         return 0;
 561 }
 562
 563 int
 564 kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
 565 {
 566         if (args->flags & ~(KVM_IRQFD_FLAG_DEASSIGN | KVM_IRQFD_FLAG_RESAMPLE))
 567                 return -EINVAL;
 568
 569         if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
 570                 return kvm_irqfd_deassign(kvm, args);
 571
 572         return kvm_irqfd_assign(kvm, args);
 573 }
 574
 575 /*
 576  * This function is called as the kvm VM fd is being released. Shutdown all
 577  * irqfds that still remain open
 578  */
 579 void
 580 kvm_irqfd_release(struct kvm *kvm)
 581 {
 582         struct kvm_kernel_irqfd *irqfd, *tmp;
 583
 584         spin_lock_irq(&kvm->irqfds.lock);
 585
 586         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
 587                 irqfd_deactivate(irqfd);
 588
 589         spin_unlock_irq(&kvm->irqfds.lock);
 590
 591         /*
 592          * Block until we know all outstanding shutdown jobs have completed
 593          * since we do not take a kvm* reference.
 594          */
 595         flush_workqueue(irqfd_cleanup_wq);
 596
 597 }
 598
 599 /*
 600  * Take note of a change in irq routing.
 601  * Caller must invoke synchronize_srcu(&kvm->irq_srcu) afterwards.
 602  */
 603 void kvm_irq_routing_update(struct kvm *kvm)
 604 {
 605         struct kvm_kernel_irqfd *irqfd;
 606
 607         spin_lock_irq(&kvm->irqfds.lock);
 608
 609         list_for_each_entry(irqfd, &kvm->irqfds.items, list) {
 610                 irqfd_update(kvm, irqfd);
 611
 612 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 613                 if (irqfd->producer) {
 614                         int ret = kvm_arch_update_irqfd_routing(
 615                                         irqfd->kvm, irqfd->producer->irq,
 616                                         irqfd->gsi, 1);
 617                         WARN_ON(ret);
 618                 }
 619 #endif
 620         }
 621
 622         spin_unlock_irq(&kvm->irqfds.lock);
 623 }
 624
 625 /*
 626  * create a host-wide workqueue for issuing deferred shutdown requests
 627  * aggregated from all vm* instances. We need our own isolated
 628  * queue to ease flushing work items when a VM exits.
 629  */
 630 int kvm_irqfd_init(void)
 631 {
 632         irqfd_cleanup_wq = alloc_workqueue("kvm-irqfd-cleanup", 0, 0);
 633         if (!irqfd_cleanup_wq)
 634                 return -ENOMEM;
 635
 636         return 0;
 637 }
 638
 639 void kvm_irqfd_exit(void)
 640 {
 641         destroy_workqueue(irqfd_cleanup_wq);
 642 }
 643 #endif
 644
 645 /*
 646  * --------------------------------------------------------------------
 647  * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
 648  *
 649  * userspace can register a PIO/MMIO address with an eventfd for receiving
 650  * notification when the memory has been touched.
 651  * --------------------------------------------------------------------
 652  */
 653
 654 struct _ioeventfd {
 655         struct list_head     list;
 656         u64                  addr;
 657         int                  length;
 658         struct eventfd_ctx  *eventfd;
 659         u64                  datamatch;
 660         struct kvm_io_device dev;
 661         u8                   bus_idx;
 662         bool                 wildcard;
 663 };
 664
 665 static inline struct _ioeventfd *
 666 to_ioeventfd(struct kvm_io_device *dev)
 667 {
 668         return container_of(dev, struct _ioeventfd, dev);
 669 }
 670
 671 static void
 672 ioeventfd_release(struct _ioeventfd *p)
 673 {
 674         eventfd_ctx_put(p->eventfd);
 675         list_del(&p->list);
 676         kfree(p);
 677 }
 678
 679 static bool
 680 ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
 681 {
 682         u64 _val;
 683
 684         if (addr != p->addr)
 685                 /* address must be precise for a hit */
 686                 return false;
 687
 688         if (!p->length)
 689                 /* length = 0 means only look at the address, so always a hit */
 690                 return true;
 691
 692         if (len != p->length)
 693                 /* address-range must be precise for a hit */
 694                 return false;
 695
 696         if (p->wildcard)
 697                 /* all else equal, wildcard is always a hit */
 698                 return true;
 699
 700         /* otherwise, we have to actually compare the data */
 701
 702         BUG_ON(!IS_ALIGNED((unsigned long)val, len));
 703
 704         switch (len) {
 705         case 1:
 706                 _val = *(u8 *)val;
 707                 break;
 708         case 2:
 709                 _val = *(u16 *)val;
 710                 break;
 711         case 4:
 712                 _val = *(u32 *)val;
 713                 break;
 714         case 8:
 715                 _val = *(u64 *)val;
 716                 break;
 717         default:
 718                 return false;
 719         }
 720
 721         return _val == p->datamatch ? true : false;
 722 }
 723
 724 /* MMIO/PIO writes trigger an event if the addr/val match */
 725 static int
 726 ioeventfd_write(struct kvm_vcpu *vcpu, struct kvm_io_device *this, gpa_t addr,
 727                 int len, const void *val)
 728 {
 729         struct _ioeventfd *p = to_ioeventfd(this);
 730
 731         if (!ioeventfd_in_range(p, addr, len, val))
 732                 return -EOPNOTSUPP;
 733
 734         eventfd_signal(p->eventfd, 1);
 735         return 0;
 736 }
 737
 738 /*
 739  * This function is called as KVM is completely shutting down.  We do not
 740  * need to worry about locking just nuke anything we have as quickly as possible
 741  */
 742 static void
 743 ioeventfd_destructor(struct kvm_io_device *this)
 744 {
 745         struct _ioeventfd *p = to_ioeventfd(this);
 746
 747         ioeventfd_release(p);
 748 }
 749
 750 static const struct kvm_io_device_ops ioeventfd_ops = {
 751         .write      = ioeventfd_write,
 752         .destructor = ioeventfd_destructor,
 753 };
 754
 755 /* assumes kvm->slots_lock held */
 756 static bool
 757 ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
 758 {
 759         struct _ioeventfd *_p;
 760
 761         list_for_each_entry(_p, &kvm->ioeventfds, list)
 762                 if (_p->bus_idx == p->bus_idx &&
 763                     _p->addr == p->addr &&
 764                     (!_p->length || !p->length ||
 765                      (_p->length == p->length &&
 766                       (_p->wildcard || p->wildcard ||
 767                        _p->datamatch == p->datamatch))))
 768                         return true;
 769
 770         return false;
 771 }
 772
 773 static enum kvm_bus ioeventfd_bus_from_flags(__u32 flags)
 774 {
 775         if (flags & KVM_IOEVENTFD_FLAG_PIO)
 776                 return KVM_PIO_BUS;
 777         if (flags & KVM_IOEVENTFD_FLAG_VIRTIO_CCW_NOTIFY)
 778                 return KVM_VIRTIO_CCW_NOTIFY_BUS;
 779         return KVM_MMIO_BUS;
 780 }
 781
 782 static int kvm_assign_ioeventfd_idx(struct kvm *kvm,
 783                                 enum kvm_bus bus_idx,
 784                                 struct kvm_ioeventfd *args)
 785 {
 786
 787         struct eventfd_ctx *eventfd;
 788         struct _ioeventfd *p;
 789         int ret;
 790
 791         eventfd = eventfd_ctx_fdget(args->fd);
 792         if (IS_ERR(eventfd))
 793                 return PTR_ERR(eventfd);
 794
 795         p = kzalloc(sizeof(*p), GFP_KERNEL);
 796         if (!p) {
 797                 ret = -ENOMEM;
 798                 goto fail;
 799         }
 800
 801         INIT_LIST_HEAD(&p->list);
 802         p->addr    = args->addr;
 803         p->bus_idx = bus_idx;
 804         p->length  = args->len;
 805         p->eventfd = eventfd;
 806
 807         /* The datamatch feature is optional, otherwise this is a wildcard */
 808         if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
 809                 p->datamatch = args->datamatch;
 810         else
 811                 p->wildcard = true;
 812
 813         mutex_lock(&kvm->slots_lock);
 814
 815         /* Verify that there isn't a match already */
 816         if (ioeventfd_check_collision(kvm, p)) {
 817                 ret = -EEXIST;
 818                 goto unlock_fail;
 819         }
 820
 821         kvm_iodevice_init(&p->dev, &ioeventfd_ops);
 822
 823         ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
 824                                       &p->dev);
 825         if (ret < 0)
 826                 goto unlock_fail;
 827
 828         kvm_get_bus(kvm, bus_idx)->ioeventfd_count++;
 829         list_add_tail(&p->list, &kvm->ioeventfds);
 830
 831         mutex_unlock(&kvm->slots_lock);
 832
 833         return 0;
 834
 835 unlock_fail:
 836         mutex_unlock(&kvm->slots_lock);
 837
 838 fail:
 839         kfree(p);
 840         eventfd_ctx_put(eventfd);
 841
 842         return ret;
 843 }
 844
 845 static int
 846 kvm_deassign_ioeventfd_idx(struct kvm *kvm, enum kvm_bus bus_idx,
 847                            struct kvm_ioeventfd *args)
 848 {
 849         struct _ioeventfd        *p, *tmp;
 850         struct eventfd_ctx       *eventfd;
 851         struct kvm_io_bus        *bus;
 852         int                       ret = -ENOENT;
 853
 854         eventfd = eventfd_ctx_fdget(args->fd);
 855         if (IS_ERR(eventfd))
 856                 return PTR_ERR(eventfd);
 857
 858         mutex_lock(&kvm->slots_lock);
 859
 860         list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
 861                 bool wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
 862
 863                 if (p->bus_idx != bus_idx ||
 864                     p->eventfd != eventfd  ||
 865                     p->addr != args->addr  ||
 866                     p->length != args->len ||
 867                     p->wildcard != wildcard)
 868                         continue;
 869
 870                 if (!p->wildcard && p->datamatch != args->datamatch)
 871                         continue;
 872
 873                 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
 874                 bus = kvm_get_bus(kvm, bus_idx);
 875                 if (bus)
 876                         bus->ioeventfd_count--;
 877                 ioeventfd_release(p);
 878                 ret = 0;
 879                 break;
 880         }
 881
 882         mutex_unlock(&kvm->slots_lock);
 883
 884         eventfd_ctx_put(eventfd);
 885
 886         return ret;
 887 }
 888
 889 static int kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 890 {
 891         enum kvm_bus bus_idx = ioeventfd_bus_from_flags(args->flags);
 892         int ret = kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
 893
 894         if (!args->len && bus_idx == KVM_MMIO_BUS)
 895                 kvm_deassign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
 896
 897         return ret;
 898 }
 899
 900 static int
 901 kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 902 {
 903         enum kvm_bus              bus_idx;
 904         int ret;
 905
 906         bus_idx = ioeventfd_bus_from_flags(args->flags);
 907         /* must be natural-word sized, or 0 to ignore length */
 908         switch (args->len) {
 909         case 0:
 910         case 1:
 911         case 2:
 912         case 4:
 913         case 8:
 914                 break;
 915         default:
 916                 return -EINVAL;
 917         }
 918
 919         /* check for range overflow */
 920         if (args->addr + args->len < args->addr)
 921                 return -EINVAL;
 922
 923         /* check for extra flags that we don't understand */
 924         if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
 925                 return -EINVAL;
 926
 927         /* ioeventfd with no length can't be combined with DATAMATCH */
 928         if (!args->len && (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH))
 929                 return -EINVAL;
 930
 931         ret = kvm_assign_ioeventfd_idx(kvm, bus_idx, args);
 932         if (ret)
 933                 goto fail;
 934
 935         /* When length is ignored, MMIO is also put on a separate bus, for
 936          * faster lookups.
 937          */
 938         if (!args->len && bus_idx == KVM_MMIO_BUS) {
 939                 ret = kvm_assign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
 940                 if (ret < 0)
 941                         goto fast_fail;
 942         }
 943
 944         return 0;
 945
 946 fast_fail:
 947         kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
 948 fail:
 949         return ret;
 950 }
 951
 952 int
 953 kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 954 {
 955         if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
 956                 return kvm_deassign_ioeventfd(kvm, args);
 957
 958         return kvm_assign_ioeventfd(kvm, args);
 959 }