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/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         struct kvm *kvm = irqfd->kvm;
 123         u64 cnt;
 124
 125         /* Make sure irqfd has been initalized in assign path. */
 126         synchronize_srcu(&kvm->irq_srcu);
 127
 128         /*
 129          * Synchronize with the wait-queue and unhook ourselves to prevent
 130          * further events.
 131          */
 132         eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
 133
 134         /*
 135          * We know no new events will be scheduled at this point, so block
 136          * until all previously outstanding events have completed
 137          */
 138         flush_work(&irqfd->inject);
 139
 140         if (irqfd->resampler) {
 141                 irqfd_resampler_shutdown(irqfd);
 142                 eventfd_ctx_put(irqfd->resamplefd);
 143         }
 144
 145         /*
 146          * It is now safe to release the object's resources
 147          */
 148 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 149         irq_bypass_unregister_consumer(&irqfd->consumer);
 150 #endif
 151         eventfd_ctx_put(irqfd->eventfd);
 152         kfree(irqfd);
 153 }
 154
 155
 156 /* assumes kvm->irqfds.lock is held */
 157 static bool
 158 irqfd_is_active(struct kvm_kernel_irqfd *irqfd)
 159 {
 160         return list_empty(&irqfd->list) ? false : true;
 161 }
 162
 163 /*
 164  * Mark the irqfd as inactive and schedule it for removal
 165  *
 166  * assumes kvm->irqfds.lock is held
 167  */
 168 static void
 169 irqfd_deactivate(struct kvm_kernel_irqfd *irqfd)
 170 {
 171         BUG_ON(!irqfd_is_active(irqfd));
 172
 173         list_del_init(&irqfd->list);
 174
 175         queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
 176 }
 177
 178 int __attribute__((weak)) kvm_arch_set_irq_inatomic(
 179                                 struct kvm_kernel_irq_routing_entry *irq,
 180                                 struct kvm *kvm, int irq_source_id,
 181                                 int level,
 182                                 bool line_status)
 183 {
 184         return -EWOULDBLOCK;
 185 }
 186
 187 /*
 188  * Called with wqh->lock held and interrupts disabled
 189  */
 190 static int
 191 irqfd_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
 192 {
 193         struct kvm_kernel_irqfd *irqfd =
 194                 container_of(wait, struct kvm_kernel_irqfd, wait);
 195         __poll_t flags = key_to_poll(key);
 196         struct kvm_kernel_irq_routing_entry irq;
 197         struct kvm *kvm = irqfd->kvm;
 198         unsigned seq;
 199         int idx;
 200
 201         if (flags & EPOLLIN) {
 202                 idx = srcu_read_lock(&kvm->irq_srcu);
 203                 do {
 204                         seq = read_seqcount_begin(&irqfd->irq_entry_sc);
 205                         irq = irqfd->irq_entry;
 206                 } while (read_seqcount_retry(&irqfd->irq_entry_sc, seq));
 207                 /* An event has been signaled, inject an interrupt */
 208                 if (kvm_arch_set_irq_inatomic(&irq, kvm,
 209                                               KVM_USERSPACE_IRQ_SOURCE_ID, 1,
 210                                               false) == -EWOULDBLOCK)
 211                         schedule_work(&irqfd->inject);
 212                 srcu_read_unlock(&kvm->irq_srcu, idx);
 213         }
 214
 215         if (flags & EPOLLHUP) {
 216                 /* The eventfd is closing, detach from KVM */
 217                 unsigned long flags;
 218
 219                 spin_lock_irqsave(&kvm->irqfds.lock, flags);
 220
 221                 /*
 222                  * We must check if someone deactivated the irqfd before
 223                  * we could acquire the irqfds.lock since the item is
 224                  * deactivated from the KVM side before it is unhooked from
 225                  * the wait-queue.  If it is already deactivated, we can
 226                  * simply return knowing the other side will cleanup for us.
 227                  * We cannot race against the irqfd going away since the
 228                  * other side is required to acquire wqh->lock, which we hold
 229                  */
 230                 if (irqfd_is_active(irqfd))
 231                         irqfd_deactivate(irqfd);
 232
 233                 spin_unlock_irqrestore(&kvm->irqfds.lock, flags);
 234         }
 235
 236         return 0;
 237 }
 238
 239 static void
 240 irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
 241                         poll_table *pt)
 242 {
 243         struct kvm_kernel_irqfd *irqfd =
 244                 container_of(pt, struct kvm_kernel_irqfd, pt);
 245         add_wait_queue(wqh, &irqfd->wait);
 246 }
 247
 248 /* Must be called under irqfds.lock */
 249 static void irqfd_update(struct kvm *kvm, struct kvm_kernel_irqfd *irqfd)
 250 {
 251         struct kvm_kernel_irq_routing_entry *e;
 252         struct kvm_kernel_irq_routing_entry entries[KVM_NR_IRQCHIPS];
 253         int n_entries;
 254
 255         n_entries = kvm_irq_map_gsi(kvm, entries, irqfd->gsi);
 256
 257         write_seqcount_begin(&irqfd->irq_entry_sc);
 258
 259         e = entries;
 260         if (n_entries == 1)
 261                 irqfd->irq_entry = *e;
 262         else
 263                 irqfd->irq_entry.type = 0;
 264
 265         write_seqcount_end(&irqfd->irq_entry_sc);
 266 }
 267
 268 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 269 void __attribute__((weak)) kvm_arch_irq_bypass_stop(
 270                                 struct irq_bypass_consumer *cons)
 271 {
 272 }
 273
 274 void __attribute__((weak)) kvm_arch_irq_bypass_start(
 275                                 struct irq_bypass_consumer *cons)
 276 {
 277 }
 278
 279 int  __attribute__((weak)) kvm_arch_update_irqfd_routing(
 280                                 struct kvm *kvm, unsigned int host_irq,
 281                                 uint32_t guest_irq, bool set)
 282 {
 283         return 0;
 284 }
 285 #endif
 286
 287 static int
 288 kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
 289 {
 290         struct kvm_kernel_irqfd *irqfd, *tmp;
 291         struct fd f;
 292         struct eventfd_ctx *eventfd = NULL, *resamplefd = NULL;
 293         int ret;
 294         __poll_t events;
 295         int idx;
 296
 297         if (!kvm_arch_intc_initialized(kvm))
 298                 return -EAGAIN;
 299
 300         irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL);
 301         if (!irqfd)
 302                 return -ENOMEM;
 303
 304         irqfd->kvm = kvm;
 305         irqfd->gsi = args->gsi;
 306         INIT_LIST_HEAD(&irqfd->list);
 307         INIT_WORK(&irqfd->inject, irqfd_inject);
 308         INIT_WORK(&irqfd->shutdown, irqfd_shutdown);
 309         seqcount_init(&irqfd->irq_entry_sc);
 310
 311         f = fdget(args->fd);
 312         if (!f.file) {
 313                 ret = -EBADF;
 314                 goto out;
 315         }
 316
 317         eventfd = eventfd_ctx_fileget(f.file);
 318         if (IS_ERR(eventfd)) {
 319                 ret = PTR_ERR(eventfd);
 320                 goto fail;
 321         }
 322
 323         irqfd->eventfd = eventfd;
 324
 325         if (args->flags & KVM_IRQFD_FLAG_RESAMPLE) {
 326                 struct kvm_kernel_irqfd_resampler *resampler;
 327
 328                 resamplefd = eventfd_ctx_fdget(args->resamplefd);
 329                 if (IS_ERR(resamplefd)) {
 330                         ret = PTR_ERR(resamplefd);
 331                         goto fail;
 332                 }
 333
 334                 irqfd->resamplefd = resamplefd;
 335                 INIT_LIST_HEAD(&irqfd->resampler_link);
 336
 337                 mutex_lock(&kvm->irqfds.resampler_lock);
 338
 339                 list_for_each_entry(resampler,
 340                                     &kvm->irqfds.resampler_list, link) {
 341                         if (resampler->notifier.gsi == irqfd->gsi) {
 342                                 irqfd->resampler = resampler;
 343                                 break;
 344                         }
 345                 }
 346
 347                 if (!irqfd->resampler) {
 348                         resampler = kzalloc(sizeof(*resampler), GFP_KERNEL);
 349                         if (!resampler) {
 350                                 ret = -ENOMEM;
 351                                 mutex_unlock(&kvm->irqfds.resampler_lock);
 352                                 goto fail;
 353                         }
 354
 355                         resampler->kvm = kvm;
 356                         INIT_LIST_HEAD(&resampler->list);
 357                         resampler->notifier.gsi = irqfd->gsi;
 358                         resampler->notifier.irq_acked = irqfd_resampler_ack;
 359                         INIT_LIST_HEAD(&resampler->link);
 360
 361                         list_add(&resampler->link, &kvm->irqfds.resampler_list);
 362                         kvm_register_irq_ack_notifier(kvm,
 363                                                       &resampler->notifier);
 364                         irqfd->resampler = resampler;
 365                 }
 366
 367                 list_add_rcu(&irqfd->resampler_link, &irqfd->resampler->list);
 368                 synchronize_srcu(&kvm->irq_srcu);
 369
 370                 mutex_unlock(&kvm->irqfds.resampler_lock);
 371         }
 372
 373         /*
 374          * Install our own custom wake-up handling so we are notified via
 375          * a callback whenever someone signals the underlying eventfd
 376          */
 377         init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
 378         init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
 379
 380         spin_lock_irq(&kvm->irqfds.lock);
 381
 382         ret = 0;
 383         list_for_each_entry(tmp, &kvm->irqfds.items, list) {
 384                 if (irqfd->eventfd != tmp->eventfd)
 385                         continue;
 386                 /* This fd is used for another irq already. */
 387                 ret = -EBUSY;
 388                 spin_unlock_irq(&kvm->irqfds.lock);
 389                 goto fail;
 390         }
 391
 392         idx = srcu_read_lock(&kvm->irq_srcu);
 393         irqfd_update(kvm, irqfd);
 394
 395         list_add_tail(&irqfd->list, &kvm->irqfds.items);
 396
 397         spin_unlock_irq(&kvm->irqfds.lock);
 398
 399         /*
 400          * Check if there was an event already pending on the eventfd
 401          * before we registered, and trigger it as if we didn't miss it.
 402          */
 403         events = vfs_poll(f.file, &irqfd->pt);
 404
 405         if (events & EPOLLIN)
 406                 schedule_work(&irqfd->inject);
 407
 408 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 409         if (kvm_arch_has_irq_bypass()) {
 410                 irqfd->consumer.token = (void *)irqfd->eventfd;
 411                 irqfd->consumer.add_producer = kvm_arch_irq_bypass_add_producer;
 412                 irqfd->consumer.del_producer = kvm_arch_irq_bypass_del_producer;
 413                 irqfd->consumer.stop = kvm_arch_irq_bypass_stop;
 414                 irqfd->consumer.start = kvm_arch_irq_bypass_start;
 415                 ret = irq_bypass_register_consumer(&irqfd->consumer);
 416                 if (ret)
 417                         pr_info("irq bypass consumer (token %p) registration fails: %d\n",
 418                                 irqfd->consumer.token, ret);
 419         }
 420 #endif
 421
 422         srcu_read_unlock(&kvm->irq_srcu, idx);
 423
 424         /*
 425          * do not drop the file until the irqfd is fully initialized, otherwise
 426          * we might race against the EPOLLHUP
 427          */
 428         fdput(f);
 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
 448 bool kvm_irq_has_notifier(struct kvm *kvm, unsigned irqchip, unsigned pin)
 449 {
 450         struct kvm_irq_ack_notifier *kian;
 451         int gsi, idx;
 452
 453         idx = srcu_read_lock(&kvm->irq_srcu);
 454         gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
 455         if (gsi != -1)
 456                 hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
 457                                          link)
 458                         if (kian->gsi == gsi) {
 459                                 srcu_read_unlock(&kvm->irq_srcu, idx);
 460                                 return true;
 461                         }
 462
 463         srcu_read_unlock(&kvm->irq_srcu, idx);
 464
 465         return false;
 466 }
 467 EXPORT_SYMBOL_GPL(kvm_irq_has_notifier);
 468
 469 void kvm_notify_acked_gsi(struct kvm *kvm, int gsi)
 470 {
 471         struct kvm_irq_ack_notifier *kian;
 472
 473         hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
 474                                  link)
 475                 if (kian->gsi == gsi)
 476                         kian->irq_acked(kian);
 477 }
 478
 479 void kvm_notify_acked_irq(struct kvm *kvm, unsigned irqchip, unsigned pin)
 480 {
 481         int gsi, idx;
 482
 483         trace_kvm_ack_irq(irqchip, pin);
 484
 485         idx = srcu_read_lock(&kvm->irq_srcu);
 486         gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
 487         if (gsi != -1)
 488                 kvm_notify_acked_gsi(kvm, gsi);
 489         srcu_read_unlock(&kvm->irq_srcu, idx);
 490 }
 491
 492 void kvm_register_irq_ack_notifier(struct kvm *kvm,
 493                                    struct kvm_irq_ack_notifier *kian)
 494 {
 495         mutex_lock(&kvm->irq_lock);
 496         hlist_add_head_rcu(&kian->link, &kvm->irq_ack_notifier_list);
 497         mutex_unlock(&kvm->irq_lock);
 498         kvm_arch_post_irq_ack_notifier_list_update(kvm);
 499 }
 500
 501 void kvm_unregister_irq_ack_notifier(struct kvm *kvm,
 502                                     struct kvm_irq_ack_notifier *kian)
 503 {
 504         mutex_lock(&kvm->irq_lock);
 505         hlist_del_init_rcu(&kian->link);
 506         mutex_unlock(&kvm->irq_lock);
 507         synchronize_srcu(&kvm->irq_srcu);
 508         kvm_arch_post_irq_ack_notifier_list_update(kvm);
 509 }
 510 #endif
 511
 512 void
 513 kvm_eventfd_init(struct kvm *kvm)
 514 {
 515 #ifdef CONFIG_HAVE_KVM_IRQFD
 516         spin_lock_init(&kvm->irqfds.lock);
 517         INIT_LIST_HEAD(&kvm->irqfds.items);
 518         INIT_LIST_HEAD(&kvm->irqfds.resampler_list);
 519         mutex_init(&kvm->irqfds.resampler_lock);
 520 #endif
 521         INIT_LIST_HEAD(&kvm->ioeventfds);
 522 }
 523
 524 #ifdef CONFIG_HAVE_KVM_IRQFD
 525 /*
 526  * shutdown any irqfd's that match fd+gsi
 527  */
 528 static int
 529 kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
 530 {
 531         struct kvm_kernel_irqfd *irqfd, *tmp;
 532         struct eventfd_ctx *eventfd;
 533
 534         eventfd = eventfd_ctx_fdget(args->fd);
 535         if (IS_ERR(eventfd))
 536                 return PTR_ERR(eventfd);
 537
 538         spin_lock_irq(&kvm->irqfds.lock);
 539
 540         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
 541                 if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
 542                         /*
 543                          * This clearing of irq_entry.type is needed for when
 544                          * another thread calls kvm_irq_routing_update before
 545                          * we flush workqueue below (we synchronize with
 546                          * kvm_irq_routing_update using irqfds.lock).
 547                          */
 548                         write_seqcount_begin(&irqfd->irq_entry_sc);
 549                         irqfd->irq_entry.type = 0;
 550                         write_seqcount_end(&irqfd->irq_entry_sc);
 551                         irqfd_deactivate(irqfd);
 552                 }
 553         }
 554
 555         spin_unlock_irq(&kvm->irqfds.lock);
 556         eventfd_ctx_put(eventfd);
 557
 558         /*
 559          * Block until we know all outstanding shutdown jobs have completed
 560          * so that we guarantee there will not be any more interrupts on this
 561          * gsi once this deassign function returns.
 562          */
 563         flush_workqueue(irqfd_cleanup_wq);
 564
 565         return 0;
 566 }
 567
 568 int
 569 kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
 570 {
 571         if (args->flags & ~(KVM_IRQFD_FLAG_DEASSIGN | KVM_IRQFD_FLAG_RESAMPLE))
 572                 return -EINVAL;
 573
 574         if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
 575                 return kvm_irqfd_deassign(kvm, args);
 576
 577         return kvm_irqfd_assign(kvm, args);
 578 }
 579
 580 /*
 581  * This function is called as the kvm VM fd is being released. Shutdown all
 582  * irqfds that still remain open
 583  */
 584 void
 585 kvm_irqfd_release(struct kvm *kvm)
 586 {
 587         struct kvm_kernel_irqfd *irqfd, *tmp;
 588
 589         spin_lock_irq(&kvm->irqfds.lock);
 590
 591         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
 592                 irqfd_deactivate(irqfd);
 593
 594         spin_unlock_irq(&kvm->irqfds.lock);
 595
 596         /*
 597          * Block until we know all outstanding shutdown jobs have completed
 598          * since we do not take a kvm* reference.
 599          */
 600         flush_workqueue(irqfd_cleanup_wq);
 601
 602 }
 603
 604 /*
 605  * Take note of a change in irq routing.
 606  * Caller must invoke synchronize_srcu(&kvm->irq_srcu) afterwards.
 607  */
 608 void kvm_irq_routing_update(struct kvm *kvm)
 609 {
 610         struct kvm_kernel_irqfd *irqfd;
 611
 612         spin_lock_irq(&kvm->irqfds.lock);
 613
 614         list_for_each_entry(irqfd, &kvm->irqfds.items, list) {
 615                 irqfd_update(kvm, irqfd);
 616
 617 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 618                 if (irqfd->producer) {
 619                         int ret = kvm_arch_update_irqfd_routing(
 620                                         irqfd->kvm, irqfd->producer->irq,
 621                                         irqfd->gsi, 1);
 622                         WARN_ON(ret);
 623                 }
 624 #endif
 625         }
 626
 627         spin_unlock_irq(&kvm->irqfds.lock);
 628 }
 629
 630 /*
 631  * create a host-wide workqueue for issuing deferred shutdown requests
 632  * aggregated from all vm* instances. We need our own isolated
 633  * queue to ease flushing work items when a VM exits.
 634  */
 635 int kvm_irqfd_init(void)
 636 {
 637         irqfd_cleanup_wq = alloc_workqueue("kvm-irqfd-cleanup", 0, 0);
 638         if (!irqfd_cleanup_wq)
 639                 return -ENOMEM;
 640
 641         return 0;
 642 }
 643
 644 void kvm_irqfd_exit(void)
 645 {
 646         destroy_workqueue(irqfd_cleanup_wq);
 647 }
 648 #endif
 649
 650 /*
 651  * --------------------------------------------------------------------
 652  * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
 653  *
 654  * userspace can register a PIO/MMIO address with an eventfd for receiving
 655  * notification when the memory has been touched.
 656  * --------------------------------------------------------------------
 657  */
 658
 659 struct _ioeventfd {
 660         struct list_head     list;
 661         u64                  addr;
 662         int                  length;
 663         struct eventfd_ctx  *eventfd;
 664         u64                  datamatch;
 665         struct kvm_io_device dev;
 666         u8                   bus_idx;
 667         bool                 wildcard;
 668 };
 669
 670 static inline struct _ioeventfd *
 671 to_ioeventfd(struct kvm_io_device *dev)
 672 {
 673         return container_of(dev, struct _ioeventfd, dev);
 674 }
 675
 676 static void
 677 ioeventfd_release(struct _ioeventfd *p)
 678 {
 679         eventfd_ctx_put(p->eventfd);
 680         list_del(&p->list);
 681         kfree(p);
 682 }
 683
 684 static bool
 685 ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
 686 {
 687         u64 _val;
 688
 689         if (addr != p->addr)
 690                 /* address must be precise for a hit */
 691                 return false;
 692
 693         if (!p->length)
 694                 /* length = 0 means only look at the address, so always a hit */
 695                 return true;
 696
 697         if (len != p->length)
 698                 /* address-range must be precise for a hit */
 699                 return false;
 700
 701         if (p->wildcard)
 702                 /* all else equal, wildcard is always a hit */
 703                 return true;
 704
 705         /* otherwise, we have to actually compare the data */
 706
 707         BUG_ON(!IS_ALIGNED((unsigned long)val, len));
 708
 709         switch (len) {
 710         case 1:
 711                 _val = *(u8 *)val;
 712                 break;
 713         case 2:
 714                 _val = *(u16 *)val;
 715                 break;
 716         case 4:
 717                 _val = *(u32 *)val;
 718                 break;
 719         case 8:
 720                 _val = *(u64 *)val;
 721                 break;
 722         default:
 723                 return false;
 724         }
 725
 726         return _val == p->datamatch ? true : false;
 727 }
 728
 729 /* MMIO/PIO writes trigger an event if the addr/val match */
 730 static int
 731 ioeventfd_write(struct kvm_vcpu *vcpu, struct kvm_io_device *this, gpa_t addr,
 732                 int len, const void *val)
 733 {
 734         struct _ioeventfd *p = to_ioeventfd(this);
 735
 736         if (!ioeventfd_in_range(p, addr, len, val))
 737                 return -EOPNOTSUPP;
 738
 739         eventfd_signal(p->eventfd, 1);
 740         return 0;
 741 }
 742
 743 /*
 744  * This function is called as KVM is completely shutting down.  We do not
 745  * need to worry about locking just nuke anything we have as quickly as possible
 746  */
 747 static void
 748 ioeventfd_destructor(struct kvm_io_device *this)
 749 {
 750         struct _ioeventfd *p = to_ioeventfd(this);
 751
 752         ioeventfd_release(p);
 753 }
 754
 755 static const struct kvm_io_device_ops ioeventfd_ops = {
 756         .write      = ioeventfd_write,
 757         .destructor = ioeventfd_destructor,
 758 };
 759
 760 /* assumes kvm->slots_lock held */
 761 static bool
 762 ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
 763 {
 764         struct _ioeventfd *_p;
 765
 766         list_for_each_entry(_p, &kvm->ioeventfds, list)
 767                 if (_p->bus_idx == p->bus_idx &&
 768                     _p->addr == p->addr &&
 769                     (!_p->length || !p->length ||
 770                      (_p->length == p->length &&
 771                       (_p->wildcard || p->wildcard ||
 772                        _p->datamatch == p->datamatch))))
 773                         return true;
 774
 775         return false;
 776 }
 777
 778 static enum kvm_bus ioeventfd_bus_from_flags(__u32 flags)
 779 {
 780         if (flags & KVM_IOEVENTFD_FLAG_PIO)
 781                 return KVM_PIO_BUS;
 782         if (flags & KVM_IOEVENTFD_FLAG_VIRTIO_CCW_NOTIFY)
 783                 return KVM_VIRTIO_CCW_NOTIFY_BUS;
 784         return KVM_MMIO_BUS;
 785 }
 786
 787 static int kvm_assign_ioeventfd_idx(struct kvm *kvm,
 788                                 enum kvm_bus bus_idx,
 789                                 struct kvm_ioeventfd *args)
 790 {
 791
 792         struct eventfd_ctx *eventfd;
 793         struct _ioeventfd *p;
 794         int ret;
 795
 796         eventfd = eventfd_ctx_fdget(args->fd);
 797         if (IS_ERR(eventfd))
 798                 return PTR_ERR(eventfd);
 799
 800         p = kzalloc(sizeof(*p), GFP_KERNEL);
 801         if (!p) {
 802                 ret = -ENOMEM;
 803                 goto fail;
 804         }
 805
 806         INIT_LIST_HEAD(&p->list);
 807         p->addr    = args->addr;
 808         p->bus_idx = bus_idx;
 809         p->length  = args->len;
 810         p->eventfd = eventfd;
 811
 812         /* The datamatch feature is optional, otherwise this is a wildcard */
 813         if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
 814                 p->datamatch = args->datamatch;
 815         else
 816                 p->wildcard = true;
 817
 818         mutex_lock(&kvm->slots_lock);
 819
 820         /* Verify that there isn't a match already */
 821         if (ioeventfd_check_collision(kvm, p)) {
 822                 ret = -EEXIST;
 823                 goto unlock_fail;
 824         }
 825
 826         kvm_iodevice_init(&p->dev, &ioeventfd_ops);
 827
 828         ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
 829                                       &p->dev);
 830         if (ret < 0)
 831                 goto unlock_fail;
 832
 833         kvm_get_bus(kvm, bus_idx)->ioeventfd_count++;
 834         list_add_tail(&p->list, &kvm->ioeventfds);
 835
 836         mutex_unlock(&kvm->slots_lock);
 837
 838         return 0;
 839
 840 unlock_fail:
 841         mutex_unlock(&kvm->slots_lock);
 842
 843 fail:
 844         kfree(p);
 845         eventfd_ctx_put(eventfd);
 846
 847         return ret;
 848 }
 849
 850 static int
 851 kvm_deassign_ioeventfd_idx(struct kvm *kvm, enum kvm_bus bus_idx,
 852                            struct kvm_ioeventfd *args)
 853 {
 854         struct _ioeventfd        *p, *tmp;
 855         struct eventfd_ctx       *eventfd;
 856         struct kvm_io_bus        *bus;
 857         int                       ret = -ENOENT;
 858
 859         eventfd = eventfd_ctx_fdget(args->fd);
 860         if (IS_ERR(eventfd))
 861                 return PTR_ERR(eventfd);
 862
 863         mutex_lock(&kvm->slots_lock);
 864
 865         list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
 866                 bool wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
 867
 868                 if (p->bus_idx != bus_idx ||
 869                     p->eventfd != eventfd  ||
 870                     p->addr != args->addr  ||
 871                     p->length != args->len ||
 872                     p->wildcard != wildcard)
 873                         continue;
 874
 875                 if (!p->wildcard && p->datamatch != args->datamatch)
 876                         continue;
 877
 878                 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
 879                 bus = kvm_get_bus(kvm, bus_idx);
 880                 if (bus)
 881                         bus->ioeventfd_count--;
 882                 ioeventfd_release(p);
 883                 ret = 0;
 884                 break;
 885         }
 886
 887         mutex_unlock(&kvm->slots_lock);
 888
 889         eventfd_ctx_put(eventfd);
 890
 891         return ret;
 892 }
 893
 894 static int kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 895 {
 896         enum kvm_bus bus_idx = ioeventfd_bus_from_flags(args->flags);
 897         int ret = kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
 898
 899         if (!args->len && bus_idx == KVM_MMIO_BUS)
 900                 kvm_deassign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
 901
 902         return ret;
 903 }
 904
 905 static int
 906 kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 907 {
 908         enum kvm_bus              bus_idx;
 909         int ret;
 910
 911         bus_idx = ioeventfd_bus_from_flags(args->flags);
 912         /* must be natural-word sized, or 0 to ignore length */
 913         switch (args->len) {
 914         case 0:
 915         case 1:
 916         case 2:
 917         case 4:
 918         case 8:
 919                 break;
 920         default:
 921                 return -EINVAL;
 922         }
 923
 924         /* check for range overflow */
 925         if (args->addr + args->len < args->addr)
 926                 return -EINVAL;
 927
 928         /* check for extra flags that we don't understand */
 929         if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
 930                 return -EINVAL;
 931
 932         /* ioeventfd with no length can't be combined with DATAMATCH */
 933         if (!args->len && (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH))
 934                 return -EINVAL;
 935
 936         ret = kvm_assign_ioeventfd_idx(kvm, bus_idx, args);
 937         if (ret)
 938                 goto fail;
 939
 940         /* When length is ignored, MMIO is also put on a separate bus, for
 941          * faster lookups.
 942          */
 943         if (!args->len && bus_idx == KVM_MMIO_BUS) {
 944                 ret = kvm_assign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
 945                 if (ret < 0)
 946                         goto fast_fail;
 947         }
 948
 949         return 0;
 950
 951 fast_fail:
 952         kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
 953 fail:
 954         return ret;
 955 }
 956
 957 int
 958 kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 959 {
 960         if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
 961                 return kvm_deassign_ioeventfd(kvm, args);
 962
 963         return kvm_assign_ioeventfd(kvm, args);
 964 }