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_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         unsigned long flags = (unsigned long)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 & POLLIN) {
 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 & POLLHUP) {
 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         unsigned int 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 = f.file->f_op->poll(f.file, &irqfd->pt);
 404
 405         if (events & POLLIN)
 406                 schedule_work(&irqfd->inject);
 407
 408 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 409         irqfd->consumer.token = (void *)irqfd->eventfd;
 410         irqfd->consumer.add_producer = kvm_arch_irq_bypass_add_producer;
 411         irqfd->consumer.del_producer = kvm_arch_irq_bypass_del_producer;
 412         irqfd->consumer.stop = kvm_arch_irq_bypass_stop;
 413         irqfd->consumer.start = kvm_arch_irq_bypass_start;
 414         ret = irq_bypass_register_consumer(&irqfd->consumer);
 415         if (ret)
 416                 pr_info("irq bypass consumer (token %p) registration fails: %d\n",
 417                                 irqfd->consumer.token, ret);
 418 #endif
 419
 420         srcu_read_unlock(&kvm->irq_srcu, idx);
 421
 422         /*
 423          * do not drop the file until the irqfd is fully initialized, otherwise
 424          * we might race against the POLLHUP
 425          */
 426         fdput(f);
 427         return 0;
 428
 429 fail:
 430         if (irqfd->resampler)
 431                 irqfd_resampler_shutdown(irqfd);
 432
 433         if (resamplefd && !IS_ERR(resamplefd))
 434                 eventfd_ctx_put(resamplefd);
 435
 436         if (eventfd && !IS_ERR(eventfd))
 437                 eventfd_ctx_put(eventfd);
 438
 439         fdput(f);
 440
 441 out:
 442         kfree(irqfd);
 443         return ret;
 444 }
 445
 446 bool kvm_irq_has_notifier(struct kvm *kvm, unsigned irqchip, unsigned pin)
 447 {
 448         struct kvm_irq_ack_notifier *kian;
 449         int gsi, idx;
 450
 451         idx = srcu_read_lock(&kvm->irq_srcu);
 452         gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
 453         if (gsi != -1)
 454                 hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
 455                                          link)
 456                         if (kian->gsi == gsi) {
 457                                 srcu_read_unlock(&kvm->irq_srcu, idx);
 458                                 return true;
 459                         }
 460
 461         srcu_read_unlock(&kvm->irq_srcu, idx);
 462
 463         return false;
 464 }
 465 EXPORT_SYMBOL_GPL(kvm_irq_has_notifier);
 466
 467 void kvm_notify_acked_gsi(struct kvm *kvm, int gsi)
 468 {
 469         struct kvm_irq_ack_notifier *kian;
 470
 471         hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
 472                                  link)
 473                 if (kian->gsi == gsi)
 474                         kian->irq_acked(kian);
 475 }
 476
 477 void kvm_notify_acked_irq(struct kvm *kvm, unsigned irqchip, unsigned pin)
 478 {
 479         int gsi, idx;
 480
 481         trace_kvm_ack_irq(irqchip, pin);
 482
 483         idx = srcu_read_lock(&kvm->irq_srcu);
 484         gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
 485         if (gsi != -1)
 486                 kvm_notify_acked_gsi(kvm, gsi);
 487         srcu_read_unlock(&kvm->irq_srcu, idx);
 488 }
 489
 490 void kvm_register_irq_ack_notifier(struct kvm *kvm,
 491                                    struct kvm_irq_ack_notifier *kian)
 492 {
 493         mutex_lock(&kvm->irq_lock);
 494         hlist_add_head_rcu(&kian->link, &kvm->irq_ack_notifier_list);
 495         mutex_unlock(&kvm->irq_lock);
 496         kvm_vcpu_request_scan_ioapic(kvm);
 497 }
 498
 499 void kvm_unregister_irq_ack_notifier(struct kvm *kvm,
 500                                     struct kvm_irq_ack_notifier *kian)
 501 {
 502         mutex_lock(&kvm->irq_lock);
 503         hlist_del_init_rcu(&kian->link);
 504         mutex_unlock(&kvm->irq_lock);
 505         synchronize_srcu(&kvm->irq_srcu);
 506         kvm_vcpu_request_scan_ioapic(kvm);
 507 }
 508 #endif
 509
 510 void
 511 kvm_eventfd_init(struct kvm *kvm)
 512 {
 513 #ifdef CONFIG_HAVE_KVM_IRQFD
 514         spin_lock_init(&kvm->irqfds.lock);
 515         INIT_LIST_HEAD(&kvm->irqfds.items);
 516         INIT_LIST_HEAD(&kvm->irqfds.resampler_list);
 517         mutex_init(&kvm->irqfds.resampler_lock);
 518 #endif
 519         INIT_LIST_HEAD(&kvm->ioeventfds);
 520 }
 521
 522 #ifdef CONFIG_HAVE_KVM_IRQFD
 523 /*
 524  * shutdown any irqfd's that match fd+gsi
 525  */
 526 static int
 527 kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
 528 {
 529         struct kvm_kernel_irqfd *irqfd, *tmp;
 530         struct eventfd_ctx *eventfd;
 531
 532         eventfd = eventfd_ctx_fdget(args->fd);
 533         if (IS_ERR(eventfd))
 534                 return PTR_ERR(eventfd);
 535
 536         spin_lock_irq(&kvm->irqfds.lock);
 537
 538         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
 539                 if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
 540                         /*
 541                          * This clearing of irq_entry.type is needed for when
 542                          * another thread calls kvm_irq_routing_update before
 543                          * we flush workqueue below (we synchronize with
 544                          * kvm_irq_routing_update using irqfds.lock).
 545                          */
 546                         write_seqcount_begin(&irqfd->irq_entry_sc);
 547                         irqfd->irq_entry.type = 0;
 548                         write_seqcount_end(&irqfd->irq_entry_sc);
 549                         irqfd_deactivate(irqfd);
 550                 }
 551         }
 552
 553         spin_unlock_irq(&kvm->irqfds.lock);
 554         eventfd_ctx_put(eventfd);
 555
 556         /*
 557          * Block until we know all outstanding shutdown jobs have completed
 558          * so that we guarantee there will not be any more interrupts on this
 559          * gsi once this deassign function returns.
 560          */
 561         flush_workqueue(irqfd_cleanup_wq);
 562
 563         return 0;
 564 }
 565
 566 int
 567 kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
 568 {
 569         if (args->flags & ~(KVM_IRQFD_FLAG_DEASSIGN | KVM_IRQFD_FLAG_RESAMPLE))
 570                 return -EINVAL;
 571
 572         if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
 573                 return kvm_irqfd_deassign(kvm, args);
 574
 575         return kvm_irqfd_assign(kvm, args);
 576 }
 577
 578 /*
 579  * This function is called as the kvm VM fd is being released. Shutdown all
 580  * irqfds that still remain open
 581  */
 582 void
 583 kvm_irqfd_release(struct kvm *kvm)
 584 {
 585         struct kvm_kernel_irqfd *irqfd, *tmp;
 586
 587         spin_lock_irq(&kvm->irqfds.lock);
 588
 589         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
 590                 irqfd_deactivate(irqfd);
 591
 592         spin_unlock_irq(&kvm->irqfds.lock);
 593
 594         /*
 595          * Block until we know all outstanding shutdown jobs have completed
 596          * since we do not take a kvm* reference.
 597          */
 598         flush_workqueue(irqfd_cleanup_wq);
 599
 600 }
 601
 602 /*
 603  * Take note of a change in irq routing.
 604  * Caller must invoke synchronize_srcu(&kvm->irq_srcu) afterwards.
 605  */
 606 void kvm_irq_routing_update(struct kvm *kvm)
 607 {
 608         struct kvm_kernel_irqfd *irqfd;
 609
 610         spin_lock_irq(&kvm->irqfds.lock);
 611
 612         list_for_each_entry(irqfd, &kvm->irqfds.items, list) {
 613                 irqfd_update(kvm, irqfd);
 614
 615 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 616                 if (irqfd->producer) {
 617                         int ret = kvm_arch_update_irqfd_routing(
 618                                         irqfd->kvm, irqfd->producer->irq,
 619                                         irqfd->gsi, 1);
 620                         WARN_ON(ret);
 621                 }
 622 #endif
 623         }
 624
 625         spin_unlock_irq(&kvm->irqfds.lock);
 626 }
 627
 628 /*
 629  * create a host-wide workqueue for issuing deferred shutdown requests
 630  * aggregated from all vm* instances. We need our own isolated single-thread
 631  * queue to prevent deadlock against flushing the normal work-queue.
 632  */
 633 int kvm_irqfd_init(void)
 634 {
 635         irqfd_cleanup_wq = create_singlethread_workqueue("kvm-irqfd-cleanup");
 636         if (!irqfd_cleanup_wq)
 637                 return -ENOMEM;
 638
 639         return 0;
 640 }
 641
 642 void kvm_irqfd_exit(void)
 643 {
 644         destroy_workqueue(irqfd_cleanup_wq);
 645 }
 646 #endif
 647
 648 /*
 649  * --------------------------------------------------------------------
 650  * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
 651  *
 652  * userspace can register a PIO/MMIO address with an eventfd for receiving
 653  * notification when the memory has been touched.
 654  * --------------------------------------------------------------------
 655  */
 656
 657 struct _ioeventfd {
 658         struct list_head     list;
 659         u64                  addr;
 660         int                  length;
 661         struct eventfd_ctx  *eventfd;
 662         u64                  datamatch;
 663         struct kvm_io_device dev;
 664         u8                   bus_idx;
 665         bool                 wildcard;
 666 };
 667
 668 static inline struct _ioeventfd *
 669 to_ioeventfd(struct kvm_io_device *dev)
 670 {
 671         return container_of(dev, struct _ioeventfd, dev);
 672 }
 673
 674 static void
 675 ioeventfd_release(struct _ioeventfd *p)
 676 {
 677         eventfd_ctx_put(p->eventfd);
 678         list_del(&p->list);
 679         kfree(p);
 680 }
 681
 682 static bool
 683 ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
 684 {
 685         u64 _val;
 686
 687         if (addr != p->addr)
 688                 /* address must be precise for a hit */
 689                 return false;
 690
 691         if (!p->length)
 692                 /* length = 0 means only look at the address, so always a hit */
 693                 return true;
 694
 695         if (len != p->length)
 696                 /* address-range must be precise for a hit */
 697                 return false;
 698
 699         if (p->wildcard)
 700                 /* all else equal, wildcard is always a hit */
 701                 return true;
 702
 703         /* otherwise, we have to actually compare the data */
 704
 705         BUG_ON(!IS_ALIGNED((unsigned long)val, len));
 706
 707         switch (len) {
 708         case 1:
 709                 _val = *(u8 *)val;
 710                 break;
 711         case 2:
 712                 _val = *(u16 *)val;
 713                 break;
 714         case 4:
 715                 _val = *(u32 *)val;
 716                 break;
 717         case 8:
 718                 _val = *(u64 *)val;
 719                 break;
 720         default:
 721                 return false;
 722         }
 723
 724         return _val == p->datamatch ? true : false;
 725 }
 726
 727 /* MMIO/PIO writes trigger an event if the addr/val match */
 728 static int
 729 ioeventfd_write(struct kvm_vcpu *vcpu, struct kvm_io_device *this, gpa_t addr,
 730                 int len, const void *val)
 731 {
 732         struct _ioeventfd *p = to_ioeventfd(this);
 733
 734         if (!ioeventfd_in_range(p, addr, len, val))
 735                 return -EOPNOTSUPP;
 736
 737         eventfd_signal(p->eventfd, 1);
 738         return 0;
 739 }
 740
 741 /*
 742  * This function is called as KVM is completely shutting down.  We do not
 743  * need to worry about locking just nuke anything we have as quickly as possible
 744  */
 745 static void
 746 ioeventfd_destructor(struct kvm_io_device *this)
 747 {
 748         struct _ioeventfd *p = to_ioeventfd(this);
 749
 750         ioeventfd_release(p);
 751 }
 752
 753 static const struct kvm_io_device_ops ioeventfd_ops = {
 754         .write      = ioeventfd_write,
 755         .destructor = ioeventfd_destructor,
 756 };
 757
 758 /* assumes kvm->slots_lock held */
 759 static bool
 760 ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
 761 {
 762         struct _ioeventfd *_p;
 763
 764         list_for_each_entry(_p, &kvm->ioeventfds, list)
 765                 if (_p->bus_idx == p->bus_idx &&
 766                     _p->addr == p->addr &&
 767                     (!_p->length || !p->length ||
 768                      (_p->length == p->length &&
 769                       (_p->wildcard || p->wildcard ||
 770                        _p->datamatch == p->datamatch))))
 771                         return true;
 772
 773         return false;
 774 }
 775
 776 static enum kvm_bus ioeventfd_bus_from_flags(__u32 flags)
 777 {
 778         if (flags & KVM_IOEVENTFD_FLAG_PIO)
 779                 return KVM_PIO_BUS;
 780         if (flags & KVM_IOEVENTFD_FLAG_VIRTIO_CCW_NOTIFY)
 781                 return KVM_VIRTIO_CCW_NOTIFY_BUS;
 782         return KVM_MMIO_BUS;
 783 }
 784
 785 static int kvm_assign_ioeventfd_idx(struct kvm *kvm,
 786                                 enum kvm_bus bus_idx,
 787                                 struct kvm_ioeventfd *args)
 788 {
 789
 790         struct eventfd_ctx *eventfd;
 791         struct _ioeventfd *p;
 792         int ret;
 793
 794         eventfd = eventfd_ctx_fdget(args->fd);
 795         if (IS_ERR(eventfd))
 796                 return PTR_ERR(eventfd);
 797
 798         p = kzalloc(sizeof(*p), GFP_KERNEL);
 799         if (!p) {
 800                 ret = -ENOMEM;
 801                 goto fail;
 802         }
 803
 804         INIT_LIST_HEAD(&p->list);
 805         p->addr    = args->addr;
 806         p->bus_idx = bus_idx;
 807         p->length  = args->len;
 808         p->eventfd = eventfd;
 809
 810         /* The datamatch feature is optional, otherwise this is a wildcard */
 811         if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
 812                 p->datamatch = args->datamatch;
 813         else
 814                 p->wildcard = true;
 815
 816         mutex_lock(&kvm->slots_lock);
 817
 818         /* Verify that there isn't a match already */
 819         if (ioeventfd_check_collision(kvm, p)) {
 820                 ret = -EEXIST;
 821                 goto unlock_fail;
 822         }
 823
 824         kvm_iodevice_init(&p->dev, &ioeventfd_ops);
 825
 826         ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
 827                                       &p->dev);
 828         if (ret < 0)
 829                 goto unlock_fail;
 830
 831         kvm->buses[bus_idx]->ioeventfd_count++;
 832         list_add_tail(&p->list, &kvm->ioeventfds);
 833
 834         mutex_unlock(&kvm->slots_lock);
 835
 836         return 0;
 837
 838 unlock_fail:
 839         mutex_unlock(&kvm->slots_lock);
 840
 841 fail:
 842         kfree(p);
 843         eventfd_ctx_put(eventfd);
 844
 845         return ret;
 846 }
 847
 848 static int
 849 kvm_deassign_ioeventfd_idx(struct kvm *kvm, enum kvm_bus bus_idx,
 850                            struct kvm_ioeventfd *args)
 851 {
 852         struct _ioeventfd        *p, *tmp;
 853         struct eventfd_ctx       *eventfd;
 854         int                       ret = -ENOENT;
 855
 856         eventfd = eventfd_ctx_fdget(args->fd);
 857         if (IS_ERR(eventfd))
 858                 return PTR_ERR(eventfd);
 859
 860         mutex_lock(&kvm->slots_lock);
 861
 862         list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
 863                 bool wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
 864
 865                 if (p->bus_idx != bus_idx ||
 866                     p->eventfd != eventfd  ||
 867                     p->addr != args->addr  ||
 868                     p->length != args->len ||
 869                     p->wildcard != wildcard)
 870                         continue;
 871
 872                 if (!p->wildcard && p->datamatch != args->datamatch)
 873                         continue;
 874
 875                 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
 876                 if (kvm->buses[bus_idx])
 877                         kvm->buses[bus_idx]->ioeventfd_count--;
 878                 ioeventfd_release(p);
 879                 ret = 0;
 880                 break;
 881         }
 882
 883         mutex_unlock(&kvm->slots_lock);
 884
 885         eventfd_ctx_put(eventfd);
 886
 887         return ret;
 888 }
 889
 890 static int kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 891 {
 892         enum kvm_bus bus_idx = ioeventfd_bus_from_flags(args->flags);
 893         int ret = kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
 894
 895         if (!args->len && bus_idx == KVM_MMIO_BUS)
 896                 kvm_deassign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
 897
 898         return ret;
 899 }
 900
 901 static int
 902 kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 903 {
 904         enum kvm_bus              bus_idx;
 905         int ret;
 906
 907         bus_idx = ioeventfd_bus_from_flags(args->flags);
 908         /* must be natural-word sized, or 0 to ignore length */
 909         switch (args->len) {
 910         case 0:
 911         case 1:
 912         case 2:
 913         case 4:
 914         case 8:
 915                 break;
 916         default:
 917                 return -EINVAL;
 918         }
 919
 920         /* check for range overflow */
 921         if (args->addr + args->len < args->addr)
 922                 return -EINVAL;
 923
 924         /* check for extra flags that we don't understand */
 925         if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
 926                 return -EINVAL;
 927
 928         /* ioeventfd with no length can't be combined with DATAMATCH */
 929         if (!args->len && (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH))
 930                 return -EINVAL;
 931
 932         ret = kvm_assign_ioeventfd_idx(kvm, bus_idx, args);
 933         if (ret)
 934                 goto fail;
 935
 936         /* When length is ignored, MMIO is also put on a separate bus, for
 937          * faster lookups.
 938          */
 939         if (!args->len && bus_idx == KVM_MMIO_BUS) {
 940                 ret = kvm_assign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
 941                 if (ret < 0)
 942                         goto fast_fail;
 943         }
 944
 945         return 0;
 946
 947 fast_fail:
 948         kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
 949 fail:
 950         return ret;
 951 }
 952
 953 int
 954 kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 955 {
 956         if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
 957                 return kvm_deassign_ioeventfd(kvm, args);
 958
 959         return kvm_assign_ioeventfd(kvm, args);
 960 }