virt/kvm/eventfd.c

   1 /*
   2  * kvm eventfd support - use eventfd objects to signal various KVM events
   3  *
   4  * Copyright 2009 Novell.  All Rights Reserved.
   5  * Copyright 2010 Red Hat, Inc. and/or its affiliates.
   6  *
   7  * Author:
   8  *      Gregory Haskins <ghaskins@novell.com>
   9  *
  10  * This file is free software; you can redistribute it and/or modify
  11  * it under the terms of version 2 of the GNU General Public License
  12  * as published by the Free Software Foundation.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU General Public License
  20  * along with this program; if not, write to the Free Software Foundation,
  21  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA.
  22  */
  23
  24 #include <linux/kvm_host.h>
  25 #include <linux/kvm.h>
  26 #include <linux/workqueue.h>
  27 #include <linux/syscalls.h>
  28 #include <linux/wait.h>
  29 #include <linux/poll.h>
  30 #include <linux/file.h>
  31 #include <linux/list.h>
  32 #include <linux/eventfd.h>
  33 #include <linux/kernel.h>
  34 #include <linux/srcu.h>
  35 #include <linux/slab.h>
  36 #include <linux/seqlock.h>
  37 #include <trace/events/kvm.h>
  38
  39 #include "irq.h"
  40 #include "iodev.h"
  41
  42 #ifdef CONFIG_HAVE_KVM_IRQFD
  43 /*
  44  * --------------------------------------------------------------------
  45  * irqfd: Allows an fd to be used to inject an interrupt to the guest
  46  *
  47  * Credit goes to Avi Kivity for the original idea.
  48  * --------------------------------------------------------------------
  49  */
  50
  51 /*
  52  * Resampling irqfds are a special variety of irqfds used to emulate
  53  * level triggered interrupts.  The interrupt is asserted on eventfd
  54  * trigger.  On acknowledgement through the irq ack notifier, the
  55  * interrupt is de-asserted and userspace is notified through the
  56  * resamplefd.  All resamplers on the same gsi are de-asserted
  57  * together, so we don't need to track the state of each individual
  58  * user.  We can also therefore share the same irq source ID.
  59  */
  60 struct _irqfd_resampler {
  61         struct kvm *kvm;
  62         /*
  63          * List of resampling struct _irqfd objects sharing this gsi.
  64          * RCU list modified under kvm->irqfds.resampler_lock
  65          */
  66         struct list_head list;
  67         struct kvm_irq_ack_notifier notifier;
  68         /*
  69          * Entry in list of kvm->irqfd.resampler_list.  Use for sharing
  70          * resamplers among irqfds on the same gsi.
  71          * Accessed and modified under kvm->irqfds.resampler_lock
  72          */
  73         struct list_head link;
  74 };
  75
  76 struct _irqfd {
  77         /* Used for MSI fast-path */
  78         struct kvm *kvm;
  79         wait_queue_t wait;
  80         /* Update side is protected by irqfds.lock */
  81         struct kvm_kernel_irq_routing_entry irq_entry;
  82         seqcount_t irq_entry_sc;
  83         /* Used for level IRQ fast-path */
  84         int gsi;
  85         struct work_struct inject;
  86         /* The resampler used by this irqfd (resampler-only) */
  87         struct _irqfd_resampler *resampler;
  88         /* Eventfd notified on resample (resampler-only) */
  89         struct eventfd_ctx *resamplefd;
  90         /* Entry in list of irqfds for a resampler (resampler-only) */
  91         struct list_head resampler_link;
  92         /* Used for setup/shutdown */
  93         struct eventfd_ctx *eventfd;
  94         struct list_head list;
  95         poll_table pt;
  96         struct work_struct shutdown;
  97 };
  98
  99 static struct workqueue_struct *irqfd_cleanup_wq;
 100
 101 static void
 102 irqfd_inject(struct work_struct *work)
 103 {
 104         struct _irqfd *irqfd = container_of(work, struct _irqfd, inject);
 105         struct kvm *kvm = irqfd->kvm;
 106
 107         if (!irqfd->resampler) {
 108                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1,
 109                                 false);
 110                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0,
 111                                 false);
 112         } else
 113                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 114                             irqfd->gsi, 1, false);
 115 }
 116
 117 /*
 118  * Since resampler irqfds share an IRQ source ID, we de-assert once
 119  * then notify all of the resampler irqfds using this GSI.  We can't
 120  * do multiple de-asserts or we risk racing with incoming re-asserts.
 121  */
 122 static void
 123 irqfd_resampler_ack(struct kvm_irq_ack_notifier *kian)
 124 {
 125         struct _irqfd_resampler *resampler;
 126         struct kvm *kvm;
 127         struct _irqfd *irqfd;
 128         int idx;
 129
 130         resampler = container_of(kian, struct _irqfd_resampler, notifier);
 131         kvm = resampler->kvm;
 132
 133         kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 134                     resampler->notifier.gsi, 0, false);
 135
 136         idx = srcu_read_lock(&kvm->irq_srcu);
 137
 138         list_for_each_entry_rcu(irqfd, &resampler->list, resampler_link)
 139                 eventfd_signal(irqfd->resamplefd, 1);
 140
 141         srcu_read_unlock(&kvm->irq_srcu, idx);
 142 }
 143
 144 static void
 145 irqfd_resampler_shutdown(struct _irqfd *irqfd)
 146 {
 147         struct _irqfd_resampler *resampler = irqfd->resampler;
 148         struct kvm *kvm = resampler->kvm;
 149
 150         mutex_lock(&kvm->irqfds.resampler_lock);
 151
 152         list_del_rcu(&irqfd->resampler_link);
 153         synchronize_srcu(&kvm->irq_srcu);
 154
 155         if (list_empty(&resampler->list)) {
 156                 list_del(&resampler->link);
 157                 kvm_unregister_irq_ack_notifier(kvm, &resampler->notifier);
 158                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 159                             resampler->notifier.gsi, 0, false);
 160                 kfree(resampler);
 161         }
 162
 163         mutex_unlock(&kvm->irqfds.resampler_lock);
 164 }
 165
 166 /*
 167  * Race-free decouple logic (ordering is critical)
 168  */
 169 static void
 170 irqfd_shutdown(struct work_struct *work)
 171 {
 172         struct _irqfd *irqfd = container_of(work, struct _irqfd, shutdown);
 173         u64 cnt;
 174
 175         /*
 176          * Synchronize with the wait-queue and unhook ourselves to prevent
 177          * further events.
 178          */
 179         eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
 180
 181         /*
 182          * We know no new events will be scheduled at this point, so block
 183          * until all previously outstanding events have completed
 184          */
 185         flush_work(&irqfd->inject);
 186
 187         if (irqfd->resampler) {
 188                 irqfd_resampler_shutdown(irqfd);
 189                 eventfd_ctx_put(irqfd->resamplefd);
 190         }
 191
 192         /*
 193          * It is now safe to release the object's resources
 194          */
 195         eventfd_ctx_put(irqfd->eventfd);
 196         kfree(irqfd);
 197 }
 198
 199
 200 /* assumes kvm->irqfds.lock is held */
 201 static bool
 202 irqfd_is_active(struct _irqfd *irqfd)
 203 {
 204         return list_empty(&irqfd->list) ? false : true;
 205 }
 206
 207 /*
 208  * Mark the irqfd as inactive and schedule it for removal
 209  *
 210  * assumes kvm->irqfds.lock is held
 211  */
 212 static void
 213 irqfd_deactivate(struct _irqfd *irqfd)
 214 {
 215         BUG_ON(!irqfd_is_active(irqfd));
 216
 217         list_del_init(&irqfd->list);
 218
 219         queue_work(irqfd_cleanup_wq, &irqfd->shutdown);
 220 }
 221
 222 /*
 223  * Called with wqh->lock held and interrupts disabled
 224  */
 225 static int
 226 irqfd_wakeup(wait_queue_t *wait, unsigned mode, int sync, void *key)
 227 {
 228         struct _irqfd *irqfd = container_of(wait, struct _irqfd, wait);
 229         unsigned long flags = (unsigned long)key;
 230         struct kvm_kernel_irq_routing_entry irq;
 231         struct kvm *kvm = irqfd->kvm;
 232         unsigned seq;
 233         int idx;
 234
 235         if (flags & POLLIN) {
 236                 idx = srcu_read_lock(&kvm->irq_srcu);
 237                 do {
 238                         seq = read_seqcount_begin(&irqfd->irq_entry_sc);
 239                         irq = irqfd->irq_entry;
 240                 } while (read_seqcount_retry(&irqfd->irq_entry_sc, seq));
 241                 /* An event has been signaled, inject an interrupt */
 242                 if (irq.type == KVM_IRQ_ROUTING_MSI)
 243                         kvm_set_msi(&irq, kvm, KVM_USERSPACE_IRQ_SOURCE_ID, 1,
 244                                         false);
 245                 else
 246                         schedule_work(&irqfd->inject);
 247                 srcu_read_unlock(&kvm->irq_srcu, idx);
 248         }
 249
 250         if (flags & POLLHUP) {
 251                 /* The eventfd is closing, detach from KVM */
 252                 unsigned long flags;
 253
 254                 spin_lock_irqsave(&kvm->irqfds.lock, flags);
 255
 256                 /*
 257                  * We must check if someone deactivated the irqfd before
 258                  * we could acquire the irqfds.lock since the item is
 259                  * deactivated from the KVM side before it is unhooked from
 260                  * the wait-queue.  If it is already deactivated, we can
 261                  * simply return knowing the other side will cleanup for us.
 262                  * We cannot race against the irqfd going away since the
 263                  * other side is required to acquire wqh->lock, which we hold
 264                  */
 265                 if (irqfd_is_active(irqfd))
 266                         irqfd_deactivate(irqfd);
 267
 268                 spin_unlock_irqrestore(&kvm->irqfds.lock, flags);
 269         }
 270
 271         return 0;
 272 }
 273
 274 static void
 275 irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
 276                         poll_table *pt)
 277 {
 278         struct _irqfd *irqfd = container_of(pt, struct _irqfd, pt);
 279         add_wait_queue(wqh, &irqfd->wait);
 280 }
 281
 282 /* Must be called under irqfds.lock */
 283 static void irqfd_update(struct kvm *kvm, struct _irqfd *irqfd)
 284 {
 285         struct kvm_kernel_irq_routing_entry *e;
 286         struct kvm_kernel_irq_routing_entry entries[KVM_NR_IRQCHIPS];
 287         int i, n_entries;
 288
 289         n_entries = kvm_irq_map_gsi(kvm, entries, irqfd->gsi);
 290
 291         write_seqcount_begin(&irqfd->irq_entry_sc);
 292
 293         irqfd->irq_entry.type = 0;
 294
 295         e = entries;
 296         for (i = 0; i < n_entries; ++i, ++e) {
 297                 /* Only fast-path MSI. */
 298                 if (e->type == KVM_IRQ_ROUTING_MSI)
 299                         irqfd->irq_entry = *e;
 300         }
 301
 302         write_seqcount_end(&irqfd->irq_entry_sc);
 303 }
 304
 305 static int
 306 kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
 307 {
 308         struct _irqfd *irqfd, *tmp;
 309         struct fd f;
 310         struct eventfd_ctx *eventfd = NULL, *resamplefd = NULL;
 311         int ret;
 312         unsigned int events;
 313         int idx;
 314
 315         irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL);
 316         if (!irqfd)
 317                 return -ENOMEM;
 318
 319         irqfd->kvm = kvm;
 320         irqfd->gsi = args->gsi;
 321         INIT_LIST_HEAD(&irqfd->list);
 322         INIT_WORK(&irqfd->inject, irqfd_inject);
 323         INIT_WORK(&irqfd->shutdown, irqfd_shutdown);
 324         seqcount_init(&irqfd->irq_entry_sc);
 325
 326         f = fdget(args->fd);
 327         if (!f.file) {
 328                 ret = -EBADF;
 329                 goto out;
 330         }
 331
 332         eventfd = eventfd_ctx_fileget(f.file);
 333         if (IS_ERR(eventfd)) {
 334                 ret = PTR_ERR(eventfd);
 335                 goto fail;
 336         }
 337
 338         irqfd->eventfd = eventfd;
 339
 340         if (args->flags & KVM_IRQFD_FLAG_RESAMPLE) {
 341                 struct _irqfd_resampler *resampler;
 342
 343                 resamplefd = eventfd_ctx_fdget(args->resamplefd);
 344                 if (IS_ERR(resamplefd)) {
 345                         ret = PTR_ERR(resamplefd);
 346                         goto fail;
 347                 }
 348
 349                 irqfd->resamplefd = resamplefd;
 350                 INIT_LIST_HEAD(&irqfd->resampler_link);
 351
 352                 mutex_lock(&kvm->irqfds.resampler_lock);
 353
 354                 list_for_each_entry(resampler,
 355                                     &kvm->irqfds.resampler_list, link) {
 356                         if (resampler->notifier.gsi == irqfd->gsi) {
 357                                 irqfd->resampler = resampler;
 358                                 break;
 359                         }
 360                 }
 361
 362                 if (!irqfd->resampler) {
 363                         resampler = kzalloc(sizeof(*resampler), GFP_KERNEL);
 364                         if (!resampler) {
 365                                 ret = -ENOMEM;
 366                                 mutex_unlock(&kvm->irqfds.resampler_lock);
 367                                 goto fail;
 368                         }
 369
 370                         resampler->kvm = kvm;
 371                         INIT_LIST_HEAD(&resampler->list);
 372                         resampler->notifier.gsi = irqfd->gsi;
 373                         resampler->notifier.irq_acked = irqfd_resampler_ack;
 374                         INIT_LIST_HEAD(&resampler->link);
 375
 376                         list_add(&resampler->link, &kvm->irqfds.resampler_list);
 377                         kvm_register_irq_ack_notifier(kvm,
 378                                                       &resampler->notifier);
 379                         irqfd->resampler = resampler;
 380                 }
 381
 382                 list_add_rcu(&irqfd->resampler_link, &irqfd->resampler->list);
 383                 synchronize_srcu(&kvm->irq_srcu);
 384
 385                 mutex_unlock(&kvm->irqfds.resampler_lock);
 386         }
 387
 388         /*
 389          * Install our own custom wake-up handling so we are notified via
 390          * a callback whenever someone signals the underlying eventfd
 391          */
 392         init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
 393         init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
 394
 395         spin_lock_irq(&kvm->irqfds.lock);
 396
 397         ret = 0;
 398         list_for_each_entry(tmp, &kvm->irqfds.items, list) {
 399                 if (irqfd->eventfd != tmp->eventfd)
 400                         continue;
 401                 /* This fd is used for another irq already. */
 402                 ret = -EBUSY;
 403                 spin_unlock_irq(&kvm->irqfds.lock);
 404                 goto fail;
 405         }
 406
 407         idx = srcu_read_lock(&kvm->irq_srcu);
 408         irqfd_update(kvm, irqfd);
 409         srcu_read_unlock(&kvm->irq_srcu, idx);
 410
 411         list_add_tail(&irqfd->list, &kvm->irqfds.items);
 412
 413         spin_unlock_irq(&kvm->irqfds.lock);
 414
 415         /*
 416          * Check if there was an event already pending on the eventfd
 417          * before we registered, and trigger it as if we didn't miss it.
 418          */
 419         events = f.file->f_op->poll(f.file, &irqfd->pt);
 420
 421         if (events & POLLIN)
 422                 schedule_work(&irqfd->inject);
 423
 424         /*
 425          * do not drop the file until the irqfd is fully initialized, otherwise
 426          * we might race against the POLLHUP
 427          */
 428         fdput(f);
 429
 430         return 0;
 431
 432 fail:
 433         if (irqfd->resampler)
 434                 irqfd_resampler_shutdown(irqfd);
 435
 436         if (resamplefd && !IS_ERR(resamplefd))
 437                 eventfd_ctx_put(resamplefd);
 438
 439         if (eventfd && !IS_ERR(eventfd))
 440                 eventfd_ctx_put(eventfd);
 441
 442         fdput(f);
 443
 444 out:
 445         kfree(irqfd);
 446         return ret;
 447 }
 448
 449 bool kvm_irq_has_notifier(struct kvm *kvm, unsigned irqchip, unsigned pin)
 450 {
 451         struct kvm_irq_ack_notifier *kian;
 452         int gsi, idx;
 453
 454         idx = srcu_read_lock(&kvm->irq_srcu);
 455         gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
 456         if (gsi != -1)
 457                 hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
 458                                          link)
 459                         if (kian->gsi == gsi) {
 460                                 srcu_read_unlock(&kvm->irq_srcu, idx);
 461                                 return true;
 462                         }
 463
 464         srcu_read_unlock(&kvm->irq_srcu, idx);
 465
 466         return false;
 467 }
 468 EXPORT_SYMBOL_GPL(kvm_irq_has_notifier);
 469
 470 void kvm_notify_acked_irq(struct kvm *kvm, unsigned irqchip, unsigned pin)
 471 {
 472         struct kvm_irq_ack_notifier *kian;
 473         int gsi, idx;
 474
 475         trace_kvm_ack_irq(irqchip, pin);
 476
 477         idx = srcu_read_lock(&kvm->irq_srcu);
 478         gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
 479         if (gsi != -1)
 480                 hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
 481                                          link)
 482                         if (kian->gsi == gsi)
 483                                 kian->irq_acked(kian);
 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 #ifdef __KVM_HAVE_IOAPIC
 494         kvm_vcpu_request_scan_ioapic(kvm);
 495 #endif
 496 }
 497
 498 void kvm_unregister_irq_ack_notifier(struct kvm *kvm,
 499                                     struct kvm_irq_ack_notifier *kian)
 500 {
 501         mutex_lock(&kvm->irq_lock);
 502         hlist_del_init_rcu(&kian->link);
 503         mutex_unlock(&kvm->irq_lock);
 504         synchronize_srcu(&kvm->irq_srcu);
 505 #ifdef __KVM_HAVE_IOAPIC
 506         kvm_vcpu_request_scan_ioapic(kvm);
 507 #endif
 508 }
 509 #endif
 510
 511 void
 512 kvm_eventfd_init(struct kvm *kvm)
 513 {
 514 #ifdef CONFIG_HAVE_KVM_IRQFD
 515         spin_lock_init(&kvm->irqfds.lock);
 516         INIT_LIST_HEAD(&kvm->irqfds.items);
 517         INIT_LIST_HEAD(&kvm->irqfds.resampler_list);
 518         mutex_init(&kvm->irqfds.resampler_lock);
 519 #endif
 520         INIT_LIST_HEAD(&kvm->ioeventfds);
 521 }
 522
 523 #ifdef CONFIG_HAVE_KVM_IRQFD
 524 /*
 525  * shutdown any irqfd's that match fd+gsi
 526  */
 527 static int
 528 kvm_irqfd_deassign(struct kvm *kvm, struct kvm_irqfd *args)
 529 {
 530         struct _irqfd *irqfd, *tmp;
 531         struct eventfd_ctx *eventfd;
 532
 533         eventfd = eventfd_ctx_fdget(args->fd);
 534         if (IS_ERR(eventfd))
 535                 return PTR_ERR(eventfd);
 536
 537         spin_lock_irq(&kvm->irqfds.lock);
 538
 539         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list) {
 540                 if (irqfd->eventfd == eventfd && irqfd->gsi == args->gsi) {
 541                         /*
 542                          * This clearing of irq_entry.type is needed for when
 543                          * another thread calls kvm_irq_routing_update before
 544                          * we flush workqueue below (we synchronize with
 545                          * kvm_irq_routing_update using irqfds.lock).
 546                          */
 547                         write_seqcount_begin(&irqfd->irq_entry_sc);
 548                         irqfd->irq_entry.type = 0;
 549                         write_seqcount_end(&irqfd->irq_entry_sc);
 550                         irqfd_deactivate(irqfd);
 551                 }
 552         }
 553
 554         spin_unlock_irq(&kvm->irqfds.lock);
 555         eventfd_ctx_put(eventfd);
 556
 557         /*
 558          * Block until we know all outstanding shutdown jobs have completed
 559          * so that we guarantee there will not be any more interrupts on this
 560          * gsi once this deassign function returns.
 561          */
 562         flush_workqueue(irqfd_cleanup_wq);
 563
 564         return 0;
 565 }
 566
 567 int
 568 kvm_irqfd(struct kvm *kvm, struct kvm_irqfd *args)
 569 {
 570         if (args->flags & ~(KVM_IRQFD_FLAG_DEASSIGN | KVM_IRQFD_FLAG_RESAMPLE))
 571                 return -EINVAL;
 572
 573         if (args->flags & KVM_IRQFD_FLAG_DEASSIGN)
 574                 return kvm_irqfd_deassign(kvm, args);
 575
 576         return kvm_irqfd_assign(kvm, args);
 577 }
 578
 579 /*
 580  * This function is called as the kvm VM fd is being released. Shutdown all
 581  * irqfds that still remain open
 582  */
 583 void
 584 kvm_irqfd_release(struct kvm *kvm)
 585 {
 586         struct _irqfd *irqfd, *tmp;
 587
 588         spin_lock_irq(&kvm->irqfds.lock);
 589
 590         list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds.items, list)
 591                 irqfd_deactivate(irqfd);
 592
 593         spin_unlock_irq(&kvm->irqfds.lock);
 594
 595         /*
 596          * Block until we know all outstanding shutdown jobs have completed
 597          * since we do not take a kvm* reference.
 598          */
 599         flush_workqueue(irqfd_cleanup_wq);
 600
 601 }
 602
 603 /*
 604  * Take note of a change in irq routing.
 605  * Caller must invoke synchronize_srcu(&kvm->irq_srcu) afterwards.
 606  */
 607 void kvm_irq_routing_update(struct kvm *kvm)
 608 {
 609         struct _irqfd *irqfd;
 610
 611         spin_lock_irq(&kvm->irqfds.lock);
 612
 613         list_for_each_entry(irqfd, &kvm->irqfds.items, list)
 614                 irqfd_update(kvm, irqfd);
 615
 616         spin_unlock_irq(&kvm->irqfds.lock);
 617 }
 618
 619 /*
 620  * create a host-wide workqueue for issuing deferred shutdown requests
 621  * aggregated from all vm* instances. We need our own isolated single-thread
 622  * queue to prevent deadlock against flushing the normal work-queue.
 623  */
 624 int kvm_irqfd_init(void)
 625 {
 626         irqfd_cleanup_wq = create_singlethread_workqueue("kvm-irqfd-cleanup");
 627         if (!irqfd_cleanup_wq)
 628                 return -ENOMEM;
 629
 630         return 0;
 631 }
 632
 633 void kvm_irqfd_exit(void)
 634 {
 635         destroy_workqueue(irqfd_cleanup_wq);
 636 }
 637 #endif
 638
 639 /*
 640  * --------------------------------------------------------------------
 641  * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
 642  *
 643  * userspace can register a PIO/MMIO address with an eventfd for receiving
 644  * notification when the memory has been touched.
 645  * --------------------------------------------------------------------
 646  */
 647
 648 struct _ioeventfd {
 649         struct list_head     list;
 650         u64                  addr;
 651         int                  length;
 652         struct eventfd_ctx  *eventfd;
 653         u64                  datamatch;
 654         struct kvm_io_device dev;
 655         u8                   bus_idx;
 656         bool                 wildcard;
 657 };
 658
 659 static inline struct _ioeventfd *
 660 to_ioeventfd(struct kvm_io_device *dev)
 661 {
 662         return container_of(dev, struct _ioeventfd, dev);
 663 }
 664
 665 static void
 666 ioeventfd_release(struct _ioeventfd *p)
 667 {
 668         eventfd_ctx_put(p->eventfd);
 669         list_del(&p->list);
 670         kfree(p);
 671 }
 672
 673 static bool
 674 ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
 675 {
 676         u64 _val;
 677
 678         if (addr != p->addr)
 679                 /* address must be precise for a hit */
 680                 return false;
 681
 682         if (!p->length)
 683                 /* length = 0 means only look at the address, so always a hit */
 684                 return true;
 685
 686         if (len != p->length)
 687                 /* address-range must be precise for a hit */
 688                 return false;
 689
 690         if (p->wildcard)
 691                 /* all else equal, wildcard is always a hit */
 692                 return true;
 693
 694         /* otherwise, we have to actually compare the data */
 695
 696         BUG_ON(!IS_ALIGNED((unsigned long)val, len));
 697
 698         switch (len) {
 699         case 1:
 700                 _val = *(u8 *)val;
 701                 break;
 702         case 2:
 703                 _val = *(u16 *)val;
 704                 break;
 705         case 4:
 706                 _val = *(u32 *)val;
 707                 break;
 708         case 8:
 709                 _val = *(u64 *)val;
 710                 break;
 711         default:
 712                 return false;
 713         }
 714
 715         return _val == p->datamatch ? true : false;
 716 }
 717
 718 /* MMIO/PIO writes trigger an event if the addr/val match */
 719 static int
 720 ioeventfd_write(struct kvm_io_device *this, gpa_t addr, int len,
 721                 const void *val)
 722 {
 723         struct _ioeventfd *p = to_ioeventfd(this);
 724
 725         if (!ioeventfd_in_range(p, addr, len, val))
 726                 return -EOPNOTSUPP;
 727
 728         eventfd_signal(p->eventfd, 1);
 729         return 0;
 730 }
 731
 732 /*
 733  * This function is called as KVM is completely shutting down.  We do not
 734  * need to worry about locking just nuke anything we have as quickly as possible
 735  */
 736 static void
 737 ioeventfd_destructor(struct kvm_io_device *this)
 738 {
 739         struct _ioeventfd *p = to_ioeventfd(this);
 740
 741         ioeventfd_release(p);
 742 }
 743
 744 static const struct kvm_io_device_ops ioeventfd_ops = {
 745         .write      = ioeventfd_write,
 746         .destructor = ioeventfd_destructor,
 747 };
 748
 749 /* assumes kvm->slots_lock held */
 750 static bool
 751 ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
 752 {
 753         struct _ioeventfd *_p;
 754
 755         list_for_each_entry(_p, &kvm->ioeventfds, list)
 756                 if (_p->bus_idx == p->bus_idx &&
 757                     _p->addr == p->addr &&
 758                     (!_p->length || !p->length ||
 759                      (_p->length == p->length &&
 760                       (_p->wildcard || p->wildcard ||
 761                        _p->datamatch == p->datamatch))))
 762                         return true;
 763
 764         return false;
 765 }
 766
 767 static enum kvm_bus ioeventfd_bus_from_flags(__u32 flags)
 768 {
 769         if (flags & KVM_IOEVENTFD_FLAG_PIO)
 770                 return KVM_PIO_BUS;
 771         if (flags & KVM_IOEVENTFD_FLAG_VIRTIO_CCW_NOTIFY)
 772                 return KVM_VIRTIO_CCW_NOTIFY_BUS;
 773         return KVM_MMIO_BUS;
 774 }
 775
 776 static int
 777 kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 778 {
 779         enum kvm_bus              bus_idx;
 780         struct _ioeventfd        *p;
 781         struct eventfd_ctx       *eventfd;
 782         int                       ret;
 783
 784         bus_idx = ioeventfd_bus_from_flags(args->flags);
 785         /* must be natural-word sized, or 0 to ignore length */
 786         switch (args->len) {
 787         case 0:
 788         case 1:
 789         case 2:
 790         case 4:
 791         case 8:
 792                 break;
 793         default:
 794                 return -EINVAL;
 795         }
 796
 797         /* check for range overflow */
 798         if (args->addr + args->len < args->addr)
 799                 return -EINVAL;
 800
 801         /* check for extra flags that we don't understand */
 802         if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
 803                 return -EINVAL;
 804
 805         /* ioeventfd with no length can't be combined with DATAMATCH */
 806         if (!args->len &&
 807             args->flags & (KVM_IOEVENTFD_FLAG_PIO |
 808                            KVM_IOEVENTFD_FLAG_DATAMATCH))
 809                 return -EINVAL;
 810
 811         eventfd = eventfd_ctx_fdget(args->fd);
 812         if (IS_ERR(eventfd))
 813                 return PTR_ERR(eventfd);
 814
 815         p = kzalloc(sizeof(*p), GFP_KERNEL);
 816         if (!p) {
 817                 ret = -ENOMEM;
 818                 goto fail;
 819         }
 820
 821         INIT_LIST_HEAD(&p->list);
 822         p->addr    = args->addr;
 823         p->bus_idx = bus_idx;
 824         p->length  = args->len;
 825         p->eventfd = eventfd;
 826
 827         /* The datamatch feature is optional, otherwise this is a wildcard */
 828         if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
 829                 p->datamatch = args->datamatch;
 830         else
 831                 p->wildcard = true;
 832
 833         mutex_lock(&kvm->slots_lock);
 834
 835         /* Verify that there isn't a match already */
 836         if (ioeventfd_check_collision(kvm, p)) {
 837                 ret = -EEXIST;
 838                 goto unlock_fail;
 839         }
 840
 841         kvm_iodevice_init(&p->dev, &ioeventfd_ops);
 842
 843         ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
 844                                       &p->dev);
 845         if (ret < 0)
 846                 goto unlock_fail;
 847
 848         /* When length is ignored, MMIO is also put on a separate bus, for
 849          * faster lookups.
 850          */
 851         if (!args->len && !(args->flags & KVM_IOEVENTFD_FLAG_PIO)) {
 852                 ret = kvm_io_bus_register_dev(kvm, KVM_FAST_MMIO_BUS,
 853                                               p->addr, 0, &p->dev);
 854                 if (ret < 0)
 855                         goto register_fail;
 856         }
 857
 858         kvm->buses[bus_idx]->ioeventfd_count++;
 859         list_add_tail(&p->list, &kvm->ioeventfds);
 860
 861         mutex_unlock(&kvm->slots_lock);
 862
 863         return 0;
 864
 865 register_fail:
 866         kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
 867 unlock_fail:
 868         mutex_unlock(&kvm->slots_lock);
 869
 870 fail:
 871         kfree(p);
 872         eventfd_ctx_put(eventfd);
 873
 874         return ret;
 875 }
 876
 877 static int
 878 kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 879 {
 880         enum kvm_bus              bus_idx;
 881         struct _ioeventfd        *p, *tmp;
 882         struct eventfd_ctx       *eventfd;
 883         int                       ret = -ENOENT;
 884
 885         bus_idx = ioeventfd_bus_from_flags(args->flags);
 886         eventfd = eventfd_ctx_fdget(args->fd);
 887         if (IS_ERR(eventfd))
 888                 return PTR_ERR(eventfd);
 889
 890         mutex_lock(&kvm->slots_lock);
 891
 892         list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
 893                 bool wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
 894
 895                 if (p->bus_idx != bus_idx ||
 896                     p->eventfd != eventfd  ||
 897                     p->addr != args->addr  ||
 898                     p->length != args->len ||
 899                     p->wildcard != wildcard)
 900                         continue;
 901
 902                 if (!p->wildcard && p->datamatch != args->datamatch)
 903                         continue;
 904
 905                 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
 906                 if (!p->length) {
 907                         kvm_io_bus_unregister_dev(kvm, KVM_FAST_MMIO_BUS,
 908                                                   &p->dev);
 909                 }
 910                 kvm->buses[bus_idx]->ioeventfd_count--;
 911                 ioeventfd_release(p);
 912                 ret = 0;
 913                 break;
 914         }
 915
 916         mutex_unlock(&kvm->slots_lock);
 917
 918         eventfd_ctx_put(eventfd);
 919
 920         return ret;
 921 }
 922
 923 int
 924 kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 925 {
 926         if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
 927                 return kvm_deassign_ioeventfd(kvm, args);
 928
 929         return kvm_assign_ioeventfd(kvm, args);
 930 }