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