virt/kvm/eventfd.c

   1 /*
   2  * kvm eventfd support - use eventfd objects to signal various KVM events
   3  *
   4  * Copyright 2009 Novell.  All Rights Reserved.
   5  * Copyright 2010 Red Hat, Inc. and/or its affiliates.
   6  *
   7  * Author:
   8  *      Gregory Haskins <[email protected]>
   9  *
  10  * This file is free software; you can redistribute it and/or modify
  11  * it under the terms of version 2 of the GNU General Public License
  12  * as published by the Free Software Foundation.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU General Public License
  20  * along with this program; if not, write to the Free Software Foundation,
  21  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA.
  22  */
  23
  24 #include <linux/kvm_host.h>
  25 #include <linux/kvm.h>
  26 #include <linux/kvm_irqfd.h>
  27 #include <linux/workqueue.h>
  28 #include <linux/syscalls.h>
  29 #include <linux/wait.h>
  30 #include <linux/poll.h>
  31 #include <linux/file.h>
  32 #include <linux/list.h>
  33 #include <linux/eventfd.h>
  34 #include <linux/kernel.h>
  35 #include <linux/srcu.h>
  36 #include <linux/slab.h>
  37 #include <linux/seqlock.h>
  38 #include <linux/irqbypass.h>
  39 #include <trace/events/kvm.h>
  40
  41 #include <kvm/iodev.h>
  42
  43 #ifdef CONFIG_HAVE_KVM_IRQFD
  44
  45
  46 static void
  47 irqfd_inject(struct work_struct *work)
  48 {
  49         struct kvm_kernel_irqfd *irqfd =
  50                 container_of(work, struct kvm_kernel_irqfd, inject);
  51         struct kvm *kvm = irqfd->kvm;
  52
  53         if (!irqfd->resampler) {
  54                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1,
  55                                 false);
  56                 kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0,
  57                                 false);
  58         } else
  59                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
  60                             irqfd->gsi, 1, false);
  61 }
  62
  63 /*
  64  * Since resampler irqfds share an IRQ source ID, we de-assert once
  65  * then notify all of the resampler irqfds using this GSI.  We can't
  66  * do multiple de-asserts or we risk racing with incoming re-asserts.
  67  */
  68 static void
  69 irqfd_resampler_ack(struct kvm_irq_ack_notifier *kian)
  70 {
  71         struct kvm_kernel_irqfd_resampler *resampler;
  72         struct kvm *kvm;
  73         struct kvm_kernel_irqfd *irqfd;
  74         int idx;
  75
  76         resampler = container_of(kian,
  77                         struct kvm_kernel_irqfd_resampler, notifier);
  78         kvm = resampler->kvm;
  79
  80         kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
  81                     resampler->notifier.gsi, 0, false);
  82
  83         idx = srcu_read_lock(&kvm->irq_srcu);
  84
  85         list_for_each_entry_rcu(irqfd, &resampler->list, resampler_link)
  86                 eventfd_signal(irqfd->resamplefd, 1);
  87
  88         srcu_read_unlock(&kvm->irq_srcu, idx);
  89 }
  90
  91 static void
  92 irqfd_resampler_shutdown(struct kvm_kernel_irqfd *irqfd)
  93 {
  94         struct kvm_kernel_irqfd_resampler *resampler = irqfd->resampler;
  95         struct kvm *kvm = resampler->kvm;
  96
  97         mutex_lock(&kvm->irqfds.resampler_lock);
  98
  99         list_del_rcu(&irqfd->resampler_link);
 100         synchronize_srcu(&kvm->irq_srcu);
 101
 102         if (list_empty(&resampler->list)) {
 103                 list_del(&resampler->link);
 104                 kvm_unregister_irq_ack_notifier(kvm, &resampler->notifier);
 105                 kvm_set_irq(kvm, KVM_IRQFD_RESAMPLE_IRQ_SOURCE_ID,
 106                             resampler->notifier.gsi, 0, false);
 107                 kfree(resampler);
 108         }
 109
 110         mutex_unlock(&kvm->irqfds.resampler_lock);
 111 }
 112
 113 /*
 114  * Race-free decouple logic (ordering is critical)
 115  */
 116 static void
 117 irqfd_shutdown(struct work_struct *work)
 118 {
 119         struct kvm_kernel_irqfd *irqfd =
 120                 container_of(work, struct kvm_kernel_irqfd, shutdown);
 121         u64 cnt;
 122
 123         /*
 124          * Synchronize with the wait-queue and unhook ourselves to prevent
 125          * further events.
 126          */
 127         eventfd_ctx_remove_wait_queue(irqfd->eventfd, &irqfd->wait, &cnt);
 128
 129         /*
 130          * We know no new events will be scheduled at this point, so block
 131          * until all previously outstanding events have completed
 132          */
 133         flush_work(&irqfd->inject);
 134
 135         if (irqfd->resampler) {
 136                 irqfd_resampler_shutdown(irqfd);
 137                 eventfd_ctx_put(irqfd->resamplefd);
 138         }
 139
 140         /*
 141          * It is now safe to release the object's resources
 142          */
 143 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 144         irq_bypass_unregister_consumer(&irqfd->consumer);
 145 #endif
 146         eventfd_ctx_put(irqfd->eventfd);
 147         kfree(irqfd);
 148 }
 149
 150
 151 /* assumes kvm->irqfds.lock is held */
 152 static bool
 153 irqfd_is_active(struct kvm_kernel_irqfd *irqfd)
 154 {
 155         return list_empty(&irqfd->list) ? false : true;
 156 }
 157
 158 /*
 159  * Mark the irqfd as inactive and schedule it for removal
 160  *
 161  * assumes kvm->irqfds.lock is held
 162  */
 163 static void
 164 irqfd_deactivate(struct kvm_kernel_irqfd *irqfd)
 165 {
 166         BUG_ON(!irqfd_is_active(irqfd));
 167
 168         list_del_init(&irqfd->list);
 169
 170         schedule_work(&irqfd->shutdown);
 171 }
 172
 173 int __attribute__((weak)) kvm_arch_set_irq_inatomic(
 174                                 struct kvm_kernel_irq_routing_entry *irq,
 175                                 struct kvm *kvm, int irq_source_id,
 176                                 int level,
 177                                 bool line_status)
 178 {
 179         return -EWOULDBLOCK;
 180 }
 181
 182 /*
 183  * Called with wqh->lock held and interrupts disabled
 184  */
 185 static int
 186 irqfd_wakeup(wait_queue_t *wait, unsigned mode, int sync, void *key)
 187 {
 188         struct kvm_kernel_irqfd *irqfd =
 189                 container_of(wait, struct kvm_kernel_irqfd, wait);
 190         unsigned long flags = (unsigned long)key;
 191         struct kvm_kernel_irq_routing_entry irq;
 192         struct kvm *kvm = irqfd->kvm;
 193         unsigned seq;
 194         int idx;
 195
 196         if (flags & POLLIN) {
 197                 idx = srcu_read_lock(&kvm->irq_srcu);
 198                 do {
 199                         seq = read_seqcount_begin(&irqfd->irq_entry_sc);
 200                         irq = irqfd->irq_entry;
 201                 } while (read_seqcount_retry(&irqfd->irq_entry_sc, seq));
 202                 /* An event has been signaled, inject an interrupt */
 203                 if (kvm_arch_set_irq_inatomic(&irq, kvm,
 204                                               KVM_USERSPACE_IRQ_SOURCE_ID, 1,
 205                                               false) == -EWOULDBLOCK)
 206                         schedule_work(&irqfd->inject);
 207                 srcu_read_unlock(&kvm->irq_srcu, idx);
 208         }
 209
 210         if (flags & POLLHUP) {
 211                 /* The eventfd is closing, detach from KVM */
 212                 unsigned long flags;
 213
 214                 spin_lock_irqsave(&kvm->irqfds.lock, flags);
 215
 216                 /*
 217                  * We must check if someone deactivated the irqfd before
 218                  * we could acquire the irqfds.lock since the item is
 219                  * deactivated from the KVM side before it is unhooked from
 220                  * the wait-queue.  If it is already deactivated, we can
 221                  * simply return knowing the other side will cleanup for us.
 222                  * We cannot race against the irqfd going away since the
 223                  * other side is required to acquire wqh->lock, which we hold
 224                  */
 225                 if (irqfd_is_active(irqfd))
 226                         irqfd_deactivate(irqfd);
 227
 228                 spin_unlock_irqrestore(&kvm->irqfds.lock, flags);
 229         }
 230
 231         return 0;
 232 }
 233
 234 static void
 235 irqfd_ptable_queue_proc(struct file *file, wait_queue_head_t *wqh,
 236                         poll_table *pt)
 237 {
 238         struct kvm_kernel_irqfd *irqfd =
 239                 container_of(pt, struct kvm_kernel_irqfd, pt);
 240         add_wait_queue(wqh, &irqfd->wait);
 241 }
 242
 243 /* Must be called under irqfds.lock */
 244 static void irqfd_update(struct kvm *kvm, struct kvm_kernel_irqfd *irqfd)
 245 {
 246         struct kvm_kernel_irq_routing_entry *e;
 247         struct kvm_kernel_irq_routing_entry entries[KVM_NR_IRQCHIPS];
 248         int n_entries;
 249
 250         n_entries = kvm_irq_map_gsi(kvm, entries, irqfd->gsi);
 251
 252         write_seqcount_begin(&irqfd->irq_entry_sc);
 253
 254         e = entries;
 255         if (n_entries == 1)
 256                 irqfd->irq_entry = *e;
 257         else
 258                 irqfd->irq_entry.type = 0;
 259
 260         write_seqcount_end(&irqfd->irq_entry_sc);
 261 }
 262
 263 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 264 void __attribute__((weak)) kvm_arch_irq_bypass_stop(
 265                                 struct irq_bypass_consumer *cons)
 266 {
 267 }
 268
 269 void __attribute__((weak)) kvm_arch_irq_bypass_start(
 270                                 struct irq_bypass_consumer *cons)
 271 {
 272 }
 273
 274 int  __attribute__((weak)) kvm_arch_update_irqfd_routing(
 275                                 struct kvm *kvm, unsigned int host_irq,
 276                                 uint32_t guest_irq, bool set)
 277 {
 278         return 0;
 279 }
 280 #endif
 281
 282 static int
 283 kvm_irqfd_assign(struct kvm *kvm, struct kvm_irqfd *args)
 284 {
 285         struct kvm_kernel_irqfd *irqfd, *tmp;
 286         struct fd f;
 287         struct eventfd_ctx *eventfd = NULL, *resamplefd = NULL;
 288         int ret;
 289         unsigned int events;
 290         int idx;
 291
 292         if (!kvm_arch_intc_initialized(kvm))
 293                 return -EAGAIN;
 294
 295         irqfd = kzalloc(sizeof(*irqfd), GFP_KERNEL);
 296         if (!irqfd)
 297                 return -ENOMEM;
 298
 299         irqfd->kvm = kvm;
 300         irqfd->gsi = args->gsi;
 301         INIT_LIST_HEAD(&irqfd->list);
 302         INIT_WORK(&irqfd->inject, irqfd_inject);
 303         INIT_WORK(&irqfd->shutdown, irqfd_shutdown);
 304         seqcount_init(&irqfd->irq_entry_sc);
 305
 306         f = fdget(args->fd);
 307         if (!f.file) {
 308                 ret = -EBADF;
 309                 goto out;
 310         }
 311
 312         eventfd = eventfd_ctx_fileget(f.file);
 313         if (IS_ERR(eventfd)) {
 314                 ret = PTR_ERR(eventfd);
 315                 goto fail;
 316         }
 317
 318         irqfd->eventfd = eventfd;
 319
 320         if (args->flags & KVM_IRQFD_FLAG_RESAMPLE) {
 321                 struct kvm_kernel_irqfd_resampler *resampler;
 322
 323                 resamplefd = eventfd_ctx_fdget(args->resamplefd);
 324                 if (IS_ERR(resamplefd)) {
 325                         ret = PTR_ERR(resamplefd);
 326                         goto fail;
 327                 }
 328
 329                 irqfd->resamplefd = resamplefd;
 330                 INIT_LIST_HEAD(&irqfd->resampler_link);
 331
 332                 mutex_lock(&kvm->irqfds.resampler_lock);
 333
 334                 list_for_each_entry(resampler,
 335                                     &kvm->irqfds.resampler_list, link) {
 336                         if (resampler->notifier.gsi == irqfd->gsi) {
 337                                 irqfd->resampler = resampler;
 338                                 break;
 339                         }
 340                 }
 341
 342                 if (!irqfd->resampler) {
 343                         resampler = kzalloc(sizeof(*resampler), GFP_KERNEL);
 344                         if (!resampler) {
 345                                 ret = -ENOMEM;
 346                                 mutex_unlock(&kvm->irqfds.resampler_lock);
 347                                 goto fail;
 348                         }
 349
 350                         resampler->kvm = kvm;
 351                         INIT_LIST_HEAD(&resampler->list);
 352                         resampler->notifier.gsi = irqfd->gsi;
 353                         resampler->notifier.irq_acked = irqfd_resampler_ack;
 354                         INIT_LIST_HEAD(&resampler->link);
 355
 356                         list_add(&resampler->link, &kvm->irqfds.resampler_list);
 357                         kvm_register_irq_ack_notifier(kvm,
 358                                                       &resampler->notifier);
 359                         irqfd->resampler = resampler;
 360                 }
 361
 362                 list_add_rcu(&irqfd->resampler_link, &irqfd->resampler->list);
 363                 synchronize_srcu(&kvm->irq_srcu);
 364
 365                 mutex_unlock(&kvm->irqfds.resampler_lock);
 366         }
 367
 368         /*
 369          * Install our own custom wake-up handling so we are notified via
 370          * a callback whenever someone signals the underlying eventfd
 371          */
 372         init_waitqueue_func_entry(&irqfd->wait, irqfd_wakeup);
 373         init_poll_funcptr(&irqfd->pt, irqfd_ptable_queue_proc);
 374
 375         spin_lock_irq(&kvm->irqfds.lock);
 376
 377         ret = 0;
 378         list_for_each_entry(tmp, &kvm->irqfds.items, list) {
 379                 if (irqfd->eventfd != tmp->eventfd)
 380                         continue;
 381                 /* This fd is used for another irq already. */
 382                 ret = -EBUSY;
 383                 spin_unlock_irq(&kvm->irqfds.lock);
 384                 goto fail;
 385         }
 386
 387         idx = srcu_read_lock(&kvm->irq_srcu);
 388         irqfd_update(kvm, irqfd);
 389         srcu_read_unlock(&kvm->irq_srcu, idx);
 390
 391         list_add_tail(&irqfd->list, &kvm->irqfds.items);
 392
 393         spin_unlock_irq(&kvm->irqfds.lock);
 394
 395         /*
 396          * Check if there was an event already pending on the eventfd
 397          * before we registered, and trigger it as if we didn't miss it.
 398          */
 399         events = f.file->f_op->poll(f.file, &irqfd->pt);
 400
 401         if (events & POLLIN)
 402                 schedule_work(&irqfd->inject);
 403
 404         /*
 405          * do not drop the file until the irqfd is fully initialized, otherwise
 406          * we might race against the POLLHUP
 407          */
 408         fdput(f);
 409 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 410         if (kvm_arch_has_irq_bypass()) {
 411                 irqfd->consumer.token = (void *)irqfd->eventfd;
 412                 irqfd->consumer.add_producer = kvm_arch_irq_bypass_add_producer;
 413                 irqfd->consumer.del_producer = kvm_arch_irq_bypass_del_producer;
 414                 irqfd->consumer.stop = kvm_arch_irq_bypass_stop;
 415                 irqfd->consumer.start = kvm_arch_irq_bypass_start;
 416                 ret = irq_bypass_register_consumer(&irqfd->consumer);
 417                 if (ret)
 418                         pr_info("irq bypass consumer (token %p) registration fails: %d\n",
 419                                 irqfd->consumer.token, ret);
 420         }
 421 #endif
 422
 423         return 0;
 424
 425 fail:
 426         if (irqfd->resampler)
 427                 irqfd_resampler_shutdown(irqfd);
 428
 429         if (resamplefd && !IS_ERR(resamplefd))
 430                 eventfd_ctx_put(resamplefd);
 431
 432         if (eventfd && !IS_ERR(eventfd))
 433                 eventfd_ctx_put(eventfd);
 434
 435         fdput(f);
 436
 437 out:
 438         kfree(irqfd);
 439         return ret;
 440 }
 441
 442 bool kvm_irq_has_notifier(struct kvm *kvm, unsigned irqchip, unsigned pin)
 443 {
 444         struct kvm_irq_ack_notifier *kian;
 445         int gsi, idx;
 446
 447         idx = srcu_read_lock(&kvm->irq_srcu);
 448         gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
 449         if (gsi != -1)
 450                 hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
 451                                          link)
 452                         if (kian->gsi == gsi) {
 453                                 srcu_read_unlock(&kvm->irq_srcu, idx);
 454                                 return true;
 455                         }
 456
 457         srcu_read_unlock(&kvm->irq_srcu, idx);
 458
 459         return false;
 460 }
 461 EXPORT_SYMBOL_GPL(kvm_irq_has_notifier);
 462
 463 void kvm_notify_acked_gsi(struct kvm *kvm, int gsi)
 464 {
 465         struct kvm_irq_ack_notifier *kian;
 466
 467         hlist_for_each_entry_rcu(kian, &kvm->irq_ack_notifier_list,
 468                                  link)
 469                 if (kian->gsi == gsi)
 470                         kian->irq_acked(kian);
 471 }
 472
 473 void kvm_notify_acked_irq(struct kvm *kvm, unsigned irqchip, unsigned pin)
 474 {
 475         int gsi, idx;
 476
 477         trace_kvm_ack_irq(irqchip, pin);
 478
 479         idx = srcu_read_lock(&kvm->irq_srcu);
 480         gsi = kvm_irq_map_chip_pin(kvm, irqchip, pin);
 481         if (gsi != -1)
 482                 kvm_notify_acked_gsi(kvm, gsi);
 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 kvm_kernel_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_work(&irqfd->shutdown);
 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 kvm_kernel_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_work(&irqfd->shutdown);
 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 kvm_kernel_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 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
 612                 if (irqfd->producer) {
 613                         int ret = kvm_arch_update_irqfd_routing(
 614                                         irqfd->kvm, irqfd->producer->irq,
 615                                         irqfd->gsi, 1);
 616                         WARN_ON(ret);
 617                 }
 618 #endif
 619         }
 620
 621         spin_unlock_irq(&kvm->irqfds.lock);
 622 }
 623
 624 void kvm_irqfd_exit(void)
 625 {
 626 }
 627 #endif
 628
 629 /*
 630  * --------------------------------------------------------------------
 631  * ioeventfd: translate a PIO/MMIO memory write to an eventfd signal.
 632  *
 633  * userspace can register a PIO/MMIO address with an eventfd for receiving
 634  * notification when the memory has been touched.
 635  * --------------------------------------------------------------------
 636  */
 637
 638 struct _ioeventfd {
 639         struct list_head     list;
 640         u64                  addr;
 641         int                  length;
 642         struct eventfd_ctx  *eventfd;
 643         u64                  datamatch;
 644         struct kvm_io_device dev;
 645         u8                   bus_idx;
 646         bool                 wildcard;
 647 };
 648
 649 static inline struct _ioeventfd *
 650 to_ioeventfd(struct kvm_io_device *dev)
 651 {
 652         return container_of(dev, struct _ioeventfd, dev);
 653 }
 654
 655 static void
 656 ioeventfd_release(struct _ioeventfd *p)
 657 {
 658         eventfd_ctx_put(p->eventfd);
 659         list_del(&p->list);
 660         kfree(p);
 661 }
 662
 663 static bool
 664 ioeventfd_in_range(struct _ioeventfd *p, gpa_t addr, int len, const void *val)
 665 {
 666         u64 _val;
 667
 668         if (addr != p->addr)
 669                 /* address must be precise for a hit */
 670                 return false;
 671
 672         if (!p->length)
 673                 /* length = 0 means only look at the address, so always a hit */
 674                 return true;
 675
 676         if (len != p->length)
 677                 /* address-range must be precise for a hit */
 678                 return false;
 679
 680         if (p->wildcard)
 681                 /* all else equal, wildcard is always a hit */
 682                 return true;
 683
 684         /* otherwise, we have to actually compare the data */
 685
 686         BUG_ON(!IS_ALIGNED((unsigned long)val, len));
 687
 688         switch (len) {
 689         case 1:
 690                 _val = *(u8 *)val;
 691                 break;
 692         case 2:
 693                 _val = *(u16 *)val;
 694                 break;
 695         case 4:
 696                 _val = *(u32 *)val;
 697                 break;
 698         case 8:
 699                 _val = *(u64 *)val;
 700                 break;
 701         default:
 702                 return false;
 703         }
 704
 705         return _val == p->datamatch ? true : false;
 706 }
 707
 708 /* MMIO/PIO writes trigger an event if the addr/val match */
 709 static int
 710 ioeventfd_write(struct kvm_vcpu *vcpu, struct kvm_io_device *this, gpa_t addr,
 711                 int len, const void *val)
 712 {
 713         struct _ioeventfd *p = to_ioeventfd(this);
 714
 715         if (!ioeventfd_in_range(p, addr, len, val))
 716                 return -EOPNOTSUPP;
 717
 718         eventfd_signal(p->eventfd, 1);
 719         return 0;
 720 }
 721
 722 /*
 723  * This function is called as KVM is completely shutting down.  We do not
 724  * need to worry about locking just nuke anything we have as quickly as possible
 725  */
 726 static void
 727 ioeventfd_destructor(struct kvm_io_device *this)
 728 {
 729         struct _ioeventfd *p = to_ioeventfd(this);
 730
 731         ioeventfd_release(p);
 732 }
 733
 734 static const struct kvm_io_device_ops ioeventfd_ops = {
 735         .write      = ioeventfd_write,
 736         .destructor = ioeventfd_destructor,
 737 };
 738
 739 /* assumes kvm->slots_lock held */
 740 static bool
 741 ioeventfd_check_collision(struct kvm *kvm, struct _ioeventfd *p)
 742 {
 743         struct _ioeventfd *_p;
 744
 745         list_for_each_entry(_p, &kvm->ioeventfds, list)
 746                 if (_p->bus_idx == p->bus_idx &&
 747                     _p->addr == p->addr &&
 748                     (!_p->length || !p->length ||
 749                      (_p->length == p->length &&
 750                       (_p->wildcard || p->wildcard ||
 751                        _p->datamatch == p->datamatch))))
 752                         return true;
 753
 754         return false;
 755 }
 756
 757 static enum kvm_bus ioeventfd_bus_from_flags(__u32 flags)
 758 {
 759         if (flags & KVM_IOEVENTFD_FLAG_PIO)
 760                 return KVM_PIO_BUS;
 761         if (flags & KVM_IOEVENTFD_FLAG_VIRTIO_CCW_NOTIFY)
 762                 return KVM_VIRTIO_CCW_NOTIFY_BUS;
 763         return KVM_MMIO_BUS;
 764 }
 765
 766 static int kvm_assign_ioeventfd_idx(struct kvm *kvm,
 767                                 enum kvm_bus bus_idx,
 768                                 struct kvm_ioeventfd *args)
 769 {
 770
 771         struct eventfd_ctx *eventfd;
 772         struct _ioeventfd *p;
 773         int ret;
 774
 775         eventfd = eventfd_ctx_fdget(args->fd);
 776         if (IS_ERR(eventfd))
 777                 return PTR_ERR(eventfd);
 778
 779         p = kzalloc(sizeof(*p), GFP_KERNEL);
 780         if (!p) {
 781                 ret = -ENOMEM;
 782                 goto fail;
 783         }
 784
 785         INIT_LIST_HEAD(&p->list);
 786         p->addr    = args->addr;
 787         p->bus_idx = bus_idx;
 788         p->length  = args->len;
 789         p->eventfd = eventfd;
 790
 791         /* The datamatch feature is optional, otherwise this is a wildcard */
 792         if (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH)
 793                 p->datamatch = args->datamatch;
 794         else
 795                 p->wildcard = true;
 796
 797         mutex_lock(&kvm->slots_lock);
 798
 799         /* Verify that there isn't a match already */
 800         if (ioeventfd_check_collision(kvm, p)) {
 801                 ret = -EEXIST;
 802                 goto unlock_fail;
 803         }
 804
 805         kvm_iodevice_init(&p->dev, &ioeventfd_ops);
 806
 807         ret = kvm_io_bus_register_dev(kvm, bus_idx, p->addr, p->length,
 808                                       &p->dev);
 809         if (ret < 0)
 810                 goto unlock_fail;
 811
 812         kvm->buses[bus_idx]->ioeventfd_count++;
 813         list_add_tail(&p->list, &kvm->ioeventfds);
 814
 815         mutex_unlock(&kvm->slots_lock);
 816
 817         return 0;
 818
 819 unlock_fail:
 820         mutex_unlock(&kvm->slots_lock);
 821
 822 fail:
 823         kfree(p);
 824         eventfd_ctx_put(eventfd);
 825
 826         return ret;
 827 }
 828
 829 static int
 830 kvm_deassign_ioeventfd_idx(struct kvm *kvm, enum kvm_bus bus_idx,
 831                            struct kvm_ioeventfd *args)
 832 {
 833         struct _ioeventfd        *p, *tmp;
 834         struct eventfd_ctx       *eventfd;
 835         int                       ret = -ENOENT;
 836
 837         eventfd = eventfd_ctx_fdget(args->fd);
 838         if (IS_ERR(eventfd))
 839                 return PTR_ERR(eventfd);
 840
 841         mutex_lock(&kvm->slots_lock);
 842
 843         list_for_each_entry_safe(p, tmp, &kvm->ioeventfds, list) {
 844                 bool wildcard = !(args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH);
 845
 846                 if (p->bus_idx != bus_idx ||
 847                     p->eventfd != eventfd  ||
 848                     p->addr != args->addr  ||
 849                     p->length != args->len ||
 850                     p->wildcard != wildcard)
 851                         continue;
 852
 853                 if (!p->wildcard && p->datamatch != args->datamatch)
 854                         continue;
 855
 856                 kvm_io_bus_unregister_dev(kvm, bus_idx, &p->dev);
 857                 kvm->buses[bus_idx]->ioeventfd_count--;
 858                 ioeventfd_release(p);
 859                 ret = 0;
 860                 break;
 861         }
 862
 863         mutex_unlock(&kvm->slots_lock);
 864
 865         eventfd_ctx_put(eventfd);
 866
 867         return ret;
 868 }
 869
 870 static int kvm_deassign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 871 {
 872         enum kvm_bus bus_idx = ioeventfd_bus_from_flags(args->flags);
 873         int ret = kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
 874
 875         if (!args->len && bus_idx == KVM_MMIO_BUS)
 876                 kvm_deassign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
 877
 878         return ret;
 879 }
 880
 881 static int
 882 kvm_assign_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 883 {
 884         enum kvm_bus              bus_idx;
 885         int ret;
 886
 887         bus_idx = ioeventfd_bus_from_flags(args->flags);
 888         /* must be natural-word sized, or 0 to ignore length */
 889         switch (args->len) {
 890         case 0:
 891         case 1:
 892         case 2:
 893         case 4:
 894         case 8:
 895                 break;
 896         default:
 897                 return -EINVAL;
 898         }
 899
 900         /* check for range overflow */
 901         if (args->addr + args->len < args->addr)
 902                 return -EINVAL;
 903
 904         /* check for extra flags that we don't understand */
 905         if (args->flags & ~KVM_IOEVENTFD_VALID_FLAG_MASK)
 906                 return -EINVAL;
 907
 908         /* ioeventfd with no length can't be combined with DATAMATCH */
 909         if (!args->len && (args->flags & KVM_IOEVENTFD_FLAG_DATAMATCH))
 910                 return -EINVAL;
 911
 912         ret = kvm_assign_ioeventfd_idx(kvm, bus_idx, args);
 913         if (ret)
 914                 goto fail;
 915
 916         /* When length is ignored, MMIO is also put on a separate bus, for
 917          * faster lookups.
 918          */
 919         if (!args->len && bus_idx == KVM_MMIO_BUS) {
 920                 ret = kvm_assign_ioeventfd_idx(kvm, KVM_FAST_MMIO_BUS, args);
 921                 if (ret < 0)
 922                         goto fast_fail;
 923         }
 924
 925         return 0;
 926
 927 fast_fail:
 928         kvm_deassign_ioeventfd_idx(kvm, bus_idx, args);
 929 fail:
 930         return ret;
 931 }
 932
 933 int
 934 kvm_ioeventfd(struct kvm *kvm, struct kvm_ioeventfd *args)
 935 {
 936         if (args->flags & KVM_IOEVENTFD_FLAG_DEASSIGN)
 937                 return kvm_deassign_ioeventfd(kvm, args);
 938
 939         return kvm_assign_ioeventfd(kvm, args);
 940 }