kernel/arch/x86/kvm/assigned-dev.c

   1 /*
   2  * Kernel-based Virtual Machine - device assignment support
   3  *
   4  * Copyright (C) 2010 Red Hat, Inc. and/or its affiliates.
   5  *
   6  * This work is licensed under the terms of the GNU GPL, version 2.  See
   7  * the COPYING file in the top-level directory.
   8  *
   9  */
  10
  11 #include <linux/kvm_host.h>
  12 #include <linux/kvm.h>
  13 #include <linux/uaccess.h>
  14 #include <linux/vmalloc.h>
  15 #include <linux/errno.h>
  16 #include <linux/spinlock.h>
  17 #include <linux/pci.h>
  18 #include <linux/interrupt.h>
  19 #include <linux/slab.h>
  20 #include <linux/namei.h>
  21 #include <linux/fs.h>
  22 #include "irq.h"
  23 #include "assigned-dev.h"
  24
  25 struct kvm_assigned_dev_kernel {
  26         struct kvm_irq_ack_notifier ack_notifier;
  27         struct list_head list;
  28         int assigned_dev_id;
  29         int host_segnr;
  30         int host_busnr;
  31         int host_devfn;
  32         unsigned int entries_nr;
  33         int host_irq;
  34         bool host_irq_disabled;
  35         bool pci_2_3;
  36         struct msix_entry *host_msix_entries;
  37         int guest_irq;
  38         struct msix_entry *guest_msix_entries;
  39         unsigned long irq_requested_type;
  40         int irq_source_id;
  41         int flags;
  42         struct pci_dev *dev;
  43         struct kvm *kvm;
  44         spinlock_t intx_lock;
  45         spinlock_t intx_mask_lock;
  46         char irq_name[32];
  47         struct pci_saved_state *pci_saved_state;
  48 };
  49
  50 static struct kvm_assigned_dev_kernel *kvm_find_assigned_dev(struct list_head *head,
  51                                                       int assigned_dev_id)
  52 {
  53         struct list_head *ptr;
  54         struct kvm_assigned_dev_kernel *match;
  55
  56         list_for_each(ptr, head) {
  57                 match = list_entry(ptr, struct kvm_assigned_dev_kernel, list);
  58                 if (match->assigned_dev_id == assigned_dev_id)
  59                         return match;
  60         }
  61         return NULL;
  62 }
  63
  64 static int find_index_from_host_irq(struct kvm_assigned_dev_kernel
  65                                     *assigned_dev, int irq)
  66 {
  67         int i, index;
  68         struct msix_entry *host_msix_entries;
  69
  70         host_msix_entries = assigned_dev->host_msix_entries;
  71
  72         index = -1;
  73         for (i = 0; i < assigned_dev->entries_nr; i++)
  74                 if (irq == host_msix_entries[i].vector) {
  75                         index = i;
  76                         break;
  77                 }
  78         if (index < 0)
  79                 printk(KERN_WARNING "Fail to find correlated MSI-X entry!\n");
  80
  81         return index;
  82 }
  83
  84 static irqreturn_t kvm_assigned_dev_intx(int irq, void *dev_id)
  85 {
  86         struct kvm_assigned_dev_kernel *assigned_dev = dev_id;
  87         int ret;
  88
  89         spin_lock(&assigned_dev->intx_lock);
  90         if (pci_check_and_mask_intx(assigned_dev->dev)) {
  91                 assigned_dev->host_irq_disabled = true;
  92                 ret = IRQ_WAKE_THREAD;
  93         } else
  94                 ret = IRQ_NONE;
  95         spin_unlock(&assigned_dev->intx_lock);
  96
  97         return ret;
  98 }
  99
 100 static void
 101 kvm_assigned_dev_raise_guest_irq(struct kvm_assigned_dev_kernel *assigned_dev,
 102                                  int vector)
 103 {
 104         if (unlikely(assigned_dev->irq_requested_type &
 105                      KVM_DEV_IRQ_GUEST_INTX)) {
 106                 spin_lock(&assigned_dev->intx_mask_lock);
 107                 if (!(assigned_dev->flags & KVM_DEV_ASSIGN_MASK_INTX))
 108                         kvm_set_irq(assigned_dev->kvm,
 109                                     assigned_dev->irq_source_id, vector, 1,
 110                                     false);
 111                 spin_unlock(&assigned_dev->intx_mask_lock);
 112         } else
 113                 kvm_set_irq(assigned_dev->kvm, assigned_dev->irq_source_id,
 114                             vector, 1, false);
 115 }
 116
 117 static irqreturn_t kvm_assigned_dev_thread_intx(int irq, void *dev_id)
 118 {
 119         struct kvm_assigned_dev_kernel *assigned_dev = dev_id;
 120
 121         if (!(assigned_dev->flags & KVM_DEV_ASSIGN_PCI_2_3)) {
 122                 spin_lock_irq(&assigned_dev->intx_lock);
 123                 disable_irq_nosync(irq);
 124                 assigned_dev->host_irq_disabled = true;
 125                 spin_unlock_irq(&assigned_dev->intx_lock);
 126         }
 127
 128         kvm_assigned_dev_raise_guest_irq(assigned_dev,
 129                                          assigned_dev->guest_irq);
 130
 131         return IRQ_HANDLED;
 132 }
 133
 134 #ifdef __KVM_HAVE_MSI
 135 static irqreturn_t kvm_assigned_dev_msi(int irq, void *dev_id)
 136 {
 137         struct kvm_assigned_dev_kernel *assigned_dev = dev_id;
 138         int ret = kvm_set_irq_inatomic(assigned_dev->kvm,
 139                                        assigned_dev->irq_source_id,
 140                                        assigned_dev->guest_irq, 1);
 141         return unlikely(ret == -EWOULDBLOCK) ? IRQ_WAKE_THREAD : IRQ_HANDLED;
 142 }
 143
 144 static irqreturn_t kvm_assigned_dev_thread_msi(int irq, void *dev_id)
 145 {
 146         struct kvm_assigned_dev_kernel *assigned_dev = dev_id;
 147
 148         kvm_assigned_dev_raise_guest_irq(assigned_dev,
 149                                          assigned_dev->guest_irq);
 150
 151         return IRQ_HANDLED;
 152 }
 153 #endif
 154
 155 #ifdef __KVM_HAVE_MSIX
 156 static irqreturn_t kvm_assigned_dev_msix(int irq, void *dev_id)
 157 {
 158         struct kvm_assigned_dev_kernel *assigned_dev = dev_id;
 159         int index = find_index_from_host_irq(assigned_dev, irq);
 160         u32 vector;
 161         int ret = 0;
 162
 163         if (index >= 0) {
 164                 vector = assigned_dev->guest_msix_entries[index].vector;
 165                 ret = kvm_set_irq_inatomic(assigned_dev->kvm,
 166                                            assigned_dev->irq_source_id,
 167                                            vector, 1);
 168         }
 169
 170         return unlikely(ret == -EWOULDBLOCK) ? IRQ_WAKE_THREAD : IRQ_HANDLED;
 171 }
 172
 173 static irqreturn_t kvm_assigned_dev_thread_msix(int irq, void *dev_id)
 174 {
 175         struct kvm_assigned_dev_kernel *assigned_dev = dev_id;
 176         int index = find_index_from_host_irq(assigned_dev, irq);
 177         u32 vector;
 178
 179         if (index >= 0) {
 180                 vector = assigned_dev->guest_msix_entries[index].vector;
 181                 kvm_assigned_dev_raise_guest_irq(assigned_dev, vector);
 182         }
 183
 184         return IRQ_HANDLED;
 185 }
 186 #endif
 187
 188 /* Ack the irq line for an assigned device */
 189 static void kvm_assigned_dev_ack_irq(struct kvm_irq_ack_notifier *kian)
 190 {
 191         struct kvm_assigned_dev_kernel *dev =
 192                 container_of(kian, struct kvm_assigned_dev_kernel,
 193                              ack_notifier);
 194
 195         kvm_set_irq(dev->kvm, dev->irq_source_id, dev->guest_irq, 0, false);
 196
 197         spin_lock(&dev->intx_mask_lock);
 198
 199         if (!(dev->flags & KVM_DEV_ASSIGN_MASK_INTX)) {
 200                 bool reassert = false;
 201
 202                 spin_lock_irq(&dev->intx_lock);
 203                 /*
 204                  * The guest IRQ may be shared so this ack can come from an
 205                  * IRQ for another guest device.
 206                  */
 207                 if (dev->host_irq_disabled) {
 208                         if (!(dev->flags & KVM_DEV_ASSIGN_PCI_2_3))
 209                                 enable_irq(dev->host_irq);
 210                         else if (!pci_check_and_unmask_intx(dev->dev))
 211                                 reassert = true;
 212                         dev->host_irq_disabled = reassert;
 213                 }
 214                 spin_unlock_irq(&dev->intx_lock);
 215
 216                 if (reassert)
 217                         kvm_set_irq(dev->kvm, dev->irq_source_id,
 218                                     dev->guest_irq, 1, false);
 219         }
 220
 221         spin_unlock(&dev->intx_mask_lock);
 222 }
 223
 224 static void deassign_guest_irq(struct kvm *kvm,
 225                                struct kvm_assigned_dev_kernel *assigned_dev)
 226 {
 227         if (assigned_dev->ack_notifier.gsi != -1)
 228                 kvm_unregister_irq_ack_notifier(kvm,
 229                                                 &assigned_dev->ack_notifier);
 230
 231         kvm_set_irq(assigned_dev->kvm, assigned_dev->irq_source_id,
 232                     assigned_dev->guest_irq, 0, false);
 233
 234         if (assigned_dev->irq_source_id != -1)
 235                 kvm_free_irq_source_id(kvm, assigned_dev->irq_source_id);
 236         assigned_dev->irq_source_id = -1;
 237         assigned_dev->irq_requested_type &= ~(KVM_DEV_IRQ_GUEST_MASK);
 238 }
 239
 240 /* The function implicit hold kvm->lock mutex due to cancel_work_sync() */
 241 static void deassign_host_irq(struct kvm *kvm,
 242                               struct kvm_assigned_dev_kernel *assigned_dev)
 243 {
 244         /*
 245          * We disable irq here to prevent further events.
 246          *
 247          * Notice this maybe result in nested disable if the interrupt type is
 248          * INTx, but it's OK for we are going to free it.
 249          *
 250          * If this function is a part of VM destroy, please ensure that till
 251          * now, the kvm state is still legal for probably we also have to wait
 252          * on a currently running IRQ handler.
 253          */
 254         if (assigned_dev->irq_requested_type & KVM_DEV_IRQ_HOST_MSIX) {
 255                 int i;
 256                 for (i = 0; i < assigned_dev->entries_nr; i++)
 257                         disable_irq(assigned_dev->host_msix_entries[i].vector);
 258
 259                 for (i = 0; i < assigned_dev->entries_nr; i++)
 260                         free_irq(assigned_dev->host_msix_entries[i].vector,
 261                                  assigned_dev);
 262
 263                 assigned_dev->entries_nr = 0;
 264                 kfree(assigned_dev->host_msix_entries);
 265                 kfree(assigned_dev->guest_msix_entries);
 266                 pci_disable_msix(assigned_dev->dev);
 267         } else {
 268                 /* Deal with MSI and INTx */
 269                 if ((assigned_dev->irq_requested_type &
 270                      KVM_DEV_IRQ_HOST_INTX) &&
 271                     (assigned_dev->flags & KVM_DEV_ASSIGN_PCI_2_3)) {
 272                         spin_lock_irq(&assigned_dev->intx_lock);
 273                         pci_intx(assigned_dev->dev, false);
 274                         spin_unlock_irq(&assigned_dev->intx_lock);
 275                         synchronize_irq(assigned_dev->host_irq);
 276                 } else
 277                         disable_irq(assigned_dev->host_irq);
 278
 279                 free_irq(assigned_dev->host_irq, assigned_dev);
 280
 281                 if (assigned_dev->irq_requested_type & KVM_DEV_IRQ_HOST_MSI)
 282                         pci_disable_msi(assigned_dev->dev);
 283         }
 284
 285         assigned_dev->irq_requested_type &= ~(KVM_DEV_IRQ_HOST_MASK);
 286 }
 287
 288 static int kvm_deassign_irq(struct kvm *kvm,
 289                             struct kvm_assigned_dev_kernel *assigned_dev,
 290                             unsigned long irq_requested_type)
 291 {
 292         unsigned long guest_irq_type, host_irq_type;
 293
 294         if (!irqchip_in_kernel(kvm))
 295                 return -EINVAL;
 296         /* no irq assignment to deassign */
 297         if (!assigned_dev->irq_requested_type)
 298                 return -ENXIO;
 299
 300         host_irq_type = irq_requested_type & KVM_DEV_IRQ_HOST_MASK;
 301         guest_irq_type = irq_requested_type & KVM_DEV_IRQ_GUEST_MASK;
 302
 303         if (host_irq_type)
 304                 deassign_host_irq(kvm, assigned_dev);
 305         if (guest_irq_type)
 306                 deassign_guest_irq(kvm, assigned_dev);
 307
 308         return 0;
 309 }
 310
 311 static void kvm_free_assigned_irq(struct kvm *kvm,
 312                                   struct kvm_assigned_dev_kernel *assigned_dev)
 313 {
 314         kvm_deassign_irq(kvm, assigned_dev, assigned_dev->irq_requested_type);
 315 }
 316
 317 static void kvm_free_assigned_device(struct kvm *kvm,
 318                                      struct kvm_assigned_dev_kernel
 319                                      *assigned_dev)
 320 {
 321         kvm_free_assigned_irq(kvm, assigned_dev);
 322
 323         pci_reset_function(assigned_dev->dev);
 324         if (pci_load_and_free_saved_state(assigned_dev->dev,
 325                                           &assigned_dev->pci_saved_state))
 326                 printk(KERN_INFO "%s: Couldn't reload %s saved state\n",
 327                        __func__, dev_name(&assigned_dev->dev->dev));
 328         else
 329                 pci_restore_state(assigned_dev->dev);
 330
 331         pci_clear_dev_assigned(assigned_dev->dev);
 332
 333         pci_release_regions(assigned_dev->dev);
 334         pci_disable_device(assigned_dev->dev);
 335         pci_dev_put(assigned_dev->dev);
 336
 337         list_del(&assigned_dev->list);
 338         kfree(assigned_dev);
 339 }
 340
 341 void kvm_free_all_assigned_devices(struct kvm *kvm)
 342 {
 343         struct list_head *ptr, *ptr2;
 344         struct kvm_assigned_dev_kernel *assigned_dev;
 345
 346         list_for_each_safe(ptr, ptr2, &kvm->arch.assigned_dev_head) {
 347                 assigned_dev = list_entry(ptr,
 348                                           struct kvm_assigned_dev_kernel,
 349                                           list);
 350
 351                 kvm_free_assigned_device(kvm, assigned_dev);
 352         }
 353 }
 354
 355 static int assigned_device_enable_host_intx(struct kvm *kvm,
 356                                             struct kvm_assigned_dev_kernel *dev)
 357 {
 358         irq_handler_t irq_handler;
 359         unsigned long flags;
 360
 361         dev->host_irq = dev->dev->irq;
 362
 363         /*
 364          * We can only share the IRQ line with other host devices if we are
 365          * able to disable the IRQ source at device-level - independently of
 366          * the guest driver. Otherwise host devices may suffer from unbounded
 367          * IRQ latencies when the guest keeps the line asserted.
 368          */
 369         if (dev->flags & KVM_DEV_ASSIGN_PCI_2_3) {
 370                 irq_handler = kvm_assigned_dev_intx;
 371                 flags = IRQF_SHARED;
 372         } else {
 373                 irq_handler = NULL;
 374                 flags = IRQF_ONESHOT;
 375         }
 376         if (request_threaded_irq(dev->host_irq, irq_handler,
 377                                  kvm_assigned_dev_thread_intx, flags,
 378                                  dev->irq_name, dev))
 379                 return -EIO;
 380
 381         if (dev->flags & KVM_DEV_ASSIGN_PCI_2_3) {
 382                 spin_lock_irq(&dev->intx_lock);
 383                 pci_intx(dev->dev, true);
 384                 spin_unlock_irq(&dev->intx_lock);
 385         }
 386         return 0;
 387 }
 388
 389 #ifdef __KVM_HAVE_MSI
 390 static int assigned_device_enable_host_msi(struct kvm *kvm,
 391                                            struct kvm_assigned_dev_kernel *dev)
 392 {
 393         int r;
 394
 395         if (!dev->dev->msi_enabled) {
 396                 r = pci_enable_msi(dev->dev);
 397                 if (r)
 398                         return r;
 399         }
 400
 401         dev->host_irq = dev->dev->irq;
 402         if (request_threaded_irq(dev->host_irq, kvm_assigned_dev_msi,
 403                                  kvm_assigned_dev_thread_msi, 0,
 404                                  dev->irq_name, dev)) {
 405                 pci_disable_msi(dev->dev);
 406                 return -EIO;
 407         }
 408
 409         return 0;
 410 }
 411 #endif
 412
 413 #ifdef __KVM_HAVE_MSIX
 414 static int assigned_device_enable_host_msix(struct kvm *kvm,
 415                                             struct kvm_assigned_dev_kernel *dev)
 416 {
 417         int i, r = -EINVAL;
 418
 419         /* host_msix_entries and guest_msix_entries should have been
 420          * initialized */
 421         if (dev->entries_nr == 0)
 422                 return r;
 423
 424         r = pci_enable_msix_exact(dev->dev,
 425                                   dev->host_msix_entries, dev->entries_nr);
 426         if (r)
 427                 return r;
 428
 429         for (i = 0; i < dev->entries_nr; i++) {
 430                 r = request_threaded_irq(dev->host_msix_entries[i].vector,
 431                                          kvm_assigned_dev_msix,
 432                                          kvm_assigned_dev_thread_msix,
 433                                          0, dev->irq_name, dev);
 434                 if (r)
 435                         goto err;
 436         }
 437
 438         return 0;
 439 err:
 440         for (i -= 1; i >= 0; i--)
 441                 free_irq(dev->host_msix_entries[i].vector, dev);
 442         pci_disable_msix(dev->dev);
 443         return r;
 444 }
 445
 446 #endif
 447
 448 static int assigned_device_enable_guest_intx(struct kvm *kvm,
 449                                 struct kvm_assigned_dev_kernel *dev,
 450                                 struct kvm_assigned_irq *irq)
 451 {
 452         dev->guest_irq = irq->guest_irq;
 453         dev->ack_notifier.gsi = irq->guest_irq;
 454         return 0;
 455 }
 456
 457 #ifdef __KVM_HAVE_MSI
 458 static int assigned_device_enable_guest_msi(struct kvm *kvm,
 459                         struct kvm_assigned_dev_kernel *dev,
 460                         struct kvm_assigned_irq *irq)
 461 {
 462         dev->guest_irq = irq->guest_irq;
 463         dev->ack_notifier.gsi = -1;
 464         return 0;
 465 }
 466 #endif
 467
 468 #ifdef __KVM_HAVE_MSIX
 469 static int assigned_device_enable_guest_msix(struct kvm *kvm,
 470                         struct kvm_assigned_dev_kernel *dev,
 471                         struct kvm_assigned_irq *irq)
 472 {
 473         dev->guest_irq = irq->guest_irq;
 474         dev->ack_notifier.gsi = -1;
 475         return 0;
 476 }
 477 #endif
 478
 479 static int assign_host_irq(struct kvm *kvm,
 480                            struct kvm_assigned_dev_kernel *dev,
 481                            __u32 host_irq_type)
 482 {
 483         int r = -EEXIST;
 484
 485         if (dev->irq_requested_type & KVM_DEV_IRQ_HOST_MASK)
 486                 return r;
 487
 488         snprintf(dev->irq_name, sizeof(dev->irq_name), "kvm:%s",
 489                  pci_name(dev->dev));
 490
 491         switch (host_irq_type) {
 492         case KVM_DEV_IRQ_HOST_INTX:
 493                 r = assigned_device_enable_host_intx(kvm, dev);
 494                 break;
 495 #ifdef __KVM_HAVE_MSI
 496         case KVM_DEV_IRQ_HOST_MSI:
 497                 r = assigned_device_enable_host_msi(kvm, dev);
 498                 break;
 499 #endif
 500 #ifdef __KVM_HAVE_MSIX
 501         case KVM_DEV_IRQ_HOST_MSIX:
 502                 r = assigned_device_enable_host_msix(kvm, dev);
 503                 break;
 504 #endif
 505         default:
 506                 r = -EINVAL;
 507         }
 508         dev->host_irq_disabled = false;
 509
 510         if (!r)
 511                 dev->irq_requested_type |= host_irq_type;
 512
 513         return r;
 514 }
 515
 516 static int assign_guest_irq(struct kvm *kvm,
 517                             struct kvm_assigned_dev_kernel *dev,
 518                             struct kvm_assigned_irq *irq,
 519                             unsigned long guest_irq_type)
 520 {
 521         int id;
 522         int r = -EEXIST;
 523
 524         if (dev->irq_requested_type & KVM_DEV_IRQ_GUEST_MASK)
 525                 return r;
 526
 527         id = kvm_request_irq_source_id(kvm);
 528         if (id < 0)
 529                 return id;
 530
 531         dev->irq_source_id = id;
 532
 533         switch (guest_irq_type) {
 534         case KVM_DEV_IRQ_GUEST_INTX:
 535                 r = assigned_device_enable_guest_intx(kvm, dev, irq);
 536                 break;
 537 #ifdef __KVM_HAVE_MSI
 538         case KVM_DEV_IRQ_GUEST_MSI:
 539                 r = assigned_device_enable_guest_msi(kvm, dev, irq);
 540                 break;
 541 #endif
 542 #ifdef __KVM_HAVE_MSIX
 543         case KVM_DEV_IRQ_GUEST_MSIX:
 544                 r = assigned_device_enable_guest_msix(kvm, dev, irq);
 545                 break;
 546 #endif
 547         default:
 548                 r = -EINVAL;
 549         }
 550
 551         if (!r) {
 552                 dev->irq_requested_type |= guest_irq_type;
 553                 if (dev->ack_notifier.gsi != -1)
 554                         kvm_register_irq_ack_notifier(kvm, &dev->ack_notifier);
 555         } else {
 556                 kvm_free_irq_source_id(kvm, dev->irq_source_id);
 557                 dev->irq_source_id = -1;
 558         }
 559
 560         return r;
 561 }
 562
 563 /* TODO Deal with KVM_DEV_IRQ_ASSIGNED_MASK_MSIX */
 564 static int kvm_vm_ioctl_assign_irq(struct kvm *kvm,
 565                                    struct kvm_assigned_irq *assigned_irq)
 566 {
 567         int r = -EINVAL;
 568         struct kvm_assigned_dev_kernel *match;
 569         unsigned long host_irq_type, guest_irq_type;
 570
 571         if (!irqchip_in_kernel(kvm))
 572                 return r;
 573
 574         mutex_lock(&kvm->lock);
 575         r = -ENODEV;
 576         match = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
 577                                       assigned_irq->assigned_dev_id);
 578         if (!match)
 579                 goto out;
 580
 581         host_irq_type = (assigned_irq->flags & KVM_DEV_IRQ_HOST_MASK);
 582         guest_irq_type = (assigned_irq->flags & KVM_DEV_IRQ_GUEST_MASK);
 583
 584         r = -EINVAL;
 585         /* can only assign one type at a time */
 586         if (hweight_long(host_irq_type) > 1)
 587                 goto out;
 588         if (hweight_long(guest_irq_type) > 1)
 589                 goto out;
 590         if (host_irq_type == 0 && guest_irq_type == 0)
 591                 goto out;
 592
 593         r = 0;
 594         if (host_irq_type)
 595                 r = assign_host_irq(kvm, match, host_irq_type);
 596         if (r)
 597                 goto out;
 598
 599         if (guest_irq_type)
 600                 r = assign_guest_irq(kvm, match, assigned_irq, guest_irq_type);
 601 out:
 602         mutex_unlock(&kvm->lock);
 603         return r;
 604 }
 605
 606 static int kvm_vm_ioctl_deassign_dev_irq(struct kvm *kvm,
 607                                          struct kvm_assigned_irq
 608                                          *assigned_irq)
 609 {
 610         int r = -ENODEV;
 611         struct kvm_assigned_dev_kernel *match;
 612         unsigned long irq_type;
 613
 614         mutex_lock(&kvm->lock);
 615
 616         match = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
 617                                       assigned_irq->assigned_dev_id);
 618         if (!match)
 619                 goto out;
 620
 621         irq_type = assigned_irq->flags & (KVM_DEV_IRQ_HOST_MASK |
 622                                           KVM_DEV_IRQ_GUEST_MASK);
 623         r = kvm_deassign_irq(kvm, match, irq_type);
 624 out:
 625         mutex_unlock(&kvm->lock);
 626         return r;
 627 }
 628
 629 /*
 630  * We want to test whether the caller has been granted permissions to
 631  * use this device.  To be able to configure and control the device,
 632  * the user needs access to PCI configuration space and BAR resources.
 633  * These are accessed through PCI sysfs.  PCI config space is often
 634  * passed to the process calling this ioctl via file descriptor, so we
 635  * can't rely on access to that file.  We can check for permissions
 636  * on each of the BAR resource files, which is a pretty clear
 637  * indicator that the user has been granted access to the device.
 638  */
 639 static int probe_sysfs_permissions(struct pci_dev *dev)
 640 {
 641 #ifdef CONFIG_SYSFS
 642         int i;
 643         bool bar_found = false;
 644
 645         for (i = PCI_STD_RESOURCES; i <= PCI_STD_RESOURCE_END; i++) {
 646                 char *kpath, *syspath;
 647                 struct path path;
 648                 struct inode *inode;
 649                 int r;
 650
 651                 if (!pci_resource_len(dev, i))
 652                         continue;
 653
 654                 kpath = kobject_get_path(&dev->dev.kobj, GFP_KERNEL);
 655                 if (!kpath)
 656                         return -ENOMEM;
 657
 658                 /* Per sysfs-rules, sysfs is always at /sys */
 659                 syspath = kasprintf(GFP_KERNEL, "/sys%s/resource%d", kpath, i);
 660                 kfree(kpath);
 661                 if (!syspath)
 662                         return -ENOMEM;
 663
 664                 r = kern_path(syspath, LOOKUP_FOLLOW, &path);
 665                 kfree(syspath);
 666                 if (r)
 667                         return r;
 668
 669                 inode = d_backing_inode(path.dentry);
 670
 671                 r = inode_permission(inode, MAY_READ | MAY_WRITE | MAY_ACCESS);
 672                 path_put(&path);
 673                 if (r)
 674                         return r;
 675
 676                 bar_found = true;
 677         }
 678
 679         /* If no resources, probably something special */
 680         if (!bar_found)
 681                 return -EPERM;
 682
 683         return 0;
 684 #else
 685         return -EINVAL; /* No way to control the device without sysfs */
 686 #endif
 687 }
 688
 689 static int kvm_vm_ioctl_assign_device(struct kvm *kvm,
 690                                       struct kvm_assigned_pci_dev *assigned_dev)
 691 {
 692         int r = 0, idx;
 693         struct kvm_assigned_dev_kernel *match;
 694         struct pci_dev *dev;
 695
 696         if (!(assigned_dev->flags & KVM_DEV_ASSIGN_ENABLE_IOMMU))
 697                 return -EINVAL;
 698
 699         mutex_lock(&kvm->lock);
 700         idx = srcu_read_lock(&kvm->srcu);
 701
 702         match = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
 703                                       assigned_dev->assigned_dev_id);
 704         if (match) {
 705                 /* device already assigned */
 706                 r = -EEXIST;
 707                 goto out;
 708         }
 709
 710         match = kzalloc(sizeof(struct kvm_assigned_dev_kernel), GFP_KERNEL);
 711         if (match == NULL) {
 712                 printk(KERN_INFO "%s: Couldn't allocate memory\n",
 713                        __func__);
 714                 r = -ENOMEM;
 715                 goto out;
 716         }
 717         dev = pci_get_domain_bus_and_slot(assigned_dev->segnr,
 718                                    assigned_dev->busnr,
 719                                    assigned_dev->devfn);
 720         if (!dev) {
 721                 printk(KERN_INFO "%s: host device not found\n", __func__);
 722                 r = -EINVAL;
 723                 goto out_free;
 724         }
 725
 726         /* Don't allow bridges to be assigned */
 727         if (dev->hdr_type != PCI_HEADER_TYPE_NORMAL) {
 728                 r = -EPERM;
 729                 goto out_put;
 730         }
 731
 732         r = probe_sysfs_permissions(dev);
 733         if (r)
 734                 goto out_put;
 735
 736         if (pci_enable_device(dev)) {
 737                 printk(KERN_INFO "%s: Could not enable PCI device\n", __func__);
 738                 r = -EBUSY;
 739                 goto out_put;
 740         }
 741         r = pci_request_regions(dev, "kvm_assigned_device");
 742         if (r) {
 743                 printk(KERN_INFO "%s: Could not get access to device regions\n",
 744                        __func__);
 745                 goto out_disable;
 746         }
 747
 748         pci_reset_function(dev);
 749         pci_save_state(dev);
 750         match->pci_saved_state = pci_store_saved_state(dev);
 751         if (!match->pci_saved_state)
 752                 printk(KERN_DEBUG "%s: Couldn't store %s saved state\n",
 753                        __func__, dev_name(&dev->dev));
 754
 755         if (!pci_intx_mask_supported(dev))
 756                 assigned_dev->flags &= ~KVM_DEV_ASSIGN_PCI_2_3;
 757
 758         match->assigned_dev_id = assigned_dev->assigned_dev_id;
 759         match->host_segnr = assigned_dev->segnr;
 760         match->host_busnr = assigned_dev->busnr;
 761         match->host_devfn = assigned_dev->devfn;
 762         match->flags = assigned_dev->flags;
 763         match->dev = dev;
 764         spin_lock_init(&match->intx_lock);
 765         spin_lock_init(&match->intx_mask_lock);
 766         match->irq_source_id = -1;
 767         match->kvm = kvm;
 768         match->ack_notifier.irq_acked = kvm_assigned_dev_ack_irq;
 769
 770         list_add(&match->list, &kvm->arch.assigned_dev_head);
 771
 772         if (!kvm->arch.iommu_domain) {
 773                 r = kvm_iommu_map_guest(kvm);
 774                 if (r)
 775                         goto out_list_del;
 776         }
 777         r = kvm_assign_device(kvm, match->dev);
 778         if (r)
 779                 goto out_list_del;
 780
 781 out:
 782         srcu_read_unlock(&kvm->srcu, idx);
 783         mutex_unlock(&kvm->lock);
 784         return r;
 785 out_list_del:
 786         if (pci_load_and_free_saved_state(dev, &match->pci_saved_state))
 787                 printk(KERN_INFO "%s: Couldn't reload %s saved state\n",
 788                        __func__, dev_name(&dev->dev));
 789         list_del(&match->list);
 790         pci_release_regions(dev);
 791 out_disable:
 792         pci_disable_device(dev);
 793 out_put:
 794         pci_dev_put(dev);
 795 out_free:
 796         kfree(match);
 797         srcu_read_unlock(&kvm->srcu, idx);
 798         mutex_unlock(&kvm->lock);
 799         return r;
 800 }
 801
 802 static int kvm_vm_ioctl_deassign_device(struct kvm *kvm,
 803                 struct kvm_assigned_pci_dev *assigned_dev)
 804 {
 805         int r = 0;
 806         struct kvm_assigned_dev_kernel *match;
 807
 808         mutex_lock(&kvm->lock);
 809
 810         match = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
 811                                       assigned_dev->assigned_dev_id);
 812         if (!match) {
 813                 printk(KERN_INFO "%s: device hasn't been assigned before, "
 814                   "so cannot be deassigned\n", __func__);
 815                 r = -EINVAL;
 816                 goto out;
 817         }
 818
 819         kvm_deassign_device(kvm, match->dev);
 820
 821         kvm_free_assigned_device(kvm, match);
 822
 823 out:
 824         mutex_unlock(&kvm->lock);
 825         return r;
 826 }
 827
 828
 829 #ifdef __KVM_HAVE_MSIX
 830 static int kvm_vm_ioctl_set_msix_nr(struct kvm *kvm,
 831                                     struct kvm_assigned_msix_nr *entry_nr)
 832 {
 833         int r = 0;
 834         struct kvm_assigned_dev_kernel *adev;
 835
 836         mutex_lock(&kvm->lock);
 837
 838         adev = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
 839                                       entry_nr->assigned_dev_id);
 840         if (!adev) {
 841                 r = -EINVAL;
 842                 goto msix_nr_out;
 843         }
 844
 845         if (adev->entries_nr == 0) {
 846                 adev->entries_nr = entry_nr->entry_nr;
 847                 if (adev->entries_nr == 0 ||
 848                     adev->entries_nr > KVM_MAX_MSIX_PER_DEV) {
 849                         r = -EINVAL;
 850                         goto msix_nr_out;
 851                 }
 852
 853                 adev->host_msix_entries = kzalloc(sizeof(struct msix_entry) *
 854                                                 entry_nr->entry_nr,
 855                                                 GFP_KERNEL);
 856                 if (!adev->host_msix_entries) {
 857                         r = -ENOMEM;
 858                         goto msix_nr_out;
 859                 }
 860                 adev->guest_msix_entries =
 861                         kzalloc(sizeof(struct msix_entry) * entry_nr->entry_nr,
 862                                 GFP_KERNEL);
 863                 if (!adev->guest_msix_entries) {
 864                         kfree(adev->host_msix_entries);
 865                         r = -ENOMEM;
 866                         goto msix_nr_out;
 867                 }
 868         } else /* Not allowed set MSI-X number twice */
 869                 r = -EINVAL;
 870 msix_nr_out:
 871         mutex_unlock(&kvm->lock);
 872         return r;
 873 }
 874
 875 static int kvm_vm_ioctl_set_msix_entry(struct kvm *kvm,
 876                                        struct kvm_assigned_msix_entry *entry)
 877 {
 878         int r = 0, i;
 879         struct kvm_assigned_dev_kernel *adev;
 880
 881         mutex_lock(&kvm->lock);
 882
 883         adev = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
 884                                       entry->assigned_dev_id);
 885
 886         if (!adev) {
 887                 r = -EINVAL;
 888                 goto msix_entry_out;
 889         }
 890
 891         for (i = 0; i < adev->entries_nr; i++)
 892                 if (adev->guest_msix_entries[i].vector == 0 ||
 893                     adev->guest_msix_entries[i].entry == entry->entry) {
 894                         adev->guest_msix_entries[i].entry = entry->entry;
 895                         adev->guest_msix_entries[i].vector = entry->gsi;
 896                         adev->host_msix_entries[i].entry = entry->entry;
 897                         break;
 898                 }
 899         if (i == adev->entries_nr) {
 900                 r = -ENOSPC;
 901                 goto msix_entry_out;
 902         }
 903
 904 msix_entry_out:
 905         mutex_unlock(&kvm->lock);
 906
 907         return r;
 908 }
 909 #endif
 910
 911 static int kvm_vm_ioctl_set_pci_irq_mask(struct kvm *kvm,
 912                 struct kvm_assigned_pci_dev *assigned_dev)
 913 {
 914         int r = 0;
 915         struct kvm_assigned_dev_kernel *match;
 916
 917         mutex_lock(&kvm->lock);
 918
 919         match = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
 920                                       assigned_dev->assigned_dev_id);
 921         if (!match) {
 922                 r = -ENODEV;
 923                 goto out;
 924         }
 925
 926         spin_lock(&match->intx_mask_lock);
 927
 928         match->flags &= ~KVM_DEV_ASSIGN_MASK_INTX;
 929         match->flags |= assigned_dev->flags & KVM_DEV_ASSIGN_MASK_INTX;
 930
 931         if (match->irq_requested_type & KVM_DEV_IRQ_GUEST_INTX) {
 932                 if (assigned_dev->flags & KVM_DEV_ASSIGN_MASK_INTX) {
 933                         kvm_set_irq(match->kvm, match->irq_source_id,
 934                                     match->guest_irq, 0, false);
 935                         /*
 936                          * Masking at hardware-level is performed on demand,
 937                          * i.e. when an IRQ actually arrives at the host.
 938                          */
 939                 } else if (!(assigned_dev->flags & KVM_DEV_ASSIGN_PCI_2_3)) {
 940                         /*
 941                          * Unmask the IRQ line if required. Unmasking at
 942                          * device level will be performed by user space.
 943                          */
 944                         spin_lock_irq(&match->intx_lock);
 945                         if (match->host_irq_disabled) {
 946                                 enable_irq(match->host_irq);
 947                                 match->host_irq_disabled = false;
 948                         }
 949                         spin_unlock_irq(&match->intx_lock);
 950                 }
 951         }
 952
 953         spin_unlock(&match->intx_mask_lock);
 954
 955 out:
 956         mutex_unlock(&kvm->lock);
 957         return r;
 958 }
 959
 960 long kvm_vm_ioctl_assigned_device(struct kvm *kvm, unsigned ioctl,
 961                                   unsigned long arg)
 962 {
 963         void __user *argp = (void __user *)arg;
 964         int r;
 965
 966         switch (ioctl) {
 967         case KVM_ASSIGN_PCI_DEVICE: {
 968                 struct kvm_assigned_pci_dev assigned_dev;
 969
 970                 r = -EFAULT;
 971                 if (copy_from_user(&assigned_dev, argp, sizeof assigned_dev))
 972                         goto out;
 973                 r = kvm_vm_ioctl_assign_device(kvm, &assigned_dev);
 974                 if (r)
 975                         goto out;
 976                 break;
 977         }
 978         case KVM_ASSIGN_IRQ: {
 979                 r = -EOPNOTSUPP;
 980                 break;
 981         }
 982         case KVM_ASSIGN_DEV_IRQ: {
 983                 struct kvm_assigned_irq assigned_irq;
 984
 985                 r = -EFAULT;
 986                 if (copy_from_user(&assigned_irq, argp, sizeof assigned_irq))
 987                         goto out;
 988                 r = kvm_vm_ioctl_assign_irq(kvm, &assigned_irq);
 989                 if (r)
 990                         goto out;
 991                 break;
 992         }
 993         case KVM_DEASSIGN_DEV_IRQ: {
 994                 struct kvm_assigned_irq assigned_irq;
 995
 996                 r = -EFAULT;
 997                 if (copy_from_user(&assigned_irq, argp, sizeof assigned_irq))
 998                         goto out;
 999                 r = kvm_vm_ioctl_deassign_dev_irq(kvm, &assigned_irq);
1000                 if (r)
1001                         goto out;
1002                 break;
1003         }
1004         case KVM_DEASSIGN_PCI_DEVICE: {
1005                 struct kvm_assigned_pci_dev assigned_dev;
1006
1007                 r = -EFAULT;
1008                 if (copy_from_user(&assigned_dev, argp, sizeof assigned_dev))
1009                         goto out;
1010                 r = kvm_vm_ioctl_deassign_device(kvm, &assigned_dev);
1011                 if (r)
1012                         goto out;
1013                 break;
1014         }
1015 #ifdef __KVM_HAVE_MSIX
1016         case KVM_ASSIGN_SET_MSIX_NR: {
1017                 struct kvm_assigned_msix_nr entry_nr;
1018                 r = -EFAULT;
1019                 if (copy_from_user(&entry_nr, argp, sizeof entry_nr))
1020                         goto out;
1021                 r = kvm_vm_ioctl_set_msix_nr(kvm, &entry_nr);
1022                 if (r)
1023                         goto out;
1024                 break;
1025         }
1026         case KVM_ASSIGN_SET_MSIX_ENTRY: {
1027                 struct kvm_assigned_msix_entry entry;
1028                 r = -EFAULT;
1029                 if (copy_from_user(&entry, argp, sizeof entry))
1030                         goto out;
1031                 r = kvm_vm_ioctl_set_msix_entry(kvm, &entry);
1032                 if (r)
1033                         goto out;
1034                 break;
1035         }
1036 #endif
1037         case KVM_ASSIGN_SET_INTX_MASK: {
1038                 struct kvm_assigned_pci_dev assigned_dev;
1039
1040                 r = -EFAULT;
1041                 if (copy_from_user(&assigned_dev, argp, sizeof assigned_dev))
1042                         goto out;
1043                 r = kvm_vm_ioctl_set_pci_irq_mask(kvm, &assigned_dev);
1044                 break;
1045         }
1046         default:
1047                 r = -ENOTTY;
1048                 break;
1049         }
1050 out:
1051         return r;
1052 }