drivers/vfio/group.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * VFIO core
   4  *
   5  * Copyright (C) 2012 Red Hat, Inc.  All rights reserved.
   6  *     Author: Alex Williamson <alex.williamson@redhat.com>
   7  *
   8  * Derived from original vfio:
   9  * Copyright 2010 Cisco Systems, Inc.  All rights reserved.
  10  * Author: Tom Lyon, pugs@cisco.com
  11  */
  12
  13 #include <linux/vfio.h>
  14 #include <linux/iommufd.h>
  15 #include <linux/anon_inodes.h>
  16 #include "vfio.h"
  17
  18 static struct vfio {
  19         struct class                    *class;
  20         struct list_head                group_list;
  21         struct mutex                    group_lock; /* locks group_list */
  22         struct ida                      group_ida;
  23         dev_t                           group_devt;
  24 } vfio;
  25
  26 static struct vfio_device *vfio_device_get_from_name(struct vfio_group *group,
  27                                                      char *buf)
  28 {
  29         struct vfio_device *it, *device = ERR_PTR(-ENODEV);
  30
  31         mutex_lock(&group->device_lock);
  32         list_for_each_entry(it, &group->device_list, group_next) {
  33                 int ret;
  34
  35                 if (it->ops->match) {
  36                         ret = it->ops->match(it, buf);
  37                         if (ret < 0) {
  38                                 device = ERR_PTR(ret);
  39                                 break;
  40                         }
  41                 } else {
  42                         ret = !strcmp(dev_name(it->dev), buf);
  43                 }
  44
  45                 if (ret && vfio_device_try_get_registration(it)) {
  46                         device = it;
  47                         break;
  48                 }
  49         }
  50         mutex_unlock(&group->device_lock);
  51
  52         return device;
  53 }
  54
  55 /*
  56  * VFIO Group fd, /dev/vfio/$GROUP
  57  */
  58 static bool vfio_group_has_iommu(struct vfio_group *group)
  59 {
  60         lockdep_assert_held(&group->group_lock);
  61         /*
  62          * There can only be users if there is a container, and if there is a
  63          * container there must be users.
  64          */
  65         WARN_ON(!group->container != !group->container_users);
  66
  67         return group->container || group->iommufd;
  68 }
  69
  70 /*
  71  * VFIO_GROUP_UNSET_CONTAINER should fail if there are other users or
  72  * if there was no container to unset.  Since the ioctl is called on
  73  * the group, we know that still exists, therefore the only valid
  74  * transition here is 1->0.
  75  */
  76 static int vfio_group_ioctl_unset_container(struct vfio_group *group)
  77 {
  78         int ret = 0;
  79
  80         mutex_lock(&group->group_lock);
  81         if (!vfio_group_has_iommu(group)) {
  82                 ret = -EINVAL;
  83                 goto out_unlock;
  84         }
  85         if (group->container) {
  86                 if (group->container_users != 1) {
  87                         ret = -EBUSY;
  88                         goto out_unlock;
  89                 }
  90                 vfio_group_detach_container(group);
  91         }
  92         if (group->iommufd) {
  93                 iommufd_ctx_put(group->iommufd);
  94                 group->iommufd = NULL;
  95         }
  96
  97 out_unlock:
  98         mutex_unlock(&group->group_lock);
  99         return ret;
 100 }
 101
 102 static int vfio_group_ioctl_set_container(struct vfio_group *group,
 103                                           int __user *arg)
 104 {
 105         struct vfio_container *container;
 106         struct iommufd_ctx *iommufd;
 107         int ret;
 108         int fd;
 109
 110         if (get_user(fd, arg))
 111                 return -EFAULT;
 112
 113         CLASS(fd, f)(fd);
 114         if (fd_empty(f))
 115                 return -EBADF;
 116
 117         mutex_lock(&group->group_lock);
 118         if (vfio_group_has_iommu(group)) {
 119                 ret = -EINVAL;
 120                 goto out_unlock;
 121         }
 122         if (!group->iommu_group) {
 123                 ret = -ENODEV;
 124                 goto out_unlock;
 125         }
 126
 127         container = vfio_container_from_file(fd_file(f));
 128         if (container) {
 129                 ret = vfio_container_attach_group(container, group);
 130                 goto out_unlock;
 131         }
 132
 133         iommufd = iommufd_ctx_from_file(fd_file(f));
 134         if (!IS_ERR(iommufd)) {
 135                 if (IS_ENABLED(CONFIG_VFIO_NOIOMMU) &&
 136                     group->type == VFIO_NO_IOMMU)
 137                         ret = iommufd_vfio_compat_set_no_iommu(iommufd);
 138                 else
 139                         ret = iommufd_vfio_compat_ioas_create(iommufd);
 140
 141                 if (ret) {
 142                         iommufd_ctx_put(iommufd);
 143                         goto out_unlock;
 144                 }
 145
 146                 group->iommufd = iommufd;
 147                 goto out_unlock;
 148         }
 149
 150         /* The FD passed is not recognized. */
 151         ret = -EBADFD;
 152
 153 out_unlock:
 154         mutex_unlock(&group->group_lock);
 155         return ret;
 156 }
 157
 158 static void vfio_device_group_get_kvm_safe(struct vfio_device *device)
 159 {
 160         spin_lock(&device->group->kvm_ref_lock);
 161         vfio_device_get_kvm_safe(device, device->group->kvm);
 162         spin_unlock(&device->group->kvm_ref_lock);
 163 }
 164
 165 static int vfio_df_group_open(struct vfio_device_file *df)
 166 {
 167         struct vfio_device *device = df->device;
 168         int ret;
 169
 170         mutex_lock(&device->group->group_lock);
 171         if (!vfio_group_has_iommu(device->group)) {
 172                 ret = -EINVAL;
 173                 goto out_unlock;
 174         }
 175
 176         mutex_lock(&device->dev_set->lock);
 177
 178         /*
 179          * Before the first device open, get the KVM pointer currently
 180          * associated with the group (if there is one) and obtain a reference
 181          * now that will be held until the open_count reaches 0 again.  Save
 182          * the pointer in the device for use by drivers.
 183          */
 184         if (device->open_count == 0)
 185                 vfio_device_group_get_kvm_safe(device);
 186
 187         df->iommufd = device->group->iommufd;
 188         if (df->iommufd && vfio_device_is_noiommu(device) && device->open_count == 0) {
 189                 /*
 190                  * Require no compat ioas to be assigned to proceed.  The basic
 191                  * statement is that the user cannot have done something that
 192                  * implies they expected translation to exist
 193                  */
 194                 if (!capable(CAP_SYS_RAWIO) ||
 195                     vfio_iommufd_device_has_compat_ioas(device, df->iommufd))
 196                         ret = -EPERM;
 197                 else
 198                         ret = 0;
 199                 goto out_put_kvm;
 200         }
 201
 202         ret = vfio_df_open(df);
 203         if (ret)
 204                 goto out_put_kvm;
 205
 206         if (df->iommufd && device->open_count == 1) {
 207                 ret = vfio_iommufd_compat_attach_ioas(device, df->iommufd);
 208                 if (ret)
 209                         goto out_close_device;
 210         }
 211
 212         /*
 213          * Paired with smp_load_acquire() in vfio_device_fops::ioctl/
 214          * read/write/mmap and vfio_file_has_device_access()
 215          */
 216         smp_store_release(&df->access_granted, true);
 217
 218         mutex_unlock(&device->dev_set->lock);
 219         mutex_unlock(&device->group->group_lock);
 220         return 0;
 221
 222 out_close_device:
 223         vfio_df_close(df);
 224 out_put_kvm:
 225         df->iommufd = NULL;
 226         if (device->open_count == 0)
 227                 vfio_device_put_kvm(device);
 228         mutex_unlock(&device->dev_set->lock);
 229 out_unlock:
 230         mutex_unlock(&device->group->group_lock);
 231         return ret;
 232 }
 233
 234 void vfio_df_group_close(struct vfio_device_file *df)
 235 {
 236         struct vfio_device *device = df->device;
 237
 238         mutex_lock(&device->group->group_lock);
 239         mutex_lock(&device->dev_set->lock);
 240
 241         vfio_df_close(df);
 242         df->iommufd = NULL;
 243
 244         if (device->open_count == 0)
 245                 vfio_device_put_kvm(device);
 246
 247         mutex_unlock(&device->dev_set->lock);
 248         mutex_unlock(&device->group->group_lock);
 249 }
 250
 251 static struct file *vfio_device_open_file(struct vfio_device *device)
 252 {
 253         struct vfio_device_file *df;
 254         struct file *filep;
 255         int ret;
 256
 257         df = vfio_allocate_device_file(device);
 258         if (IS_ERR(df)) {
 259                 ret = PTR_ERR(df);
 260                 goto err_out;
 261         }
 262
 263         df->group = device->group;
 264
 265         ret = vfio_df_group_open(df);
 266         if (ret)
 267                 goto err_free;
 268
 269         /*
 270          * We can't use anon_inode_getfd() because we need to modify
 271          * the f_mode flags directly to allow more than just ioctls
 272          */
 273         filep = anon_inode_getfile("[vfio-device]", &vfio_device_fops,
 274                                    df, O_RDWR);
 275         if (IS_ERR(filep)) {
 276                 ret = PTR_ERR(filep);
 277                 goto err_close_device;
 278         }
 279
 280         /*
 281          * TODO: add an anon_inode interface to do this.
 282          * Appears to be missing by lack of need rather than
 283          * explicitly prevented.  Now there's need.
 284          */
 285         filep->f_mode |= (FMODE_PREAD | FMODE_PWRITE);
 286
 287         /*
 288          * Use the pseudo fs inode on the device to link all mmaps
 289          * to the same address space, allowing us to unmap all vmas
 290          * associated to this device using unmap_mapping_range().
 291          */
 292         filep->f_mapping = device->inode->i_mapping;
 293
 294         if (device->group->type == VFIO_NO_IOMMU)
 295                 dev_warn(device->dev, "vfio-noiommu device opened by user "
 296                          "(%s:%d)\n", current->comm, task_pid_nr(current));
 297         /*
 298          * On success the ref of device is moved to the file and
 299          * put in vfio_device_fops_release()
 300          */
 301         return filep;
 302
 303 err_close_device:
 304         vfio_df_group_close(df);
 305 err_free:
 306         kfree(df);
 307 err_out:
 308         return ERR_PTR(ret);
 309 }
 310
 311 static int vfio_group_ioctl_get_device_fd(struct vfio_group *group,
 312                                           char __user *arg)
 313 {
 314         struct vfio_device *device;
 315         struct file *filep;
 316         char *buf;
 317         int fdno;
 318         int ret;
 319
 320         buf = strndup_user(arg, PAGE_SIZE);
 321         if (IS_ERR(buf))
 322                 return PTR_ERR(buf);
 323
 324         device = vfio_device_get_from_name(group, buf);
 325         kfree(buf);
 326         if (IS_ERR(device))
 327                 return PTR_ERR(device);
 328
 329         fdno = get_unused_fd_flags(O_CLOEXEC);
 330         if (fdno < 0) {
 331                 ret = fdno;
 332                 goto err_put_device;
 333         }
 334
 335         filep = vfio_device_open_file(device);
 336         if (IS_ERR(filep)) {
 337                 ret = PTR_ERR(filep);
 338                 goto err_put_fdno;
 339         }
 340
 341         fd_install(fdno, filep);
 342         return fdno;
 343
 344 err_put_fdno:
 345         put_unused_fd(fdno);
 346 err_put_device:
 347         vfio_device_put_registration(device);
 348         return ret;
 349 }
 350
 351 static int vfio_group_ioctl_get_status(struct vfio_group *group,
 352                                        struct vfio_group_status __user *arg)
 353 {
 354         unsigned long minsz = offsetofend(struct vfio_group_status, flags);
 355         struct vfio_group_status status;
 356
 357         if (copy_from_user(&status, arg, minsz))
 358                 return -EFAULT;
 359
 360         if (status.argsz < minsz)
 361                 return -EINVAL;
 362
 363         status.flags = 0;
 364
 365         mutex_lock(&group->group_lock);
 366         if (!group->iommu_group) {
 367                 mutex_unlock(&group->group_lock);
 368                 return -ENODEV;
 369         }
 370
 371         /*
 372          * With the container FD the iommu_group_claim_dma_owner() is done
 373          * during SET_CONTAINER but for IOMMFD this is done during
 374          * VFIO_GROUP_GET_DEVICE_FD. Meaning that with iommufd
 375          * VFIO_GROUP_FLAGS_VIABLE could be set but GET_DEVICE_FD will fail due
 376          * to viability.
 377          */
 378         if (vfio_group_has_iommu(group))
 379                 status.flags |= VFIO_GROUP_FLAGS_CONTAINER_SET |
 380                                 VFIO_GROUP_FLAGS_VIABLE;
 381         else if (!iommu_group_dma_owner_claimed(group->iommu_group))
 382                 status.flags |= VFIO_GROUP_FLAGS_VIABLE;
 383         mutex_unlock(&group->group_lock);
 384
 385         if (copy_to_user(arg, &status, minsz))
 386                 return -EFAULT;
 387         return 0;
 388 }
 389
 390 static long vfio_group_fops_unl_ioctl(struct file *filep,
 391                                       unsigned int cmd, unsigned long arg)
 392 {
 393         struct vfio_group *group = filep->private_data;
 394         void __user *uarg = (void __user *)arg;
 395
 396         switch (cmd) {
 397         case VFIO_GROUP_GET_DEVICE_FD:
 398                 return vfio_group_ioctl_get_device_fd(group, uarg);
 399         case VFIO_GROUP_GET_STATUS:
 400                 return vfio_group_ioctl_get_status(group, uarg);
 401         case VFIO_GROUP_SET_CONTAINER:
 402                 return vfio_group_ioctl_set_container(group, uarg);
 403         case VFIO_GROUP_UNSET_CONTAINER:
 404                 return vfio_group_ioctl_unset_container(group);
 405         default:
 406                 return -ENOTTY;
 407         }
 408 }
 409
 410 int vfio_device_block_group(struct vfio_device *device)
 411 {
 412         struct vfio_group *group = device->group;
 413         int ret = 0;
 414
 415         mutex_lock(&group->group_lock);
 416         if (group->opened_file) {
 417                 ret = -EBUSY;
 418                 goto out_unlock;
 419         }
 420
 421         group->cdev_device_open_cnt++;
 422
 423 out_unlock:
 424         mutex_unlock(&group->group_lock);
 425         return ret;
 426 }
 427
 428 void vfio_device_unblock_group(struct vfio_device *device)
 429 {
 430         struct vfio_group *group = device->group;
 431
 432         mutex_lock(&group->group_lock);
 433         group->cdev_device_open_cnt--;
 434         mutex_unlock(&group->group_lock);
 435 }
 436
 437 static int vfio_group_fops_open(struct inode *inode, struct file *filep)
 438 {
 439         struct vfio_group *group =
 440                 container_of(inode->i_cdev, struct vfio_group, cdev);
 441         int ret;
 442
 443         mutex_lock(&group->group_lock);
 444
 445         /*
 446          * drivers can be zero if this races with vfio_device_remove_group(), it
 447          * will be stable at 0 under the group rwsem
 448          */
 449         if (refcount_read(&group->drivers) == 0) {
 450                 ret = -ENODEV;
 451                 goto out_unlock;
 452         }
 453
 454         if (group->type == VFIO_NO_IOMMU && !capable(CAP_SYS_RAWIO)) {
 455                 ret = -EPERM;
 456                 goto out_unlock;
 457         }
 458
 459         if (group->cdev_device_open_cnt) {
 460                 ret = -EBUSY;
 461                 goto out_unlock;
 462         }
 463
 464         /*
 465          * Do we need multiple instances of the group open?  Seems not.
 466          */
 467         if (group->opened_file) {
 468                 ret = -EBUSY;
 469                 goto out_unlock;
 470         }
 471         group->opened_file = filep;
 472         filep->private_data = group;
 473         ret = 0;
 474 out_unlock:
 475         mutex_unlock(&group->group_lock);
 476         return ret;
 477 }
 478
 479 static int vfio_group_fops_release(struct inode *inode, struct file *filep)
 480 {
 481         struct vfio_group *group = filep->private_data;
 482
 483         filep->private_data = NULL;
 484
 485         mutex_lock(&group->group_lock);
 486         /*
 487          * Device FDs hold a group file reference, therefore the group release
 488          * is only called when there are no open devices.
 489          */
 490         WARN_ON(group->notifier.head);
 491         if (group->container)
 492                 vfio_group_detach_container(group);
 493         if (group->iommufd) {
 494                 iommufd_ctx_put(group->iommufd);
 495                 group->iommufd = NULL;
 496         }
 497         group->opened_file = NULL;
 498         mutex_unlock(&group->group_lock);
 499         return 0;
 500 }
 501
 502 static const struct file_operations vfio_group_fops = {
 503         .owner          = THIS_MODULE,
 504         .unlocked_ioctl = vfio_group_fops_unl_ioctl,
 505         .compat_ioctl   = compat_ptr_ioctl,
 506         .open           = vfio_group_fops_open,
 507         .release        = vfio_group_fops_release,
 508 };
 509
 510 /*
 511  * Group objects - create, release, get, put, search
 512  */
 513 static struct vfio_group *
 514 vfio_group_find_from_iommu(struct iommu_group *iommu_group)
 515 {
 516         struct vfio_group *group;
 517
 518         lockdep_assert_held(&vfio.group_lock);
 519
 520         /*
 521          * group->iommu_group from the vfio.group_list cannot be NULL
 522          * under the vfio.group_lock.
 523          */
 524         list_for_each_entry(group, &vfio.group_list, vfio_next) {
 525                 if (group->iommu_group == iommu_group)
 526                         return group;
 527         }
 528         return NULL;
 529 }
 530
 531 static void vfio_group_release(struct device *dev)
 532 {
 533         struct vfio_group *group = container_of(dev, struct vfio_group, dev);
 534
 535         mutex_destroy(&group->device_lock);
 536         mutex_destroy(&group->group_lock);
 537         WARN_ON(group->iommu_group);
 538         WARN_ON(group->cdev_device_open_cnt);
 539         ida_free(&vfio.group_ida, MINOR(group->dev.devt));
 540         kfree(group);
 541 }
 542
 543 static struct vfio_group *vfio_group_alloc(struct iommu_group *iommu_group,
 544                                            enum vfio_group_type type)
 545 {
 546         struct vfio_group *group;
 547         int minor;
 548
 549         group = kzalloc(sizeof(*group), GFP_KERNEL);
 550         if (!group)
 551                 return ERR_PTR(-ENOMEM);
 552
 553         minor = ida_alloc_max(&vfio.group_ida, MINORMASK, GFP_KERNEL);
 554         if (minor < 0) {
 555                 kfree(group);
 556                 return ERR_PTR(minor);
 557         }
 558
 559         device_initialize(&group->dev);
 560         group->dev.devt = MKDEV(MAJOR(vfio.group_devt), minor);
 561         group->dev.class = vfio.class;
 562         group->dev.release = vfio_group_release;
 563         cdev_init(&group->cdev, &vfio_group_fops);
 564         group->cdev.owner = THIS_MODULE;
 565
 566         refcount_set(&group->drivers, 1);
 567         mutex_init(&group->group_lock);
 568         spin_lock_init(&group->kvm_ref_lock);
 569         INIT_LIST_HEAD(&group->device_list);
 570         mutex_init(&group->device_lock);
 571         group->iommu_group = iommu_group;
 572         /* put in vfio_group_release() */
 573         iommu_group_ref_get(iommu_group);
 574         group->type = type;
 575         BLOCKING_INIT_NOTIFIER_HEAD(&group->notifier);
 576
 577         return group;
 578 }
 579
 580 static struct vfio_group *vfio_create_group(struct iommu_group *iommu_group,
 581                 enum vfio_group_type type)
 582 {
 583         struct vfio_group *group;
 584         struct vfio_group *ret;
 585         int err;
 586
 587         lockdep_assert_held(&vfio.group_lock);
 588
 589         group = vfio_group_alloc(iommu_group, type);
 590         if (IS_ERR(group))
 591                 return group;
 592
 593         err = dev_set_name(&group->dev, "%s%d",
 594                            group->type == VFIO_NO_IOMMU ? "noiommu-" : "",
 595                            iommu_group_id(iommu_group));
 596         if (err) {
 597                 ret = ERR_PTR(err);
 598                 goto err_put;
 599         }
 600
 601         err = cdev_device_add(&group->cdev, &group->dev);
 602         if (err) {
 603                 ret = ERR_PTR(err);
 604                 goto err_put;
 605         }
 606
 607         list_add(&group->vfio_next, &vfio.group_list);
 608
 609         return group;
 610
 611 err_put:
 612         put_device(&group->dev);
 613         return ret;
 614 }
 615
 616 static struct vfio_group *vfio_noiommu_group_alloc(struct device *dev,
 617                 enum vfio_group_type type)
 618 {
 619         struct iommu_group *iommu_group;
 620         struct vfio_group *group;
 621         int ret;
 622
 623         iommu_group = iommu_group_alloc();
 624         if (IS_ERR(iommu_group))
 625                 return ERR_CAST(iommu_group);
 626
 627         ret = iommu_group_set_name(iommu_group, "vfio-noiommu");
 628         if (ret)
 629                 goto out_put_group;
 630         ret = iommu_group_add_device(iommu_group, dev);
 631         if (ret)
 632                 goto out_put_group;
 633
 634         mutex_lock(&vfio.group_lock);
 635         group = vfio_create_group(iommu_group, type);
 636         mutex_unlock(&vfio.group_lock);
 637         if (IS_ERR(group)) {
 638                 ret = PTR_ERR(group);
 639                 goto out_remove_device;
 640         }
 641         iommu_group_put(iommu_group);
 642         return group;
 643
 644 out_remove_device:
 645         iommu_group_remove_device(dev);
 646 out_put_group:
 647         iommu_group_put(iommu_group);
 648         return ERR_PTR(ret);
 649 }
 650
 651 static bool vfio_group_has_device(struct vfio_group *group, struct device *dev)
 652 {
 653         struct vfio_device *device;
 654
 655         mutex_lock(&group->device_lock);
 656         list_for_each_entry(device, &group->device_list, group_next) {
 657                 if (device->dev == dev) {
 658                         mutex_unlock(&group->device_lock);
 659                         return true;
 660                 }
 661         }
 662         mutex_unlock(&group->device_lock);
 663         return false;
 664 }
 665
 666 static struct vfio_group *vfio_group_find_or_alloc(struct device *dev)
 667 {
 668         struct iommu_group *iommu_group;
 669         struct vfio_group *group;
 670
 671         iommu_group = iommu_group_get(dev);
 672         if (!iommu_group && vfio_noiommu) {
 673                 /*
 674                  * With noiommu enabled, create an IOMMU group for devices that
 675                  * don't already have one, implying no IOMMU hardware/driver
 676                  * exists.  Taint the kernel because we're about to give a DMA
 677                  * capable device to a user without IOMMU protection.
 678                  */
 679                 group = vfio_noiommu_group_alloc(dev, VFIO_NO_IOMMU);
 680                 if (!IS_ERR(group)) {
 681                         add_taint(TAINT_USER, LOCKDEP_STILL_OK);
 682                         dev_warn(dev, "Adding kernel taint for vfio-noiommu group on device\n");
 683                 }
 684                 return group;
 685         }
 686
 687         if (!iommu_group)
 688                 return ERR_PTR(-EINVAL);
 689
 690         mutex_lock(&vfio.group_lock);
 691         group = vfio_group_find_from_iommu(iommu_group);
 692         if (group) {
 693                 if (WARN_ON(vfio_group_has_device(group, dev)))
 694                         group = ERR_PTR(-EINVAL);
 695                 else
 696                         refcount_inc(&group->drivers);
 697         } else {
 698                 group = vfio_create_group(iommu_group, VFIO_IOMMU);
 699         }
 700         mutex_unlock(&vfio.group_lock);
 701
 702         /* The vfio_group holds a reference to the iommu_group */
 703         iommu_group_put(iommu_group);
 704         return group;
 705 }
 706
 707 int vfio_device_set_group(struct vfio_device *device,
 708                           enum vfio_group_type type)
 709 {
 710         struct vfio_group *group;
 711
 712         if (type == VFIO_IOMMU)
 713                 group = vfio_group_find_or_alloc(device->dev);
 714         else
 715                 group = vfio_noiommu_group_alloc(device->dev, type);
 716
 717         if (IS_ERR(group))
 718                 return PTR_ERR(group);
 719
 720         /* Our reference on group is moved to the device */
 721         device->group = group;
 722         return 0;
 723 }
 724
 725 void vfio_device_remove_group(struct vfio_device *device)
 726 {
 727         struct vfio_group *group = device->group;
 728         struct iommu_group *iommu_group;
 729
 730         if (group->type == VFIO_NO_IOMMU || group->type == VFIO_EMULATED_IOMMU)
 731                 iommu_group_remove_device(device->dev);
 732
 733         /* Pairs with vfio_create_group() / vfio_group_get_from_iommu() */
 734         if (!refcount_dec_and_mutex_lock(&group->drivers, &vfio.group_lock))
 735                 return;
 736         list_del(&group->vfio_next);
 737
 738         /*
 739          * We could concurrently probe another driver in the group that might
 740          * race vfio_device_remove_group() with vfio_get_group(), so we have to
 741          * ensure that the sysfs is all cleaned up under lock otherwise the
 742          * cdev_device_add() will fail due to the name aready existing.
 743          */
 744         cdev_device_del(&group->cdev, &group->dev);
 745
 746         mutex_lock(&group->group_lock);
 747         /*
 748          * These data structures all have paired operations that can only be
 749          * undone when the caller holds a live reference on the device. Since
 750          * all pairs must be undone these WARN_ON's indicate some caller did not
 751          * properly hold the group reference.
 752          */
 753         WARN_ON(!list_empty(&group->device_list));
 754         WARN_ON(group->notifier.head);
 755
 756         /*
 757          * Revoke all users of group->iommu_group. At this point we know there
 758          * are no devices active because we are unplugging the last one. Setting
 759          * iommu_group to NULL blocks all new users.
 760          */
 761         if (group->container)
 762                 vfio_group_detach_container(group);
 763         iommu_group = group->iommu_group;
 764         group->iommu_group = NULL;
 765         mutex_unlock(&group->group_lock);
 766         mutex_unlock(&vfio.group_lock);
 767
 768         iommu_group_put(iommu_group);
 769         put_device(&group->dev);
 770 }
 771
 772 void vfio_device_group_register(struct vfio_device *device)
 773 {
 774         mutex_lock(&device->group->device_lock);
 775         list_add(&device->group_next, &device->group->device_list);
 776         mutex_unlock(&device->group->device_lock);
 777 }
 778
 779 void vfio_device_group_unregister(struct vfio_device *device)
 780 {
 781         mutex_lock(&device->group->device_lock);
 782         list_del(&device->group_next);
 783         mutex_unlock(&device->group->device_lock);
 784 }
 785
 786 int vfio_device_group_use_iommu(struct vfio_device *device)
 787 {
 788         struct vfio_group *group = device->group;
 789         int ret = 0;
 790
 791         lockdep_assert_held(&group->group_lock);
 792
 793         if (WARN_ON(!group->container))
 794                 return -EINVAL;
 795
 796         ret = vfio_group_use_container(group);
 797         if (ret)
 798                 return ret;
 799         vfio_device_container_register(device);
 800         return 0;
 801 }
 802
 803 void vfio_device_group_unuse_iommu(struct vfio_device *device)
 804 {
 805         struct vfio_group *group = device->group;
 806
 807         lockdep_assert_held(&group->group_lock);
 808
 809         if (WARN_ON(!group->container))
 810                 return;
 811
 812         vfio_device_container_unregister(device);
 813         vfio_group_unuse_container(group);
 814 }
 815
 816 bool vfio_device_has_container(struct vfio_device *device)
 817 {
 818         return device->group->container;
 819 }
 820
 821 struct vfio_group *vfio_group_from_file(struct file *file)
 822 {
 823         struct vfio_group *group = file->private_data;
 824
 825         if (file->f_op != &vfio_group_fops)
 826                 return NULL;
 827         return group;
 828 }
 829
 830 /**
 831  * vfio_file_iommu_group - Return the struct iommu_group for the vfio group file
 832  * @file: VFIO group file
 833  *
 834  * The returned iommu_group is valid as long as a ref is held on the file. This
 835  * returns a reference on the group. This function is deprecated, only the SPAPR
 836  * path in kvm should call it.
 837  */
 838 struct iommu_group *vfio_file_iommu_group(struct file *file)
 839 {
 840         struct vfio_group *group = vfio_group_from_file(file);
 841         struct iommu_group *iommu_group = NULL;
 842
 843         if (!IS_ENABLED(CONFIG_SPAPR_TCE_IOMMU))
 844                 return NULL;
 845
 846         if (!group)
 847                 return NULL;
 848
 849         mutex_lock(&group->group_lock);
 850         if (group->iommu_group) {
 851                 iommu_group = group->iommu_group;
 852                 iommu_group_ref_get(iommu_group);
 853         }
 854         mutex_unlock(&group->group_lock);
 855         return iommu_group;
 856 }
 857 EXPORT_SYMBOL_GPL(vfio_file_iommu_group);
 858
 859 /**
 860  * vfio_file_is_group - True if the file is a vfio group file
 861  * @file: VFIO group file
 862  */
 863 bool vfio_file_is_group(struct file *file)
 864 {
 865         return vfio_group_from_file(file);
 866 }
 867 EXPORT_SYMBOL_GPL(vfio_file_is_group);
 868
 869 bool vfio_group_enforced_coherent(struct vfio_group *group)
 870 {
 871         struct vfio_device *device;
 872         bool ret = true;
 873
 874         /*
 875          * If the device does not have IOMMU_CAP_ENFORCE_CACHE_COHERENCY then
 876          * any domain later attached to it will also not support it. If the cap
 877          * is set then the iommu_domain eventually attached to the device/group
 878          * must use a domain with enforce_cache_coherency().
 879          */
 880         mutex_lock(&group->device_lock);
 881         list_for_each_entry(device, &group->device_list, group_next) {
 882                 if (!device_iommu_capable(device->dev,
 883                                           IOMMU_CAP_ENFORCE_CACHE_COHERENCY)) {
 884                         ret = false;
 885                         break;
 886                 }
 887         }
 888         mutex_unlock(&group->device_lock);
 889         return ret;
 890 }
 891
 892 void vfio_group_set_kvm(struct vfio_group *group, struct kvm *kvm)
 893 {
 894         spin_lock(&group->kvm_ref_lock);
 895         group->kvm = kvm;
 896         spin_unlock(&group->kvm_ref_lock);
 897 }
 898
 899 /**
 900  * vfio_file_has_dev - True if the VFIO file is a handle for device
 901  * @file: VFIO file to check
 902  * @device: Device that must be part of the file
 903  *
 904  * Returns true if given file has permission to manipulate the given device.
 905  */
 906 bool vfio_file_has_dev(struct file *file, struct vfio_device *device)
 907 {
 908         struct vfio_group *group = vfio_group_from_file(file);
 909
 910         if (!group)
 911                 return false;
 912
 913         return group == device->group;
 914 }
 915 EXPORT_SYMBOL_GPL(vfio_file_has_dev);
 916
 917 static char *vfio_devnode(const struct device *dev, umode_t *mode)
 918 {
 919         return kasprintf(GFP_KERNEL, "vfio/%s", dev_name(dev));
 920 }
 921
 922 int __init vfio_group_init(void)
 923 {
 924         int ret;
 925
 926         ida_init(&vfio.group_ida);
 927         mutex_init(&vfio.group_lock);
 928         INIT_LIST_HEAD(&vfio.group_list);
 929
 930         ret = vfio_container_init();
 931         if (ret)
 932                 return ret;
 933
 934         /* /dev/vfio/$GROUP */
 935         vfio.class = class_create("vfio");
 936         if (IS_ERR(vfio.class)) {
 937                 ret = PTR_ERR(vfio.class);
 938                 goto err_group_class;
 939         }
 940
 941         vfio.class->devnode = vfio_devnode;
 942
 943         ret = alloc_chrdev_region(&vfio.group_devt, 0, MINORMASK + 1, "vfio");
 944         if (ret)
 945                 goto err_alloc_chrdev;
 946         return 0;
 947
 948 err_alloc_chrdev:
 949         class_destroy(vfio.class);
 950         vfio.class = NULL;
 951 err_group_class:
 952         vfio_container_cleanup();
 953         return ret;
 954 }
 955
 956 void vfio_group_cleanup(void)
 957 {
 958         WARN_ON(!list_empty(&vfio.group_list));
 959         ida_destroy(&vfio.group_ida);
 960         unregister_chrdev_region(vfio.group_devt, MINORMASK + 1);
 961         class_destroy(vfio.class);
 962         vfio.class = NULL;
 963         vfio_container_cleanup();
 964 }