| 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 <[email protected]> |
| 7 |
* |
| 8 |
* Derived from original vfio: |
| 9 |
* Copyright 2010 Cisco Systems, Inc. All rights reserved. |
| 10 |
* Author: Tom Lyon, [email protected] |
| 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 |
goto out_put_kvm; |
| 198 |
} |
| 199 |
} |
| 200 |
|
| 201 |
ret = vfio_df_open(df); |
| 202 |
if (ret) |
| 203 |
goto out_put_kvm; |
| 204 |
|
| 205 |
if (df->iommufd && device->open_count == 1) { |
| 206 |
ret = vfio_iommufd_compat_attach_ioas(device, df->iommufd); |
| 207 |
if (ret) |
| 208 |
goto out_close_device; |
| 209 |
} |
| 210 |
|
| 211 |
/* |
| 212 |
* Paired with smp_load_acquire() in vfio_device_fops::ioctl/ |
| 213 |
* read/write/mmap and vfio_file_has_device_access() |
| 214 |
*/ |
| 215 |
smp_store_release(&df->access_granted, true); |
| 216 |
|
| 217 |
mutex_unlock(&device->dev_set->lock); |
| 218 |
mutex_unlock(&device->group->group_lock); |
| 219 |
return 0; |
| 220 |
|
| 221 |
out_close_device: |
| 222 |
vfio_df_close(df); |
| 223 |
out_put_kvm: |
| 224 |
df->iommufd = NULL; |
| 225 |
if (device->open_count == 0) |
| 226 |
vfio_device_put_kvm(device); |
| 227 |
mutex_unlock(&device->dev_set->lock); |
| 228 |
out_unlock: |
| 229 |
mutex_unlock(&device->group->group_lock); |
| 230 |
return ret; |
| 231 |
} |
| 232 |
|
| 233 |
void vfio_df_group_close(struct vfio_device_file *df) |
| 234 |
{ |
| 235 |
struct vfio_device *device = df->device; |
| 236 |
|
| 237 |
mutex_lock(&device->group->group_lock); |
| 238 |
mutex_lock(&device->dev_set->lock); |
| 239 |
|
| 240 |
vfio_df_close(df); |
| 241 |
df->iommufd = NULL; |
| 242 |
|
| 243 |
if (device->open_count == 0) |
| 244 |
vfio_device_put_kvm(device); |
| 245 |
|
| 246 |
mutex_unlock(&device->dev_set->lock); |
| 247 |
mutex_unlock(&device->group->group_lock); |
| 248 |
} |
| 249 |
|
| 250 |
static struct file *vfio_device_open_file(struct vfio_device *device) |
| 251 |
{ |
| 252 |
struct vfio_device_file *df; |
| 253 |
struct file *filep; |
| 254 |
int ret; |
| 255 |
|
| 256 |
df = vfio_allocate_device_file(device); |
| 257 |
if (IS_ERR(df)) { |
| 258 |
ret = PTR_ERR(df); |
| 259 |
goto err_out; |
| 260 |
} |
| 261 |
|
| 262 |
df->group = device->group; |
| 263 |
|
| 264 |
ret = vfio_df_group_open(df); |
| 265 |
if (ret) |
| 266 |
goto err_free; |
| 267 |
|
| 268 |
filep = anon_inode_getfile_fmode("[vfio-device]", &vfio_device_fops, |
| 269 |
df, O_RDWR, FMODE_PREAD | FMODE_PWRITE); |
| 270 |
if (IS_ERR(filep)) { |
| 271 |
ret = PTR_ERR(filep); |
| 272 |
goto err_close_device; |
| 273 |
} |
| 274 |
/* |
| 275 |
* Use the pseudo fs inode on the device to link all mmaps |
| 276 |
* to the same address space, allowing us to unmap all vmas |
| 277 |
* associated to this device using unmap_mapping_range(). |
| 278 |
*/ |
| 279 |
filep->f_mapping = device->inode->i_mapping; |
| 280 |
|
| 281 |
if (device->group->type == VFIO_NO_IOMMU) |
| 282 |
dev_warn(device->dev, "vfio-noiommu device opened by user " |
| 283 |
"(%s:%d)\n", current->comm, task_pid_nr(current)); |
| 284 |
/* |
| 285 |
* On success the ref of device is moved to the file and |
| 286 |
* put in vfio_device_fops_release() |
| 287 |
*/ |
| 288 |
return filep; |
| 289 |
|
| 290 |
err_close_device: |
| 291 |
vfio_df_group_close(df); |
| 292 |
err_free: |
| 293 |
kfree(df); |
| 294 |
err_out: |
| 295 |
return ERR_PTR(ret); |
| 296 |
} |
| 297 |
|
| 298 |
static int vfio_group_ioctl_get_device_fd(struct vfio_group *group, |
| 299 |
char __user *arg) |
| 300 |
{ |
| 301 |
struct vfio_device *device; |
| 302 |
char *buf; |
| 303 |
int fd; |
| 304 |
|
| 305 |
buf = strndup_user(arg, PAGE_SIZE); |
| 306 |
if (IS_ERR(buf)) |
| 307 |
return PTR_ERR(buf); |
| 308 |
|
| 309 |
device = vfio_device_get_from_name(group, buf); |
| 310 |
kfree(buf); |
| 311 |
if (IS_ERR(device)) |
| 312 |
return PTR_ERR(device); |
| 313 |
|
| 314 |
fd = FD_ADD(O_CLOEXEC, vfio_device_open_file(device)); |
| 315 |
if (fd < 0) |
| 316 |
vfio_device_put_registration(device); |
| 317 |
return fd; |
| 318 |
} |
| 319 |
|
| 320 |
static int vfio_group_ioctl_get_status(struct vfio_group *group, |
| 321 |
struct vfio_group_status __user *arg) |
| 322 |
{ |
| 323 |
unsigned long minsz = offsetofend(struct vfio_group_status, flags); |
| 324 |
struct vfio_group_status status; |
| 325 |
|
| 326 |
if (copy_from_user(&status, arg, minsz)) |
| 327 |
return -EFAULT; |
| 328 |
|
| 329 |
if (status.argsz < minsz) |
| 330 |
return -EINVAL; |
| 331 |
|
| 332 |
status.flags = 0; |
| 333 |
|
| 334 |
mutex_lock(&group->group_lock); |
| 335 |
if (!group->iommu_group) { |
| 336 |
mutex_unlock(&group->group_lock); |
| 337 |
return -ENODEV; |
| 338 |
} |
| 339 |
|
| 340 |
/* |
| 341 |
* With the container FD the iommu_group_claim_dma_owner() is done |
| 342 |
* during SET_CONTAINER but for IOMMFD this is done during |
| 343 |
* VFIO_GROUP_GET_DEVICE_FD. Meaning that with iommufd |
| 344 |
* VFIO_GROUP_FLAGS_VIABLE could be set but GET_DEVICE_FD will fail due |
| 345 |
* to viability. |
| 346 |
*/ |
| 347 |
if (vfio_group_has_iommu(group)) |
| 348 |
status.flags |= VFIO_GROUP_FLAGS_CONTAINER_SET | |
| 349 |
VFIO_GROUP_FLAGS_VIABLE; |
| 350 |
else if (!iommu_group_dma_owner_claimed(group->iommu_group)) |
| 351 |
status.flags |= VFIO_GROUP_FLAGS_VIABLE; |
| 352 |
mutex_unlock(&group->group_lock); |
| 353 |
|
| 354 |
if (copy_to_user(arg, &status, minsz)) |
| 355 |
return -EFAULT; |
| 356 |
return 0; |
| 357 |
} |
| 358 |
|
| 359 |
static long vfio_group_fops_unl_ioctl(struct file *filep, |
| 360 |
unsigned int cmd, unsigned long arg) |
| 361 |
{ |
| 362 |
struct vfio_group *group = filep->private_data; |
| 363 |
void __user *uarg = (void __user *)arg; |
| 364 |
|
| 365 |
switch (cmd) { |
| 366 |
case VFIO_GROUP_GET_DEVICE_FD: |
| 367 |
return vfio_group_ioctl_get_device_fd(group, uarg); |
| 368 |
case VFIO_GROUP_GET_STATUS: |
| 369 |
return vfio_group_ioctl_get_status(group, uarg); |
| 370 |
case VFIO_GROUP_SET_CONTAINER: |
| 371 |
return vfio_group_ioctl_set_container(group, uarg); |
| 372 |
case VFIO_GROUP_UNSET_CONTAINER: |
| 373 |
return vfio_group_ioctl_unset_container(group); |
| 374 |
default: |
| 375 |
return -ENOTTY; |
| 376 |
} |
| 377 |
} |
| 378 |
|
| 379 |
int vfio_device_block_group(struct vfio_device *device) |
| 380 |
{ |
| 381 |
struct vfio_group *group = device->group; |
| 382 |
int ret = 0; |
| 383 |
|
| 384 |
mutex_lock(&group->group_lock); |
| 385 |
if (group->opened_file) { |
| 386 |
ret = -EBUSY; |
| 387 |
goto out_unlock; |
| 388 |
} |
| 389 |
|
| 390 |
group->cdev_device_open_cnt++; |
| 391 |
|
| 392 |
out_unlock: |
| 393 |
mutex_unlock(&group->group_lock); |
| 394 |
return ret; |
| 395 |
} |
| 396 |
|
| 397 |
void vfio_device_unblock_group(struct vfio_device *device) |
| 398 |
{ |
| 399 |
struct vfio_group *group = device->group; |
| 400 |
|
| 401 |
mutex_lock(&group->group_lock); |
| 402 |
group->cdev_device_open_cnt--; |
| 403 |
mutex_unlock(&group->group_lock); |
| 404 |
} |
| 405 |
|
| 406 |
static int vfio_group_fops_open(struct inode *inode, struct file *filep) |
| 407 |
{ |
| 408 |
struct vfio_group *group = |
| 409 |
container_of(inode->i_cdev, struct vfio_group, cdev); |
| 410 |
int ret; |
| 411 |
|
| 412 |
mutex_lock(&group->group_lock); |
| 413 |
|
| 414 |
/* |
| 415 |
* drivers can be zero if this races with vfio_device_remove_group(), it |
| 416 |
* will be stable at 0 under the group rwsem |
| 417 |
*/ |
| 418 |
if (refcount_read(&group->drivers) == 0) { |
| 419 |
ret = -ENODEV; |
| 420 |
goto out_unlock; |
| 421 |
} |
| 422 |
|
| 423 |
if (group->type == VFIO_NO_IOMMU && !capable(CAP_SYS_RAWIO)) { |
| 424 |
ret = -EPERM; |
| 425 |
goto out_unlock; |
| 426 |
} |
| 427 |
|
| 428 |
if (group->cdev_device_open_cnt) { |
| 429 |
ret = -EBUSY; |
| 430 |
goto out_unlock; |
| 431 |
} |
| 432 |
|
| 433 |
/* |
| 434 |
* Do we need multiple instances of the group open? Seems not. |
| 435 |
*/ |
| 436 |
if (group->opened_file) { |
| 437 |
ret = -EBUSY; |
| 438 |
goto out_unlock; |
| 439 |
} |
| 440 |
group->opened_file = filep; |
| 441 |
filep->private_data = group; |
| 442 |
ret = 0; |
| 443 |
out_unlock: |
| 444 |
mutex_unlock(&group->group_lock); |
| 445 |
return ret; |
| 446 |
} |
| 447 |
|
| 448 |
static int vfio_group_fops_release(struct inode *inode, struct file *filep) |
| 449 |
{ |
| 450 |
struct vfio_group *group = filep->private_data; |
| 451 |
|
| 452 |
filep->private_data = NULL; |
| 453 |
|
| 454 |
mutex_lock(&group->group_lock); |
| 455 |
/* |
| 456 |
* Device FDs hold a group file reference, therefore the group release |
| 457 |
* is only called when there are no open devices. |
| 458 |
*/ |
| 459 |
WARN_ON(group->notifier.head); |
| 460 |
if (group->container) |
| 461 |
vfio_group_detach_container(group); |
| 462 |
if (group->iommufd) { |
| 463 |
iommufd_ctx_put(group->iommufd); |
| 464 |
group->iommufd = NULL; |
| 465 |
} |
| 466 |
group->opened_file = NULL; |
| 467 |
mutex_unlock(&group->group_lock); |
| 468 |
return 0; |
| 469 |
} |
| 470 |
|
| 471 |
static const struct file_operations vfio_group_fops = { |
| 472 |
.owner = THIS_MODULE, |
| 473 |
.unlocked_ioctl = vfio_group_fops_unl_ioctl, |
| 474 |
.compat_ioctl = compat_ptr_ioctl, |
| 475 |
.open = vfio_group_fops_open, |
| 476 |
.release = vfio_group_fops_release, |
| 477 |
}; |
| 478 |
|
| 479 |
/* |
| 480 |
* Group objects - create, release, get, put, search |
| 481 |
*/ |
| 482 |
static struct vfio_group * |
| 483 |
vfio_group_find_from_iommu(struct iommu_group *iommu_group) |
| 484 |
{ |
| 485 |
struct vfio_group *group; |
| 486 |
|
| 487 |
lockdep_assert_held(&vfio.group_lock); |
| 488 |
|
| 489 |
/* |
| 490 |
* group->iommu_group from the vfio.group_list cannot be NULL |
| 491 |
* under the vfio.group_lock. |
| 492 |
*/ |
| 493 |
list_for_each_entry(group, &vfio.group_list, vfio_next) { |
| 494 |
if (group->iommu_group == iommu_group) |
| 495 |
return group; |
| 496 |
} |
| 497 |
return NULL; |
| 498 |
} |
| 499 |
|
| 500 |
static void vfio_group_release(struct device *dev) |
| 501 |
{ |
| 502 |
struct vfio_group *group = container_of(dev, struct vfio_group, dev); |
| 503 |
|
| 504 |
mutex_destroy(&group->device_lock); |
| 505 |
mutex_destroy(&group->group_lock); |
| 506 |
WARN_ON(group->iommu_group); |
| 507 |
WARN_ON(group->cdev_device_open_cnt); |
| 508 |
ida_free(&vfio.group_ida, MINOR(group->dev.devt)); |
| 509 |
kfree(group); |
| 510 |
} |
| 511 |
|
| 512 |
static struct vfio_group *vfio_group_alloc(struct iommu_group *iommu_group, |
| 513 |
enum vfio_group_type type) |
| 514 |
{ |
| 515 |
struct vfio_group *group; |
| 516 |
int minor; |
| 517 |
|
| 518 |
group = kzalloc_obj(*group); |
| 519 |
if (!group) |
| 520 |
return ERR_PTR(-ENOMEM); |
| 521 |
|
| 522 |
minor = ida_alloc_max(&vfio.group_ida, MINORMASK, GFP_KERNEL); |
| 523 |
if (minor < 0) { |
| 524 |
kfree(group); |
| 525 |
return ERR_PTR(minor); |
| 526 |
} |
| 527 |
|
| 528 |
device_initialize(&group->dev); |
| 529 |
group->dev.devt = MKDEV(MAJOR(vfio.group_devt), minor); |
| 530 |
group->dev.class = vfio.class; |
| 531 |
group->dev.release = vfio_group_release; |
| 532 |
cdev_init(&group->cdev, &vfio_group_fops); |
| 533 |
group->cdev.owner = THIS_MODULE; |
| 534 |
|
| 535 |
refcount_set(&group->drivers, 1); |
| 536 |
mutex_init(&group->group_lock); |
| 537 |
spin_lock_init(&group->kvm_ref_lock); |
| 538 |
INIT_LIST_HEAD(&group->device_list); |
| 539 |
mutex_init(&group->device_lock); |
| 540 |
group->iommu_group = iommu_group; |
| 541 |
/* put in vfio_group_release() */ |
| 542 |
iommu_group_ref_get(iommu_group); |
| 543 |
group->type = type; |
| 544 |
BLOCKING_INIT_NOTIFIER_HEAD(&group->notifier); |
| 545 |
|
| 546 |
return group; |
| 547 |
} |
| 548 |
|
| 549 |
static struct vfio_group *vfio_create_group(struct iommu_group *iommu_group, |
| 550 |
enum vfio_group_type type) |
| 551 |
{ |
| 552 |
struct vfio_group *group; |
| 553 |
struct vfio_group *ret; |
| 554 |
int err; |
| 555 |
|
| 556 |
lockdep_assert_held(&vfio.group_lock); |
| 557 |
|
| 558 |
group = vfio_group_alloc(iommu_group, type); |
| 559 |
if (IS_ERR(group)) |
| 560 |
return group; |
| 561 |
|
| 562 |
err = dev_set_name(&group->dev, "%s%d", |
| 563 |
group->type == VFIO_NO_IOMMU ? "noiommu-" : "", |
| 564 |
iommu_group_id(iommu_group)); |
| 565 |
if (err) { |
| 566 |
ret = ERR_PTR(err); |
| 567 |
goto err_put; |
| 568 |
} |
| 569 |
|
| 570 |
err = cdev_device_add(&group->cdev, &group->dev); |
| 571 |
if (err) { |
| 572 |
ret = ERR_PTR(err); |
| 573 |
goto err_put; |
| 574 |
} |
| 575 |
|
| 576 |
list_add(&group->vfio_next, &vfio.group_list); |
| 577 |
|
| 578 |
return group; |
| 579 |
|
| 580 |
err_put: |
| 581 |
put_device(&group->dev); |
| 582 |
return ret; |
| 583 |
} |
| 584 |
|
| 585 |
static struct vfio_group *vfio_noiommu_group_alloc(struct device *dev, |
| 586 |
enum vfio_group_type type) |
| 587 |
{ |
| 588 |
struct iommu_group *iommu_group; |
| 589 |
struct vfio_group *group; |
| 590 |
int ret; |
| 591 |
|
| 592 |
iommu_group = iommu_group_alloc(); |
| 593 |
if (IS_ERR(iommu_group)) |
| 594 |
return ERR_CAST(iommu_group); |
| 595 |
|
| 596 |
ret = iommu_group_set_name(iommu_group, "vfio-noiommu"); |
| 597 |
if (ret) |
| 598 |
goto out_put_group; |
| 599 |
ret = iommu_group_add_device(iommu_group, dev); |
| 600 |
if (ret) |
| 601 |
goto out_put_group; |
| 602 |
|
| 603 |
mutex_lock(&vfio.group_lock); |
| 604 |
group = vfio_create_group(iommu_group, type); |
| 605 |
mutex_unlock(&vfio.group_lock); |
| 606 |
if (IS_ERR(group)) { |
| 607 |
ret = PTR_ERR(group); |
| 608 |
goto out_remove_device; |
| 609 |
} |
| 610 |
iommu_group_put(iommu_group); |
| 611 |
return group; |
| 612 |
|
| 613 |
out_remove_device: |
| 614 |
iommu_group_remove_device(dev); |
| 615 |
out_put_group: |
| 616 |
iommu_group_put(iommu_group); |
| 617 |
return ERR_PTR(ret); |
| 618 |
} |
| 619 |
|
| 620 |
static bool vfio_group_has_device(struct vfio_group *group, struct device *dev) |
| 621 |
{ |
| 622 |
struct vfio_device *device; |
| 623 |
|
| 624 |
mutex_lock(&group->device_lock); |
| 625 |
list_for_each_entry(device, &group->device_list, group_next) { |
| 626 |
if (device->dev == dev) { |
| 627 |
mutex_unlock(&group->device_lock); |
| 628 |
return true; |
| 629 |
} |
| 630 |
} |
| 631 |
mutex_unlock(&group->device_lock); |
| 632 |
return false; |
| 633 |
} |
| 634 |
|
| 635 |
static struct vfio_group *vfio_group_find_or_alloc(struct device *dev) |
| 636 |
{ |
| 637 |
struct iommu_group *iommu_group; |
| 638 |
struct vfio_group *group; |
| 639 |
|
| 640 |
iommu_group = iommu_group_get(dev); |
| 641 |
if (!iommu_group && vfio_noiommu) { |
| 642 |
/* |
| 643 |
* With noiommu enabled, create an IOMMU group for devices that |
| 644 |
* don't already have one, implying no IOMMU hardware/driver |
| 645 |
* exists. Taint the kernel because we're about to give a DMA |
| 646 |
* capable device to a user without IOMMU protection. |
| 647 |
*/ |
| 648 |
group = vfio_noiommu_group_alloc(dev, VFIO_NO_IOMMU); |
| 649 |
if (!IS_ERR(group)) { |
| 650 |
add_taint(TAINT_USER, LOCKDEP_STILL_OK); |
| 651 |
dev_warn(dev, "Adding kernel taint for vfio-noiommu group on device\n"); |
| 652 |
} |
| 653 |
return group; |
| 654 |
} |
| 655 |
|
| 656 |
if (!iommu_group) |
| 657 |
return ERR_PTR(-EINVAL); |
| 658 |
|
| 659 |
mutex_lock(&vfio.group_lock); |
| 660 |
group = vfio_group_find_from_iommu(iommu_group); |
| 661 |
if (group) { |
| 662 |
if (WARN_ON(vfio_group_has_device(group, dev))) |
| 663 |
group = ERR_PTR(-EINVAL); |
| 664 |
else |
| 665 |
refcount_inc(&group->drivers); |
| 666 |
} else { |
| 667 |
group = vfio_create_group(iommu_group, VFIO_IOMMU); |
| 668 |
} |
| 669 |
mutex_unlock(&vfio.group_lock); |
| 670 |
|
| 671 |
/* The vfio_group holds a reference to the iommu_group */ |
| 672 |
iommu_group_put(iommu_group); |
| 673 |
return group; |
| 674 |
} |
| 675 |
|
| 676 |
int vfio_device_set_group(struct vfio_device *device, |
| 677 |
enum vfio_group_type type) |
| 678 |
{ |
| 679 |
struct vfio_group *group; |
| 680 |
|
| 681 |
if (type == VFIO_IOMMU) |
| 682 |
group = vfio_group_find_or_alloc(device->dev); |
| 683 |
else |
| 684 |
group = vfio_noiommu_group_alloc(device->dev, type); |
| 685 |
|
| 686 |
if (IS_ERR(group)) |
| 687 |
return PTR_ERR(group); |
| 688 |
|
| 689 |
/* Our reference on group is moved to the device */ |
| 690 |
device->group = group; |
| 691 |
return 0; |
| 692 |
} |
| 693 |
|
| 694 |
void vfio_device_remove_group(struct vfio_device *device) |
| 695 |
{ |
| 696 |
struct vfio_group *group = device->group; |
| 697 |
struct iommu_group *iommu_group; |
| 698 |
|
| 699 |
if (group->type == VFIO_NO_IOMMU || group->type == VFIO_EMULATED_IOMMU) |
| 700 |
iommu_group_remove_device(device->dev); |
| 701 |
|
| 702 |
/* Pairs with vfio_create_group() / vfio_group_get_from_iommu() */ |
| 703 |
if (!refcount_dec_and_mutex_lock(&group->drivers, &vfio.group_lock)) |
| 704 |
return; |
| 705 |
list_del(&group->vfio_next); |
| 706 |
|
| 707 |
/* |
| 708 |
* We could concurrently probe another driver in the group that might |
| 709 |
* race vfio_device_remove_group() with vfio_get_group(), so we have to |
| 710 |
* ensure that the sysfs is all cleaned up under lock otherwise the |
| 711 |
* cdev_device_add() will fail due to the name aready existing. |
| 712 |
*/ |
| 713 |
cdev_device_del(&group->cdev, &group->dev); |
| 714 |
|
| 715 |
mutex_lock(&group->group_lock); |
| 716 |
/* |
| 717 |
* These data structures all have paired operations that can only be |
| 718 |
* undone when the caller holds a live reference on the device. Since |
| 719 |
* all pairs must be undone these WARN_ON's indicate some caller did not |
| 720 |
* properly hold the group reference. |
| 721 |
*/ |
| 722 |
WARN_ON(!list_empty(&group->device_list)); |
| 723 |
WARN_ON(group->notifier.head); |
| 724 |
|
| 725 |
/* |
| 726 |
* Revoke all users of group->iommu_group. At this point we know there |
| 727 |
* are no devices active because we are unplugging the last one. Setting |
| 728 |
* iommu_group to NULL blocks all new users. |
| 729 |
*/ |
| 730 |
if (group->container) |
| 731 |
vfio_group_detach_container(group); |
| 732 |
iommu_group = group->iommu_group; |
| 733 |
group->iommu_group = NULL; |
| 734 |
mutex_unlock(&group->group_lock); |
| 735 |
mutex_unlock(&vfio.group_lock); |
| 736 |
|
| 737 |
iommu_group_put(iommu_group); |
| 738 |
put_device(&group->dev); |
| 739 |
} |
| 740 |
|
| 741 |
void vfio_device_group_register(struct vfio_device *device) |
| 742 |
{ |
| 743 |
mutex_lock(&device->group->device_lock); |
| 744 |
list_add(&device->group_next, &device->group->device_list); |
| 745 |
mutex_unlock(&device->group->device_lock); |
| 746 |
} |
| 747 |
|
| 748 |
void vfio_device_group_unregister(struct vfio_device *device) |
| 749 |
{ |
| 750 |
mutex_lock(&device->group->device_lock); |
| 751 |
list_del(&device->group_next); |
| 752 |
mutex_unlock(&device->group->device_lock); |
| 753 |
} |
| 754 |
|
| 755 |
int vfio_device_group_use_iommu(struct vfio_device *device) |
| 756 |
{ |
| 757 |
struct vfio_group *group = device->group; |
| 758 |
int ret = 0; |
| 759 |
|
| 760 |
lockdep_assert_held(&group->group_lock); |
| 761 |
|
| 762 |
if (WARN_ON(!group->container)) |
| 763 |
return -EINVAL; |
| 764 |
|
| 765 |
ret = vfio_group_use_container(group); |
| 766 |
if (ret) |
| 767 |
return ret; |
| 768 |
vfio_device_container_register(device); |
| 769 |
return 0; |
| 770 |
} |
| 771 |
|
| 772 |
void vfio_device_group_unuse_iommu(struct vfio_device *device) |
| 773 |
{ |
| 774 |
struct vfio_group *group = device->group; |
| 775 |
|
| 776 |
lockdep_assert_held(&group->group_lock); |
| 777 |
|
| 778 |
if (WARN_ON(!group->container)) |
| 779 |
return; |
| 780 |
|
| 781 |
vfio_device_container_unregister(device); |
| 782 |
vfio_group_unuse_container(group); |
| 783 |
} |
| 784 |
|
| 785 |
bool vfio_device_has_container(struct vfio_device *device) |
| 786 |
{ |
| 787 |
return device->group->container; |
| 788 |
} |
| 789 |
|
| 790 |
struct vfio_group *vfio_group_from_file(struct file *file) |
| 791 |
{ |
| 792 |
struct vfio_group *group = file->private_data; |
| 793 |
|
| 794 |
if (file->f_op != &vfio_group_fops) |
| 795 |
return NULL; |
| 796 |
return group; |
| 797 |
} |
| 798 |
|
| 799 |
/** |
| 800 |
* vfio_file_iommu_group - Return the struct iommu_group for the vfio group file |
| 801 |
* @file: VFIO group file |
| 802 |
* |
| 803 |
* The returned iommu_group is valid as long as a ref is held on the file. This |
| 804 |
* returns a reference on the group. This function is deprecated, only the SPAPR |
| 805 |
* path in kvm should call it. |
| 806 |
*/ |
| 807 |
struct iommu_group *vfio_file_iommu_group(struct file *file) |
| 808 |
{ |
| 809 |
struct vfio_group *group = vfio_group_from_file(file); |
| 810 |
struct iommu_group *iommu_group = NULL; |
| 811 |
|
| 812 |
if (!IS_ENABLED(CONFIG_SPAPR_TCE_IOMMU)) |
| 813 |
return NULL; |
| 814 |
|
| 815 |
if (!group) |
| 816 |
return NULL; |
| 817 |
|
| 818 |
mutex_lock(&group->group_lock); |
| 819 |
if (group->iommu_group) { |
| 820 |
iommu_group = group->iommu_group; |
| 821 |
iommu_group_ref_get(iommu_group); |
| 822 |
} |
| 823 |
mutex_unlock(&group->group_lock); |
| 824 |
return iommu_group; |
| 825 |
} |
| 826 |
EXPORT_SYMBOL_GPL(vfio_file_iommu_group); |
| 827 |
|
| 828 |
/** |
| 829 |
* vfio_file_is_group - True if the file is a vfio group file |
| 830 |
* @file: VFIO group file |
| 831 |
*/ |
| 832 |
bool vfio_file_is_group(struct file *file) |
| 833 |
{ |
| 834 |
return vfio_group_from_file(file); |
| 835 |
} |
| 836 |
EXPORT_SYMBOL_GPL(vfio_file_is_group); |
| 837 |
|
| 838 |
bool vfio_group_enforced_coherent(struct vfio_group *group) |
| 839 |
{ |
| 840 |
struct vfio_device *device; |
| 841 |
bool ret = true; |
| 842 |
|
| 843 |
/* |
| 844 |
* If the device does not have IOMMU_CAP_ENFORCE_CACHE_COHERENCY then |
| 845 |
* any domain later attached to it will also not support it. If the cap |
| 846 |
* is set then the iommu_domain eventually attached to the device/group |
| 847 |
* must use a domain with enforce_cache_coherency(). |
| 848 |
*/ |
| 849 |
mutex_lock(&group->device_lock); |
| 850 |
list_for_each_entry(device, &group->device_list, group_next) { |
| 851 |
if (!device_iommu_capable(device->dev, |
| 852 |
IOMMU_CAP_ENFORCE_CACHE_COHERENCY)) { |
| 853 |
ret = false; |
| 854 |
break; |
| 855 |
} |
| 856 |
} |
| 857 |
mutex_unlock(&group->device_lock); |
| 858 |
return ret; |
| 859 |
} |
| 860 |
|
| 861 |
void vfio_group_set_kvm(struct vfio_group *group, struct kvm *kvm) |
| 862 |
{ |
| 863 |
spin_lock(&group->kvm_ref_lock); |
| 864 |
group->kvm = kvm; |
| 865 |
spin_unlock(&group->kvm_ref_lock); |
| 866 |
} |
| 867 |
|
| 868 |
/** |
| 869 |
* vfio_file_has_dev - True if the VFIO file is a handle for device |
| 870 |
* @file: VFIO file to check |
| 871 |
* @device: Device that must be part of the file |
| 872 |
* |
| 873 |
* Returns true if given file has permission to manipulate the given device. |
| 874 |
*/ |
| 875 |
bool vfio_file_has_dev(struct file *file, struct vfio_device *device) |
| 876 |
{ |
| 877 |
struct vfio_group *group = vfio_group_from_file(file); |
| 878 |
|
| 879 |
if (!group) |
| 880 |
return false; |
| 881 |
|
| 882 |
return group == device->group; |
| 883 |
} |
| 884 |
EXPORT_SYMBOL_GPL(vfio_file_has_dev); |
| 885 |
|
| 886 |
static char *vfio_devnode(const struct device *dev, umode_t *mode) |
| 887 |
{ |
| 888 |
return kasprintf(GFP_KERNEL, "vfio/%s", dev_name(dev)); |
| 889 |
} |
| 890 |
|
| 891 |
int __init vfio_group_init(void) |
| 892 |
{ |
| 893 |
int ret; |
| 894 |
|
| 895 |
ida_init(&vfio.group_ida); |
| 896 |
mutex_init(&vfio.group_lock); |
| 897 |
INIT_LIST_HEAD(&vfio.group_list); |
| 898 |
|
| 899 |
ret = vfio_container_init(); |
| 900 |
if (ret) |
| 901 |
return ret; |
| 902 |
|
| 903 |
/* /dev/vfio/$GROUP */ |
| 904 |
vfio.class = class_create("vfio"); |
| 905 |
if (IS_ERR(vfio.class)) { |
| 906 |
ret = PTR_ERR(vfio.class); |
| 907 |
goto err_group_class; |
| 908 |
} |
| 909 |
|
| 910 |
vfio.class->devnode = vfio_devnode; |
| 911 |
|
| 912 |
ret = alloc_chrdev_region(&vfio.group_devt, 0, MINORMASK + 1, "vfio"); |
| 913 |
if (ret) |
| 914 |
goto err_alloc_chrdev; |
| 915 |
return 0; |
| 916 |
|
| 917 |
err_alloc_chrdev: |
| 918 |
class_destroy(vfio.class); |
| 919 |
vfio.class = NULL; |
| 920 |
err_group_class: |
| 921 |
vfio_container_cleanup(); |
| 922 |
return ret; |
| 923 |
} |
| 924 |
|
| 925 |
void vfio_group_cleanup(void) |
| 926 |
{ |
| 927 |
WARN_ON(!list_empty(&vfio.group_list)); |
| 928 |
ida_destroy(&vfio.group_ida); |
| 929 |
unregister_chrdev_region(vfio.group_devt, MINORMASK + 1); |
| 930 |
class_destroy(vfio.class); |
| 931 |
vfio.class = NULL; |
| 932 |
vfio_container_cleanup(); |
| 933 |
} |