aboutsummaryrefslogtreecommitdiffstats
path: root/drivers/vfio/vfio.c
diff options
context:
space:
mode:
Diffstat (limited to 'drivers/vfio/vfio.c')
-rw-r--r--drivers/vfio/vfio.c188
1 files changed, 7 insertions, 181 deletions
diff --git a/drivers/vfio/vfio.c b/drivers/vfio/vfio.c
index de632da2e22f..6070b793cbcb 100644
--- a/drivers/vfio/vfio.c
+++ b/drivers/vfio/vfio.c
@@ -62,7 +62,6 @@ struct vfio_container {
62 struct rw_semaphore group_lock; 62 struct rw_semaphore group_lock;
63 struct vfio_iommu_driver *iommu_driver; 63 struct vfio_iommu_driver *iommu_driver;
64 void *iommu_data; 64 void *iommu_data;
65 bool noiommu;
66}; 65};
67 66
68struct vfio_unbound_dev { 67struct vfio_unbound_dev {
@@ -85,7 +84,6 @@ struct vfio_group {
85 struct list_head unbound_list; 84 struct list_head unbound_list;
86 struct mutex unbound_lock; 85 struct mutex unbound_lock;
87 atomic_t opened; 86 atomic_t opened;
88 bool noiommu;
89}; 87};
90 88
91struct vfio_device { 89struct vfio_device {
@@ -97,147 +95,6 @@ struct vfio_device {
97 void *device_data; 95 void *device_data;
98}; 96};
99 97
100#ifdef CONFIG_VFIO_NOIOMMU
101static bool noiommu __read_mostly;
102module_param_named(enable_unsafe_noiommu_support,
103 noiommu, bool, S_IRUGO | S_IWUSR);
104MODULE_PARM_DESC(enable_unsafe_noiommu_mode, "Enable UNSAFE, no-IOMMU mode. This mode provides no device isolation, no DMA translation, no host kernel protection, cannot be used for device assignment to virtual machines, requires RAWIO permissions, and will taint the kernel. If you do not know what this is for, step away. (default: false)");
105#endif
106
107/*
108 * vfio_iommu_group_{get,put} are only intended for VFIO bus driver probe
109 * and remove functions, any use cases other than acquiring the first
110 * reference for the purpose of calling vfio_add_group_dev() or removing
111 * that symmetric reference after vfio_del_group_dev() should use the raw
112 * iommu_group_{get,put} functions. In particular, vfio_iommu_group_put()
113 * removes the device from the dummy group and cannot be nested.
114 */
115struct iommu_group *vfio_iommu_group_get(struct device *dev)
116{
117 struct iommu_group *group;
118 int __maybe_unused ret;
119
120 group = iommu_group_get(dev);
121
122#ifdef CONFIG_VFIO_NOIOMMU
123 /*
124 * With noiommu enabled, an IOMMU group will be created for a device
125 * that doesn't already have one and doesn't have an iommu_ops on their
126 * bus. We use iommu_present() again in the main code to detect these
127 * fake groups.
128 */
129 if (group || !noiommu || iommu_present(dev->bus))
130 return group;
131
132 group = iommu_group_alloc();
133 if (IS_ERR(group))
134 return NULL;
135
136 iommu_group_set_name(group, "vfio-noiommu");
137 ret = iommu_group_add_device(group, dev);
138 iommu_group_put(group);
139 if (ret)
140 return NULL;
141
142 /*
143 * Where to taint? At this point we've added an IOMMU group for a
144 * device that is not backed by iommu_ops, therefore any iommu_
145 * callback using iommu_ops can legitimately Oops. So, while we may
146 * be about to give a DMA capable device to a user without IOMMU
147 * protection, which is clearly taint-worthy, let's go ahead and do
148 * it here.
149 */
150 add_taint(TAINT_USER, LOCKDEP_STILL_OK);
151 dev_warn(dev, "Adding kernel taint for vfio-noiommu group on device\n");
152#endif
153
154 return group;
155}
156EXPORT_SYMBOL_GPL(vfio_iommu_group_get);
157
158void vfio_iommu_group_put(struct iommu_group *group, struct device *dev)
159{
160#ifdef CONFIG_VFIO_NOIOMMU
161 if (!iommu_present(dev->bus))
162 iommu_group_remove_device(dev);
163#endif
164
165 iommu_group_put(group);
166}
167EXPORT_SYMBOL_GPL(vfio_iommu_group_put);
168
169#ifdef CONFIG_VFIO_NOIOMMU
170static void *vfio_noiommu_open(unsigned long arg)
171{
172 if (arg != VFIO_NOIOMMU_IOMMU)
173 return ERR_PTR(-EINVAL);
174 if (!capable(CAP_SYS_RAWIO))
175 return ERR_PTR(-EPERM);
176
177 return NULL;
178}
179
180static void vfio_noiommu_release(void *iommu_data)
181{
182}
183
184static long vfio_noiommu_ioctl(void *iommu_data,
185 unsigned int cmd, unsigned long arg)
186{
187 if (cmd == VFIO_CHECK_EXTENSION)
188 return arg == VFIO_NOIOMMU_IOMMU ? 1 : 0;
189
190 return -ENOTTY;
191}
192
193static int vfio_iommu_present(struct device *dev, void *unused)
194{
195 return iommu_present(dev->bus) ? 1 : 0;
196}
197
198static int vfio_noiommu_attach_group(void *iommu_data,
199 struct iommu_group *iommu_group)
200{
201 return iommu_group_for_each_dev(iommu_group, NULL,
202 vfio_iommu_present) ? -EINVAL : 0;
203}
204
205static void vfio_noiommu_detach_group(void *iommu_data,
206 struct iommu_group *iommu_group)
207{
208}
209
210static struct vfio_iommu_driver_ops vfio_noiommu_ops = {
211 .name = "vfio-noiommu",
212 .owner = THIS_MODULE,
213 .open = vfio_noiommu_open,
214 .release = vfio_noiommu_release,
215 .ioctl = vfio_noiommu_ioctl,
216 .attach_group = vfio_noiommu_attach_group,
217 .detach_group = vfio_noiommu_detach_group,
218};
219
220static struct vfio_iommu_driver vfio_noiommu_driver = {
221 .ops = &vfio_noiommu_ops,
222};
223
224/*
225 * Wrap IOMMU drivers, the noiommu driver is the one and only driver for
226 * noiommu groups (and thus containers) and not available for normal groups.
227 */
228#define vfio_for_each_iommu_driver(con, pos) \
229 for (pos = con->noiommu ? &vfio_noiommu_driver : \
230 list_first_entry(&vfio.iommu_drivers_list, \
231 struct vfio_iommu_driver, vfio_next); \
232 (con->noiommu ? pos != NULL : \
233 &pos->vfio_next != &vfio.iommu_drivers_list); \
234 pos = con->noiommu ? NULL : list_next_entry(pos, vfio_next))
235#else
236#define vfio_for_each_iommu_driver(con, pos) \
237 list_for_each_entry(pos, &vfio.iommu_drivers_list, vfio_next)
238#endif
239
240
241/** 98/**
242 * IOMMU driver registration 99 * IOMMU driver registration
243 */ 100 */
@@ -342,8 +199,7 @@ static void vfio_group_unlock_and_free(struct vfio_group *group)
342/** 199/**
343 * Group objects - create, release, get, put, search 200 * Group objects - create, release, get, put, search
344 */ 201 */
345static struct vfio_group *vfio_create_group(struct iommu_group *iommu_group, 202static struct vfio_group *vfio_create_group(struct iommu_group *iommu_group)
346 bool noiommu)
347{ 203{
348 struct vfio_group *group, *tmp; 204 struct vfio_group *group, *tmp;
349 struct device *dev; 205 struct device *dev;
@@ -361,7 +217,6 @@ static struct vfio_group *vfio_create_group(struct iommu_group *iommu_group,
361 atomic_set(&group->container_users, 0); 217 atomic_set(&group->container_users, 0);
362 atomic_set(&group->opened, 0); 218 atomic_set(&group->opened, 0);
363 group->iommu_group = iommu_group; 219 group->iommu_group = iommu_group;
364 group->noiommu = noiommu;
365 220
366 group->nb.notifier_call = vfio_iommu_group_notifier; 221 group->nb.notifier_call = vfio_iommu_group_notifier;
367 222
@@ -397,8 +252,7 @@ static struct vfio_group *vfio_create_group(struct iommu_group *iommu_group,
397 252
398 dev = device_create(vfio.class, NULL, 253 dev = device_create(vfio.class, NULL,
399 MKDEV(MAJOR(vfio.group_devt), minor), 254 MKDEV(MAJOR(vfio.group_devt), minor),
400 group, "%s%d", noiommu ? "noiommu-" : "", 255 group, "%d", iommu_group_id(iommu_group));
401 iommu_group_id(iommu_group));
402 if (IS_ERR(dev)) { 256 if (IS_ERR(dev)) {
403 vfio_free_group_minor(minor); 257 vfio_free_group_minor(minor);
404 vfio_group_unlock_and_free(group); 258 vfio_group_unlock_and_free(group);
@@ -682,7 +536,7 @@ static int vfio_group_nb_add_dev(struct vfio_group *group, struct device *dev)
682 return 0; 536 return 0;
683 537
684 /* TODO Prevent device auto probing */ 538 /* TODO Prevent device auto probing */
685 WARN("Device %s added to live group %d!\n", dev_name(dev), 539 WARN(1, "Device %s added to live group %d!\n", dev_name(dev),
686 iommu_group_id(group->iommu_group)); 540 iommu_group_id(group->iommu_group));
687 541
688 return 0; 542 return 0;
@@ -786,8 +640,7 @@ int vfio_add_group_dev(struct device *dev,
786 640
787 group = vfio_group_get_from_iommu(iommu_group); 641 group = vfio_group_get_from_iommu(iommu_group);
788 if (!group) { 642 if (!group) {
789 group = vfio_create_group(iommu_group, 643 group = vfio_create_group(iommu_group);
790 !iommu_present(dev->bus));
791 if (IS_ERR(group)) { 644 if (IS_ERR(group)) {
792 iommu_group_put(iommu_group); 645 iommu_group_put(iommu_group);
793 return PTR_ERR(group); 646 return PTR_ERR(group);
@@ -999,7 +852,8 @@ static long vfio_ioctl_check_extension(struct vfio_container *container,
999 */ 852 */
1000 if (!driver) { 853 if (!driver) {
1001 mutex_lock(&vfio.iommu_drivers_lock); 854 mutex_lock(&vfio.iommu_drivers_lock);
1002 vfio_for_each_iommu_driver(container, driver) { 855 list_for_each_entry(driver, &vfio.iommu_drivers_list,
856 vfio_next) {
1003 if (!try_module_get(driver->ops->owner)) 857 if (!try_module_get(driver->ops->owner))
1004 continue; 858 continue;
1005 859
@@ -1068,7 +922,7 @@ static long vfio_ioctl_set_iommu(struct vfio_container *container,
1068 } 922 }
1069 923
1070 mutex_lock(&vfio.iommu_drivers_lock); 924 mutex_lock(&vfio.iommu_drivers_lock);
1071 vfio_for_each_iommu_driver(container, driver) { 925 list_for_each_entry(driver, &vfio.iommu_drivers_list, vfio_next) {
1072 void *data; 926 void *data;
1073 927
1074 if (!try_module_get(driver->ops->owner)) 928 if (!try_module_get(driver->ops->owner))
@@ -1333,9 +1187,6 @@ static int vfio_group_set_container(struct vfio_group *group, int container_fd)
1333 if (atomic_read(&group->container_users)) 1187 if (atomic_read(&group->container_users))
1334 return -EINVAL; 1188 return -EINVAL;
1335 1189
1336 if (group->noiommu && !capable(CAP_SYS_RAWIO))
1337 return -EPERM;
1338
1339 f = fdget(container_fd); 1190 f = fdget(container_fd);
1340 if (!f.file) 1191 if (!f.file)
1341 return -EBADF; 1192 return -EBADF;
@@ -1351,13 +1202,6 @@ static int vfio_group_set_container(struct vfio_group *group, int container_fd)
1351 1202
1352 down_write(&container->group_lock); 1203 down_write(&container->group_lock);
1353 1204
1354 /* Real groups and fake groups cannot mix */
1355 if (!list_empty(&container->group_list) &&
1356 container->noiommu != group->noiommu) {
1357 ret = -EPERM;
1358 goto unlock_out;
1359 }
1360
1361 driver = container->iommu_driver; 1205 driver = container->iommu_driver;
1362 if (driver) { 1206 if (driver) {
1363 ret = driver->ops->attach_group(container->iommu_data, 1207 ret = driver->ops->attach_group(container->iommu_data,
@@ -1367,7 +1211,6 @@ static int vfio_group_set_container(struct vfio_group *group, int container_fd)
1367 } 1211 }
1368 1212
1369 group->container = container; 1213 group->container = container;
1370 container->noiommu = group->noiommu;
1371 list_add(&group->container_next, &container->group_list); 1214 list_add(&group->container_next, &container->group_list);
1372 1215
1373 /* Get a reference on the container and mark a user within the group */ 1216 /* Get a reference on the container and mark a user within the group */
@@ -1398,9 +1241,6 @@ static int vfio_group_get_device_fd(struct vfio_group *group, char *buf)
1398 !group->container->iommu_driver || !vfio_group_viable(group)) 1241 !group->container->iommu_driver || !vfio_group_viable(group))
1399 return -EINVAL; 1242 return -EINVAL;
1400 1243
1401 if (group->noiommu && !capable(CAP_SYS_RAWIO))
1402 return -EPERM;
1403
1404 device = vfio_device_get_from_name(group, buf); 1244 device = vfio_device_get_from_name(group, buf);
1405 if (!device) 1245 if (!device)
1406 return -ENODEV; 1246 return -ENODEV;
@@ -1443,10 +1283,6 @@ static int vfio_group_get_device_fd(struct vfio_group *group, char *buf)
1443 1283
1444 fd_install(ret, filep); 1284 fd_install(ret, filep);
1445 1285
1446 if (group->noiommu)
1447 dev_warn(device->dev, "vfio-noiommu device opened by user "
1448 "(%s:%d)\n", current->comm, task_pid_nr(current));
1449
1450 return ret; 1286 return ret;
1451} 1287}
1452 1288
@@ -1535,11 +1371,6 @@ static int vfio_group_fops_open(struct inode *inode, struct file *filep)
1535 if (!group) 1371 if (!group)
1536 return -ENODEV; 1372 return -ENODEV;
1537 1373
1538 if (group->noiommu && !capable(CAP_SYS_RAWIO)) {
1539 vfio_group_put(group);
1540 return -EPERM;
1541 }
1542
1543 /* Do we need multiple instances of the group open? Seems not. */ 1374 /* Do we need multiple instances of the group open? Seems not. */
1544 opened = atomic_cmpxchg(&group->opened, 0, 1); 1375 opened = atomic_cmpxchg(&group->opened, 0, 1);
1545 if (opened) { 1376 if (opened) {
@@ -1702,11 +1533,6 @@ struct vfio_group *vfio_group_get_external_user(struct file *filep)
1702 if (!atomic_inc_not_zero(&group->container_users)) 1533 if (!atomic_inc_not_zero(&group->container_users))
1703 return ERR_PTR(-EINVAL); 1534 return ERR_PTR(-EINVAL);
1704 1535
1705 if (group->noiommu) {
1706 atomic_dec(&group->container_users);
1707 return ERR_PTR(-EPERM);
1708 }
1709
1710 if (!group->container->iommu_driver || 1536 if (!group->container->iommu_driver ||
1711 !vfio_group_viable(group)) { 1537 !vfio_group_viable(group)) {
1712 atomic_dec(&group->container_users); 1538 atomic_dec(&group->container_users);