Commit | Line | Data |
---|---|---|
cba3345c AW |
1 | /* |
2 | * VFIO core | |
3 | * | |
4 | * Copyright (C) 2012 Red Hat, Inc. All rights reserved. | |
5 | * Author: Alex Williamson <alex.williamson@redhat.com> | |
6 | * | |
7 | * This program is free software; you can redistribute it and/or modify | |
8 | * it under the terms of the GNU General Public License version 2 as | |
9 | * published by the Free Software Foundation. | |
10 | * | |
11 | * Derived from original vfio: | |
12 | * Copyright 2010 Cisco Systems, Inc. All rights reserved. | |
13 | * Author: Tom Lyon, pugs@cisco.com | |
14 | */ | |
15 | ||
16 | #include <linux/cdev.h> | |
17 | #include <linux/compat.h> | |
18 | #include <linux/device.h> | |
19 | #include <linux/file.h> | |
20 | #include <linux/anon_inodes.h> | |
21 | #include <linux/fs.h> | |
22 | #include <linux/idr.h> | |
23 | #include <linux/iommu.h> | |
24 | #include <linux/list.h> | |
d1099901 | 25 | #include <linux/miscdevice.h> |
cba3345c AW |
26 | #include <linux/module.h> |
27 | #include <linux/mutex.h> | |
9587f44a | 28 | #include <linux/rwsem.h> |
cba3345c AW |
29 | #include <linux/sched.h> |
30 | #include <linux/slab.h> | |
664e9386 | 31 | #include <linux/stat.h> |
cba3345c AW |
32 | #include <linux/string.h> |
33 | #include <linux/uaccess.h> | |
34 | #include <linux/vfio.h> | |
35 | #include <linux/wait.h> | |
36 | ||
37 | #define DRIVER_VERSION "0.3" | |
38 | #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>" | |
39 | #define DRIVER_DESC "VFIO - User Level meta-driver" | |
40 | ||
41 | static struct vfio { | |
42 | struct class *class; | |
43 | struct list_head iommu_drivers_list; | |
44 | struct mutex iommu_drivers_lock; | |
45 | struct list_head group_list; | |
46 | struct idr group_idr; | |
47 | struct mutex group_lock; | |
48 | struct cdev group_cdev; | |
d1099901 | 49 | dev_t group_devt; |
cba3345c AW |
50 | wait_queue_head_t release_q; |
51 | } vfio; | |
52 | ||
53 | struct vfio_iommu_driver { | |
54 | const struct vfio_iommu_driver_ops *ops; | |
55 | struct list_head vfio_next; | |
56 | }; | |
57 | ||
58 | struct vfio_container { | |
59 | struct kref kref; | |
60 | struct list_head group_list; | |
9587f44a | 61 | struct rw_semaphore group_lock; |
cba3345c AW |
62 | struct vfio_iommu_driver *iommu_driver; |
63 | void *iommu_data; | |
64 | }; | |
65 | ||
66 | struct vfio_group { | |
67 | struct kref kref; | |
68 | int minor; | |
69 | atomic_t container_users; | |
70 | struct iommu_group *iommu_group; | |
71 | struct vfio_container *container; | |
72 | struct list_head device_list; | |
73 | struct mutex device_lock; | |
74 | struct device *dev; | |
75 | struct notifier_block nb; | |
76 | struct list_head vfio_next; | |
77 | struct list_head container_next; | |
6d6768c6 | 78 | atomic_t opened; |
cba3345c AW |
79 | }; |
80 | ||
81 | struct vfio_device { | |
82 | struct kref kref; | |
83 | struct device *dev; | |
84 | const struct vfio_device_ops *ops; | |
85 | struct vfio_group *group; | |
86 | struct list_head group_next; | |
87 | void *device_data; | |
88 | }; | |
89 | ||
90 | /** | |
91 | * IOMMU driver registration | |
92 | */ | |
93 | int vfio_register_iommu_driver(const struct vfio_iommu_driver_ops *ops) | |
94 | { | |
95 | struct vfio_iommu_driver *driver, *tmp; | |
96 | ||
97 | driver = kzalloc(sizeof(*driver), GFP_KERNEL); | |
98 | if (!driver) | |
99 | return -ENOMEM; | |
100 | ||
101 | driver->ops = ops; | |
102 | ||
103 | mutex_lock(&vfio.iommu_drivers_lock); | |
104 | ||
105 | /* Check for duplicates */ | |
106 | list_for_each_entry(tmp, &vfio.iommu_drivers_list, vfio_next) { | |
107 | if (tmp->ops == ops) { | |
108 | mutex_unlock(&vfio.iommu_drivers_lock); | |
109 | kfree(driver); | |
110 | return -EINVAL; | |
111 | } | |
112 | } | |
113 | ||
114 | list_add(&driver->vfio_next, &vfio.iommu_drivers_list); | |
115 | ||
116 | mutex_unlock(&vfio.iommu_drivers_lock); | |
117 | ||
118 | return 0; | |
119 | } | |
120 | EXPORT_SYMBOL_GPL(vfio_register_iommu_driver); | |
121 | ||
122 | void vfio_unregister_iommu_driver(const struct vfio_iommu_driver_ops *ops) | |
123 | { | |
124 | struct vfio_iommu_driver *driver; | |
125 | ||
126 | mutex_lock(&vfio.iommu_drivers_lock); | |
127 | list_for_each_entry(driver, &vfio.iommu_drivers_list, vfio_next) { | |
128 | if (driver->ops == ops) { | |
129 | list_del(&driver->vfio_next); | |
130 | mutex_unlock(&vfio.iommu_drivers_lock); | |
131 | kfree(driver); | |
132 | return; | |
133 | } | |
134 | } | |
135 | mutex_unlock(&vfio.iommu_drivers_lock); | |
136 | } | |
137 | EXPORT_SYMBOL_GPL(vfio_unregister_iommu_driver); | |
138 | ||
139 | /** | |
140 | * Group minor allocation/free - both called with vfio.group_lock held | |
141 | */ | |
142 | static int vfio_alloc_group_minor(struct vfio_group *group) | |
143 | { | |
d1099901 | 144 | return idr_alloc(&vfio.group_idr, group, 0, MINORMASK + 1, GFP_KERNEL); |
cba3345c AW |
145 | } |
146 | ||
147 | static void vfio_free_group_minor(int minor) | |
148 | { | |
149 | idr_remove(&vfio.group_idr, minor); | |
150 | } | |
151 | ||
152 | static int vfio_iommu_group_notifier(struct notifier_block *nb, | |
153 | unsigned long action, void *data); | |
154 | static void vfio_group_get(struct vfio_group *group); | |
155 | ||
156 | /** | |
157 | * Container objects - containers are created when /dev/vfio/vfio is | |
158 | * opened, but their lifecycle extends until the last user is done, so | |
159 | * it's freed via kref. Must support container/group/device being | |
160 | * closed in any order. | |
161 | */ | |
162 | static void vfio_container_get(struct vfio_container *container) | |
163 | { | |
164 | kref_get(&container->kref); | |
165 | } | |
166 | ||
167 | static void vfio_container_release(struct kref *kref) | |
168 | { | |
169 | struct vfio_container *container; | |
170 | container = container_of(kref, struct vfio_container, kref); | |
171 | ||
172 | kfree(container); | |
173 | } | |
174 | ||
175 | static void vfio_container_put(struct vfio_container *container) | |
176 | { | |
177 | kref_put(&container->kref, vfio_container_release); | |
178 | } | |
179 | ||
9df7b25a JL |
180 | static void vfio_group_unlock_and_free(struct vfio_group *group) |
181 | { | |
182 | mutex_unlock(&vfio.group_lock); | |
183 | /* | |
184 | * Unregister outside of lock. A spurious callback is harmless now | |
185 | * that the group is no longer in vfio.group_list. | |
186 | */ | |
187 | iommu_group_unregister_notifier(group->iommu_group, &group->nb); | |
188 | kfree(group); | |
189 | } | |
190 | ||
cba3345c AW |
191 | /** |
192 | * Group objects - create, release, get, put, search | |
193 | */ | |
194 | static struct vfio_group *vfio_create_group(struct iommu_group *iommu_group) | |
195 | { | |
196 | struct vfio_group *group, *tmp; | |
197 | struct device *dev; | |
198 | int ret, minor; | |
199 | ||
200 | group = kzalloc(sizeof(*group), GFP_KERNEL); | |
201 | if (!group) | |
202 | return ERR_PTR(-ENOMEM); | |
203 | ||
204 | kref_init(&group->kref); | |
205 | INIT_LIST_HEAD(&group->device_list); | |
206 | mutex_init(&group->device_lock); | |
207 | atomic_set(&group->container_users, 0); | |
6d6768c6 | 208 | atomic_set(&group->opened, 0); |
cba3345c AW |
209 | group->iommu_group = iommu_group; |
210 | ||
211 | group->nb.notifier_call = vfio_iommu_group_notifier; | |
212 | ||
213 | /* | |
214 | * blocking notifiers acquire a rwsem around registering and hold | |
215 | * it around callback. Therefore, need to register outside of | |
216 | * vfio.group_lock to avoid A-B/B-A contention. Our callback won't | |
217 | * do anything unless it can find the group in vfio.group_list, so | |
218 | * no harm in registering early. | |
219 | */ | |
220 | ret = iommu_group_register_notifier(iommu_group, &group->nb); | |
221 | if (ret) { | |
222 | kfree(group); | |
223 | return ERR_PTR(ret); | |
224 | } | |
225 | ||
226 | mutex_lock(&vfio.group_lock); | |
227 | ||
228 | minor = vfio_alloc_group_minor(group); | |
229 | if (minor < 0) { | |
9df7b25a | 230 | vfio_group_unlock_and_free(group); |
cba3345c AW |
231 | return ERR_PTR(minor); |
232 | } | |
233 | ||
234 | /* Did we race creating this group? */ | |
235 | list_for_each_entry(tmp, &vfio.group_list, vfio_next) { | |
236 | if (tmp->iommu_group == iommu_group) { | |
237 | vfio_group_get(tmp); | |
238 | vfio_free_group_minor(minor); | |
9df7b25a | 239 | vfio_group_unlock_and_free(group); |
cba3345c AW |
240 | return tmp; |
241 | } | |
242 | } | |
243 | ||
d1099901 AW |
244 | dev = device_create(vfio.class, NULL, |
245 | MKDEV(MAJOR(vfio.group_devt), minor), | |
cba3345c AW |
246 | group, "%d", iommu_group_id(iommu_group)); |
247 | if (IS_ERR(dev)) { | |
248 | vfio_free_group_minor(minor); | |
9df7b25a | 249 | vfio_group_unlock_and_free(group); |
cba3345c AW |
250 | return (struct vfio_group *)dev; /* ERR_PTR */ |
251 | } | |
252 | ||
253 | group->minor = minor; | |
254 | group->dev = dev; | |
255 | ||
256 | list_add(&group->vfio_next, &vfio.group_list); | |
257 | ||
258 | mutex_unlock(&vfio.group_lock); | |
259 | ||
260 | return group; | |
261 | } | |
262 | ||
6d2cd3ce | 263 | /* called with vfio.group_lock held */ |
cba3345c AW |
264 | static void vfio_group_release(struct kref *kref) |
265 | { | |
266 | struct vfio_group *group = container_of(kref, struct vfio_group, kref); | |
267 | ||
268 | WARN_ON(!list_empty(&group->device_list)); | |
269 | ||
d1099901 | 270 | device_destroy(vfio.class, MKDEV(MAJOR(vfio.group_devt), group->minor)); |
cba3345c AW |
271 | list_del(&group->vfio_next); |
272 | vfio_free_group_minor(group->minor); | |
9df7b25a | 273 | vfio_group_unlock_and_free(group); |
cba3345c AW |
274 | } |
275 | ||
276 | static void vfio_group_put(struct vfio_group *group) | |
277 | { | |
6d2cd3ce | 278 | kref_put_mutex(&group->kref, vfio_group_release, &vfio.group_lock); |
cba3345c AW |
279 | } |
280 | ||
281 | /* Assume group_lock or group reference is held */ | |
282 | static void vfio_group_get(struct vfio_group *group) | |
283 | { | |
284 | kref_get(&group->kref); | |
285 | } | |
286 | ||
287 | /* | |
288 | * Not really a try as we will sleep for mutex, but we need to make | |
289 | * sure the group pointer is valid under lock and get a reference. | |
290 | */ | |
291 | static struct vfio_group *vfio_group_try_get(struct vfio_group *group) | |
292 | { | |
293 | struct vfio_group *target = group; | |
294 | ||
295 | mutex_lock(&vfio.group_lock); | |
296 | list_for_each_entry(group, &vfio.group_list, vfio_next) { | |
297 | if (group == target) { | |
298 | vfio_group_get(group); | |
299 | mutex_unlock(&vfio.group_lock); | |
300 | return group; | |
301 | } | |
302 | } | |
303 | mutex_unlock(&vfio.group_lock); | |
304 | ||
305 | return NULL; | |
306 | } | |
307 | ||
308 | static | |
309 | struct vfio_group *vfio_group_get_from_iommu(struct iommu_group *iommu_group) | |
310 | { | |
311 | struct vfio_group *group; | |
312 | ||
313 | mutex_lock(&vfio.group_lock); | |
314 | list_for_each_entry(group, &vfio.group_list, vfio_next) { | |
315 | if (group->iommu_group == iommu_group) { | |
316 | vfio_group_get(group); | |
317 | mutex_unlock(&vfio.group_lock); | |
318 | return group; | |
319 | } | |
320 | } | |
321 | mutex_unlock(&vfio.group_lock); | |
322 | ||
323 | return NULL; | |
324 | } | |
325 | ||
326 | static struct vfio_group *vfio_group_get_from_minor(int minor) | |
327 | { | |
328 | struct vfio_group *group; | |
329 | ||
330 | mutex_lock(&vfio.group_lock); | |
331 | group = idr_find(&vfio.group_idr, minor); | |
332 | if (!group) { | |
333 | mutex_unlock(&vfio.group_lock); | |
334 | return NULL; | |
335 | } | |
336 | vfio_group_get(group); | |
337 | mutex_unlock(&vfio.group_lock); | |
338 | ||
339 | return group; | |
340 | } | |
341 | ||
342 | /** | |
343 | * Device objects - create, release, get, put, search | |
344 | */ | |
345 | static | |
346 | struct vfio_device *vfio_group_create_device(struct vfio_group *group, | |
347 | struct device *dev, | |
348 | const struct vfio_device_ops *ops, | |
349 | void *device_data) | |
350 | { | |
351 | struct vfio_device *device; | |
cba3345c AW |
352 | |
353 | device = kzalloc(sizeof(*device), GFP_KERNEL); | |
354 | if (!device) | |
355 | return ERR_PTR(-ENOMEM); | |
356 | ||
357 | kref_init(&device->kref); | |
358 | device->dev = dev; | |
359 | device->group = group; | |
360 | device->ops = ops; | |
361 | device->device_data = device_data; | |
8283b491 | 362 | dev_set_drvdata(dev, device); |
cba3345c AW |
363 | |
364 | /* No need to get group_lock, caller has group reference */ | |
365 | vfio_group_get(group); | |
366 | ||
367 | mutex_lock(&group->device_lock); | |
368 | list_add(&device->group_next, &group->device_list); | |
369 | mutex_unlock(&group->device_lock); | |
370 | ||
371 | return device; | |
372 | } | |
373 | ||
374 | static void vfio_device_release(struct kref *kref) | |
375 | { | |
376 | struct vfio_device *device = container_of(kref, | |
377 | struct vfio_device, kref); | |
378 | struct vfio_group *group = device->group; | |
379 | ||
cba3345c AW |
380 | list_del(&device->group_next); |
381 | mutex_unlock(&group->device_lock); | |
382 | ||
383 | dev_set_drvdata(device->dev, NULL); | |
384 | ||
385 | kfree(device); | |
386 | ||
387 | /* vfio_del_group_dev may be waiting for this device */ | |
388 | wake_up(&vfio.release_q); | |
389 | } | |
390 | ||
391 | /* Device reference always implies a group reference */ | |
44f50716 | 392 | void vfio_device_put(struct vfio_device *device) |
cba3345c | 393 | { |
934ad4c2 | 394 | struct vfio_group *group = device->group; |
90b1253e | 395 | kref_put_mutex(&device->kref, vfio_device_release, &group->device_lock); |
934ad4c2 | 396 | vfio_group_put(group); |
cba3345c | 397 | } |
44f50716 | 398 | EXPORT_SYMBOL_GPL(vfio_device_put); |
cba3345c AW |
399 | |
400 | static void vfio_device_get(struct vfio_device *device) | |
401 | { | |
402 | vfio_group_get(device->group); | |
403 | kref_get(&device->kref); | |
404 | } | |
405 | ||
406 | static struct vfio_device *vfio_group_get_device(struct vfio_group *group, | |
407 | struct device *dev) | |
408 | { | |
409 | struct vfio_device *device; | |
410 | ||
411 | mutex_lock(&group->device_lock); | |
412 | list_for_each_entry(device, &group->device_list, group_next) { | |
413 | if (device->dev == dev) { | |
414 | vfio_device_get(device); | |
415 | mutex_unlock(&group->device_lock); | |
416 | return device; | |
417 | } | |
418 | } | |
419 | mutex_unlock(&group->device_lock); | |
420 | return NULL; | |
421 | } | |
422 | ||
423 | /* | |
424 | * Whitelist some drivers that we know are safe (no dma) or just sit on | |
425 | * a device. It's not always practical to leave a device within a group | |
426 | * driverless as it could get re-bound to something unsafe. | |
427 | */ | |
2b489a45 | 428 | static const char * const vfio_driver_whitelist[] = { "pci-stub", "pcieport" }; |
cba3345c AW |
429 | |
430 | static bool vfio_whitelisted_driver(struct device_driver *drv) | |
431 | { | |
432 | int i; | |
433 | ||
434 | for (i = 0; i < ARRAY_SIZE(vfio_driver_whitelist); i++) { | |
435 | if (!strcmp(drv->name, vfio_driver_whitelist[i])) | |
436 | return true; | |
437 | } | |
438 | ||
439 | return false; | |
440 | } | |
441 | ||
442 | /* | |
443 | * A vfio group is viable for use by userspace if all devices are either | |
444 | * driver-less or bound to a vfio or whitelisted driver. We test the | |
445 | * latter by the existence of a struct vfio_device matching the dev. | |
446 | */ | |
447 | static int vfio_dev_viable(struct device *dev, void *data) | |
448 | { | |
449 | struct vfio_group *group = data; | |
450 | struct vfio_device *device; | |
de2b3eea | 451 | struct device_driver *drv = ACCESS_ONCE(dev->driver); |
cba3345c | 452 | |
de2b3eea | 453 | if (!drv || vfio_whitelisted_driver(drv)) |
cba3345c AW |
454 | return 0; |
455 | ||
456 | device = vfio_group_get_device(group, dev); | |
457 | if (device) { | |
458 | vfio_device_put(device); | |
459 | return 0; | |
460 | } | |
461 | ||
462 | return -EINVAL; | |
463 | } | |
464 | ||
465 | /** | |
466 | * Async device support | |
467 | */ | |
468 | static int vfio_group_nb_add_dev(struct vfio_group *group, struct device *dev) | |
469 | { | |
470 | struct vfio_device *device; | |
471 | ||
472 | /* Do we already know about it? We shouldn't */ | |
473 | device = vfio_group_get_device(group, dev); | |
474 | if (WARN_ON_ONCE(device)) { | |
475 | vfio_device_put(device); | |
476 | return 0; | |
477 | } | |
478 | ||
479 | /* Nothing to do for idle groups */ | |
480 | if (!atomic_read(&group->container_users)) | |
481 | return 0; | |
482 | ||
483 | /* TODO Prevent device auto probing */ | |
484 | WARN("Device %s added to live group %d!\n", dev_name(dev), | |
485 | iommu_group_id(group->iommu_group)); | |
486 | ||
487 | return 0; | |
488 | } | |
489 | ||
cba3345c AW |
490 | static int vfio_group_nb_verify(struct vfio_group *group, struct device *dev) |
491 | { | |
492 | /* We don't care what happens when the group isn't in use */ | |
493 | if (!atomic_read(&group->container_users)) | |
494 | return 0; | |
495 | ||
496 | return vfio_dev_viable(dev, group); | |
497 | } | |
498 | ||
499 | static int vfio_iommu_group_notifier(struct notifier_block *nb, | |
500 | unsigned long action, void *data) | |
501 | { | |
502 | struct vfio_group *group = container_of(nb, struct vfio_group, nb); | |
503 | struct device *dev = data; | |
504 | ||
505 | /* | |
c6401930 AW |
506 | * Need to go through a group_lock lookup to get a reference or we |
507 | * risk racing a group being removed. Ignore spurious notifies. | |
cba3345c AW |
508 | */ |
509 | group = vfio_group_try_get(group); | |
c6401930 | 510 | if (!group) |
cba3345c AW |
511 | return NOTIFY_OK; |
512 | ||
513 | switch (action) { | |
514 | case IOMMU_GROUP_NOTIFY_ADD_DEVICE: | |
515 | vfio_group_nb_add_dev(group, dev); | |
516 | break; | |
517 | case IOMMU_GROUP_NOTIFY_DEL_DEVICE: | |
de9c7602 AW |
518 | /* |
519 | * Nothing to do here. If the device is in use, then the | |
520 | * vfio sub-driver should block the remove callback until | |
521 | * it is unused. If the device is unused or attached to a | |
522 | * stub driver, then it should be released and we don't | |
523 | * care that it will be going away. | |
524 | */ | |
cba3345c AW |
525 | break; |
526 | case IOMMU_GROUP_NOTIFY_BIND_DRIVER: | |
527 | pr_debug("%s: Device %s, group %d binding to driver\n", | |
528 | __func__, dev_name(dev), | |
529 | iommu_group_id(group->iommu_group)); | |
530 | break; | |
531 | case IOMMU_GROUP_NOTIFY_BOUND_DRIVER: | |
532 | pr_debug("%s: Device %s, group %d bound to driver %s\n", | |
533 | __func__, dev_name(dev), | |
534 | iommu_group_id(group->iommu_group), dev->driver->name); | |
535 | BUG_ON(vfio_group_nb_verify(group, dev)); | |
536 | break; | |
537 | case IOMMU_GROUP_NOTIFY_UNBIND_DRIVER: | |
538 | pr_debug("%s: Device %s, group %d unbinding from driver %s\n", | |
539 | __func__, dev_name(dev), | |
540 | iommu_group_id(group->iommu_group), dev->driver->name); | |
541 | break; | |
542 | case IOMMU_GROUP_NOTIFY_UNBOUND_DRIVER: | |
543 | pr_debug("%s: Device %s, group %d unbound from driver\n", | |
544 | __func__, dev_name(dev), | |
545 | iommu_group_id(group->iommu_group)); | |
546 | /* | |
547 | * XXX An unbound device in a live group is ok, but we'd | |
548 | * really like to avoid the above BUG_ON by preventing other | |
549 | * drivers from binding to it. Once that occurs, we have to | |
550 | * stop the system to maintain isolation. At a minimum, we'd | |
551 | * want a toggle to disable driver auto probe for this device. | |
552 | */ | |
553 | break; | |
554 | } | |
555 | ||
556 | vfio_group_put(group); | |
557 | return NOTIFY_OK; | |
558 | } | |
559 | ||
560 | /** | |
561 | * VFIO driver API | |
562 | */ | |
563 | int vfio_add_group_dev(struct device *dev, | |
564 | const struct vfio_device_ops *ops, void *device_data) | |
565 | { | |
566 | struct iommu_group *iommu_group; | |
567 | struct vfio_group *group; | |
568 | struct vfio_device *device; | |
569 | ||
570 | iommu_group = iommu_group_get(dev); | |
571 | if (!iommu_group) | |
572 | return -EINVAL; | |
573 | ||
574 | group = vfio_group_get_from_iommu(iommu_group); | |
575 | if (!group) { | |
576 | group = vfio_create_group(iommu_group); | |
577 | if (IS_ERR(group)) { | |
578 | iommu_group_put(iommu_group); | |
579 | return PTR_ERR(group); | |
580 | } | |
581 | } | |
582 | ||
583 | device = vfio_group_get_device(group, dev); | |
584 | if (device) { | |
585 | WARN(1, "Device %s already exists on group %d\n", | |
586 | dev_name(dev), iommu_group_id(iommu_group)); | |
587 | vfio_device_put(device); | |
588 | vfio_group_put(group); | |
589 | iommu_group_put(iommu_group); | |
590 | return -EBUSY; | |
591 | } | |
592 | ||
593 | device = vfio_group_create_device(group, dev, ops, device_data); | |
594 | if (IS_ERR(device)) { | |
595 | vfio_group_put(group); | |
596 | iommu_group_put(iommu_group); | |
597 | return PTR_ERR(device); | |
598 | } | |
599 | ||
600 | /* | |
601 | * Added device holds reference to iommu_group and vfio_device | |
602 | * (which in turn holds reference to vfio_group). Drop extra | |
603 | * group reference used while acquiring device. | |
604 | */ | |
605 | vfio_group_put(group); | |
606 | ||
607 | return 0; | |
608 | } | |
609 | EXPORT_SYMBOL_GPL(vfio_add_group_dev); | |
610 | ||
44f50716 VMP |
611 | /** |
612 | * Get a reference to the vfio_device for a device that is known to | |
613 | * be bound to a vfio driver. The driver implicitly holds a | |
614 | * vfio_device reference between vfio_add_group_dev and | |
615 | * vfio_del_group_dev. We can therefore use drvdata to increment | |
616 | * that reference from the struct device. This additional | |
617 | * reference must be released by calling vfio_device_put. | |
618 | */ | |
619 | struct vfio_device *vfio_device_get_from_dev(struct device *dev) | |
620 | { | |
621 | struct vfio_device *device = dev_get_drvdata(dev); | |
622 | ||
623 | vfio_device_get(device); | |
624 | ||
625 | return device; | |
626 | } | |
627 | EXPORT_SYMBOL_GPL(vfio_device_get_from_dev); | |
628 | ||
629 | /* | |
630 | * Caller must hold a reference to the vfio_device | |
631 | */ | |
632 | void *vfio_device_data(struct vfio_device *device) | |
633 | { | |
634 | return device->device_data; | |
635 | } | |
636 | EXPORT_SYMBOL_GPL(vfio_device_data); | |
637 | ||
e014e944 AW |
638 | /* Given a referenced group, check if it contains the device */ |
639 | static bool vfio_dev_present(struct vfio_group *group, struct device *dev) | |
cba3345c | 640 | { |
cba3345c AW |
641 | struct vfio_device *device; |
642 | ||
cba3345c | 643 | device = vfio_group_get_device(group, dev); |
e014e944 | 644 | if (!device) |
cba3345c | 645 | return false; |
cba3345c AW |
646 | |
647 | vfio_device_put(device); | |
cba3345c AW |
648 | return true; |
649 | } | |
650 | ||
651 | /* | |
652 | * Decrement the device reference count and wait for the device to be | |
653 | * removed. Open file descriptors for the device... */ | |
654 | void *vfio_del_group_dev(struct device *dev) | |
655 | { | |
656 | struct vfio_device *device = dev_get_drvdata(dev); | |
657 | struct vfio_group *group = device->group; | |
658 | struct iommu_group *iommu_group = group->iommu_group; | |
659 | void *device_data = device->device_data; | |
660 | ||
e014e944 AW |
661 | /* |
662 | * The group exists so long as we have a device reference. Get | |
663 | * a group reference and use it to scan for the device going away. | |
664 | */ | |
665 | vfio_group_get(group); | |
666 | ||
cba3345c AW |
667 | vfio_device_put(device); |
668 | ||
669 | /* TODO send a signal to encourage this to be released */ | |
e014e944 AW |
670 | wait_event(vfio.release_q, !vfio_dev_present(group, dev)); |
671 | ||
672 | vfio_group_put(group); | |
cba3345c AW |
673 | |
674 | iommu_group_put(iommu_group); | |
675 | ||
676 | return device_data; | |
677 | } | |
678 | EXPORT_SYMBOL_GPL(vfio_del_group_dev); | |
679 | ||
680 | /** | |
681 | * VFIO base fd, /dev/vfio/vfio | |
682 | */ | |
683 | static long vfio_ioctl_check_extension(struct vfio_container *container, | |
684 | unsigned long arg) | |
685 | { | |
0b43c082 | 686 | struct vfio_iommu_driver *driver; |
cba3345c AW |
687 | long ret = 0; |
688 | ||
0b43c082 AW |
689 | down_read(&container->group_lock); |
690 | ||
691 | driver = container->iommu_driver; | |
692 | ||
cba3345c AW |
693 | switch (arg) { |
694 | /* No base extensions yet */ | |
695 | default: | |
696 | /* | |
697 | * If no driver is set, poll all registered drivers for | |
698 | * extensions and return the first positive result. If | |
699 | * a driver is already set, further queries will be passed | |
700 | * only to that driver. | |
701 | */ | |
702 | if (!driver) { | |
703 | mutex_lock(&vfio.iommu_drivers_lock); | |
704 | list_for_each_entry(driver, &vfio.iommu_drivers_list, | |
705 | vfio_next) { | |
706 | if (!try_module_get(driver->ops->owner)) | |
707 | continue; | |
708 | ||
709 | ret = driver->ops->ioctl(NULL, | |
710 | VFIO_CHECK_EXTENSION, | |
711 | arg); | |
712 | module_put(driver->ops->owner); | |
713 | if (ret > 0) | |
714 | break; | |
715 | } | |
716 | mutex_unlock(&vfio.iommu_drivers_lock); | |
717 | } else | |
718 | ret = driver->ops->ioctl(container->iommu_data, | |
719 | VFIO_CHECK_EXTENSION, arg); | |
720 | } | |
721 | ||
0b43c082 AW |
722 | up_read(&container->group_lock); |
723 | ||
cba3345c AW |
724 | return ret; |
725 | } | |
726 | ||
9587f44a | 727 | /* hold write lock on container->group_lock */ |
cba3345c AW |
728 | static int __vfio_container_attach_groups(struct vfio_container *container, |
729 | struct vfio_iommu_driver *driver, | |
730 | void *data) | |
731 | { | |
732 | struct vfio_group *group; | |
733 | int ret = -ENODEV; | |
734 | ||
735 | list_for_each_entry(group, &container->group_list, container_next) { | |
736 | ret = driver->ops->attach_group(data, group->iommu_group); | |
737 | if (ret) | |
738 | goto unwind; | |
739 | } | |
740 | ||
741 | return ret; | |
742 | ||
743 | unwind: | |
744 | list_for_each_entry_continue_reverse(group, &container->group_list, | |
745 | container_next) { | |
746 | driver->ops->detach_group(data, group->iommu_group); | |
747 | } | |
748 | ||
749 | return ret; | |
750 | } | |
751 | ||
752 | static long vfio_ioctl_set_iommu(struct vfio_container *container, | |
753 | unsigned long arg) | |
754 | { | |
755 | struct vfio_iommu_driver *driver; | |
756 | long ret = -ENODEV; | |
757 | ||
9587f44a | 758 | down_write(&container->group_lock); |
cba3345c AW |
759 | |
760 | /* | |
761 | * The container is designed to be an unprivileged interface while | |
762 | * the group can be assigned to specific users. Therefore, only by | |
763 | * adding a group to a container does the user get the privilege of | |
764 | * enabling the iommu, which may allocate finite resources. There | |
765 | * is no unset_iommu, but by removing all the groups from a container, | |
766 | * the container is deprivileged and returns to an unset state. | |
767 | */ | |
768 | if (list_empty(&container->group_list) || container->iommu_driver) { | |
9587f44a | 769 | up_write(&container->group_lock); |
cba3345c AW |
770 | return -EINVAL; |
771 | } | |
772 | ||
773 | mutex_lock(&vfio.iommu_drivers_lock); | |
774 | list_for_each_entry(driver, &vfio.iommu_drivers_list, vfio_next) { | |
775 | void *data; | |
776 | ||
777 | if (!try_module_get(driver->ops->owner)) | |
778 | continue; | |
779 | ||
780 | /* | |
781 | * The arg magic for SET_IOMMU is the same as CHECK_EXTENSION, | |
782 | * so test which iommu driver reported support for this | |
783 | * extension and call open on them. We also pass them the | |
784 | * magic, allowing a single driver to support multiple | |
785 | * interfaces if they'd like. | |
786 | */ | |
787 | if (driver->ops->ioctl(NULL, VFIO_CHECK_EXTENSION, arg) <= 0) { | |
788 | module_put(driver->ops->owner); | |
789 | continue; | |
790 | } | |
791 | ||
792 | /* module reference holds the driver we're working on */ | |
793 | mutex_unlock(&vfio.iommu_drivers_lock); | |
794 | ||
795 | data = driver->ops->open(arg); | |
796 | if (IS_ERR(data)) { | |
797 | ret = PTR_ERR(data); | |
798 | module_put(driver->ops->owner); | |
799 | goto skip_drivers_unlock; | |
800 | } | |
801 | ||
802 | ret = __vfio_container_attach_groups(container, driver, data); | |
803 | if (!ret) { | |
804 | container->iommu_driver = driver; | |
805 | container->iommu_data = data; | |
806 | } else { | |
807 | driver->ops->release(data); | |
808 | module_put(driver->ops->owner); | |
809 | } | |
810 | ||
811 | goto skip_drivers_unlock; | |
812 | } | |
813 | ||
814 | mutex_unlock(&vfio.iommu_drivers_lock); | |
815 | skip_drivers_unlock: | |
9587f44a | 816 | up_write(&container->group_lock); |
cba3345c AW |
817 | |
818 | return ret; | |
819 | } | |
820 | ||
821 | static long vfio_fops_unl_ioctl(struct file *filep, | |
822 | unsigned int cmd, unsigned long arg) | |
823 | { | |
824 | struct vfio_container *container = filep->private_data; | |
825 | struct vfio_iommu_driver *driver; | |
826 | void *data; | |
827 | long ret = -EINVAL; | |
828 | ||
829 | if (!container) | |
830 | return ret; | |
831 | ||
cba3345c AW |
832 | switch (cmd) { |
833 | case VFIO_GET_API_VERSION: | |
834 | ret = VFIO_API_VERSION; | |
835 | break; | |
836 | case VFIO_CHECK_EXTENSION: | |
837 | ret = vfio_ioctl_check_extension(container, arg); | |
838 | break; | |
839 | case VFIO_SET_IOMMU: | |
840 | ret = vfio_ioctl_set_iommu(container, arg); | |
841 | break; | |
842 | default: | |
0b43c082 AW |
843 | down_read(&container->group_lock); |
844 | ||
845 | driver = container->iommu_driver; | |
846 | data = container->iommu_data; | |
847 | ||
cba3345c AW |
848 | if (driver) /* passthrough all unrecognized ioctls */ |
849 | ret = driver->ops->ioctl(data, cmd, arg); | |
0b43c082 AW |
850 | |
851 | up_read(&container->group_lock); | |
cba3345c AW |
852 | } |
853 | ||
854 | return ret; | |
855 | } | |
856 | ||
857 | #ifdef CONFIG_COMPAT | |
858 | static long vfio_fops_compat_ioctl(struct file *filep, | |
859 | unsigned int cmd, unsigned long arg) | |
860 | { | |
861 | arg = (unsigned long)compat_ptr(arg); | |
862 | return vfio_fops_unl_ioctl(filep, cmd, arg); | |
863 | } | |
864 | #endif /* CONFIG_COMPAT */ | |
865 | ||
866 | static int vfio_fops_open(struct inode *inode, struct file *filep) | |
867 | { | |
868 | struct vfio_container *container; | |
869 | ||
870 | container = kzalloc(sizeof(*container), GFP_KERNEL); | |
871 | if (!container) | |
872 | return -ENOMEM; | |
873 | ||
874 | INIT_LIST_HEAD(&container->group_list); | |
9587f44a | 875 | init_rwsem(&container->group_lock); |
cba3345c AW |
876 | kref_init(&container->kref); |
877 | ||
878 | filep->private_data = container; | |
879 | ||
880 | return 0; | |
881 | } | |
882 | ||
883 | static int vfio_fops_release(struct inode *inode, struct file *filep) | |
884 | { | |
885 | struct vfio_container *container = filep->private_data; | |
886 | ||
887 | filep->private_data = NULL; | |
888 | ||
889 | vfio_container_put(container); | |
890 | ||
891 | return 0; | |
892 | } | |
893 | ||
894 | /* | |
895 | * Once an iommu driver is set, we optionally pass read/write/mmap | |
896 | * on to the driver, allowing management interfaces beyond ioctl. | |
897 | */ | |
898 | static ssize_t vfio_fops_read(struct file *filep, char __user *buf, | |
899 | size_t count, loff_t *ppos) | |
900 | { | |
901 | struct vfio_container *container = filep->private_data; | |
0b43c082 AW |
902 | struct vfio_iommu_driver *driver; |
903 | ssize_t ret = -EINVAL; | |
cba3345c | 904 | |
0b43c082 AW |
905 | down_read(&container->group_lock); |
906 | ||
907 | driver = container->iommu_driver; | |
908 | if (likely(driver && driver->ops->read)) | |
909 | ret = driver->ops->read(container->iommu_data, | |
910 | buf, count, ppos); | |
cba3345c | 911 | |
0b43c082 AW |
912 | up_read(&container->group_lock); |
913 | ||
914 | return ret; | |
cba3345c AW |
915 | } |
916 | ||
917 | static ssize_t vfio_fops_write(struct file *filep, const char __user *buf, | |
918 | size_t count, loff_t *ppos) | |
919 | { | |
920 | struct vfio_container *container = filep->private_data; | |
0b43c082 AW |
921 | struct vfio_iommu_driver *driver; |
922 | ssize_t ret = -EINVAL; | |
cba3345c | 923 | |
0b43c082 AW |
924 | down_read(&container->group_lock); |
925 | ||
926 | driver = container->iommu_driver; | |
927 | if (likely(driver && driver->ops->write)) | |
928 | ret = driver->ops->write(container->iommu_data, | |
929 | buf, count, ppos); | |
930 | ||
931 | up_read(&container->group_lock); | |
cba3345c | 932 | |
0b43c082 | 933 | return ret; |
cba3345c AW |
934 | } |
935 | ||
936 | static int vfio_fops_mmap(struct file *filep, struct vm_area_struct *vma) | |
937 | { | |
938 | struct vfio_container *container = filep->private_data; | |
0b43c082 AW |
939 | struct vfio_iommu_driver *driver; |
940 | int ret = -EINVAL; | |
cba3345c | 941 | |
0b43c082 | 942 | down_read(&container->group_lock); |
cba3345c | 943 | |
0b43c082 AW |
944 | driver = container->iommu_driver; |
945 | if (likely(driver && driver->ops->mmap)) | |
946 | ret = driver->ops->mmap(container->iommu_data, vma); | |
947 | ||
948 | up_read(&container->group_lock); | |
949 | ||
950 | return ret; | |
cba3345c AW |
951 | } |
952 | ||
953 | static const struct file_operations vfio_fops = { | |
954 | .owner = THIS_MODULE, | |
955 | .open = vfio_fops_open, | |
956 | .release = vfio_fops_release, | |
957 | .read = vfio_fops_read, | |
958 | .write = vfio_fops_write, | |
959 | .unlocked_ioctl = vfio_fops_unl_ioctl, | |
960 | #ifdef CONFIG_COMPAT | |
961 | .compat_ioctl = vfio_fops_compat_ioctl, | |
962 | #endif | |
963 | .mmap = vfio_fops_mmap, | |
964 | }; | |
965 | ||
966 | /** | |
967 | * VFIO Group fd, /dev/vfio/$GROUP | |
968 | */ | |
969 | static void __vfio_group_unset_container(struct vfio_group *group) | |
970 | { | |
971 | struct vfio_container *container = group->container; | |
972 | struct vfio_iommu_driver *driver; | |
973 | ||
9587f44a | 974 | down_write(&container->group_lock); |
cba3345c AW |
975 | |
976 | driver = container->iommu_driver; | |
977 | if (driver) | |
978 | driver->ops->detach_group(container->iommu_data, | |
979 | group->iommu_group); | |
980 | ||
981 | group->container = NULL; | |
982 | list_del(&group->container_next); | |
983 | ||
984 | /* Detaching the last group deprivileges a container, remove iommu */ | |
985 | if (driver && list_empty(&container->group_list)) { | |
986 | driver->ops->release(container->iommu_data); | |
987 | module_put(driver->ops->owner); | |
988 | container->iommu_driver = NULL; | |
989 | container->iommu_data = NULL; | |
990 | } | |
991 | ||
9587f44a | 992 | up_write(&container->group_lock); |
cba3345c AW |
993 | |
994 | vfio_container_put(container); | |
995 | } | |
996 | ||
997 | /* | |
998 | * VFIO_GROUP_UNSET_CONTAINER should fail if there are other users or | |
999 | * if there was no container to unset. Since the ioctl is called on | |
1000 | * the group, we know that still exists, therefore the only valid | |
1001 | * transition here is 1->0. | |
1002 | */ | |
1003 | static int vfio_group_unset_container(struct vfio_group *group) | |
1004 | { | |
1005 | int users = atomic_cmpxchg(&group->container_users, 1, 0); | |
1006 | ||
1007 | if (!users) | |
1008 | return -EINVAL; | |
1009 | if (users != 1) | |
1010 | return -EBUSY; | |
1011 | ||
1012 | __vfio_group_unset_container(group); | |
1013 | ||
1014 | return 0; | |
1015 | } | |
1016 | ||
1017 | /* | |
1018 | * When removing container users, anything that removes the last user | |
1019 | * implicitly removes the group from the container. That is, if the | |
1020 | * group file descriptor is closed, as well as any device file descriptors, | |
1021 | * the group is free. | |
1022 | */ | |
1023 | static void vfio_group_try_dissolve_container(struct vfio_group *group) | |
1024 | { | |
1025 | if (0 == atomic_dec_if_positive(&group->container_users)) | |
1026 | __vfio_group_unset_container(group); | |
1027 | } | |
1028 | ||
1029 | static int vfio_group_set_container(struct vfio_group *group, int container_fd) | |
1030 | { | |
2903ff01 | 1031 | struct fd f; |
cba3345c AW |
1032 | struct vfio_container *container; |
1033 | struct vfio_iommu_driver *driver; | |
2903ff01 | 1034 | int ret = 0; |
cba3345c AW |
1035 | |
1036 | if (atomic_read(&group->container_users)) | |
1037 | return -EINVAL; | |
1038 | ||
2903ff01 AV |
1039 | f = fdget(container_fd); |
1040 | if (!f.file) | |
cba3345c AW |
1041 | return -EBADF; |
1042 | ||
1043 | /* Sanity check, is this really our fd? */ | |
2903ff01 AV |
1044 | if (f.file->f_op != &vfio_fops) { |
1045 | fdput(f); | |
cba3345c AW |
1046 | return -EINVAL; |
1047 | } | |
1048 | ||
2903ff01 | 1049 | container = f.file->private_data; |
cba3345c AW |
1050 | WARN_ON(!container); /* fget ensures we don't race vfio_release */ |
1051 | ||
9587f44a | 1052 | down_write(&container->group_lock); |
cba3345c AW |
1053 | |
1054 | driver = container->iommu_driver; | |
1055 | if (driver) { | |
1056 | ret = driver->ops->attach_group(container->iommu_data, | |
1057 | group->iommu_group); | |
1058 | if (ret) | |
1059 | goto unlock_out; | |
1060 | } | |
1061 | ||
1062 | group->container = container; | |
1063 | list_add(&group->container_next, &container->group_list); | |
1064 | ||
1065 | /* Get a reference on the container and mark a user within the group */ | |
1066 | vfio_container_get(container); | |
1067 | atomic_inc(&group->container_users); | |
1068 | ||
1069 | unlock_out: | |
9587f44a | 1070 | up_write(&container->group_lock); |
2903ff01 | 1071 | fdput(f); |
cba3345c AW |
1072 | return ret; |
1073 | } | |
1074 | ||
1075 | static bool vfio_group_viable(struct vfio_group *group) | |
1076 | { | |
1077 | return (iommu_group_for_each_dev(group->iommu_group, | |
1078 | group, vfio_dev_viable) == 0); | |
1079 | } | |
1080 | ||
1081 | static const struct file_operations vfio_device_fops; | |
1082 | ||
1083 | static int vfio_group_get_device_fd(struct vfio_group *group, char *buf) | |
1084 | { | |
1085 | struct vfio_device *device; | |
1086 | struct file *filep; | |
1087 | int ret = -ENODEV; | |
1088 | ||
1089 | if (0 == atomic_read(&group->container_users) || | |
1090 | !group->container->iommu_driver || !vfio_group_viable(group)) | |
1091 | return -EINVAL; | |
1092 | ||
1093 | mutex_lock(&group->device_lock); | |
1094 | list_for_each_entry(device, &group->device_list, group_next) { | |
1095 | if (strcmp(dev_name(device->dev), buf)) | |
1096 | continue; | |
1097 | ||
1098 | ret = device->ops->open(device->device_data); | |
1099 | if (ret) | |
1100 | break; | |
1101 | /* | |
1102 | * We can't use anon_inode_getfd() because we need to modify | |
1103 | * the f_mode flags directly to allow more than just ioctls | |
1104 | */ | |
5d042fbd | 1105 | ret = get_unused_fd_flags(O_CLOEXEC); |
cba3345c AW |
1106 | if (ret < 0) { |
1107 | device->ops->release(device->device_data); | |
1108 | break; | |
1109 | } | |
1110 | ||
1111 | filep = anon_inode_getfile("[vfio-device]", &vfio_device_fops, | |
1112 | device, O_RDWR); | |
1113 | if (IS_ERR(filep)) { | |
1114 | put_unused_fd(ret); | |
1115 | ret = PTR_ERR(filep); | |
1116 | device->ops->release(device->device_data); | |
1117 | break; | |
1118 | } | |
1119 | ||
1120 | /* | |
1121 | * TODO: add an anon_inode interface to do this. | |
1122 | * Appears to be missing by lack of need rather than | |
1123 | * explicitly prevented. Now there's need. | |
1124 | */ | |
1125 | filep->f_mode |= (FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE); | |
1126 | ||
cba3345c AW |
1127 | vfio_device_get(device); |
1128 | atomic_inc(&group->container_users); | |
31605deb AV |
1129 | |
1130 | fd_install(ret, filep); | |
cba3345c AW |
1131 | break; |
1132 | } | |
1133 | mutex_unlock(&group->device_lock); | |
1134 | ||
1135 | return ret; | |
1136 | } | |
1137 | ||
1138 | static long vfio_group_fops_unl_ioctl(struct file *filep, | |
1139 | unsigned int cmd, unsigned long arg) | |
1140 | { | |
1141 | struct vfio_group *group = filep->private_data; | |
1142 | long ret = -ENOTTY; | |
1143 | ||
1144 | switch (cmd) { | |
1145 | case VFIO_GROUP_GET_STATUS: | |
1146 | { | |
1147 | struct vfio_group_status status; | |
1148 | unsigned long minsz; | |
1149 | ||
1150 | minsz = offsetofend(struct vfio_group_status, flags); | |
1151 | ||
1152 | if (copy_from_user(&status, (void __user *)arg, minsz)) | |
1153 | return -EFAULT; | |
1154 | ||
1155 | if (status.argsz < minsz) | |
1156 | return -EINVAL; | |
1157 | ||
1158 | status.flags = 0; | |
1159 | ||
1160 | if (vfio_group_viable(group)) | |
1161 | status.flags |= VFIO_GROUP_FLAGS_VIABLE; | |
1162 | ||
1163 | if (group->container) | |
1164 | status.flags |= VFIO_GROUP_FLAGS_CONTAINER_SET; | |
1165 | ||
1166 | if (copy_to_user((void __user *)arg, &status, minsz)) | |
1167 | return -EFAULT; | |
1168 | ||
1169 | ret = 0; | |
1170 | break; | |
1171 | } | |
1172 | case VFIO_GROUP_SET_CONTAINER: | |
1173 | { | |
1174 | int fd; | |
1175 | ||
1176 | if (get_user(fd, (int __user *)arg)) | |
1177 | return -EFAULT; | |
1178 | ||
1179 | if (fd < 0) | |
1180 | return -EINVAL; | |
1181 | ||
1182 | ret = vfio_group_set_container(group, fd); | |
1183 | break; | |
1184 | } | |
1185 | case VFIO_GROUP_UNSET_CONTAINER: | |
1186 | ret = vfio_group_unset_container(group); | |
1187 | break; | |
1188 | case VFIO_GROUP_GET_DEVICE_FD: | |
1189 | { | |
1190 | char *buf; | |
1191 | ||
1192 | buf = strndup_user((const char __user *)arg, PAGE_SIZE); | |
1193 | if (IS_ERR(buf)) | |
1194 | return PTR_ERR(buf); | |
1195 | ||
1196 | ret = vfio_group_get_device_fd(group, buf); | |
1197 | kfree(buf); | |
1198 | break; | |
1199 | } | |
1200 | } | |
1201 | ||
1202 | return ret; | |
1203 | } | |
1204 | ||
1205 | #ifdef CONFIG_COMPAT | |
1206 | static long vfio_group_fops_compat_ioctl(struct file *filep, | |
1207 | unsigned int cmd, unsigned long arg) | |
1208 | { | |
1209 | arg = (unsigned long)compat_ptr(arg); | |
1210 | return vfio_group_fops_unl_ioctl(filep, cmd, arg); | |
1211 | } | |
1212 | #endif /* CONFIG_COMPAT */ | |
1213 | ||
1214 | static int vfio_group_fops_open(struct inode *inode, struct file *filep) | |
1215 | { | |
1216 | struct vfio_group *group; | |
6d6768c6 | 1217 | int opened; |
cba3345c AW |
1218 | |
1219 | group = vfio_group_get_from_minor(iminor(inode)); | |
1220 | if (!group) | |
1221 | return -ENODEV; | |
1222 | ||
6d6768c6 AW |
1223 | /* Do we need multiple instances of the group open? Seems not. */ |
1224 | opened = atomic_cmpxchg(&group->opened, 0, 1); | |
1225 | if (opened) { | |
1226 | vfio_group_put(group); | |
1227 | return -EBUSY; | |
1228 | } | |
1229 | ||
1230 | /* Is something still in use from a previous open? */ | |
cba3345c | 1231 | if (group->container) { |
6d6768c6 | 1232 | atomic_dec(&group->opened); |
cba3345c AW |
1233 | vfio_group_put(group); |
1234 | return -EBUSY; | |
1235 | } | |
1236 | ||
1237 | filep->private_data = group; | |
1238 | ||
1239 | return 0; | |
1240 | } | |
1241 | ||
1242 | static int vfio_group_fops_release(struct inode *inode, struct file *filep) | |
1243 | { | |
1244 | struct vfio_group *group = filep->private_data; | |
1245 | ||
1246 | filep->private_data = NULL; | |
1247 | ||
1248 | vfio_group_try_dissolve_container(group); | |
1249 | ||
6d6768c6 AW |
1250 | atomic_dec(&group->opened); |
1251 | ||
cba3345c AW |
1252 | vfio_group_put(group); |
1253 | ||
1254 | return 0; | |
1255 | } | |
1256 | ||
1257 | static const struct file_operations vfio_group_fops = { | |
1258 | .owner = THIS_MODULE, | |
1259 | .unlocked_ioctl = vfio_group_fops_unl_ioctl, | |
1260 | #ifdef CONFIG_COMPAT | |
1261 | .compat_ioctl = vfio_group_fops_compat_ioctl, | |
1262 | #endif | |
1263 | .open = vfio_group_fops_open, | |
1264 | .release = vfio_group_fops_release, | |
1265 | }; | |
1266 | ||
1267 | /** | |
1268 | * VFIO Device fd | |
1269 | */ | |
1270 | static int vfio_device_fops_release(struct inode *inode, struct file *filep) | |
1271 | { | |
1272 | struct vfio_device *device = filep->private_data; | |
1273 | ||
1274 | device->ops->release(device->device_data); | |
1275 | ||
1276 | vfio_group_try_dissolve_container(device->group); | |
1277 | ||
1278 | vfio_device_put(device); | |
1279 | ||
1280 | return 0; | |
1281 | } | |
1282 | ||
1283 | static long vfio_device_fops_unl_ioctl(struct file *filep, | |
1284 | unsigned int cmd, unsigned long arg) | |
1285 | { | |
1286 | struct vfio_device *device = filep->private_data; | |
1287 | ||
1288 | if (unlikely(!device->ops->ioctl)) | |
1289 | return -EINVAL; | |
1290 | ||
1291 | return device->ops->ioctl(device->device_data, cmd, arg); | |
1292 | } | |
1293 | ||
1294 | static ssize_t vfio_device_fops_read(struct file *filep, char __user *buf, | |
1295 | size_t count, loff_t *ppos) | |
1296 | { | |
1297 | struct vfio_device *device = filep->private_data; | |
1298 | ||
1299 | if (unlikely(!device->ops->read)) | |
1300 | return -EINVAL; | |
1301 | ||
1302 | return device->ops->read(device->device_data, buf, count, ppos); | |
1303 | } | |
1304 | ||
1305 | static ssize_t vfio_device_fops_write(struct file *filep, | |
1306 | const char __user *buf, | |
1307 | size_t count, loff_t *ppos) | |
1308 | { | |
1309 | struct vfio_device *device = filep->private_data; | |
1310 | ||
1311 | if (unlikely(!device->ops->write)) | |
1312 | return -EINVAL; | |
1313 | ||
1314 | return device->ops->write(device->device_data, buf, count, ppos); | |
1315 | } | |
1316 | ||
1317 | static int vfio_device_fops_mmap(struct file *filep, struct vm_area_struct *vma) | |
1318 | { | |
1319 | struct vfio_device *device = filep->private_data; | |
1320 | ||
1321 | if (unlikely(!device->ops->mmap)) | |
1322 | return -EINVAL; | |
1323 | ||
1324 | return device->ops->mmap(device->device_data, vma); | |
1325 | } | |
1326 | ||
1327 | #ifdef CONFIG_COMPAT | |
1328 | static long vfio_device_fops_compat_ioctl(struct file *filep, | |
1329 | unsigned int cmd, unsigned long arg) | |
1330 | { | |
1331 | arg = (unsigned long)compat_ptr(arg); | |
1332 | return vfio_device_fops_unl_ioctl(filep, cmd, arg); | |
1333 | } | |
1334 | #endif /* CONFIG_COMPAT */ | |
1335 | ||
1336 | static const struct file_operations vfio_device_fops = { | |
1337 | .owner = THIS_MODULE, | |
1338 | .release = vfio_device_fops_release, | |
1339 | .read = vfio_device_fops_read, | |
1340 | .write = vfio_device_fops_write, | |
1341 | .unlocked_ioctl = vfio_device_fops_unl_ioctl, | |
1342 | #ifdef CONFIG_COMPAT | |
1343 | .compat_ioctl = vfio_device_fops_compat_ioctl, | |
1344 | #endif | |
1345 | .mmap = vfio_device_fops_mmap, | |
1346 | }; | |
1347 | ||
6cdd9782 AK |
1348 | /** |
1349 | * External user API, exported by symbols to be linked dynamically. | |
1350 | * | |
1351 | * The protocol includes: | |
1352 | * 1. do normal VFIO init operation: | |
1353 | * - opening a new container; | |
1354 | * - attaching group(s) to it; | |
1355 | * - setting an IOMMU driver for a container. | |
1356 | * When IOMMU is set for a container, all groups in it are | |
1357 | * considered ready to use by an external user. | |
1358 | * | |
1359 | * 2. User space passes a group fd to an external user. | |
1360 | * The external user calls vfio_group_get_external_user() | |
1361 | * to verify that: | |
1362 | * - the group is initialized; | |
1363 | * - IOMMU is set for it. | |
1364 | * If both checks passed, vfio_group_get_external_user() | |
1365 | * increments the container user counter to prevent | |
1366 | * the VFIO group from disposal before KVM exits. | |
1367 | * | |
1368 | * 3. The external user calls vfio_external_user_iommu_id() | |
1369 | * to know an IOMMU ID. | |
1370 | * | |
1371 | * 4. When the external KVM finishes, it calls | |
1372 | * vfio_group_put_external_user() to release the VFIO group. | |
1373 | * This call decrements the container user counter. | |
1374 | */ | |
1375 | struct vfio_group *vfio_group_get_external_user(struct file *filep) | |
1376 | { | |
1377 | struct vfio_group *group = filep->private_data; | |
1378 | ||
1379 | if (filep->f_op != &vfio_group_fops) | |
1380 | return ERR_PTR(-EINVAL); | |
1381 | ||
1382 | if (!atomic_inc_not_zero(&group->container_users)) | |
1383 | return ERR_PTR(-EINVAL); | |
1384 | ||
1385 | if (!group->container->iommu_driver || | |
1386 | !vfio_group_viable(group)) { | |
1387 | atomic_dec(&group->container_users); | |
1388 | return ERR_PTR(-EINVAL); | |
1389 | } | |
1390 | ||
1391 | vfio_group_get(group); | |
1392 | ||
1393 | return group; | |
1394 | } | |
1395 | EXPORT_SYMBOL_GPL(vfio_group_get_external_user); | |
1396 | ||
1397 | void vfio_group_put_external_user(struct vfio_group *group) | |
1398 | { | |
1399 | vfio_group_put(group); | |
1400 | vfio_group_try_dissolve_container(group); | |
1401 | } | |
1402 | EXPORT_SYMBOL_GPL(vfio_group_put_external_user); | |
1403 | ||
1404 | int vfio_external_user_iommu_id(struct vfio_group *group) | |
1405 | { | |
1406 | return iommu_group_id(group->iommu_group); | |
1407 | } | |
1408 | EXPORT_SYMBOL_GPL(vfio_external_user_iommu_id); | |
1409 | ||
88d7ab89 AW |
1410 | long vfio_external_check_extension(struct vfio_group *group, unsigned long arg) |
1411 | { | |
1412 | return vfio_ioctl_check_extension(group->container, arg); | |
1413 | } | |
1414 | EXPORT_SYMBOL_GPL(vfio_external_check_extension); | |
1415 | ||
cba3345c AW |
1416 | /** |
1417 | * Module/class support | |
1418 | */ | |
1419 | static char *vfio_devnode(struct device *dev, umode_t *mode) | |
1420 | { | |
1421 | return kasprintf(GFP_KERNEL, "vfio/%s", dev_name(dev)); | |
1422 | } | |
1423 | ||
d1099901 AW |
1424 | static struct miscdevice vfio_dev = { |
1425 | .minor = VFIO_MINOR, | |
1426 | .name = "vfio", | |
1427 | .fops = &vfio_fops, | |
1428 | .nodename = "vfio/vfio", | |
1429 | .mode = S_IRUGO | S_IWUGO, | |
1430 | }; | |
1431 | ||
cba3345c AW |
1432 | static int __init vfio_init(void) |
1433 | { | |
1434 | int ret; | |
1435 | ||
1436 | idr_init(&vfio.group_idr); | |
1437 | mutex_init(&vfio.group_lock); | |
1438 | mutex_init(&vfio.iommu_drivers_lock); | |
1439 | INIT_LIST_HEAD(&vfio.group_list); | |
1440 | INIT_LIST_HEAD(&vfio.iommu_drivers_list); | |
1441 | init_waitqueue_head(&vfio.release_q); | |
1442 | ||
d1099901 AW |
1443 | ret = misc_register(&vfio_dev); |
1444 | if (ret) { | |
1445 | pr_err("vfio: misc device register failed\n"); | |
1446 | return ret; | |
1447 | } | |
1448 | ||
1449 | /* /dev/vfio/$GROUP */ | |
cba3345c AW |
1450 | vfio.class = class_create(THIS_MODULE, "vfio"); |
1451 | if (IS_ERR(vfio.class)) { | |
1452 | ret = PTR_ERR(vfio.class); | |
1453 | goto err_class; | |
1454 | } | |
1455 | ||
1456 | vfio.class->devnode = vfio_devnode; | |
1457 | ||
d1099901 | 1458 | ret = alloc_chrdev_region(&vfio.group_devt, 0, MINORMASK, "vfio"); |
cba3345c | 1459 | if (ret) |
d1099901 | 1460 | goto err_alloc_chrdev; |
cba3345c | 1461 | |
cba3345c | 1462 | cdev_init(&vfio.group_cdev, &vfio_group_fops); |
d1099901 | 1463 | ret = cdev_add(&vfio.group_cdev, vfio.group_devt, MINORMASK); |
cba3345c | 1464 | if (ret) |
d1099901 | 1465 | goto err_cdev_add; |
cba3345c AW |
1466 | |
1467 | pr_info(DRIVER_DESC " version: " DRIVER_VERSION "\n"); | |
1468 | ||
73fa0d10 AW |
1469 | /* |
1470 | * Attempt to load known iommu-drivers. This gives us a working | |
1471 | * environment without the user needing to explicitly load iommu | |
1472 | * drivers. | |
1473 | */ | |
1474 | request_module_nowait("vfio_iommu_type1"); | |
5ffd229c | 1475 | request_module_nowait("vfio_iommu_spapr_tce"); |
73fa0d10 | 1476 | |
cba3345c AW |
1477 | return 0; |
1478 | ||
d1099901 AW |
1479 | err_cdev_add: |
1480 | unregister_chrdev_region(vfio.group_devt, MINORMASK); | |
1481 | err_alloc_chrdev: | |
cba3345c AW |
1482 | class_destroy(vfio.class); |
1483 | vfio.class = NULL; | |
1484 | err_class: | |
d1099901 | 1485 | misc_deregister(&vfio_dev); |
cba3345c AW |
1486 | return ret; |
1487 | } | |
1488 | ||
1489 | static void __exit vfio_cleanup(void) | |
1490 | { | |
1491 | WARN_ON(!list_empty(&vfio.group_list)); | |
1492 | ||
1493 | idr_destroy(&vfio.group_idr); | |
1494 | cdev_del(&vfio.group_cdev); | |
d1099901 | 1495 | unregister_chrdev_region(vfio.group_devt, MINORMASK); |
cba3345c AW |
1496 | class_destroy(vfio.class); |
1497 | vfio.class = NULL; | |
d1099901 | 1498 | misc_deregister(&vfio_dev); |
cba3345c AW |
1499 | } |
1500 | ||
1501 | module_init(vfio_init); | |
1502 | module_exit(vfio_cleanup); | |
1503 | ||
1504 | MODULE_VERSION(DRIVER_VERSION); | |
1505 | MODULE_LICENSE("GPL v2"); | |
1506 | MODULE_AUTHOR(DRIVER_AUTHOR); | |
1507 | MODULE_DESCRIPTION(DRIVER_DESC); | |
d1099901 AW |
1508 | MODULE_ALIAS_MISCDEV(VFIO_MINOR); |
1509 | MODULE_ALIAS("devname:vfio/vfio"); |