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 | ||
60720a0f AW |
66 | struct vfio_unbound_dev { |
67 | struct device *dev; | |
68 | struct list_head unbound_next; | |
69 | }; | |
70 | ||
cba3345c AW |
71 | struct vfio_group { |
72 | struct kref kref; | |
73 | int minor; | |
74 | atomic_t container_users; | |
75 | struct iommu_group *iommu_group; | |
76 | struct vfio_container *container; | |
77 | struct list_head device_list; | |
78 | struct mutex device_lock; | |
79 | struct device *dev; | |
80 | struct notifier_block nb; | |
81 | struct list_head vfio_next; | |
82 | struct list_head container_next; | |
60720a0f AW |
83 | struct list_head unbound_list; |
84 | struct mutex unbound_lock; | |
6d6768c6 | 85 | atomic_t opened; |
cba3345c AW |
86 | }; |
87 | ||
88 | struct vfio_device { | |
89 | struct kref kref; | |
90 | struct device *dev; | |
91 | const struct vfio_device_ops *ops; | |
92 | struct vfio_group *group; | |
93 | struct list_head group_next; | |
94 | void *device_data; | |
95 | }; | |
96 | ||
97 | /** | |
98 | * IOMMU driver registration | |
99 | */ | |
100 | int vfio_register_iommu_driver(const struct vfio_iommu_driver_ops *ops) | |
101 | { | |
102 | struct vfio_iommu_driver *driver, *tmp; | |
103 | ||
104 | driver = kzalloc(sizeof(*driver), GFP_KERNEL); | |
105 | if (!driver) | |
106 | return -ENOMEM; | |
107 | ||
108 | driver->ops = ops; | |
109 | ||
110 | mutex_lock(&vfio.iommu_drivers_lock); | |
111 | ||
112 | /* Check for duplicates */ | |
113 | list_for_each_entry(tmp, &vfio.iommu_drivers_list, vfio_next) { | |
114 | if (tmp->ops == ops) { | |
115 | mutex_unlock(&vfio.iommu_drivers_lock); | |
116 | kfree(driver); | |
117 | return -EINVAL; | |
118 | } | |
119 | } | |
120 | ||
121 | list_add(&driver->vfio_next, &vfio.iommu_drivers_list); | |
122 | ||
123 | mutex_unlock(&vfio.iommu_drivers_lock); | |
124 | ||
125 | return 0; | |
126 | } | |
127 | EXPORT_SYMBOL_GPL(vfio_register_iommu_driver); | |
128 | ||
129 | void vfio_unregister_iommu_driver(const struct vfio_iommu_driver_ops *ops) | |
130 | { | |
131 | struct vfio_iommu_driver *driver; | |
132 | ||
133 | mutex_lock(&vfio.iommu_drivers_lock); | |
134 | list_for_each_entry(driver, &vfio.iommu_drivers_list, vfio_next) { | |
135 | if (driver->ops == ops) { | |
136 | list_del(&driver->vfio_next); | |
137 | mutex_unlock(&vfio.iommu_drivers_lock); | |
138 | kfree(driver); | |
139 | return; | |
140 | } | |
141 | } | |
142 | mutex_unlock(&vfio.iommu_drivers_lock); | |
143 | } | |
144 | EXPORT_SYMBOL_GPL(vfio_unregister_iommu_driver); | |
145 | ||
146 | /** | |
147 | * Group minor allocation/free - both called with vfio.group_lock held | |
148 | */ | |
149 | static int vfio_alloc_group_minor(struct vfio_group *group) | |
150 | { | |
d1099901 | 151 | return idr_alloc(&vfio.group_idr, group, 0, MINORMASK + 1, GFP_KERNEL); |
cba3345c AW |
152 | } |
153 | ||
154 | static void vfio_free_group_minor(int minor) | |
155 | { | |
156 | idr_remove(&vfio.group_idr, minor); | |
157 | } | |
158 | ||
159 | static int vfio_iommu_group_notifier(struct notifier_block *nb, | |
160 | unsigned long action, void *data); | |
161 | static void vfio_group_get(struct vfio_group *group); | |
162 | ||
163 | /** | |
164 | * Container objects - containers are created when /dev/vfio/vfio is | |
165 | * opened, but their lifecycle extends until the last user is done, so | |
166 | * it's freed via kref. Must support container/group/device being | |
167 | * closed in any order. | |
168 | */ | |
169 | static void vfio_container_get(struct vfio_container *container) | |
170 | { | |
171 | kref_get(&container->kref); | |
172 | } | |
173 | ||
174 | static void vfio_container_release(struct kref *kref) | |
175 | { | |
176 | struct vfio_container *container; | |
177 | container = container_of(kref, struct vfio_container, kref); | |
178 | ||
179 | kfree(container); | |
180 | } | |
181 | ||
182 | static void vfio_container_put(struct vfio_container *container) | |
183 | { | |
184 | kref_put(&container->kref, vfio_container_release); | |
185 | } | |
186 | ||
9df7b25a JL |
187 | static void vfio_group_unlock_and_free(struct vfio_group *group) |
188 | { | |
189 | mutex_unlock(&vfio.group_lock); | |
190 | /* | |
191 | * Unregister outside of lock. A spurious callback is harmless now | |
192 | * that the group is no longer in vfio.group_list. | |
193 | */ | |
194 | iommu_group_unregister_notifier(group->iommu_group, &group->nb); | |
195 | kfree(group); | |
196 | } | |
197 | ||
cba3345c AW |
198 | /** |
199 | * Group objects - create, release, get, put, search | |
200 | */ | |
201 | static struct vfio_group *vfio_create_group(struct iommu_group *iommu_group) | |
202 | { | |
203 | struct vfio_group *group, *tmp; | |
204 | struct device *dev; | |
205 | int ret, minor; | |
206 | ||
207 | group = kzalloc(sizeof(*group), GFP_KERNEL); | |
208 | if (!group) | |
209 | return ERR_PTR(-ENOMEM); | |
210 | ||
211 | kref_init(&group->kref); | |
212 | INIT_LIST_HEAD(&group->device_list); | |
213 | mutex_init(&group->device_lock); | |
60720a0f AW |
214 | INIT_LIST_HEAD(&group->unbound_list); |
215 | mutex_init(&group->unbound_lock); | |
cba3345c | 216 | atomic_set(&group->container_users, 0); |
6d6768c6 | 217 | atomic_set(&group->opened, 0); |
cba3345c AW |
218 | group->iommu_group = iommu_group; |
219 | ||
220 | group->nb.notifier_call = vfio_iommu_group_notifier; | |
221 | ||
222 | /* | |
223 | * blocking notifiers acquire a rwsem around registering and hold | |
224 | * it around callback. Therefore, need to register outside of | |
225 | * vfio.group_lock to avoid A-B/B-A contention. Our callback won't | |
226 | * do anything unless it can find the group in vfio.group_list, so | |
227 | * no harm in registering early. | |
228 | */ | |
229 | ret = iommu_group_register_notifier(iommu_group, &group->nb); | |
230 | if (ret) { | |
231 | kfree(group); | |
232 | return ERR_PTR(ret); | |
233 | } | |
234 | ||
235 | mutex_lock(&vfio.group_lock); | |
236 | ||
cba3345c AW |
237 | /* Did we race creating this group? */ |
238 | list_for_each_entry(tmp, &vfio.group_list, vfio_next) { | |
239 | if (tmp->iommu_group == iommu_group) { | |
240 | vfio_group_get(tmp); | |
9df7b25a | 241 | vfio_group_unlock_and_free(group); |
cba3345c AW |
242 | return tmp; |
243 | } | |
244 | } | |
245 | ||
2f51bf4b ZL |
246 | minor = vfio_alloc_group_minor(group); |
247 | if (minor < 0) { | |
248 | vfio_group_unlock_and_free(group); | |
249 | return ERR_PTR(minor); | |
250 | } | |
251 | ||
d1099901 AW |
252 | dev = device_create(vfio.class, NULL, |
253 | MKDEV(MAJOR(vfio.group_devt), minor), | |
cba3345c AW |
254 | group, "%d", iommu_group_id(iommu_group)); |
255 | if (IS_ERR(dev)) { | |
256 | vfio_free_group_minor(minor); | |
9df7b25a | 257 | vfio_group_unlock_and_free(group); |
cba3345c AW |
258 | return (struct vfio_group *)dev; /* ERR_PTR */ |
259 | } | |
260 | ||
261 | group->minor = minor; | |
262 | group->dev = dev; | |
263 | ||
264 | list_add(&group->vfio_next, &vfio.group_list); | |
265 | ||
266 | mutex_unlock(&vfio.group_lock); | |
267 | ||
268 | return group; | |
269 | } | |
270 | ||
6d2cd3ce | 271 | /* called with vfio.group_lock held */ |
cba3345c AW |
272 | static void vfio_group_release(struct kref *kref) |
273 | { | |
274 | struct vfio_group *group = container_of(kref, struct vfio_group, kref); | |
60720a0f | 275 | struct vfio_unbound_dev *unbound, *tmp; |
4a68810d | 276 | struct iommu_group *iommu_group = group->iommu_group; |
cba3345c AW |
277 | |
278 | WARN_ON(!list_empty(&group->device_list)); | |
279 | ||
60720a0f AW |
280 | list_for_each_entry_safe(unbound, tmp, |
281 | &group->unbound_list, unbound_next) { | |
282 | list_del(&unbound->unbound_next); | |
283 | kfree(unbound); | |
284 | } | |
285 | ||
d1099901 | 286 | device_destroy(vfio.class, MKDEV(MAJOR(vfio.group_devt), group->minor)); |
cba3345c AW |
287 | list_del(&group->vfio_next); |
288 | vfio_free_group_minor(group->minor); | |
9df7b25a | 289 | vfio_group_unlock_and_free(group); |
4a68810d | 290 | iommu_group_put(iommu_group); |
cba3345c AW |
291 | } |
292 | ||
293 | static void vfio_group_put(struct vfio_group *group) | |
294 | { | |
6d2cd3ce | 295 | kref_put_mutex(&group->kref, vfio_group_release, &vfio.group_lock); |
cba3345c AW |
296 | } |
297 | ||
298 | /* Assume group_lock or group reference is held */ | |
299 | static void vfio_group_get(struct vfio_group *group) | |
300 | { | |
301 | kref_get(&group->kref); | |
302 | } | |
303 | ||
304 | /* | |
305 | * Not really a try as we will sleep for mutex, but we need to make | |
306 | * sure the group pointer is valid under lock and get a reference. | |
307 | */ | |
308 | static struct vfio_group *vfio_group_try_get(struct vfio_group *group) | |
309 | { | |
310 | struct vfio_group *target = group; | |
311 | ||
312 | mutex_lock(&vfio.group_lock); | |
313 | list_for_each_entry(group, &vfio.group_list, vfio_next) { | |
314 | if (group == target) { | |
315 | vfio_group_get(group); | |
316 | mutex_unlock(&vfio.group_lock); | |
317 | return group; | |
318 | } | |
319 | } | |
320 | mutex_unlock(&vfio.group_lock); | |
321 | ||
322 | return NULL; | |
323 | } | |
324 | ||
325 | static | |
326 | struct vfio_group *vfio_group_get_from_iommu(struct iommu_group *iommu_group) | |
327 | { | |
328 | struct vfio_group *group; | |
329 | ||
330 | mutex_lock(&vfio.group_lock); | |
331 | list_for_each_entry(group, &vfio.group_list, vfio_next) { | |
332 | if (group->iommu_group == iommu_group) { | |
333 | vfio_group_get(group); | |
334 | mutex_unlock(&vfio.group_lock); | |
335 | return group; | |
336 | } | |
337 | } | |
338 | mutex_unlock(&vfio.group_lock); | |
339 | ||
340 | return NULL; | |
341 | } | |
342 | ||
343 | static struct vfio_group *vfio_group_get_from_minor(int minor) | |
344 | { | |
345 | struct vfio_group *group; | |
346 | ||
347 | mutex_lock(&vfio.group_lock); | |
348 | group = idr_find(&vfio.group_idr, minor); | |
349 | if (!group) { | |
350 | mutex_unlock(&vfio.group_lock); | |
351 | return NULL; | |
352 | } | |
353 | vfio_group_get(group); | |
354 | mutex_unlock(&vfio.group_lock); | |
355 | ||
356 | return group; | |
357 | } | |
358 | ||
359 | /** | |
360 | * Device objects - create, release, get, put, search | |
361 | */ | |
362 | static | |
363 | struct vfio_device *vfio_group_create_device(struct vfio_group *group, | |
364 | struct device *dev, | |
365 | const struct vfio_device_ops *ops, | |
366 | void *device_data) | |
367 | { | |
368 | struct vfio_device *device; | |
cba3345c AW |
369 | |
370 | device = kzalloc(sizeof(*device), GFP_KERNEL); | |
371 | if (!device) | |
372 | return ERR_PTR(-ENOMEM); | |
373 | ||
374 | kref_init(&device->kref); | |
375 | device->dev = dev; | |
376 | device->group = group; | |
377 | device->ops = ops; | |
378 | device->device_data = device_data; | |
8283b491 | 379 | dev_set_drvdata(dev, device); |
cba3345c AW |
380 | |
381 | /* No need to get group_lock, caller has group reference */ | |
382 | vfio_group_get(group); | |
383 | ||
384 | mutex_lock(&group->device_lock); | |
385 | list_add(&device->group_next, &group->device_list); | |
386 | mutex_unlock(&group->device_lock); | |
387 | ||
388 | return device; | |
389 | } | |
390 | ||
391 | static void vfio_device_release(struct kref *kref) | |
392 | { | |
393 | struct vfio_device *device = container_of(kref, | |
394 | struct vfio_device, kref); | |
395 | struct vfio_group *group = device->group; | |
396 | ||
cba3345c AW |
397 | list_del(&device->group_next); |
398 | mutex_unlock(&group->device_lock); | |
399 | ||
400 | dev_set_drvdata(device->dev, NULL); | |
401 | ||
402 | kfree(device); | |
403 | ||
404 | /* vfio_del_group_dev may be waiting for this device */ | |
405 | wake_up(&vfio.release_q); | |
406 | } | |
407 | ||
408 | /* Device reference always implies a group reference */ | |
44f50716 | 409 | void vfio_device_put(struct vfio_device *device) |
cba3345c | 410 | { |
934ad4c2 | 411 | struct vfio_group *group = device->group; |
90b1253e | 412 | kref_put_mutex(&device->kref, vfio_device_release, &group->device_lock); |
934ad4c2 | 413 | vfio_group_put(group); |
cba3345c | 414 | } |
44f50716 | 415 | EXPORT_SYMBOL_GPL(vfio_device_put); |
cba3345c AW |
416 | |
417 | static void vfio_device_get(struct vfio_device *device) | |
418 | { | |
419 | vfio_group_get(device->group); | |
420 | kref_get(&device->kref); | |
421 | } | |
422 | ||
423 | static struct vfio_device *vfio_group_get_device(struct vfio_group *group, | |
424 | struct device *dev) | |
425 | { | |
426 | struct vfio_device *device; | |
427 | ||
428 | mutex_lock(&group->device_lock); | |
429 | list_for_each_entry(device, &group->device_list, group_next) { | |
430 | if (device->dev == dev) { | |
431 | vfio_device_get(device); | |
432 | mutex_unlock(&group->device_lock); | |
433 | return device; | |
434 | } | |
435 | } | |
436 | mutex_unlock(&group->device_lock); | |
437 | return NULL; | |
438 | } | |
439 | ||
440 | /* | |
441 | * Whitelist some drivers that we know are safe (no dma) or just sit on | |
442 | * a device. It's not always practical to leave a device within a group | |
443 | * driverless as it could get re-bound to something unsafe. | |
444 | */ | |
2b489a45 | 445 | static const char * const vfio_driver_whitelist[] = { "pci-stub", "pcieport" }; |
cba3345c AW |
446 | |
447 | static bool vfio_whitelisted_driver(struct device_driver *drv) | |
448 | { | |
449 | int i; | |
450 | ||
451 | for (i = 0; i < ARRAY_SIZE(vfio_driver_whitelist); i++) { | |
452 | if (!strcmp(drv->name, vfio_driver_whitelist[i])) | |
453 | return true; | |
454 | } | |
455 | ||
456 | return false; | |
457 | } | |
458 | ||
459 | /* | |
60720a0f AW |
460 | * A vfio group is viable for use by userspace if all devices are in |
461 | * one of the following states: | |
462 | * - driver-less | |
463 | * - bound to a vfio driver | |
464 | * - bound to a whitelisted driver | |
465 | * | |
466 | * We use two methods to determine whether a device is bound to a vfio | |
467 | * driver. The first is to test whether the device exists in the vfio | |
468 | * group. The second is to test if the device exists on the group | |
469 | * unbound_list, indicating it's in the middle of transitioning from | |
470 | * a vfio driver to driver-less. | |
cba3345c AW |
471 | */ |
472 | static int vfio_dev_viable(struct device *dev, void *data) | |
473 | { | |
474 | struct vfio_group *group = data; | |
475 | struct vfio_device *device; | |
de2b3eea | 476 | struct device_driver *drv = ACCESS_ONCE(dev->driver); |
60720a0f AW |
477 | struct vfio_unbound_dev *unbound; |
478 | int ret = -EINVAL; | |
479 | ||
480 | mutex_lock(&group->unbound_lock); | |
481 | list_for_each_entry(unbound, &group->unbound_list, unbound_next) { | |
482 | if (dev == unbound->dev) { | |
483 | ret = 0; | |
484 | break; | |
485 | } | |
486 | } | |
487 | mutex_unlock(&group->unbound_lock); | |
cba3345c | 488 | |
60720a0f | 489 | if (!ret || !drv || vfio_whitelisted_driver(drv)) |
cba3345c AW |
490 | return 0; |
491 | ||
492 | device = vfio_group_get_device(group, dev); | |
493 | if (device) { | |
494 | vfio_device_put(device); | |
495 | return 0; | |
496 | } | |
497 | ||
60720a0f | 498 | return ret; |
cba3345c AW |
499 | } |
500 | ||
501 | /** | |
502 | * Async device support | |
503 | */ | |
504 | static int vfio_group_nb_add_dev(struct vfio_group *group, struct device *dev) | |
505 | { | |
506 | struct vfio_device *device; | |
507 | ||
508 | /* Do we already know about it? We shouldn't */ | |
509 | device = vfio_group_get_device(group, dev); | |
510 | if (WARN_ON_ONCE(device)) { | |
511 | vfio_device_put(device); | |
512 | return 0; | |
513 | } | |
514 | ||
515 | /* Nothing to do for idle groups */ | |
516 | if (!atomic_read(&group->container_users)) | |
517 | return 0; | |
518 | ||
519 | /* TODO Prevent device auto probing */ | |
520 | WARN("Device %s added to live group %d!\n", dev_name(dev), | |
521 | iommu_group_id(group->iommu_group)); | |
522 | ||
523 | return 0; | |
524 | } | |
525 | ||
cba3345c AW |
526 | static int vfio_group_nb_verify(struct vfio_group *group, struct device *dev) |
527 | { | |
528 | /* We don't care what happens when the group isn't in use */ | |
529 | if (!atomic_read(&group->container_users)) | |
530 | return 0; | |
531 | ||
532 | return vfio_dev_viable(dev, group); | |
533 | } | |
534 | ||
535 | static int vfio_iommu_group_notifier(struct notifier_block *nb, | |
536 | unsigned long action, void *data) | |
537 | { | |
538 | struct vfio_group *group = container_of(nb, struct vfio_group, nb); | |
539 | struct device *dev = data; | |
60720a0f | 540 | struct vfio_unbound_dev *unbound; |
cba3345c AW |
541 | |
542 | /* | |
c6401930 AW |
543 | * Need to go through a group_lock lookup to get a reference or we |
544 | * risk racing a group being removed. Ignore spurious notifies. | |
cba3345c AW |
545 | */ |
546 | group = vfio_group_try_get(group); | |
c6401930 | 547 | if (!group) |
cba3345c AW |
548 | return NOTIFY_OK; |
549 | ||
550 | switch (action) { | |
551 | case IOMMU_GROUP_NOTIFY_ADD_DEVICE: | |
552 | vfio_group_nb_add_dev(group, dev); | |
553 | break; | |
554 | case IOMMU_GROUP_NOTIFY_DEL_DEVICE: | |
de9c7602 AW |
555 | /* |
556 | * Nothing to do here. If the device is in use, then the | |
557 | * vfio sub-driver should block the remove callback until | |
558 | * it is unused. If the device is unused or attached to a | |
559 | * stub driver, then it should be released and we don't | |
560 | * care that it will be going away. | |
561 | */ | |
cba3345c AW |
562 | break; |
563 | case IOMMU_GROUP_NOTIFY_BIND_DRIVER: | |
564 | pr_debug("%s: Device %s, group %d binding to driver\n", | |
565 | __func__, dev_name(dev), | |
566 | iommu_group_id(group->iommu_group)); | |
567 | break; | |
568 | case IOMMU_GROUP_NOTIFY_BOUND_DRIVER: | |
569 | pr_debug("%s: Device %s, group %d bound to driver %s\n", | |
570 | __func__, dev_name(dev), | |
571 | iommu_group_id(group->iommu_group), dev->driver->name); | |
572 | BUG_ON(vfio_group_nb_verify(group, dev)); | |
573 | break; | |
574 | case IOMMU_GROUP_NOTIFY_UNBIND_DRIVER: | |
575 | pr_debug("%s: Device %s, group %d unbinding from driver %s\n", | |
576 | __func__, dev_name(dev), | |
577 | iommu_group_id(group->iommu_group), dev->driver->name); | |
578 | break; | |
579 | case IOMMU_GROUP_NOTIFY_UNBOUND_DRIVER: | |
580 | pr_debug("%s: Device %s, group %d unbound from driver\n", | |
581 | __func__, dev_name(dev), | |
582 | iommu_group_id(group->iommu_group)); | |
583 | /* | |
584 | * XXX An unbound device in a live group is ok, but we'd | |
585 | * really like to avoid the above BUG_ON by preventing other | |
586 | * drivers from binding to it. Once that occurs, we have to | |
587 | * stop the system to maintain isolation. At a minimum, we'd | |
588 | * want a toggle to disable driver auto probe for this device. | |
589 | */ | |
60720a0f AW |
590 | |
591 | mutex_lock(&group->unbound_lock); | |
592 | list_for_each_entry(unbound, | |
593 | &group->unbound_list, unbound_next) { | |
594 | if (dev == unbound->dev) { | |
595 | list_del(&unbound->unbound_next); | |
596 | kfree(unbound); | |
597 | break; | |
598 | } | |
599 | } | |
600 | mutex_unlock(&group->unbound_lock); | |
cba3345c AW |
601 | break; |
602 | } | |
603 | ||
604 | vfio_group_put(group); | |
605 | return NOTIFY_OK; | |
606 | } | |
607 | ||
608 | /** | |
609 | * VFIO driver API | |
610 | */ | |
611 | int vfio_add_group_dev(struct device *dev, | |
612 | const struct vfio_device_ops *ops, void *device_data) | |
613 | { | |
614 | struct iommu_group *iommu_group; | |
615 | struct vfio_group *group; | |
616 | struct vfio_device *device; | |
617 | ||
618 | iommu_group = iommu_group_get(dev); | |
619 | if (!iommu_group) | |
620 | return -EINVAL; | |
621 | ||
622 | group = vfio_group_get_from_iommu(iommu_group); | |
623 | if (!group) { | |
624 | group = vfio_create_group(iommu_group); | |
625 | if (IS_ERR(group)) { | |
626 | iommu_group_put(iommu_group); | |
627 | return PTR_ERR(group); | |
628 | } | |
4a68810d AW |
629 | } else { |
630 | /* | |
631 | * A found vfio_group already holds a reference to the | |
632 | * iommu_group. A created vfio_group keeps the reference. | |
633 | */ | |
634 | iommu_group_put(iommu_group); | |
cba3345c AW |
635 | } |
636 | ||
637 | device = vfio_group_get_device(group, dev); | |
638 | if (device) { | |
639 | WARN(1, "Device %s already exists on group %d\n", | |
640 | dev_name(dev), iommu_group_id(iommu_group)); | |
641 | vfio_device_put(device); | |
642 | vfio_group_put(group); | |
cba3345c AW |
643 | return -EBUSY; |
644 | } | |
645 | ||
646 | device = vfio_group_create_device(group, dev, ops, device_data); | |
647 | if (IS_ERR(device)) { | |
648 | vfio_group_put(group); | |
cba3345c AW |
649 | return PTR_ERR(device); |
650 | } | |
651 | ||
652 | /* | |
4a68810d AW |
653 | * Drop all but the vfio_device reference. The vfio_device holds |
654 | * a reference to the vfio_group, which holds a reference to the | |
655 | * iommu_group. | |
cba3345c AW |
656 | */ |
657 | vfio_group_put(group); | |
658 | ||
659 | return 0; | |
660 | } | |
661 | EXPORT_SYMBOL_GPL(vfio_add_group_dev); | |
662 | ||
44f50716 VMP |
663 | /** |
664 | * Get a reference to the vfio_device for a device that is known to | |
665 | * be bound to a vfio driver. The driver implicitly holds a | |
666 | * vfio_device reference between vfio_add_group_dev and | |
667 | * vfio_del_group_dev. We can therefore use drvdata to increment | |
668 | * that reference from the struct device. This additional | |
669 | * reference must be released by calling vfio_device_put. | |
670 | */ | |
671 | struct vfio_device *vfio_device_get_from_dev(struct device *dev) | |
672 | { | |
673 | struct vfio_device *device = dev_get_drvdata(dev); | |
674 | ||
675 | vfio_device_get(device); | |
676 | ||
677 | return device; | |
678 | } | |
679 | EXPORT_SYMBOL_GPL(vfio_device_get_from_dev); | |
680 | ||
681 | /* | |
682 | * Caller must hold a reference to the vfio_device | |
683 | */ | |
684 | void *vfio_device_data(struct vfio_device *device) | |
685 | { | |
686 | return device->device_data; | |
687 | } | |
688 | EXPORT_SYMBOL_GPL(vfio_device_data); | |
689 | ||
e014e944 AW |
690 | /* Given a referenced group, check if it contains the device */ |
691 | static bool vfio_dev_present(struct vfio_group *group, struct device *dev) | |
cba3345c | 692 | { |
cba3345c AW |
693 | struct vfio_device *device; |
694 | ||
cba3345c | 695 | device = vfio_group_get_device(group, dev); |
e014e944 | 696 | if (!device) |
cba3345c | 697 | return false; |
cba3345c AW |
698 | |
699 | vfio_device_put(device); | |
cba3345c AW |
700 | return true; |
701 | } | |
702 | ||
703 | /* | |
704 | * Decrement the device reference count and wait for the device to be | |
705 | * removed. Open file descriptors for the device... */ | |
706 | void *vfio_del_group_dev(struct device *dev) | |
707 | { | |
708 | struct vfio_device *device = dev_get_drvdata(dev); | |
709 | struct vfio_group *group = device->group; | |
cba3345c | 710 | void *device_data = device->device_data; |
60720a0f | 711 | struct vfio_unbound_dev *unbound; |
13060b64 | 712 | unsigned int i = 0; |
cba3345c | 713 | |
e014e944 AW |
714 | /* |
715 | * The group exists so long as we have a device reference. Get | |
716 | * a group reference and use it to scan for the device going away. | |
717 | */ | |
718 | vfio_group_get(group); | |
719 | ||
60720a0f AW |
720 | /* |
721 | * When the device is removed from the group, the group suddenly | |
722 | * becomes non-viable; the device has a driver (until the unbind | |
723 | * completes), but it's not present in the group. This is bad news | |
724 | * for any external users that need to re-acquire a group reference | |
725 | * in order to match and release their existing reference. To | |
726 | * solve this, we track such devices on the unbound_list to bridge | |
727 | * the gap until they're fully unbound. | |
728 | */ | |
729 | unbound = kzalloc(sizeof(*unbound), GFP_KERNEL); | |
730 | if (unbound) { | |
731 | unbound->dev = dev; | |
732 | mutex_lock(&group->unbound_lock); | |
733 | list_add(&unbound->unbound_next, &group->unbound_list); | |
734 | mutex_unlock(&group->unbound_lock); | |
735 | } | |
736 | WARN_ON(!unbound); | |
737 | ||
cba3345c AW |
738 | vfio_device_put(device); |
739 | ||
13060b64 AW |
740 | /* |
741 | * If the device is still present in the group after the above | |
742 | * 'put', then it is in use and we need to request it from the | |
743 | * bus driver. The driver may in turn need to request the | |
744 | * device from the user. We send the request on an arbitrary | |
745 | * interval with counter to allow the driver to take escalating | |
746 | * measures to release the device if it has the ability to do so. | |
747 | */ | |
748 | do { | |
749 | device = vfio_group_get_device(group, dev); | |
750 | if (!device) | |
751 | break; | |
752 | ||
753 | if (device->ops->request) | |
754 | device->ops->request(device_data, i++); | |
755 | ||
756 | vfio_device_put(device); | |
757 | ||
758 | } while (wait_event_interruptible_timeout(vfio.release_q, | |
759 | !vfio_dev_present(group, dev), | |
760 | HZ * 10) <= 0); | |
e014e944 AW |
761 | |
762 | vfio_group_put(group); | |
cba3345c | 763 | |
cba3345c AW |
764 | return device_data; |
765 | } | |
766 | EXPORT_SYMBOL_GPL(vfio_del_group_dev); | |
767 | ||
768 | /** | |
769 | * VFIO base fd, /dev/vfio/vfio | |
770 | */ | |
771 | static long vfio_ioctl_check_extension(struct vfio_container *container, | |
772 | unsigned long arg) | |
773 | { | |
0b43c082 | 774 | struct vfio_iommu_driver *driver; |
cba3345c AW |
775 | long ret = 0; |
776 | ||
0b43c082 AW |
777 | down_read(&container->group_lock); |
778 | ||
779 | driver = container->iommu_driver; | |
780 | ||
cba3345c AW |
781 | switch (arg) { |
782 | /* No base extensions yet */ | |
783 | default: | |
784 | /* | |
785 | * If no driver is set, poll all registered drivers for | |
786 | * extensions and return the first positive result. If | |
787 | * a driver is already set, further queries will be passed | |
788 | * only to that driver. | |
789 | */ | |
790 | if (!driver) { | |
791 | mutex_lock(&vfio.iommu_drivers_lock); | |
792 | list_for_each_entry(driver, &vfio.iommu_drivers_list, | |
793 | vfio_next) { | |
794 | if (!try_module_get(driver->ops->owner)) | |
795 | continue; | |
796 | ||
797 | ret = driver->ops->ioctl(NULL, | |
798 | VFIO_CHECK_EXTENSION, | |
799 | arg); | |
800 | module_put(driver->ops->owner); | |
801 | if (ret > 0) | |
802 | break; | |
803 | } | |
804 | mutex_unlock(&vfio.iommu_drivers_lock); | |
805 | } else | |
806 | ret = driver->ops->ioctl(container->iommu_data, | |
807 | VFIO_CHECK_EXTENSION, arg); | |
808 | } | |
809 | ||
0b43c082 AW |
810 | up_read(&container->group_lock); |
811 | ||
cba3345c AW |
812 | return ret; |
813 | } | |
814 | ||
9587f44a | 815 | /* hold write lock on container->group_lock */ |
cba3345c AW |
816 | static int __vfio_container_attach_groups(struct vfio_container *container, |
817 | struct vfio_iommu_driver *driver, | |
818 | void *data) | |
819 | { | |
820 | struct vfio_group *group; | |
821 | int ret = -ENODEV; | |
822 | ||
823 | list_for_each_entry(group, &container->group_list, container_next) { | |
824 | ret = driver->ops->attach_group(data, group->iommu_group); | |
825 | if (ret) | |
826 | goto unwind; | |
827 | } | |
828 | ||
829 | return ret; | |
830 | ||
831 | unwind: | |
832 | list_for_each_entry_continue_reverse(group, &container->group_list, | |
833 | container_next) { | |
834 | driver->ops->detach_group(data, group->iommu_group); | |
835 | } | |
836 | ||
837 | return ret; | |
838 | } | |
839 | ||
840 | static long vfio_ioctl_set_iommu(struct vfio_container *container, | |
841 | unsigned long arg) | |
842 | { | |
843 | struct vfio_iommu_driver *driver; | |
844 | long ret = -ENODEV; | |
845 | ||
9587f44a | 846 | down_write(&container->group_lock); |
cba3345c AW |
847 | |
848 | /* | |
849 | * The container is designed to be an unprivileged interface while | |
850 | * the group can be assigned to specific users. Therefore, only by | |
851 | * adding a group to a container does the user get the privilege of | |
852 | * enabling the iommu, which may allocate finite resources. There | |
853 | * is no unset_iommu, but by removing all the groups from a container, | |
854 | * the container is deprivileged and returns to an unset state. | |
855 | */ | |
856 | if (list_empty(&container->group_list) || container->iommu_driver) { | |
9587f44a | 857 | up_write(&container->group_lock); |
cba3345c AW |
858 | return -EINVAL; |
859 | } | |
860 | ||
861 | mutex_lock(&vfio.iommu_drivers_lock); | |
862 | list_for_each_entry(driver, &vfio.iommu_drivers_list, vfio_next) { | |
863 | void *data; | |
864 | ||
865 | if (!try_module_get(driver->ops->owner)) | |
866 | continue; | |
867 | ||
868 | /* | |
869 | * The arg magic for SET_IOMMU is the same as CHECK_EXTENSION, | |
870 | * so test which iommu driver reported support for this | |
871 | * extension and call open on them. We also pass them the | |
872 | * magic, allowing a single driver to support multiple | |
873 | * interfaces if they'd like. | |
874 | */ | |
875 | if (driver->ops->ioctl(NULL, VFIO_CHECK_EXTENSION, arg) <= 0) { | |
876 | module_put(driver->ops->owner); | |
877 | continue; | |
878 | } | |
879 | ||
880 | /* module reference holds the driver we're working on */ | |
881 | mutex_unlock(&vfio.iommu_drivers_lock); | |
882 | ||
883 | data = driver->ops->open(arg); | |
884 | if (IS_ERR(data)) { | |
885 | ret = PTR_ERR(data); | |
886 | module_put(driver->ops->owner); | |
887 | goto skip_drivers_unlock; | |
888 | } | |
889 | ||
890 | ret = __vfio_container_attach_groups(container, driver, data); | |
891 | if (!ret) { | |
892 | container->iommu_driver = driver; | |
893 | container->iommu_data = data; | |
894 | } else { | |
895 | driver->ops->release(data); | |
896 | module_put(driver->ops->owner); | |
897 | } | |
898 | ||
899 | goto skip_drivers_unlock; | |
900 | } | |
901 | ||
902 | mutex_unlock(&vfio.iommu_drivers_lock); | |
903 | skip_drivers_unlock: | |
9587f44a | 904 | up_write(&container->group_lock); |
cba3345c AW |
905 | |
906 | return ret; | |
907 | } | |
908 | ||
909 | static long vfio_fops_unl_ioctl(struct file *filep, | |
910 | unsigned int cmd, unsigned long arg) | |
911 | { | |
912 | struct vfio_container *container = filep->private_data; | |
913 | struct vfio_iommu_driver *driver; | |
914 | void *data; | |
915 | long ret = -EINVAL; | |
916 | ||
917 | if (!container) | |
918 | return ret; | |
919 | ||
cba3345c AW |
920 | switch (cmd) { |
921 | case VFIO_GET_API_VERSION: | |
922 | ret = VFIO_API_VERSION; | |
923 | break; | |
924 | case VFIO_CHECK_EXTENSION: | |
925 | ret = vfio_ioctl_check_extension(container, arg); | |
926 | break; | |
927 | case VFIO_SET_IOMMU: | |
928 | ret = vfio_ioctl_set_iommu(container, arg); | |
929 | break; | |
930 | default: | |
0b43c082 AW |
931 | down_read(&container->group_lock); |
932 | ||
933 | driver = container->iommu_driver; | |
934 | data = container->iommu_data; | |
935 | ||
cba3345c AW |
936 | if (driver) /* passthrough all unrecognized ioctls */ |
937 | ret = driver->ops->ioctl(data, cmd, arg); | |
0b43c082 AW |
938 | |
939 | up_read(&container->group_lock); | |
cba3345c AW |
940 | } |
941 | ||
942 | return ret; | |
943 | } | |
944 | ||
945 | #ifdef CONFIG_COMPAT | |
946 | static long vfio_fops_compat_ioctl(struct file *filep, | |
947 | unsigned int cmd, unsigned long arg) | |
948 | { | |
949 | arg = (unsigned long)compat_ptr(arg); | |
950 | return vfio_fops_unl_ioctl(filep, cmd, arg); | |
951 | } | |
952 | #endif /* CONFIG_COMPAT */ | |
953 | ||
954 | static int vfio_fops_open(struct inode *inode, struct file *filep) | |
955 | { | |
956 | struct vfio_container *container; | |
957 | ||
958 | container = kzalloc(sizeof(*container), GFP_KERNEL); | |
959 | if (!container) | |
960 | return -ENOMEM; | |
961 | ||
962 | INIT_LIST_HEAD(&container->group_list); | |
9587f44a | 963 | init_rwsem(&container->group_lock); |
cba3345c AW |
964 | kref_init(&container->kref); |
965 | ||
966 | filep->private_data = container; | |
967 | ||
968 | return 0; | |
969 | } | |
970 | ||
971 | static int vfio_fops_release(struct inode *inode, struct file *filep) | |
972 | { | |
973 | struct vfio_container *container = filep->private_data; | |
974 | ||
975 | filep->private_data = NULL; | |
976 | ||
977 | vfio_container_put(container); | |
978 | ||
979 | return 0; | |
980 | } | |
981 | ||
982 | /* | |
983 | * Once an iommu driver is set, we optionally pass read/write/mmap | |
984 | * on to the driver, allowing management interfaces beyond ioctl. | |
985 | */ | |
986 | static ssize_t vfio_fops_read(struct file *filep, char __user *buf, | |
987 | size_t count, loff_t *ppos) | |
988 | { | |
989 | struct vfio_container *container = filep->private_data; | |
0b43c082 AW |
990 | struct vfio_iommu_driver *driver; |
991 | ssize_t ret = -EINVAL; | |
cba3345c | 992 | |
0b43c082 AW |
993 | down_read(&container->group_lock); |
994 | ||
995 | driver = container->iommu_driver; | |
996 | if (likely(driver && driver->ops->read)) | |
997 | ret = driver->ops->read(container->iommu_data, | |
998 | buf, count, ppos); | |
cba3345c | 999 | |
0b43c082 AW |
1000 | up_read(&container->group_lock); |
1001 | ||
1002 | return ret; | |
cba3345c AW |
1003 | } |
1004 | ||
1005 | static ssize_t vfio_fops_write(struct file *filep, const char __user *buf, | |
1006 | size_t count, loff_t *ppos) | |
1007 | { | |
1008 | struct vfio_container *container = filep->private_data; | |
0b43c082 AW |
1009 | struct vfio_iommu_driver *driver; |
1010 | ssize_t ret = -EINVAL; | |
cba3345c | 1011 | |
0b43c082 AW |
1012 | down_read(&container->group_lock); |
1013 | ||
1014 | driver = container->iommu_driver; | |
1015 | if (likely(driver && driver->ops->write)) | |
1016 | ret = driver->ops->write(container->iommu_data, | |
1017 | buf, count, ppos); | |
1018 | ||
1019 | up_read(&container->group_lock); | |
cba3345c | 1020 | |
0b43c082 | 1021 | return ret; |
cba3345c AW |
1022 | } |
1023 | ||
1024 | static int vfio_fops_mmap(struct file *filep, struct vm_area_struct *vma) | |
1025 | { | |
1026 | struct vfio_container *container = filep->private_data; | |
0b43c082 AW |
1027 | struct vfio_iommu_driver *driver; |
1028 | int ret = -EINVAL; | |
cba3345c | 1029 | |
0b43c082 | 1030 | down_read(&container->group_lock); |
cba3345c | 1031 | |
0b43c082 AW |
1032 | driver = container->iommu_driver; |
1033 | if (likely(driver && driver->ops->mmap)) | |
1034 | ret = driver->ops->mmap(container->iommu_data, vma); | |
1035 | ||
1036 | up_read(&container->group_lock); | |
1037 | ||
1038 | return ret; | |
cba3345c AW |
1039 | } |
1040 | ||
1041 | static const struct file_operations vfio_fops = { | |
1042 | .owner = THIS_MODULE, | |
1043 | .open = vfio_fops_open, | |
1044 | .release = vfio_fops_release, | |
1045 | .read = vfio_fops_read, | |
1046 | .write = vfio_fops_write, | |
1047 | .unlocked_ioctl = vfio_fops_unl_ioctl, | |
1048 | #ifdef CONFIG_COMPAT | |
1049 | .compat_ioctl = vfio_fops_compat_ioctl, | |
1050 | #endif | |
1051 | .mmap = vfio_fops_mmap, | |
1052 | }; | |
1053 | ||
1054 | /** | |
1055 | * VFIO Group fd, /dev/vfio/$GROUP | |
1056 | */ | |
1057 | static void __vfio_group_unset_container(struct vfio_group *group) | |
1058 | { | |
1059 | struct vfio_container *container = group->container; | |
1060 | struct vfio_iommu_driver *driver; | |
1061 | ||
9587f44a | 1062 | down_write(&container->group_lock); |
cba3345c AW |
1063 | |
1064 | driver = container->iommu_driver; | |
1065 | if (driver) | |
1066 | driver->ops->detach_group(container->iommu_data, | |
1067 | group->iommu_group); | |
1068 | ||
1069 | group->container = NULL; | |
1070 | list_del(&group->container_next); | |
1071 | ||
1072 | /* Detaching the last group deprivileges a container, remove iommu */ | |
1073 | if (driver && list_empty(&container->group_list)) { | |
1074 | driver->ops->release(container->iommu_data); | |
1075 | module_put(driver->ops->owner); | |
1076 | container->iommu_driver = NULL; | |
1077 | container->iommu_data = NULL; | |
1078 | } | |
1079 | ||
9587f44a | 1080 | up_write(&container->group_lock); |
cba3345c AW |
1081 | |
1082 | vfio_container_put(container); | |
1083 | } | |
1084 | ||
1085 | /* | |
1086 | * VFIO_GROUP_UNSET_CONTAINER should fail if there are other users or | |
1087 | * if there was no container to unset. Since the ioctl is called on | |
1088 | * the group, we know that still exists, therefore the only valid | |
1089 | * transition here is 1->0. | |
1090 | */ | |
1091 | static int vfio_group_unset_container(struct vfio_group *group) | |
1092 | { | |
1093 | int users = atomic_cmpxchg(&group->container_users, 1, 0); | |
1094 | ||
1095 | if (!users) | |
1096 | return -EINVAL; | |
1097 | if (users != 1) | |
1098 | return -EBUSY; | |
1099 | ||
1100 | __vfio_group_unset_container(group); | |
1101 | ||
1102 | return 0; | |
1103 | } | |
1104 | ||
1105 | /* | |
1106 | * When removing container users, anything that removes the last user | |
1107 | * implicitly removes the group from the container. That is, if the | |
1108 | * group file descriptor is closed, as well as any device file descriptors, | |
1109 | * the group is free. | |
1110 | */ | |
1111 | static void vfio_group_try_dissolve_container(struct vfio_group *group) | |
1112 | { | |
1113 | if (0 == atomic_dec_if_positive(&group->container_users)) | |
1114 | __vfio_group_unset_container(group); | |
1115 | } | |
1116 | ||
1117 | static int vfio_group_set_container(struct vfio_group *group, int container_fd) | |
1118 | { | |
2903ff01 | 1119 | struct fd f; |
cba3345c AW |
1120 | struct vfio_container *container; |
1121 | struct vfio_iommu_driver *driver; | |
2903ff01 | 1122 | int ret = 0; |
cba3345c AW |
1123 | |
1124 | if (atomic_read(&group->container_users)) | |
1125 | return -EINVAL; | |
1126 | ||
2903ff01 AV |
1127 | f = fdget(container_fd); |
1128 | if (!f.file) | |
cba3345c AW |
1129 | return -EBADF; |
1130 | ||
1131 | /* Sanity check, is this really our fd? */ | |
2903ff01 AV |
1132 | if (f.file->f_op != &vfio_fops) { |
1133 | fdput(f); | |
cba3345c AW |
1134 | return -EINVAL; |
1135 | } | |
1136 | ||
2903ff01 | 1137 | container = f.file->private_data; |
cba3345c AW |
1138 | WARN_ON(!container); /* fget ensures we don't race vfio_release */ |
1139 | ||
9587f44a | 1140 | down_write(&container->group_lock); |
cba3345c AW |
1141 | |
1142 | driver = container->iommu_driver; | |
1143 | if (driver) { | |
1144 | ret = driver->ops->attach_group(container->iommu_data, | |
1145 | group->iommu_group); | |
1146 | if (ret) | |
1147 | goto unlock_out; | |
1148 | } | |
1149 | ||
1150 | group->container = container; | |
1151 | list_add(&group->container_next, &container->group_list); | |
1152 | ||
1153 | /* Get a reference on the container and mark a user within the group */ | |
1154 | vfio_container_get(container); | |
1155 | atomic_inc(&group->container_users); | |
1156 | ||
1157 | unlock_out: | |
9587f44a | 1158 | up_write(&container->group_lock); |
2903ff01 | 1159 | fdput(f); |
cba3345c AW |
1160 | return ret; |
1161 | } | |
1162 | ||
1163 | static bool vfio_group_viable(struct vfio_group *group) | |
1164 | { | |
1165 | return (iommu_group_for_each_dev(group->iommu_group, | |
1166 | group, vfio_dev_viable) == 0); | |
1167 | } | |
1168 | ||
1169 | static const struct file_operations vfio_device_fops; | |
1170 | ||
1171 | static int vfio_group_get_device_fd(struct vfio_group *group, char *buf) | |
1172 | { | |
1173 | struct vfio_device *device; | |
1174 | struct file *filep; | |
1175 | int ret = -ENODEV; | |
1176 | ||
1177 | if (0 == atomic_read(&group->container_users) || | |
1178 | !group->container->iommu_driver || !vfio_group_viable(group)) | |
1179 | return -EINVAL; | |
1180 | ||
1181 | mutex_lock(&group->device_lock); | |
1182 | list_for_each_entry(device, &group->device_list, group_next) { | |
1183 | if (strcmp(dev_name(device->dev), buf)) | |
1184 | continue; | |
1185 | ||
1186 | ret = device->ops->open(device->device_data); | |
1187 | if (ret) | |
1188 | break; | |
1189 | /* | |
1190 | * We can't use anon_inode_getfd() because we need to modify | |
1191 | * the f_mode flags directly to allow more than just ioctls | |
1192 | */ | |
5d042fbd | 1193 | ret = get_unused_fd_flags(O_CLOEXEC); |
cba3345c AW |
1194 | if (ret < 0) { |
1195 | device->ops->release(device->device_data); | |
1196 | break; | |
1197 | } | |
1198 | ||
1199 | filep = anon_inode_getfile("[vfio-device]", &vfio_device_fops, | |
1200 | device, O_RDWR); | |
1201 | if (IS_ERR(filep)) { | |
1202 | put_unused_fd(ret); | |
1203 | ret = PTR_ERR(filep); | |
1204 | device->ops->release(device->device_data); | |
1205 | break; | |
1206 | } | |
1207 | ||
1208 | /* | |
1209 | * TODO: add an anon_inode interface to do this. | |
1210 | * Appears to be missing by lack of need rather than | |
1211 | * explicitly prevented. Now there's need. | |
1212 | */ | |
1213 | filep->f_mode |= (FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE); | |
1214 | ||
cba3345c AW |
1215 | vfio_device_get(device); |
1216 | atomic_inc(&group->container_users); | |
31605deb AV |
1217 | |
1218 | fd_install(ret, filep); | |
cba3345c AW |
1219 | break; |
1220 | } | |
1221 | mutex_unlock(&group->device_lock); | |
1222 | ||
1223 | return ret; | |
1224 | } | |
1225 | ||
1226 | static long vfio_group_fops_unl_ioctl(struct file *filep, | |
1227 | unsigned int cmd, unsigned long arg) | |
1228 | { | |
1229 | struct vfio_group *group = filep->private_data; | |
1230 | long ret = -ENOTTY; | |
1231 | ||
1232 | switch (cmd) { | |
1233 | case VFIO_GROUP_GET_STATUS: | |
1234 | { | |
1235 | struct vfio_group_status status; | |
1236 | unsigned long minsz; | |
1237 | ||
1238 | minsz = offsetofend(struct vfio_group_status, flags); | |
1239 | ||
1240 | if (copy_from_user(&status, (void __user *)arg, minsz)) | |
1241 | return -EFAULT; | |
1242 | ||
1243 | if (status.argsz < minsz) | |
1244 | return -EINVAL; | |
1245 | ||
1246 | status.flags = 0; | |
1247 | ||
1248 | if (vfio_group_viable(group)) | |
1249 | status.flags |= VFIO_GROUP_FLAGS_VIABLE; | |
1250 | ||
1251 | if (group->container) | |
1252 | status.flags |= VFIO_GROUP_FLAGS_CONTAINER_SET; | |
1253 | ||
1254 | if (copy_to_user((void __user *)arg, &status, minsz)) | |
1255 | return -EFAULT; | |
1256 | ||
1257 | ret = 0; | |
1258 | break; | |
1259 | } | |
1260 | case VFIO_GROUP_SET_CONTAINER: | |
1261 | { | |
1262 | int fd; | |
1263 | ||
1264 | if (get_user(fd, (int __user *)arg)) | |
1265 | return -EFAULT; | |
1266 | ||
1267 | if (fd < 0) | |
1268 | return -EINVAL; | |
1269 | ||
1270 | ret = vfio_group_set_container(group, fd); | |
1271 | break; | |
1272 | } | |
1273 | case VFIO_GROUP_UNSET_CONTAINER: | |
1274 | ret = vfio_group_unset_container(group); | |
1275 | break; | |
1276 | case VFIO_GROUP_GET_DEVICE_FD: | |
1277 | { | |
1278 | char *buf; | |
1279 | ||
1280 | buf = strndup_user((const char __user *)arg, PAGE_SIZE); | |
1281 | if (IS_ERR(buf)) | |
1282 | return PTR_ERR(buf); | |
1283 | ||
1284 | ret = vfio_group_get_device_fd(group, buf); | |
1285 | kfree(buf); | |
1286 | break; | |
1287 | } | |
1288 | } | |
1289 | ||
1290 | return ret; | |
1291 | } | |
1292 | ||
1293 | #ifdef CONFIG_COMPAT | |
1294 | static long vfio_group_fops_compat_ioctl(struct file *filep, | |
1295 | unsigned int cmd, unsigned long arg) | |
1296 | { | |
1297 | arg = (unsigned long)compat_ptr(arg); | |
1298 | return vfio_group_fops_unl_ioctl(filep, cmd, arg); | |
1299 | } | |
1300 | #endif /* CONFIG_COMPAT */ | |
1301 | ||
1302 | static int vfio_group_fops_open(struct inode *inode, struct file *filep) | |
1303 | { | |
1304 | struct vfio_group *group; | |
6d6768c6 | 1305 | int opened; |
cba3345c AW |
1306 | |
1307 | group = vfio_group_get_from_minor(iminor(inode)); | |
1308 | if (!group) | |
1309 | return -ENODEV; | |
1310 | ||
6d6768c6 AW |
1311 | /* Do we need multiple instances of the group open? Seems not. */ |
1312 | opened = atomic_cmpxchg(&group->opened, 0, 1); | |
1313 | if (opened) { | |
1314 | vfio_group_put(group); | |
1315 | return -EBUSY; | |
1316 | } | |
1317 | ||
1318 | /* Is something still in use from a previous open? */ | |
cba3345c | 1319 | if (group->container) { |
6d6768c6 | 1320 | atomic_dec(&group->opened); |
cba3345c AW |
1321 | vfio_group_put(group); |
1322 | return -EBUSY; | |
1323 | } | |
1324 | ||
1325 | filep->private_data = group; | |
1326 | ||
1327 | return 0; | |
1328 | } | |
1329 | ||
1330 | static int vfio_group_fops_release(struct inode *inode, struct file *filep) | |
1331 | { | |
1332 | struct vfio_group *group = filep->private_data; | |
1333 | ||
1334 | filep->private_data = NULL; | |
1335 | ||
1336 | vfio_group_try_dissolve_container(group); | |
1337 | ||
6d6768c6 AW |
1338 | atomic_dec(&group->opened); |
1339 | ||
cba3345c AW |
1340 | vfio_group_put(group); |
1341 | ||
1342 | return 0; | |
1343 | } | |
1344 | ||
1345 | static const struct file_operations vfio_group_fops = { | |
1346 | .owner = THIS_MODULE, | |
1347 | .unlocked_ioctl = vfio_group_fops_unl_ioctl, | |
1348 | #ifdef CONFIG_COMPAT | |
1349 | .compat_ioctl = vfio_group_fops_compat_ioctl, | |
1350 | #endif | |
1351 | .open = vfio_group_fops_open, | |
1352 | .release = vfio_group_fops_release, | |
1353 | }; | |
1354 | ||
1355 | /** | |
1356 | * VFIO Device fd | |
1357 | */ | |
1358 | static int vfio_device_fops_release(struct inode *inode, struct file *filep) | |
1359 | { | |
1360 | struct vfio_device *device = filep->private_data; | |
1361 | ||
1362 | device->ops->release(device->device_data); | |
1363 | ||
1364 | vfio_group_try_dissolve_container(device->group); | |
1365 | ||
1366 | vfio_device_put(device); | |
1367 | ||
1368 | return 0; | |
1369 | } | |
1370 | ||
1371 | static long vfio_device_fops_unl_ioctl(struct file *filep, | |
1372 | unsigned int cmd, unsigned long arg) | |
1373 | { | |
1374 | struct vfio_device *device = filep->private_data; | |
1375 | ||
1376 | if (unlikely(!device->ops->ioctl)) | |
1377 | return -EINVAL; | |
1378 | ||
1379 | return device->ops->ioctl(device->device_data, cmd, arg); | |
1380 | } | |
1381 | ||
1382 | static ssize_t vfio_device_fops_read(struct file *filep, char __user *buf, | |
1383 | size_t count, loff_t *ppos) | |
1384 | { | |
1385 | struct vfio_device *device = filep->private_data; | |
1386 | ||
1387 | if (unlikely(!device->ops->read)) | |
1388 | return -EINVAL; | |
1389 | ||
1390 | return device->ops->read(device->device_data, buf, count, ppos); | |
1391 | } | |
1392 | ||
1393 | static ssize_t vfio_device_fops_write(struct file *filep, | |
1394 | const char __user *buf, | |
1395 | size_t count, loff_t *ppos) | |
1396 | { | |
1397 | struct vfio_device *device = filep->private_data; | |
1398 | ||
1399 | if (unlikely(!device->ops->write)) | |
1400 | return -EINVAL; | |
1401 | ||
1402 | return device->ops->write(device->device_data, buf, count, ppos); | |
1403 | } | |
1404 | ||
1405 | static int vfio_device_fops_mmap(struct file *filep, struct vm_area_struct *vma) | |
1406 | { | |
1407 | struct vfio_device *device = filep->private_data; | |
1408 | ||
1409 | if (unlikely(!device->ops->mmap)) | |
1410 | return -EINVAL; | |
1411 | ||
1412 | return device->ops->mmap(device->device_data, vma); | |
1413 | } | |
1414 | ||
1415 | #ifdef CONFIG_COMPAT | |
1416 | static long vfio_device_fops_compat_ioctl(struct file *filep, | |
1417 | unsigned int cmd, unsigned long arg) | |
1418 | { | |
1419 | arg = (unsigned long)compat_ptr(arg); | |
1420 | return vfio_device_fops_unl_ioctl(filep, cmd, arg); | |
1421 | } | |
1422 | #endif /* CONFIG_COMPAT */ | |
1423 | ||
1424 | static const struct file_operations vfio_device_fops = { | |
1425 | .owner = THIS_MODULE, | |
1426 | .release = vfio_device_fops_release, | |
1427 | .read = vfio_device_fops_read, | |
1428 | .write = vfio_device_fops_write, | |
1429 | .unlocked_ioctl = vfio_device_fops_unl_ioctl, | |
1430 | #ifdef CONFIG_COMPAT | |
1431 | .compat_ioctl = vfio_device_fops_compat_ioctl, | |
1432 | #endif | |
1433 | .mmap = vfio_device_fops_mmap, | |
1434 | }; | |
1435 | ||
6cdd9782 AK |
1436 | /** |
1437 | * External user API, exported by symbols to be linked dynamically. | |
1438 | * | |
1439 | * The protocol includes: | |
1440 | * 1. do normal VFIO init operation: | |
1441 | * - opening a new container; | |
1442 | * - attaching group(s) to it; | |
1443 | * - setting an IOMMU driver for a container. | |
1444 | * When IOMMU is set for a container, all groups in it are | |
1445 | * considered ready to use by an external user. | |
1446 | * | |
1447 | * 2. User space passes a group fd to an external user. | |
1448 | * The external user calls vfio_group_get_external_user() | |
1449 | * to verify that: | |
1450 | * - the group is initialized; | |
1451 | * - IOMMU is set for it. | |
1452 | * If both checks passed, vfio_group_get_external_user() | |
1453 | * increments the container user counter to prevent | |
1454 | * the VFIO group from disposal before KVM exits. | |
1455 | * | |
1456 | * 3. The external user calls vfio_external_user_iommu_id() | |
1457 | * to know an IOMMU ID. | |
1458 | * | |
1459 | * 4. When the external KVM finishes, it calls | |
1460 | * vfio_group_put_external_user() to release the VFIO group. | |
1461 | * This call decrements the container user counter. | |
1462 | */ | |
1463 | struct vfio_group *vfio_group_get_external_user(struct file *filep) | |
1464 | { | |
1465 | struct vfio_group *group = filep->private_data; | |
1466 | ||
1467 | if (filep->f_op != &vfio_group_fops) | |
1468 | return ERR_PTR(-EINVAL); | |
1469 | ||
1470 | if (!atomic_inc_not_zero(&group->container_users)) | |
1471 | return ERR_PTR(-EINVAL); | |
1472 | ||
1473 | if (!group->container->iommu_driver || | |
1474 | !vfio_group_viable(group)) { | |
1475 | atomic_dec(&group->container_users); | |
1476 | return ERR_PTR(-EINVAL); | |
1477 | } | |
1478 | ||
1479 | vfio_group_get(group); | |
1480 | ||
1481 | return group; | |
1482 | } | |
1483 | EXPORT_SYMBOL_GPL(vfio_group_get_external_user); | |
1484 | ||
1485 | void vfio_group_put_external_user(struct vfio_group *group) | |
1486 | { | |
1487 | vfio_group_put(group); | |
1488 | vfio_group_try_dissolve_container(group); | |
1489 | } | |
1490 | EXPORT_SYMBOL_GPL(vfio_group_put_external_user); | |
1491 | ||
1492 | int vfio_external_user_iommu_id(struct vfio_group *group) | |
1493 | { | |
1494 | return iommu_group_id(group->iommu_group); | |
1495 | } | |
1496 | EXPORT_SYMBOL_GPL(vfio_external_user_iommu_id); | |
1497 | ||
88d7ab89 AW |
1498 | long vfio_external_check_extension(struct vfio_group *group, unsigned long arg) |
1499 | { | |
1500 | return vfio_ioctl_check_extension(group->container, arg); | |
1501 | } | |
1502 | EXPORT_SYMBOL_GPL(vfio_external_check_extension); | |
1503 | ||
cba3345c AW |
1504 | /** |
1505 | * Module/class support | |
1506 | */ | |
1507 | static char *vfio_devnode(struct device *dev, umode_t *mode) | |
1508 | { | |
1509 | return kasprintf(GFP_KERNEL, "vfio/%s", dev_name(dev)); | |
1510 | } | |
1511 | ||
d1099901 AW |
1512 | static struct miscdevice vfio_dev = { |
1513 | .minor = VFIO_MINOR, | |
1514 | .name = "vfio", | |
1515 | .fops = &vfio_fops, | |
1516 | .nodename = "vfio/vfio", | |
1517 | .mode = S_IRUGO | S_IWUGO, | |
1518 | }; | |
1519 | ||
cba3345c AW |
1520 | static int __init vfio_init(void) |
1521 | { | |
1522 | int ret; | |
1523 | ||
1524 | idr_init(&vfio.group_idr); | |
1525 | mutex_init(&vfio.group_lock); | |
1526 | mutex_init(&vfio.iommu_drivers_lock); | |
1527 | INIT_LIST_HEAD(&vfio.group_list); | |
1528 | INIT_LIST_HEAD(&vfio.iommu_drivers_list); | |
1529 | init_waitqueue_head(&vfio.release_q); | |
1530 | ||
d1099901 AW |
1531 | ret = misc_register(&vfio_dev); |
1532 | if (ret) { | |
1533 | pr_err("vfio: misc device register failed\n"); | |
1534 | return ret; | |
1535 | } | |
1536 | ||
1537 | /* /dev/vfio/$GROUP */ | |
cba3345c AW |
1538 | vfio.class = class_create(THIS_MODULE, "vfio"); |
1539 | if (IS_ERR(vfio.class)) { | |
1540 | ret = PTR_ERR(vfio.class); | |
1541 | goto err_class; | |
1542 | } | |
1543 | ||
1544 | vfio.class->devnode = vfio_devnode; | |
1545 | ||
d1099901 | 1546 | ret = alloc_chrdev_region(&vfio.group_devt, 0, MINORMASK, "vfio"); |
cba3345c | 1547 | if (ret) |
d1099901 | 1548 | goto err_alloc_chrdev; |
cba3345c | 1549 | |
cba3345c | 1550 | cdev_init(&vfio.group_cdev, &vfio_group_fops); |
d1099901 | 1551 | ret = cdev_add(&vfio.group_cdev, vfio.group_devt, MINORMASK); |
cba3345c | 1552 | if (ret) |
d1099901 | 1553 | goto err_cdev_add; |
cba3345c AW |
1554 | |
1555 | pr_info(DRIVER_DESC " version: " DRIVER_VERSION "\n"); | |
1556 | ||
73fa0d10 AW |
1557 | /* |
1558 | * Attempt to load known iommu-drivers. This gives us a working | |
1559 | * environment without the user needing to explicitly load iommu | |
1560 | * drivers. | |
1561 | */ | |
1562 | request_module_nowait("vfio_iommu_type1"); | |
5ffd229c | 1563 | request_module_nowait("vfio_iommu_spapr_tce"); |
73fa0d10 | 1564 | |
cba3345c AW |
1565 | return 0; |
1566 | ||
d1099901 AW |
1567 | err_cdev_add: |
1568 | unregister_chrdev_region(vfio.group_devt, MINORMASK); | |
1569 | err_alloc_chrdev: | |
cba3345c AW |
1570 | class_destroy(vfio.class); |
1571 | vfio.class = NULL; | |
1572 | err_class: | |
d1099901 | 1573 | misc_deregister(&vfio_dev); |
cba3345c AW |
1574 | return ret; |
1575 | } | |
1576 | ||
1577 | static void __exit vfio_cleanup(void) | |
1578 | { | |
1579 | WARN_ON(!list_empty(&vfio.group_list)); | |
1580 | ||
1581 | idr_destroy(&vfio.group_idr); | |
1582 | cdev_del(&vfio.group_cdev); | |
d1099901 | 1583 | unregister_chrdev_region(vfio.group_devt, MINORMASK); |
cba3345c AW |
1584 | class_destroy(vfio.class); |
1585 | vfio.class = NULL; | |
d1099901 | 1586 | misc_deregister(&vfio_dev); |
cba3345c AW |
1587 | } |
1588 | ||
1589 | module_init(vfio_init); | |
1590 | module_exit(vfio_cleanup); | |
1591 | ||
1592 | MODULE_VERSION(DRIVER_VERSION); | |
1593 | MODULE_LICENSE("GPL v2"); | |
1594 | MODULE_AUTHOR(DRIVER_AUTHOR); | |
1595 | MODULE_DESCRIPTION(DRIVER_DESC); | |
d1099901 AW |
1596 | MODULE_ALIAS_MISCDEV(VFIO_MINOR); |
1597 | MODULE_ALIAS("devname:vfio/vfio"); |