Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * VFIO-KVM bridge pseudo device
4 *
5 * Copyright (C) 2013 Red Hat, Inc. All rights reserved.
6 * Author: Alex Williamson <alex.williamson@redhat.com>
7 */
8
9#include <linux/errno.h>
10#include <linux/file.h>
11#include <linux/kvm_host.h>
12#include <linux/list.h>
13#include <linux/module.h>
14#include <linux/mutex.h>
15#include <linux/slab.h>
16#include <linux/uaccess.h>
17#include <linux/vfio.h>
18#include "vfio.h"
19
20#ifdef CONFIG_SPAPR_TCE_IOMMU
21#include <asm/kvm_ppc.h>
22#endif
23
24struct kvm_vfio_group {
25 struct list_head node;
26 struct file *file;
27#ifdef CONFIG_SPAPR_TCE_IOMMU
28 struct iommu_group *iommu_group;
29#endif
30};
31
32struct kvm_vfio {
33 struct list_head group_list;
34 struct mutex lock;
35 bool noncoherent;
36};
37
38static void kvm_vfio_file_set_kvm(struct file *file, struct kvm *kvm)
39{
40 void (*fn)(struct file *file, struct kvm *kvm);
41
42 fn = symbol_get(vfio_file_set_kvm);
43 if (!fn)
44 return;
45
46 fn(file, kvm);
47
48 symbol_put(vfio_file_set_kvm);
49}
50
51static bool kvm_vfio_file_enforced_coherent(struct file *file)
52{
53 bool (*fn)(struct file *file);
54 bool ret;
55
56 fn = symbol_get(vfio_file_enforced_coherent);
57 if (!fn)
58 return false;
59
60 ret = fn(file);
61
62 symbol_put(vfio_file_enforced_coherent);
63
64 return ret;
65}
66
67static bool kvm_vfio_file_is_group(struct file *file)
68{
69 bool (*fn)(struct file *file);
70 bool ret;
71
72 fn = symbol_get(vfio_file_is_group);
73 if (!fn)
74 return false;
75
76 ret = fn(file);
77
78 symbol_put(vfio_file_is_group);
79
80 return ret;
81}
82
83#ifdef CONFIG_SPAPR_TCE_IOMMU
84static struct iommu_group *kvm_vfio_file_iommu_group(struct file *file)
85{
86 struct iommu_group *(*fn)(struct file *file);
87 struct iommu_group *ret;
88
89 fn = symbol_get(vfio_file_iommu_group);
90 if (!fn)
91 return NULL;
92
93 ret = fn(file);
94
95 symbol_put(vfio_file_iommu_group);
96
97 return ret;
98}
99
100static void kvm_spapr_tce_release_vfio_group(struct kvm *kvm,
101 struct kvm_vfio_group *kvg)
102{
103 if (WARN_ON_ONCE(!kvg->iommu_group))
104 return;
105
106 kvm_spapr_tce_release_iommu_group(kvm, kvg->iommu_group);
107 iommu_group_put(kvg->iommu_group);
108 kvg->iommu_group = NULL;
109}
110#endif
111
112/*
113 * Groups can use the same or different IOMMU domains. If the same then
114 * adding a new group may change the coherency of groups we've previously
115 * been told about. We don't want to care about any of that so we retest
116 * each group and bail as soon as we find one that's noncoherent. This
117 * means we only ever [un]register_noncoherent_dma once for the whole device.
118 */
119static void kvm_vfio_update_coherency(struct kvm_device *dev)
120{
121 struct kvm_vfio *kv = dev->private;
122 bool noncoherent = false;
123 struct kvm_vfio_group *kvg;
124
125 mutex_lock(&kv->lock);
126
127 list_for_each_entry(kvg, &kv->group_list, node) {
128 if (!kvm_vfio_file_enforced_coherent(kvg->file)) {
129 noncoherent = true;
130 break;
131 }
132 }
133
134 if (noncoherent != kv->noncoherent) {
135 kv->noncoherent = noncoherent;
136
137 if (kv->noncoherent)
138 kvm_arch_register_noncoherent_dma(dev->kvm);
139 else
140 kvm_arch_unregister_noncoherent_dma(dev->kvm);
141 }
142
143 mutex_unlock(&kv->lock);
144}
145
146static int kvm_vfio_group_add(struct kvm_device *dev, unsigned int fd)
147{
148 struct kvm_vfio *kv = dev->private;
149 struct kvm_vfio_group *kvg;
150 struct file *filp;
151 int ret;
152
153 filp = fget(fd);
154 if (!filp)
155 return -EBADF;
156
157 /* Ensure the FD is a vfio group FD.*/
158 if (!kvm_vfio_file_is_group(filp)) {
159 ret = -EINVAL;
160 goto err_fput;
161 }
162
163 mutex_lock(&kv->lock);
164
165 list_for_each_entry(kvg, &kv->group_list, node) {
166 if (kvg->file == filp) {
167 ret = -EEXIST;
168 goto err_unlock;
169 }
170 }
171
172 kvg = kzalloc(sizeof(*kvg), GFP_KERNEL_ACCOUNT);
173 if (!kvg) {
174 ret = -ENOMEM;
175 goto err_unlock;
176 }
177
178 kvg->file = filp;
179 list_add_tail(&kvg->node, &kv->group_list);
180
181 kvm_arch_start_assignment(dev->kvm);
182
183 mutex_unlock(&kv->lock);
184
185 kvm_vfio_file_set_kvm(kvg->file, dev->kvm);
186 kvm_vfio_update_coherency(dev);
187
188 return 0;
189err_unlock:
190 mutex_unlock(&kv->lock);
191err_fput:
192 fput(filp);
193 return ret;
194}
195
196static int kvm_vfio_group_del(struct kvm_device *dev, unsigned int fd)
197{
198 struct kvm_vfio *kv = dev->private;
199 struct kvm_vfio_group *kvg;
200 struct fd f;
201 int ret;
202
203 f = fdget(fd);
204 if (!f.file)
205 return -EBADF;
206
207 ret = -ENOENT;
208
209 mutex_lock(&kv->lock);
210
211 list_for_each_entry(kvg, &kv->group_list, node) {
212 if (kvg->file != f.file)
213 continue;
214
215 list_del(&kvg->node);
216 kvm_arch_end_assignment(dev->kvm);
217#ifdef CONFIG_SPAPR_TCE_IOMMU
218 kvm_spapr_tce_release_vfio_group(dev->kvm, kvg);
219#endif
220 kvm_vfio_file_set_kvm(kvg->file, NULL);
221 fput(kvg->file);
222 kfree(kvg);
223 ret = 0;
224 break;
225 }
226
227 mutex_unlock(&kv->lock);
228
229 fdput(f);
230
231 kvm_vfio_update_coherency(dev);
232
233 return ret;
234}
235
236#ifdef CONFIG_SPAPR_TCE_IOMMU
237static int kvm_vfio_group_set_spapr_tce(struct kvm_device *dev,
238 void __user *arg)
239{
240 struct kvm_vfio_spapr_tce param;
241 struct kvm_vfio *kv = dev->private;
242 struct kvm_vfio_group *kvg;
243 struct fd f;
244 int ret;
245
246 if (copy_from_user(¶m, arg, sizeof(struct kvm_vfio_spapr_tce)))
247 return -EFAULT;
248
249 f = fdget(param.groupfd);
250 if (!f.file)
251 return -EBADF;
252
253 ret = -ENOENT;
254
255 mutex_lock(&kv->lock);
256
257 list_for_each_entry(kvg, &kv->group_list, node) {
258 if (kvg->file != f.file)
259 continue;
260
261 if (!kvg->iommu_group) {
262 kvg->iommu_group = kvm_vfio_file_iommu_group(kvg->file);
263 if (WARN_ON_ONCE(!kvg->iommu_group)) {
264 ret = -EIO;
265 goto err_fdput;
266 }
267 }
268
269 ret = kvm_spapr_tce_attach_iommu_group(dev->kvm, param.tablefd,
270 kvg->iommu_group);
271 break;
272 }
273
274err_fdput:
275 mutex_unlock(&kv->lock);
276 fdput(f);
277 return ret;
278}
279#endif
280
281static int kvm_vfio_set_group(struct kvm_device *dev, long attr,
282 void __user *arg)
283{
284 int32_t __user *argp = arg;
285 int32_t fd;
286
287 switch (attr) {
288 case KVM_DEV_VFIO_GROUP_ADD:
289 if (get_user(fd, argp))
290 return -EFAULT;
291 return kvm_vfio_group_add(dev, fd);
292
293 case KVM_DEV_VFIO_GROUP_DEL:
294 if (get_user(fd, argp))
295 return -EFAULT;
296 return kvm_vfio_group_del(dev, fd);
297
298#ifdef CONFIG_SPAPR_TCE_IOMMU
299 case KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE:
300 return kvm_vfio_group_set_spapr_tce(dev, arg);
301#endif
302 }
303
304 return -ENXIO;
305}
306
307static int kvm_vfio_set_attr(struct kvm_device *dev,
308 struct kvm_device_attr *attr)
309{
310 switch (attr->group) {
311 case KVM_DEV_VFIO_GROUP:
312 return kvm_vfio_set_group(dev, attr->attr,
313 u64_to_user_ptr(attr->addr));
314 }
315
316 return -ENXIO;
317}
318
319static int kvm_vfio_has_attr(struct kvm_device *dev,
320 struct kvm_device_attr *attr)
321{
322 switch (attr->group) {
323 case KVM_DEV_VFIO_GROUP:
324 switch (attr->attr) {
325 case KVM_DEV_VFIO_GROUP_ADD:
326 case KVM_DEV_VFIO_GROUP_DEL:
327#ifdef CONFIG_SPAPR_TCE_IOMMU
328 case KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE:
329#endif
330 return 0;
331 }
332
333 break;
334 }
335
336 return -ENXIO;
337}
338
339static void kvm_vfio_release(struct kvm_device *dev)
340{
341 struct kvm_vfio *kv = dev->private;
342 struct kvm_vfio_group *kvg, *tmp;
343
344 list_for_each_entry_safe(kvg, tmp, &kv->group_list, node) {
345#ifdef CONFIG_SPAPR_TCE_IOMMU
346 kvm_spapr_tce_release_vfio_group(dev->kvm, kvg);
347#endif
348 kvm_vfio_file_set_kvm(kvg->file, NULL);
349 fput(kvg->file);
350 list_del(&kvg->node);
351 kfree(kvg);
352 kvm_arch_end_assignment(dev->kvm);
353 }
354
355 kvm_vfio_update_coherency(dev);
356
357 kfree(kv);
358 kfree(dev); /* alloc by kvm_ioctl_create_device, free by .release */
359}
360
361static int kvm_vfio_create(struct kvm_device *dev, u32 type);
362
363static struct kvm_device_ops kvm_vfio_ops = {
364 .name = "kvm-vfio",
365 .create = kvm_vfio_create,
366 .release = kvm_vfio_release,
367 .set_attr = kvm_vfio_set_attr,
368 .has_attr = kvm_vfio_has_attr,
369};
370
371static int kvm_vfio_create(struct kvm_device *dev, u32 type)
372{
373 struct kvm_device *tmp;
374 struct kvm_vfio *kv;
375
376 /* Only one VFIO "device" per VM */
377 list_for_each_entry(tmp, &dev->kvm->devices, vm_node)
378 if (tmp->ops == &kvm_vfio_ops)
379 return -EBUSY;
380
381 kv = kzalloc(sizeof(*kv), GFP_KERNEL_ACCOUNT);
382 if (!kv)
383 return -ENOMEM;
384
385 INIT_LIST_HEAD(&kv->group_list);
386 mutex_init(&kv->lock);
387
388 dev->private = kv;
389
390 return 0;
391}
392
393int kvm_vfio_ops_init(void)
394{
395 return kvm_register_device_ops(&kvm_vfio_ops, KVM_DEV_TYPE_VFIO);
396}
397
398void kvm_vfio_ops_exit(void)
399{
400 kvm_unregister_device_ops(KVM_DEV_TYPE_VFIO);
401}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * VFIO-KVM bridge pseudo device
4 *
5 * Copyright (C) 2013 Red Hat, Inc. All rights reserved.
6 * Author: Alex Williamson <alex.williamson@redhat.com>
7 */
8
9#include <linux/errno.h>
10#include <linux/file.h>
11#include <linux/kvm_host.h>
12#include <linux/list.h>
13#include <linux/module.h>
14#include <linux/mutex.h>
15#include <linux/slab.h>
16#include <linux/uaccess.h>
17#include <linux/vfio.h>
18#include "vfio.h"
19
20#ifdef CONFIG_SPAPR_TCE_IOMMU
21#include <asm/kvm_ppc.h>
22#endif
23
24struct kvm_vfio_file {
25 struct list_head node;
26 struct file *file;
27#ifdef CONFIG_SPAPR_TCE_IOMMU
28 struct iommu_group *iommu_group;
29#endif
30};
31
32struct kvm_vfio {
33 struct list_head file_list;
34 struct mutex lock;
35 bool noncoherent;
36};
37
38static void kvm_vfio_file_set_kvm(struct file *file, struct kvm *kvm)
39{
40 void (*fn)(struct file *file, struct kvm *kvm);
41
42 fn = symbol_get(vfio_file_set_kvm);
43 if (!fn)
44 return;
45
46 fn(file, kvm);
47
48 symbol_put(vfio_file_set_kvm);
49}
50
51static bool kvm_vfio_file_enforced_coherent(struct file *file)
52{
53 bool (*fn)(struct file *file);
54 bool ret;
55
56 fn = symbol_get(vfio_file_enforced_coherent);
57 if (!fn)
58 return false;
59
60 ret = fn(file);
61
62 symbol_put(vfio_file_enforced_coherent);
63
64 return ret;
65}
66
67static bool kvm_vfio_file_is_valid(struct file *file)
68{
69 bool (*fn)(struct file *file);
70 bool ret;
71
72 fn = symbol_get(vfio_file_is_valid);
73 if (!fn)
74 return false;
75
76 ret = fn(file);
77
78 symbol_put(vfio_file_is_valid);
79
80 return ret;
81}
82
83#ifdef CONFIG_SPAPR_TCE_IOMMU
84static struct iommu_group *kvm_vfio_file_iommu_group(struct file *file)
85{
86 struct iommu_group *(*fn)(struct file *file);
87 struct iommu_group *ret;
88
89 fn = symbol_get(vfio_file_iommu_group);
90 if (!fn)
91 return NULL;
92
93 ret = fn(file);
94
95 symbol_put(vfio_file_iommu_group);
96
97 return ret;
98}
99
100static void kvm_spapr_tce_release_vfio_group(struct kvm *kvm,
101 struct kvm_vfio_file *kvf)
102{
103 if (WARN_ON_ONCE(!kvf->iommu_group))
104 return;
105
106 kvm_spapr_tce_release_iommu_group(kvm, kvf->iommu_group);
107 iommu_group_put(kvf->iommu_group);
108 kvf->iommu_group = NULL;
109}
110#endif
111
112/*
113 * Groups/devices can use the same or different IOMMU domains. If the same
114 * then adding a new group/device may change the coherency of groups/devices
115 * we've previously been told about. We don't want to care about any of
116 * that so we retest each group/device and bail as soon as we find one that's
117 * noncoherent. This means we only ever [un]register_noncoherent_dma once
118 * for the whole device.
119 */
120static void kvm_vfio_update_coherency(struct kvm_device *dev)
121{
122 struct kvm_vfio *kv = dev->private;
123 bool noncoherent = false;
124 struct kvm_vfio_file *kvf;
125
126 list_for_each_entry(kvf, &kv->file_list, node) {
127 if (!kvm_vfio_file_enforced_coherent(kvf->file)) {
128 noncoherent = true;
129 break;
130 }
131 }
132
133 if (noncoherent != kv->noncoherent) {
134 kv->noncoherent = noncoherent;
135
136 if (kv->noncoherent)
137 kvm_arch_register_noncoherent_dma(dev->kvm);
138 else
139 kvm_arch_unregister_noncoherent_dma(dev->kvm);
140 }
141}
142
143static int kvm_vfio_file_add(struct kvm_device *dev, unsigned int fd)
144{
145 struct kvm_vfio *kv = dev->private;
146 struct kvm_vfio_file *kvf;
147 struct file *filp;
148 int ret = 0;
149
150 filp = fget(fd);
151 if (!filp)
152 return -EBADF;
153
154 /* Ensure the FD is a vfio FD. */
155 if (!kvm_vfio_file_is_valid(filp)) {
156 ret = -EINVAL;
157 goto out_fput;
158 }
159
160 mutex_lock(&kv->lock);
161
162 list_for_each_entry(kvf, &kv->file_list, node) {
163 if (kvf->file == filp) {
164 ret = -EEXIST;
165 goto out_unlock;
166 }
167 }
168
169 kvf = kzalloc(sizeof(*kvf), GFP_KERNEL_ACCOUNT);
170 if (!kvf) {
171 ret = -ENOMEM;
172 goto out_unlock;
173 }
174
175 kvf->file = get_file(filp);
176 list_add_tail(&kvf->node, &kv->file_list);
177
178 kvm_arch_start_assignment(dev->kvm);
179 kvm_vfio_file_set_kvm(kvf->file, dev->kvm);
180 kvm_vfio_update_coherency(dev);
181
182out_unlock:
183 mutex_unlock(&kv->lock);
184out_fput:
185 fput(filp);
186 return ret;
187}
188
189static int kvm_vfio_file_del(struct kvm_device *dev, unsigned int fd)
190{
191 struct kvm_vfio *kv = dev->private;
192 struct kvm_vfio_file *kvf;
193 CLASS(fd, f)(fd);
194 int ret;
195
196 if (fd_empty(f))
197 return -EBADF;
198
199 ret = -ENOENT;
200
201 mutex_lock(&kv->lock);
202
203 list_for_each_entry(kvf, &kv->file_list, node) {
204 if (kvf->file != fd_file(f))
205 continue;
206
207 list_del(&kvf->node);
208 kvm_arch_end_assignment(dev->kvm);
209#ifdef CONFIG_SPAPR_TCE_IOMMU
210 kvm_spapr_tce_release_vfio_group(dev->kvm, kvf);
211#endif
212 kvm_vfio_file_set_kvm(kvf->file, NULL);
213 fput(kvf->file);
214 kfree(kvf);
215 ret = 0;
216 break;
217 }
218
219 kvm_vfio_update_coherency(dev);
220
221 mutex_unlock(&kv->lock);
222 return ret;
223}
224
225#ifdef CONFIG_SPAPR_TCE_IOMMU
226static int kvm_vfio_file_set_spapr_tce(struct kvm_device *dev,
227 void __user *arg)
228{
229 struct kvm_vfio_spapr_tce param;
230 struct kvm_vfio *kv = dev->private;
231 struct kvm_vfio_file *kvf;
232 int ret;
233
234 if (copy_from_user(¶m, arg, sizeof(struct kvm_vfio_spapr_tce)))
235 return -EFAULT;
236
237 CLASS(fd, f)(param.groupfd);
238 if (fd_empty(f))
239 return -EBADF;
240
241 ret = -ENOENT;
242
243 mutex_lock(&kv->lock);
244
245 list_for_each_entry(kvf, &kv->file_list, node) {
246 if (kvf->file != fd_file(f))
247 continue;
248
249 if (!kvf->iommu_group) {
250 kvf->iommu_group = kvm_vfio_file_iommu_group(kvf->file);
251 if (WARN_ON_ONCE(!kvf->iommu_group)) {
252 ret = -EIO;
253 goto err_fdput;
254 }
255 }
256
257 ret = kvm_spapr_tce_attach_iommu_group(dev->kvm, param.tablefd,
258 kvf->iommu_group);
259 break;
260 }
261
262err_fdput:
263 mutex_unlock(&kv->lock);
264 return ret;
265}
266#endif
267
268static int kvm_vfio_set_file(struct kvm_device *dev, long attr,
269 void __user *arg)
270{
271 int32_t __user *argp = arg;
272 int32_t fd;
273
274 switch (attr) {
275 case KVM_DEV_VFIO_FILE_ADD:
276 if (get_user(fd, argp))
277 return -EFAULT;
278 return kvm_vfio_file_add(dev, fd);
279
280 case KVM_DEV_VFIO_FILE_DEL:
281 if (get_user(fd, argp))
282 return -EFAULT;
283 return kvm_vfio_file_del(dev, fd);
284
285#ifdef CONFIG_SPAPR_TCE_IOMMU
286 case KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE:
287 return kvm_vfio_file_set_spapr_tce(dev, arg);
288#endif
289 }
290
291 return -ENXIO;
292}
293
294static int kvm_vfio_set_attr(struct kvm_device *dev,
295 struct kvm_device_attr *attr)
296{
297 switch (attr->group) {
298 case KVM_DEV_VFIO_FILE:
299 return kvm_vfio_set_file(dev, attr->attr,
300 u64_to_user_ptr(attr->addr));
301 }
302
303 return -ENXIO;
304}
305
306static int kvm_vfio_has_attr(struct kvm_device *dev,
307 struct kvm_device_attr *attr)
308{
309 switch (attr->group) {
310 case KVM_DEV_VFIO_FILE:
311 switch (attr->attr) {
312 case KVM_DEV_VFIO_FILE_ADD:
313 case KVM_DEV_VFIO_FILE_DEL:
314#ifdef CONFIG_SPAPR_TCE_IOMMU
315 case KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE:
316#endif
317 return 0;
318 }
319
320 break;
321 }
322
323 return -ENXIO;
324}
325
326static void kvm_vfio_release(struct kvm_device *dev)
327{
328 struct kvm_vfio *kv = dev->private;
329 struct kvm_vfio_file *kvf, *tmp;
330
331 list_for_each_entry_safe(kvf, tmp, &kv->file_list, node) {
332#ifdef CONFIG_SPAPR_TCE_IOMMU
333 kvm_spapr_tce_release_vfio_group(dev->kvm, kvf);
334#endif
335 kvm_vfio_file_set_kvm(kvf->file, NULL);
336 fput(kvf->file);
337 list_del(&kvf->node);
338 kfree(kvf);
339 kvm_arch_end_assignment(dev->kvm);
340 }
341
342 kvm_vfio_update_coherency(dev);
343
344 kfree(kv);
345 kfree(dev); /* alloc by kvm_ioctl_create_device, free by .release */
346}
347
348static int kvm_vfio_create(struct kvm_device *dev, u32 type);
349
350static const struct kvm_device_ops kvm_vfio_ops = {
351 .name = "kvm-vfio",
352 .create = kvm_vfio_create,
353 .release = kvm_vfio_release,
354 .set_attr = kvm_vfio_set_attr,
355 .has_attr = kvm_vfio_has_attr,
356};
357
358static int kvm_vfio_create(struct kvm_device *dev, u32 type)
359{
360 struct kvm_device *tmp;
361 struct kvm_vfio *kv;
362
363 lockdep_assert_held(&dev->kvm->lock);
364
365 /* Only one VFIO "device" per VM */
366 list_for_each_entry(tmp, &dev->kvm->devices, vm_node)
367 if (tmp->ops == &kvm_vfio_ops)
368 return -EBUSY;
369
370 kv = kzalloc(sizeof(*kv), GFP_KERNEL_ACCOUNT);
371 if (!kv)
372 return -ENOMEM;
373
374 INIT_LIST_HEAD(&kv->file_list);
375 mutex_init(&kv->lock);
376
377 dev->private = kv;
378
379 return 0;
380}
381
382int kvm_vfio_ops_init(void)
383{
384 return kvm_register_device_ops(&kvm_vfio_ops, KVM_DEV_TYPE_VFIO);
385}
386
387void kvm_vfio_ops_exit(void)
388{
389 kvm_unregister_device_ops(KVM_DEV_TYPE_VFIO);
390}