Loading...
1/*
2 * VFIO-KVM bridge pseudo device
3 *
4 * Copyright (C) 2013 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
12#include <linux/errno.h>
13#include <linux/file.h>
14#include <linux/kvm_host.h>
15#include <linux/list.h>
16#include <linux/module.h>
17#include <linux/mutex.h>
18#include <linux/slab.h>
19#include <linux/uaccess.h>
20#include <linux/vfio.h>
21#include "vfio.h"
22
23struct kvm_vfio_group {
24 struct list_head node;
25 struct vfio_group *vfio_group;
26};
27
28struct kvm_vfio {
29 struct list_head group_list;
30 struct mutex lock;
31 bool noncoherent;
32};
33
34static struct vfio_group *kvm_vfio_group_get_external_user(struct file *filep)
35{
36 struct vfio_group *vfio_group;
37 struct vfio_group *(*fn)(struct file *);
38
39 fn = symbol_get(vfio_group_get_external_user);
40 if (!fn)
41 return ERR_PTR(-EINVAL);
42
43 vfio_group = fn(filep);
44
45 symbol_put(vfio_group_get_external_user);
46
47 return vfio_group;
48}
49
50static void kvm_vfio_group_put_external_user(struct vfio_group *vfio_group)
51{
52 void (*fn)(struct vfio_group *);
53
54 fn = symbol_get(vfio_group_put_external_user);
55 if (!fn)
56 return;
57
58 fn(vfio_group);
59
60 symbol_put(vfio_group_put_external_user);
61}
62
63static bool kvm_vfio_group_is_coherent(struct vfio_group *vfio_group)
64{
65 long (*fn)(struct vfio_group *, unsigned long);
66 long ret;
67
68 fn = symbol_get(vfio_external_check_extension);
69 if (!fn)
70 return false;
71
72 ret = fn(vfio_group, VFIO_DMA_CC_IOMMU);
73
74 symbol_put(vfio_external_check_extension);
75
76 return ret > 0;
77}
78
79/*
80 * Groups can use the same or different IOMMU domains. If the same then
81 * adding a new group may change the coherency of groups we've previously
82 * been told about. We don't want to care about any of that so we retest
83 * each group and bail as soon as we find one that's noncoherent. This
84 * means we only ever [un]register_noncoherent_dma once for the whole device.
85 */
86static void kvm_vfio_update_coherency(struct kvm_device *dev)
87{
88 struct kvm_vfio *kv = dev->private;
89 bool noncoherent = false;
90 struct kvm_vfio_group *kvg;
91
92 mutex_lock(&kv->lock);
93
94 list_for_each_entry(kvg, &kv->group_list, node) {
95 if (!kvm_vfio_group_is_coherent(kvg->vfio_group)) {
96 noncoherent = true;
97 break;
98 }
99 }
100
101 if (noncoherent != kv->noncoherent) {
102 kv->noncoherent = noncoherent;
103
104 if (kv->noncoherent)
105 kvm_arch_register_noncoherent_dma(dev->kvm);
106 else
107 kvm_arch_unregister_noncoherent_dma(dev->kvm);
108 }
109
110 mutex_unlock(&kv->lock);
111}
112
113static int kvm_vfio_set_group(struct kvm_device *dev, long attr, u64 arg)
114{
115 struct kvm_vfio *kv = dev->private;
116 struct vfio_group *vfio_group;
117 struct kvm_vfio_group *kvg;
118 int32_t __user *argp = (int32_t __user *)(unsigned long)arg;
119 struct fd f;
120 int32_t fd;
121 int ret;
122
123 switch (attr) {
124 case KVM_DEV_VFIO_GROUP_ADD:
125 if (get_user(fd, argp))
126 return -EFAULT;
127
128 f = fdget(fd);
129 if (!f.file)
130 return -EBADF;
131
132 vfio_group = kvm_vfio_group_get_external_user(f.file);
133 fdput(f);
134
135 if (IS_ERR(vfio_group))
136 return PTR_ERR(vfio_group);
137
138 mutex_lock(&kv->lock);
139
140 list_for_each_entry(kvg, &kv->group_list, node) {
141 if (kvg->vfio_group == vfio_group) {
142 mutex_unlock(&kv->lock);
143 kvm_vfio_group_put_external_user(vfio_group);
144 return -EEXIST;
145 }
146 }
147
148 kvg = kzalloc(sizeof(*kvg), GFP_KERNEL);
149 if (!kvg) {
150 mutex_unlock(&kv->lock);
151 kvm_vfio_group_put_external_user(vfio_group);
152 return -ENOMEM;
153 }
154
155 list_add_tail(&kvg->node, &kv->group_list);
156 kvg->vfio_group = vfio_group;
157
158 kvm_arch_start_assignment(dev->kvm);
159
160 mutex_unlock(&kv->lock);
161
162 kvm_vfio_update_coherency(dev);
163
164 return 0;
165
166 case KVM_DEV_VFIO_GROUP_DEL:
167 if (get_user(fd, argp))
168 return -EFAULT;
169
170 f = fdget(fd);
171 if (!f.file)
172 return -EBADF;
173
174 vfio_group = kvm_vfio_group_get_external_user(f.file);
175 fdput(f);
176
177 if (IS_ERR(vfio_group))
178 return PTR_ERR(vfio_group);
179
180 ret = -ENOENT;
181
182 mutex_lock(&kv->lock);
183
184 list_for_each_entry(kvg, &kv->group_list, node) {
185 if (kvg->vfio_group != vfio_group)
186 continue;
187
188 list_del(&kvg->node);
189 kvm_vfio_group_put_external_user(kvg->vfio_group);
190 kfree(kvg);
191 ret = 0;
192 break;
193 }
194
195 kvm_arch_end_assignment(dev->kvm);
196
197 mutex_unlock(&kv->lock);
198
199 kvm_vfio_group_put_external_user(vfio_group);
200
201 kvm_vfio_update_coherency(dev);
202
203 return ret;
204 }
205
206 return -ENXIO;
207}
208
209static int kvm_vfio_set_attr(struct kvm_device *dev,
210 struct kvm_device_attr *attr)
211{
212 switch (attr->group) {
213 case KVM_DEV_VFIO_GROUP:
214 return kvm_vfio_set_group(dev, attr->attr, attr->addr);
215 }
216
217 return -ENXIO;
218}
219
220static int kvm_vfio_has_attr(struct kvm_device *dev,
221 struct kvm_device_attr *attr)
222{
223 switch (attr->group) {
224 case KVM_DEV_VFIO_GROUP:
225 switch (attr->attr) {
226 case KVM_DEV_VFIO_GROUP_ADD:
227 case KVM_DEV_VFIO_GROUP_DEL:
228 return 0;
229 }
230
231 break;
232 }
233
234 return -ENXIO;
235}
236
237static void kvm_vfio_destroy(struct kvm_device *dev)
238{
239 struct kvm_vfio *kv = dev->private;
240 struct kvm_vfio_group *kvg, *tmp;
241
242 list_for_each_entry_safe(kvg, tmp, &kv->group_list, node) {
243 kvm_vfio_group_put_external_user(kvg->vfio_group);
244 list_del(&kvg->node);
245 kfree(kvg);
246 kvm_arch_end_assignment(dev->kvm);
247 }
248
249 kvm_vfio_update_coherency(dev);
250
251 kfree(kv);
252 kfree(dev); /* alloc by kvm_ioctl_create_device, free by .destroy */
253}
254
255static int kvm_vfio_create(struct kvm_device *dev, u32 type);
256
257static struct kvm_device_ops kvm_vfio_ops = {
258 .name = "kvm-vfio",
259 .create = kvm_vfio_create,
260 .destroy = kvm_vfio_destroy,
261 .set_attr = kvm_vfio_set_attr,
262 .has_attr = kvm_vfio_has_attr,
263};
264
265static int kvm_vfio_create(struct kvm_device *dev, u32 type)
266{
267 struct kvm_device *tmp;
268 struct kvm_vfio *kv;
269
270 /* Only one VFIO "device" per VM */
271 list_for_each_entry(tmp, &dev->kvm->devices, vm_node)
272 if (tmp->ops == &kvm_vfio_ops)
273 return -EBUSY;
274
275 kv = kzalloc(sizeof(*kv), GFP_KERNEL);
276 if (!kv)
277 return -ENOMEM;
278
279 INIT_LIST_HEAD(&kv->group_list);
280 mutex_init(&kv->lock);
281
282 dev->private = kv;
283
284 return 0;
285}
286
287int kvm_vfio_ops_init(void)
288{
289 return kvm_register_device_ops(&kvm_vfio_ops, KVM_DEV_TYPE_VFIO);
290}
291
292void kvm_vfio_ops_exit(void)
293{
294 kvm_unregister_device_ops(KVM_DEV_TYPE_VFIO);
295}
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 vfio_group *vfio_group;
27};
28
29struct kvm_vfio {
30 struct list_head group_list;
31 struct mutex lock;
32 bool noncoherent;
33};
34
35static struct vfio_group *kvm_vfio_group_get_external_user(struct file *filep)
36{
37 struct vfio_group *vfio_group;
38 struct vfio_group *(*fn)(struct file *);
39
40 fn = symbol_get(vfio_group_get_external_user);
41 if (!fn)
42 return ERR_PTR(-EINVAL);
43
44 vfio_group = fn(filep);
45
46 symbol_put(vfio_group_get_external_user);
47
48 return vfio_group;
49}
50
51static bool kvm_vfio_external_group_match_file(struct vfio_group *group,
52 struct file *filep)
53{
54 bool ret, (*fn)(struct vfio_group *, struct file *);
55
56 fn = symbol_get(vfio_external_group_match_file);
57 if (!fn)
58 return false;
59
60 ret = fn(group, filep);
61
62 symbol_put(vfio_external_group_match_file);
63
64 return ret;
65}
66
67static void kvm_vfio_group_put_external_user(struct vfio_group *vfio_group)
68{
69 void (*fn)(struct vfio_group *);
70
71 fn = symbol_get(vfio_group_put_external_user);
72 if (!fn)
73 return;
74
75 fn(vfio_group);
76
77 symbol_put(vfio_group_put_external_user);
78}
79
80static void kvm_vfio_group_set_kvm(struct vfio_group *group, struct kvm *kvm)
81{
82 void (*fn)(struct vfio_group *, struct kvm *);
83
84 fn = symbol_get(vfio_group_set_kvm);
85 if (!fn)
86 return;
87
88 fn(group, kvm);
89
90 symbol_put(vfio_group_set_kvm);
91}
92
93static bool kvm_vfio_group_is_coherent(struct vfio_group *vfio_group)
94{
95 long (*fn)(struct vfio_group *, unsigned long);
96 long ret;
97
98 fn = symbol_get(vfio_external_check_extension);
99 if (!fn)
100 return false;
101
102 ret = fn(vfio_group, VFIO_DMA_CC_IOMMU);
103
104 symbol_put(vfio_external_check_extension);
105
106 return ret > 0;
107}
108
109#ifdef CONFIG_SPAPR_TCE_IOMMU
110static int kvm_vfio_external_user_iommu_id(struct vfio_group *vfio_group)
111{
112 int (*fn)(struct vfio_group *);
113 int ret = -EINVAL;
114
115 fn = symbol_get(vfio_external_user_iommu_id);
116 if (!fn)
117 return ret;
118
119 ret = fn(vfio_group);
120
121 symbol_put(vfio_external_user_iommu_id);
122
123 return ret;
124}
125
126static struct iommu_group *kvm_vfio_group_get_iommu_group(
127 struct vfio_group *group)
128{
129 int group_id = kvm_vfio_external_user_iommu_id(group);
130
131 if (group_id < 0)
132 return NULL;
133
134 return iommu_group_get_by_id(group_id);
135}
136
137static void kvm_spapr_tce_release_vfio_group(struct kvm *kvm,
138 struct vfio_group *vfio_group)
139{
140 struct iommu_group *grp = kvm_vfio_group_get_iommu_group(vfio_group);
141
142 if (WARN_ON_ONCE(!grp))
143 return;
144
145 kvm_spapr_tce_release_iommu_group(kvm, grp);
146 iommu_group_put(grp);
147}
148#endif
149
150/*
151 * Groups can use the same or different IOMMU domains. If the same then
152 * adding a new group may change the coherency of groups we've previously
153 * been told about. We don't want to care about any of that so we retest
154 * each group and bail as soon as we find one that's noncoherent. This
155 * means we only ever [un]register_noncoherent_dma once for the whole device.
156 */
157static void kvm_vfio_update_coherency(struct kvm_device *dev)
158{
159 struct kvm_vfio *kv = dev->private;
160 bool noncoherent = false;
161 struct kvm_vfio_group *kvg;
162
163 mutex_lock(&kv->lock);
164
165 list_for_each_entry(kvg, &kv->group_list, node) {
166 if (!kvm_vfio_group_is_coherent(kvg->vfio_group)) {
167 noncoherent = true;
168 break;
169 }
170 }
171
172 if (noncoherent != kv->noncoherent) {
173 kv->noncoherent = noncoherent;
174
175 if (kv->noncoherent)
176 kvm_arch_register_noncoherent_dma(dev->kvm);
177 else
178 kvm_arch_unregister_noncoherent_dma(dev->kvm);
179 }
180
181 mutex_unlock(&kv->lock);
182}
183
184static int kvm_vfio_set_group(struct kvm_device *dev, long attr, u64 arg)
185{
186 struct kvm_vfio *kv = dev->private;
187 struct vfio_group *vfio_group;
188 struct kvm_vfio_group *kvg;
189 int32_t __user *argp = (int32_t __user *)(unsigned long)arg;
190 struct fd f;
191 int32_t fd;
192 int ret;
193
194 switch (attr) {
195 case KVM_DEV_VFIO_GROUP_ADD:
196 if (get_user(fd, argp))
197 return -EFAULT;
198
199 f = fdget(fd);
200 if (!f.file)
201 return -EBADF;
202
203 vfio_group = kvm_vfio_group_get_external_user(f.file);
204 fdput(f);
205
206 if (IS_ERR(vfio_group))
207 return PTR_ERR(vfio_group);
208
209 mutex_lock(&kv->lock);
210
211 list_for_each_entry(kvg, &kv->group_list, node) {
212 if (kvg->vfio_group == vfio_group) {
213 mutex_unlock(&kv->lock);
214 kvm_vfio_group_put_external_user(vfio_group);
215 return -EEXIST;
216 }
217 }
218
219 kvg = kzalloc(sizeof(*kvg), GFP_KERNEL_ACCOUNT);
220 if (!kvg) {
221 mutex_unlock(&kv->lock);
222 kvm_vfio_group_put_external_user(vfio_group);
223 return -ENOMEM;
224 }
225
226 list_add_tail(&kvg->node, &kv->group_list);
227 kvg->vfio_group = vfio_group;
228
229 kvm_arch_start_assignment(dev->kvm);
230
231 mutex_unlock(&kv->lock);
232
233 kvm_vfio_group_set_kvm(vfio_group, dev->kvm);
234
235 kvm_vfio_update_coherency(dev);
236
237 return 0;
238
239 case KVM_DEV_VFIO_GROUP_DEL:
240 if (get_user(fd, argp))
241 return -EFAULT;
242
243 f = fdget(fd);
244 if (!f.file)
245 return -EBADF;
246
247 ret = -ENOENT;
248
249 mutex_lock(&kv->lock);
250
251 list_for_each_entry(kvg, &kv->group_list, node) {
252 if (!kvm_vfio_external_group_match_file(kvg->vfio_group,
253 f.file))
254 continue;
255
256 list_del(&kvg->node);
257 kvm_arch_end_assignment(dev->kvm);
258#ifdef CONFIG_SPAPR_TCE_IOMMU
259 kvm_spapr_tce_release_vfio_group(dev->kvm,
260 kvg->vfio_group);
261#endif
262 kvm_vfio_group_set_kvm(kvg->vfio_group, NULL);
263 kvm_vfio_group_put_external_user(kvg->vfio_group);
264 kfree(kvg);
265 ret = 0;
266 break;
267 }
268
269 mutex_unlock(&kv->lock);
270
271 fdput(f);
272
273 kvm_vfio_update_coherency(dev);
274
275 return ret;
276
277#ifdef CONFIG_SPAPR_TCE_IOMMU
278 case KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE: {
279 struct kvm_vfio_spapr_tce param;
280 struct kvm_vfio *kv = dev->private;
281 struct vfio_group *vfio_group;
282 struct kvm_vfio_group *kvg;
283 struct fd f;
284 struct iommu_group *grp;
285
286 if (copy_from_user(¶m, (void __user *)arg,
287 sizeof(struct kvm_vfio_spapr_tce)))
288 return -EFAULT;
289
290 f = fdget(param.groupfd);
291 if (!f.file)
292 return -EBADF;
293
294 vfio_group = kvm_vfio_group_get_external_user(f.file);
295 fdput(f);
296
297 if (IS_ERR(vfio_group))
298 return PTR_ERR(vfio_group);
299
300 grp = kvm_vfio_group_get_iommu_group(vfio_group);
301 if (WARN_ON_ONCE(!grp)) {
302 kvm_vfio_group_put_external_user(vfio_group);
303 return -EIO;
304 }
305
306 ret = -ENOENT;
307
308 mutex_lock(&kv->lock);
309
310 list_for_each_entry(kvg, &kv->group_list, node) {
311 if (kvg->vfio_group != vfio_group)
312 continue;
313
314 ret = kvm_spapr_tce_attach_iommu_group(dev->kvm,
315 param.tablefd, grp);
316 break;
317 }
318
319 mutex_unlock(&kv->lock);
320
321 iommu_group_put(grp);
322 kvm_vfio_group_put_external_user(vfio_group);
323
324 return ret;
325 }
326#endif /* CONFIG_SPAPR_TCE_IOMMU */
327 }
328
329 return -ENXIO;
330}
331
332static int kvm_vfio_set_attr(struct kvm_device *dev,
333 struct kvm_device_attr *attr)
334{
335 switch (attr->group) {
336 case KVM_DEV_VFIO_GROUP:
337 return kvm_vfio_set_group(dev, attr->attr, attr->addr);
338 }
339
340 return -ENXIO;
341}
342
343static int kvm_vfio_has_attr(struct kvm_device *dev,
344 struct kvm_device_attr *attr)
345{
346 switch (attr->group) {
347 case KVM_DEV_VFIO_GROUP:
348 switch (attr->attr) {
349 case KVM_DEV_VFIO_GROUP_ADD:
350 case KVM_DEV_VFIO_GROUP_DEL:
351#ifdef CONFIG_SPAPR_TCE_IOMMU
352 case KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE:
353#endif
354 return 0;
355 }
356
357 break;
358 }
359
360 return -ENXIO;
361}
362
363static void kvm_vfio_destroy(struct kvm_device *dev)
364{
365 struct kvm_vfio *kv = dev->private;
366 struct kvm_vfio_group *kvg, *tmp;
367
368 list_for_each_entry_safe(kvg, tmp, &kv->group_list, node) {
369#ifdef CONFIG_SPAPR_TCE_IOMMU
370 kvm_spapr_tce_release_vfio_group(dev->kvm, kvg->vfio_group);
371#endif
372 kvm_vfio_group_set_kvm(kvg->vfio_group, NULL);
373 kvm_vfio_group_put_external_user(kvg->vfio_group);
374 list_del(&kvg->node);
375 kfree(kvg);
376 kvm_arch_end_assignment(dev->kvm);
377 }
378
379 kvm_vfio_update_coherency(dev);
380
381 kfree(kv);
382 kfree(dev); /* alloc by kvm_ioctl_create_device, free by .destroy */
383}
384
385static int kvm_vfio_create(struct kvm_device *dev, u32 type);
386
387static struct kvm_device_ops kvm_vfio_ops = {
388 .name = "kvm-vfio",
389 .create = kvm_vfio_create,
390 .destroy = kvm_vfio_destroy,
391 .set_attr = kvm_vfio_set_attr,
392 .has_attr = kvm_vfio_has_attr,
393};
394
395static int kvm_vfio_create(struct kvm_device *dev, u32 type)
396{
397 struct kvm_device *tmp;
398 struct kvm_vfio *kv;
399
400 /* Only one VFIO "device" per VM */
401 list_for_each_entry(tmp, &dev->kvm->devices, vm_node)
402 if (tmp->ops == &kvm_vfio_ops)
403 return -EBUSY;
404
405 kv = kzalloc(sizeof(*kv), GFP_KERNEL_ACCOUNT);
406 if (!kv)
407 return -ENOMEM;
408
409 INIT_LIST_HEAD(&kv->group_list);
410 mutex_init(&kv->lock);
411
412 dev->private = kv;
413
414 return 0;
415}
416
417int kvm_vfio_ops_init(void)
418{
419 return kvm_register_device_ops(&kvm_vfio_ops, KVM_DEV_TYPE_VFIO);
420}
421
422void kvm_vfio_ops_exit(void)
423{
424 kvm_unregister_device_ops(KVM_DEV_TYPE_VFIO);
425}