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
24 struct 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
32 struct kvm_vfio {
33 struct list_head file_list;
34 struct mutex lock;
35 bool noncoherent;
36 };
37
kvm_vfio_file_set_kvm(struct file * file,struct kvm * kvm)38 static 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
kvm_vfio_file_enforced_coherent(struct file * file)51 static 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
kvm_vfio_file_is_valid(struct file * file)67 static 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
kvm_vfio_file_iommu_group(struct file * file)84 static 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
kvm_spapr_tce_release_vfio_group(struct kvm * kvm,struct kvm_vfio_file * kvf)100 static 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 */
kvm_vfio_update_coherency(struct kvm_device * dev)120 static 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
kvm_vfio_file_add(struct kvm_device * dev,unsigned int fd)143 static 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_vfio_file_set_kvm(kvf->file, dev->kvm);
179 kvm_vfio_update_coherency(dev);
180
181 out_unlock:
182 mutex_unlock(&kv->lock);
183 out_fput:
184 fput(filp);
185 return ret;
186 }
187
kvm_vfio_file_del(struct kvm_device * dev,unsigned int fd)188 static int kvm_vfio_file_del(struct kvm_device *dev, unsigned int fd)
189 {
190 struct kvm_vfio *kv = dev->private;
191 struct kvm_vfio_file *kvf;
192 CLASS(fd, f)(fd);
193 int ret;
194
195 if (fd_empty(f))
196 return -EBADF;
197
198 ret = -ENOENT;
199
200 mutex_lock(&kv->lock);
201
202 list_for_each_entry(kvf, &kv->file_list, node) {
203 if (kvf->file != fd_file(f))
204 continue;
205
206 list_del(&kvf->node);
207 #ifdef CONFIG_SPAPR_TCE_IOMMU
208 kvm_spapr_tce_release_vfio_group(dev->kvm, kvf);
209 #endif
210 kvm_vfio_file_set_kvm(kvf->file, NULL);
211 fput(kvf->file);
212 kfree(kvf);
213 ret = 0;
214 break;
215 }
216
217 kvm_vfio_update_coherency(dev);
218
219 mutex_unlock(&kv->lock);
220 return ret;
221 }
222
223 #ifdef CONFIG_SPAPR_TCE_IOMMU
kvm_vfio_file_set_spapr_tce(struct kvm_device * dev,void __user * arg)224 static int kvm_vfio_file_set_spapr_tce(struct kvm_device *dev,
225 void __user *arg)
226 {
227 struct kvm_vfio_spapr_tce param;
228 struct kvm_vfio *kv = dev->private;
229 struct kvm_vfio_file *kvf;
230 int ret;
231
232 if (copy_from_user(¶m, arg, sizeof(struct kvm_vfio_spapr_tce)))
233 return -EFAULT;
234
235 CLASS(fd, f)(param.groupfd);
236 if (fd_empty(f))
237 return -EBADF;
238
239 ret = -ENOENT;
240
241 mutex_lock(&kv->lock);
242
243 list_for_each_entry(kvf, &kv->file_list, node) {
244 if (kvf->file != fd_file(f))
245 continue;
246
247 if (!kvf->iommu_group) {
248 kvf->iommu_group = kvm_vfio_file_iommu_group(kvf->file);
249 if (WARN_ON_ONCE(!kvf->iommu_group)) {
250 ret = -EIO;
251 goto err_fdput;
252 }
253 }
254
255 ret = kvm_spapr_tce_attach_iommu_group(dev->kvm, param.tablefd,
256 kvf->iommu_group);
257 break;
258 }
259
260 err_fdput:
261 mutex_unlock(&kv->lock);
262 return ret;
263 }
264 #endif
265
kvm_vfio_set_file(struct kvm_device * dev,long attr,void __user * arg)266 static int kvm_vfio_set_file(struct kvm_device *dev, long attr,
267 void __user *arg)
268 {
269 int32_t __user *argp = arg;
270 int32_t fd;
271
272 switch (attr) {
273 case KVM_DEV_VFIO_FILE_ADD:
274 if (get_user(fd, argp))
275 return -EFAULT;
276 return kvm_vfio_file_add(dev, fd);
277
278 case KVM_DEV_VFIO_FILE_DEL:
279 if (get_user(fd, argp))
280 return -EFAULT;
281 return kvm_vfio_file_del(dev, fd);
282
283 #ifdef CONFIG_SPAPR_TCE_IOMMU
284 case KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE:
285 return kvm_vfio_file_set_spapr_tce(dev, arg);
286 #endif
287 }
288
289 return -ENXIO;
290 }
291
kvm_vfio_set_attr(struct kvm_device * dev,struct kvm_device_attr * attr)292 static int kvm_vfio_set_attr(struct kvm_device *dev,
293 struct kvm_device_attr *attr)
294 {
295 switch (attr->group) {
296 case KVM_DEV_VFIO_FILE:
297 return kvm_vfio_set_file(dev, attr->attr,
298 u64_to_user_ptr(attr->addr));
299 }
300
301 return -ENXIO;
302 }
303
kvm_vfio_has_attr(struct kvm_device * dev,struct kvm_device_attr * attr)304 static int kvm_vfio_has_attr(struct kvm_device *dev,
305 struct kvm_device_attr *attr)
306 {
307 switch (attr->group) {
308 case KVM_DEV_VFIO_FILE:
309 switch (attr->attr) {
310 case KVM_DEV_VFIO_FILE_ADD:
311 case KVM_DEV_VFIO_FILE_DEL:
312 #ifdef CONFIG_SPAPR_TCE_IOMMU
313 case KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE:
314 #endif
315 return 0;
316 }
317
318 break;
319 }
320
321 return -ENXIO;
322 }
323
kvm_vfio_release(struct kvm_device * dev)324 static void kvm_vfio_release(struct kvm_device *dev)
325 {
326 struct kvm_vfio *kv = dev->private;
327 struct kvm_vfio_file *kvf, *tmp;
328
329 list_for_each_entry_safe(kvf, tmp, &kv->file_list, node) {
330 #ifdef CONFIG_SPAPR_TCE_IOMMU
331 kvm_spapr_tce_release_vfio_group(dev->kvm, kvf);
332 #endif
333 kvm_vfio_file_set_kvm(kvf->file, NULL);
334 fput(kvf->file);
335 list_del(&kvf->node);
336 kfree(kvf);
337 }
338
339 kvm_vfio_update_coherency(dev);
340
341 kfree(kv);
342 kfree(dev); /* alloc by kvm_ioctl_create_device, free by .release */
343 }
344
345 static int kvm_vfio_create(struct kvm_device *dev, u32 type);
346
347 static const struct kvm_device_ops kvm_vfio_ops = {
348 .name = "kvm-vfio",
349 .create = kvm_vfio_create,
350 .release = kvm_vfio_release,
351 .set_attr = kvm_vfio_set_attr,
352 .has_attr = kvm_vfio_has_attr,
353 };
354
kvm_vfio_create(struct kvm_device * dev,u32 type)355 static int kvm_vfio_create(struct kvm_device *dev, u32 type)
356 {
357 struct kvm_device *tmp;
358 struct kvm_vfio *kv;
359
360 lockdep_assert_held(&dev->kvm->lock);
361
362 /* Only one VFIO "device" per VM */
363 list_for_each_entry(tmp, &dev->kvm->devices, vm_node)
364 if (tmp->ops == &kvm_vfio_ops)
365 return -EBUSY;
366
367 kv = kzalloc(sizeof(*kv), GFP_KERNEL_ACCOUNT);
368 if (!kv)
369 return -ENOMEM;
370
371 INIT_LIST_HEAD(&kv->file_list);
372 mutex_init(&kv->lock);
373
374 dev->private = kv;
375
376 return 0;
377 }
378
kvm_vfio_ops_init(void)379 int kvm_vfio_ops_init(void)
380 {
381 return kvm_register_device_ops(&kvm_vfio_ops, KVM_DEV_TYPE_VFIO);
382 }
383
kvm_vfio_ops_exit(void)384 void kvm_vfio_ops_exit(void)
385 {
386 kvm_unregister_device_ops(KVM_DEV_TYPE_VFIO);
387 }
388