projects/linux·blob master · refs

359 lines · 7.3 KiB

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 <[email protected]>
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 __free(fput) = NULL;
148
149 filp = fget(fd);
150 if (!filp)
151 return -EBADF;
152
153 /* Ensure the FD is a vfio FD. */
154 if (!kvm_vfio_file_is_valid(filp))
155 return -EINVAL;
156
157 guard(mutex)(&kv->lock);
158
159 list_for_each_entry(kvf, &kv->file_list, node) {
160 if (kvf->file == filp)
161 return -EEXIST;
162 }
163
164 kvf = kzalloc_obj(*kvf, GFP_KERNEL_ACCOUNT);
165 if (!kvf)
166 return -ENOMEM;
167
168 kvf->file = get_file(filp);
169 list_add_tail(&kvf->node, &kv->file_list);
170
171 kvm_vfio_file_set_kvm(kvf->file, dev->kvm);
172 kvm_vfio_update_coherency(dev);
173
174 return 0;
175}
176
177static void kvm_vfio_file_free(struct kvm_device *dev, struct kvm_vfio_file *kvf)
178{
179#ifdef CONFIG_SPAPR_TCE_IOMMU
180 kvm_spapr_tce_release_vfio_group(dev->kvm, kvf);
181#endif
182 kvm_vfio_file_set_kvm(kvf->file, NULL);
183 fput(kvf->file);
184 list_del(&kvf->node);
185 kfree(kvf);
186}
187
188static 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
194 if (fd_empty(f))
195 return -EBADF;
196
197 guard(mutex)(&kv->lock);
198
199 list_for_each_entry(kvf, &kv->file_list, node) {
200 if (kvf->file == fd_file(f)) {
201 kvm_vfio_file_free(dev, kvf);
202 kvm_vfio_update_coherency(dev);
203 return 0;
204 }
205 }
206
207 return -ENOENT;
208}
209
210#ifdef CONFIG_SPAPR_TCE_IOMMU
211static int kvm_vfio_file_set_spapr_tce(struct kvm_device *dev,
212 void __user *arg)
213{
214 struct kvm_vfio_spapr_tce param;
215 struct kvm_vfio *kv = dev->private;
216 struct kvm_vfio_file *kvf;
217
218 if (copy_from_user(&param, arg, sizeof(struct kvm_vfio_spapr_tce)))
219 return -EFAULT;
220
221 CLASS(fd, f)(param.groupfd);
222 if (fd_empty(f))
223 return -EBADF;
224
225 guard(mutex)(&kv->lock);
226
227 list_for_each_entry(kvf, &kv->file_list, node) {
228 if (kvf->file != fd_file(f))
229 continue;
230
231 if (!kvf->iommu_group) {
232 kvf->iommu_group = kvm_vfio_file_iommu_group(kvf->file);
233 if (WARN_ON_ONCE(!kvf->iommu_group))
234 return -EIO;
235 }
236
237 return kvm_spapr_tce_attach_iommu_group(dev->kvm, param.tablefd,
238 kvf->iommu_group);
239 }
240
241 return -ENOENT;
242}
243#endif
244
245static int kvm_vfio_set_file(struct kvm_device *dev, long attr,
246 void __user *arg)
247{
248 int32_t __user *argp = arg;
249 int32_t fd;
250
251 switch (attr) {
252 case KVM_DEV_VFIO_FILE_ADD:
253 if (get_user(fd, argp))
254 return -EFAULT;
255 return kvm_vfio_file_add(dev, fd);
256
257 case KVM_DEV_VFIO_FILE_DEL:
258 if (get_user(fd, argp))
259 return -EFAULT;
260 return kvm_vfio_file_del(dev, fd);
261
262#ifdef CONFIG_SPAPR_TCE_IOMMU
263 case KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE:
264 return kvm_vfio_file_set_spapr_tce(dev, arg);
265#endif
266 }
267
268 return -ENXIO;
269}
270
271static int kvm_vfio_set_attr(struct kvm_device *dev,
272 struct kvm_device_attr *attr)
273{
274 switch (attr->group) {
275 case KVM_DEV_VFIO_FILE:
276 return kvm_vfio_set_file(dev, attr->attr,
277 u64_to_user_ptr(attr->addr));
278 }
279
280 return -ENXIO;
281}
282
283static int kvm_vfio_has_attr(struct kvm_device *dev,
284 struct kvm_device_attr *attr)
285{
286 switch (attr->group) {
287 case KVM_DEV_VFIO_FILE:
288 switch (attr->attr) {
289 case KVM_DEV_VFIO_FILE_ADD:
290 case KVM_DEV_VFIO_FILE_DEL:
291#ifdef CONFIG_SPAPR_TCE_IOMMU
292 case KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE:
293#endif
294 return 0;
295 }
296
297 break;
298 }
299
300 return -ENXIO;
301}
302
303static void kvm_vfio_release(struct kvm_device *dev)
304{
305 struct kvm_vfio *kv = dev->private;
306 struct kvm_vfio_file *kvf, *tmp;
307
308 list_for_each_entry_safe(kvf, tmp, &kv->file_list, node)
309 kvm_vfio_file_free(dev, kvf);
310
311 kvm_vfio_update_coherency(dev);
312
313 kfree(kv);
314 kfree(dev); /* alloc by kvm_ioctl_create_device, free by .release */
315}
316
317static int kvm_vfio_create(struct kvm_device *dev, u32 type);
318
319static const struct kvm_device_ops kvm_vfio_ops = {
320 .name = "kvm-vfio",
321 .create = kvm_vfio_create,
322 .release = kvm_vfio_release,
323 .set_attr = kvm_vfio_set_attr,
324 .has_attr = kvm_vfio_has_attr,
325};
326
327static int kvm_vfio_create(struct kvm_device *dev, u32 type)
328{
329 struct kvm_device *tmp;
330 struct kvm_vfio *kv;
331
332 lockdep_assert_held(&dev->kvm->lock);
333
334 /* Only one VFIO "device" per VM */
335 list_for_each_entry(tmp, &dev->kvm->devices, vm_node)
336 if (tmp->ops == &kvm_vfio_ops)
337 return -EBUSY;
338
339 kv = kzalloc_obj(*kv, GFP_KERNEL_ACCOUNT);
340 if (!kv)
341 return -ENOMEM;
342
343 INIT_LIST_HEAD(&kv->file_list);
344 mutex_init(&kv->lock);
345
346 dev->private = kv;
347
348 return 0;
349}
350
351int kvm_vfio_ops_init(void)
352{
353 return kvm_register_device_ops(&kvm_vfio_ops, KVM_DEV_TYPE_VFIO);
354}
355
356void kvm_vfio_ops_exit(void)
357{
358 kvm_unregister_device_ops(KVM_DEV_TYPE_VFIO);
359}