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 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 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 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 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 } 99 100 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 */ 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 mutex_lock(&kv->lock); 127 128 list_for_each_entry(kvf, &kv->file_list, node) { 129 if (!kvm_vfio_file_enforced_coherent(kvf->file)) { 130 noncoherent = true; 131 break; 132 } 133 } 134 135 if (noncoherent != kv->noncoherent) { 136 kv->noncoherent = noncoherent; 137 138 if (kv->noncoherent) 139 kvm_arch_register_noncoherent_dma(dev->kvm); 140 else 141 kvm_arch_unregister_noncoherent_dma(dev->kvm); 142 } 143 144 mutex_unlock(&kv->lock); 145 } 146 147 static int kvm_vfio_file_add(struct kvm_device *dev, unsigned int fd) 148 { 149 struct kvm_vfio *kv = dev->private; 150 struct kvm_vfio_file *kvf; 151 struct file *filp; 152 int ret; 153 154 filp = fget(fd); 155 if (!filp) 156 return -EBADF; 157 158 /* Ensure the FD is a vfio FD. */ 159 if (!kvm_vfio_file_is_valid(filp)) { 160 ret = -EINVAL; 161 goto err_fput; 162 } 163 164 mutex_lock(&kv->lock); 165 166 list_for_each_entry(kvf, &kv->file_list, node) { 167 if (kvf->file == filp) { 168 ret = -EEXIST; 169 goto err_unlock; 170 } 171 } 172 173 kvf = kzalloc(sizeof(*kvf), GFP_KERNEL_ACCOUNT); 174 if (!kvf) { 175 ret = -ENOMEM; 176 goto err_unlock; 177 } 178 179 kvf->file = filp; 180 list_add_tail(&kvf->node, &kv->file_list); 181 182 kvm_arch_start_assignment(dev->kvm); 183 184 mutex_unlock(&kv->lock); 185 186 kvm_vfio_file_set_kvm(kvf->file, dev->kvm); 187 kvm_vfio_update_coherency(dev); 188 189 return 0; 190 err_unlock: 191 mutex_unlock(&kv->lock); 192 err_fput: 193 fput(filp); 194 return ret; 195 } 196 197 static int kvm_vfio_file_del(struct kvm_device *dev, unsigned int fd) 198 { 199 struct kvm_vfio *kv = dev->private; 200 struct kvm_vfio_file *kvf; 201 struct fd f; 202 int ret; 203 204 f = fdget(fd); 205 if (!f.file) 206 return -EBADF; 207 208 ret = -ENOENT; 209 210 mutex_lock(&kv->lock); 211 212 list_for_each_entry(kvf, &kv->file_list, node) { 213 if (kvf->file != f.file) 214 continue; 215 216 list_del(&kvf->node); 217 kvm_arch_end_assignment(dev->kvm); 218 #ifdef CONFIG_SPAPR_TCE_IOMMU 219 kvm_spapr_tce_release_vfio_group(dev->kvm, kvf); 220 #endif 221 kvm_vfio_file_set_kvm(kvf->file, NULL); 222 fput(kvf->file); 223 kfree(kvf); 224 ret = 0; 225 break; 226 } 227 228 mutex_unlock(&kv->lock); 229 230 fdput(f); 231 232 kvm_vfio_update_coherency(dev); 233 234 return ret; 235 } 236 237 #ifdef CONFIG_SPAPR_TCE_IOMMU 238 static int kvm_vfio_file_set_spapr_tce(struct kvm_device *dev, 239 void __user *arg) 240 { 241 struct kvm_vfio_spapr_tce param; 242 struct kvm_vfio *kv = dev->private; 243 struct kvm_vfio_file *kvf; 244 struct fd f; 245 int ret; 246 247 if (copy_from_user(¶m, arg, sizeof(struct kvm_vfio_spapr_tce))) 248 return -EFAULT; 249 250 f = fdget(param.groupfd); 251 if (!f.file) 252 return -EBADF; 253 254 ret = -ENOENT; 255 256 mutex_lock(&kv->lock); 257 258 list_for_each_entry(kvf, &kv->file_list, node) { 259 if (kvf->file != f.file) 260 continue; 261 262 if (!kvf->iommu_group) { 263 kvf->iommu_group = kvm_vfio_file_iommu_group(kvf->file); 264 if (WARN_ON_ONCE(!kvf->iommu_group)) { 265 ret = -EIO; 266 goto err_fdput; 267 } 268 } 269 270 ret = kvm_spapr_tce_attach_iommu_group(dev->kvm, param.tablefd, 271 kvf->iommu_group); 272 break; 273 } 274 275 err_fdput: 276 mutex_unlock(&kv->lock); 277 fdput(f); 278 return ret; 279 } 280 #endif 281 282 static int kvm_vfio_set_file(struct kvm_device *dev, long attr, 283 void __user *arg) 284 { 285 int32_t __user *argp = arg; 286 int32_t fd; 287 288 switch (attr) { 289 case KVM_DEV_VFIO_FILE_ADD: 290 if (get_user(fd, argp)) 291 return -EFAULT; 292 return kvm_vfio_file_add(dev, fd); 293 294 case KVM_DEV_VFIO_FILE_DEL: 295 if (get_user(fd, argp)) 296 return -EFAULT; 297 return kvm_vfio_file_del(dev, fd); 298 299 #ifdef CONFIG_SPAPR_TCE_IOMMU 300 case KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE: 301 return kvm_vfio_file_set_spapr_tce(dev, arg); 302 #endif 303 } 304 305 return -ENXIO; 306 } 307 308 static int kvm_vfio_set_attr(struct kvm_device *dev, 309 struct kvm_device_attr *attr) 310 { 311 switch (attr->group) { 312 case KVM_DEV_VFIO_FILE: 313 return kvm_vfio_set_file(dev, attr->attr, 314 u64_to_user_ptr(attr->addr)); 315 } 316 317 return -ENXIO; 318 } 319 320 static int kvm_vfio_has_attr(struct kvm_device *dev, 321 struct kvm_device_attr *attr) 322 { 323 switch (attr->group) { 324 case KVM_DEV_VFIO_FILE: 325 switch (attr->attr) { 326 case KVM_DEV_VFIO_FILE_ADD: 327 case KVM_DEV_VFIO_FILE_DEL: 328 #ifdef CONFIG_SPAPR_TCE_IOMMU 329 case KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE: 330 #endif 331 return 0; 332 } 333 334 break; 335 } 336 337 return -ENXIO; 338 } 339 340 static void kvm_vfio_release(struct kvm_device *dev) 341 { 342 struct kvm_vfio *kv = dev->private; 343 struct kvm_vfio_file *kvf, *tmp; 344 345 list_for_each_entry_safe(kvf, tmp, &kv->file_list, node) { 346 #ifdef CONFIG_SPAPR_TCE_IOMMU 347 kvm_spapr_tce_release_vfio_group(dev->kvm, kvf); 348 #endif 349 kvm_vfio_file_set_kvm(kvf->file, NULL); 350 fput(kvf->file); 351 list_del(&kvf->node); 352 kfree(kvf); 353 kvm_arch_end_assignment(dev->kvm); 354 } 355 356 kvm_vfio_update_coherency(dev); 357 358 kfree(kv); 359 kfree(dev); /* alloc by kvm_ioctl_create_device, free by .release */ 360 } 361 362 static int kvm_vfio_create(struct kvm_device *dev, u32 type); 363 364 static struct kvm_device_ops kvm_vfio_ops = { 365 .name = "kvm-vfio", 366 .create = kvm_vfio_create, 367 .release = kvm_vfio_release, 368 .set_attr = kvm_vfio_set_attr, 369 .has_attr = kvm_vfio_has_attr, 370 }; 371 372 static int kvm_vfio_create(struct kvm_device *dev, u32 type) 373 { 374 struct kvm_device *tmp; 375 struct kvm_vfio *kv; 376 377 /* Only one VFIO "device" per VM */ 378 list_for_each_entry(tmp, &dev->kvm->devices, vm_node) 379 if (tmp->ops == &kvm_vfio_ops) 380 return -EBUSY; 381 382 kv = kzalloc(sizeof(*kv), GFP_KERNEL_ACCOUNT); 383 if (!kv) 384 return -ENOMEM; 385 386 INIT_LIST_HEAD(&kv->file_list); 387 mutex_init(&kv->lock); 388 389 dev->private = kv; 390 391 return 0; 392 } 393 394 int kvm_vfio_ops_init(void) 395 { 396 return kvm_register_device_ops(&kvm_vfio_ops, KVM_DEV_TYPE_VFIO); 397 } 398 399 void kvm_vfio_ops_exit(void) 400 { 401 kvm_unregister_device_ops(KVM_DEV_TYPE_VFIO); 402 } 403