virt/kvm/vfio.c

   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
  23 #ifdef CONFIG_SPAPR_TCE_IOMMU
  24 #include <asm/kvm_ppc.h>
  25 #endif
  26
  27 struct kvm_vfio_group {
  28         struct list_head node;
  29         struct vfio_group *vfio_group;
  30 };
  31
  32 struct kvm_vfio {
  33         struct list_head group_list;
  34         struct mutex lock;
  35         bool noncoherent;
  36 };
  37
  38 static struct vfio_group *kvm_vfio_group_get_external_user(struct file *filep)
  39 {
  40         struct vfio_group *vfio_group;
  41         struct vfio_group *(*fn)(struct file *);
  42
  43         fn = symbol_get(vfio_group_get_external_user);
  44         if (!fn)
  45                 return ERR_PTR(-EINVAL);
  46
  47         vfio_group = fn(filep);
  48
  49         symbol_put(vfio_group_get_external_user);
  50
  51         return vfio_group;
  52 }
  53
  54 static bool kvm_vfio_external_group_match_file(struct vfio_group *group,
  55                                                struct file *filep)
  56 {
  57         bool ret, (*fn)(struct vfio_group *, struct file *);
  58
  59         fn = symbol_get(vfio_external_group_match_file);
  60         if (!fn)
  61                 return false;
  62
  63         ret = fn(group, filep);
  64
  65         symbol_put(vfio_external_group_match_file);
  66
  67         return ret;
  68 }
  69
  70 static void kvm_vfio_group_put_external_user(struct vfio_group *vfio_group)
  71 {
  72         void (*fn)(struct vfio_group *);
  73
  74         fn = symbol_get(vfio_group_put_external_user);
  75         if (!fn)
  76                 return;
  77
  78         fn(vfio_group);
  79
  80         symbol_put(vfio_group_put_external_user);
  81 }
  82
  83 static void kvm_vfio_group_set_kvm(struct vfio_group *group, struct kvm *kvm)
  84 {
  85         void (*fn)(struct vfio_group *, struct kvm *);
  86
  87         fn = symbol_get(vfio_group_set_kvm);
  88         if (!fn)
  89                 return;
  90
  91         fn(group, kvm);
  92
  93         symbol_put(vfio_group_set_kvm);
  94 }
  95
  96 static bool kvm_vfio_group_is_coherent(struct vfio_group *vfio_group)
  97 {
  98         long (*fn)(struct vfio_group *, unsigned long);
  99         long ret;
 100
 101         fn = symbol_get(vfio_external_check_extension);
 102         if (!fn)
 103                 return false;
 104
 105         ret = fn(vfio_group, VFIO_DMA_CC_IOMMU);
 106
 107         symbol_put(vfio_external_check_extension);
 108
 109         return ret > 0;
 110 }
 111
 112 #ifdef CONFIG_SPAPR_TCE_IOMMU
 113 static int kvm_vfio_external_user_iommu_id(struct vfio_group *vfio_group)
 114 {
 115         int (*fn)(struct vfio_group *);
 116         int ret = -EINVAL;
 117
 118         fn = symbol_get(vfio_external_user_iommu_id);
 119         if (!fn)
 120                 return ret;
 121
 122         ret = fn(vfio_group);
 123
 124         symbol_put(vfio_external_user_iommu_id);
 125
 126         return ret;
 127 }
 128
 129 static struct iommu_group *kvm_vfio_group_get_iommu_group(
 130                 struct vfio_group *group)
 131 {
 132         int group_id = kvm_vfio_external_user_iommu_id(group);
 133
 134         if (group_id < 0)
 135                 return NULL;
 136
 137         return iommu_group_get_by_id(group_id);
 138 }
 139
 140 static void kvm_spapr_tce_release_vfio_group(struct kvm *kvm,
 141                 struct vfio_group *vfio_group)
 142 {
 143         struct iommu_group *grp = kvm_vfio_group_get_iommu_group(vfio_group);
 144
 145         if (WARN_ON_ONCE(!grp))
 146                 return;
 147
 148         kvm_spapr_tce_release_iommu_group(kvm, grp);
 149         iommu_group_put(grp);
 150 }
 151 #endif
 152
 153 /*
 154  * Groups can use the same or different IOMMU domains.  If the same then
 155  * adding a new group may change the coherency of groups we've previously
 156  * been told about.  We don't want to care about any of that so we retest
 157  * each group and bail as soon as we find one that's noncoherent.  This
 158  * means we only ever [un]register_noncoherent_dma once for the whole device.
 159  */
 160 static void kvm_vfio_update_coherency(struct kvm_device *dev)
 161 {
 162         struct kvm_vfio *kv = dev->private;
 163         bool noncoherent = false;
 164         struct kvm_vfio_group *kvg;
 165
 166         mutex_lock(&kv->lock);
 167
 168         list_for_each_entry(kvg, &kv->group_list, node) {
 169                 if (!kvm_vfio_group_is_coherent(kvg->vfio_group)) {
 170                         noncoherent = true;
 171                         break;
 172                 }
 173         }
 174
 175         if (noncoherent != kv->noncoherent) {
 176                 kv->noncoherent = noncoherent;
 177
 178                 if (kv->noncoherent)
 179                         kvm_arch_register_noncoherent_dma(dev->kvm);
 180                 else
 181                         kvm_arch_unregister_noncoherent_dma(dev->kvm);
 182         }
 183
 184         mutex_unlock(&kv->lock);
 185 }
 186
 187 static int kvm_vfio_set_group(struct kvm_device *dev, long attr, u64 arg)
 188 {
 189         struct kvm_vfio *kv = dev->private;
 190         struct vfio_group *vfio_group;
 191         struct kvm_vfio_group *kvg;
 192         int32_t __user *argp = (int32_t __user *)(unsigned long)arg;
 193         struct fd f;
 194         int32_t fd;
 195         int ret;
 196
 197         switch (attr) {
 198         case KVM_DEV_VFIO_GROUP_ADD:
 199                 if (get_user(fd, argp))
 200                         return -EFAULT;
 201
 202                 f = fdget(fd);
 203                 if (!f.file)
 204                         return -EBADF;
 205
 206                 vfio_group = kvm_vfio_group_get_external_user(f.file);
 207                 fdput(f);
 208
 209                 if (IS_ERR(vfio_group))
 210                         return PTR_ERR(vfio_group);
 211
 212                 mutex_lock(&kv->lock);
 213
 214                 list_for_each_entry(kvg, &kv->group_list, node) {
 215                         if (kvg->vfio_group == vfio_group) {
 216                                 mutex_unlock(&kv->lock);
 217                                 kvm_vfio_group_put_external_user(vfio_group);
 218                                 return -EEXIST;
 219                         }
 220                 }
 221
 222                 kvg = kzalloc(sizeof(*kvg), GFP_KERNEL);
 223                 if (!kvg) {
 224                         mutex_unlock(&kv->lock);
 225                         kvm_vfio_group_put_external_user(vfio_group);
 226                         return -ENOMEM;
 227                 }
 228
 229                 list_add_tail(&kvg->node, &kv->group_list);
 230                 kvg->vfio_group = vfio_group;
 231
 232                 kvm_arch_start_assignment(dev->kvm);
 233
 234                 mutex_unlock(&kv->lock);
 235
 236                 kvm_vfio_group_set_kvm(vfio_group, dev->kvm);
 237
 238                 kvm_vfio_update_coherency(dev);
 239
 240                 return 0;
 241
 242         case KVM_DEV_VFIO_GROUP_DEL:
 243                 if (get_user(fd, argp))
 244                         return -EFAULT;
 245
 246                 f = fdget(fd);
 247                 if (!f.file)
 248                         return -EBADF;
 249
 250                 ret = -ENOENT;
 251
 252                 mutex_lock(&kv->lock);
 253
 254                 list_for_each_entry(kvg, &kv->group_list, node) {
 255                         if (!kvm_vfio_external_group_match_file(kvg->vfio_group,
 256                                                                 f.file))
 257                                 continue;
 258
 259                         list_del(&kvg->node);
 260                         kvm_arch_end_assignment(dev->kvm);
 261 #ifdef CONFIG_SPAPR_TCE_IOMMU
 262                         kvm_spapr_tce_release_vfio_group(dev->kvm,
 263                                                          kvg->vfio_group);
 264 #endif
 265                         kvm_vfio_group_set_kvm(kvg->vfio_group, NULL);
 266                         kvm_vfio_group_put_external_user(kvg->vfio_group);
 267                         kfree(kvg);
 268                         ret = 0;
 269                         break;
 270                 }
 271
 272                 mutex_unlock(&kv->lock);
 273
 274                 fdput(f);
 275
 276                 kvm_vfio_update_coherency(dev);
 277
 278                 return ret;
 279
 280 #ifdef CONFIG_SPAPR_TCE_IOMMU
 281         case KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE: {
 282                 struct kvm_vfio_spapr_tce param;
 283                 struct kvm_vfio *kv = dev->private;
 284                 struct vfio_group *vfio_group;
 285                 struct kvm_vfio_group *kvg;
 286                 struct fd f;
 287                 struct iommu_group *grp;
 288
 289                 if (copy_from_user(&param, (void __user *)arg,
 290                                 sizeof(struct kvm_vfio_spapr_tce)))
 291                         return -EFAULT;
 292
 293                 f = fdget(param.groupfd);
 294                 if (!f.file)
 295                         return -EBADF;
 296
 297                 vfio_group = kvm_vfio_group_get_external_user(f.file);
 298                 fdput(f);
 299
 300                 if (IS_ERR(vfio_group))
 301                         return PTR_ERR(vfio_group);
 302
 303                 grp = kvm_vfio_group_get_iommu_group(vfio_group);
 304                 if (WARN_ON_ONCE(!grp)) {
 305                         kvm_vfio_group_put_external_user(vfio_group);
 306                         return -EIO;
 307                 }
 308
 309                 ret = -ENOENT;
 310
 311                 mutex_lock(&kv->lock);
 312
 313                 list_for_each_entry(kvg, &kv->group_list, node) {
 314                         if (kvg->vfio_group != vfio_group)
 315                                 continue;
 316
 317                         ret = kvm_spapr_tce_attach_iommu_group(dev->kvm,
 318                                         param.tablefd, grp);
 319                         break;
 320                 }
 321
 322                 mutex_unlock(&kv->lock);
 323
 324                 iommu_group_put(grp);
 325                 kvm_vfio_group_put_external_user(vfio_group);
 326
 327                 return ret;
 328         }
 329 #endif /* CONFIG_SPAPR_TCE_IOMMU */
 330         }
 331
 332         return -ENXIO;
 333 }
 334
 335 static int kvm_vfio_set_attr(struct kvm_device *dev,
 336                              struct kvm_device_attr *attr)
 337 {
 338         switch (attr->group) {
 339         case KVM_DEV_VFIO_GROUP:
 340                 return kvm_vfio_set_group(dev, attr->attr, attr->addr);
 341         }
 342
 343         return -ENXIO;
 344 }
 345
 346 static int kvm_vfio_has_attr(struct kvm_device *dev,
 347                              struct kvm_device_attr *attr)
 348 {
 349         switch (attr->group) {
 350         case KVM_DEV_VFIO_GROUP:
 351                 switch (attr->attr) {
 352                 case KVM_DEV_VFIO_GROUP_ADD:
 353                 case KVM_DEV_VFIO_GROUP_DEL:
 354 #ifdef CONFIG_SPAPR_TCE_IOMMU
 355                 case KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE:
 356 #endif
 357                         return 0;
 358                 }
 359
 360                 break;
 361         }
 362
 363         return -ENXIO;
 364 }
 365
 366 static void kvm_vfio_destroy(struct kvm_device *dev)
 367 {
 368         struct kvm_vfio *kv = dev->private;
 369         struct kvm_vfio_group *kvg, *tmp;
 370
 371         list_for_each_entry_safe(kvg, tmp, &kv->group_list, node) {
 372 #ifdef CONFIG_SPAPR_TCE_IOMMU
 373                 kvm_spapr_tce_release_vfio_group(dev->kvm, kvg->vfio_group);
 374 #endif
 375                 kvm_vfio_group_set_kvm(kvg->vfio_group, NULL);
 376                 kvm_vfio_group_put_external_user(kvg->vfio_group);
 377                 list_del(&kvg->node);
 378                 kfree(kvg);
 379                 kvm_arch_end_assignment(dev->kvm);
 380         }
 381
 382         kvm_vfio_update_coherency(dev);
 383
 384         kfree(kv);
 385         kfree(dev); /* alloc by kvm_ioctl_create_device, free by .destroy */
 386 }
 387
 388 static int kvm_vfio_create(struct kvm_device *dev, u32 type);
 389
 390 static struct kvm_device_ops kvm_vfio_ops = {
 391         .name = "kvm-vfio",
 392         .create = kvm_vfio_create,
 393         .destroy = kvm_vfio_destroy,
 394         .set_attr = kvm_vfio_set_attr,
 395         .has_attr = kvm_vfio_has_attr,
 396 };
 397
 398 static int kvm_vfio_create(struct kvm_device *dev, u32 type)
 399 {
 400         struct kvm_device *tmp;
 401         struct kvm_vfio *kv;
 402
 403         /* Only one VFIO "device" per VM */
 404         list_for_each_entry(tmp, &dev->kvm->devices, vm_node)
 405                 if (tmp->ops == &kvm_vfio_ops)
 406                         return -EBUSY;
 407
 408         kv = kzalloc(sizeof(*kv), GFP_KERNEL);
 409         if (!kv)
 410                 return -ENOMEM;
 411
 412         INIT_LIST_HEAD(&kv->group_list);
 413         mutex_init(&kv->lock);
 414
 415         dev->private = kv;
 416
 417         return 0;
 418 }
 419
 420 int kvm_vfio_ops_init(void)
 421 {
 422         return kvm_register_device_ops(&kvm_vfio_ops, KVM_DEV_TYPE_VFIO);
 423 }
 424
 425 void kvm_vfio_ops_exit(void)
 426 {
 427         kvm_unregister_device_ops(KVM_DEV_TYPE_VFIO);
 428 }