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