vfio.c 8.8 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425
  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. #include <linux/errno.h>
  9. #include <linux/file.h>
  10. #include <linux/kvm_host.h>
  11. #include <linux/list.h>
  12. #include <linux/module.h>
  13. #include <linux/mutex.h>
  14. #include <linux/slab.h>
  15. #include <linux/uaccess.h>
  16. #include <linux/vfio.h>
  17. #include "vfio.h"
  18. #ifdef CONFIG_SPAPR_TCE_IOMMU
  19. #include <asm/kvm_ppc.h>
  20. #endif
  21. struct kvm_vfio_group {
  22. struct list_head node;
  23. struct vfio_group *vfio_group;
  24. };
  25. struct kvm_vfio {
  26. struct list_head group_list;
  27. struct mutex lock;
  28. bool noncoherent;
  29. };
  30. static struct vfio_group *kvm_vfio_group_get_external_user(struct file *filep)
  31. {
  32. struct vfio_group *vfio_group;
  33. struct vfio_group *(*fn)(struct file *);
  34. fn = symbol_get(vfio_group_get_external_user);
  35. if (!fn)
  36. return ERR_PTR(-EINVAL);
  37. vfio_group = fn(filep);
  38. symbol_put(vfio_group_get_external_user);
  39. return vfio_group;
  40. }
  41. static bool kvm_vfio_external_group_match_file(struct vfio_group *group,
  42. struct file *filep)
  43. {
  44. bool ret, (*fn)(struct vfio_group *, struct file *);
  45. fn = symbol_get(vfio_external_group_match_file);
  46. if (!fn)
  47. return false;
  48. ret = fn(group, filep);
  49. symbol_put(vfio_external_group_match_file);
  50. return ret;
  51. }
  52. static void kvm_vfio_group_put_external_user(struct vfio_group *vfio_group)
  53. {
  54. void (*fn)(struct vfio_group *);
  55. fn = symbol_get(vfio_group_put_external_user);
  56. if (!fn)
  57. return;
  58. fn(vfio_group);
  59. symbol_put(vfio_group_put_external_user);
  60. }
  61. static void kvm_vfio_group_set_kvm(struct vfio_group *group, struct kvm *kvm)
  62. {
  63. void (*fn)(struct vfio_group *, struct kvm *);
  64. fn = symbol_get(vfio_group_set_kvm);
  65. if (!fn)
  66. return;
  67. fn(group, kvm);
  68. symbol_put(vfio_group_set_kvm);
  69. }
  70. static bool kvm_vfio_group_is_coherent(struct vfio_group *vfio_group)
  71. {
  72. long (*fn)(struct vfio_group *, unsigned long);
  73. long ret;
  74. fn = symbol_get(vfio_external_check_extension);
  75. if (!fn)
  76. return false;
  77. ret = fn(vfio_group, VFIO_DMA_CC_IOMMU);
  78. symbol_put(vfio_external_check_extension);
  79. return ret > 0;
  80. }
  81. #ifdef CONFIG_SPAPR_TCE_IOMMU
  82. static int kvm_vfio_external_user_iommu_id(struct vfio_group *vfio_group)
  83. {
  84. int (*fn)(struct vfio_group *);
  85. int ret = -EINVAL;
  86. fn = symbol_get(vfio_external_user_iommu_id);
  87. if (!fn)
  88. return ret;
  89. ret = fn(vfio_group);
  90. symbol_put(vfio_external_user_iommu_id);
  91. return ret;
  92. }
  93. static struct iommu_group *kvm_vfio_group_get_iommu_group(
  94. struct vfio_group *group)
  95. {
  96. int group_id = kvm_vfio_external_user_iommu_id(group);
  97. if (group_id < 0)
  98. return NULL;
  99. return iommu_group_get_by_id(group_id);
  100. }
  101. static void kvm_spapr_tce_release_vfio_group(struct kvm *kvm,
  102. struct vfio_group *vfio_group)
  103. {
  104. struct iommu_group *grp = kvm_vfio_group_get_iommu_group(vfio_group);
  105. if (WARN_ON_ONCE(!grp))
  106. return;
  107. kvm_spapr_tce_release_iommu_group(kvm, grp);
  108. iommu_group_put(grp);
  109. }
  110. #endif
  111. /*
  112. * Groups can use the same or different IOMMU domains. If the same then
  113. * adding a new group may change the coherency of groups we've previously
  114. * been told about. We don't want to care about any of that so we retest
  115. * each group and bail as soon as we find one that's noncoherent. This
  116. * means we only ever [un]register_noncoherent_dma once for the whole device.
  117. */
  118. static void kvm_vfio_update_coherency(struct kvm_device *dev)
  119. {
  120. struct kvm_vfio *kv = dev->private;
  121. bool noncoherent = false;
  122. struct kvm_vfio_group *kvg;
  123. mutex_lock(&kv->lock);
  124. list_for_each_entry(kvg, &kv->group_list, node) {
  125. if (!kvm_vfio_group_is_coherent(kvg->vfio_group)) {
  126. noncoherent = true;
  127. break;
  128. }
  129. }
  130. if (noncoherent != kv->noncoherent) {
  131. kv->noncoherent = noncoherent;
  132. if (kv->noncoherent)
  133. kvm_arch_register_noncoherent_dma(dev->kvm);
  134. else
  135. kvm_arch_unregister_noncoherent_dma(dev->kvm);
  136. }
  137. mutex_unlock(&kv->lock);
  138. }
  139. static int kvm_vfio_set_group(struct kvm_device *dev, long attr, u64 arg)
  140. {
  141. struct kvm_vfio *kv = dev->private;
  142. struct vfio_group *vfio_group;
  143. struct kvm_vfio_group *kvg;
  144. int32_t __user *argp = (int32_t __user *)(unsigned long)arg;
  145. struct fd f;
  146. int32_t fd;
  147. int ret;
  148. switch (attr) {
  149. case KVM_DEV_VFIO_GROUP_ADD:
  150. if (get_user(fd, argp))
  151. return -EFAULT;
  152. f = fdget(fd);
  153. if (!f.file)
  154. return -EBADF;
  155. vfio_group = kvm_vfio_group_get_external_user(f.file);
  156. fdput(f);
  157. if (IS_ERR(vfio_group))
  158. return PTR_ERR(vfio_group);
  159. mutex_lock(&kv->lock);
  160. list_for_each_entry(kvg, &kv->group_list, node) {
  161. if (kvg->vfio_group == vfio_group) {
  162. mutex_unlock(&kv->lock);
  163. kvm_vfio_group_put_external_user(vfio_group);
  164. return -EEXIST;
  165. }
  166. }
  167. kvg = kzalloc(sizeof(*kvg), GFP_KERNEL_ACCOUNT);
  168. if (!kvg) {
  169. mutex_unlock(&kv->lock);
  170. kvm_vfio_group_put_external_user(vfio_group);
  171. return -ENOMEM;
  172. }
  173. list_add_tail(&kvg->node, &kv->group_list);
  174. kvg->vfio_group = vfio_group;
  175. kvm_arch_start_assignment(dev->kvm);
  176. mutex_unlock(&kv->lock);
  177. kvm_vfio_group_set_kvm(vfio_group, dev->kvm);
  178. kvm_vfio_update_coherency(dev);
  179. return 0;
  180. case KVM_DEV_VFIO_GROUP_DEL:
  181. if (get_user(fd, argp))
  182. return -EFAULT;
  183. f = fdget(fd);
  184. if (!f.file)
  185. return -EBADF;
  186. ret = -ENOENT;
  187. mutex_lock(&kv->lock);
  188. list_for_each_entry(kvg, &kv->group_list, node) {
  189. if (!kvm_vfio_external_group_match_file(kvg->vfio_group,
  190. f.file))
  191. continue;
  192. list_del(&kvg->node);
  193. kvm_arch_end_assignment(dev->kvm);
  194. #ifdef CONFIG_SPAPR_TCE_IOMMU
  195. kvm_spapr_tce_release_vfio_group(dev->kvm,
  196. kvg->vfio_group);
  197. #endif
  198. kvm_vfio_group_set_kvm(kvg->vfio_group, NULL);
  199. kvm_vfio_group_put_external_user(kvg->vfio_group);
  200. kfree(kvg);
  201. ret = 0;
  202. break;
  203. }
  204. mutex_unlock(&kv->lock);
  205. fdput(f);
  206. kvm_vfio_update_coherency(dev);
  207. return ret;
  208. #ifdef CONFIG_SPAPR_TCE_IOMMU
  209. case KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE: {
  210. struct kvm_vfio_spapr_tce param;
  211. struct kvm_vfio *kv = dev->private;
  212. struct vfio_group *vfio_group;
  213. struct kvm_vfio_group *kvg;
  214. struct fd f;
  215. struct iommu_group *grp;
  216. if (copy_from_user(&param, (void __user *)arg,
  217. sizeof(struct kvm_vfio_spapr_tce)))
  218. return -EFAULT;
  219. f = fdget(param.groupfd);
  220. if (!f.file)
  221. return -EBADF;
  222. vfio_group = kvm_vfio_group_get_external_user(f.file);
  223. fdput(f);
  224. if (IS_ERR(vfio_group))
  225. return PTR_ERR(vfio_group);
  226. grp = kvm_vfio_group_get_iommu_group(vfio_group);
  227. if (WARN_ON_ONCE(!grp)) {
  228. kvm_vfio_group_put_external_user(vfio_group);
  229. return -EIO;
  230. }
  231. ret = -ENOENT;
  232. mutex_lock(&kv->lock);
  233. list_for_each_entry(kvg, &kv->group_list, node) {
  234. if (kvg->vfio_group != vfio_group)
  235. continue;
  236. ret = kvm_spapr_tce_attach_iommu_group(dev->kvm,
  237. param.tablefd, grp);
  238. break;
  239. }
  240. mutex_unlock(&kv->lock);
  241. iommu_group_put(grp);
  242. kvm_vfio_group_put_external_user(vfio_group);
  243. return ret;
  244. }
  245. #endif /* CONFIG_SPAPR_TCE_IOMMU */
  246. }
  247. return -ENXIO;
  248. }
  249. static int kvm_vfio_set_attr(struct kvm_device *dev,
  250. struct kvm_device_attr *attr)
  251. {
  252. switch (attr->group) {
  253. case KVM_DEV_VFIO_GROUP:
  254. return kvm_vfio_set_group(dev, attr->attr, attr->addr);
  255. }
  256. return -ENXIO;
  257. }
  258. static int kvm_vfio_has_attr(struct kvm_device *dev,
  259. struct kvm_device_attr *attr)
  260. {
  261. switch (attr->group) {
  262. case KVM_DEV_VFIO_GROUP:
  263. switch (attr->attr) {
  264. case KVM_DEV_VFIO_GROUP_ADD:
  265. case KVM_DEV_VFIO_GROUP_DEL:
  266. #ifdef CONFIG_SPAPR_TCE_IOMMU
  267. case KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE:
  268. #endif
  269. return 0;
  270. }
  271. break;
  272. }
  273. return -ENXIO;
  274. }
  275. static void kvm_vfio_destroy(struct kvm_device *dev)
  276. {
  277. struct kvm_vfio *kv = dev->private;
  278. struct kvm_vfio_group *kvg, *tmp;
  279. list_for_each_entry_safe(kvg, tmp, &kv->group_list, node) {
  280. #ifdef CONFIG_SPAPR_TCE_IOMMU
  281. kvm_spapr_tce_release_vfio_group(dev->kvm, kvg->vfio_group);
  282. #endif
  283. kvm_vfio_group_set_kvm(kvg->vfio_group, NULL);
  284. kvm_vfio_group_put_external_user(kvg->vfio_group);
  285. list_del(&kvg->node);
  286. kfree(kvg);
  287. kvm_arch_end_assignment(dev->kvm);
  288. }
  289. kvm_vfio_update_coherency(dev);
  290. kfree(kv);
  291. kfree(dev); /* alloc by kvm_ioctl_create_device, free by .destroy */
  292. }
  293. static int kvm_vfio_create(struct kvm_device *dev, u32 type);
  294. static struct kvm_device_ops kvm_vfio_ops = {
  295. .name = "kvm-vfio",
  296. .create = kvm_vfio_create,
  297. .destroy = kvm_vfio_destroy,
  298. .set_attr = kvm_vfio_set_attr,
  299. .has_attr = kvm_vfio_has_attr,
  300. };
  301. static int kvm_vfio_create(struct kvm_device *dev, u32 type)
  302. {
  303. struct kvm_device *tmp;
  304. struct kvm_vfio *kv;
  305. /* Only one VFIO "device" per VM */
  306. list_for_each_entry(tmp, &dev->kvm->devices, vm_node)
  307. if (tmp->ops == &kvm_vfio_ops)
  308. return -EBUSY;
  309. kv = kzalloc(sizeof(*kv), GFP_KERNEL_ACCOUNT);
  310. if (!kv)
  311. return -ENOMEM;
  312. INIT_LIST_HEAD(&kv->group_list);
  313. mutex_init(&kv->lock);
  314. dev->private = kv;
  315. return 0;
  316. }
  317. int kvm_vfio_ops_init(void)
  318. {
  319. return kvm_register_device_ops(&kvm_vfio_ops, KVM_DEV_TYPE_VFIO);
  320. }
  321. void kvm_vfio_ops_exit(void)
  322. {
  323. kvm_unregister_device_ops(KVM_DEV_TYPE_VFIO);
  324. }