]>
Commit | Line | Data |
---|---|---|
88f5bff3 | 1 | #include <linux/btrfs.h> |
3ded6d6b | 2 | #include <linux/capability.h> |
aa269008 JB |
3 | #include <linux/cred.h> |
4 | #include <linux/mount.h> | |
88f5bff3 | 5 | #include <linux/fdtable.h> |
aa269008 JB |
6 | #include <linux/file.h> |
7 | #include <linux/fs.h> | |
8 | #include <linux/namei.h> | |
9 | #include <linux/module.h> | |
10 | #include <linux/kernel.h> | |
11 | #include <linux/magic.h> | |
12 | #include <linux/parser.h> | |
3ded6d6b | 13 | #include <linux/security.h> |
aa269008 JB |
14 | #include <linux/seq_file.h> |
15 | #include <linux/statfs.h> | |
16 | #include <linux/slab.h> | |
17 | #include <linux/user_namespace.h> | |
18 | #include <linux/uidgid.h> | |
19 | #include <linux/xattr.h> | |
3ded6d6b CB |
20 | #include <linux/posix_acl.h> |
21 | #include <linux/posix_acl_xattr.h> | |
22 | #include <linux/uio.h> | |
5b0d135d | 23 | #include <linux/fiemap.h> |
aa269008 JB |
24 | |
25 | struct shiftfs_super_info { | |
26 | struct vfsmount *mnt; | |
27 | struct user_namespace *userns; | |
3ded6d6b CB |
28 | /* creds of process who created the super block */ |
29 | const struct cred *creator_cred; | |
aa269008 | 30 | bool mark; |
3ded6d6b | 31 | unsigned int passthrough; |
14ba9a56 | 32 | unsigned int passthrough_mark; |
aa269008 JB |
33 | }; |
34 | ||
3ded6d6b CB |
35 | static void shiftfs_fill_inode(struct inode *inode, unsigned long ino, |
36 | umode_t mode, dev_t dev, struct dentry *dentry); | |
37 | ||
38 | #define SHIFTFS_PASSTHROUGH_NONE 0 | |
39 | #define SHIFTFS_PASSTHROUGH_STAT 1 | |
88f5bff3 CB |
40 | #define SHIFTFS_PASSTHROUGH_IOCTL 2 |
41 | #define SHIFTFS_PASSTHROUGH_ALL \ | |
42 | (SHIFTFS_PASSTHROUGH_STAT | SHIFTFS_PASSTHROUGH_IOCTL) | |
43 | ||
44 | static inline bool shiftfs_passthrough_ioctls(struct shiftfs_super_info *info) | |
45 | { | |
46 | if (!(info->passthrough & SHIFTFS_PASSTHROUGH_IOCTL)) | |
47 | return false; | |
48 | ||
88f5bff3 CB |
49 | return true; |
50 | } | |
3ded6d6b CB |
51 | |
52 | static inline bool shiftfs_passthrough_statfs(struct shiftfs_super_info *info) | |
53 | { | |
54 | if (!(info->passthrough & SHIFTFS_PASSTHROUGH_STAT)) | |
55 | return false; | |
56 | ||
3ded6d6b CB |
57 | return true; |
58 | } | |
aa269008 JB |
59 | |
60 | enum { | |
61 | OPT_MARK, | |
3ded6d6b | 62 | OPT_PASSTHROUGH, |
aa269008 JB |
63 | OPT_LAST, |
64 | }; | |
65 | ||
66 | /* global filesystem options */ | |
67 | static const match_table_t tokens = { | |
68 | { OPT_MARK, "mark" }, | |
3ded6d6b | 69 | { OPT_PASSTHROUGH, "passthrough=%u" }, |
aa269008 JB |
70 | { OPT_LAST, NULL } |
71 | }; | |
72 | ||
3ded6d6b | 73 | static const struct cred *shiftfs_override_creds(const struct super_block *sb) |
aa269008 | 74 | { |
3ded6d6b | 75 | struct shiftfs_super_info *sbinfo = sb->s_fs_info; |
aa269008 | 76 | |
3ded6d6b CB |
77 | return override_creds(sbinfo->creator_cred); |
78 | } | |
79 | ||
80 | static inline void shiftfs_revert_object_creds(const struct cred *oldcred, | |
81 | struct cred *newcred) | |
82 | { | |
83 | revert_creds(oldcred); | |
84 | put_cred(newcred); | |
85 | } | |
86 | ||
ef687aea SF |
87 | static kuid_t shift_kuid(struct user_namespace *from, struct user_namespace *to, |
88 | kuid_t kuid) | |
89 | { | |
90 | uid_t uid = from_kuid(from, kuid); | |
91 | return make_kuid(to, uid); | |
92 | } | |
93 | ||
94 | static kgid_t shift_kgid(struct user_namespace *from, struct user_namespace *to, | |
95 | kgid_t kgid) | |
96 | { | |
97 | gid_t gid = from_kgid(from, kgid); | |
98 | return make_kgid(to, gid); | |
99 | } | |
100 | ||
3ded6d6b CB |
101 | static int shiftfs_override_object_creds(const struct super_block *sb, |
102 | const struct cred **oldcred, | |
103 | struct cred **newcred, | |
104 | struct dentry *dentry, umode_t mode, | |
105 | bool hardlink) | |
106 | { | |
ef687aea | 107 | struct shiftfs_super_info *sbinfo = sb->s_fs_info; |
3ded6d6b CB |
108 | kuid_t fsuid = current_fsuid(); |
109 | kgid_t fsgid = current_fsgid(); | |
110 | ||
111 | *oldcred = shiftfs_override_creds(sb); | |
112 | ||
113 | *newcred = prepare_creds(); | |
114 | if (!*newcred) { | |
115 | revert_creds(*oldcred); | |
116 | return -ENOMEM; | |
117 | } | |
118 | ||
ef687aea SF |
119 | (*newcred)->fsuid = shift_kuid(sb->s_user_ns, sbinfo->userns, fsuid); |
120 | (*newcred)->fsgid = shift_kgid(sb->s_user_ns, sbinfo->userns, fsgid); | |
3ded6d6b CB |
121 | |
122 | if (!hardlink) { | |
123 | int err = security_dentry_create_files_as(dentry, mode, | |
124 | &dentry->d_name, | |
125 | *oldcred, *newcred); | |
126 | if (err) { | |
127 | shiftfs_revert_object_creds(*oldcred, *newcred); | |
128 | return err; | |
129 | } | |
130 | } | |
aa269008 | 131 | |
3ded6d6b CB |
132 | put_cred(override_creds(*newcred)); |
133 | return 0; | |
134 | } | |
aa269008 | 135 | |
3ded6d6b CB |
136 | static void shiftfs_copyattr(struct inode *from, struct inode *to) |
137 | { | |
138 | struct user_namespace *from_ns = from->i_sb->s_user_ns; | |
139 | struct user_namespace *to_ns = to->i_sb->s_user_ns; | |
140 | ||
141 | to->i_uid = shift_kuid(from_ns, to_ns, from->i_uid); | |
142 | to->i_gid = shift_kgid(from_ns, to_ns, from->i_gid); | |
143 | to->i_mode = from->i_mode; | |
144 | to->i_atime = from->i_atime; | |
145 | to->i_mtime = from->i_mtime; | |
146 | to->i_ctime = from->i_ctime; | |
147 | i_size_write(to, i_size_read(from)); | |
148 | } | |
aa269008 | 149 | |
3ded6d6b CB |
150 | static void shiftfs_copyflags(struct inode *from, struct inode *to) |
151 | { | |
152 | unsigned int mask = S_SYNC | S_IMMUTABLE | S_APPEND | S_NOATIME; | |
aa269008 | 153 | |
3ded6d6b | 154 | inode_set_flags(to, from->i_flags & mask, mask); |
aa269008 JB |
155 | } |
156 | ||
3ded6d6b | 157 | static void shiftfs_file_accessed(struct file *file) |
aa269008 | 158 | { |
3ded6d6b CB |
159 | struct inode *upperi, *loweri; |
160 | ||
161 | if (file->f_flags & O_NOATIME) | |
aa269008 JB |
162 | return; |
163 | ||
3ded6d6b CB |
164 | upperi = file_inode(file); |
165 | loweri = upperi->i_private; | |
166 | ||
167 | if (!loweri) | |
168 | return; | |
169 | ||
170 | upperi->i_mtime = loweri->i_mtime; | |
171 | upperi->i_ctime = loweri->i_ctime; | |
172 | ||
173 | touch_atime(&file->f_path); | |
aa269008 JB |
174 | } |
175 | ||
3ded6d6b CB |
176 | static int shiftfs_parse_mount_options(struct shiftfs_super_info *sbinfo, |
177 | char *options) | |
aa269008 JB |
178 | { |
179 | char *p; | |
180 | substring_t args[MAX_OPT_ARGS]; | |
181 | ||
3ded6d6b CB |
182 | sbinfo->mark = false; |
183 | sbinfo->passthrough = 0; | |
aa269008 JB |
184 | |
185 | while ((p = strsep(&options, ",")) != NULL) { | |
3ded6d6b | 186 | int err, intarg, token; |
aa269008 JB |
187 | |
188 | if (!*p) | |
189 | continue; | |
190 | ||
191 | token = match_token(p, tokens, args); | |
192 | switch (token) { | |
193 | case OPT_MARK: | |
3ded6d6b CB |
194 | sbinfo->mark = true; |
195 | break; | |
196 | case OPT_PASSTHROUGH: | |
197 | err = match_int(&args[0], &intarg); | |
198 | if (err) | |
199 | return err; | |
200 | ||
201 | if (intarg & ~SHIFTFS_PASSTHROUGH_ALL) | |
202 | return -EINVAL; | |
203 | ||
204 | sbinfo->passthrough = intarg; | |
aa269008 JB |
205 | break; |
206 | default: | |
207 | return -EINVAL; | |
208 | } | |
209 | } | |
3ded6d6b | 210 | |
aa269008 JB |
211 | return 0; |
212 | } | |
213 | ||
214 | static void shiftfs_d_release(struct dentry *dentry) | |
215 | { | |
3ded6d6b | 216 | struct dentry *lowerd = dentry->d_fsdata; |
aa269008 | 217 | |
3ded6d6b CB |
218 | if (lowerd) |
219 | dput(lowerd); | |
aa269008 JB |
220 | } |
221 | ||
222 | static struct dentry *shiftfs_d_real(struct dentry *dentry, | |
223 | const struct inode *inode) | |
224 | { | |
3ded6d6b CB |
225 | struct dentry *lowerd = dentry->d_fsdata; |
226 | ||
227 | if (inode && d_inode(dentry) == inode) | |
228 | return dentry; | |
aa269008 | 229 | |
3ded6d6b CB |
230 | lowerd = d_real(lowerd, inode); |
231 | if (lowerd && (!inode || inode == d_inode(lowerd))) | |
232 | return lowerd; | |
aa269008 | 233 | |
3ded6d6b CB |
234 | WARN(1, "shiftfs_d_real(%pd4, %s:%lu): real dentry not found\n", dentry, |
235 | inode ? inode->i_sb->s_id : "NULL", inode ? inode->i_ino : 0); | |
236 | return dentry; | |
aa269008 JB |
237 | } |
238 | ||
239 | static int shiftfs_d_weak_revalidate(struct dentry *dentry, unsigned int flags) | |
240 | { | |
3ded6d6b CB |
241 | int err = 1; |
242 | struct dentry *lowerd = dentry->d_fsdata; | |
aa269008 | 243 | |
3ded6d6b | 244 | if (d_is_negative(lowerd) != d_is_negative(dentry)) |
aa269008 JB |
245 | return 0; |
246 | ||
3ded6d6b CB |
247 | if ((lowerd->d_flags & DCACHE_OP_WEAK_REVALIDATE)) |
248 | err = lowerd->d_op->d_weak_revalidate(lowerd, flags); | |
aa269008 | 249 | |
3ded6d6b CB |
250 | if (d_really_is_positive(dentry)) { |
251 | struct inode *inode = d_inode(dentry); | |
252 | struct inode *loweri = d_inode(lowerd); | |
253 | ||
254 | shiftfs_copyattr(loweri, inode); | |
3ded6d6b CB |
255 | } |
256 | ||
257 | return err; | |
aa269008 JB |
258 | } |
259 | ||
260 | static int shiftfs_d_revalidate(struct dentry *dentry, unsigned int flags) | |
261 | { | |
3ded6d6b CB |
262 | int err = 1; |
263 | struct dentry *lowerd = dentry->d_fsdata; | |
aa269008 | 264 | |
3ded6d6b CB |
265 | if (d_unhashed(lowerd) || |
266 | ((d_is_negative(lowerd) != d_is_negative(dentry)))) | |
aa269008 JB |
267 | return 0; |
268 | ||
3ded6d6b CB |
269 | if (flags & LOOKUP_RCU) |
270 | return -ECHILD; | |
aa269008 | 271 | |
3ded6d6b CB |
272 | if ((lowerd->d_flags & DCACHE_OP_REVALIDATE)) |
273 | err = lowerd->d_op->d_revalidate(lowerd, flags); | |
aa269008 | 274 | |
3ded6d6b CB |
275 | if (d_really_is_positive(dentry)) { |
276 | struct inode *inode = d_inode(dentry); | |
277 | struct inode *loweri = d_inode(lowerd); | |
aa269008 | 278 | |
3ded6d6b | 279 | shiftfs_copyattr(loweri, inode); |
3ded6d6b | 280 | } |
aa269008 | 281 | |
3ded6d6b | 282 | return err; |
aa269008 JB |
283 | } |
284 | ||
285 | static const struct dentry_operations shiftfs_dentry_ops = { | |
3ded6d6b CB |
286 | .d_release = shiftfs_d_release, |
287 | .d_real = shiftfs_d_real, | |
288 | .d_revalidate = shiftfs_d_revalidate, | |
aa269008 JB |
289 | .d_weak_revalidate = shiftfs_d_weak_revalidate, |
290 | }; | |
291 | ||
aa269008 JB |
292 | static const char *shiftfs_get_link(struct dentry *dentry, struct inode *inode, |
293 | struct delayed_call *done) | |
294 | { | |
3ded6d6b CB |
295 | const char *p; |
296 | const struct cred *oldcred; | |
297 | struct dentry *lowerd; | |
aa269008 | 298 | |
3ded6d6b CB |
299 | /* RCU lookup not supported */ |
300 | if (!dentry) | |
aa269008 | 301 | return ERR_PTR(-ECHILD); |
3ded6d6b CB |
302 | |
303 | lowerd = dentry->d_fsdata; | |
304 | oldcred = shiftfs_override_creds(dentry->d_sb); | |
305 | p = vfs_get_link(lowerd, done); | |
306 | revert_creds(oldcred); | |
307 | ||
308 | return p; | |
aa269008 JB |
309 | } |
310 | ||
d347e71d AR |
311 | static int shiftfs_setxattr(struct user_namespace *ns, |
312 | struct dentry *dentry, struct inode *inode, | |
aa269008 JB |
313 | const char *name, const void *value, |
314 | size_t size, int flags) | |
315 | { | |
3ded6d6b CB |
316 | struct dentry *lowerd = dentry->d_fsdata; |
317 | int err; | |
318 | const struct cred *oldcred; | |
319 | ||
320 | oldcred = shiftfs_override_creds(dentry->d_sb); | |
d347e71d | 321 | err = vfs_setxattr(ns, lowerd, name, value, size, flags); |
3ded6d6b | 322 | revert_creds(oldcred); |
aa269008 | 323 | |
3ded6d6b | 324 | shiftfs_copyattr(lowerd->d_inode, inode); |
aa269008 JB |
325 | |
326 | return err; | |
327 | } | |
328 | ||
329 | static int shiftfs_xattr_get(const struct xattr_handler *handler, | |
330 | struct dentry *dentry, struct inode *inode, | |
331 | const char *name, void *value, size_t size) | |
332 | { | |
3ded6d6b | 333 | struct dentry *lowerd = dentry->d_fsdata; |
aa269008 | 334 | int err; |
3ded6d6b | 335 | const struct cred *oldcred; |
aa269008 | 336 | |
3ded6d6b | 337 | oldcred = shiftfs_override_creds(dentry->d_sb); |
d347e71d | 338 | err = vfs_getxattr(&init_user_ns, lowerd, name, value, size); |
3ded6d6b | 339 | revert_creds(oldcred); |
aa269008 JB |
340 | |
341 | return err; | |
342 | } | |
343 | ||
344 | static ssize_t shiftfs_listxattr(struct dentry *dentry, char *list, | |
345 | size_t size) | |
346 | { | |
3ded6d6b | 347 | struct dentry *lowerd = dentry->d_fsdata; |
aa269008 | 348 | int err; |
3ded6d6b | 349 | const struct cred *oldcred; |
aa269008 | 350 | |
3ded6d6b CB |
351 | oldcred = shiftfs_override_creds(dentry->d_sb); |
352 | err = vfs_listxattr(lowerd, list, size); | |
353 | revert_creds(oldcred); | |
aa269008 JB |
354 | |
355 | return err; | |
356 | } | |
357 | ||
d347e71d AR |
358 | static int shiftfs_removexattr(struct user_namespace *ns, |
359 | struct dentry *dentry, const char *name) | |
aa269008 | 360 | { |
3ded6d6b | 361 | struct dentry *lowerd = dentry->d_fsdata; |
aa269008 | 362 | int err; |
3ded6d6b CB |
363 | const struct cred *oldcred; |
364 | ||
365 | oldcred = shiftfs_override_creds(dentry->d_sb); | |
d347e71d | 366 | err = vfs_removexattr(ns, lowerd, name); |
3ded6d6b | 367 | revert_creds(oldcred); |
aa269008 | 368 | |
3ded6d6b CB |
369 | /* update c/mtime */ |
370 | shiftfs_copyattr(lowerd->d_inode, d_inode(dentry)); | |
aa269008 JB |
371 | |
372 | return err; | |
373 | } | |
374 | ||
375 | static int shiftfs_xattr_set(const struct xattr_handler *handler, | |
d347e71d | 376 | struct user_namespace *ns, |
aa269008 JB |
377 | struct dentry *dentry, struct inode *inode, |
378 | const char *name, const void *value, size_t size, | |
379 | int flags) | |
380 | { | |
381 | if (!value) | |
d347e71d AR |
382 | return shiftfs_removexattr(ns, dentry, name); |
383 | return shiftfs_setxattr(ns, dentry, inode, name, value, size, flags); | |
aa269008 JB |
384 | } |
385 | ||
3ded6d6b | 386 | static int shiftfs_inode_test(struct inode *inode, void *data) |
aa269008 | 387 | { |
3ded6d6b CB |
388 | return inode->i_private == data; |
389 | } | |
aa269008 | 390 | |
3ded6d6b CB |
391 | static int shiftfs_inode_set(struct inode *inode, void *data) |
392 | { | |
393 | inode->i_private = data; | |
394 | return 0; | |
aa269008 JB |
395 | } |
396 | ||
d347e71d AR |
397 | static int shiftfs_create_object(struct user_namespace *ns, |
398 | struct inode *diri, struct dentry *dentry, | |
3ded6d6b CB |
399 | umode_t mode, const char *symlink, |
400 | struct dentry *hardlink, bool excl) | |
aa269008 | 401 | { |
aa269008 | 402 | int err; |
3ded6d6b CB |
403 | const struct cred *oldcred; |
404 | struct cred *newcred; | |
405 | void *loweri_iop_ptr = NULL; | |
406 | umode_t modei = mode; | |
407 | struct super_block *dir_sb = diri->i_sb; | |
408 | struct dentry *lowerd_new = dentry->d_fsdata; | |
409 | struct inode *inode = NULL, *loweri_dir = diri->i_private; | |
410 | const struct inode_operations *loweri_dir_iop = loweri_dir->i_op; | |
411 | struct dentry *lowerd_link = NULL; | |
aa269008 JB |
412 | |
413 | if (hardlink) { | |
3ded6d6b | 414 | loweri_iop_ptr = loweri_dir_iop->link; |
aa269008 JB |
415 | } else { |
416 | switch (mode & S_IFMT) { | |
417 | case S_IFDIR: | |
3ded6d6b | 418 | loweri_iop_ptr = loweri_dir_iop->mkdir; |
aa269008 JB |
419 | break; |
420 | case S_IFREG: | |
3ded6d6b | 421 | loweri_iop_ptr = loweri_dir_iop->create; |
aa269008 JB |
422 | break; |
423 | case S_IFLNK: | |
3ded6d6b CB |
424 | loweri_iop_ptr = loweri_dir_iop->symlink; |
425 | break; | |
426 | case S_IFSOCK: | |
427 | /* fall through */ | |
428 | case S_IFIFO: | |
429 | loweri_iop_ptr = loweri_dir_iop->mknod; | |
430 | break; | |
aa269008 JB |
431 | } |
432 | } | |
3ded6d6b CB |
433 | if (!loweri_iop_ptr) { |
434 | err = -EINVAL; | |
435 | goto out_iput; | |
436 | } | |
aa269008 | 437 | |
3ded6d6b | 438 | inode_lock_nested(loweri_dir, I_MUTEX_PARENT); |
aa269008 | 439 | |
3ded6d6b CB |
440 | if (!hardlink) { |
441 | inode = new_inode(dir_sb); | |
442 | if (!inode) { | |
443 | err = -ENOMEM; | |
444 | goto out_iput; | |
445 | } | |
446 | ||
447 | /* | |
448 | * new_inode() will have added the new inode to the super | |
449 | * block's list of inodes. Further below we will call | |
450 | * inode_insert5() Which would perform the same operation again | |
451 | * thereby corrupting the list. To avoid this raise I_CREATING | |
452 | * in i_state which will cause inode_insert5() to skip this | |
453 | * step. I_CREATING will be cleared by d_instantiate_new() | |
454 | * below. | |
455 | */ | |
456 | spin_lock(&inode->i_lock); | |
457 | inode->i_state |= I_CREATING; | |
458 | spin_unlock(&inode->i_lock); | |
aa269008 | 459 | |
d347e71d | 460 | inode_init_owner(ns, inode, diri, mode); |
3ded6d6b CB |
461 | modei = inode->i_mode; |
462 | } | |
aa269008 | 463 | |
3ded6d6b CB |
464 | err = shiftfs_override_object_creds(dentry->d_sb, &oldcred, &newcred, |
465 | dentry, modei, hardlink != NULL); | |
466 | if (err) | |
467 | goto out_iput; | |
aa269008 | 468 | |
aa269008 | 469 | if (hardlink) { |
3ded6d6b | 470 | lowerd_link = hardlink->d_fsdata; |
d347e71d | 471 | err = vfs_link(lowerd_link, ns, loweri_dir, lowerd_new, NULL); |
aa269008 | 472 | } else { |
3ded6d6b | 473 | switch (modei & S_IFMT) { |
aa269008 | 474 | case S_IFDIR: |
d347e71d | 475 | err = vfs_mkdir(ns, loweri_dir, lowerd_new, modei); |
aa269008 JB |
476 | break; |
477 | case S_IFREG: | |
d347e71d | 478 | err = vfs_create(ns, loweri_dir, lowerd_new, modei, excl); |
aa269008 JB |
479 | break; |
480 | case S_IFLNK: | |
d347e71d | 481 | err = vfs_symlink(ns, loweri_dir, lowerd_new, symlink); |
3ded6d6b CB |
482 | break; |
483 | case S_IFSOCK: | |
484 | /* fall through */ | |
485 | case S_IFIFO: | |
d347e71d | 486 | err = vfs_mknod(ns, loweri_dir, lowerd_new, modei, 0); |
3ded6d6b CB |
487 | break; |
488 | default: | |
489 | err = -EINVAL; | |
490 | break; | |
aa269008 JB |
491 | } |
492 | } | |
493 | ||
3ded6d6b | 494 | shiftfs_revert_object_creds(oldcred, newcred); |
aa269008 | 495 | |
3ded6d6b CB |
496 | if (!err && WARN_ON(!lowerd_new->d_inode)) |
497 | err = -EIO; | |
aa269008 | 498 | if (err) |
3ded6d6b CB |
499 | goto out_iput; |
500 | ||
501 | if (hardlink) { | |
502 | inode = d_inode(hardlink); | |
503 | ihold(inode); | |
504 | ||
505 | /* copy up times from lower inode */ | |
506 | shiftfs_copyattr(d_inode(lowerd_link), inode); | |
507 | set_nlink(d_inode(hardlink), d_inode(lowerd_link)->i_nlink); | |
508 | d_instantiate(dentry, inode); | |
509 | } else { | |
510 | struct inode *inode_tmp; | |
511 | struct inode *loweri_new = d_inode(lowerd_new); | |
512 | ||
513 | inode_tmp = inode_insert5(inode, (unsigned long)loweri_new, | |
514 | shiftfs_inode_test, shiftfs_inode_set, | |
515 | loweri_new); | |
516 | if (unlikely(inode_tmp != inode)) { | |
517 | pr_err_ratelimited("shiftfs: newly created inode found in cache\n"); | |
518 | iput(inode_tmp); | |
519 | err = -EINVAL; | |
520 | goto out_iput; | |
521 | } | |
aa269008 | 522 | |
3ded6d6b CB |
523 | ihold(loweri_new); |
524 | shiftfs_fill_inode(inode, loweri_new->i_ino, loweri_new->i_mode, | |
525 | 0, lowerd_new); | |
526 | d_instantiate_new(dentry, inode); | |
527 | } | |
aa269008 | 528 | |
3ded6d6b CB |
529 | shiftfs_copyattr(loweri_dir, diri); |
530 | if (loweri_iop_ptr == loweri_dir_iop->mkdir) | |
531 | set_nlink(diri, loweri_dir->i_nlink); | |
aa269008 | 532 | |
3ded6d6b | 533 | inode = NULL; |
aa269008 | 534 | |
3ded6d6b CB |
535 | out_iput: |
536 | iput(inode); | |
537 | inode_unlock(loweri_dir); | |
aa269008 JB |
538 | |
539 | return err; | |
540 | } | |
541 | ||
d347e71d AR |
542 | static int shiftfs_create(struct user_namespace *ns, |
543 | struct inode *dir, struct dentry *dentry, | |
aa269008 JB |
544 | umode_t mode, bool excl) |
545 | { | |
546 | mode |= S_IFREG; | |
547 | ||
d347e71d | 548 | return shiftfs_create_object(ns, dir, dentry, mode, NULL, NULL, excl); |
aa269008 JB |
549 | } |
550 | ||
d347e71d | 551 | static int shiftfs_mkdir(struct user_namespace *ns, struct inode *dir, struct dentry *dentry, |
aa269008 JB |
552 | umode_t mode) |
553 | { | |
554 | mode |= S_IFDIR; | |
555 | ||
d347e71d | 556 | return shiftfs_create_object(ns, dir, dentry, mode, NULL, NULL, false); |
aa269008 JB |
557 | } |
558 | ||
559 | static int shiftfs_link(struct dentry *hardlink, struct inode *dir, | |
560 | struct dentry *dentry) | |
561 | { | |
d347e71d | 562 | return shiftfs_create_object(&init_user_ns, dir, dentry, 0, NULL, hardlink, false); |
3ded6d6b CB |
563 | } |
564 | ||
d347e71d AR |
565 | static int shiftfs_mknod(struct user_namespace *ns, |
566 | struct inode *dir, struct dentry *dentry, umode_t mode, | |
3ded6d6b CB |
567 | dev_t rdev) |
568 | { | |
569 | if (!S_ISFIFO(mode) && !S_ISSOCK(mode)) | |
570 | return -EPERM; | |
571 | ||
d347e71d | 572 | return shiftfs_create_object(ns, dir, dentry, mode, NULL, NULL, false); |
aa269008 JB |
573 | } |
574 | ||
d347e71d | 575 | static int shiftfs_symlink(struct user_namespace *ns, struct inode *dir, struct dentry *dentry, |
aa269008 JB |
576 | const char *symlink) |
577 | { | |
d347e71d | 578 | return shiftfs_create_object(ns, dir, dentry, S_IFLNK, symlink, NULL, false); |
aa269008 JB |
579 | } |
580 | ||
581 | static int shiftfs_rm(struct inode *dir, struct dentry *dentry, bool rmdir) | |
582 | { | |
3ded6d6b CB |
583 | struct dentry *lowerd = dentry->d_fsdata; |
584 | struct inode *loweri = dir->i_private; | |
b8abaae8 | 585 | struct inode *inode = d_inode(dentry); |
aa269008 | 586 | int err; |
3ded6d6b | 587 | const struct cred *oldcred; |
aa269008 | 588 | |
a7a5232f | 589 | dget(lowerd); |
3ded6d6b CB |
590 | oldcred = shiftfs_override_creds(dentry->d_sb); |
591 | inode_lock_nested(loweri, I_MUTEX_PARENT); | |
aa269008 | 592 | if (rmdir) |
d347e71d | 593 | err = vfs_rmdir(&init_user_ns, loweri, lowerd); |
aa269008 | 594 | else |
d347e71d | 595 | err = vfs_unlink(&init_user_ns, loweri, lowerd, NULL); |
3ded6d6b | 596 | revert_creds(oldcred); |
aa269008 | 597 | |
b8abaae8 | 598 | if (!err) { |
3ded6d6b CB |
599 | d_drop(dentry); |
600 | ||
b8abaae8 CB |
601 | if (rmdir) |
602 | clear_nlink(inode); | |
603 | else | |
604 | drop_nlink(inode); | |
605 | } | |
606 | inode_unlock(loweri); | |
607 | ||
608 | shiftfs_copyattr(loweri, dir); | |
a7a5232f | 609 | dput(lowerd); |
aa269008 JB |
610 | |
611 | return err; | |
612 | } | |
613 | ||
614 | static int shiftfs_unlink(struct inode *dir, struct dentry *dentry) | |
615 | { | |
616 | return shiftfs_rm(dir, dentry, false); | |
617 | } | |
618 | ||
619 | static int shiftfs_rmdir(struct inode *dir, struct dentry *dentry) | |
620 | { | |
621 | return shiftfs_rm(dir, dentry, true); | |
622 | } | |
623 | ||
d347e71d AR |
624 | static int shiftfs_rename(struct user_namespace *ns, |
625 | struct inode *olddir, struct dentry *old, | |
aa269008 JB |
626 | struct inode *newdir, struct dentry *new, |
627 | unsigned int flags) | |
628 | { | |
3ded6d6b CB |
629 | struct dentry *lowerd_dir_old = old->d_parent->d_fsdata, |
630 | *lowerd_dir_new = new->d_parent->d_fsdata, | |
631 | *lowerd_old = old->d_fsdata, *lowerd_new = new->d_fsdata, | |
632 | *trapd; | |
633 | struct inode *loweri_dir_old = lowerd_dir_old->d_inode, | |
634 | *loweri_dir_new = lowerd_dir_new->d_inode; | |
d347e71d AR |
635 | struct renamedata rd = { |
636 | .old_mnt_userns = ns, | |
637 | .old_dir = loweri_dir_old, | |
638 | .old_dentry = lowerd_old, | |
639 | .new_mnt_userns = ns, | |
640 | .new_dir = loweri_dir_new, | |
641 | .new_dentry = lowerd_new, | |
642 | }; | |
aa269008 | 643 | int err = -EINVAL; |
3ded6d6b | 644 | const struct cred *oldcred; |
aa269008 | 645 | |
3ded6d6b | 646 | trapd = lock_rename(lowerd_dir_new, lowerd_dir_old); |
aa269008 | 647 | |
3ded6d6b | 648 | if (trapd == lowerd_old || trapd == lowerd_new) |
aa269008 JB |
649 | goto out_unlock; |
650 | ||
3ded6d6b | 651 | oldcred = shiftfs_override_creds(old->d_sb); |
d347e71d | 652 | err = vfs_rename(&rd); |
3ded6d6b | 653 | revert_creds(oldcred); |
aa269008 | 654 | |
3ded6d6b CB |
655 | shiftfs_copyattr(loweri_dir_old, olddir); |
656 | shiftfs_copyattr(loweri_dir_new, newdir); | |
aa269008 | 657 | |
3ded6d6b CB |
658 | out_unlock: |
659 | unlock_rename(lowerd_dir_new, lowerd_dir_old); | |
aa269008 JB |
660 | |
661 | return err; | |
662 | } | |
663 | ||
664 | static struct dentry *shiftfs_lookup(struct inode *dir, struct dentry *dentry, | |
665 | unsigned int flags) | |
666 | { | |
3ded6d6b CB |
667 | struct dentry *new; |
668 | struct inode *newi; | |
669 | const struct cred *oldcred; | |
670 | struct dentry *lowerd = dentry->d_parent->d_fsdata; | |
671 | struct inode *inode = NULL, *loweri = lowerd->d_inode; | |
672 | ||
673 | inode_lock(loweri); | |
674 | oldcred = shiftfs_override_creds(dentry->d_sb); | |
675 | new = lookup_one_len(dentry->d_name.name, lowerd, dentry->d_name.len); | |
676 | revert_creds(oldcred); | |
677 | inode_unlock(loweri); | |
aa269008 JB |
678 | |
679 | if (IS_ERR(new)) | |
680 | return new; | |
681 | ||
682 | dentry->d_fsdata = new; | |
683 | ||
3ded6d6b CB |
684 | newi = new->d_inode; |
685 | if (!newi) | |
aa269008 JB |
686 | goto out; |
687 | ||
3ded6d6b CB |
688 | inode = iget5_locked(dentry->d_sb, (unsigned long)newi, |
689 | shiftfs_inode_test, shiftfs_inode_set, newi); | |
690 | if (!inode) { | |
aa269008 JB |
691 | dput(new); |
692 | return ERR_PTR(-ENOMEM); | |
693 | } | |
3ded6d6b CB |
694 | if (inode->i_state & I_NEW) { |
695 | /* | |
696 | * inode->i_private set by shiftfs_inode_set(), but we still | |
697 | * need to take a reference | |
698 | */ | |
699 | ihold(newi); | |
700 | shiftfs_fill_inode(inode, newi->i_ino, newi->i_mode, 0, new); | |
701 | unlock_new_inode(inode); | |
702 | } | |
aa269008 | 703 | |
3ded6d6b CB |
704 | out: |
705 | return d_splice_alias(inode, dentry); | |
aa269008 JB |
706 | } |
707 | ||
d347e71d | 708 | static int shiftfs_permission(struct user_namespace *ns, struct inode *inode, int mask) |
aa269008 | 709 | { |
aa269008 | 710 | int err; |
3ded6d6b CB |
711 | const struct cred *oldcred; |
712 | struct inode *loweri = inode->i_private; | |
aa269008 | 713 | |
3ded6d6b CB |
714 | if (!loweri) { |
715 | WARN_ON(!(mask & MAY_NOT_BLOCK)); | |
aa269008 | 716 | return -ECHILD; |
3ded6d6b | 717 | } |
aa269008 | 718 | |
d347e71d | 719 | err = generic_permission(ns, inode, mask); |
3ded6d6b CB |
720 | if (err) |
721 | return err; | |
722 | ||
723 | oldcred = shiftfs_override_creds(inode->i_sb); | |
d347e71d | 724 | err = inode_permission(ns, loweri, mask); |
3ded6d6b CB |
725 | revert_creds(oldcred); |
726 | ||
727 | return err; | |
728 | } | |
729 | ||
730 | static int shiftfs_fiemap(struct inode *inode, | |
731 | struct fiemap_extent_info *fieinfo, u64 start, | |
732 | u64 len) | |
733 | { | |
734 | int err; | |
735 | const struct cred *oldcred; | |
736 | struct inode *loweri = inode->i_private; | |
737 | ||
738 | if (!loweri->i_op->fiemap) | |
739 | return -EOPNOTSUPP; | |
740 | ||
741 | oldcred = shiftfs_override_creds(inode->i_sb); | |
742 | if (fieinfo->fi_flags & FIEMAP_FLAG_SYNC) | |
743 | filemap_write_and_wait(loweri->i_mapping); | |
744 | err = loweri->i_op->fiemap(loweri, fieinfo, start, len); | |
745 | revert_creds(oldcred); | |
746 | ||
747 | return err; | |
748 | } | |
749 | ||
d347e71d AR |
750 | static int shiftfs_tmpfile(struct user_namespace *ns, |
751 | struct inode *dir, struct dentry *dentry, | |
3ded6d6b CB |
752 | umode_t mode) |
753 | { | |
754 | int err; | |
755 | const struct cred *oldcred; | |
756 | struct dentry *lowerd = dentry->d_fsdata; | |
757 | struct inode *loweri = dir->i_private; | |
758 | ||
759 | if (!loweri->i_op->tmpfile) | |
760 | return -EOPNOTSUPP; | |
761 | ||
762 | oldcred = shiftfs_override_creds(dir->i_sb); | |
d347e71d | 763 | err = loweri->i_op->tmpfile(ns, loweri, lowerd, mode); |
3ded6d6b | 764 | revert_creds(oldcred); |
aa269008 JB |
765 | |
766 | return err; | |
767 | } | |
768 | ||
d347e71d | 769 | static int shiftfs_setattr(struct user_namespace *ns, struct dentry *dentry, struct iattr *attr) |
aa269008 | 770 | { |
3ded6d6b CB |
771 | struct dentry *lowerd = dentry->d_fsdata; |
772 | struct inode *loweri = lowerd->d_inode; | |
25af0f23 | 773 | struct iattr newattr; |
3ded6d6b | 774 | const struct cred *oldcred; |
aa269008 | 775 | struct super_block *sb = dentry->d_sb; |
ef687aea | 776 | struct shiftfs_super_info *sbinfo = sb->s_fs_info; |
aa269008 JB |
777 | int err; |
778 | ||
d347e71d | 779 | err = setattr_prepare(ns, dentry, attr); |
3ded6d6b CB |
780 | if (err) |
781 | return err; | |
782 | ||
25af0f23 | 783 | newattr = *attr; |
ef687aea SF |
784 | newattr.ia_uid = shift_kuid(sb->s_user_ns, sbinfo->userns, attr->ia_uid); |
785 | newattr.ia_gid = shift_kgid(sb->s_user_ns, sbinfo->userns, attr->ia_gid); | |
aa269008 | 786 | |
25af0f23 SF |
787 | /* |
788 | * mode change is for clearing setuid/setgid bits. Allow lower fs | |
789 | * to interpret this in its own way. | |
790 | */ | |
791 | if (newattr.ia_valid & (ATTR_KILL_SUID|ATTR_KILL_SGID)) | |
792 | newattr.ia_valid &= ~ATTR_MODE; | |
793 | ||
3ded6d6b CB |
794 | inode_lock(loweri); |
795 | oldcred = shiftfs_override_creds(dentry->d_sb); | |
d347e71d | 796 | err = notify_change(ns, lowerd, &newattr, NULL); |
3ded6d6b CB |
797 | revert_creds(oldcred); |
798 | inode_unlock(loweri); | |
aa269008 | 799 | |
3ded6d6b | 800 | shiftfs_copyattr(loweri, d_inode(dentry)); |
aa269008 | 801 | |
3ded6d6b | 802 | return err; |
aa269008 JB |
803 | } |
804 | ||
d347e71d AR |
805 | static int shiftfs_getattr(struct user_namespace *ns, |
806 | const struct path *path, struct kstat *stat, | |
aa269008 JB |
807 | u32 request_mask, unsigned int query_flags) |
808 | { | |
809 | struct inode *inode = path->dentry->d_inode; | |
3ded6d6b CB |
810 | struct dentry *lowerd = path->dentry->d_fsdata; |
811 | struct inode *loweri = lowerd->d_inode; | |
812 | struct shiftfs_super_info *info = path->dentry->d_sb->s_fs_info; | |
813 | struct path newpath = { .mnt = info->mnt, .dentry = lowerd }; | |
814 | struct user_namespace *from_ns = loweri->i_sb->s_user_ns; | |
815 | struct user_namespace *to_ns = inode->i_sb->s_user_ns; | |
816 | const struct cred *oldcred; | |
817 | int err; | |
818 | ||
819 | oldcred = shiftfs_override_creds(inode->i_sb); | |
820 | err = vfs_getattr(&newpath, stat, request_mask, query_flags); | |
821 | revert_creds(oldcred); | |
aa269008 JB |
822 | |
823 | if (err) | |
824 | return err; | |
825 | ||
826 | /* transform the underlying id */ | |
3ded6d6b CB |
827 | stat->uid = shift_kuid(from_ns, to_ns, stat->uid); |
828 | stat->gid = shift_kgid(from_ns, to_ns, stat->gid); | |
aa269008 JB |
829 | return 0; |
830 | } | |
831 | ||
3ded6d6b | 832 | #ifdef CONFIG_SHIFT_FS_POSIX_ACL |
aa269008 | 833 | |
3ded6d6b CB |
834 | static int |
835 | shift_acl_ids(struct user_namespace *from, struct user_namespace *to, | |
836 | struct posix_acl *acl) | |
aa269008 | 837 | { |
3ded6d6b CB |
838 | int i; |
839 | ||
840 | for (i = 0; i < acl->a_count; i++) { | |
841 | struct posix_acl_entry *e = &acl->a_entries[i]; | |
842 | switch(e->e_tag) { | |
843 | case ACL_USER: | |
844 | e->e_uid = shift_kuid(from, to, e->e_uid); | |
845 | if (!uid_valid(e->e_uid)) | |
846 | return -EOVERFLOW; | |
847 | break; | |
848 | case ACL_GROUP: | |
849 | e->e_gid = shift_kgid(from, to, e->e_gid); | |
850 | if (!gid_valid(e->e_gid)) | |
851 | return -EOVERFLOW; | |
852 | break; | |
853 | } | |
854 | } | |
855 | return 0; | |
856 | } | |
aa269008 | 857 | |
3ded6d6b CB |
858 | static void |
859 | shift_acl_xattr_ids(struct user_namespace *from, struct user_namespace *to, | |
860 | void *value, size_t size) | |
861 | { | |
862 | struct posix_acl_xattr_header *header = value; | |
863 | struct posix_acl_xattr_entry *entry = (void *)(header + 1), *end; | |
864 | int count; | |
865 | kuid_t kuid; | |
866 | kgid_t kgid; | |
aa269008 | 867 | |
3ded6d6b CB |
868 | if (!value) |
869 | return; | |
870 | if (size < sizeof(struct posix_acl_xattr_header)) | |
871 | return; | |
872 | if (header->a_version != cpu_to_le32(POSIX_ACL_XATTR_VERSION)) | |
873 | return; | |
aa269008 | 874 | |
3ded6d6b CB |
875 | count = posix_acl_xattr_count(size); |
876 | if (count < 0) | |
877 | return; | |
878 | if (count == 0) | |
879 | return; | |
aa269008 | 880 | |
3ded6d6b CB |
881 | for (end = entry + count; entry != end; entry++) { |
882 | switch(le16_to_cpu(entry->e_tag)) { | |
883 | case ACL_USER: | |
884 | kuid = make_kuid(&init_user_ns, le32_to_cpu(entry->e_id)); | |
885 | kuid = shift_kuid(from, to, kuid); | |
886 | entry->e_id = cpu_to_le32(from_kuid(&init_user_ns, kuid)); | |
887 | break; | |
888 | case ACL_GROUP: | |
d347e71d | 889 | kgid = make_kgid(from, le32_to_cpu(entry->e_id)); |
3ded6d6b | 890 | kgid = shift_kgid(from, to, kgid); |
d347e71d | 891 | entry->e_id = cpu_to_le32(from_kgid(from, kgid)); |
3ded6d6b CB |
892 | break; |
893 | default: | |
894 | break; | |
895 | } | |
896 | } | |
aa269008 JB |
897 | } |
898 | ||
d347e71d AR |
899 | static struct posix_acl * |
900 | shiftfs_get_acl(struct inode *inode, int type, bool rcu) | |
aa269008 | 901 | { |
3ded6d6b CB |
902 | struct inode *loweri = inode->i_private; |
903 | const struct cred *oldcred; | |
904 | struct posix_acl *lower_acl, *acl = NULL; | |
905 | struct user_namespace *from_ns = loweri->i_sb->s_user_ns; | |
906 | struct user_namespace *to_ns = inode->i_sb->s_user_ns; | |
907 | int size; | |
908 | int err; | |
aa269008 | 909 | |
d347e71d AR |
910 | if (rcu) |
911 | return ERR_PTR(-ECHILD); | |
912 | ||
3ded6d6b CB |
913 | if (!IS_POSIXACL(loweri)) |
914 | return NULL; | |
aa269008 | 915 | |
3ded6d6b CB |
916 | oldcred = shiftfs_override_creds(inode->i_sb); |
917 | lower_acl = get_acl(loweri, type); | |
918 | revert_creds(oldcred); | |
aa269008 | 919 | |
3ded6d6b CB |
920 | if (lower_acl && !IS_ERR(lower_acl)) { |
921 | /* XXX: export posix_acl_clone? */ | |
922 | size = sizeof(struct posix_acl) + | |
923 | lower_acl->a_count * sizeof(struct posix_acl_entry); | |
924 | acl = kmemdup(lower_acl, size, GFP_KERNEL); | |
925 | posix_acl_release(lower_acl); | |
aa269008 | 926 | |
3ded6d6b CB |
927 | if (!acl) |
928 | return ERR_PTR(-ENOMEM); | |
aa269008 | 929 | |
3ded6d6b | 930 | refcount_set(&acl->a_refcount, 1); |
aa269008 | 931 | |
3ded6d6b CB |
932 | err = shift_acl_ids(from_ns, to_ns, acl); |
933 | if (err) { | |
934 | kfree(acl); | |
935 | return ERR_PTR(err); | |
936 | } | |
937 | } | |
938 | ||
939 | return acl; | |
aa269008 JB |
940 | } |
941 | ||
3ded6d6b CB |
942 | static int |
943 | shiftfs_posix_acl_xattr_get(const struct xattr_handler *handler, | |
944 | struct dentry *dentry, struct inode *inode, | |
945 | const char *name, void *buffer, size_t size) | |
aa269008 | 946 | { |
3ded6d6b CB |
947 | struct inode *loweri = inode->i_private; |
948 | int ret; | |
949 | ||
950 | ret = shiftfs_xattr_get(NULL, dentry, inode, handler->name, | |
951 | buffer, size); | |
952 | if (ret < 0) | |
953 | return ret; | |
aa269008 | 954 | |
3ded6d6b CB |
955 | inode_lock(loweri); |
956 | shift_acl_xattr_ids(loweri->i_sb->s_user_ns, inode->i_sb->s_user_ns, | |
957 | buffer, size); | |
958 | inode_unlock(loweri); | |
959 | return ret; | |
aa269008 JB |
960 | } |
961 | ||
3ded6d6b CB |
962 | static int |
963 | shiftfs_posix_acl_xattr_set(const struct xattr_handler *handler, | |
d347e71d | 964 | struct user_namespace *ns, |
3ded6d6b CB |
965 | struct dentry *dentry, struct inode *inode, |
966 | const char *name, const void *value, | |
967 | size_t size, int flags) | |
968 | { | |
969 | struct inode *loweri = inode->i_private; | |
970 | int err; | |
aa269008 | 971 | |
3ded6d6b CB |
972 | if (!IS_POSIXACL(loweri) || !loweri->i_op->set_acl) |
973 | return -EOPNOTSUPP; | |
974 | if (handler->flags == ACL_TYPE_DEFAULT && !S_ISDIR(inode->i_mode)) | |
975 | return value ? -EACCES : 0; | |
d347e71d | 976 | if (!inode_owner_or_capable(ns, inode)) |
3ded6d6b CB |
977 | return -EPERM; |
978 | ||
979 | if (value) { | |
980 | shift_acl_xattr_ids(inode->i_sb->s_user_ns, | |
981 | loweri->i_sb->s_user_ns, | |
982 | (void *)value, size); | |
d347e71d | 983 | err = shiftfs_setxattr(ns, dentry, inode, handler->name, value, |
3ded6d6b CB |
984 | size, flags); |
985 | } else { | |
d347e71d | 986 | err = shiftfs_removexattr(ns, dentry, handler->name); |
3ded6d6b | 987 | } |
aa269008 | 988 | |
3ded6d6b CB |
989 | if (!err) |
990 | shiftfs_copyattr(loweri, inode); | |
991 | ||
992 | return err; | |
993 | } | |
994 | ||
995 | static const struct xattr_handler | |
996 | shiftfs_posix_acl_access_xattr_handler = { | |
997 | .name = XATTR_NAME_POSIX_ACL_ACCESS, | |
998 | .flags = ACL_TYPE_ACCESS, | |
999 | .get = shiftfs_posix_acl_xattr_get, | |
1000 | .set = shiftfs_posix_acl_xattr_set, | |
aa269008 JB |
1001 | }; |
1002 | ||
3ded6d6b CB |
1003 | static const struct xattr_handler |
1004 | shiftfs_posix_acl_default_xattr_handler = { | |
1005 | .name = XATTR_NAME_POSIX_ACL_DEFAULT, | |
1006 | .flags = ACL_TYPE_DEFAULT, | |
1007 | .get = shiftfs_posix_acl_xattr_get, | |
1008 | .set = shiftfs_posix_acl_xattr_set, | |
aa269008 JB |
1009 | }; |
1010 | ||
3ded6d6b | 1011 | #else /* !CONFIG_SHIFT_FS_POSIX_ACL */ |
aa269008 | 1012 | |
3ded6d6b | 1013 | #define shiftfs_get_acl NULL |
aa269008 | 1014 | |
3ded6d6b | 1015 | #endif /* CONFIG_SHIFT_FS_POSIX_ACL */ |
aa269008 | 1016 | |
3ded6d6b CB |
1017 | static const struct inode_operations shiftfs_dir_inode_operations = { |
1018 | .lookup = shiftfs_lookup, | |
1019 | .mkdir = shiftfs_mkdir, | |
1020 | .symlink = shiftfs_symlink, | |
1021 | .unlink = shiftfs_unlink, | |
1022 | .rmdir = shiftfs_rmdir, | |
1023 | .rename = shiftfs_rename, | |
1024 | .link = shiftfs_link, | |
1025 | .setattr = shiftfs_setattr, | |
1026 | .create = shiftfs_create, | |
1027 | .mknod = shiftfs_mknod, | |
1028 | .permission = shiftfs_permission, | |
1029 | .getattr = shiftfs_getattr, | |
1030 | .listxattr = shiftfs_listxattr, | |
1031 | .get_acl = shiftfs_get_acl, | |
1032 | }; | |
1033 | ||
1034 | static const struct inode_operations shiftfs_file_inode_operations = { | |
1035 | .fiemap = shiftfs_fiemap, | |
1036 | .getattr = shiftfs_getattr, | |
1037 | .get_acl = shiftfs_get_acl, | |
1038 | .listxattr = shiftfs_listxattr, | |
1039 | .permission = shiftfs_permission, | |
1040 | .setattr = shiftfs_setattr, | |
1041 | .tmpfile = shiftfs_tmpfile, | |
1042 | }; | |
1043 | ||
1044 | static const struct inode_operations shiftfs_special_inode_operations = { | |
1045 | .getattr = shiftfs_getattr, | |
1046 | .get_acl = shiftfs_get_acl, | |
1047 | .listxattr = shiftfs_listxattr, | |
1048 | .permission = shiftfs_permission, | |
1049 | .setattr = shiftfs_setattr, | |
1050 | }; | |
1051 | ||
1052 | static const struct inode_operations shiftfs_symlink_inode_operations = { | |
1053 | .getattr = shiftfs_getattr, | |
1054 | .get_link = shiftfs_get_link, | |
1055 | .listxattr = shiftfs_listxattr, | |
1056 | .setattr = shiftfs_setattr, | |
1057 | }; | |
1058 | ||
1059 | static struct file *shiftfs_open_realfile(const struct file *file, | |
b4de76a8 | 1060 | struct inode *realinode) |
3ded6d6b | 1061 | { |
b4de76a8 CB |
1062 | struct file *realfile; |
1063 | const struct cred *old_cred; | |
3ded6d6b | 1064 | struct inode *inode = file_inode(file); |
b4de76a8 | 1065 | struct dentry *lowerd = file->f_path.dentry->d_fsdata; |
3ded6d6b | 1066 | struct shiftfs_super_info *info = inode->i_sb->s_fs_info; |
b4de76a8 | 1067 | struct path realpath = { .mnt = info->mnt, .dentry = lowerd }; |
3ded6d6b | 1068 | |
b4de76a8 CB |
1069 | old_cred = shiftfs_override_creds(inode->i_sb); |
1070 | realfile = open_with_fake_path(&realpath, file->f_flags, realinode, | |
1071 | info->creator_cred); | |
1072 | revert_creds(old_cred); | |
3ded6d6b | 1073 | |
b4de76a8 | 1074 | return realfile; |
3ded6d6b CB |
1075 | } |
1076 | ||
1077 | #define SHIFTFS_SETFL_MASK (O_APPEND | O_NONBLOCK | O_NDELAY | O_DIRECT) | |
1078 | ||
1079 | static int shiftfs_change_flags(struct file *file, unsigned int flags) | |
1080 | { | |
1081 | struct inode *inode = file_inode(file); | |
1082 | int err; | |
1083 | ||
1084 | /* if some flag changed that cannot be changed then something's amiss */ | |
1085 | if (WARN_ON((file->f_flags ^ flags) & ~SHIFTFS_SETFL_MASK)) | |
1086 | return -EIO; | |
1087 | ||
1088 | flags &= SHIFTFS_SETFL_MASK; | |
1089 | ||
1090 | if (((flags ^ file->f_flags) & O_APPEND) && IS_APPEND(inode)) | |
1091 | return -EPERM; | |
1092 | ||
1093 | if (flags & O_DIRECT) { | |
1094 | if (!file->f_mapping->a_ops || | |
1095 | !file->f_mapping->a_ops->direct_IO) | |
1096 | return -EINVAL; | |
1097 | } | |
1098 | ||
1099 | if (file->f_op->check_flags) { | |
1100 | err = file->f_op->check_flags(flags); | |
1101 | if (err) | |
1102 | return err; | |
1103 | } | |
1104 | ||
1105 | spin_lock(&file->f_lock); | |
1106 | file->f_flags = (file->f_flags & ~SHIFTFS_SETFL_MASK) | flags; | |
1107 | spin_unlock(&file->f_lock); | |
1108 | ||
1109 | return 0; | |
1110 | } | |
1111 | ||
3ded6d6b CB |
1112 | static int shiftfs_open(struct inode *inode, struct file *file) |
1113 | { | |
3ded6d6b | 1114 | struct file *realfile; |
3ded6d6b | 1115 | |
b4de76a8 CB |
1116 | realfile = shiftfs_open_realfile(file, inode->i_private); |
1117 | if (IS_ERR(realfile)) | |
3ded6d6b | 1118 | return PTR_ERR(realfile); |
3ded6d6b | 1119 | |
b4de76a8 | 1120 | file->private_data = realfile; |
71d7f9d9 CB |
1121 | /* For O_DIRECT dentry_open() checks f_mapping->a_ops->direct_IO. */ |
1122 | file->f_mapping = realfile->f_mapping; | |
1123 | ||
3ded6d6b CB |
1124 | return 0; |
1125 | } | |
1126 | ||
b4de76a8 | 1127 | static int shiftfs_dir_open(struct inode *inode, struct file *file) |
3ded6d6b | 1128 | { |
b4de76a8 CB |
1129 | struct file *realfile; |
1130 | const struct cred *oldcred; | |
1131 | struct dentry *lowerd = file->f_path.dentry->d_fsdata; | |
1132 | struct shiftfs_super_info *info = inode->i_sb->s_fs_info; | |
1133 | struct path realpath = { .mnt = info->mnt, .dentry = lowerd }; | |
1134 | ||
1135 | oldcred = shiftfs_override_creds(file->f_path.dentry->d_sb); | |
1136 | realfile = dentry_open(&realpath, file->f_flags | O_NOATIME, | |
1137 | info->creator_cred); | |
1138 | revert_creds(oldcred); | |
1139 | if (IS_ERR(realfile)) | |
1140 | return PTR_ERR(realfile); | |
3ded6d6b | 1141 | |
b4de76a8 | 1142 | file->private_data = realfile; |
3ded6d6b | 1143 | |
b4de76a8 CB |
1144 | return 0; |
1145 | } | |
1146 | ||
1147 | static int shiftfs_release(struct inode *inode, struct file *file) | |
1148 | { | |
1149 | struct file *realfile = file->private_data; | |
1150 | ||
1151 | if (realfile) | |
1152 | fput(realfile); | |
3ded6d6b CB |
1153 | |
1154 | return 0; | |
1155 | } | |
1156 | ||
b4de76a8 CB |
1157 | static int shiftfs_dir_release(struct inode *inode, struct file *file) |
1158 | { | |
1159 | return shiftfs_release(inode, file); | |
1160 | } | |
1161 | ||
958069e8 CB |
1162 | static loff_t shiftfs_dir_llseek(struct file *file, loff_t offset, int whence) |
1163 | { | |
b4de76a8 | 1164 | struct file *realfile = file->private_data; |
958069e8 CB |
1165 | |
1166 | return vfs_llseek(realfile, offset, whence); | |
1167 | } | |
1168 | ||
1169 | static loff_t shiftfs_file_llseek(struct file *file, loff_t offset, int whence) | |
3ded6d6b CB |
1170 | { |
1171 | struct inode *realinode = file_inode(file)->i_private; | |
1172 | ||
1173 | return generic_file_llseek_size(file, offset, whence, | |
1174 | realinode->i_sb->s_maxbytes, | |
1175 | i_size_read(realinode)); | |
1176 | } | |
1177 | ||
1178 | /* XXX: Need to figure out what to to about atime updates, maybe other | |
1179 | * timestamps too ... ref. ovl_file_accessed() */ | |
1180 | ||
1181 | static rwf_t shiftfs_iocb_to_rwf(struct kiocb *iocb) | |
1182 | { | |
1183 | int ifl = iocb->ki_flags; | |
1184 | rwf_t flags = 0; | |
1185 | ||
1186 | if (ifl & IOCB_NOWAIT) | |
1187 | flags |= RWF_NOWAIT; | |
1188 | if (ifl & IOCB_HIPRI) | |
1189 | flags |= RWF_HIPRI; | |
1190 | if (ifl & IOCB_DSYNC) | |
1191 | flags |= RWF_DSYNC; | |
1192 | if (ifl & IOCB_SYNC) | |
1193 | flags |= RWF_SYNC; | |
1194 | ||
1195 | return flags; | |
1196 | } | |
1197 | ||
392d05f6 CB |
1198 | static int shiftfs_real_fdget(const struct file *file, struct fd *lowerfd) |
1199 | { | |
1200 | struct file *realfile; | |
1201 | ||
1202 | if (file->f_op->open != shiftfs_open && | |
1203 | file->f_op->open != shiftfs_dir_open) | |
1204 | return -EINVAL; | |
1205 | ||
1206 | realfile = file->private_data; | |
1207 | lowerfd->flags = 0; | |
1208 | lowerfd->file = realfile; | |
1209 | ||
1210 | /* Did the flags change since open? */ | |
1211 | if (unlikely(file->f_flags & ~lowerfd->file->f_flags)) | |
1212 | return shiftfs_change_flags(lowerfd->file, file->f_flags); | |
1213 | ||
1214 | return 0; | |
1215 | } | |
1216 | ||
3ded6d6b CB |
1217 | static ssize_t shiftfs_read_iter(struct kiocb *iocb, struct iov_iter *iter) |
1218 | { | |
1219 | struct file *file = iocb->ki_filp; | |
1220 | struct fd lowerfd; | |
1221 | const struct cred *oldcred; | |
1222 | ssize_t ret; | |
1223 | ||
1224 | if (!iov_iter_count(iter)) | |
1225 | return 0; | |
1226 | ||
1227 | ret = shiftfs_real_fdget(file, &lowerfd); | |
1228 | if (ret) | |
1229 | return ret; | |
1230 | ||
1231 | oldcred = shiftfs_override_creds(file->f_path.dentry->d_sb); | |
1232 | ret = vfs_iter_read(lowerfd.file, iter, &iocb->ki_pos, | |
1233 | shiftfs_iocb_to_rwf(iocb)); | |
1234 | revert_creds(oldcred); | |
1235 | ||
1236 | shiftfs_file_accessed(file); | |
1237 | ||
1238 | fdput(lowerfd); | |
1239 | return ret; | |
1240 | } | |
1241 | ||
1242 | static ssize_t shiftfs_write_iter(struct kiocb *iocb, struct iov_iter *iter) | |
1243 | { | |
1244 | struct file *file = iocb->ki_filp; | |
1245 | struct inode *inode = file_inode(file); | |
1246 | struct fd lowerfd; | |
1247 | const struct cred *oldcred; | |
1248 | ssize_t ret; | |
1249 | ||
1250 | if (!iov_iter_count(iter)) | |
1251 | return 0; | |
1252 | ||
1253 | inode_lock(inode); | |
1254 | /* Update mode */ | |
1255 | shiftfs_copyattr(inode->i_private, inode); | |
1256 | ret = file_remove_privs(file); | |
1257 | if (ret) | |
1258 | goto out_unlock; | |
1259 | ||
1260 | ret = shiftfs_real_fdget(file, &lowerfd); | |
1261 | if (ret) | |
1262 | goto out_unlock; | |
1263 | ||
1264 | oldcred = shiftfs_override_creds(file->f_path.dentry->d_sb); | |
1265 | file_start_write(lowerfd.file); | |
1266 | ret = vfs_iter_write(lowerfd.file, iter, &iocb->ki_pos, | |
1267 | shiftfs_iocb_to_rwf(iocb)); | |
1268 | file_end_write(lowerfd.file); | |
1269 | revert_creds(oldcred); | |
1270 | ||
1271 | /* Update size */ | |
1272 | shiftfs_copyattr(inode->i_private, inode); | |
1273 | ||
1274 | fdput(lowerfd); | |
1275 | ||
1276 | out_unlock: | |
1277 | inode_unlock(inode); | |
1278 | return ret; | |
1279 | } | |
1280 | ||
1281 | static int shiftfs_fsync(struct file *file, loff_t start, loff_t end, | |
1282 | int datasync) | |
1283 | { | |
1284 | struct fd lowerfd; | |
1285 | const struct cred *oldcred; | |
1286 | int ret; | |
1287 | ||
1288 | ret = shiftfs_real_fdget(file, &lowerfd); | |
1289 | if (ret) | |
1290 | return ret; | |
1291 | ||
1292 | oldcred = shiftfs_override_creds(file->f_path.dentry->d_sb); | |
1293 | ret = vfs_fsync_range(lowerfd.file, start, end, datasync); | |
1294 | revert_creds(oldcred); | |
1295 | ||
1296 | fdput(lowerfd); | |
1297 | return ret; | |
1298 | } | |
1299 | ||
1300 | static int shiftfs_mmap(struct file *file, struct vm_area_struct *vma) | |
1301 | { | |
b4de76a8 | 1302 | struct file *realfile = file->private_data; |
3ded6d6b CB |
1303 | const struct cred *oldcred; |
1304 | int ret; | |
1305 | ||
1306 | if (!realfile->f_op->mmap) | |
1307 | return -ENODEV; | |
1308 | ||
1309 | if (WARN_ON(file != vma->vm_file)) | |
1310 | return -EIO; | |
1311 | ||
1312 | oldcred = shiftfs_override_creds(file->f_path.dentry->d_sb); | |
1313 | vma->vm_file = get_file(realfile); | |
1314 | ret = call_mmap(vma->vm_file, vma); | |
1315 | revert_creds(oldcred); | |
1316 | ||
1317 | shiftfs_file_accessed(file); | |
1318 | ||
3a60b1d0 SF |
1319 | if (ret) { |
1320 | /* | |
1321 | * Drop refcount from new vm_file value and restore original | |
1322 | * vm_file value | |
1323 | */ | |
1324 | vma->vm_file = file; | |
1325 | fput(realfile); | |
1326 | } else { | |
1327 | /* Drop refcount from previous vm_file value */ | |
1328 | fput(file); | |
1329 | } | |
3ded6d6b CB |
1330 | |
1331 | return ret; | |
1332 | } | |
1333 | ||
1334 | static long shiftfs_fallocate(struct file *file, int mode, loff_t offset, | |
1335 | loff_t len) | |
1336 | { | |
1337 | struct inode *inode = file_inode(file); | |
1338 | struct inode *loweri = inode->i_private; | |
1339 | struct fd lowerfd; | |
1340 | const struct cred *oldcred; | |
1341 | int ret; | |
1342 | ||
1343 | ret = shiftfs_real_fdget(file, &lowerfd); | |
1344 | if (ret) | |
1345 | return ret; | |
1346 | ||
1347 | oldcred = shiftfs_override_creds(file->f_path.dentry->d_sb); | |
1348 | ret = vfs_fallocate(lowerfd.file, mode, offset, len); | |
1349 | revert_creds(oldcred); | |
1350 | ||
1351 | /* Update size */ | |
1352 | shiftfs_copyattr(loweri, inode); | |
1353 | ||
1354 | fdput(lowerfd); | |
1355 | return ret; | |
1356 | } | |
1357 | ||
1358 | static int shiftfs_fadvise(struct file *file, loff_t offset, loff_t len, | |
1359 | int advice) | |
1360 | { | |
1361 | struct fd lowerfd; | |
1362 | const struct cred *oldcred; | |
1363 | int ret; | |
1364 | ||
1365 | ret = shiftfs_real_fdget(file, &lowerfd); | |
1366 | if (ret) | |
1367 | return ret; | |
1368 | ||
1369 | oldcred = shiftfs_override_creds(file->f_path.dentry->d_sb); | |
1370 | ret = vfs_fadvise(lowerfd.file, offset, len, advice); | |
1371 | revert_creds(oldcred); | |
1372 | ||
1373 | fdput(lowerfd); | |
1374 | return ret; | |
1375 | } | |
1376 | ||
87a7b993 | 1377 | static int shiftfs_override_ioctl_creds(int cmd, const struct super_block *sb, |
3ded6d6b CB |
1378 | const struct cred **oldcred, |
1379 | struct cred **newcred) | |
1380 | { | |
ef687aea | 1381 | struct shiftfs_super_info *sbinfo = sb->s_fs_info; |
3ded6d6b CB |
1382 | kuid_t fsuid = current_fsuid(); |
1383 | kgid_t fsgid = current_fsgid(); | |
1384 | ||
1385 | *oldcred = shiftfs_override_creds(sb); | |
1386 | ||
1387 | *newcred = prepare_creds(); | |
1388 | if (!*newcred) { | |
1389 | revert_creds(*oldcred); | |
1390 | return -ENOMEM; | |
1391 | } | |
1392 | ||
ef687aea SF |
1393 | (*newcred)->fsuid = shift_kuid(sb->s_user_ns, sbinfo->userns, fsuid); |
1394 | (*newcred)->fsgid = shift_kgid(sb->s_user_ns, sbinfo->userns, fsgid); | |
3ded6d6b CB |
1395 | |
1396 | /* clear all caps to prevent bypassing capable() checks */ | |
1397 | cap_clear((*newcred)->cap_bset); | |
1398 | cap_clear((*newcred)->cap_effective); | |
1399 | cap_clear((*newcred)->cap_inheritable); | |
1400 | cap_clear((*newcred)->cap_permitted); | |
1401 | ||
87a7b993 CB |
1402 | if (cmd == BTRFS_IOC_SNAP_DESTROY) { |
1403 | kuid_t kuid_root = make_kuid(sb->s_user_ns, 0); | |
1404 | /* | |
1405 | * Allow the root user in the container to remove subvolumes | |
1406 | * from other users. | |
1407 | */ | |
1408 | if (uid_valid(kuid_root) && uid_eq(fsuid, kuid_root)) | |
1409 | cap_raise((*newcred)->cap_effective, CAP_DAC_OVERRIDE); | |
1410 | } | |
1411 | ||
3ded6d6b CB |
1412 | put_cred(override_creds(*newcred)); |
1413 | return 0; | |
1414 | } | |
1415 | ||
1416 | static inline void shiftfs_revert_ioctl_creds(const struct cred *oldcred, | |
1417 | struct cred *newcred) | |
1418 | { | |
1419 | return shiftfs_revert_object_creds(oldcred, newcred); | |
1420 | } | |
1421 | ||
88f5bff3 CB |
1422 | static inline bool is_btrfs_snap_ioctl(int cmd) |
1423 | { | |
1424 | if ((cmd == BTRFS_IOC_SNAP_CREATE) || (cmd == BTRFS_IOC_SNAP_CREATE_V2)) | |
1425 | return true; | |
1426 | ||
1427 | return false; | |
1428 | } | |
1429 | ||
89126b92 | 1430 | static int shiftfs_btrfs_ioctl_fd_restore(int cmd, int fd, void __user *arg, |
88f5bff3 CB |
1431 | struct btrfs_ioctl_vol_args *v1, |
1432 | struct btrfs_ioctl_vol_args_v2 *v2) | |
1433 | { | |
1434 | int ret; | |
1435 | ||
1436 | if (!is_btrfs_snap_ioctl(cmd)) | |
1437 | return 0; | |
1438 | ||
1439 | if (cmd == BTRFS_IOC_SNAP_CREATE) | |
1440 | ret = copy_to_user(arg, v1, sizeof(*v1)); | |
1441 | else | |
1442 | ret = copy_to_user(arg, v2, sizeof(*v2)); | |
1443 | ||
a6e9d80a | 1444 | close_fd(fd); |
88f5bff3 CB |
1445 | kfree(v1); |
1446 | kfree(v2); | |
1447 | ||
902ca980 | 1448 | return ret ? -EFAULT: 0; |
88f5bff3 CB |
1449 | } |
1450 | ||
1451 | static int shiftfs_btrfs_ioctl_fd_replace(int cmd, void __user *arg, | |
1452 | struct btrfs_ioctl_vol_args **b1, | |
1453 | struct btrfs_ioctl_vol_args_v2 **b2, | |
88f5bff3 CB |
1454 | int *newfd) |
1455 | { | |
1456 | int oldfd, ret; | |
1457 | struct fd src; | |
89126b92 | 1458 | struct fd lfd = {}; |
88f5bff3 CB |
1459 | struct btrfs_ioctl_vol_args *v1 = NULL; |
1460 | struct btrfs_ioctl_vol_args_v2 *v2 = NULL; | |
1461 | ||
c88d1b2c SF |
1462 | *b1 = NULL; |
1463 | *b2 = NULL; | |
1464 | ||
88f5bff3 CB |
1465 | if (!is_btrfs_snap_ioctl(cmd)) |
1466 | return 0; | |
1467 | ||
1468 | if (cmd == BTRFS_IOC_SNAP_CREATE) { | |
1469 | v1 = memdup_user(arg, sizeof(*v1)); | |
1470 | if (IS_ERR(v1)) | |
1471 | return PTR_ERR(v1); | |
1472 | oldfd = v1->fd; | |
88f5bff3 CB |
1473 | } else { |
1474 | v2 = memdup_user(arg, sizeof(*v2)); | |
1475 | if (IS_ERR(v2)) | |
1476 | return PTR_ERR(v2); | |
1477 | oldfd = v2->fd; | |
88f5bff3 CB |
1478 | } |
1479 | ||
1480 | src = fdget(oldfd); | |
c88d1b2c SF |
1481 | if (!src.file) { |
1482 | ret = -EINVAL; | |
1483 | goto err_free; | |
1484 | } | |
88f5bff3 | 1485 | |
89126b92 SF |
1486 | ret = shiftfs_real_fdget(src.file, &lfd); |
1487 | if (ret) { | |
1488 | fdput(src); | |
c88d1b2c | 1489 | goto err_free; |
89126b92 SF |
1490 | } |
1491 | ||
1492 | /* | |
1493 | * shiftfs_real_fdget() does not take a reference to lfd.file, so | |
1494 | * take a reference here to offset the one which will be put by | |
a6e9d80a | 1495 | * close_fd(), and make sure that reference is put on fdput(lfd). |
89126b92 SF |
1496 | */ |
1497 | get_file(lfd.file); | |
1498 | lfd.flags |= FDPUT_FPUT; | |
1499 | fdput(src); | |
88f5bff3 | 1500 | |
89126b92 | 1501 | *newfd = get_unused_fd_flags(lfd.file->f_flags); |
88f5bff3 | 1502 | if (*newfd < 0) { |
89126b92 | 1503 | fdput(lfd); |
c88d1b2c SF |
1504 | ret = *newfd; |
1505 | goto err_free; | |
88f5bff3 CB |
1506 | } |
1507 | ||
89126b92 | 1508 | fd_install(*newfd, lfd.file); |
88f5bff3 CB |
1509 | |
1510 | if (cmd == BTRFS_IOC_SNAP_CREATE) { | |
1511 | v1->fd = *newfd; | |
1512 | ret = copy_to_user(arg, v1, sizeof(*v1)); | |
1513 | v1->fd = oldfd; | |
1514 | } else { | |
1515 | v2->fd = *newfd; | |
1516 | ret = copy_to_user(arg, v2, sizeof(*v2)); | |
1517 | v2->fd = oldfd; | |
1518 | } | |
1519 | ||
c88d1b2c SF |
1520 | if (!ret) { |
1521 | *b1 = v1; | |
1522 | *b2 = v2; | |
1523 | } else { | |
89126b92 | 1524 | shiftfs_btrfs_ioctl_fd_restore(cmd, *newfd, arg, v1, v2); |
902ca980 | 1525 | ret = -EFAULT; |
c88d1b2c SF |
1526 | } |
1527 | ||
1528 | return ret; | |
1529 | ||
1530 | err_free: | |
1531 | kfree(v1); | |
1532 | kfree(v2); | |
88f5bff3 CB |
1533 | |
1534 | return ret; | |
1535 | } | |
1536 | ||
3ded6d6b CB |
1537 | static long shiftfs_real_ioctl(struct file *file, unsigned int cmd, |
1538 | unsigned long arg) | |
1539 | { | |
3ded6d6b CB |
1540 | struct fd lowerfd; |
1541 | struct cred *newcred; | |
1542 | const struct cred *oldcred; | |
88f5bff3 CB |
1543 | int newfd = -EBADF; |
1544 | long err = 0, ret = 0; | |
1545 | void __user *argp = (void __user *)arg; | |
3ded6d6b | 1546 | struct super_block *sb = file->f_path.dentry->d_sb; |
88f5bff3 CB |
1547 | struct btrfs_ioctl_vol_args *btrfs_v1 = NULL; |
1548 | struct btrfs_ioctl_vol_args_v2 *btrfs_v2 = NULL; | |
1549 | ||
1550 | ret = shiftfs_btrfs_ioctl_fd_replace(cmd, argp, &btrfs_v1, &btrfs_v2, | |
89126b92 | 1551 | &newfd); |
88f5bff3 CB |
1552 | if (ret < 0) |
1553 | return ret; | |
3ded6d6b CB |
1554 | |
1555 | ret = shiftfs_real_fdget(file, &lowerfd); | |
1556 | if (ret) | |
88f5bff3 | 1557 | goto out_restore; |
3ded6d6b | 1558 | |
87a7b993 | 1559 | ret = shiftfs_override_ioctl_creds(cmd, sb, &oldcred, &newcred); |
3ded6d6b CB |
1560 | if (ret) |
1561 | goto out_fdput; | |
1562 | ||
1563 | ret = vfs_ioctl(lowerfd.file, cmd, arg); | |
1564 | ||
1565 | shiftfs_revert_ioctl_creds(oldcred, newcred); | |
1566 | ||
1567 | shiftfs_copyattr(file_inode(lowerfd.file), file_inode(file)); | |
1568 | shiftfs_copyflags(file_inode(lowerfd.file), file_inode(file)); | |
1569 | ||
1570 | out_fdput: | |
1571 | fdput(lowerfd); | |
1572 | ||
88f5bff3 | 1573 | out_restore: |
89126b92 | 1574 | err = shiftfs_btrfs_ioctl_fd_restore(cmd, newfd, argp, |
88f5bff3 CB |
1575 | btrfs_v1, btrfs_v2); |
1576 | if (!ret) | |
1577 | ret = err; | |
1578 | ||
3ded6d6b CB |
1579 | return ret; |
1580 | } | |
1581 | ||
4a60080e | 1582 | static bool in_ioctl_whitelist(int flag, unsigned long arg) |
88f5bff3 | 1583 | { |
4a60080e CB |
1584 | void __user *argp = (void __user *)arg; |
1585 | u64 flags = 0; | |
1586 | ||
88f5bff3 | 1587 | switch (flag) { |
4a60080e CB |
1588 | case BTRFS_IOC_FS_INFO: |
1589 | return true; | |
88f5bff3 CB |
1590 | case BTRFS_IOC_SNAP_CREATE: |
1591 | return true; | |
1592 | case BTRFS_IOC_SNAP_CREATE_V2: | |
1593 | return true; | |
1594 | case BTRFS_IOC_SUBVOL_CREATE: | |
1595 | return true; | |
1596 | case BTRFS_IOC_SUBVOL_CREATE_V2: | |
4a60080e CB |
1597 | return true; |
1598 | case BTRFS_IOC_SUBVOL_GETFLAGS: | |
1599 | return true; | |
1600 | case BTRFS_IOC_SUBVOL_SETFLAGS: | |
26d1a2e4 | 1601 | if (copy_from_user(&flags, argp, sizeof(flags))) |
4a60080e CB |
1602 | return false; |
1603 | ||
1604 | if (flags & ~BTRFS_SUBVOL_RDONLY) | |
1605 | return false; | |
1606 | ||
88f5bff3 CB |
1607 | return true; |
1608 | case BTRFS_IOC_SNAP_DESTROY: | |
1609 | return true; | |
1610 | } | |
1611 | ||
1612 | return false; | |
1613 | } | |
1614 | ||
3ded6d6b CB |
1615 | static long shiftfs_ioctl(struct file *file, unsigned int cmd, |
1616 | unsigned long arg) | |
1617 | { | |
1618 | switch (cmd) { | |
1619 | case FS_IOC_GETVERSION: | |
1620 | /* fall through */ | |
1621 | case FS_IOC_GETFLAGS: | |
1622 | /* fall through */ | |
1623 | case FS_IOC_SETFLAGS: | |
1624 | break; | |
1625 | default: | |
4a60080e | 1626 | if (!in_ioctl_whitelist(cmd, arg) || |
88f5bff3 CB |
1627 | !shiftfs_passthrough_ioctls(file->f_path.dentry->d_sb->s_fs_info)) |
1628 | return -ENOTTY; | |
3ded6d6b CB |
1629 | } |
1630 | ||
1631 | return shiftfs_real_ioctl(file, cmd, arg); | |
1632 | } | |
1633 | ||
1634 | static long shiftfs_compat_ioctl(struct file *file, unsigned int cmd, | |
1635 | unsigned long arg) | |
1636 | { | |
1637 | switch (cmd) { | |
1638 | case FS_IOC32_GETVERSION: | |
1639 | /* fall through */ | |
1640 | case FS_IOC32_GETFLAGS: | |
1641 | /* fall through */ | |
1642 | case FS_IOC32_SETFLAGS: | |
1643 | break; | |
1644 | default: | |
4a60080e | 1645 | if (!in_ioctl_whitelist(cmd, arg) || |
88f5bff3 CB |
1646 | !shiftfs_passthrough_ioctls(file->f_path.dentry->d_sb->s_fs_info)) |
1647 | return -ENOIOCTLCMD; | |
3ded6d6b CB |
1648 | } |
1649 | ||
1650 | return shiftfs_real_ioctl(file, cmd, arg); | |
1651 | } | |
1652 | ||
1653 | enum shiftfs_copyop { | |
1654 | SHIFTFS_COPY, | |
1655 | SHIFTFS_CLONE, | |
1656 | SHIFTFS_DEDUPE, | |
1657 | }; | |
1658 | ||
1659 | static ssize_t shiftfs_copyfile(struct file *file_in, loff_t pos_in, | |
1660 | struct file *file_out, loff_t pos_out, u64 len, | |
1661 | unsigned int flags, enum shiftfs_copyop op) | |
1662 | { | |
1663 | ssize_t ret; | |
1664 | struct fd real_in, real_out; | |
1665 | const struct cred *oldcred; | |
1666 | struct inode *inode_out = file_inode(file_out); | |
1667 | struct inode *loweri = inode_out->i_private; | |
1668 | ||
1669 | ret = shiftfs_real_fdget(file_out, &real_out); | |
1670 | if (ret) | |
1671 | return ret; | |
1672 | ||
1673 | ret = shiftfs_real_fdget(file_in, &real_in); | |
1674 | if (ret) { | |
1675 | fdput(real_out); | |
1676 | return ret; | |
1677 | } | |
1678 | ||
1679 | oldcred = shiftfs_override_creds(inode_out->i_sb); | |
1680 | switch (op) { | |
1681 | case SHIFTFS_COPY: | |
1682 | ret = vfs_copy_file_range(real_in.file, pos_in, real_out.file, | |
1683 | pos_out, len, flags); | |
1684 | break; | |
1685 | ||
1686 | case SHIFTFS_CLONE: | |
1687 | ret = vfs_clone_file_range(real_in.file, pos_in, real_out.file, | |
1688 | pos_out, len, flags); | |
1689 | break; | |
1690 | ||
1691 | case SHIFTFS_DEDUPE: | |
1692 | ret = vfs_dedupe_file_range_one(real_in.file, pos_in, | |
1693 | real_out.file, pos_out, len, | |
1694 | flags); | |
1695 | break; | |
1696 | } | |
1697 | revert_creds(oldcred); | |
1698 | ||
1699 | /* Update size */ | |
1700 | shiftfs_copyattr(loweri, inode_out); | |
1701 | ||
1702 | fdput(real_in); | |
1703 | fdput(real_out); | |
1704 | ||
1705 | return ret; | |
1706 | } | |
1707 | ||
1708 | static ssize_t shiftfs_copy_file_range(struct file *file_in, loff_t pos_in, | |
1709 | struct file *file_out, loff_t pos_out, | |
1710 | size_t len, unsigned int flags) | |
1711 | { | |
1712 | return shiftfs_copyfile(file_in, pos_in, file_out, pos_out, len, flags, | |
1713 | SHIFTFS_COPY); | |
1714 | } | |
1715 | ||
1716 | static loff_t shiftfs_remap_file_range(struct file *file_in, loff_t pos_in, | |
1717 | struct file *file_out, loff_t pos_out, | |
1718 | loff_t len, unsigned int remap_flags) | |
1719 | { | |
1720 | enum shiftfs_copyop op; | |
1721 | ||
1722 | if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY)) | |
1723 | return -EINVAL; | |
1724 | ||
1725 | if (remap_flags & REMAP_FILE_DEDUP) | |
1726 | op = SHIFTFS_DEDUPE; | |
1727 | else | |
1728 | op = SHIFTFS_CLONE; | |
1729 | ||
1730 | return shiftfs_copyfile(file_in, pos_in, file_out, pos_out, len, | |
1731 | remap_flags, op); | |
1732 | } | |
1733 | ||
1734 | static int shiftfs_iterate_shared(struct file *file, struct dir_context *ctx) | |
1735 | { | |
1736 | const struct cred *oldcred; | |
1737 | int err = -ENOTDIR; | |
b4de76a8 | 1738 | struct file *realfile = file->private_data; |
3ded6d6b CB |
1739 | |
1740 | oldcred = shiftfs_override_creds(file->f_path.dentry->d_sb); | |
1741 | err = iterate_dir(realfile, ctx); | |
1742 | revert_creds(oldcred); | |
1743 | ||
1744 | return err; | |
1745 | } | |
1746 | ||
1747 | const struct file_operations shiftfs_file_operations = { | |
1748 | .open = shiftfs_open, | |
1749 | .release = shiftfs_release, | |
958069e8 | 1750 | .llseek = shiftfs_file_llseek, |
3ded6d6b CB |
1751 | .read_iter = shiftfs_read_iter, |
1752 | .write_iter = shiftfs_write_iter, | |
1753 | .fsync = shiftfs_fsync, | |
1754 | .mmap = shiftfs_mmap, | |
1755 | .fallocate = shiftfs_fallocate, | |
1756 | .fadvise = shiftfs_fadvise, | |
1757 | .unlocked_ioctl = shiftfs_ioctl, | |
1758 | .compat_ioctl = shiftfs_compat_ioctl, | |
1759 | .copy_file_range = shiftfs_copy_file_range, | |
1760 | .remap_file_range = shiftfs_remap_file_range, | |
37c48396 CB |
1761 | .splice_read = generic_file_splice_read, |
1762 | .splice_write = iter_file_splice_write, | |
3ded6d6b CB |
1763 | }; |
1764 | ||
1765 | const struct file_operations shiftfs_dir_operations = { | |
b4de76a8 CB |
1766 | .open = shiftfs_dir_open, |
1767 | .release = shiftfs_dir_release, | |
3ded6d6b CB |
1768 | .compat_ioctl = shiftfs_compat_ioctl, |
1769 | .fsync = shiftfs_fsync, | |
1770 | .iterate_shared = shiftfs_iterate_shared, | |
958069e8 | 1771 | .llseek = shiftfs_dir_llseek, |
3ded6d6b | 1772 | .read = generic_read_dir, |
3ded6d6b CB |
1773 | .unlocked_ioctl = shiftfs_ioctl, |
1774 | }; | |
1775 | ||
1776 | static const struct address_space_operations shiftfs_aops = { | |
1777 | /* For O_DIRECT dentry_open() checks f_mapping->a_ops->direct_IO */ | |
1778 | .direct_IO = noop_direct_IO, | |
1779 | }; | |
1780 | ||
1781 | static void shiftfs_fill_inode(struct inode *inode, unsigned long ino, | |
1782 | umode_t mode, dev_t dev, struct dentry *dentry) | |
1783 | { | |
1784 | struct inode *loweri; | |
1785 | ||
1786 | inode->i_ino = ino; | |
1787 | inode->i_flags |= S_NOCMTIME; | |
1788 | ||
1789 | mode &= S_IFMT; | |
1790 | inode->i_mode = mode; | |
1791 | switch (mode & S_IFMT) { | |
1792 | case S_IFDIR: | |
1793 | inode->i_op = &shiftfs_dir_inode_operations; | |
1794 | inode->i_fop = &shiftfs_dir_operations; | |
1795 | break; | |
1796 | case S_IFLNK: | |
1797 | inode->i_op = &shiftfs_symlink_inode_operations; | |
1798 | break; | |
1799 | case S_IFREG: | |
1800 | inode->i_op = &shiftfs_file_inode_operations; | |
1801 | inode->i_fop = &shiftfs_file_operations; | |
1802 | inode->i_mapping->a_ops = &shiftfs_aops; | |
1803 | break; | |
1804 | default: | |
1805 | inode->i_op = &shiftfs_special_inode_operations; | |
1806 | init_special_inode(inode, mode, dev); | |
1807 | break; | |
1808 | } | |
1809 | ||
1810 | if (!dentry) | |
1811 | return; | |
1812 | ||
1813 | loweri = dentry->d_inode; | |
1814 | if (!loweri->i_op->get_link) | |
1815 | inode->i_opflags |= IOP_NOFOLLOW; | |
1816 | ||
1817 | shiftfs_copyattr(loweri, inode); | |
1818 | shiftfs_copyflags(loweri, inode); | |
1819 | set_nlink(inode, loweri->i_nlink); | |
1820 | } | |
1821 | ||
1822 | static int shiftfs_show_options(struct seq_file *m, struct dentry *dentry) | |
1823 | { | |
1824 | struct super_block *sb = dentry->d_sb; | |
1825 | struct shiftfs_super_info *sbinfo = sb->s_fs_info; | |
1826 | ||
1827 | if (sbinfo->mark) | |
1828 | seq_show_option(m, "mark", NULL); | |
1829 | ||
1830 | if (sbinfo->passthrough) | |
1831 | seq_printf(m, ",passthrough=%u", sbinfo->passthrough); | |
1832 | ||
1833 | return 0; | |
1834 | } | |
1835 | ||
1836 | static int shiftfs_statfs(struct dentry *dentry, struct kstatfs *buf) | |
1837 | { | |
1838 | struct super_block *sb = dentry->d_sb; | |
1839 | struct shiftfs_super_info *sbinfo = sb->s_fs_info; | |
1840 | struct dentry *root = sb->s_root; | |
1841 | struct dentry *realroot = root->d_fsdata; | |
1842 | struct path realpath = { .mnt = sbinfo->mnt, .dentry = realroot }; | |
1843 | int err; | |
1844 | ||
1845 | err = vfs_statfs(&realpath, buf); | |
aa269008 | 1846 | if (err) |
3ded6d6b | 1847 | return err; |
aa269008 | 1848 | |
3ded6d6b CB |
1849 | if (!shiftfs_passthrough_statfs(sbinfo)) |
1850 | buf->f_type = sb->s_magic; | |
aa269008 | 1851 | |
3ded6d6b CB |
1852 | return 0; |
1853 | } | |
aa269008 | 1854 | |
3ded6d6b CB |
1855 | static void shiftfs_evict_inode(struct inode *inode) |
1856 | { | |
1857 | struct inode *loweri = inode->i_private; | |
1858 | ||
1859 | clear_inode(inode); | |
1860 | ||
1861 | if (loweri) | |
1862 | iput(loweri); | |
1863 | } | |
1864 | ||
1865 | static void shiftfs_put_super(struct super_block *sb) | |
1866 | { | |
1867 | struct shiftfs_super_info *sbinfo = sb->s_fs_info; | |
1868 | ||
1869 | if (sbinfo) { | |
1870 | mntput(sbinfo->mnt); | |
1871 | put_cred(sbinfo->creator_cred); | |
1872 | kfree(sbinfo); | |
1873 | } | |
1874 | } | |
1875 | ||
1876 | static const struct xattr_handler shiftfs_xattr_handler = { | |
1877 | .prefix = "", | |
1878 | .get = shiftfs_xattr_get, | |
1879 | .set = shiftfs_xattr_set, | |
1880 | }; | |
1881 | ||
1882 | const struct xattr_handler *shiftfs_xattr_handlers[] = { | |
1883 | #ifdef CONFIG_SHIFT_FS_POSIX_ACL | |
1884 | &shiftfs_posix_acl_access_xattr_handler, | |
1885 | &shiftfs_posix_acl_default_xattr_handler, | |
1886 | #endif | |
1887 | &shiftfs_xattr_handler, | |
1888 | NULL | |
1889 | }; | |
1890 | ||
1891 | static inline bool passthrough_is_subset(int old_flags, int new_flags) | |
1892 | { | |
1893 | if ((new_flags & old_flags) != new_flags) | |
1894 | return false; | |
1895 | ||
1896 | return true; | |
1897 | } | |
1898 | ||
40b373d6 CB |
1899 | static int shiftfs_super_check_flags(unsigned long old_flags, |
1900 | unsigned long new_flags) | |
1901 | { | |
1902 | if ((old_flags & SB_RDONLY) && !(new_flags & SB_RDONLY)) | |
1903 | return -EPERM; | |
1904 | ||
1905 | if ((old_flags & SB_NOSUID) && !(new_flags & SB_NOSUID)) | |
1906 | return -EPERM; | |
1907 | ||
1908 | if ((old_flags & SB_NODEV) && !(new_flags & SB_NODEV)) | |
1909 | return -EPERM; | |
1910 | ||
1911 | if ((old_flags & SB_NOEXEC) && !(new_flags & SB_NOEXEC)) | |
1912 | return -EPERM; | |
1913 | ||
1914 | if ((old_flags & SB_NOATIME) && !(new_flags & SB_NOATIME)) | |
1915 | return -EPERM; | |
1916 | ||
1917 | if ((old_flags & SB_NODIRATIME) && !(new_flags & SB_NODIRATIME)) | |
1918 | return -EPERM; | |
1919 | ||
1920 | if (!(old_flags & SB_POSIXACL) && (new_flags & SB_POSIXACL)) | |
1921 | return -EPERM; | |
1922 | ||
1923 | return 0; | |
1924 | } | |
1925 | ||
3ded6d6b CB |
1926 | static int shiftfs_remount(struct super_block *sb, int *flags, char *data) |
1927 | { | |
1928 | int err; | |
1929 | struct shiftfs_super_info new = {}; | |
1930 | struct shiftfs_super_info *info = sb->s_fs_info; | |
1931 | ||
1932 | err = shiftfs_parse_mount_options(&new, data); | |
aa269008 | 1933 | if (err) |
3ded6d6b CB |
1934 | return err; |
1935 | ||
40b373d6 CB |
1936 | err = shiftfs_super_check_flags(sb->s_flags, *flags); |
1937 | if (err) | |
1938 | return err; | |
1939 | ||
3ded6d6b CB |
1940 | /* Mark mount option cannot be changed. */ |
1941 | if (info->mark || (info->mark != new.mark)) | |
1942 | return -EPERM; | |
1943 | ||
1944 | if (info->passthrough != new.passthrough) { | |
1945 | /* Don't allow exceeding passthrough options of mark mount. */ | |
14ba9a56 | 1946 | if (!passthrough_is_subset(info->passthrough_mark, |
3ded6d6b CB |
1947 | info->passthrough)) |
1948 | return -EPERM; | |
1949 | ||
1950 | info->passthrough = new.passthrough; | |
1951 | } | |
1952 | ||
1953 | return 0; | |
1954 | } | |
aa269008 | 1955 | |
3ded6d6b CB |
1956 | static const struct super_operations shiftfs_super_ops = { |
1957 | .put_super = shiftfs_put_super, | |
1958 | .show_options = shiftfs_show_options, | |
1959 | .statfs = shiftfs_statfs, | |
1960 | .remount_fs = shiftfs_remount, | |
1961 | .evict_inode = shiftfs_evict_inode, | |
1962 | }; | |
1963 | ||
1964 | struct shiftfs_data { | |
1965 | void *data; | |
1966 | const char *path; | |
1967 | }; | |
1968 | ||
40b373d6 CB |
1969 | static void shiftfs_super_force_flags(struct super_block *sb, |
1970 | unsigned long lower_flags) | |
1971 | { | |
1972 | sb->s_flags |= lower_flags & (SB_RDONLY | SB_NOSUID | SB_NODEV | | |
1973 | SB_NOEXEC | SB_NOATIME | SB_NODIRATIME); | |
1974 | ||
1975 | if (!(lower_flags & SB_POSIXACL)) | |
1976 | sb->s_flags &= ~SB_POSIXACL; | |
1977 | } | |
1978 | ||
3ded6d6b CB |
1979 | static int shiftfs_fill_super(struct super_block *sb, void *raw_data, |
1980 | int silent) | |
1981 | { | |
1982 | int err; | |
1983 | struct path path = {}; | |
1984 | struct shiftfs_super_info *sbinfo_mp; | |
1985 | char *name = NULL; | |
1986 | struct inode *inode = NULL; | |
1987 | struct dentry *dentry = NULL; | |
1988 | struct shiftfs_data *data = raw_data; | |
1989 | struct shiftfs_super_info *sbinfo = NULL; | |
1990 | ||
1991 | if (!data->path) | |
1992 | return -EINVAL; | |
1993 | ||
1994 | sb->s_fs_info = kzalloc(sizeof(*sbinfo), GFP_KERNEL); | |
1995 | if (!sb->s_fs_info) | |
1996 | return -ENOMEM; | |
1997 | sbinfo = sb->s_fs_info; | |
1998 | ||
1999 | err = shiftfs_parse_mount_options(sbinfo, data->data); | |
2000 | if (err) | |
2001 | return err; | |
2002 | ||
2003 | /* to mount a mark, must be userns admin */ | |
2004 | if (!sbinfo->mark && !ns_capable(current_user_ns(), CAP_SYS_ADMIN)) | |
2005 | return -EPERM; | |
2006 | ||
2007 | name = kstrdup(data->path, GFP_KERNEL); | |
2008 | if (!name) | |
2009 | return -ENOMEM; | |
2010 | ||
2011 | err = kern_path(name, LOOKUP_FOLLOW, &path); | |
2012 | if (err) | |
2013 | goto out_free_name; | |
aa269008 JB |
2014 | |
2015 | if (!S_ISDIR(path.dentry->d_inode->i_mode)) { | |
2016 | err = -ENOTDIR; | |
3ded6d6b | 2017 | goto out_put_path; |
aa269008 JB |
2018 | } |
2019 | ||
40b373d6 CB |
2020 | sb->s_flags |= SB_POSIXACL; |
2021 | ||
3ded6d6b | 2022 | if (sbinfo->mark) { |
9f4c882a | 2023 | struct cred *cred_tmp; |
3ded6d6b CB |
2024 | struct super_block *lower_sb = path.mnt->mnt_sb; |
2025 | ||
2026 | /* to mark a mount point, must root wrt lower s_user_ns */ | |
2027 | if (!ns_capable(lower_sb->s_user_ns, CAP_SYS_ADMIN)) { | |
2028 | err = -EPERM; | |
2029 | goto out_put_path; | |
2030 | } | |
aa269008 | 2031 | |
aa269008 JB |
2032 | /* |
2033 | * this part is visible unshifted, so make sure no | |
2034 | * executables that could be used to give suid | |
2035 | * privileges | |
2036 | */ | |
2037 | sb->s_iflags = SB_I_NOEXEC; | |
aa269008 | 2038 | |
40b373d6 CB |
2039 | shiftfs_super_force_flags(sb, lower_sb->s_flags); |
2040 | ||
aa269008 | 2041 | /* |
3ded6d6b CB |
2042 | * Handle nesting of shiftfs mounts by referring this mark |
2043 | * mount back to the original mark mount. This is more | |
2044 | * efficient and alleviates concerns about stack depth. | |
aa269008 | 2045 | */ |
3ded6d6b CB |
2046 | if (lower_sb->s_magic == SHIFTFS_MAGIC) { |
2047 | sbinfo_mp = lower_sb->s_fs_info; | |
2048 | ||
2049 | /* Doesn't make sense to mark a mark mount */ | |
2050 | if (sbinfo_mp->mark) { | |
2051 | err = -EINVAL; | |
2052 | goto out_put_path; | |
2053 | } | |
2054 | ||
2055 | if (!passthrough_is_subset(sbinfo_mp->passthrough, | |
2056 | sbinfo->passthrough)) { | |
2057 | err = -EPERM; | |
2058 | goto out_put_path; | |
2059 | } | |
2060 | ||
2061 | sbinfo->mnt = mntget(sbinfo_mp->mnt); | |
2062 | dentry = dget(path.dentry->d_fsdata); | |
14ba9a56 CB |
2063 | /* |
2064 | * Copy up the passthrough mount options from the | |
2065 | * parent mark mountpoint. | |
2066 | */ | |
2067 | sbinfo->passthrough_mark = sbinfo_mp->passthrough_mark; | |
b91d316e | 2068 | sbinfo->creator_cred = get_cred(sbinfo_mp->creator_cred); |
3ded6d6b CB |
2069 | } else { |
2070 | sbinfo->mnt = mntget(path.mnt); | |
2071 | dentry = dget(path.dentry); | |
14ba9a56 CB |
2072 | /* |
2073 | * For a new mark passthrough_mark and passthrough | |
2074 | * are identical. | |
2075 | */ | |
2076 | sbinfo->passthrough_mark = sbinfo->passthrough; | |
3ded6d6b | 2077 | |
b91d316e CB |
2078 | cred_tmp = prepare_creds(); |
2079 | if (!cred_tmp) { | |
2080 | err = -ENOMEM; | |
2081 | goto out_put_path; | |
2082 | } | |
2083 | /* Don't override disk quota limits or use reserved space. */ | |
2084 | cap_lower(cred_tmp->cap_effective, CAP_SYS_RESOURCE); | |
2085 | sbinfo->creator_cred = cred_tmp; | |
3ded6d6b CB |
2086 | } |
2087 | } else { | |
2088 | /* | |
2089 | * This leg executes if we're admin capable in the namespace, | |
2090 | * so be very careful. | |
2091 | */ | |
2092 | err = -EPERM; | |
aa269008 | 2093 | if (path.dentry->d_sb->s_magic != SHIFTFS_MAGIC) |
3ded6d6b CB |
2094 | goto out_put_path; |
2095 | ||
2096 | sbinfo_mp = path.dentry->d_sb->s_fs_info; | |
2097 | if (!sbinfo_mp->mark) | |
2098 | goto out_put_path; | |
2099 | ||
2100 | if (!passthrough_is_subset(sbinfo_mp->passthrough, | |
2101 | sbinfo->passthrough)) | |
2102 | goto out_put_path; | |
2103 | ||
2104 | sbinfo->mnt = mntget(sbinfo_mp->mnt); | |
2105 | sbinfo->creator_cred = get_cred(sbinfo_mp->creator_cred); | |
aa269008 | 2106 | dentry = dget(path.dentry->d_fsdata); |
14ba9a56 CB |
2107 | /* |
2108 | * Copy up passthrough settings from mark mountpoint so we can | |
2109 | * verify when the overlay wants to remount with different | |
2110 | * passthrough settings. | |
2111 | */ | |
2112 | sbinfo->passthrough_mark = sbinfo_mp->passthrough; | |
40b373d6 | 2113 | shiftfs_super_force_flags(sb, path.mnt->mnt_sb->s_flags); |
3ded6d6b CB |
2114 | } |
2115 | ||
2116 | sb->s_stack_depth = dentry->d_sb->s_stack_depth + 1; | |
2117 | if (sb->s_stack_depth > FILESYSTEM_MAX_STACK_DEPTH) { | |
2118 | printk(KERN_ERR "shiftfs: maximum stacking depth exceeded\n"); | |
2119 | err = -EINVAL; | |
2120 | goto out_put_path; | |
2121 | } | |
2122 | ||
2123 | inode = new_inode(sb); | |
2124 | if (!inode) { | |
2125 | err = -ENOMEM; | |
2126 | goto out_put_path; | |
aa269008 | 2127 | } |
3ded6d6b CB |
2128 | shiftfs_fill_inode(inode, dentry->d_inode->i_ino, S_IFDIR, 0, dentry); |
2129 | ||
2130 | ihold(dentry->d_inode); | |
2131 | inode->i_private = dentry->d_inode; | |
2132 | ||
aa269008 | 2133 | sb->s_magic = SHIFTFS_MAGIC; |
ed1034ae | 2134 | sb->s_maxbytes = MAX_LFS_FILESIZE; |
aa269008 JB |
2135 | sb->s_op = &shiftfs_super_ops; |
2136 | sb->s_xattr = shiftfs_xattr_handlers; | |
2137 | sb->s_d_op = &shiftfs_dentry_ops; | |
3ded6d6b CB |
2138 | sb->s_root = d_make_root(inode); |
2139 | if (!sb->s_root) { | |
2140 | err = -ENOMEM; | |
2141 | goto out_put_path; | |
2142 | } | |
2143 | ||
aa269008 | 2144 | sb->s_root->d_fsdata = dentry; |
3ded6d6b CB |
2145 | sbinfo->userns = get_user_ns(dentry->d_sb->s_user_ns); |
2146 | shiftfs_copyattr(dentry->d_inode, sb->s_root->d_inode); | |
aa269008 | 2147 | |
3ded6d6b CB |
2148 | dentry = NULL; |
2149 | err = 0; | |
aa269008 | 2150 | |
3ded6d6b | 2151 | out_put_path: |
aa269008 | 2152 | path_put(&path); |
3ded6d6b CB |
2153 | |
2154 | out_free_name: | |
aa269008 | 2155 | kfree(name); |
3ded6d6b CB |
2156 | |
2157 | dput(dentry); | |
2158 | ||
aa269008 JB |
2159 | return err; |
2160 | } | |
2161 | ||
2162 | static struct dentry *shiftfs_mount(struct file_system_type *fs_type, | |
2163 | int flags, const char *dev_name, void *data) | |
2164 | { | |
2165 | struct shiftfs_data d = { data, dev_name }; | |
2166 | ||
2167 | return mount_nodev(fs_type, flags, &d, shiftfs_fill_super); | |
2168 | } | |
2169 | ||
2170 | static struct file_system_type shiftfs_type = { | |
2171 | .owner = THIS_MODULE, | |
2172 | .name = "shiftfs", | |
2173 | .mount = shiftfs_mount, | |
2174 | .kill_sb = kill_anon_super, | |
2175 | .fs_flags = FS_USERNS_MOUNT, | |
2176 | }; | |
2177 | ||
2178 | static int __init shiftfs_init(void) | |
2179 | { | |
2180 | return register_filesystem(&shiftfs_type); | |
2181 | } | |
2182 | ||
2183 | static void __exit shiftfs_exit(void) | |
2184 | { | |
2185 | unregister_filesystem(&shiftfs_type); | |
2186 | } | |
2187 | ||
2188 | MODULE_ALIAS_FS("shiftfs"); | |
2189 | MODULE_AUTHOR("James Bottomley"); | |
3ded6d6b CB |
2190 | MODULE_AUTHOR("Seth Forshee <seth.forshee@canonical.com>"); |
2191 | MODULE_AUTHOR("Christian Brauner <christian.brauner@ubuntu.com>"); | |
2192 | MODULE_DESCRIPTION("id shifting filesystem"); | |
aa269008 JB |
2193 | MODULE_LICENSE("GPL v2"); |
2194 | module_init(shiftfs_init) | |
2195 | module_exit(shiftfs_exit) |