xref: /linux/fs/exportfs/expfs.c (revision f2c6dbd220170c2396fb019ead67fbada1e23ebd)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) Neil Brown 2002
4  * Copyright (C) Christoph Hellwig 2007
5  *
6  * This file contains the code mapping from inodes to NFS file handles,
7  * and for mapping back from file handles to dentries.
8  *
9  * For details on why we do all the strange and hairy things in here
10  * take a look at Documentation/filesystems/nfs/exporting.rst.
11  */
12 #include <linux/exportfs.h>
13 #include <linux/fs.h>
14 #include <linux/file.h>
15 #include <linux/module.h>
16 #include <linux/mount.h>
17 #include <linux/namei.h>
18 #include <linux/sched.h>
19 #include <linux/cred.h>
20 
21 #define dprintk(fmt, args...) pr_debug(fmt, ##args)
22 
23 
24 static int get_name(const struct path *path, char *name, struct dentry *child);
25 
26 
27 static int exportfs_get_name(struct vfsmount *mnt, struct dentry *dir,
28 		char *name, struct dentry *child)
29 {
30 	const struct export_operations *nop = dir->d_sb->s_export_op;
31 	struct path path = {.mnt = mnt, .dentry = dir};
32 
33 	if (nop->get_name)
34 		return nop->get_name(dir, name, child);
35 	else
36 		return get_name(&path, name, child);
37 }
38 
39 /*
40  * Check if the dentry or any of it's aliases is acceptable.
41  */
42 static struct dentry *
43 find_acceptable_alias(struct dentry *result,
44 		int (*acceptable)(void *context, struct dentry *dentry),
45 		void *context)
46 {
47 	struct dentry *dentry, *toput = NULL;
48 	struct inode *inode;
49 
50 	if (acceptable(context, result))
51 		return result;
52 
53 	inode = result->d_inode;
54 	spin_lock(&inode->i_lock);
55 	hlist_for_each_entry(dentry, &inode->i_dentry, d_u.d_alias) {
56 		dget(dentry);
57 		spin_unlock(&inode->i_lock);
58 		if (toput)
59 			dput(toput);
60 		if (dentry != result && acceptable(context, dentry)) {
61 			dput(result);
62 			return dentry;
63 		}
64 		spin_lock(&inode->i_lock);
65 		toput = dentry;
66 	}
67 	spin_unlock(&inode->i_lock);
68 
69 	if (toput)
70 		dput(toput);
71 	return NULL;
72 }
73 
74 static bool dentry_connected(struct dentry *dentry)
75 {
76 	dget(dentry);
77 	while (dentry->d_flags & DCACHE_DISCONNECTED) {
78 		struct dentry *parent = dget_parent(dentry);
79 
80 		dput(dentry);
81 		if (dentry == parent) {
82 			dput(parent);
83 			return false;
84 		}
85 		dentry = parent;
86 	}
87 	dput(dentry);
88 	return true;
89 }
90 
91 static void clear_disconnected(struct dentry *dentry)
92 {
93 	dget(dentry);
94 	while (dentry->d_flags & DCACHE_DISCONNECTED) {
95 		struct dentry *parent = dget_parent(dentry);
96 
97 		WARN_ON_ONCE(IS_ROOT(dentry));
98 
99 		spin_lock(&dentry->d_lock);
100 		dentry->d_flags &= ~DCACHE_DISCONNECTED;
101 		spin_unlock(&dentry->d_lock);
102 
103 		dput(dentry);
104 		dentry = parent;
105 	}
106 	dput(dentry);
107 }
108 
109 /*
110  * Reconnect a directory dentry with its parent.
111  *
112  * This can return a dentry, or NULL, or an error.
113  *
114  * In the first case the returned dentry is the parent of the given
115  * dentry, and may itself need to be reconnected to its parent.
116  *
117  * In the NULL case, a concurrent VFS operation has either renamed or
118  * removed this directory.  The concurrent operation has reconnected our
119  * dentry, so we no longer need to.
120  */
121 static struct dentry *reconnect_one(struct vfsmount *mnt,
122 		struct dentry *dentry, char *nbuf)
123 {
124 	struct dentry *parent;
125 	struct dentry *tmp;
126 	int err;
127 
128 	parent = ERR_PTR(-EACCES);
129 	inode_lock(dentry->d_inode);
130 	if (mnt->mnt_sb->s_export_op->get_parent)
131 		parent = mnt->mnt_sb->s_export_op->get_parent(dentry);
132 	inode_unlock(dentry->d_inode);
133 
134 	if (IS_ERR(parent)) {
135 		dprintk("get_parent of %lu failed, err %ld\n",
136 			dentry->d_inode->i_ino, PTR_ERR(parent));
137 		return parent;
138 	}
139 
140 	dprintk("%s: find name of %lu in %lu\n", __func__,
141 		dentry->d_inode->i_ino, parent->d_inode->i_ino);
142 	err = exportfs_get_name(mnt, parent, nbuf, dentry);
143 	if (err == -ENOENT)
144 		goto out_reconnected;
145 	if (err)
146 		goto out_err;
147 	dprintk("%s: found name: %s\n", __func__, nbuf);
148 	tmp = lookup_one_unlocked(mnt_idmap(mnt), nbuf, parent, strlen(nbuf));
149 	if (IS_ERR(tmp)) {
150 		dprintk("lookup failed: %ld\n", PTR_ERR(tmp));
151 		err = PTR_ERR(tmp);
152 		goto out_err;
153 	}
154 	if (tmp != dentry) {
155 		/*
156 		 * Somebody has renamed it since exportfs_get_name();
157 		 * great, since it could've only been renamed if it
158 		 * got looked up and thus connected, and it would
159 		 * remain connected afterwards.  We are done.
160 		 */
161 		dput(tmp);
162 		goto out_reconnected;
163 	}
164 	dput(tmp);
165 	if (IS_ROOT(dentry)) {
166 		err = -ESTALE;
167 		goto out_err;
168 	}
169 	return parent;
170 
171 out_err:
172 	dput(parent);
173 	return ERR_PTR(err);
174 out_reconnected:
175 	dput(parent);
176 	/*
177 	 * Someone must have renamed our entry into another parent, in
178 	 * which case it has been reconnected by the rename.
179 	 *
180 	 * Or someone removed it entirely, in which case filehandle
181 	 * lookup will succeed but the directory is now IS_DEAD and
182 	 * subsequent operations on it will fail.
183 	 *
184 	 * Alternatively, maybe there was no race at all, and the
185 	 * filesystem is just corrupt and gave us a parent that doesn't
186 	 * actually contain any entry pointing to this inode.  So,
187 	 * double check that this worked and return -ESTALE if not:
188 	 */
189 	if (!dentry_connected(dentry))
190 		return ERR_PTR(-ESTALE);
191 	return NULL;
192 }
193 
194 /*
195  * Make sure target_dir is fully connected to the dentry tree.
196  *
197  * On successful return, DCACHE_DISCONNECTED will be cleared on
198  * target_dir, and target_dir->d_parent->...->d_parent will reach the
199  * root of the filesystem.
200  *
201  * Whenever DCACHE_DISCONNECTED is unset, target_dir is fully connected.
202  * But the converse is not true: target_dir may have DCACHE_DISCONNECTED
203  * set but already be connected.  In that case we'll verify the
204  * connection to root and then clear the flag.
205  *
206  * Note that target_dir could be removed by a concurrent operation.  In
207  * that case reconnect_path may still succeed with target_dir fully
208  * connected, but further operations using the filehandle will fail when
209  * necessary (due to S_DEAD being set on the directory).
210  */
211 static int
212 reconnect_path(struct vfsmount *mnt, struct dentry *target_dir, char *nbuf)
213 {
214 	struct dentry *dentry, *parent;
215 
216 	dentry = dget(target_dir);
217 
218 	while (dentry->d_flags & DCACHE_DISCONNECTED) {
219 		BUG_ON(dentry == mnt->mnt_sb->s_root);
220 
221 		if (IS_ROOT(dentry))
222 			parent = reconnect_one(mnt, dentry, nbuf);
223 		else
224 			parent = dget_parent(dentry);
225 
226 		if (!parent)
227 			break;
228 		dput(dentry);
229 		if (IS_ERR(parent))
230 			return PTR_ERR(parent);
231 		dentry = parent;
232 	}
233 	dput(dentry);
234 	clear_disconnected(target_dir);
235 	return 0;
236 }
237 
238 struct getdents_callback {
239 	struct dir_context ctx;
240 	char *name;		/* name that was found. It already points to a
241 				   buffer NAME_MAX+1 is size */
242 	u64 ino;		/* the inum we are looking for */
243 	int found;		/* inode matched? */
244 	int sequence;		/* sequence counter */
245 };
246 
247 /*
248  * A rather strange filldir function to capture
249  * the name matching the specified inode number.
250  */
251 static bool filldir_one(struct dir_context *ctx, const char *name, int len,
252 			loff_t pos, u64 ino, unsigned int d_type)
253 {
254 	struct getdents_callback *buf =
255 		container_of(ctx, struct getdents_callback, ctx);
256 
257 	buf->sequence++;
258 	if (buf->ino == ino && len <= NAME_MAX && !is_dot_dotdot(name, len)) {
259 		memcpy(buf->name, name, len);
260 		buf->name[len] = '\0';
261 		buf->found = 1;
262 		return false;	// no more
263 	}
264 	return true;
265 }
266 
267 /**
268  * get_name - default export_operations->get_name function
269  * @path:   the directory in which to find a name
270  * @name:   a pointer to a %NAME_MAX+1 char buffer to store the name
271  * @child:  the dentry for the child directory.
272  *
273  * calls readdir on the parent until it finds an entry with
274  * the same inode number as the child, and returns that.
275  */
276 static int get_name(const struct path *path, char *name, struct dentry *child)
277 {
278 	const struct cred *cred = current_cred();
279 	struct inode *dir = path->dentry->d_inode;
280 	int error;
281 	struct file *file;
282 	struct kstat stat;
283 	struct path child_path = {
284 		.mnt = path->mnt,
285 		.dentry = child,
286 	};
287 	struct getdents_callback buffer = {
288 		.ctx.actor = filldir_one,
289 		.name = name,
290 	};
291 
292 	error = -ENOTDIR;
293 	if (!dir || !S_ISDIR(dir->i_mode))
294 		goto out;
295 	error = -EINVAL;
296 	if (!dir->i_fop)
297 		goto out;
298 	/*
299 	 * inode->i_ino is unsigned long, kstat->ino is u64, so the
300 	 * former would be insufficient on 32-bit hosts when the
301 	 * filesystem supports 64-bit inode numbers.  So we need to
302 	 * actually call ->getattr, not just read i_ino:
303 	 */
304 	error = vfs_getattr_nosec(&child_path, &stat,
305 				  STATX_INO, AT_STATX_SYNC_AS_STAT);
306 	if (error)
307 		return error;
308 	buffer.ino = stat.ino;
309 	/*
310 	 * Open the directory ...
311 	 */
312 	file = dentry_open(path, O_RDONLY, cred);
313 	error = PTR_ERR(file);
314 	if (IS_ERR(file))
315 		goto out;
316 
317 	error = -EINVAL;
318 	if (!file->f_op->iterate_shared)
319 		goto out_close;
320 
321 	buffer.sequence = 0;
322 	while (1) {
323 		int old_seq = buffer.sequence;
324 
325 		error = iterate_dir(file, &buffer.ctx);
326 		if (buffer.found) {
327 			error = 0;
328 			break;
329 		}
330 
331 		if (error < 0)
332 			break;
333 
334 		error = -ENOENT;
335 		if (old_seq == buffer.sequence)
336 			break;
337 	}
338 
339 out_close:
340 	fput(file);
341 out:
342 	return error;
343 }
344 
345 #define FILEID_INO64_GEN_LEN 3
346 
347 /**
348  * exportfs_encode_ino64_fid - encode non-decodeable 64bit ino file id
349  * @inode:   the object to encode
350  * @fid:     where to store the file handle fragment
351  * @max_len: maximum length to store there (in 4 byte units)
352  *
353  * This generic function is used to encode a non-decodeable file id for
354  * fanotify for filesystems that do not support NFS export.
355  */
356 static int exportfs_encode_ino64_fid(struct inode *inode, struct fid *fid,
357 				     int *max_len)
358 {
359 	if (*max_len < FILEID_INO64_GEN_LEN) {
360 		*max_len = FILEID_INO64_GEN_LEN;
361 		return FILEID_INVALID;
362 	}
363 
364 	fid->i64.ino = inode->i_ino;
365 	fid->i64.gen = inode->i_generation;
366 	*max_len = FILEID_INO64_GEN_LEN;
367 
368 	return FILEID_INO64_GEN;
369 }
370 
371 /**
372  * exportfs_encode_inode_fh - encode a file handle from inode
373  * @inode:   the object to encode
374  * @fid:     where to store the file handle fragment
375  * @max_len: maximum length to store there
376  * @parent:  parent directory inode, if wanted
377  * @flags:   properties of the requested file handle
378  *
379  * Returns an enum fid_type or a negative errno.
380  */
381 int exportfs_encode_inode_fh(struct inode *inode, struct fid *fid,
382 			     int *max_len, struct inode *parent, int flags)
383 {
384 	const struct export_operations *nop = inode->i_sb->s_export_op;
385 
386 	if (!exportfs_can_encode_fh(nop, flags))
387 		return -EOPNOTSUPP;
388 
389 	if (!nop && (flags & EXPORT_FH_FID))
390 		return exportfs_encode_ino64_fid(inode, fid, max_len);
391 
392 	return nop->encode_fh(inode, fid->raw, max_len, parent);
393 }
394 EXPORT_SYMBOL_GPL(exportfs_encode_inode_fh);
395 
396 /**
397  * exportfs_encode_fh - encode a file handle from dentry
398  * @dentry:  the object to encode
399  * @fid:     where to store the file handle fragment
400  * @max_len: maximum length to store there
401  * @flags:   properties of the requested file handle
402  *
403  * Returns an enum fid_type or a negative errno.
404  */
405 int exportfs_encode_fh(struct dentry *dentry, struct fid *fid, int *max_len,
406 		       int flags)
407 {
408 	int error;
409 	struct dentry *p = NULL;
410 	struct inode *inode = dentry->d_inode, *parent = NULL;
411 
412 	if ((flags & EXPORT_FH_CONNECTABLE) && !S_ISDIR(inode->i_mode)) {
413 		p = dget_parent(dentry);
414 		/*
415 		 * note that while p might've ceased to be our parent already,
416 		 * it's still pinned by and still positive.
417 		 */
418 		parent = p->d_inode;
419 	}
420 
421 	error = exportfs_encode_inode_fh(inode, fid, max_len, parent, flags);
422 	dput(p);
423 
424 	return error;
425 }
426 EXPORT_SYMBOL_GPL(exportfs_encode_fh);
427 
428 struct dentry *
429 exportfs_decode_fh_raw(struct vfsmount *mnt, struct fid *fid, int fh_len,
430 		       int fileid_type, unsigned int flags,
431 		       int (*acceptable)(void *, struct dentry *),
432 		       void *context)
433 {
434 	const struct export_operations *nop = mnt->mnt_sb->s_export_op;
435 	struct dentry *result, *alias;
436 	char nbuf[NAME_MAX+1];
437 	int err;
438 
439 	/*
440 	 * Try to get any dentry for the given file handle from the filesystem.
441 	 */
442 	if (!exportfs_can_decode_fh(nop))
443 		return ERR_PTR(-ESTALE);
444 	result = nop->fh_to_dentry(mnt->mnt_sb, fid, fh_len, fileid_type);
445 	if (IS_ERR_OR_NULL(result))
446 		return result;
447 
448 	if ((flags & EXPORT_FH_DIR_ONLY) && !d_is_dir(result)) {
449 		err = -ENOTDIR;
450 		goto err_result;
451 	}
452 
453 	/*
454 	 * If no acceptance criteria was specified by caller, a disconnected
455 	 * dentry is also accepatable. Callers may use this mode to query if
456 	 * file handle is stale or to get a reference to an inode without
457 	 * risking the high overhead caused by directory reconnect.
458 	 */
459 	if (!acceptable)
460 		return result;
461 
462 	if (d_is_dir(result)) {
463 		/*
464 		 * This request is for a directory.
465 		 *
466 		 * On the positive side there is only one dentry for each
467 		 * directory inode.  On the negative side this implies that we
468 		 * to ensure our dentry is connected all the way up to the
469 		 * filesystem root.
470 		 */
471 		if (result->d_flags & DCACHE_DISCONNECTED) {
472 			err = reconnect_path(mnt, result, nbuf);
473 			if (err)
474 				goto err_result;
475 		}
476 
477 		if (!acceptable(context, result)) {
478 			err = -EACCES;
479 			goto err_result;
480 		}
481 
482 		return result;
483 	} else {
484 		/*
485 		 * It's not a directory.  Life is a little more complicated.
486 		 */
487 		struct dentry *target_dir, *nresult;
488 
489 		/*
490 		 * See if either the dentry we just got from the filesystem
491 		 * or any alias for it is acceptable.  This is always true
492 		 * if this filesystem is exported without the subtreecheck
493 		 * option.  If the filesystem is exported with the subtree
494 		 * check option there's a fair chance we need to look at
495 		 * the parent directory in the file handle and make sure
496 		 * it's connected to the filesystem root.
497 		 */
498 		alias = find_acceptable_alias(result, acceptable, context);
499 		if (alias)
500 			return alias;
501 
502 		/*
503 		 * Try to extract a dentry for the parent directory from the
504 		 * file handle.  If this fails we'll have to give up.
505 		 */
506 		err = -ESTALE;
507 		if (!nop->fh_to_parent)
508 			goto err_result;
509 
510 		target_dir = nop->fh_to_parent(mnt->mnt_sb, fid,
511 				fh_len, fileid_type);
512 		if (!target_dir)
513 			goto err_result;
514 		err = PTR_ERR(target_dir);
515 		if (IS_ERR(target_dir))
516 			goto err_result;
517 
518 		/*
519 		 * And as usual we need to make sure the parent directory is
520 		 * connected to the filesystem root.  The VFS really doesn't
521 		 * like disconnected directories..
522 		 */
523 		err = reconnect_path(mnt, target_dir, nbuf);
524 		if (err) {
525 			dput(target_dir);
526 			goto err_result;
527 		}
528 
529 		/*
530 		 * Now that we've got both a well-connected parent and a
531 		 * dentry for the inode we're after, make sure that our
532 		 * inode is actually connected to the parent.
533 		 */
534 		err = exportfs_get_name(mnt, target_dir, nbuf, result);
535 		if (err) {
536 			dput(target_dir);
537 			goto err_result;
538 		}
539 
540 		inode_lock(target_dir->d_inode);
541 		nresult = lookup_one(mnt_idmap(mnt), nbuf,
542 				     target_dir, strlen(nbuf));
543 		if (!IS_ERR(nresult)) {
544 			if (unlikely(nresult->d_inode != result->d_inode)) {
545 				dput(nresult);
546 				nresult = ERR_PTR(-ESTALE);
547 			}
548 		}
549 		inode_unlock(target_dir->d_inode);
550 		/*
551 		 * At this point we are done with the parent, but it's pinned
552 		 * by the child dentry anyway.
553 		 */
554 		dput(target_dir);
555 
556 		if (IS_ERR(nresult)) {
557 			err = PTR_ERR(nresult);
558 			goto err_result;
559 		}
560 		dput(result);
561 		result = nresult;
562 
563 		/*
564 		 * And finally make sure the dentry is actually acceptable
565 		 * to NFSD.
566 		 */
567 		alias = find_acceptable_alias(result, acceptable, context);
568 		if (!alias) {
569 			err = -EACCES;
570 			goto err_result;
571 		}
572 
573 		return alias;
574 	}
575 
576  err_result:
577 	dput(result);
578 	return ERR_PTR(err);
579 }
580 EXPORT_SYMBOL_GPL(exportfs_decode_fh_raw);
581 
582 struct dentry *exportfs_decode_fh(struct vfsmount *mnt, struct fid *fid,
583 				  int fh_len, int fileid_type,
584 				  int (*acceptable)(void *, struct dentry *),
585 				  void *context)
586 {
587 	struct dentry *ret;
588 
589 	ret = exportfs_decode_fh_raw(mnt, fid, fh_len, fileid_type, 0,
590 				     acceptable, context);
591 	if (IS_ERR_OR_NULL(ret)) {
592 		if (ret == ERR_PTR(-ENOMEM))
593 			return ret;
594 		return ERR_PTR(-ESTALE);
595 	}
596 	return ret;
597 }
598 EXPORT_SYMBOL_GPL(exportfs_decode_fh);
599 
600 MODULE_LICENSE("GPL");
601