inode.c revision 332f606b
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 *
4 * Copyright (C) 2011 Novell Inc.
5 */
6
7#include <linux/fs.h>
8#include <linux/slab.h>
9#include <linux/cred.h>
10#include <linux/xattr.h>
11#include <linux/posix_acl.h>
12#include <linux/ratelimit.h>
13#include <linux/fiemap.h>
14#include <linux/fileattr.h>
15#include <linux/security.h>
16#include <linux/namei.h>
17#include "overlayfs.h"
18
19
20int ovl_setattr(struct user_namespace *mnt_userns, struct dentry *dentry,
21		struct iattr *attr)
22{
23	int err;
24	bool full_copy_up = false;
25	struct dentry *upperdentry;
26	const struct cred *old_cred;
27
28	err = setattr_prepare(&init_user_ns, dentry, attr);
29	if (err)
30		return err;
31
32	err = ovl_want_write(dentry);
33	if (err)
34		goto out;
35
36	if (attr->ia_valid & ATTR_SIZE) {
37		/* Truncate should trigger data copy up as well */
38		full_copy_up = true;
39	}
40
41	if (!full_copy_up)
42		err = ovl_copy_up(dentry);
43	else
44		err = ovl_copy_up_with_data(dentry);
45	if (!err) {
46		struct inode *winode = NULL;
47
48		upperdentry = ovl_dentry_upper(dentry);
49
50		if (attr->ia_valid & ATTR_SIZE) {
51			winode = d_inode(upperdentry);
52			err = get_write_access(winode);
53			if (err)
54				goto out_drop_write;
55		}
56
57		if (attr->ia_valid & (ATTR_KILL_SUID|ATTR_KILL_SGID))
58			attr->ia_valid &= ~ATTR_MODE;
59
60		/*
61		 * We might have to translate ovl file into real file object
62		 * once use cases emerge.  For now, simply don't let underlying
63		 * filesystem rely on attr->ia_file
64		 */
65		attr->ia_valid &= ~ATTR_FILE;
66
67		/*
68		 * If open(O_TRUNC) is done, VFS calls ->setattr with ATTR_OPEN
69		 * set.  Overlayfs does not pass O_TRUNC flag to underlying
70		 * filesystem during open -> do not pass ATTR_OPEN.  This
71		 * disables optimization in fuse which assumes open(O_TRUNC)
72		 * already set file size to 0.  But we never passed O_TRUNC to
73		 * fuse.  So by clearing ATTR_OPEN, fuse will be forced to send
74		 * setattr request to server.
75		 */
76		attr->ia_valid &= ~ATTR_OPEN;
77
78		inode_lock(upperdentry->d_inode);
79		old_cred = ovl_override_creds(dentry->d_sb);
80		err = notify_change(&init_user_ns, upperdentry, attr, NULL);
81		revert_creds(old_cred);
82		if (!err)
83			ovl_copyattr(upperdentry->d_inode, dentry->d_inode);
84		inode_unlock(upperdentry->d_inode);
85
86		if (winode)
87			put_write_access(winode);
88	}
89out_drop_write:
90	ovl_drop_write(dentry);
91out:
92	return err;
93}
94
95static void ovl_map_dev_ino(struct dentry *dentry, struct kstat *stat, int fsid)
96{
97	bool samefs = ovl_same_fs(dentry->d_sb);
98	unsigned int xinobits = ovl_xino_bits(dentry->d_sb);
99	unsigned int xinoshift = 64 - xinobits;
100
101	if (samefs) {
102		/*
103		 * When all layers are on the same fs, all real inode
104		 * number are unique, so we use the overlay st_dev,
105		 * which is friendly to du -x.
106		 */
107		stat->dev = dentry->d_sb->s_dev;
108		return;
109	} else if (xinobits) {
110		/*
111		 * All inode numbers of underlying fs should not be using the
112		 * high xinobits, so we use high xinobits to partition the
113		 * overlay st_ino address space. The high bits holds the fsid
114		 * (upper fsid is 0). The lowest xinobit is reserved for mapping
115		 * the non-persistent inode numbers range in case of overflow.
116		 * This way all overlay inode numbers are unique and use the
117		 * overlay st_dev.
118		 */
119		if (likely(!(stat->ino >> xinoshift))) {
120			stat->ino |= ((u64)fsid) << (xinoshift + 1);
121			stat->dev = dentry->d_sb->s_dev;
122			return;
123		} else if (ovl_xino_warn(dentry->d_sb)) {
124			pr_warn_ratelimited("inode number too big (%pd2, ino=%llu, xinobits=%d)\n",
125					    dentry, stat->ino, xinobits);
126		}
127	}
128
129	/* The inode could not be mapped to a unified st_ino address space */
130	if (S_ISDIR(dentry->d_inode->i_mode)) {
131		/*
132		 * Always use the overlay st_dev for directories, so 'find
133		 * -xdev' will scan the entire overlay mount and won't cross the
134		 * overlay mount boundaries.
135		 *
136		 * If not all layers are on the same fs the pair {real st_ino;
137		 * overlay st_dev} is not unique, so use the non persistent
138		 * overlay st_ino for directories.
139		 */
140		stat->dev = dentry->d_sb->s_dev;
141		stat->ino = dentry->d_inode->i_ino;
142	} else {
143		/*
144		 * For non-samefs setup, if we cannot map all layers st_ino
145		 * to a unified address space, we need to make sure that st_dev
146		 * is unique per underlying fs, so we use the unique anonymous
147		 * bdev assigned to the underlying fs.
148		 */
149		stat->dev = OVL_FS(dentry->d_sb)->fs[fsid].pseudo_dev;
150	}
151}
152
153int ovl_getattr(struct user_namespace *mnt_userns, const struct path *path,
154		struct kstat *stat, u32 request_mask, unsigned int flags)
155{
156	struct dentry *dentry = path->dentry;
157	enum ovl_path_type type;
158	struct path realpath;
159	const struct cred *old_cred;
160	struct inode *inode = d_inode(dentry);
161	bool is_dir = S_ISDIR(inode->i_mode);
162	int fsid = 0;
163	int err;
164	bool metacopy_blocks = false;
165
166	metacopy_blocks = ovl_is_metacopy_dentry(dentry);
167
168	type = ovl_path_real(dentry, &realpath);
169	old_cred = ovl_override_creds(dentry->d_sb);
170	err = vfs_getattr(&realpath, stat, request_mask, flags);
171	if (err)
172		goto out;
173
174	/* Report the effective immutable/append-only STATX flags */
175	generic_fill_statx_attr(inode, stat);
176
177	/*
178	 * For non-dir or same fs, we use st_ino of the copy up origin.
179	 * This guaranties constant st_dev/st_ino across copy up.
180	 * With xino feature and non-samefs, we use st_ino of the copy up
181	 * origin masked with high bits that represent the layer id.
182	 *
183	 * If lower filesystem supports NFS file handles, this also guaranties
184	 * persistent st_ino across mount cycle.
185	 */
186	if (!is_dir || ovl_same_dev(dentry->d_sb)) {
187		if (!OVL_TYPE_UPPER(type)) {
188			fsid = ovl_layer_lower(dentry)->fsid;
189		} else if (OVL_TYPE_ORIGIN(type)) {
190			struct kstat lowerstat;
191			u32 lowermask = STATX_INO | STATX_BLOCKS |
192					(!is_dir ? STATX_NLINK : 0);
193
194			ovl_path_lower(dentry, &realpath);
195			err = vfs_getattr(&realpath, &lowerstat,
196					  lowermask, flags);
197			if (err)
198				goto out;
199
200			/*
201			 * Lower hardlinks may be broken on copy up to different
202			 * upper files, so we cannot use the lower origin st_ino
203			 * for those different files, even for the same fs case.
204			 *
205			 * Similarly, several redirected dirs can point to the
206			 * same dir on a lower layer. With the "verify_lower"
207			 * feature, we do not use the lower origin st_ino, if
208			 * we haven't verified that this redirect is unique.
209			 *
210			 * With inodes index enabled, it is safe to use st_ino
211			 * of an indexed origin. The index validates that the
212			 * upper hardlink is not broken and that a redirected
213			 * dir is the only redirect to that origin.
214			 */
215			if (ovl_test_flag(OVL_INDEX, d_inode(dentry)) ||
216			    (!ovl_verify_lower(dentry->d_sb) &&
217			     (is_dir || lowerstat.nlink == 1))) {
218				fsid = ovl_layer_lower(dentry)->fsid;
219				stat->ino = lowerstat.ino;
220			}
221
222			/*
223			 * If we are querying a metacopy dentry and lower
224			 * dentry is data dentry, then use the blocks we
225			 * queried just now. We don't have to do additional
226			 * vfs_getattr(). If lower itself is metacopy, then
227			 * additional vfs_getattr() is unavoidable.
228			 */
229			if (metacopy_blocks &&
230			    realpath.dentry == ovl_dentry_lowerdata(dentry)) {
231				stat->blocks = lowerstat.blocks;
232				metacopy_blocks = false;
233			}
234		}
235
236		if (metacopy_blocks) {
237			/*
238			 * If lower is not same as lowerdata or if there was
239			 * no origin on upper, we can end up here.
240			 */
241			struct kstat lowerdatastat;
242			u32 lowermask = STATX_BLOCKS;
243
244			ovl_path_lowerdata(dentry, &realpath);
245			err = vfs_getattr(&realpath, &lowerdatastat,
246					  lowermask, flags);
247			if (err)
248				goto out;
249			stat->blocks = lowerdatastat.blocks;
250		}
251	}
252
253	ovl_map_dev_ino(dentry, stat, fsid);
254
255	/*
256	 * It's probably not worth it to count subdirs to get the
257	 * correct link count.  nlink=1 seems to pacify 'find' and
258	 * other utilities.
259	 */
260	if (is_dir && OVL_TYPE_MERGE(type))
261		stat->nlink = 1;
262
263	/*
264	 * Return the overlay inode nlinks for indexed upper inodes.
265	 * Overlay inode nlink counts the union of the upper hardlinks
266	 * and non-covered lower hardlinks. It does not include the upper
267	 * index hardlink.
268	 */
269	if (!is_dir && ovl_test_flag(OVL_INDEX, d_inode(dentry)))
270		stat->nlink = dentry->d_inode->i_nlink;
271
272out:
273	revert_creds(old_cred);
274
275	return err;
276}
277
278int ovl_permission(struct user_namespace *mnt_userns,
279		   struct inode *inode, int mask)
280{
281	struct inode *upperinode = ovl_inode_upper(inode);
282	struct inode *realinode = upperinode ?: ovl_inode_lower(inode);
283	const struct cred *old_cred;
284	int err;
285
286	/* Careful in RCU walk mode */
287	if (!realinode) {
288		WARN_ON(!(mask & MAY_NOT_BLOCK));
289		return -ECHILD;
290	}
291
292	/*
293	 * Check overlay inode with the creds of task and underlying inode
294	 * with creds of mounter
295	 */
296	err = generic_permission(&init_user_ns, inode, mask);
297	if (err)
298		return err;
299
300	old_cred = ovl_override_creds(inode->i_sb);
301	if (!upperinode &&
302	    !special_file(realinode->i_mode) && mask & MAY_WRITE) {
303		mask &= ~(MAY_WRITE | MAY_APPEND);
304		/* Make sure mounter can read file for copy up later */
305		mask |= MAY_READ;
306	}
307	err = inode_permission(&init_user_ns, realinode, mask);
308	revert_creds(old_cred);
309
310	return err;
311}
312
313static const char *ovl_get_link(struct dentry *dentry,
314				struct inode *inode,
315				struct delayed_call *done)
316{
317	const struct cred *old_cred;
318	const char *p;
319
320	if (!dentry)
321		return ERR_PTR(-ECHILD);
322
323	old_cred = ovl_override_creds(dentry->d_sb);
324	p = vfs_get_link(ovl_dentry_real(dentry), done);
325	revert_creds(old_cred);
326	return p;
327}
328
329bool ovl_is_private_xattr(struct super_block *sb, const char *name)
330{
331	struct ovl_fs *ofs = sb->s_fs_info;
332
333	if (ofs->config.userxattr)
334		return strncmp(name, OVL_XATTR_USER_PREFIX,
335			       sizeof(OVL_XATTR_USER_PREFIX) - 1) == 0;
336	else
337		return strncmp(name, OVL_XATTR_TRUSTED_PREFIX,
338			       sizeof(OVL_XATTR_TRUSTED_PREFIX) - 1) == 0;
339}
340
341int ovl_xattr_set(struct dentry *dentry, struct inode *inode, const char *name,
342		  const void *value, size_t size, int flags)
343{
344	int err;
345	struct dentry *upperdentry = ovl_i_dentry_upper(inode);
346	struct dentry *realdentry = upperdentry ?: ovl_dentry_lower(dentry);
347	const struct cred *old_cred;
348
349	err = ovl_want_write(dentry);
350	if (err)
351		goto out;
352
353	if (!value && !upperdentry) {
354		old_cred = ovl_override_creds(dentry->d_sb);
355		err = vfs_getxattr(&init_user_ns, realdentry, name, NULL, 0);
356		revert_creds(old_cred);
357		if (err < 0)
358			goto out_drop_write;
359	}
360
361	if (!upperdentry) {
362		err = ovl_copy_up(dentry);
363		if (err)
364			goto out_drop_write;
365
366		realdentry = ovl_dentry_upper(dentry);
367	}
368
369	old_cred = ovl_override_creds(dentry->d_sb);
370	if (value)
371		err = vfs_setxattr(&init_user_ns, realdentry, name, value, size,
372				   flags);
373	else {
374		WARN_ON(flags != XATTR_REPLACE);
375		err = vfs_removexattr(&init_user_ns, realdentry, name);
376	}
377	revert_creds(old_cred);
378
379	/* copy c/mtime */
380	ovl_copyattr(d_inode(realdentry), inode);
381
382out_drop_write:
383	ovl_drop_write(dentry);
384out:
385	return err;
386}
387
388int ovl_xattr_get(struct dentry *dentry, struct inode *inode, const char *name,
389		  void *value, size_t size)
390{
391	ssize_t res;
392	const struct cred *old_cred;
393	struct dentry *realdentry =
394		ovl_i_dentry_upper(inode) ?: ovl_dentry_lower(dentry);
395
396	old_cred = ovl_override_creds(dentry->d_sb);
397	res = vfs_getxattr(&init_user_ns, realdentry, name, value, size);
398	revert_creds(old_cred);
399	return res;
400}
401
402static bool ovl_can_list(struct super_block *sb, const char *s)
403{
404	/* Never list private (.overlay) */
405	if (ovl_is_private_xattr(sb, s))
406		return false;
407
408	/* List all non-trusted xattrs */
409	if (strncmp(s, XATTR_TRUSTED_PREFIX, XATTR_TRUSTED_PREFIX_LEN) != 0)
410		return true;
411
412	/* list other trusted for superuser only */
413	return ns_capable_noaudit(&init_user_ns, CAP_SYS_ADMIN);
414}
415
416ssize_t ovl_listxattr(struct dentry *dentry, char *list, size_t size)
417{
418	struct dentry *realdentry = ovl_dentry_real(dentry);
419	ssize_t res;
420	size_t len;
421	char *s;
422	const struct cred *old_cred;
423
424	old_cred = ovl_override_creds(dentry->d_sb);
425	res = vfs_listxattr(realdentry, list, size);
426	revert_creds(old_cred);
427	if (res <= 0 || size == 0)
428		return res;
429
430	/* filter out private xattrs */
431	for (s = list, len = res; len;) {
432		size_t slen = strnlen(s, len) + 1;
433
434		/* underlying fs providing us with an broken xattr list? */
435		if (WARN_ON(slen > len))
436			return -EIO;
437
438		len -= slen;
439		if (!ovl_can_list(dentry->d_sb, s)) {
440			res -= slen;
441			memmove(s, s + slen, len);
442		} else {
443			s += slen;
444		}
445	}
446
447	return res;
448}
449
450struct posix_acl *ovl_get_acl(struct inode *inode, int type, bool rcu)
451{
452	struct inode *realinode = ovl_inode_real(inode);
453	const struct cred *old_cred;
454	struct posix_acl *acl;
455
456	if (!IS_ENABLED(CONFIG_FS_POSIX_ACL) || !IS_POSIXACL(realinode))
457		return NULL;
458
459	if (rcu)
460		return get_cached_acl_rcu(realinode, type);
461
462	old_cred = ovl_override_creds(inode->i_sb);
463	acl = get_acl(realinode, type);
464	revert_creds(old_cred);
465
466	return acl;
467}
468
469int ovl_update_time(struct inode *inode, struct timespec64 *ts, int flags)
470{
471	if (flags & S_ATIME) {
472		struct ovl_fs *ofs = inode->i_sb->s_fs_info;
473		struct path upperpath = {
474			.mnt = ovl_upper_mnt(ofs),
475			.dentry = ovl_upperdentry_dereference(OVL_I(inode)),
476		};
477
478		if (upperpath.dentry) {
479			touch_atime(&upperpath);
480			inode->i_atime = d_inode(upperpath.dentry)->i_atime;
481		}
482	}
483	return 0;
484}
485
486static int ovl_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
487		      u64 start, u64 len)
488{
489	int err;
490	struct inode *realinode = ovl_inode_realdata(inode);
491	const struct cred *old_cred;
492
493	if (!realinode->i_op->fiemap)
494		return -EOPNOTSUPP;
495
496	old_cred = ovl_override_creds(inode->i_sb);
497	err = realinode->i_op->fiemap(realinode, fieinfo, start, len);
498	revert_creds(old_cred);
499
500	return err;
501}
502
503/*
504 * Work around the fact that security_file_ioctl() takes a file argument.
505 * Introducing security_inode_fileattr_get/set() hooks would solve this issue
506 * properly.
507 */
508static int ovl_security_fileattr(struct path *realpath, struct fileattr *fa,
509				 bool set)
510{
511	struct file *file;
512	unsigned int cmd;
513	int err;
514
515	file = dentry_open(realpath, O_RDONLY, current_cred());
516	if (IS_ERR(file))
517		return PTR_ERR(file);
518
519	if (set)
520		cmd = fa->fsx_valid ? FS_IOC_FSSETXATTR : FS_IOC_SETFLAGS;
521	else
522		cmd = fa->fsx_valid ? FS_IOC_FSGETXATTR : FS_IOC_GETFLAGS;
523
524	err = security_file_ioctl(file, cmd, 0);
525	fput(file);
526
527	return err;
528}
529
530int ovl_real_fileattr_set(struct path *realpath, struct fileattr *fa)
531{
532	int err;
533
534	err = ovl_security_fileattr(realpath, fa, true);
535	if (err)
536		return err;
537
538	return vfs_fileattr_set(&init_user_ns, realpath->dentry, fa);
539}
540
541int ovl_fileattr_set(struct user_namespace *mnt_userns,
542		     struct dentry *dentry, struct fileattr *fa)
543{
544	struct inode *inode = d_inode(dentry);
545	struct path upperpath;
546	const struct cred *old_cred;
547	unsigned int flags;
548	int err;
549
550	err = ovl_want_write(dentry);
551	if (err)
552		goto out;
553
554	err = ovl_copy_up(dentry);
555	if (!err) {
556		ovl_path_real(dentry, &upperpath);
557
558		old_cred = ovl_override_creds(inode->i_sb);
559		/*
560		 * Store immutable/append-only flags in xattr and clear them
561		 * in upper fileattr (in case they were set by older kernel)
562		 * so children of "ovl-immutable" directories lower aliases of
563		 * "ovl-immutable" hardlinks could be copied up.
564		 * Clear xattr when flags are cleared.
565		 */
566		err = ovl_set_protattr(inode, upperpath.dentry, fa);
567		if (!err)
568			err = ovl_real_fileattr_set(&upperpath, fa);
569		revert_creds(old_cred);
570
571		/*
572		 * Merge real inode flags with inode flags read from
573		 * overlay.protattr xattr
574		 */
575		flags = ovl_inode_real(inode)->i_flags & OVL_COPY_I_FLAGS_MASK;
576
577		BUILD_BUG_ON(OVL_PROT_I_FLAGS_MASK & ~OVL_COPY_I_FLAGS_MASK);
578		flags |= inode->i_flags & OVL_PROT_I_FLAGS_MASK;
579		inode_set_flags(inode, flags, OVL_COPY_I_FLAGS_MASK);
580
581		/* Update ctime */
582		ovl_copyattr(ovl_inode_real(inode), inode);
583	}
584	ovl_drop_write(dentry);
585out:
586	return err;
587}
588
589/* Convert inode protection flags to fileattr flags */
590static void ovl_fileattr_prot_flags(struct inode *inode, struct fileattr *fa)
591{
592	BUILD_BUG_ON(OVL_PROT_FS_FLAGS_MASK & ~FS_COMMON_FL);
593	BUILD_BUG_ON(OVL_PROT_FSX_FLAGS_MASK & ~FS_XFLAG_COMMON);
594
595	if (inode->i_flags & S_APPEND) {
596		fa->flags |= FS_APPEND_FL;
597		fa->fsx_xflags |= FS_XFLAG_APPEND;
598	}
599	if (inode->i_flags & S_IMMUTABLE) {
600		fa->flags |= FS_IMMUTABLE_FL;
601		fa->fsx_xflags |= FS_XFLAG_IMMUTABLE;
602	}
603}
604
605int ovl_real_fileattr_get(struct path *realpath, struct fileattr *fa)
606{
607	int err;
608
609	err = ovl_security_fileattr(realpath, fa, false);
610	if (err)
611		return err;
612
613	return vfs_fileattr_get(realpath->dentry, fa);
614}
615
616int ovl_fileattr_get(struct dentry *dentry, struct fileattr *fa)
617{
618	struct inode *inode = d_inode(dentry);
619	struct path realpath;
620	const struct cred *old_cred;
621	int err;
622
623	ovl_path_real(dentry, &realpath);
624
625	old_cred = ovl_override_creds(inode->i_sb);
626	err = ovl_real_fileattr_get(&realpath, fa);
627	ovl_fileattr_prot_flags(inode, fa);
628	revert_creds(old_cred);
629
630	return err;
631}
632
633static const struct inode_operations ovl_file_inode_operations = {
634	.setattr	= ovl_setattr,
635	.permission	= ovl_permission,
636	.getattr	= ovl_getattr,
637	.listxattr	= ovl_listxattr,
638	.get_acl	= ovl_get_acl,
639	.update_time	= ovl_update_time,
640	.fiemap		= ovl_fiemap,
641	.fileattr_get	= ovl_fileattr_get,
642	.fileattr_set	= ovl_fileattr_set,
643};
644
645static const struct inode_operations ovl_symlink_inode_operations = {
646	.setattr	= ovl_setattr,
647	.get_link	= ovl_get_link,
648	.getattr	= ovl_getattr,
649	.listxattr	= ovl_listxattr,
650	.update_time	= ovl_update_time,
651};
652
653static const struct inode_operations ovl_special_inode_operations = {
654	.setattr	= ovl_setattr,
655	.permission	= ovl_permission,
656	.getattr	= ovl_getattr,
657	.listxattr	= ovl_listxattr,
658	.get_acl	= ovl_get_acl,
659	.update_time	= ovl_update_time,
660};
661
662static const struct address_space_operations ovl_aops = {
663	/* For O_DIRECT dentry_open() checks f_mapping->a_ops->direct_IO */
664	.direct_IO		= noop_direct_IO,
665};
666
667/*
668 * It is possible to stack overlayfs instance on top of another
669 * overlayfs instance as lower layer. We need to annotate the
670 * stackable i_mutex locks according to stack level of the super
671 * block instance. An overlayfs instance can never be in stack
672 * depth 0 (there is always a real fs below it).  An overlayfs
673 * inode lock will use the lockdep annotation ovl_i_mutex_key[depth].
674 *
675 * For example, here is a snip from /proc/lockdep_chains after
676 * dir_iterate of nested overlayfs:
677 *
678 * [...] &ovl_i_mutex_dir_key[depth]   (stack_depth=2)
679 * [...] &ovl_i_mutex_dir_key[depth]#2 (stack_depth=1)
680 * [...] &type->i_mutex_dir_key        (stack_depth=0)
681 *
682 * Locking order w.r.t ovl_want_write() is important for nested overlayfs.
683 *
684 * This chain is valid:
685 * - inode->i_rwsem			(inode_lock[2])
686 * - upper_mnt->mnt_sb->s_writers	(ovl_want_write[0])
687 * - OVL_I(inode)->lock			(ovl_inode_lock[2])
688 * - OVL_I(lowerinode)->lock		(ovl_inode_lock[1])
689 *
690 * And this chain is valid:
691 * - inode->i_rwsem			(inode_lock[2])
692 * - OVL_I(inode)->lock			(ovl_inode_lock[2])
693 * - lowerinode->i_rwsem		(inode_lock[1])
694 * - OVL_I(lowerinode)->lock		(ovl_inode_lock[1])
695 *
696 * But lowerinode->i_rwsem SHOULD NOT be acquired while ovl_want_write() is
697 * held, because it is in reverse order of the non-nested case using the same
698 * upper fs:
699 * - inode->i_rwsem			(inode_lock[1])
700 * - upper_mnt->mnt_sb->s_writers	(ovl_want_write[0])
701 * - OVL_I(inode)->lock			(ovl_inode_lock[1])
702 */
703#define OVL_MAX_NESTING FILESYSTEM_MAX_STACK_DEPTH
704
705static inline void ovl_lockdep_annotate_inode_mutex_key(struct inode *inode)
706{
707#ifdef CONFIG_LOCKDEP
708	static struct lock_class_key ovl_i_mutex_key[OVL_MAX_NESTING];
709	static struct lock_class_key ovl_i_mutex_dir_key[OVL_MAX_NESTING];
710	static struct lock_class_key ovl_i_lock_key[OVL_MAX_NESTING];
711
712	int depth = inode->i_sb->s_stack_depth - 1;
713
714	if (WARN_ON_ONCE(depth < 0 || depth >= OVL_MAX_NESTING))
715		depth = 0;
716
717	if (S_ISDIR(inode->i_mode))
718		lockdep_set_class(&inode->i_rwsem, &ovl_i_mutex_dir_key[depth]);
719	else
720		lockdep_set_class(&inode->i_rwsem, &ovl_i_mutex_key[depth]);
721
722	lockdep_set_class(&OVL_I(inode)->lock, &ovl_i_lock_key[depth]);
723#endif
724}
725
726static void ovl_next_ino(struct inode *inode)
727{
728	struct ovl_fs *ofs = inode->i_sb->s_fs_info;
729
730	inode->i_ino = atomic_long_inc_return(&ofs->last_ino);
731	if (unlikely(!inode->i_ino))
732		inode->i_ino = atomic_long_inc_return(&ofs->last_ino);
733}
734
735static void ovl_map_ino(struct inode *inode, unsigned long ino, int fsid)
736{
737	int xinobits = ovl_xino_bits(inode->i_sb);
738	unsigned int xinoshift = 64 - xinobits;
739
740	/*
741	 * When d_ino is consistent with st_ino (samefs or i_ino has enough
742	 * bits to encode layer), set the same value used for st_ino to i_ino,
743	 * so inode number exposed via /proc/locks and a like will be
744	 * consistent with d_ino and st_ino values. An i_ino value inconsistent
745	 * with d_ino also causes nfsd readdirplus to fail.
746	 */
747	inode->i_ino = ino;
748	if (ovl_same_fs(inode->i_sb)) {
749		return;
750	} else if (xinobits && likely(!(ino >> xinoshift))) {
751		inode->i_ino |= (unsigned long)fsid << (xinoshift + 1);
752		return;
753	}
754
755	/*
756	 * For directory inodes on non-samefs with xino disabled or xino
757	 * overflow, we allocate a non-persistent inode number, to be used for
758	 * resolving st_ino collisions in ovl_map_dev_ino().
759	 *
760	 * To avoid ino collision with legitimate xino values from upper
761	 * layer (fsid 0), use the lowest xinobit to map the non
762	 * persistent inode numbers to the unified st_ino address space.
763	 */
764	if (S_ISDIR(inode->i_mode)) {
765		ovl_next_ino(inode);
766		if (xinobits) {
767			inode->i_ino &= ~0UL >> xinobits;
768			inode->i_ino |= 1UL << xinoshift;
769		}
770	}
771}
772
773void ovl_inode_init(struct inode *inode, struct ovl_inode_params *oip,
774		    unsigned long ino, int fsid)
775{
776	struct inode *realinode;
777
778	if (oip->upperdentry)
779		OVL_I(inode)->__upperdentry = oip->upperdentry;
780	if (oip->lowerpath && oip->lowerpath->dentry)
781		OVL_I(inode)->lower = igrab(d_inode(oip->lowerpath->dentry));
782	if (oip->lowerdata)
783		OVL_I(inode)->lowerdata = igrab(d_inode(oip->lowerdata));
784
785	realinode = ovl_inode_real(inode);
786	ovl_copyattr(realinode, inode);
787	ovl_copyflags(realinode, inode);
788	ovl_map_ino(inode, ino, fsid);
789}
790
791static void ovl_fill_inode(struct inode *inode, umode_t mode, dev_t rdev)
792{
793	inode->i_mode = mode;
794	inode->i_flags |= S_NOCMTIME;
795#ifdef CONFIG_FS_POSIX_ACL
796	inode->i_acl = inode->i_default_acl = ACL_DONT_CACHE;
797#endif
798
799	ovl_lockdep_annotate_inode_mutex_key(inode);
800
801	switch (mode & S_IFMT) {
802	case S_IFREG:
803		inode->i_op = &ovl_file_inode_operations;
804		inode->i_fop = &ovl_file_operations;
805		inode->i_mapping->a_ops = &ovl_aops;
806		break;
807
808	case S_IFDIR:
809		inode->i_op = &ovl_dir_inode_operations;
810		inode->i_fop = &ovl_dir_operations;
811		break;
812
813	case S_IFLNK:
814		inode->i_op = &ovl_symlink_inode_operations;
815		break;
816
817	default:
818		inode->i_op = &ovl_special_inode_operations;
819		init_special_inode(inode, mode, rdev);
820		break;
821	}
822}
823
824/*
825 * With inodes index enabled, an overlay inode nlink counts the union of upper
826 * hardlinks and non-covered lower hardlinks. During the lifetime of a non-pure
827 * upper inode, the following nlink modifying operations can happen:
828 *
829 * 1. Lower hardlink copy up
830 * 2. Upper hardlink created, unlinked or renamed over
831 * 3. Lower hardlink whiteout or renamed over
832 *
833 * For the first, copy up case, the union nlink does not change, whether the
834 * operation succeeds or fails, but the upper inode nlink may change.
835 * Therefore, before copy up, we store the union nlink value relative to the
836 * lower inode nlink in the index inode xattr .overlay.nlink.
837 *
838 * For the second, upper hardlink case, the union nlink should be incremented
839 * or decremented IFF the operation succeeds, aligned with nlink change of the
840 * upper inode. Therefore, before link/unlink/rename, we store the union nlink
841 * value relative to the upper inode nlink in the index inode.
842 *
843 * For the last, lower cover up case, we simplify things by preceding the
844 * whiteout or cover up with copy up. This makes sure that there is an index
845 * upper inode where the nlink xattr can be stored before the copied up upper
846 * entry is unlink.
847 */
848#define OVL_NLINK_ADD_UPPER	(1 << 0)
849
850/*
851 * On-disk format for indexed nlink:
852 *
853 * nlink relative to the upper inode - "U[+-]NUM"
854 * nlink relative to the lower inode - "L[+-]NUM"
855 */
856
857static int ovl_set_nlink_common(struct dentry *dentry,
858				struct dentry *realdentry, const char *format)
859{
860	struct inode *inode = d_inode(dentry);
861	struct inode *realinode = d_inode(realdentry);
862	char buf[13];
863	int len;
864
865	len = snprintf(buf, sizeof(buf), format,
866		       (int) (inode->i_nlink - realinode->i_nlink));
867
868	if (WARN_ON(len >= sizeof(buf)))
869		return -EIO;
870
871	return ovl_do_setxattr(OVL_FS(inode->i_sb), ovl_dentry_upper(dentry),
872			       OVL_XATTR_NLINK, buf, len);
873}
874
875int ovl_set_nlink_upper(struct dentry *dentry)
876{
877	return ovl_set_nlink_common(dentry, ovl_dentry_upper(dentry), "U%+i");
878}
879
880int ovl_set_nlink_lower(struct dentry *dentry)
881{
882	return ovl_set_nlink_common(dentry, ovl_dentry_lower(dentry), "L%+i");
883}
884
885unsigned int ovl_get_nlink(struct ovl_fs *ofs, struct dentry *lowerdentry,
886			   struct dentry *upperdentry,
887			   unsigned int fallback)
888{
889	int nlink_diff;
890	int nlink;
891	char buf[13];
892	int err;
893
894	if (!lowerdentry || !upperdentry || d_inode(lowerdentry)->i_nlink == 1)
895		return fallback;
896
897	err = ovl_do_getxattr(ofs, upperdentry, OVL_XATTR_NLINK,
898			      &buf, sizeof(buf) - 1);
899	if (err < 0)
900		goto fail;
901
902	buf[err] = '\0';
903	if ((buf[0] != 'L' && buf[0] != 'U') ||
904	    (buf[1] != '+' && buf[1] != '-'))
905		goto fail;
906
907	err = kstrtoint(buf + 1, 10, &nlink_diff);
908	if (err < 0)
909		goto fail;
910
911	nlink = d_inode(buf[0] == 'L' ? lowerdentry : upperdentry)->i_nlink;
912	nlink += nlink_diff;
913
914	if (nlink <= 0)
915		goto fail;
916
917	return nlink;
918
919fail:
920	pr_warn_ratelimited("failed to get index nlink (%pd2, err=%i)\n",
921			    upperdentry, err);
922	return fallback;
923}
924
925struct inode *ovl_new_inode(struct super_block *sb, umode_t mode, dev_t rdev)
926{
927	struct inode *inode;
928
929	inode = new_inode(sb);
930	if (inode)
931		ovl_fill_inode(inode, mode, rdev);
932
933	return inode;
934}
935
936static int ovl_inode_test(struct inode *inode, void *data)
937{
938	return inode->i_private == data;
939}
940
941static int ovl_inode_set(struct inode *inode, void *data)
942{
943	inode->i_private = data;
944	return 0;
945}
946
947static bool ovl_verify_inode(struct inode *inode, struct dentry *lowerdentry,
948			     struct dentry *upperdentry, bool strict)
949{
950	/*
951	 * For directories, @strict verify from lookup path performs consistency
952	 * checks, so NULL lower/upper in dentry must match NULL lower/upper in
953	 * inode. Non @strict verify from NFS handle decode path passes NULL for
954	 * 'unknown' lower/upper.
955	 */
956	if (S_ISDIR(inode->i_mode) && strict) {
957		/* Real lower dir moved to upper layer under us? */
958		if (!lowerdentry && ovl_inode_lower(inode))
959			return false;
960
961		/* Lookup of an uncovered redirect origin? */
962		if (!upperdentry && ovl_inode_upper(inode))
963			return false;
964	}
965
966	/*
967	 * Allow non-NULL lower inode in ovl_inode even if lowerdentry is NULL.
968	 * This happens when finding a copied up overlay inode for a renamed
969	 * or hardlinked overlay dentry and lower dentry cannot be followed
970	 * by origin because lower fs does not support file handles.
971	 */
972	if (lowerdentry && ovl_inode_lower(inode) != d_inode(lowerdentry))
973		return false;
974
975	/*
976	 * Allow non-NULL __upperdentry in inode even if upperdentry is NULL.
977	 * This happens when finding a lower alias for a copied up hard link.
978	 */
979	if (upperdentry && ovl_inode_upper(inode) != d_inode(upperdentry))
980		return false;
981
982	return true;
983}
984
985struct inode *ovl_lookup_inode(struct super_block *sb, struct dentry *real,
986			       bool is_upper)
987{
988	struct inode *inode, *key = d_inode(real);
989
990	inode = ilookup5(sb, (unsigned long) key, ovl_inode_test, key);
991	if (!inode)
992		return NULL;
993
994	if (!ovl_verify_inode(inode, is_upper ? NULL : real,
995			      is_upper ? real : NULL, false)) {
996		iput(inode);
997		return ERR_PTR(-ESTALE);
998	}
999
1000	return inode;
1001}
1002
1003bool ovl_lookup_trap_inode(struct super_block *sb, struct dentry *dir)
1004{
1005	struct inode *key = d_inode(dir);
1006	struct inode *trap;
1007	bool res;
1008
1009	trap = ilookup5(sb, (unsigned long) key, ovl_inode_test, key);
1010	if (!trap)
1011		return false;
1012
1013	res = IS_DEADDIR(trap) && !ovl_inode_upper(trap) &&
1014				  !ovl_inode_lower(trap);
1015
1016	iput(trap);
1017	return res;
1018}
1019
1020/*
1021 * Create an inode cache entry for layer root dir, that will intentionally
1022 * fail ovl_verify_inode(), so any lookup that will find some layer root
1023 * will fail.
1024 */
1025struct inode *ovl_get_trap_inode(struct super_block *sb, struct dentry *dir)
1026{
1027	struct inode *key = d_inode(dir);
1028	struct inode *trap;
1029
1030	if (!d_is_dir(dir))
1031		return ERR_PTR(-ENOTDIR);
1032
1033	trap = iget5_locked(sb, (unsigned long) key, ovl_inode_test,
1034			    ovl_inode_set, key);
1035	if (!trap)
1036		return ERR_PTR(-ENOMEM);
1037
1038	if (!(trap->i_state & I_NEW)) {
1039		/* Conflicting layer roots? */
1040		iput(trap);
1041		return ERR_PTR(-ELOOP);
1042	}
1043
1044	trap->i_mode = S_IFDIR;
1045	trap->i_flags = S_DEAD;
1046	unlock_new_inode(trap);
1047
1048	return trap;
1049}
1050
1051/*
1052 * Does overlay inode need to be hashed by lower inode?
1053 */
1054static bool ovl_hash_bylower(struct super_block *sb, struct dentry *upper,
1055			     struct dentry *lower, bool index)
1056{
1057	struct ovl_fs *ofs = sb->s_fs_info;
1058
1059	/* No, if pure upper */
1060	if (!lower)
1061		return false;
1062
1063	/* Yes, if already indexed */
1064	if (index)
1065		return true;
1066
1067	/* Yes, if won't be copied up */
1068	if (!ovl_upper_mnt(ofs))
1069		return true;
1070
1071	/* No, if lower hardlink is or will be broken on copy up */
1072	if ((upper || !ovl_indexdir(sb)) &&
1073	    !d_is_dir(lower) && d_inode(lower)->i_nlink > 1)
1074		return false;
1075
1076	/* No, if non-indexed upper with NFS export */
1077	if (sb->s_export_op && upper)
1078		return false;
1079
1080	/* Otherwise, hash by lower inode for fsnotify */
1081	return true;
1082}
1083
1084static struct inode *ovl_iget5(struct super_block *sb, struct inode *newinode,
1085			       struct inode *key)
1086{
1087	return newinode ? inode_insert5(newinode, (unsigned long) key,
1088					 ovl_inode_test, ovl_inode_set, key) :
1089			  iget5_locked(sb, (unsigned long) key,
1090				       ovl_inode_test, ovl_inode_set, key);
1091}
1092
1093struct inode *ovl_get_inode(struct super_block *sb,
1094			    struct ovl_inode_params *oip)
1095{
1096	struct ovl_fs *ofs = OVL_FS(sb);
1097	struct dentry *upperdentry = oip->upperdentry;
1098	struct ovl_path *lowerpath = oip->lowerpath;
1099	struct inode *realinode = upperdentry ? d_inode(upperdentry) : NULL;
1100	struct inode *inode;
1101	struct dentry *lowerdentry = lowerpath ? lowerpath->dentry : NULL;
1102	bool bylower = ovl_hash_bylower(sb, upperdentry, lowerdentry,
1103					oip->index);
1104	int fsid = bylower ? lowerpath->layer->fsid : 0;
1105	bool is_dir;
1106	unsigned long ino = 0;
1107	int err = oip->newinode ? -EEXIST : -ENOMEM;
1108
1109	if (!realinode)
1110		realinode = d_inode(lowerdentry);
1111
1112	/*
1113	 * Copy up origin (lower) may exist for non-indexed upper, but we must
1114	 * not use lower as hash key if this is a broken hardlink.
1115	 */
1116	is_dir = S_ISDIR(realinode->i_mode);
1117	if (upperdentry || bylower) {
1118		struct inode *key = d_inode(bylower ? lowerdentry :
1119						      upperdentry);
1120		unsigned int nlink = is_dir ? 1 : realinode->i_nlink;
1121
1122		inode = ovl_iget5(sb, oip->newinode, key);
1123		if (!inode)
1124			goto out_err;
1125		if (!(inode->i_state & I_NEW)) {
1126			/*
1127			 * Verify that the underlying files stored in the inode
1128			 * match those in the dentry.
1129			 */
1130			if (!ovl_verify_inode(inode, lowerdentry, upperdentry,
1131					      true)) {
1132				iput(inode);
1133				err = -ESTALE;
1134				goto out_err;
1135			}
1136
1137			dput(upperdentry);
1138			kfree(oip->redirect);
1139			goto out;
1140		}
1141
1142		/* Recalculate nlink for non-dir due to indexing */
1143		if (!is_dir)
1144			nlink = ovl_get_nlink(ofs, lowerdentry, upperdentry,
1145					      nlink);
1146		set_nlink(inode, nlink);
1147		ino = key->i_ino;
1148	} else {
1149		/* Lower hardlink that will be broken on copy up */
1150		inode = new_inode(sb);
1151		if (!inode) {
1152			err = -ENOMEM;
1153			goto out_err;
1154		}
1155		ino = realinode->i_ino;
1156		fsid = lowerpath->layer->fsid;
1157	}
1158	ovl_fill_inode(inode, realinode->i_mode, realinode->i_rdev);
1159	ovl_inode_init(inode, oip, ino, fsid);
1160
1161	if (upperdentry && ovl_is_impuredir(sb, upperdentry))
1162		ovl_set_flag(OVL_IMPURE, inode);
1163
1164	if (oip->index)
1165		ovl_set_flag(OVL_INDEX, inode);
1166
1167	OVL_I(inode)->redirect = oip->redirect;
1168
1169	if (bylower)
1170		ovl_set_flag(OVL_CONST_INO, inode);
1171
1172	/* Check for non-merge dir that may have whiteouts */
1173	if (is_dir) {
1174		if (((upperdentry && lowerdentry) || oip->numlower > 1) ||
1175		    ovl_check_origin_xattr(ofs, upperdentry ?: lowerdentry)) {
1176			ovl_set_flag(OVL_WHITEOUTS, inode);
1177		}
1178	}
1179
1180	/* Check for immutable/append-only inode flags in xattr */
1181	if (upperdentry)
1182		ovl_check_protattr(inode, upperdentry);
1183
1184	if (inode->i_state & I_NEW)
1185		unlock_new_inode(inode);
1186out:
1187	return inode;
1188
1189out_err:
1190	pr_warn_ratelimited("failed to get inode (%i)\n", err);
1191	inode = ERR_PTR(err);
1192	goto out;
1193}
1194