1 // SPDX-License-Identifier: GPL-2.0-only
4 * Copyright (C) 2011 Novell Inc.
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 <linux/posix_acl.h>
18 #include <linux/posix_acl_xattr.h>
19 #include "overlayfs.h"
22 int ovl_setattr(struct mnt_idmap *idmap, struct dentry *dentry,
26 struct ovl_fs *ofs = OVL_FS(dentry->d_sb);
27 bool full_copy_up = false;
28 struct dentry *upperdentry;
29 const struct cred *old_cred;
31 err = setattr_prepare(&nop_mnt_idmap, dentry, attr);
35 if (attr->ia_valid & ATTR_SIZE) {
36 /* Truncate should trigger data copy up as well */
41 err = ovl_copy_up(dentry);
43 err = ovl_copy_up_with_data(dentry);
45 struct inode *winode = NULL;
47 upperdentry = ovl_dentry_upper(dentry);
49 if (attr->ia_valid & ATTR_SIZE) {
50 winode = d_inode(upperdentry);
51 err = get_write_access(winode);
56 if (attr->ia_valid & (ATTR_KILL_SUID|ATTR_KILL_SGID))
57 attr->ia_valid &= ~ATTR_MODE;
60 * We might have to translate ovl file into real file object
61 * once use cases emerge. For now, simply don't let underlying
62 * filesystem rely on attr->ia_file
64 attr->ia_valid &= ~ATTR_FILE;
67 * If open(O_TRUNC) is done, VFS calls ->setattr with ATTR_OPEN
68 * set. Overlayfs does not pass O_TRUNC flag to underlying
69 * filesystem during open -> do not pass ATTR_OPEN. This
70 * disables optimization in fuse which assumes open(O_TRUNC)
71 * already set file size to 0. But we never passed O_TRUNC to
72 * fuse. So by clearing ATTR_OPEN, fuse will be forced to send
73 * setattr request to server.
75 attr->ia_valid &= ~ATTR_OPEN;
77 err = ovl_want_write(dentry);
81 inode_lock(upperdentry->d_inode);
82 old_cred = ovl_override_creds(dentry->d_sb);
83 err = ovl_do_notify_change(ofs, upperdentry, attr);
84 revert_creds(old_cred);
86 ovl_copyattr(dentry->d_inode);
87 inode_unlock(upperdentry->d_inode);
88 ovl_drop_write(dentry);
92 put_write_access(winode);
98 static void ovl_map_dev_ino(struct dentry *dentry, struct kstat *stat, int fsid)
100 struct ovl_fs *ofs = OVL_FS(dentry->d_sb);
101 bool samefs = ovl_same_fs(ofs);
102 unsigned int xinobits = ovl_xino_bits(ofs);
103 unsigned int xinoshift = 64 - xinobits;
107 * When all layers are on the same fs, all real inode
108 * number are unique, so we use the overlay st_dev,
109 * which is friendly to du -x.
111 stat->dev = dentry->d_sb->s_dev;
113 } else if (xinobits) {
115 * All inode numbers of underlying fs should not be using the
116 * high xinobits, so we use high xinobits to partition the
117 * overlay st_ino address space. The high bits holds the fsid
118 * (upper fsid is 0). The lowest xinobit is reserved for mapping
119 * the non-persistent inode numbers range in case of overflow.
120 * This way all overlay inode numbers are unique and use the
123 if (likely(!(stat->ino >> xinoshift))) {
124 stat->ino |= ((u64)fsid) << (xinoshift + 1);
125 stat->dev = dentry->d_sb->s_dev;
127 } else if (ovl_xino_warn(ofs)) {
128 pr_warn_ratelimited("inode number too big (%pd2, ino=%llu, xinobits=%d)\n",
129 dentry, stat->ino, xinobits);
133 /* The inode could not be mapped to a unified st_ino address space */
134 if (S_ISDIR(dentry->d_inode->i_mode)) {
136 * Always use the overlay st_dev for directories, so 'find
137 * -xdev' will scan the entire overlay mount and won't cross the
138 * overlay mount boundaries.
140 * If not all layers are on the same fs the pair {real st_ino;
141 * overlay st_dev} is not unique, so use the non persistent
142 * overlay st_ino for directories.
144 stat->dev = dentry->d_sb->s_dev;
145 stat->ino = dentry->d_inode->i_ino;
148 * For non-samefs setup, if we cannot map all layers st_ino
149 * to a unified address space, we need to make sure that st_dev
150 * is unique per underlying fs, so we use the unique anonymous
151 * bdev assigned to the underlying fs.
153 stat->dev = ofs->fs[fsid].pseudo_dev;
157 int ovl_getattr(struct mnt_idmap *idmap, const struct path *path,
158 struct kstat *stat, u32 request_mask, unsigned int flags)
160 struct dentry *dentry = path->dentry;
161 enum ovl_path_type type;
162 struct path realpath;
163 const struct cred *old_cred;
164 struct inode *inode = d_inode(dentry);
165 bool is_dir = S_ISDIR(inode->i_mode);
168 bool metacopy_blocks = false;
170 metacopy_blocks = ovl_is_metacopy_dentry(dentry);
172 type = ovl_path_real(dentry, &realpath);
173 old_cred = ovl_override_creds(dentry->d_sb);
174 err = ovl_do_getattr(&realpath, stat, request_mask, flags);
178 /* Report the effective immutable/append-only STATX flags */
179 generic_fill_statx_attr(inode, stat);
182 * For non-dir or same fs, we use st_ino of the copy up origin.
183 * This guaranties constant st_dev/st_ino across copy up.
184 * With xino feature and non-samefs, we use st_ino of the copy up
185 * origin masked with high bits that represent the layer id.
187 * If lower filesystem supports NFS file handles, this also guaranties
188 * persistent st_ino across mount cycle.
190 if (!is_dir || ovl_same_dev(OVL_FS(dentry->d_sb))) {
191 if (!OVL_TYPE_UPPER(type)) {
192 fsid = ovl_layer_lower(dentry)->fsid;
193 } else if (OVL_TYPE_ORIGIN(type)) {
194 struct kstat lowerstat;
195 u32 lowermask = STATX_INO | STATX_BLOCKS |
196 (!is_dir ? STATX_NLINK : 0);
198 ovl_path_lower(dentry, &realpath);
199 err = ovl_do_getattr(&realpath, &lowerstat, lowermask,
205 * Lower hardlinks may be broken on copy up to different
206 * upper files, so we cannot use the lower origin st_ino
207 * for those different files, even for the same fs case.
209 * Similarly, several redirected dirs can point to the
210 * same dir on a lower layer. With the "verify_lower"
211 * feature, we do not use the lower origin st_ino, if
212 * we haven't verified that this redirect is unique.
214 * With inodes index enabled, it is safe to use st_ino
215 * of an indexed origin. The index validates that the
216 * upper hardlink is not broken and that a redirected
217 * dir is the only redirect to that origin.
219 if (ovl_test_flag(OVL_INDEX, d_inode(dentry)) ||
220 (!ovl_verify_lower(dentry->d_sb) &&
221 (is_dir || lowerstat.nlink == 1))) {
222 fsid = ovl_layer_lower(dentry)->fsid;
223 stat->ino = lowerstat.ino;
227 * If we are querying a metacopy dentry and lower
228 * dentry is data dentry, then use the blocks we
229 * queried just now. We don't have to do additional
230 * vfs_getattr(). If lower itself is metacopy, then
231 * additional vfs_getattr() is unavoidable.
233 if (metacopy_blocks &&
234 realpath.dentry == ovl_dentry_lowerdata(dentry)) {
235 stat->blocks = lowerstat.blocks;
236 metacopy_blocks = false;
240 if (metacopy_blocks) {
242 * If lower is not same as lowerdata or if there was
243 * no origin on upper, we can end up here.
244 * With lazy lowerdata lookup, guess lowerdata blocks
245 * from size to avoid lowerdata lookup on stat(2).
247 struct kstat lowerdatastat;
248 u32 lowermask = STATX_BLOCKS;
250 ovl_path_lowerdata(dentry, &realpath);
251 if (realpath.dentry) {
252 err = ovl_do_getattr(&realpath, &lowerdatastat,
257 lowerdatastat.blocks =
258 round_up(stat->size, stat->blksize) >> 9;
260 stat->blocks = lowerdatastat.blocks;
264 ovl_map_dev_ino(dentry, stat, fsid);
267 * It's probably not worth it to count subdirs to get the
268 * correct link count. nlink=1 seems to pacify 'find' and
271 if (is_dir && OVL_TYPE_MERGE(type))
275 * Return the overlay inode nlinks for indexed upper inodes.
276 * Overlay inode nlink counts the union of the upper hardlinks
277 * and non-covered lower hardlinks. It does not include the upper
280 if (!is_dir && ovl_test_flag(OVL_INDEX, d_inode(dentry)))
281 stat->nlink = dentry->d_inode->i_nlink;
284 revert_creds(old_cred);
289 int ovl_permission(struct mnt_idmap *idmap,
290 struct inode *inode, int mask)
292 struct inode *upperinode = ovl_inode_upper(inode);
293 struct inode *realinode;
294 struct path realpath;
295 const struct cred *old_cred;
298 /* Careful in RCU walk mode */
299 realinode = ovl_i_path_real(inode, &realpath);
301 WARN_ON(!(mask & MAY_NOT_BLOCK));
306 * Check overlay inode with the creds of task and underlying inode
307 * with creds of mounter
309 err = generic_permission(&nop_mnt_idmap, inode, mask);
313 old_cred = ovl_override_creds(inode->i_sb);
315 !special_file(realinode->i_mode) && mask & MAY_WRITE) {
316 mask &= ~(MAY_WRITE | MAY_APPEND);
317 /* Make sure mounter can read file for copy up later */
320 err = inode_permission(mnt_idmap(realpath.mnt), realinode, mask);
321 revert_creds(old_cred);
326 static const char *ovl_get_link(struct dentry *dentry,
328 struct delayed_call *done)
330 const struct cred *old_cred;
334 return ERR_PTR(-ECHILD);
336 old_cred = ovl_override_creds(dentry->d_sb);
337 p = vfs_get_link(ovl_dentry_real(dentry), done);
338 revert_creds(old_cred);
342 #ifdef CONFIG_FS_POSIX_ACL
344 * Apply the idmapping of the layer to POSIX ACLs. The caller must pass a clone
345 * of the POSIX ACLs retrieved from the lower layer to this function to not
346 * alter the POSIX ACLs for the underlying filesystem.
348 static void ovl_idmap_posix_acl(const struct inode *realinode,
349 struct mnt_idmap *idmap,
350 struct posix_acl *acl)
352 struct user_namespace *fs_userns = i_user_ns(realinode);
354 for (unsigned int i = 0; i < acl->a_count; i++) {
358 struct posix_acl_entry *e = &acl->a_entries[i];
361 vfsuid = make_vfsuid(idmap, fs_userns, e->e_uid);
362 e->e_uid = vfsuid_into_kuid(vfsuid);
365 vfsgid = make_vfsgid(idmap, fs_userns, e->e_gid);
366 e->e_gid = vfsgid_into_kgid(vfsgid);
373 * The @noperm argument is used to skip permission checking and is a temporary
374 * measure. Quoting Miklos from an earlier discussion:
376 * > So there are two paths to getting an acl:
377 * > 1) permission checking and 2) retrieving the value via getxattr(2).
378 * > This is a similar situation as reading a symlink vs. following it.
379 * > When following a symlink overlayfs always reads the link on the
380 * > underlying fs just as if it was a readlink(2) call, calling
381 * > security_inode_readlink() instead of security_inode_follow_link().
382 * > This is logical: we are reading the link from the underlying storage,
383 * > and following it on overlayfs.
385 * > Applying the same logic to acl: we do need to call the
386 * > security_inode_getxattr() on the underlying fs, even if just want to
387 * > check permissions on overlay. This is currently not done, which is an
390 * > Maybe adding the check to ovl_get_acl() is the right way to go, but
391 * > I'm a little afraid of a performance regression. Will look into that.
393 * Until we have made a decision allow this helper to take the @noperm
394 * argument. We should hopefully be able to remove it soon.
396 struct posix_acl *ovl_get_acl_path(const struct path *path,
397 const char *acl_name, bool noperm)
399 struct posix_acl *real_acl, *clone;
400 struct mnt_idmap *idmap;
401 struct inode *realinode = d_inode(path->dentry);
403 idmap = mnt_idmap(path->mnt);
406 real_acl = get_inode_acl(realinode, posix_acl_type(acl_name));
408 real_acl = vfs_get_acl(idmap, path->dentry, acl_name);
409 if (IS_ERR_OR_NULL(real_acl))
412 if (!is_idmapped_mnt(path->mnt))
416 * We cannot alter the ACLs returned from the relevant layer as that
417 * would alter the cached values filesystem wide for the lower
418 * filesystem. Instead we can clone the ACLs and then apply the
419 * relevant idmapping of the layer.
421 clone = posix_acl_clone(real_acl, GFP_KERNEL);
422 posix_acl_release(real_acl); /* release original acl */
424 return ERR_PTR(-ENOMEM);
426 ovl_idmap_posix_acl(realinode, idmap, clone);
431 * When the relevant layer is an idmapped mount we need to take the idmapping
432 * of the layer into account and translate any ACL_{GROUP,USER} values
433 * according to the idmapped mount.
435 * We cannot alter the ACLs returned from the relevant layer as that would
436 * alter the cached values filesystem wide for the lower filesystem. Instead we
437 * can clone the ACLs and then apply the relevant idmapping of the layer.
439 * This is obviously only relevant when idmapped layers are used.
441 struct posix_acl *do_ovl_get_acl(struct mnt_idmap *idmap,
442 struct inode *inode, int type,
443 bool rcu, bool noperm)
445 struct inode *realinode;
446 struct posix_acl *acl;
447 struct path realpath;
449 /* Careful in RCU walk mode */
450 realinode = ovl_i_path_real(inode, &realpath);
453 return ERR_PTR(-ECHILD);
456 if (!IS_POSIXACL(realinode))
461 * If the layer is idmapped drop out of RCU path walk
462 * so we can clone the ACLs.
464 if (is_idmapped_mnt(realpath.mnt))
465 return ERR_PTR(-ECHILD);
467 acl = get_cached_acl_rcu(realinode, type);
469 const struct cred *old_cred;
471 old_cred = ovl_override_creds(inode->i_sb);
472 acl = ovl_get_acl_path(&realpath, posix_acl_xattr_name(type), noperm);
473 revert_creds(old_cred);
479 static int ovl_set_or_remove_acl(struct dentry *dentry, struct inode *inode,
480 struct posix_acl *acl, int type)
483 struct path realpath;
484 const char *acl_name;
485 const struct cred *old_cred;
486 struct ovl_fs *ofs = OVL_FS(dentry->d_sb);
487 struct dentry *upperdentry = ovl_dentry_upper(dentry);
488 struct dentry *realdentry = upperdentry ?: ovl_dentry_lower(dentry);
491 * If ACL is to be removed from a lower file, check if it exists in
492 * the first place before copying it up.
494 acl_name = posix_acl_xattr_name(type);
495 if (!acl && !upperdentry) {
496 struct posix_acl *real_acl;
498 ovl_path_lower(dentry, &realpath);
499 old_cred = ovl_override_creds(dentry->d_sb);
500 real_acl = vfs_get_acl(mnt_idmap(realpath.mnt), realdentry,
502 revert_creds(old_cred);
503 if (IS_ERR(real_acl)) {
504 err = PTR_ERR(real_acl);
507 posix_acl_release(real_acl);
511 err = ovl_copy_up(dentry);
515 realdentry = ovl_dentry_upper(dentry);
518 err = ovl_want_write(dentry);
522 old_cred = ovl_override_creds(dentry->d_sb);
524 err = ovl_do_set_acl(ofs, realdentry, acl_name, acl);
526 err = ovl_do_remove_acl(ofs, realdentry, acl_name);
527 revert_creds(old_cred);
528 ovl_drop_write(dentry);
536 int ovl_set_acl(struct mnt_idmap *idmap, struct dentry *dentry,
537 struct posix_acl *acl, int type)
540 struct inode *inode = d_inode(dentry);
541 struct dentry *workdir = ovl_workdir(dentry);
542 struct inode *realinode = ovl_inode_real(inode);
544 if (!IS_POSIXACL(d_inode(workdir)))
546 if (!realinode->i_op->set_acl)
548 if (type == ACL_TYPE_DEFAULT && !S_ISDIR(inode->i_mode))
549 return acl ? -EACCES : 0;
550 if (!inode_owner_or_capable(&nop_mnt_idmap, inode))
554 * Check if sgid bit needs to be cleared (actual setacl operation will
555 * be done with mounter's capabilities and so that won't do it for us).
557 if (unlikely(inode->i_mode & S_ISGID) && type == ACL_TYPE_ACCESS &&
558 !in_group_p(inode->i_gid) &&
559 !capable_wrt_inode_uidgid(&nop_mnt_idmap, inode, CAP_FSETID)) {
560 struct iattr iattr = { .ia_valid = ATTR_KILL_SGID };
562 err = ovl_setattr(&nop_mnt_idmap, dentry, &iattr);
567 return ovl_set_or_remove_acl(dentry, inode, acl, type);
571 int ovl_update_time(struct inode *inode, int flags)
573 if (flags & S_ATIME) {
574 struct ovl_fs *ofs = OVL_FS(inode->i_sb);
575 struct path upperpath = {
576 .mnt = ovl_upper_mnt(ofs),
577 .dentry = ovl_upperdentry_dereference(OVL_I(inode)),
580 if (upperpath.dentry) {
581 touch_atime(&upperpath);
582 inode_set_atime_to_ts(inode,
583 inode_get_atime(d_inode(upperpath.dentry)));
589 static int ovl_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
593 struct inode *realinode = ovl_inode_realdata(inode);
594 const struct cred *old_cred;
599 if (!realinode->i_op->fiemap)
602 old_cred = ovl_override_creds(inode->i_sb);
603 err = realinode->i_op->fiemap(realinode, fieinfo, start, len);
604 revert_creds(old_cred);
610 * Work around the fact that security_file_ioctl() takes a file argument.
611 * Introducing security_inode_fileattr_get/set() hooks would solve this issue
614 static int ovl_security_fileattr(const struct path *realpath, struct fileattr *fa,
621 file = dentry_open(realpath, O_RDONLY, current_cred());
623 return PTR_ERR(file);
626 cmd = fa->fsx_valid ? FS_IOC_FSSETXATTR : FS_IOC_SETFLAGS;
628 cmd = fa->fsx_valid ? FS_IOC_FSGETXATTR : FS_IOC_GETFLAGS;
630 err = security_file_ioctl(file, cmd, 0);
636 int ovl_real_fileattr_set(const struct path *realpath, struct fileattr *fa)
640 err = ovl_security_fileattr(realpath, fa, true);
644 return vfs_fileattr_set(mnt_idmap(realpath->mnt), realpath->dentry, fa);
647 int ovl_fileattr_set(struct mnt_idmap *idmap,
648 struct dentry *dentry, struct fileattr *fa)
650 struct inode *inode = d_inode(dentry);
651 struct path upperpath;
652 const struct cred *old_cred;
656 err = ovl_copy_up(dentry);
658 ovl_path_real(dentry, &upperpath);
660 err = ovl_want_write(dentry);
664 old_cred = ovl_override_creds(inode->i_sb);
666 * Store immutable/append-only flags in xattr and clear them
667 * in upper fileattr (in case they were set by older kernel)
668 * so children of "ovl-immutable" directories lower aliases of
669 * "ovl-immutable" hardlinks could be copied up.
670 * Clear xattr when flags are cleared.
672 err = ovl_set_protattr(inode, upperpath.dentry, fa);
674 err = ovl_real_fileattr_set(&upperpath, fa);
675 revert_creds(old_cred);
676 ovl_drop_write(dentry);
679 * Merge real inode flags with inode flags read from
680 * overlay.protattr xattr
682 flags = ovl_inode_real(inode)->i_flags & OVL_COPY_I_FLAGS_MASK;
684 BUILD_BUG_ON(OVL_PROT_I_FLAGS_MASK & ~OVL_COPY_I_FLAGS_MASK);
685 flags |= inode->i_flags & OVL_PROT_I_FLAGS_MASK;
686 inode_set_flags(inode, flags, OVL_COPY_I_FLAGS_MASK);
695 /* Convert inode protection flags to fileattr flags */
696 static void ovl_fileattr_prot_flags(struct inode *inode, struct fileattr *fa)
698 BUILD_BUG_ON(OVL_PROT_FS_FLAGS_MASK & ~FS_COMMON_FL);
699 BUILD_BUG_ON(OVL_PROT_FSX_FLAGS_MASK & ~FS_XFLAG_COMMON);
701 if (inode->i_flags & S_APPEND) {
702 fa->flags |= FS_APPEND_FL;
703 fa->fsx_xflags |= FS_XFLAG_APPEND;
705 if (inode->i_flags & S_IMMUTABLE) {
706 fa->flags |= FS_IMMUTABLE_FL;
707 fa->fsx_xflags |= FS_XFLAG_IMMUTABLE;
711 int ovl_real_fileattr_get(const struct path *realpath, struct fileattr *fa)
715 err = ovl_security_fileattr(realpath, fa, false);
719 err = vfs_fileattr_get(realpath->dentry, fa);
720 if (err == -ENOIOCTLCMD)
725 int ovl_fileattr_get(struct dentry *dentry, struct fileattr *fa)
727 struct inode *inode = d_inode(dentry);
728 struct path realpath;
729 const struct cred *old_cred;
732 ovl_path_real(dentry, &realpath);
734 old_cred = ovl_override_creds(inode->i_sb);
735 err = ovl_real_fileattr_get(&realpath, fa);
736 ovl_fileattr_prot_flags(inode, fa);
737 revert_creds(old_cred);
742 static const struct inode_operations ovl_file_inode_operations = {
743 .setattr = ovl_setattr,
744 .permission = ovl_permission,
745 .getattr = ovl_getattr,
746 .listxattr = ovl_listxattr,
747 .get_inode_acl = ovl_get_inode_acl,
748 .get_acl = ovl_get_acl,
749 .set_acl = ovl_set_acl,
750 .update_time = ovl_update_time,
751 .fiemap = ovl_fiemap,
752 .fileattr_get = ovl_fileattr_get,
753 .fileattr_set = ovl_fileattr_set,
756 static const struct inode_operations ovl_symlink_inode_operations = {
757 .setattr = ovl_setattr,
758 .get_link = ovl_get_link,
759 .getattr = ovl_getattr,
760 .listxattr = ovl_listxattr,
761 .update_time = ovl_update_time,
764 static const struct inode_operations ovl_special_inode_operations = {
765 .setattr = ovl_setattr,
766 .permission = ovl_permission,
767 .getattr = ovl_getattr,
768 .listxattr = ovl_listxattr,
769 .get_inode_acl = ovl_get_inode_acl,
770 .get_acl = ovl_get_acl,
771 .set_acl = ovl_set_acl,
772 .update_time = ovl_update_time,
775 static const struct address_space_operations ovl_aops = {
776 /* For O_DIRECT dentry_open() checks f_mapping->a_ops->direct_IO */
777 .direct_IO = noop_direct_IO,
781 * It is possible to stack overlayfs instance on top of another
782 * overlayfs instance as lower layer. We need to annotate the
783 * stackable i_mutex locks according to stack level of the super
784 * block instance. An overlayfs instance can never be in stack
785 * depth 0 (there is always a real fs below it). An overlayfs
786 * inode lock will use the lockdep annotation ovl_i_mutex_key[depth].
788 * For example, here is a snip from /proc/lockdep_chains after
789 * dir_iterate of nested overlayfs:
791 * [...] &ovl_i_mutex_dir_key[depth] (stack_depth=2)
792 * [...] &ovl_i_mutex_dir_key[depth]#2 (stack_depth=1)
793 * [...] &type->i_mutex_dir_key (stack_depth=0)
795 * Locking order w.r.t ovl_want_write() is important for nested overlayfs.
797 * This chain is valid:
798 * - inode->i_rwsem (inode_lock[2])
799 * - upper_mnt->mnt_sb->s_writers (ovl_want_write[0])
800 * - OVL_I(inode)->lock (ovl_inode_lock[2])
801 * - OVL_I(lowerinode)->lock (ovl_inode_lock[1])
803 * And this chain is valid:
804 * - inode->i_rwsem (inode_lock[2])
805 * - OVL_I(inode)->lock (ovl_inode_lock[2])
806 * - lowerinode->i_rwsem (inode_lock[1])
807 * - OVL_I(lowerinode)->lock (ovl_inode_lock[1])
809 * But lowerinode->i_rwsem SHOULD NOT be acquired while ovl_want_write() is
810 * held, because it is in reverse order of the non-nested case using the same
812 * - inode->i_rwsem (inode_lock[1])
813 * - upper_mnt->mnt_sb->s_writers (ovl_want_write[0])
814 * - OVL_I(inode)->lock (ovl_inode_lock[1])
816 #define OVL_MAX_NESTING FILESYSTEM_MAX_STACK_DEPTH
818 static inline void ovl_lockdep_annotate_inode_mutex_key(struct inode *inode)
820 #ifdef CONFIG_LOCKDEP
821 static struct lock_class_key ovl_i_mutex_key[OVL_MAX_NESTING];
822 static struct lock_class_key ovl_i_mutex_dir_key[OVL_MAX_NESTING];
823 static struct lock_class_key ovl_i_lock_key[OVL_MAX_NESTING];
825 int depth = inode->i_sb->s_stack_depth - 1;
827 if (WARN_ON_ONCE(depth < 0 || depth >= OVL_MAX_NESTING))
830 if (S_ISDIR(inode->i_mode))
831 lockdep_set_class(&inode->i_rwsem, &ovl_i_mutex_dir_key[depth]);
833 lockdep_set_class(&inode->i_rwsem, &ovl_i_mutex_key[depth]);
835 lockdep_set_class(&OVL_I(inode)->lock, &ovl_i_lock_key[depth]);
839 static void ovl_next_ino(struct inode *inode)
841 struct ovl_fs *ofs = OVL_FS(inode->i_sb);
843 inode->i_ino = atomic_long_inc_return(&ofs->last_ino);
844 if (unlikely(!inode->i_ino))
845 inode->i_ino = atomic_long_inc_return(&ofs->last_ino);
848 static void ovl_map_ino(struct inode *inode, unsigned long ino, int fsid)
850 struct ovl_fs *ofs = OVL_FS(inode->i_sb);
851 int xinobits = ovl_xino_bits(ofs);
852 unsigned int xinoshift = 64 - xinobits;
855 * When d_ino is consistent with st_ino (samefs or i_ino has enough
856 * bits to encode layer), set the same value used for st_ino to i_ino,
857 * so inode number exposed via /proc/locks and a like will be
858 * consistent with d_ino and st_ino values. An i_ino value inconsistent
859 * with d_ino also causes nfsd readdirplus to fail.
862 if (ovl_same_fs(ofs)) {
864 } else if (xinobits && likely(!(ino >> xinoshift))) {
865 inode->i_ino |= (unsigned long)fsid << (xinoshift + 1);
870 * For directory inodes on non-samefs with xino disabled or xino
871 * overflow, we allocate a non-persistent inode number, to be used for
872 * resolving st_ino collisions in ovl_map_dev_ino().
874 * To avoid ino collision with legitimate xino values from upper
875 * layer (fsid 0), use the lowest xinobit to map the non
876 * persistent inode numbers to the unified st_ino address space.
878 if (S_ISDIR(inode->i_mode)) {
881 inode->i_ino &= ~0UL >> xinobits;
882 inode->i_ino |= 1UL << xinoshift;
887 void ovl_inode_init(struct inode *inode, struct ovl_inode_params *oip,
888 unsigned long ino, int fsid)
890 struct inode *realinode;
891 struct ovl_inode *oi = OVL_I(inode);
893 oi->__upperdentry = oip->upperdentry;
895 oi->redirect = oip->redirect;
896 oi->lowerdata_redirect = oip->lowerdata_redirect;
898 realinode = ovl_inode_real(inode);
900 ovl_copyflags(realinode, inode);
901 ovl_map_ino(inode, ino, fsid);
904 static void ovl_fill_inode(struct inode *inode, umode_t mode, dev_t rdev)
906 inode->i_mode = mode;
907 inode->i_flags |= S_NOCMTIME;
908 #ifdef CONFIG_FS_POSIX_ACL
909 inode->i_acl = inode->i_default_acl = ACL_DONT_CACHE;
912 ovl_lockdep_annotate_inode_mutex_key(inode);
914 switch (mode & S_IFMT) {
916 inode->i_op = &ovl_file_inode_operations;
917 inode->i_fop = &ovl_file_operations;
918 inode->i_mapping->a_ops = &ovl_aops;
922 inode->i_op = &ovl_dir_inode_operations;
923 inode->i_fop = &ovl_dir_operations;
927 inode->i_op = &ovl_symlink_inode_operations;
931 inode->i_op = &ovl_special_inode_operations;
932 init_special_inode(inode, mode, rdev);
938 * With inodes index enabled, an overlay inode nlink counts the union of upper
939 * hardlinks and non-covered lower hardlinks. During the lifetime of a non-pure
940 * upper inode, the following nlink modifying operations can happen:
942 * 1. Lower hardlink copy up
943 * 2. Upper hardlink created, unlinked or renamed over
944 * 3. Lower hardlink whiteout or renamed over
946 * For the first, copy up case, the union nlink does not change, whether the
947 * operation succeeds or fails, but the upper inode nlink may change.
948 * Therefore, before copy up, we store the union nlink value relative to the
949 * lower inode nlink in the index inode xattr .overlay.nlink.
951 * For the second, upper hardlink case, the union nlink should be incremented
952 * or decremented IFF the operation succeeds, aligned with nlink change of the
953 * upper inode. Therefore, before link/unlink/rename, we store the union nlink
954 * value relative to the upper inode nlink in the index inode.
956 * For the last, lower cover up case, we simplify things by preceding the
957 * whiteout or cover up with copy up. This makes sure that there is an index
958 * upper inode where the nlink xattr can be stored before the copied up upper
961 #define OVL_NLINK_ADD_UPPER (1 << 0)
964 * On-disk format for indexed nlink:
966 * nlink relative to the upper inode - "U[+-]NUM"
967 * nlink relative to the lower inode - "L[+-]NUM"
970 static int ovl_set_nlink_common(struct dentry *dentry,
971 struct dentry *realdentry, const char *format)
973 struct inode *inode = d_inode(dentry);
974 struct inode *realinode = d_inode(realdentry);
978 len = snprintf(buf, sizeof(buf), format,
979 (int) (inode->i_nlink - realinode->i_nlink));
981 if (WARN_ON(len >= sizeof(buf)))
984 return ovl_setxattr(OVL_FS(inode->i_sb), ovl_dentry_upper(dentry),
985 OVL_XATTR_NLINK, buf, len);
988 int ovl_set_nlink_upper(struct dentry *dentry)
990 return ovl_set_nlink_common(dentry, ovl_dentry_upper(dentry), "U%+i");
993 int ovl_set_nlink_lower(struct dentry *dentry)
995 return ovl_set_nlink_common(dentry, ovl_dentry_lower(dentry), "L%+i");
998 unsigned int ovl_get_nlink(struct ovl_fs *ofs, struct dentry *lowerdentry,
999 struct dentry *upperdentry,
1000 unsigned int fallback)
1007 if (!lowerdentry || !upperdentry || d_inode(lowerdentry)->i_nlink == 1)
1010 err = ovl_getxattr_upper(ofs, upperdentry, OVL_XATTR_NLINK,
1011 &buf, sizeof(buf) - 1);
1016 if ((buf[0] != 'L' && buf[0] != 'U') ||
1017 (buf[1] != '+' && buf[1] != '-'))
1020 err = kstrtoint(buf + 1, 10, &nlink_diff);
1024 nlink = d_inode(buf[0] == 'L' ? lowerdentry : upperdentry)->i_nlink;
1025 nlink += nlink_diff;
1033 pr_warn_ratelimited("failed to get index nlink (%pd2, err=%i)\n",
1038 struct inode *ovl_new_inode(struct super_block *sb, umode_t mode, dev_t rdev)
1040 struct inode *inode;
1042 inode = new_inode(sb);
1044 ovl_fill_inode(inode, mode, rdev);
1049 static int ovl_inode_test(struct inode *inode, void *data)
1051 return inode->i_private == data;
1054 static int ovl_inode_set(struct inode *inode, void *data)
1056 inode->i_private = data;
1060 static bool ovl_verify_inode(struct inode *inode, struct dentry *lowerdentry,
1061 struct dentry *upperdentry, bool strict)
1064 * For directories, @strict verify from lookup path performs consistency
1065 * checks, so NULL lower/upper in dentry must match NULL lower/upper in
1066 * inode. Non @strict verify from NFS handle decode path passes NULL for
1067 * 'unknown' lower/upper.
1069 if (S_ISDIR(inode->i_mode) && strict) {
1070 /* Real lower dir moved to upper layer under us? */
1071 if (!lowerdentry && ovl_inode_lower(inode))
1074 /* Lookup of an uncovered redirect origin? */
1075 if (!upperdentry && ovl_inode_upper(inode))
1080 * Allow non-NULL lower inode in ovl_inode even if lowerdentry is NULL.
1081 * This happens when finding a copied up overlay inode for a renamed
1082 * or hardlinked overlay dentry and lower dentry cannot be followed
1083 * by origin because lower fs does not support file handles.
1085 if (lowerdentry && ovl_inode_lower(inode) != d_inode(lowerdentry))
1089 * Allow non-NULL __upperdentry in inode even if upperdentry is NULL.
1090 * This happens when finding a lower alias for a copied up hard link.
1092 if (upperdentry && ovl_inode_upper(inode) != d_inode(upperdentry))
1098 struct inode *ovl_lookup_inode(struct super_block *sb, struct dentry *real,
1101 struct inode *inode, *key = d_inode(real);
1103 inode = ilookup5(sb, (unsigned long) key, ovl_inode_test, key);
1107 if (!ovl_verify_inode(inode, is_upper ? NULL : real,
1108 is_upper ? real : NULL, false)) {
1110 return ERR_PTR(-ESTALE);
1116 bool ovl_lookup_trap_inode(struct super_block *sb, struct dentry *dir)
1118 struct inode *key = d_inode(dir);
1122 trap = ilookup5(sb, (unsigned long) key, ovl_inode_test, key);
1126 res = IS_DEADDIR(trap) && !ovl_inode_upper(trap) &&
1127 !ovl_inode_lower(trap);
1134 * Create an inode cache entry for layer root dir, that will intentionally
1135 * fail ovl_verify_inode(), so any lookup that will find some layer root
1138 struct inode *ovl_get_trap_inode(struct super_block *sb, struct dentry *dir)
1140 struct inode *key = d_inode(dir);
1144 return ERR_PTR(-ENOTDIR);
1146 trap = iget5_locked(sb, (unsigned long) key, ovl_inode_test,
1147 ovl_inode_set, key);
1149 return ERR_PTR(-ENOMEM);
1151 if (!(trap->i_state & I_NEW)) {
1152 /* Conflicting layer roots? */
1154 return ERR_PTR(-ELOOP);
1157 trap->i_mode = S_IFDIR;
1158 trap->i_flags = S_DEAD;
1159 unlock_new_inode(trap);
1165 * Does overlay inode need to be hashed by lower inode?
1167 static bool ovl_hash_bylower(struct super_block *sb, struct dentry *upper,
1168 struct dentry *lower, bool index)
1170 struct ovl_fs *ofs = OVL_FS(sb);
1172 /* No, if pure upper */
1176 /* Yes, if already indexed */
1180 /* Yes, if won't be copied up */
1181 if (!ovl_upper_mnt(ofs))
1184 /* No, if lower hardlink is or will be broken on copy up */
1185 if ((upper || !ovl_indexdir(sb)) &&
1186 !d_is_dir(lower) && d_inode(lower)->i_nlink > 1)
1189 /* No, if non-indexed upper with NFS export */
1190 if (ofs->config.nfs_export && upper)
1193 /* Otherwise, hash by lower inode for fsnotify */
1197 static struct inode *ovl_iget5(struct super_block *sb, struct inode *newinode,
1200 return newinode ? inode_insert5(newinode, (unsigned long) key,
1201 ovl_inode_test, ovl_inode_set, key) :
1202 iget5_locked(sb, (unsigned long) key,
1203 ovl_inode_test, ovl_inode_set, key);
1206 struct inode *ovl_get_inode(struct super_block *sb,
1207 struct ovl_inode_params *oip)
1209 struct ovl_fs *ofs = OVL_FS(sb);
1210 struct dentry *upperdentry = oip->upperdentry;
1211 struct ovl_path *lowerpath = ovl_lowerpath(oip->oe);
1212 struct inode *realinode = upperdentry ? d_inode(upperdentry) : NULL;
1213 struct inode *inode;
1214 struct dentry *lowerdentry = lowerpath ? lowerpath->dentry : NULL;
1215 struct path realpath = {
1216 .dentry = upperdentry ?: lowerdentry,
1217 .mnt = upperdentry ? ovl_upper_mnt(ofs) : lowerpath->layer->mnt,
1219 bool bylower = ovl_hash_bylower(sb, upperdentry, lowerdentry,
1221 int fsid = bylower ? lowerpath->layer->fsid : 0;
1223 unsigned long ino = 0;
1224 int err = oip->newinode ? -EEXIST : -ENOMEM;
1227 realinode = d_inode(lowerdentry);
1230 * Copy up origin (lower) may exist for non-indexed upper, but we must
1231 * not use lower as hash key if this is a broken hardlink.
1233 is_dir = S_ISDIR(realinode->i_mode);
1234 if (upperdentry || bylower) {
1235 struct inode *key = d_inode(bylower ? lowerdentry :
1237 unsigned int nlink = is_dir ? 1 : realinode->i_nlink;
1239 inode = ovl_iget5(sb, oip->newinode, key);
1242 if (!(inode->i_state & I_NEW)) {
1244 * Verify that the underlying files stored in the inode
1245 * match those in the dentry.
1247 if (!ovl_verify_inode(inode, lowerdentry, upperdentry,
1255 ovl_free_entry(oip->oe);
1256 kfree(oip->redirect);
1257 kfree(oip->lowerdata_redirect);
1261 /* Recalculate nlink for non-dir due to indexing */
1263 nlink = ovl_get_nlink(ofs, lowerdentry, upperdentry,
1265 set_nlink(inode, nlink);
1268 /* Lower hardlink that will be broken on copy up */
1269 inode = new_inode(sb);
1274 ino = realinode->i_ino;
1275 fsid = lowerpath->layer->fsid;
1277 ovl_fill_inode(inode, realinode->i_mode, realinode->i_rdev);
1278 ovl_inode_init(inode, oip, ino, fsid);
1280 if (upperdentry && ovl_is_impuredir(sb, upperdentry))
1281 ovl_set_flag(OVL_IMPURE, inode);
1284 ovl_set_flag(OVL_INDEX, inode);
1287 ovl_set_flag(OVL_CONST_INO, inode);
1289 /* Check for non-merge dir that may have whiteouts */
1291 if (((upperdentry && lowerdentry) || ovl_numlower(oip->oe) > 1) ||
1292 ovl_path_check_origin_xattr(ofs, &realpath)) {
1293 ovl_set_flag(OVL_WHITEOUTS, inode);
1297 /* Check for immutable/append-only inode flags in xattr */
1299 ovl_check_protattr(inode, upperdentry);
1301 if (inode->i_state & I_NEW)
1302 unlock_new_inode(inode);
1307 pr_warn_ratelimited("failed to get inode (%i)\n", err);
1308 inode = ERR_PTR(err);