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/ratelimit.h>
12 #include <linux/fiemap.h>
13 #include <linux/fileattr.h>
14 #include <linux/security.h>
15 #include <linux/namei.h>
16 #include <linux/posix_acl.h>
17 #include <linux/posix_acl_xattr.h>
18 #include "overlayfs.h"
21 int ovl_setattr(struct mnt_idmap *idmap, struct dentry *dentry,
25 struct ovl_fs *ofs = OVL_FS(dentry->d_sb);
26 bool full_copy_up = false;
27 struct dentry *upperdentry;
28 const struct cred *old_cred;
30 err = setattr_prepare(&nop_mnt_idmap, dentry, attr);
34 if (attr->ia_valid & ATTR_SIZE) {
35 /* Truncate should trigger data copy up as well */
40 err = ovl_copy_up(dentry);
42 err = ovl_copy_up_with_data(dentry);
44 struct inode *winode = NULL;
46 upperdentry = ovl_dentry_upper(dentry);
48 if (attr->ia_valid & ATTR_SIZE) {
49 winode = d_inode(upperdentry);
50 err = get_write_access(winode);
55 if (attr->ia_valid & (ATTR_KILL_SUID|ATTR_KILL_SGID))
56 attr->ia_valid &= ~ATTR_MODE;
59 * We might have to translate ovl file into real file object
60 * once use cases emerge. For now, simply don't let underlying
61 * filesystem rely on attr->ia_file
63 attr->ia_valid &= ~ATTR_FILE;
66 * If open(O_TRUNC) is done, VFS calls ->setattr with ATTR_OPEN
67 * set. Overlayfs does not pass O_TRUNC flag to underlying
68 * filesystem during open -> do not pass ATTR_OPEN. This
69 * disables optimization in fuse which assumes open(O_TRUNC)
70 * already set file size to 0. But we never passed O_TRUNC to
71 * fuse. So by clearing ATTR_OPEN, fuse will be forced to send
72 * setattr request to server.
74 attr->ia_valid &= ~ATTR_OPEN;
76 err = ovl_want_write(dentry);
80 inode_lock(upperdentry->d_inode);
81 old_cred = ovl_override_creds(dentry->d_sb);
82 err = ovl_do_notify_change(ofs, upperdentry, attr);
83 ovl_revert_creds(old_cred);
85 ovl_copyattr(dentry->d_inode);
86 inode_unlock(upperdentry->d_inode);
87 ovl_drop_write(dentry);
91 put_write_access(winode);
97 static void ovl_map_dev_ino(struct dentry *dentry, struct kstat *stat, int fsid)
99 struct ovl_fs *ofs = OVL_FS(dentry->d_sb);
100 bool samefs = ovl_same_fs(ofs);
101 unsigned int xinobits = ovl_xino_bits(ofs);
102 unsigned int xinoshift = 64 - xinobits;
106 * When all layers are on the same fs, all real inode
107 * number are unique, so we use the overlay st_dev,
108 * which is friendly to du -x.
110 stat->dev = dentry->d_sb->s_dev;
112 } else if (xinobits) {
114 * All inode numbers of underlying fs should not be using the
115 * high xinobits, so we use high xinobits to partition the
116 * overlay st_ino address space. The high bits holds the fsid
117 * (upper fsid is 0). The lowest xinobit is reserved for mapping
118 * the non-persistent inode numbers range in case of overflow.
119 * This way all overlay inode numbers are unique and use the
122 if (likely(!(stat->ino >> xinoshift))) {
123 stat->ino |= ((u64)fsid) << (xinoshift + 1);
124 stat->dev = dentry->d_sb->s_dev;
126 } else if (ovl_xino_warn(ofs)) {
127 pr_warn_ratelimited("inode number too big (%pd2, ino=%llu, xinobits=%d)\n",
128 dentry, stat->ino, xinobits);
132 /* The inode could not be mapped to a unified st_ino address space */
133 if (S_ISDIR(dentry->d_inode->i_mode)) {
135 * Always use the overlay st_dev for directories, so 'find
136 * -xdev' will scan the entire overlay mount and won't cross the
137 * overlay mount boundaries.
139 * If not all layers are on the same fs the pair {real st_ino;
140 * overlay st_dev} is not unique, so use the non persistent
141 * overlay st_ino for directories.
143 stat->dev = dentry->d_sb->s_dev;
144 stat->ino = dentry->d_inode->i_ino;
147 * For non-samefs setup, if we cannot map all layers st_ino
148 * to a unified address space, we need to make sure that st_dev
149 * is unique per underlying fs, so we use the unique anonymous
150 * bdev assigned to the underlying fs.
152 stat->dev = ofs->fs[fsid].pseudo_dev;
156 int ovl_getattr(struct mnt_idmap *idmap, const struct path *path,
157 struct kstat *stat, u32 request_mask, unsigned int flags)
159 struct dentry *dentry = path->dentry;
160 enum ovl_path_type type;
161 struct path realpath;
162 const struct cred *old_cred;
163 struct inode *inode = d_inode(dentry);
164 bool is_dir = S_ISDIR(inode->i_mode);
167 bool metacopy_blocks = false;
169 metacopy_blocks = ovl_is_metacopy_dentry(dentry);
171 type = ovl_path_real(dentry, &realpath);
172 old_cred = ovl_override_creds(dentry->d_sb);
173 err = vfs_getattr_nosec(&realpath, stat, request_mask, flags);
177 /* Report the effective immutable/append-only STATX flags */
178 generic_fill_statx_attr(inode, stat);
181 * For non-dir or same fs, we use st_ino of the copy up origin.
182 * This guaranties constant st_dev/st_ino across copy up.
183 * With xino feature and non-samefs, we use st_ino of the copy up
184 * origin masked with high bits that represent the layer id.
186 * If lower filesystem supports NFS file handles, this also guaranties
187 * persistent st_ino across mount cycle.
189 if (!is_dir || ovl_same_dev(OVL_FS(dentry->d_sb))) {
190 if (!OVL_TYPE_UPPER(type)) {
191 fsid = ovl_layer_lower(dentry)->fsid;
192 } else if (OVL_TYPE_ORIGIN(type)) {
193 struct kstat lowerstat;
194 u32 lowermask = STATX_INO | STATX_BLOCKS |
195 (!is_dir ? STATX_NLINK : 0);
197 ovl_path_lower(dentry, &realpath);
198 err = vfs_getattr_nosec(&realpath, &lowerstat, lowermask,
204 * Lower hardlinks may be broken on copy up to different
205 * upper files, so we cannot use the lower origin st_ino
206 * for those different files, even for the same fs case.
208 * Similarly, several redirected dirs can point to the
209 * same dir on a lower layer. With the "verify_lower"
210 * feature, we do not use the lower origin st_ino, if
211 * we haven't verified that this redirect is unique.
213 * With inodes index enabled, it is safe to use st_ino
214 * of an indexed origin. The index validates that the
215 * upper hardlink is not broken and that a redirected
216 * dir is the only redirect to that origin.
218 if (ovl_test_flag(OVL_INDEX, d_inode(dentry)) ||
219 (!ovl_verify_lower(dentry->d_sb) &&
220 (is_dir || lowerstat.nlink == 1))) {
221 fsid = ovl_layer_lower(dentry)->fsid;
222 stat->ino = lowerstat.ino;
226 * If we are querying a metacopy dentry and lower
227 * dentry is data dentry, then use the blocks we
228 * queried just now. We don't have to do additional
229 * vfs_getattr(). If lower itself is metacopy, then
230 * additional vfs_getattr() is unavoidable.
232 if (metacopy_blocks &&
233 realpath.dentry == ovl_dentry_lowerdata(dentry)) {
234 stat->blocks = lowerstat.blocks;
235 metacopy_blocks = false;
239 if (metacopy_blocks) {
241 * If lower is not same as lowerdata or if there was
242 * no origin on upper, we can end up here.
243 * With lazy lowerdata lookup, guess lowerdata blocks
244 * from size to avoid lowerdata lookup on stat(2).
246 struct kstat lowerdatastat;
247 u32 lowermask = STATX_BLOCKS;
249 ovl_path_lowerdata(dentry, &realpath);
250 if (realpath.dentry) {
251 err = vfs_getattr_nosec(&realpath, &lowerdatastat,
256 lowerdatastat.blocks =
257 round_up(stat->size, stat->blksize) >> 9;
259 stat->blocks = lowerdatastat.blocks;
263 ovl_map_dev_ino(dentry, stat, fsid);
266 * It's probably not worth it to count subdirs to get the
267 * correct link count. nlink=1 seems to pacify 'find' and
270 if (is_dir && OVL_TYPE_MERGE(type))
274 * Return the overlay inode nlinks for indexed upper inodes.
275 * Overlay inode nlink counts the union of the upper hardlinks
276 * and non-covered lower hardlinks. It does not include the upper
279 if (!is_dir && ovl_test_flag(OVL_INDEX, d_inode(dentry)))
280 stat->nlink = dentry->d_inode->i_nlink;
283 ovl_revert_creds(old_cred);
288 int ovl_permission(struct mnt_idmap *idmap,
289 struct inode *inode, int mask)
291 struct inode *upperinode = ovl_inode_upper(inode);
292 struct inode *realinode;
293 struct path realpath;
294 const struct cred *old_cred;
297 /* Careful in RCU walk mode */
298 realinode = ovl_i_path_real(inode, &realpath);
300 WARN_ON(!(mask & MAY_NOT_BLOCK));
305 * Check overlay inode with the creds of task and underlying inode
306 * with creds of mounter
308 err = generic_permission(&nop_mnt_idmap, inode, mask);
312 old_cred = ovl_override_creds(inode->i_sb);
314 !special_file(realinode->i_mode) && mask & MAY_WRITE) {
315 mask &= ~(MAY_WRITE | MAY_APPEND);
316 /* Make sure mounter can read file for copy up later */
319 err = inode_permission(mnt_idmap(realpath.mnt), realinode, mask);
320 ovl_revert_creds(old_cred);
325 static const char *ovl_get_link(struct dentry *dentry,
327 struct delayed_call *done)
329 const struct cred *old_cred;
333 return ERR_PTR(-ECHILD);
335 old_cred = ovl_override_creds(dentry->d_sb);
336 p = vfs_get_link(ovl_dentry_real(dentry), done);
337 ovl_revert_creds(old_cred);
341 #ifdef CONFIG_FS_POSIX_ACL
343 * Apply the idmapping of the layer to POSIX ACLs. The caller must pass a clone
344 * of the POSIX ACLs retrieved from the lower layer to this function to not
345 * alter the POSIX ACLs for the underlying filesystem.
347 static void ovl_idmap_posix_acl(const struct inode *realinode,
348 struct mnt_idmap *idmap,
349 struct posix_acl *acl)
351 struct user_namespace *fs_userns = i_user_ns(realinode);
353 for (unsigned int i = 0; i < acl->a_count; i++) {
357 struct posix_acl_entry *e = &acl->a_entries[i];
360 vfsuid = make_vfsuid(idmap, fs_userns, e->e_uid);
361 e->e_uid = vfsuid_into_kuid(vfsuid);
364 vfsgid = make_vfsgid(idmap, fs_userns, e->e_gid);
365 e->e_gid = vfsgid_into_kgid(vfsgid);
372 * The @noperm argument is used to skip permission checking and is a temporary
373 * measure. Quoting Miklos from an earlier discussion:
375 * > So there are two paths to getting an acl:
376 * > 1) permission checking and 2) retrieving the value via getxattr(2).
377 * > This is a similar situation as reading a symlink vs. following it.
378 * > When following a symlink overlayfs always reads the link on the
379 * > underlying fs just as if it was a readlink(2) call, calling
380 * > security_inode_readlink() instead of security_inode_follow_link().
381 * > This is logical: we are reading the link from the underlying storage,
382 * > and following it on overlayfs.
384 * > Applying the same logic to acl: we do need to call the
385 * > security_inode_getxattr() on the underlying fs, even if just want to
386 * > check permissions on overlay. This is currently not done, which is an
389 * > Maybe adding the check to ovl_get_acl() is the right way to go, but
390 * > I'm a little afraid of a performance regression. Will look into that.
392 * Until we have made a decision allow this helper to take the @noperm
393 * argument. We should hopefully be able to remove it soon.
395 struct posix_acl *ovl_get_acl_path(const struct path *path,
396 const char *acl_name, bool noperm)
398 struct posix_acl *real_acl, *clone;
399 struct mnt_idmap *idmap;
400 struct inode *realinode = d_inode(path->dentry);
402 idmap = mnt_idmap(path->mnt);
405 real_acl = get_inode_acl(realinode, posix_acl_type(acl_name));
407 real_acl = vfs_get_acl(idmap, path->dentry, acl_name);
408 if (IS_ERR_OR_NULL(real_acl))
411 if (!is_idmapped_mnt(path->mnt))
415 * We cannot alter the ACLs returned from the relevant layer as that
416 * would alter the cached values filesystem wide for the lower
417 * filesystem. Instead we can clone the ACLs and then apply the
418 * relevant idmapping of the layer.
420 clone = posix_acl_clone(real_acl, GFP_KERNEL);
421 posix_acl_release(real_acl); /* release original acl */
423 return ERR_PTR(-ENOMEM);
425 ovl_idmap_posix_acl(realinode, idmap, clone);
430 * When the relevant layer is an idmapped mount we need to take the idmapping
431 * of the layer into account and translate any ACL_{GROUP,USER} values
432 * according to the idmapped mount.
434 * We cannot alter the ACLs returned from the relevant layer as that would
435 * alter the cached values filesystem wide for the lower filesystem. Instead we
436 * can clone the ACLs and then apply the relevant idmapping of the layer.
438 * This is obviously only relevant when idmapped layers are used.
440 struct posix_acl *do_ovl_get_acl(struct mnt_idmap *idmap,
441 struct inode *inode, int type,
442 bool rcu, bool noperm)
444 struct inode *realinode;
445 struct posix_acl *acl;
446 struct path realpath;
448 /* Careful in RCU walk mode */
449 realinode = ovl_i_path_real(inode, &realpath);
452 return ERR_PTR(-ECHILD);
455 if (!IS_POSIXACL(realinode))
460 * If the layer is idmapped drop out of RCU path walk
461 * so we can clone the ACLs.
463 if (is_idmapped_mnt(realpath.mnt))
464 return ERR_PTR(-ECHILD);
466 acl = get_cached_acl_rcu(realinode, type);
468 const struct cred *old_cred;
470 old_cred = ovl_override_creds(inode->i_sb);
471 acl = ovl_get_acl_path(&realpath, posix_acl_xattr_name(type), noperm);
472 ovl_revert_creds(old_cred);
478 static int ovl_set_or_remove_acl(struct dentry *dentry, struct inode *inode,
479 struct posix_acl *acl, int type)
482 struct path realpath;
483 const char *acl_name;
484 const struct cred *old_cred;
485 struct ovl_fs *ofs = OVL_FS(dentry->d_sb);
486 struct dentry *upperdentry = ovl_dentry_upper(dentry);
487 struct dentry *realdentry = upperdentry ?: ovl_dentry_lower(dentry);
490 * If ACL is to be removed from a lower file, check if it exists in
491 * the first place before copying it up.
493 acl_name = posix_acl_xattr_name(type);
494 if (!acl && !upperdentry) {
495 struct posix_acl *real_acl;
497 ovl_path_lower(dentry, &realpath);
498 old_cred = ovl_override_creds(dentry->d_sb);
499 real_acl = vfs_get_acl(mnt_idmap(realpath.mnt), realdentry,
501 ovl_revert_creds(old_cred);
502 if (IS_ERR(real_acl)) {
503 err = PTR_ERR(real_acl);
506 posix_acl_release(real_acl);
510 err = ovl_copy_up(dentry);
514 realdentry = ovl_dentry_upper(dentry);
517 err = ovl_want_write(dentry);
521 old_cred = ovl_override_creds(dentry->d_sb);
523 err = ovl_do_set_acl(ofs, realdentry, acl_name, acl);
525 err = ovl_do_remove_acl(ofs, realdentry, acl_name);
526 ovl_revert_creds(old_cred);
527 ovl_drop_write(dentry);
535 int ovl_set_acl(struct mnt_idmap *idmap, struct dentry *dentry,
536 struct posix_acl *acl, int type)
539 struct inode *inode = d_inode(dentry);
540 struct dentry *workdir = ovl_workdir(dentry);
541 struct inode *realinode = ovl_inode_real(inode);
543 if (!IS_POSIXACL(d_inode(workdir)))
545 if (!realinode->i_op->set_acl)
547 if (type == ACL_TYPE_DEFAULT && !S_ISDIR(inode->i_mode))
548 return acl ? -EACCES : 0;
549 if (!inode_owner_or_capable(&nop_mnt_idmap, inode))
553 * Check if sgid bit needs to be cleared (actual setacl operation will
554 * be done with mounter's capabilities and so that won't do it for us).
556 if (unlikely(inode->i_mode & S_ISGID) && type == ACL_TYPE_ACCESS &&
557 !in_group_p(inode->i_gid) &&
558 !capable_wrt_inode_uidgid(&nop_mnt_idmap, inode, CAP_FSETID)) {
559 struct iattr iattr = { .ia_valid = ATTR_KILL_SGID };
561 err = ovl_setattr(&nop_mnt_idmap, dentry, &iattr);
566 return ovl_set_or_remove_acl(dentry, inode, acl, type);
570 int ovl_update_time(struct inode *inode, int flags)
572 if (flags & S_ATIME) {
573 struct ovl_fs *ofs = OVL_FS(inode->i_sb);
574 struct path upperpath = {
575 .mnt = ovl_upper_mnt(ofs),
576 .dentry = ovl_upperdentry_dereference(OVL_I(inode)),
579 if (upperpath.dentry) {
580 touch_atime(&upperpath);
581 inode_set_atime_to_ts(inode,
582 inode_get_atime(d_inode(upperpath.dentry)));
588 static int ovl_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
592 struct inode *realinode = ovl_inode_realdata(inode);
593 const struct cred *old_cred;
598 if (!realinode->i_op->fiemap)
601 old_cred = ovl_override_creds(inode->i_sb);
602 err = realinode->i_op->fiemap(realinode, fieinfo, start, len);
603 ovl_revert_creds(old_cred);
609 * Work around the fact that security_file_ioctl() takes a file argument.
610 * Introducing security_inode_fileattr_get/set() hooks would solve this issue
613 static int ovl_security_fileattr(const struct path *realpath, struct fileattr *fa,
622 if (force_o_largefile())
623 flags |= O_LARGEFILE;
625 file = dentry_open(realpath, flags, current_cred());
627 return PTR_ERR(file);
630 cmd = fa->fsx_valid ? FS_IOC_FSSETXATTR : FS_IOC_SETFLAGS;
632 cmd = fa->fsx_valid ? FS_IOC_FSGETXATTR : FS_IOC_GETFLAGS;
634 err = security_file_ioctl(file, cmd, 0);
640 int ovl_real_fileattr_set(const struct path *realpath, struct fileattr *fa)
644 err = ovl_security_fileattr(realpath, fa, true);
648 return vfs_fileattr_set(mnt_idmap(realpath->mnt), realpath->dentry, fa);
651 int ovl_fileattr_set(struct mnt_idmap *idmap,
652 struct dentry *dentry, struct fileattr *fa)
654 struct inode *inode = d_inode(dentry);
655 struct path upperpath;
656 const struct cred *old_cred;
660 err = ovl_copy_up(dentry);
662 ovl_path_real(dentry, &upperpath);
664 err = ovl_want_write(dentry);
668 old_cred = ovl_override_creds(inode->i_sb);
670 * Store immutable/append-only flags in xattr and clear them
671 * in upper fileattr (in case they were set by older kernel)
672 * so children of "ovl-immutable" directories lower aliases of
673 * "ovl-immutable" hardlinks could be copied up.
674 * Clear xattr when flags are cleared.
676 err = ovl_set_protattr(inode, upperpath.dentry, fa);
678 err = ovl_real_fileattr_set(&upperpath, fa);
679 ovl_revert_creds(old_cred);
680 ovl_drop_write(dentry);
683 * Merge real inode flags with inode flags read from
684 * overlay.protattr xattr
686 flags = ovl_inode_real(inode)->i_flags & OVL_COPY_I_FLAGS_MASK;
688 BUILD_BUG_ON(OVL_PROT_I_FLAGS_MASK & ~OVL_COPY_I_FLAGS_MASK);
689 flags |= inode->i_flags & OVL_PROT_I_FLAGS_MASK;
690 inode_set_flags(inode, flags, OVL_COPY_I_FLAGS_MASK);
699 /* Convert inode protection flags to fileattr flags */
700 static void ovl_fileattr_prot_flags(struct inode *inode, struct fileattr *fa)
702 BUILD_BUG_ON(OVL_PROT_FS_FLAGS_MASK & ~FS_COMMON_FL);
703 BUILD_BUG_ON(OVL_PROT_FSX_FLAGS_MASK & ~FS_XFLAG_COMMON);
705 if (inode->i_flags & S_APPEND) {
706 fa->flags |= FS_APPEND_FL;
707 fa->fsx_xflags |= FS_XFLAG_APPEND;
709 if (inode->i_flags & S_IMMUTABLE) {
710 fa->flags |= FS_IMMUTABLE_FL;
711 fa->fsx_xflags |= FS_XFLAG_IMMUTABLE;
715 int ovl_real_fileattr_get(const struct path *realpath, struct fileattr *fa)
719 err = ovl_security_fileattr(realpath, fa, false);
723 err = vfs_fileattr_get(realpath->dentry, fa);
724 if (err == -ENOIOCTLCMD)
729 int ovl_fileattr_get(struct dentry *dentry, struct fileattr *fa)
731 struct inode *inode = d_inode(dentry);
732 struct path realpath;
733 const struct cred *old_cred;
736 ovl_path_real(dentry, &realpath);
738 old_cred = ovl_override_creds(inode->i_sb);
739 err = ovl_real_fileattr_get(&realpath, fa);
740 ovl_fileattr_prot_flags(inode, fa);
741 ovl_revert_creds(old_cred);
746 static const struct inode_operations ovl_file_inode_operations = {
747 .setattr = ovl_setattr,
748 .permission = ovl_permission,
749 .getattr = ovl_getattr,
750 .listxattr = ovl_listxattr,
751 .get_inode_acl = ovl_get_inode_acl,
752 .get_acl = ovl_get_acl,
753 .set_acl = ovl_set_acl,
754 .update_time = ovl_update_time,
755 .fiemap = ovl_fiemap,
756 .fileattr_get = ovl_fileattr_get,
757 .fileattr_set = ovl_fileattr_set,
760 static const struct inode_operations ovl_symlink_inode_operations = {
761 .setattr = ovl_setattr,
762 .get_link = ovl_get_link,
763 .getattr = ovl_getattr,
764 .listxattr = ovl_listxattr,
765 .update_time = ovl_update_time,
768 static const struct inode_operations ovl_special_inode_operations = {
769 .setattr = ovl_setattr,
770 .permission = ovl_permission,
771 .getattr = ovl_getattr,
772 .listxattr = ovl_listxattr,
773 .get_inode_acl = ovl_get_inode_acl,
774 .get_acl = ovl_get_acl,
775 .set_acl = ovl_set_acl,
776 .update_time = ovl_update_time,
779 static const struct address_space_operations ovl_aops = {
780 /* For O_DIRECT dentry_open() checks f_mapping->a_ops->direct_IO */
781 .direct_IO = noop_direct_IO,
785 * It is possible to stack overlayfs instance on top of another
786 * overlayfs instance as lower layer. We need to annotate the
787 * stackable i_mutex locks according to stack level of the super
788 * block instance. An overlayfs instance can never be in stack
789 * depth 0 (there is always a real fs below it). An overlayfs
790 * inode lock will use the lockdep annotation ovl_i_mutex_key[depth].
792 * For example, here is a snip from /proc/lockdep_chains after
793 * dir_iterate of nested overlayfs:
795 * [...] &ovl_i_mutex_dir_key[depth] (stack_depth=2)
796 * [...] &ovl_i_mutex_dir_key[depth]#2 (stack_depth=1)
797 * [...] &type->i_mutex_dir_key (stack_depth=0)
799 * Locking order w.r.t ovl_want_write() is important for nested overlayfs.
801 * This chain is valid:
802 * - inode->i_rwsem (inode_lock[2])
803 * - upper_mnt->mnt_sb->s_writers (ovl_want_write[0])
804 * - OVL_I(inode)->lock (ovl_inode_lock[2])
805 * - OVL_I(lowerinode)->lock (ovl_inode_lock[1])
807 * And this chain is valid:
808 * - inode->i_rwsem (inode_lock[2])
809 * - OVL_I(inode)->lock (ovl_inode_lock[2])
810 * - lowerinode->i_rwsem (inode_lock[1])
811 * - OVL_I(lowerinode)->lock (ovl_inode_lock[1])
813 * But lowerinode->i_rwsem SHOULD NOT be acquired while ovl_want_write() is
814 * held, because it is in reverse order of the non-nested case using the same
816 * - inode->i_rwsem (inode_lock[1])
817 * - upper_mnt->mnt_sb->s_writers (ovl_want_write[0])
818 * - OVL_I(inode)->lock (ovl_inode_lock[1])
820 #define OVL_MAX_NESTING FILESYSTEM_MAX_STACK_DEPTH
822 static inline void ovl_lockdep_annotate_inode_mutex_key(struct inode *inode)
824 #ifdef CONFIG_LOCKDEP
825 static struct lock_class_key ovl_i_mutex_key[OVL_MAX_NESTING];
826 static struct lock_class_key ovl_i_mutex_dir_key[OVL_MAX_NESTING];
827 static struct lock_class_key ovl_i_lock_key[OVL_MAX_NESTING];
829 int depth = inode->i_sb->s_stack_depth - 1;
831 if (WARN_ON_ONCE(depth < 0 || depth >= OVL_MAX_NESTING))
834 if (S_ISDIR(inode->i_mode))
835 lockdep_set_class(&inode->i_rwsem, &ovl_i_mutex_dir_key[depth]);
837 lockdep_set_class(&inode->i_rwsem, &ovl_i_mutex_key[depth]);
839 lockdep_set_class(&OVL_I(inode)->lock, &ovl_i_lock_key[depth]);
843 static void ovl_next_ino(struct inode *inode)
845 struct ovl_fs *ofs = OVL_FS(inode->i_sb);
847 inode->i_ino = atomic_long_inc_return(&ofs->last_ino);
848 if (unlikely(!inode->i_ino))
849 inode->i_ino = atomic_long_inc_return(&ofs->last_ino);
852 static void ovl_map_ino(struct inode *inode, unsigned long ino, int fsid)
854 struct ovl_fs *ofs = OVL_FS(inode->i_sb);
855 int xinobits = ovl_xino_bits(ofs);
856 unsigned int xinoshift = 64 - xinobits;
859 * When d_ino is consistent with st_ino (samefs or i_ino has enough
860 * bits to encode layer), set the same value used for st_ino to i_ino,
861 * so inode number exposed via /proc/locks and a like will be
862 * consistent with d_ino and st_ino values. An i_ino value inconsistent
863 * with d_ino also causes nfsd readdirplus to fail.
866 if (ovl_same_fs(ofs)) {
868 } else if (xinobits && likely(!(ino >> xinoshift))) {
869 inode->i_ino |= (unsigned long)fsid << (xinoshift + 1);
874 * For directory inodes on non-samefs with xino disabled or xino
875 * overflow, we allocate a non-persistent inode number, to be used for
876 * resolving st_ino collisions in ovl_map_dev_ino().
878 * To avoid ino collision with legitimate xino values from upper
879 * layer (fsid 0), use the lowest xinobit to map the non
880 * persistent inode numbers to the unified st_ino address space.
882 if (S_ISDIR(inode->i_mode)) {
885 inode->i_ino &= ~0UL >> xinobits;
886 inode->i_ino |= 1UL << xinoshift;
891 void ovl_inode_init(struct inode *inode, struct ovl_inode_params *oip,
892 unsigned long ino, int fsid)
894 struct inode *realinode;
895 struct ovl_inode *oi = OVL_I(inode);
897 oi->__upperdentry = oip->upperdentry;
899 oi->redirect = oip->redirect;
900 oi->lowerdata_redirect = oip->lowerdata_redirect;
902 realinode = ovl_inode_real(inode);
904 ovl_copyflags(realinode, inode);
905 ovl_map_ino(inode, ino, fsid);
908 static void ovl_fill_inode(struct inode *inode, umode_t mode, dev_t rdev)
910 inode->i_mode = mode;
911 inode->i_flags |= S_NOCMTIME;
912 #ifdef CONFIG_FS_POSIX_ACL
913 inode->i_acl = inode->i_default_acl = ACL_DONT_CACHE;
916 ovl_lockdep_annotate_inode_mutex_key(inode);
918 switch (mode & S_IFMT) {
920 inode->i_op = &ovl_file_inode_operations;
921 inode->i_fop = &ovl_file_operations;
922 inode->i_mapping->a_ops = &ovl_aops;
926 inode->i_op = &ovl_dir_inode_operations;
927 inode->i_fop = &ovl_dir_operations;
931 inode->i_op = &ovl_symlink_inode_operations;
935 inode->i_op = &ovl_special_inode_operations;
936 init_special_inode(inode, mode, rdev);
942 * With inodes index enabled, an overlay inode nlink counts the union of upper
943 * hardlinks and non-covered lower hardlinks. During the lifetime of a non-pure
944 * upper inode, the following nlink modifying operations can happen:
946 * 1. Lower hardlink copy up
947 * 2. Upper hardlink created, unlinked or renamed over
948 * 3. Lower hardlink whiteout or renamed over
950 * For the first, copy up case, the union nlink does not change, whether the
951 * operation succeeds or fails, but the upper inode nlink may change.
952 * Therefore, before copy up, we store the union nlink value relative to the
953 * lower inode nlink in the index inode xattr .overlay.nlink.
955 * For the second, upper hardlink case, the union nlink should be incremented
956 * or decremented IFF the operation succeeds, aligned with nlink change of the
957 * upper inode. Therefore, before link/unlink/rename, we store the union nlink
958 * value relative to the upper inode nlink in the index inode.
960 * For the last, lower cover up case, we simplify things by preceding the
961 * whiteout or cover up with copy up. This makes sure that there is an index
962 * upper inode where the nlink xattr can be stored before the copied up upper
965 #define OVL_NLINK_ADD_UPPER (1 << 0)
968 * On-disk format for indexed nlink:
970 * nlink relative to the upper inode - "U[+-]NUM"
971 * nlink relative to the lower inode - "L[+-]NUM"
974 static int ovl_set_nlink_common(struct dentry *dentry,
975 struct dentry *realdentry, const char *format)
977 struct inode *inode = d_inode(dentry);
978 struct inode *realinode = d_inode(realdentry);
982 len = snprintf(buf, sizeof(buf), format,
983 (int) (inode->i_nlink - realinode->i_nlink));
985 if (WARN_ON(len >= sizeof(buf)))
988 return ovl_setxattr(OVL_FS(inode->i_sb), ovl_dentry_upper(dentry),
989 OVL_XATTR_NLINK, buf, len);
992 int ovl_set_nlink_upper(struct dentry *dentry)
994 return ovl_set_nlink_common(dentry, ovl_dentry_upper(dentry), "U%+i");
997 int ovl_set_nlink_lower(struct dentry *dentry)
999 return ovl_set_nlink_common(dentry, ovl_dentry_lower(dentry), "L%+i");
1002 unsigned int ovl_get_nlink(struct ovl_fs *ofs, struct dentry *lowerdentry,
1003 struct dentry *upperdentry,
1004 unsigned int fallback)
1011 if (!lowerdentry || !upperdentry || d_inode(lowerdentry)->i_nlink == 1)
1014 err = ovl_getxattr_upper(ofs, upperdentry, OVL_XATTR_NLINK,
1015 &buf, sizeof(buf) - 1);
1020 if ((buf[0] != 'L' && buf[0] != 'U') ||
1021 (buf[1] != '+' && buf[1] != '-'))
1024 err = kstrtoint(buf + 1, 10, &nlink_diff);
1028 nlink = d_inode(buf[0] == 'L' ? lowerdentry : upperdentry)->i_nlink;
1029 nlink += nlink_diff;
1037 pr_warn_ratelimited("failed to get index nlink (%pd2, err=%i)\n",
1042 struct inode *ovl_new_inode(struct super_block *sb, umode_t mode, dev_t rdev)
1044 struct inode *inode;
1046 inode = new_inode(sb);
1048 ovl_fill_inode(inode, mode, rdev);
1053 static int ovl_inode_test(struct inode *inode, void *data)
1055 return inode->i_private == data;
1058 static int ovl_inode_set(struct inode *inode, void *data)
1060 inode->i_private = data;
1064 static bool ovl_verify_inode(struct inode *inode, struct dentry *lowerdentry,
1065 struct dentry *upperdentry, bool strict)
1068 * For directories, @strict verify from lookup path performs consistency
1069 * checks, so NULL lower/upper in dentry must match NULL lower/upper in
1070 * inode. Non @strict verify from NFS handle decode path passes NULL for
1071 * 'unknown' lower/upper.
1073 if (S_ISDIR(inode->i_mode) && strict) {
1074 /* Real lower dir moved to upper layer under us? */
1075 if (!lowerdentry && ovl_inode_lower(inode))
1078 /* Lookup of an uncovered redirect origin? */
1079 if (!upperdentry && ovl_inode_upper(inode))
1084 * Allow non-NULL lower inode in ovl_inode even if lowerdentry is NULL.
1085 * This happens when finding a copied up overlay inode for a renamed
1086 * or hardlinked overlay dentry and lower dentry cannot be followed
1087 * by origin because lower fs does not support file handles.
1089 if (lowerdentry && ovl_inode_lower(inode) != d_inode(lowerdentry))
1093 * Allow non-NULL __upperdentry in inode even if upperdentry is NULL.
1094 * This happens when finding a lower alias for a copied up hard link.
1096 if (upperdentry && ovl_inode_upper(inode) != d_inode(upperdentry))
1102 struct inode *ovl_lookup_inode(struct super_block *sb, struct dentry *real,
1105 struct inode *inode, *key = d_inode(real);
1107 inode = ilookup5(sb, (unsigned long) key, ovl_inode_test, key);
1111 if (!ovl_verify_inode(inode, is_upper ? NULL : real,
1112 is_upper ? real : NULL, false)) {
1114 return ERR_PTR(-ESTALE);
1120 bool ovl_lookup_trap_inode(struct super_block *sb, struct dentry *dir)
1122 struct inode *key = d_inode(dir);
1126 trap = ilookup5(sb, (unsigned long) key, ovl_inode_test, key);
1130 res = IS_DEADDIR(trap) && !ovl_inode_upper(trap) &&
1131 !ovl_inode_lower(trap);
1138 * Create an inode cache entry for layer root dir, that will intentionally
1139 * fail ovl_verify_inode(), so any lookup that will find some layer root
1142 struct inode *ovl_get_trap_inode(struct super_block *sb, struct dentry *dir)
1144 struct inode *key = d_inode(dir);
1148 return ERR_PTR(-ENOTDIR);
1150 trap = iget5_locked(sb, (unsigned long) key, ovl_inode_test,
1151 ovl_inode_set, key);
1153 return ERR_PTR(-ENOMEM);
1155 if (!(trap->i_state & I_NEW)) {
1156 /* Conflicting layer roots? */
1158 return ERR_PTR(-ELOOP);
1161 trap->i_mode = S_IFDIR;
1162 trap->i_flags = S_DEAD;
1163 unlock_new_inode(trap);
1169 * Does overlay inode need to be hashed by lower inode?
1171 static bool ovl_hash_bylower(struct super_block *sb, struct dentry *upper,
1172 struct dentry *lower, bool index)
1174 struct ovl_fs *ofs = OVL_FS(sb);
1176 /* No, if pure upper */
1180 /* Yes, if already indexed */
1184 /* Yes, if won't be copied up */
1185 if (!ovl_upper_mnt(ofs))
1188 /* No, if lower hardlink is or will be broken on copy up */
1189 if ((upper || !ovl_indexdir(sb)) &&
1190 !d_is_dir(lower) && d_inode(lower)->i_nlink > 1)
1193 /* No, if non-indexed upper with NFS export */
1194 if (ofs->config.nfs_export && upper)
1197 /* Otherwise, hash by lower inode for fsnotify */
1201 static struct inode *ovl_iget5(struct super_block *sb, struct inode *newinode,
1204 return newinode ? inode_insert5(newinode, (unsigned long) key,
1205 ovl_inode_test, ovl_inode_set, key) :
1206 iget5_locked(sb, (unsigned long) key,
1207 ovl_inode_test, ovl_inode_set, key);
1210 struct inode *ovl_get_inode(struct super_block *sb,
1211 struct ovl_inode_params *oip)
1213 struct ovl_fs *ofs = OVL_FS(sb);
1214 struct dentry *upperdentry = oip->upperdentry;
1215 struct ovl_path *lowerpath = ovl_lowerpath(oip->oe);
1216 struct inode *realinode = upperdentry ? d_inode(upperdentry) : NULL;
1217 struct inode *inode;
1218 struct dentry *lowerdentry = lowerpath ? lowerpath->dentry : NULL;
1219 struct path realpath = {
1220 .dentry = upperdentry ?: lowerdentry,
1221 .mnt = upperdentry ? ovl_upper_mnt(ofs) : lowerpath->layer->mnt,
1223 bool bylower = ovl_hash_bylower(sb, upperdentry, lowerdentry,
1225 int fsid = bylower ? lowerpath->layer->fsid : 0;
1227 unsigned long ino = 0;
1228 int err = oip->newinode ? -EEXIST : -ENOMEM;
1231 realinode = d_inode(lowerdentry);
1234 * Copy up origin (lower) may exist for non-indexed upper, but we must
1235 * not use lower as hash key if this is a broken hardlink.
1237 is_dir = S_ISDIR(realinode->i_mode);
1238 if (upperdentry || bylower) {
1239 struct inode *key = d_inode(bylower ? lowerdentry :
1241 unsigned int nlink = is_dir ? 1 : realinode->i_nlink;
1243 inode = ovl_iget5(sb, oip->newinode, key);
1246 if (!(inode->i_state & I_NEW)) {
1248 * Verify that the underlying files stored in the inode
1249 * match those in the dentry.
1251 if (!ovl_verify_inode(inode, lowerdentry, upperdentry,
1259 ovl_free_entry(oip->oe);
1260 kfree(oip->redirect);
1261 kfree(oip->lowerdata_redirect);
1265 /* Recalculate nlink for non-dir due to indexing */
1267 nlink = ovl_get_nlink(ofs, lowerdentry, upperdentry,
1269 set_nlink(inode, nlink);
1272 /* Lower hardlink that will be broken on copy up */
1273 inode = new_inode(sb);
1278 ino = realinode->i_ino;
1279 fsid = lowerpath->layer->fsid;
1281 ovl_fill_inode(inode, realinode->i_mode, realinode->i_rdev);
1282 ovl_inode_init(inode, oip, ino, fsid);
1284 if (upperdentry && ovl_is_impuredir(sb, upperdentry))
1285 ovl_set_flag(OVL_IMPURE, inode);
1288 ovl_set_flag(OVL_INDEX, inode);
1291 ovl_set_flag(OVL_CONST_INO, inode);
1293 /* Check for non-merge dir that may have whiteouts */
1295 if (((upperdentry && lowerdentry) || ovl_numlower(oip->oe) > 1) ||
1296 ovl_path_check_origin_xattr(ofs, &realpath)) {
1297 ovl_set_flag(OVL_WHITEOUTS, inode);
1301 /* Check for immutable/append-only inode flags in xattr */
1303 ovl_check_protattr(inode, upperdentry);
1305 if (inode->i_state & I_NEW)
1306 unlock_new_inode(inode);
1311 pr_warn_ratelimited("failed to get inode (%i)\n", err);
1312 inode = ERR_PTR(err);