Now that copy up of regular file is done using O_TMPFILE,
we don't need to hold rename_lock throughout copy up.

Use the copy up waitqueue to synchronize concurrent copy up
of the same file. Different regular files can be copied up
concurrently.

The upper dir inode_lock is taken instead of rename_lock,
because it is needed for lookup and later for linking the
temp file, but it is released while copying up data.
Suggested-by: NAl Viro <viro@zeniv.linux.org.uk>
Signed-off-by: NAmir Goldstein <amir73il@gmail.com>
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

The overlay sb 'copyup_wq' and overlay inode 'copying' condition
variable are about to replace the upper sb rename_lock, as finer
grained synchronization objects for concurrent copy up.
Suggested-by: NMiklos Szeredi <miklos@szeredi.hu>
Signed-off-by: NAmir Goldstein <amir73il@gmail.com>
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

In preparation for concurrent copy up, implement copy up
of regular file as O_TMPFILE that is linked to upperdir
instead of a file in workdir that is moved to upperdir.
Suggested-by: NAl Viro <viro@zeniv.linux.org.uk>
Signed-off-by: NAmir Goldstein <amir73il@gmail.com>
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

As preparation to implementing copy up with O_TMPFILE,
name the variable for dentry before final rename 'temp' and
assign it to 'newdentry' only after rename.

Also lookup upper dentry before looking up temp dentry and
move ovl_set_timestamps() into ovl_copy_up_locked(), because
that is going to be more convenient for upcoming change.
Signed-off-by: NAmir Goldstein <amir73il@gmail.com>
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

This is needed for choosing between concurrent copyup
using O_TMPFILE and legacy copyup using workdir+rename.
Signed-off-by: NAmir Goldstein <amir73il@gmail.com>
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

If kcalloc() failed, the return value of ovl_fill_super() is -EINVAL,
not -ENOMEM. So this patch sets this value to -ENOMEM before calling
kcalloc(), and sets it back to -EINVAL after calling kcalloc().
Signed-off-by: NGeliang Tang <geliangtang@gmail.com>
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

FWIW, there's a bit of abuse of struct kstat in overlayfs object
creation paths - for one thing, it ends up with a very small subset
of struct kstat (mode + rdev), for another it also needs link in
case of symlinks and ends up passing it separately.

IMO it would be better to introduce a separate object for that.

In principle, we might even lift that thing into general API and switch
 ->mkdir()/->mknod()/->symlink() to identical calling conventions.  Hell
knows, perhaps ->create() as well...
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

This removes code duplication.
Signed-off-by: NAmir Goldstein <amir73il@gmail.com>
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

The benefit of making directories opaque on creation is that lookups can
stop short when they reach the original created directory, instead of
continue lookup the entire depth of parent directory stack.

The best case is overlay with N layers, performing lookup for first level
directory, which exists only in upper.  In that case, there will be only
one lookup instead of N.
Signed-off-by: NAmir Goldstein <amir73il@gmail.com>
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

oe->opaque is set for

 a) whiteouts
 b) directories having the "trusted.overlay.opaque" xattr

Case b can be simplified, since setting the xattr always implies setting
oe->opaque.  Also once set, the opaque flag is never cleared.

Don't need to set opaque flag for non-directories.
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

Show the value of redirect_dir in /proc/mounts.
Signed-off-by: NAmir Goldstein <amir73il@gmail.com>
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

Add a module option to allow tuning the max size of absolute redirects.
Default is 256.

Size of relative redirects is naturally limited by the the underlying
filesystem's max filename length (usually 255).
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

This patch introduces a kernel config option and a module param.  Both can
be used independently to turn the default value of redirect_dir on or off.
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

Before introducing redirect_dir feature, the condition
!ovl_lower_positive(dentry) for a directory, implied that it is a pure
upper directory, which may be removed if empty.

Now that directory can be redirect, it is possible that upper does not
cover any lower (i.e. !ovl_lower_positive(dentry)), but the directory is a
merge (with redirected path) and maybe non empty.

Check for this case in ovl_remove_upper().

This change fixes the following test case from rename-pop-dir.py
of unionmount-testsuite:

    """Remove dir and rename old name"""
    d = ctx.non_empty_dir()
    d2 = ctx.no_dir()

    ctx.rmdir(d, err=ENOTEMPTY)
    ctx.rename(d, d2)
    ctx.rmdir(d, err=ENOENT)
    ctx.rmdir(d2, err=ENOTEMPTY)

./run --ov rename-pop-dir
/mnt/a/no_dir103: Expected error (Directory not empty) was not produced
Signed-off-by: NAmir Goldstein <amir73il@gmail.com>
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

Current code returns EXDEV when a directory would need to be copied up to
move.  We could copy up the directory tree in this case, but there's
another, simpler solution: point to old lower directory from moved upper
directory.

This is achieved with a "trusted.overlay.redirect" xattr storing the path
relative to the root of the overlay.  After such attribute has been set,
the directory can be moved without further actions required.

This is a backward incompatible feature, old kernels won't be able to
correctly mount an overlay containing redirected directories.
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

If a directory has the "trusted.overlay.redirect" xattr, it means that the
value of the xattr should be used to find the underlying directory on the
next lower layer.

The redirect may be relative or absolute.  Absolute redirects begin with a
slash.

A relative redirect means: instead of the current dentry's name use the
value of the redirect to find the directory in the next lower
layer. Relative redirects must not contain a slash.

An absolute redirect means: look up the directory relative to the root of
the overlay using the value of the redirect in the next lower layer.

Redirects work on lower layers as well.
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

Use a common helper for lookup of upper and lower layers.  This paves the
way for looking up directory redirects.

No functional change.
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

When the upper overlayfs checks "trusted.overlay.*" xattr on the underlying
overlayfs mount, it gets -EPERM, which confuses the upper overlayfs.

Fix this by returning -EOPNOTSUPP instead of -EPERM from
ovl_own_xattr_get() and ovl_own_xattr_set().  This behavior is consistent
with the behavior of ovl_listxattr(), which filters out the private
overlayfs xattrs.

Note: nested overlays are deprecated.  But this change makes sense
regardless: these xattrs are private to the overlay and should always be
hidden.  Hence getting and setting them should indicate this.

[SzMi: Use EOPNOTSUPP instead of ENODATA and use it for both getting and
setting "trusted.overlay." xattrs.  This is a perfectly valid error code
for "we don't support this prefix", which is the case here.]
Signed-off-by: NAmir Goldstein <amir73il@gmail.com>
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

We already calculate f_namelen in statfs as the maximum of the name lengths
provided by the filesystems taking part in the overlay.
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

fs/overlayfs/super.c is the biggest of the overlayfs source files and it
contains various utility functions as well as the rather complicated lookup
code.  Split these parts out to separate files.

Before:

 1446 fs/overlayfs/super.c

After:

  919 fs/overlayfs/super.c
  267 fs/overlayfs/namei.c
  235 fs/overlayfs/util.c
   51 fs/overlayfs/ovl_entry.h
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

If encountering a non-directory, then stop looking at lower layers.

In this case the oe->opaque flag is not set anymore, which doesn't matter
since existence of lower file is now checked at remove/rename time.
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

Check if something exists on the lower layer(s) under the target or rename
to decide if directory needs to be marked "opaque".

Marking opaque is done before the rename, and on failure the marking was
undone.  Also the opaque xattr was removed if the target didn't cover
anything.

This patch changes behavior so that removal of "opaque" is not done in
either of the above cases.  This means that directory may have the opaque
flag even if it doesn't cover anything.  However this shouldn't affect the
performance or semantics of the overalay, while simplifying the code.
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

d_is_dir() is safe to call on a negative dentry.  Use this fact to simplify
handling of the lower or merged directories.
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

The remainging uses of __OVL_PATH_PURE can be replaced by
ovl_dentry_is_opaque().
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

Currently ovl_lookup() checks existence of lower file even if there's a
non-directory on upper (which is always opaque).  This is done so that
remove can decide whether a whiteout is needed or not.

It would be better to defer this check to unlink, since most of the time
the gathered information about opaqueness will be unused.

This adds a helper ovl_lower_positive() that checks if there's anything on
the lower layer(s).

The following patches also introduce changes to how the "opaque" attribute
is updated on directories: this attribute is added when the directory is
creted or moved over a whiteout or object covering something on the lower
layer.  However following changes will allow the attribute to remain on the
directory after being moved, even if the new location doesn't cover
anything.  Because of this, we need to check lower layers even for opaque
directories, so that whiteout is only created when necessary.

This function will later be also used to decide about marking a directory
opaque, so deal with negative dentries as well.  When dealing with
negative, it's enough to check for being a whiteout

If the dentry is positive but not upper then it also obviously needs
whiteout/opaque.
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

And use it instead of ovl_dentry_is_opaque() where appropriate.
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

Since commit 07a2daab ("ovl: Copy up underlying inode's ->i_mode to
overlay inode") sticky checking on overlay inode is performed by the vfs,
so checking against sticky on underlying inode is not needed.
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

This is redundant, the vfs already performed this check (and was broken,
see commit 9409e22a ("vfs: rename: check backing inode being equal")).
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

No sense in opening special files on the underlying layers, they work just
as well if opened on the overlay.

Side effect is that it's no longer possible to connect one side of a pipe
opened on overlayfs with the other side opened on the underlying layer.
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

When copying up within the same fs, try to use vfs_clone_file_range().
This is very efficient when lower and upper are on the same fs
with file reflink support. If vfs_clone_file_range() fails for any
reason, copy up falls back to the regular data copy code.

Tested correct behavior when lower and upper are on:
1. same ext4 (copy)
2. same xfs + reflink patches + mkfs.xfs (copy)
3. same xfs + reflink patches + mkfs.xfs -m reflink=1 (reflink)
4. different xfs + reflink patches + mkfs.xfs -m reflink=1 (copy)

For comparison, on my laptop, xfstest overlay/001 (copy up of large
sparse files) takes less than 1 second in the xfs reflink setup vs.
25 seconds on the rest of the setups.
Signed-off-by: NAmir Goldstein <amir73il@gmail.com>
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

This reverts commit 03bea604.

Commit 4d0c5ba2 ("vfs: do get_write_access() on upper layer of
overlayfs") makes the writecount checks inside overlayfs superfluous, the
file is already copied up and write access acquired on the upper inode when
ovl_setattr is called with ATTR_SIZE.
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

If .readlink == NULL implies generic_readlink().

Generated by:

to_del="\.readlink.*=.*generic_readlink"
for i in `git grep -l $to_del`; do sed -i "/$to_del"/d $i; done
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>

Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Handling of recursion in d_real() is completely broken.  Recursion is only
done in the 'inode != NULL' case.  But when opening the file we have
'inode == NULL' hence d_real() will return an overlay dentry.  This won't
work since overlayfs doesn't define its own file operations, so all file
ops will fail.

Fix by doing the recursion first and the check against the inode second.

Bash script to reproduce the issue written by Quentin:

 - 8< - - - - - 8< - - - - - 8< - - - - - 8< - - - -
tmpdir=$(mktemp -d)
pushd ${tmpdir}

mkdir -p {upper,lower,work}
echo -n 'rocks' > lower/ksplice
mount -t overlay level_zero upper -o lowerdir=lower,upperdir=upper,workdir=work
cat upper/ksplice

tmpdir2=$(mktemp -d)
pushd ${tmpdir2}

mkdir -p {upper,work}
mount -t overlay level_one upper -o lowerdir=${tmpdir}/upper,upperdir=upper,workdir=work
ls -l upper/ksplice
cat upper/ksplice
 - 8< - - - - - 8< - - - - - 8< - - - - - 8< - - - - 
Reported-by: NQuentin Casasnovas <quentin.casasnovas@oracle.com>
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>
Fixes: 2d902671 ("vfs: merge .d_select_inode() into .d_real()")
Cc: <stable@vger.kernel.org> # v4.8+

Make sure the copied up file hits the disk before renaming to the final
destination.  If this is not done then the copy-up may corrupt the data in
the file in case of a crash.
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>
Cc: <stable@vger.kernel.org>

tmpfs doesn't have ->get_acl() because it only uses cached acls.

This fixes the acl tests in pjdfstest when tmpfs is used as the upper layer
of the overlay.
Reported-by: NAmir Goldstein <amir73il@gmail.com>
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>
Fixes: 39a25b2b ("ovl: define ->get_acl() for overlay inodes")
Cc: <stable@vger.kernel.org> # v4.8

This change fixes xfstest generic/375, which failed to clear the
setgid bit in the following test case on overlayfs:

  touch $testfile
  chown 100:100 $testfile
  chmod 2755 $testfile
  _runas -u 100 -g 101 -- setfacl -m u::rwx,g::rwx,o::rwx $testfile
Reported-by: NAmir Goldstein <amir73il@gmail.com>
Signed-off-by: NMiklos Szeredi <mszeredi@redhat.com>
Tested-by: NAmir Goldstein <amir73il@gmail.com>
Fixes: d837a49b ("ovl: fix POSIX ACL setting")
Cc: <stable@vger.kernel.org> # v4.8

ovl_fill_super calls ovl_new_inode to create a root inode for the new
superblock before initializing sb->s_xattr.  This wrongly causes
IOP_XATTR to be cleared in i_opflags of the new inode, causing SELinux
to log the following message:

  SELinux: (dev overlay, type overlay) has no xattr support

Fix this by initializing sb->s_xattr and similar fields before calling
ovl_new_inode.
Signed-off-by: NAndreas Gruenbacher <agruenba@redhat.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>