Symbolic links are files that act as pointers to other files. To
understand their behavior, you must first understand how hard
links work.
A hard link to a file is indistinguishable from the original file
because it is a reference to the object underlying the original
filename. (To be precise: each of the hard links to a file is a
reference to the same inode number, where an inode number is an
index into the inode table, which contains metadata about all
files on a filesystem. See stat(2).) Changes to a file are
independent of the name used to reference the file. Hard links
may not refer to directories (to prevent the possibility of loops
within the filesystem tree, which would confuse many programs)
and may not refer to files on different filesystems (because
inode numbers are not unique across filesystems).
A symbolic link is a special type of file whose contents are a
string that is the pathname of another file, the file to which
the link refers. (The contents of a symbolic link can be read
using readlink(2).) In other words, a symbolic link is a pointer
to another name, and not to an underlying object. For this
reason, symbolic links may refer to directories and may cross
filesystem boundaries.
There is no requirement that the pathname referred to by a
symbolic link should exist. A symbolic link that refers to a
pathname that does not exist is said to be a dangling link.
Because a symbolic link and its referenced object coexist in the
filesystem name space, confusion can arise in distinguishing
between the link itself and the referenced object. On historical
systems, commands and system calls adopted their own link-
following conventions in a somewhat ad-hoc fashion. Rules for a
more uniform approach, as they are implemented on Linux and other
systems, are outlined here. It is important that site-local
applications also conform to these rules, so that the user
interface can be as consistent as possible.
Magic links
There is a special class of symbolic-link-like objects known as
"magic links", which can be found in certain pseudofilesystems
such as proc(5) (examples include /proc/[pid]/exe and
/proc/[pid]/fd/*). Unlike normal symbolic links, magic links are
not resolved through pathname-expansion, but instead act as
direct references to the kernel's own representation of a file
handle. As such, these magic links allow users to access files
which cannot be referenced with normal paths (such as unlinked
files still referenced by a running program ).
Because they can bypass ordinary mount_namespaces(7)-based
restrictions, magic links have been used as attack vectors in
various exploits.
Symbolic link ownership, permissions, and timestamps
The owner and group of an existing symbolic link can be changed
using lchown(2). The only time that the ownership of a symbolic
link matters is when the link is being removed or renamed in a
directory that has the sticky bit set (see stat(2)).
The last access and last modification timestamps of a symbolic
link can be changed using utimensat(2) or lutimes(3).
On Linux, the permissions of an ordinary symbolic link are not
used in any operations; the permissions are always 0777 (read,
write, and execute for all user categories), and can't be
changed.
However, magic links do not follow this rule. They can have a
non-0777 mode, though this mode is not currently used in any
permission checks.
Obtaining a file descriptor that refers to a symbolic link
Using the combination of the O_PATH and O_NOFOLLOW flags to
open(2) yields a file descriptor that can be passed as the dirfd
argument in system calls such as fstatat(2), fchownat(2),
fchmodat(2), linkat(2), and readlinkat(2), in order to operate on
the symbolic link itself (rather than the file to which it
refers).
By default (i.e., if the AT_SYMLINK_FOLLOW flag is not
specified), if name_to_handle_at(2) is applied to a symbolic
link, it yields a handle for the symbolic link (rather than the
file to which it refers). One can then obtain a file descriptor
for the symbolic link (rather than the file to which it refers)
by specifying the O_PATH flag in a subsequent call to
open_by_handle_at(2). Again, that file descriptor can be used in
the aforementioned system calls to operate on the symbolic link
itself.
Handling of symbolic links by system calls and commands
Symbolic links are handled either by operating on the link
itself, or by operating on the object referred to by the link.
In the latter case, an application or system call is said to
follow the link. Symbolic links may refer to other symbolic
links, in which case the links are dereferenced until an object
that is not a symbolic link is found, a symbolic link that refers
to a file which does not exist is found, or a loop is detected.
(Loop detection is done by placing an upper limit on the number
of links that may be followed, and an error results if this limit
is exceeded.)
There are three separate areas that need to be discussed. They
are as follows:
1. Symbolic links used as filename arguments for system calls.
2. Symbolic links specified as command-line arguments to
utilities that are not traversing a file tree.
3. Symbolic links encountered by utilities that are traversing a
file tree (either specified on the command line or encountered
as part of the file hierarchy walk).
Before describing the treatment of symbolic links by system calls
and commands, we require some terminology. Given a pathname of
the form a/b/c, the part preceding the final slash (i.e., a/b) is
called the dirname component, and the part following the final
slash (i.e., c) is called the basename component.
Treatment of symbolic links in system calls
The first area is symbolic links used as filename arguments for
system calls.
The treatment of symbolic links within a pathname passed to a
system call is as follows:
1. Within the dirname component of a pathname, symbolic links are
always followed in nearly every system call. (This is also
true for commands.) The one exception is openat2(2), which
provides flags that can be used to explicitly prevent
following of symbolic links in the dirname component.
2. Except as noted below, all system calls follow symbolic links
in the basename component of a pathname. For example, if
there were a symbolic link slink which pointed to a file named
afile, the system call open("slink" ...) would return a file
descriptor referring to the file afile.
Various system calls do not follow links in the basename
component of a pathname, and operate on the symbolic link itself.
They are: lchown(2), lgetxattr(2), llistxattr(2),
lremovexattr(2), lsetxattr(2), lstat(2), readlink(2), rename(2),
rmdir(2), and unlink(2).
Certain other system calls optionally follow symbolic links in
the basename component of a pathname. They are: faccessat(2),
fchownat(2), fstatat(2), linkat(2), name_to_handle_at(2),
open(2), openat(2), open_by_handle_at(2), and utimensat(2); see
their manual pages for details. Because remove(3) is an alias
for unlink(2), that library function also does not follow
symbolic links. When rmdir(2) is applied to a symbolic link, it
fails with the error ENOTDIR.
link(2) warrants special discussion. POSIX.1-2001 specifies that
link(2) should dereference oldpath if it is a symbolic link.
However, Linux does not do this. (By default, Solaris is the
same, but the POSIX.1-2001 specified behavior can be obtained
with suitable compiler options.) POSIX.1-2008 changed the
specification to allow either behavior in an implementation.
Commands not traversing a file tree
The second area is symbolic links, specified as command-line
filename arguments, to commands which are not traversing a file
tree.
Except as noted below, commands follow symbolic links named as
command-line arguments. For example, if there were a symbolic
link slink which pointed to a file named afile, the command cat
slink would display the contents of the file afile.
It is important to realize that this rule includes commands which
may optionally traverse file trees; for example, the command
chown file is included in this rule, while the command chown -R
file, which performs a tree traversal, is not. (The latter is
described in the third area, below.)
If it is explicitly intended that the command operate on the
symbolic link instead of following the symbolic link—for example,
it is desired that chown slink change the ownership of the file
that slink is, whether it is a symbolic link or not—then the -h
option should be used. In the above example, chown root slink
would change the ownership of the file referred to by slink,
while chown -h root slink would change the ownership of slink
itself.
There are some exceptions to this rule:
* The mv(1) and rm(1) commands do not follow symbolic links named
as arguments, but respectively attempt to rename and delete
them. (Note, if the symbolic link references a file via a
relative path, moving it to another directory may very well
cause it to stop working, since the path may no longer be
correct.)
* The ls(1) command is also an exception to this rule. For
compatibility with historic systems (when ls(1) is not doing a
tree walk—that is, -R option is not specified), the ls(1)
command follows symbolic links named as arguments if the -H or
-L option is specified, or if the -F, -d, or -l options are not
specified. (The ls(1) command is the only command where the -H
and -L options affect its behavior even though it is not doing
a walk of a file tree.)
* The file(1) command is also an exception to this rule. The
file(1) command does not follow symbolic links named as
argument by default. The file(1) command does follow symbolic
links named as argument if the -L option is specified.
Commands traversing a file tree
The following commands either optionally or always traverse file
trees: chgrp(1), chmod(1), chown(1), cp(1), du(1), find(1),
ls(1), pax(1), rm(1), and tar(1).
It is important to realize that the following rules apply equally
to symbolic links encountered during the file tree traversal and
symbolic links listed as command-line arguments.
The first rule applies to symbolic links that reference files
other than directories. Operations that apply to symbolic links
are performed on the links themselves, but otherwise the links
are ignored.
The command rm -r slink directory will remove slink, as well as
any symbolic links encountered in the tree traversal of
directory, because symbolic links may be removed. In no case
will rm(1) affect the file referred to by slink.
The second rule applies to symbolic links that refer to
directories. Symbolic links that refer to directories are never
followed by default. This is often referred to as a "physical"
walk, as opposed to a "logical" walk (where symbolic links that
refer to directories are followed).
Certain conventions are (should be) followed as consistently as
possible by commands that perform file tree walks:
* A command can be made to follow any symbolic links named on the
command line, regardless of the type of file they reference, by
specifying the -H (for "half-logical") flag. This flag is
intended to make the command-line name space look like the
logical name space. (Note, for commands that do not always do
file tree traversals, the -H flag will be ignored if the -R
flag is not also specified.)
For example, the command chown -HR user slink will traverse the
file hierarchy rooted in the file pointed to by slink. Note,
the -H is not the same as the previously discussed -h flag.
The -H flag causes symbolic links specified on the command line
to be dereferenced for the purposes of both the action to be
performed and the tree walk, and it is as if the user had
specified the name of the file to which the symbolic link
pointed.
* A command can be made to follow any symbolic links named on the
command line, as well as any symbolic links encountered during
the traversal, regardless of the type of file they reference,
by specifying the -L (for "logical") flag. This flag is
intended to make the entire name space look like the logical
name space. (Note, for commands that do not always do file
tree traversals, the -L flag will be ignored if the -R flag is
not also specified.)
For example, the command chown -LR user slink will change the
owner of the file referred to by slink. If slink refers to a
directory, chown will traverse the file hierarchy rooted in the
directory that it references. In addition, if any symbolic
links are encountered in any file tree that chown traverses,
they will be treated in the same fashion as slink.
* A command can be made to provide the default behavior by
specifying the -P (for "physical") flag. This flag is intended
to make the entire name space look like the physical name
space.
For commands that do not by default do file tree traversals, the
-H, -L, and -P flags are ignored if the -R flag is not also
specified. In addition, you may specify the -H, -L, and -P
options more than once; the last one specified determines the
command's behavior. This is intended to permit you to alias
commands to behave one way or the other, and then override that
behavior on the command line.
The ls(1) and rm(1) commands have exceptions to these rules:
* The rm(1) command operates on the symbolic link, and not the
file it references, and therefore never follows a symbolic
link. The rm(1) command does not support the -H, -L, or -P
options.
* To maintain compatibility with historic systems, the ls(1)
command acts a little differently. If you do not specify the
-F, -d, or -l options, ls(1) will follow symbolic links
specified on the command line. If the -L flag is specified,
ls(1) follows all symbolic links, regardless of their type,
whether specified on the command line or encountered in the
tree walk.
