如何在 POSIX sh 中获取脚本目录？

小开

@City responded that

DIR=$( cd -P -- "$(dirname -- "$(command -v -- "$0")")" && pwd -P )

works. I used it too.
I found the command at https://stackoverflow.com/questions/760110/can-i-get-the-absolute-path-to-the-cu‌rrent-script-in-kornshell.

小开

最佳答案

The POSIX-shell (sh) counterpart of $BASH_SOURCE is $0. ^{see bottom for background info}

Caveat: The crucial difference is that if your script is being sourced (loaded into the current shell with .), the snippets below will not work properly. ^{explanation further below}

^{Note that I've changed DIR to dir in the snippets below, because it's better not to use all-uppercase variable names so as to avoid clashes with environment variables and special shell variables.

The CDPATH= prefix takes the place of > /dev/null in the original command: $CDPATH is set to a null string so as to ensure that cd never echoes anything.}

In the simplest case, this will do (the equivalent of the OP's command):

dir=$(CDPATH= cd -- "$(dirname -- "$0")" && pwd)

If you also want to resolve the resulting directory path to its ultimate target in case the directory and/or its components are symlinks, add -P to the pwd command:

dir=$(CDPATH= cd -- "$(dirname -- "$0")" && pwd -P)

Caveat: This is NOT the same as finding the script's own true directory of origin:
Let's say your script foo is symlinked to /usr/local/bin/foo in the $PATH, but its true path is /foodir/bin/foo.
The above will still report /usr/local/bin, because the symlink resolution (-P) is applied to the directory, /usr/local/bin, rather than to the script itself.

To find the script's own true directory of origin, you'd have to inspect the script's path to see if it's a symlink and, if so, follow the (chain of) symlinks to the ultimate target file, and then extract the directory path from the target file's canonical path.

GNU's readlink -f (better: readlink -e) could do that for you, but readlink is not a POSIX utility.
^{While BSD platforms, including macOS, have a readlink utility too, on macOS it doesn't support -f's functionality. That said, to show how simple the task becomes if readlink -f is available:
dir=$(dirname "$(readlink -f -- "$0")").}

In fact, there is no POSIX utility for resolving file symlinks. There are ways to work around that, but they're cumbersome and not fully robust:

The following, POSIX-compliant shell function implements what GNU's readlink -e does and is a reasonably robust solution that only fails in two rare edge cases:

paths with embedded newlines (very rare)
filenames containing literal string -> (also rare)

With this function, named rreadlink, defined, the following determines the script's true directory path of origin:

dir=$(dirname -- "$(rreadlink "$0")")

Note: If you're willing to assume the presence of a (non-POSIX) readlink utility - which would cover macOS, FreeBSD and Linux - a similar, but simpler solution can be found in this answer to a related question.

rreadlink() source code - place before calls to it in scripts:

rreadlink() ( # Execute the function in a *subshell* to localize variables and the effect of `cd`.


target=$1 fname= targetDir= CDPATH=


# Try to make the execution environment as predictable as possible:
# All commands below are invoked via `command`, so we must make sure that `command`
# itself is not redefined as an alias or shell function.
# (Note that command is too inconsistent across shells, so we don't use it.)
# `command` is a *builtin* in bash, dash, ksh, zsh, and some platforms do not even have
# an external utility version of it (e.g, Ubuntu).
# `command` bypasses aliases and shell functions and also finds builtins
# in bash, dash, and ksh. In zsh, option POSIX_BUILTINS must be turned on for that
# to happen.
{ \unalias command; \unset -f command; } >/dev/null 2>&1
[ -n "$ZSH_VERSION" ] && options[POSIX_BUILTINS]=on # make zsh find *builtins* with `command` too.


while :; do # Resolve potential symlinks until the ultimate target is found.
[ -L "$target" ] || [ -e "$target" ] || { command printf '%s\n' "ERROR: '$target' does not exist." >&2; return 1; }
command cd "$(command dirname -- "$target")" # Change to target dir; necessary for correct resolution of target path.
fname=$(command basename -- "$target") # Extract filename.
[ "$fname" = '/' ] && fname='' # !! curiously, `basename /` returns '/'
if [ -L "$fname" ]; then
# Extract [next] target path, which may be defined
# *relative* to the symlink's own directory.
# Note: We parse `ls -l` output to find the symlink target
#       which is the only POSIX-compliant, albeit somewhat fragile, way.
target=$(command ls -l "$fname")
target=${target#* -> }
continue # Resolve [next] symlink target.
fi
break # Ultimate target reached.
done
targetDir=$(command pwd -P) # Get canonical dir. path
# Output the ultimate target's canonical path.
# Note that we manually resolve paths ending in /. and /.. to make sure we have a normalized path.
if [ "$fname" = '.' ]; then
command printf '%s\n' "${targetDir%/}"
elif  [ "$fname" = '..' ]; then
# Caveat: something like /var/.. will resolve to /private (assuming /var@ -> /private/var), i.e. the '..' is applied
# AFTER canonicalization.
command printf '%s\n' "$(command dirname -- "${targetDir}")"
else
command printf '%s\n' "${targetDir%/}/$fname"
fi
)

To be robust and predictable, the function uses command to ensure that only shell builtins or external utilities are called (ignores overloads in the forms of aliases and functions).
It's been tested in recent versions of the following shells: bash, dash, ksh, zsh.

How to handle sourced invocations:

tl;dr:

Using POSIX features only:

You cannot determine the script's path in a sourced invocation (except in zsh, which, however, doesn't usually act as sh).
You can detect whether or not your script is being sourced ONLY if your script is being sourced directly by the shell (such as in a shell profile/initialization file; possibly via a chain of sourcings), by comparing $0 to the shell executable name/path (except in zsh, where, as noted $0 is truly the current script's path). By contrast (except in zsh), a script being sourced from another script that itself was directly invoked, contains that script's path in $0.
To solve these problems, bash, ksh, and zsh have nonstandard features that do allow determining the actual script path even in sourced scenarios and also detecting whether a script is being sourced or not; for instance, in bash, $BASH_SOURCE always contains the running script's path, whether it's being sourced or not, and [[ $0 != "$BASH_SOURCE" ]] can be used to test whether the script is being sourced.
- This answer indirectly shows these techniques, in the context of determining whether a given script is being sourced.

To show why this cannot be done, let's analyze the command from Walter A's answer:

    # NOT recommended - see discussion below.
DIR=$( cd -P -- "$(dirname -- "$(command -v -- "$0")")" && pwd -P )

(Two asides:
- Using -P twice is redundant - it's sufficient to use it with pwd.
- The command is missing silencing of cd's potential stdout output, if $CDPATH happens to be set.)
command -v -- "$0"
- command -v -- "$0" is designed to cover one additional scenario: if the script is being sourced from an interactive shell, $0 typically contains the mere filename of the shell executable (sh), in which case dirname would simply return . (because that's what dirname invariably does when given a argument without a path component). command -v -- "$0" then returns that shell's absolute path through a $PATH lookup (/bin/sh). Note, however, that login shells on some platforms (e.g., OSX) have their filename prefixed with - in $0 (-sh), in which case command -v -- "$0" doesn't work as intended (returns an empty string).
- Conversely, command -v -- "$0" can misbehave in two non-sourced scenarios in which the shell executable, sh, is directly invoked, with the script as an argument:
  - if the script itself is not executable: command -v -- "$0" may return an empty string, depending on what specific shell acts as sh on a given system: bash, ksh, and zsh return an empty string; only dash echoes $0
    ^{The POSIX spec. for command doesn't explicitly say whether command -v, when applied to a filesystem path, should only return command -v0 - which is what bash, ksh, and zsh do - but you can argue that it is implied by the very purpose of command; curiously, dash, which is usually the most compliant POSIX citizen, is deviating from the standard here. By contrast, ksh is the lone model citizen here, because it is the only one that reports executable files only command -v1 reports them with an command -v2 (albeit not normalized) path, as the spec requires.}
  - if the script is executable, but not in the $PATH, and the invocation uses its mere filename (e.g., sh myScript), command -v -- "$0" will also return the empty string, except in dash.
- Given that the script's directory cannot be determined when the script is being sourced - because $0 then doesn't contain that information (except in zsh, which doesn't usually act as sh) - there's no good solution to this problem.
  - Returning the shell executable's directory path in that situation is of limited use - it is, after all, not the script's directory - except perhaps to later use that path in a test to determine whether or not the script is being sourced.
    - A more reliable approach would be to simply test $0 directly: [ "$0" = "sh" ] || [ "$0" = "-sh" ] || [ "$0" = "/bin/sh" ]
  - However, even that doesn't work if the script is being sourced from another script (that was itself directly invoked), because $0 then simply contains the sourcing script's path.
- Given the limited usefulness of command -v -- "$0" in sourced scenarios and the fact that it breaks two non-sourced scenarios, my vote is for NOT using it, which leaves us with:
  - All non-sourced scenarios are covered.
  - In sourced invocations, you cannot determine the script's path, and at best, in limited circumstances, you can detect whether or not sourcing is occurring:
    - When sourced directly by the shell (such as from a shell profile/initialization file), $dir ends up either containing ., if the shell executable was invoked as a mere filename (applying dirname to a mere filename always returns .), or the shell executable's directory path otherwise. . cannot be reliably distinguished from a non-sourced invocation from the current directory.
    - When sourced from another script (that was itself not also sourced), $0 contains that script's path, and the script being sourced has no way of telling whether that's the case.

Background information:

POSIX defines the behavior of $0 with respect to shell scripts here.

Essentially, $0 should reflect the path of the script file as specified, which implies:

Do NOT rely on $0 containing an absolute path.
$0 contains an absolute path only if:
you explicitly specify an absolute path; e.g.:
- ~/bin/myScript (assuming the script itself is executable)
- sh ~/bin/myScript
you invoke an executable script by mere filename, which requires that it both be executable and in the $PATH; behind the scenes, the system transforms myScript into an absolute path and then executes it; e.g.:
- myScript # executes /home/jdoe/bin/myScript, for instance
In all other cases, $0 will reflect the script path as specified:
- When explicitly invoking sh with a script, this can be a mere filename (e.g., sh myScript) or a relative path (e.g., sh ./myScript)
- When invoking an executable script directly, this can be a relative path (e.g., ./myScript - note that a mere filename would only find scripts in the $PATH).

In practice, ABC0, ABC1, ABC2, and zsh all exhibit this behavior.

By contrast, POSIX does NOT mandate the value of $0 when sourcing a script (using the special built-in utility . ("dot")), so you cannot rely on it, and, in practice, behavior differs across shells.

Thus, you cannot blindly use $0 when your script is being sourced and expect standardized behavior.
- In practice, bash, dash, and ksh leave $0 untouched when sourcing scripts, meaning that $0 contains the caller's $0 value, or, more accurately, the $0 value of the most recent caller in the call chain that hasn't been sourced itself; thus, $0 may point either to the shell's executable or to the path of dash0 (directly invoked) script that sourced the current one.
- By contrast, zsh, as the lone dissenter, actually does report the current script's path in $0. Conversely, $0 will provide no indication as to whether the script is being sourced or not.
- In short: using POSIX features only, you can neither tell reliably whether the script at hand is being sourced, nor what the script at hand's path is, nor what the relationship of $0 to the current script's path is.
If you do need to handle this situation, you must identify the specific shell at hand and access its specific non-standard features:
- bash, ksh, and zsh all offer their own ways of obtaining the running script's path, even when it's being sourced.

For the sake of completeness: the value of $0 in other contexts:

Inside a shell function, POSIX mandates that $0 remain unchanged; therefore, whatever value it has outside the function, it'll have inside as well.
- In practice, bash, dash, and ksh do behave that way.
- Again, zsh is the lone dissenter and reports the function's name.
In a shell that accepted a command string via the -c option on startup, it's the first operand (non-option argument) that sets $0; e.g.:
- sh -c 'echo \$0: $0 \$1: $1' foo one # -> '$0: foo $1: one'
- bash, dash, ksh, and zsh all behave that way.
Otherwise, in a shell not executing a script file, $0 is the value of the first argument that the shell's parent process passed - typically, that's the shell's name or path (e.g. sh, or /bin/sh); this includes:
- an interactive shell
  - Caveat: some platforms, notably OSX, always create login shells when creating interactive shells, and prepend - to the shell name before placing it in $0, so as to signal to the shell that it is a _login shell; thus, by default, $0 reports -bash, not bash, in interactive shells on OSX.
- a shell that reads commands from stdin
  - this also applies to piping a script file to the shell via stdin (e.g., sh < myScript)
- bash, dash, ksh, and zsh all behave that way.

小开

if      OLDPWD=/dev/fd/0 \
cd - && ls -lLidFH ?
then    cd . <8
fi      </proc/self/fd 8<. 9<$0

there. that should enable you to change directpry through some magic links as file descriptors.

setting $OLDPWD pre-cd exports the value for the duration of the one change directory (note: ABC1 can have residual effects on ABC3 tables, but the only ABC4 about which i am aware that actually males any good use of these is kevin ahlmquists - and since herbert xu - ABC5, and maybe some bsd stuff, but what do i know?) but does not carry over any cd exports as a result of the change.

thus, $OLDPWD does not change, actually, and if it had any value at all this remains as was. $PWD is changed as a result of the first cd, and the value becomes /dev/fd/0 which points to /proc/self/fd, where a list of file descriptors for our process should be in ., to include whatever $0 is on ./2.

so we do an ls ... ? and have a look at all of the wonderful info we can get, and we go whence we came.

yay!