Bash Pitfalls

This page shows common errors that Bash programmers make. These examples are all flawed in some way.

You will save yourself from many of these pitfalls if you simply always use quotes and never use WordSplitting for any reason! Word splitting is a broken legacy misfeature inherited from the Bourne shell that's stuck on by default if you don't quote expansions. The vast majority of pitfalls are in some way related to unquoted expansions, and the ensuing word splitting and globbing that result.

1. for i in $(ls *.mp3)

One of the most common mistakes BASH programmers make is to write a loop like this:

for i in $(ls *.mp3); do    # Wrong!
    some command $i         # Wrong!
done

for i in $(ls)              # Wrong!
for i in `ls`               # Wrong!

for i in $(find . -type f)  # Wrong!
for i in `find . -type f`   # Wrong!

files=($(find . -type f))   # Wrong!
for i in ${files[@]}        # Wrong!

Yes, it would be great if you could just treat the output of ls or find as a list of filenames and iterate over it. But you cannot. This entire approach is fatally flawed, and there is no trick that can make it work. You must use an entirely different approach.

There are at least 5 problems with this:

1. If a filename contains whitespace, it undergoes WordSplitting. Assuming we have a file named 01 - Don't Eat the Yellow Snow.mp3 in the current directory, the for loop will iterate over each word in the resulting file name: 01, -, Don't, Eat, etc.

2. If a filename contains glob characters, it undergoes filename expansion ("globbing"). If ls produces any output containing a * character, the word containing it will become recognized as a pattern and substituted with a list of all filenames that match it.

3. If the command substitution returns multiple filenames, there is no way to tell where the first one ends and the second one begins. Pathnames may contain any character except NUL. Yes, this includes newlines.

4. The ls utility may mangle filenames. Depending on which platform you're on, which arguments you used (or didn't use), and whether its standard output is pointing to a terminal or not, ls may randomly decide to replace certain characters in a filename with "?", or simply not print them at all. Never try to parse the output of ls.

5. The CommandSubstitution strips all trailing newline characters from its output. That may seem desirable since ls adds a newline, but if the last filename in the list ends with a newline, `...` or $() will remove that one also.

You can't simply double-quote the substitution either:

for i in "$(ls *.mp3)"; do # Wrong!

This causes the entire output of ls to be treated as a single word. Instead of iterating over each file name, the loop will only execute once, assigning to i a string with all the filenames rammed together.

Nor can you simply change IFS to a newline. Filenames can also contain newlines.

Another variation on this theme is abusing word splitting and a for loop to (incorrectly) read lines of a file. For example:

IFS=$'\n'
for line in $(cat file); do ...     # Wrong!

This doesn't work! Especially if those lines are filenames. Bash (or any other Bourne family shell) just doesn't work this way.

In addition to all of the above, the use of ls is just plain unnecessary. It's an external command whose output is intended specifically to be read by a human, not parsed by a script.

So, what's the right way to do it?

for i in *.mp3; do    # Better! and...
    some command "$i" # ...always double-quote expansions!
done

POSIX shells such as Bash have the globbing feature specifically for this purpose -- to allow the shell to expand patterns into a list of matching filenames. There is no need to interpret the results of an external utility. Because globbing is the very last expansion step, each match of the *.mp3 pattern correctly expands to a separate word, and isn't subject to the effects of an unquoted expansion. (If you need to process files recursively, see UsingFind, or consider using the shopt -s globstar feature in Bash 4 and higher.)

Question: What happens if there are no *.mp3-files in the current directory? Then the for loop is executed once, with i="*.mp3", which is not the expected behavior! The workaround is to test whether there is a matching file:

# POSIX
for i in *.mp3; do
    [ -e "$i" ] || continue
    some command "$i"
done

Another solution is to use Bash's shopt -s nullglob feature, though this should only be done after reading the documentation and carefully considering the effect of this setting on all other globs in the script.

Note the quotes around $i in the loop body above. This leads to our second pitfall:

2. cp $file $target

What's wrong with the command shown above? Well, nothing, if you happen to know in advance that $file and $target have no white space or wildcards in them. However, the results of the expansions are still subject to WordSplitting and pathname expansion. Always double-quote parameter expansions.

cp -- "$file" "$target"

Without the double quotes, you'll get a command like cp 01 - Don't Eat the Yellow Snow.mp3 /mnt/usb, which will result in errors like cp: cannot stat `01': No such file or directory. If $file has wildcards in it (* or ? or [), they will be expanded if there are files that match them. With the double quotes, all's well, unless "$file" happens to start with a -, in which case cp thinks you're trying to feed it command line options (See pitfall #3 below.)

Even in the somewhat uncommon circumstance that you can guarantee the variable contents, it is conventional and good practice to quote parameter expansions, especially if they contain file names. Experienced script writers will always use quotes except perhaps for a small number of cases in which it is absolutely obvious from the immediate code context that a parameter contains a guaranteed safe value. Experts will most likely consider the cp command in the title always wrong. You should too.

3. Filenames with leading dashes

Filenames with leading dashes can cause many problems. Globs like *.mp3 are sorted into an expanded list (according to your current locale), and - sorts before letters in most locales. The list is then passed to some command, which may incorrectly interpret the -filename as an option. There are two major solutions to this.

One solution is to insert -- between the command (like cp) and its arguments. That tells it to stop scanning for options, and all is well:

cp -- "$file" "$target"

There are potential problems with this approach. You have to be sure to insert -- for every usage of the parameter in a context where it might possibly be interpreted as an option -- which is easy to miss and may involve a lot of redundancy.

Most well-written option parsing libraries understand this, and the programs that use them correctly should inherit that feature for free. However, still be aware that it is ultimately up to the application to recognize end of options. Some programs that manually parse options, or do it incorrectly, or use poor 3rd-party libraries may not recognize it. Standard utilities should, with a few exceptions that are specified by POSIX. echo is one example.

Another option is to ensure that your filenames always begin with a directory by using relative or absolute pathnames.

for i in ./*.mp3; do
    cp "$i" /target
    ...
done

In this case, even if we have a file whose name begins with -, the glob will ensure that the variable always contains something like ./-foo.mp3, which is perfectly safe as far as cp is concerned.

Finally, if you can guarantee that all results will have the same prefix, and are only using the variable a few times within a loop body, you can simply concatenate the prefix with the expansion. This gives a theoretical savings in generating and storing a few extra characters for each word.

for i in *.mp3; do
    cp "./$i" /target
    ...
done

4. [ $foo = "bar" ]

This is very similar to the issue in pitfall #2, but I repeat it because it's so important. In the example above, the quotes are in the wrong place. You do not need to quote a string literal in bash (unless it contains metacharacters or pattern characters). But you should quote your variables if you aren't sure whether they could contain white space or wildcards.

This example can break for several reasons:

• If a variable referenced in [ doesn't exist, or is blank, then the [ command would end up looking like:

• [ = "bar" ] # Wrong!

• ...and will throw the error: unary operator expected. (The = operator is binary, not unary, so the [ command is rather shocked to see it there.)

• If the variable contains internal whitespace, then it gets split into separate words before the [ command sees it. Thus:

• [ multiple words here = "bar" ]

• While that may look OK to you, it's a syntax error as far as [ is concerned. The correct way to write this is:

• # POSIX
  [ "$foo" = bar ] # Right!

• This works fine on POSIX-conformant implementations even if $foo begins with a -, because POSIX [ determines its action depending on the number of arguments passed to it. Only very ancient shells have a problem with this, and you shouldn't worry about them when writing new code (see the x"$foo" workaround below).

In Bash and many other ksh-like shells, there is a superior alternative which uses the [[ keyword.

# Bash / Ksh
[[ $foo == bar ]] # Right!

You don't need to quote variable references on the left-hand side of = in [[ ]] because they don't undergo word splitting or globbing, and even blank variables will be handled correctly. On the other hand, quoting them won't hurt anything either. Unlike [ and test, you may also use the identical ==. Do note however that comparisons using [[ perform pattern matching against the string on the right hand side, not just a plain string comparison. To make the string on the right literal, you must quote it if any characters that have special meaning in pattern matching contexts are used.

# Bash / Ksh
match=b*r
[[ $foo == "$match" ]] # Good! Unquoted would also match against the pattern b*r.

You may have seen code like this:

# POSIX / Bourne
[ x"$foo" = xbar ] # Ok, but usually unnecessary.

The x"$foo" hack is required for code that must run on very ancient shells which lack [[, and have a more primitive [, which gets confused if $foo begins with a -. On said older systems, [ still doesn't care whether the token on the right hand side of the = begins with a -. It just uses it literally. It's just the left-hand side that needs extra caution.

Note that shells that require this workaround are not POSIX-conforming. Even the Heirloom Bourne shell doesn't require this (probably the non-POSIX Bourne shell clone that's still most widely in use as a system shell). Such extreme portability is rarely a requirement and makes your code less readable (and uglier).

5. cd $(dirname "$f")

This is yet another quoting error. As with a variable expansion, the result of a CommandSubstitution undergoes WordSplitting and pathname expansion. So you should quote it:

cd -P -- "$(dirname -- "$f")"

What's not obvious here is how the quotes nest. A C programmer reading this would expect the first and second double-quotes to be grouped together; and then the third and fourth. But that's not the case in Bash. Bash treats the double-quotes inside the command substitution as one pair, and the double-quotes outside the substitution as another pair.

Another way of writing this: the parser treats the command substitution as a "nesting level", and the quotes inside it are separate from the quotes outside it.

6. [ "$foo" = bar && "$bar" = foo ]

You can't use && inside the old test (or [) command. The Bash parser sees && outside of [[ ]] or (( )) and breaks your command into two commands, before and after the &&. Use one of these instead:

[ bar = "$foo" ] && [ foo = "$bar" ] # Right! (POSIX)
[[ $foo = bar && $bar = foo ]]       # Also right! (Bash / Ksh)

(Note that we reversed the constant and the variable inside [ for the legacy reasons discussed in pitfall #4. We could also have reversed the [[ case, but the expansions would require quoting to prevent interpretation as a pattern.) The same thing applies to ||. Either use [[ instead, or use two [ commands.

Avoid this:

[ bar = "$foo" -a foo = "$bar" ] # Not portable.

The binary -a and -o, and ( / ) (grouping) operators are XSI extensions to the POSIX standard. All are marked as obsolescent in POSIX-2008. They should not be used in new code. One of the practical problems with [ A = B -a C = D ] (or -o) is that POSIX does not specify the results of a test or [ command with more than 4 arguments. It probably works in most shells, but you can't count on it. If you have to write for POSIX shells, then you should use two test or [ commands separated by a && operator instead.

7. [[ $foo > 7 ]]

There are multiple issues here. First, the [[ command should not be used solely for evaluating arithmetic expressions. It should be used for test expressions involving one of the supported test operators. Though technically you can do math using some of [['s operators, it only makes sense to do so in conjunction with one of the non-math test operators somewhere in the expression. If you just want to do a numeric comparison (or any other shell arithmetic), it is much better to just use (( )) instead:

# Bash / Ksh
((foo > 7))     # Right!
[[ foo -gt 7 ]] # Works, but is pointless. Most will consider it wrong. Use ((...)) or let instead.

If you use the > operator inside [[ ]], it's treated as a string comparison (test for collation order by locale), not an integer comparison. This may work sometimes, but it will fail when you least expect it. If you use > inside [ ], it's even worse: it's an output redirection. You'll get a file named 7 in your directory, and the test will succeed as long as $foo is not empty.

If strict POSIX-conformance is a requirement, and (( is not available, then the correct alternative using old-style [ is

# POSIX
[ "$foo" -gt 7 ]       # Also right!
[ $((foo > 7)) -ne 0 ] # POSIX-compatible equivalent to ((, for more general math operations.

Note that the test ... -gt command will fail in interesting ways if $foo is not an integer. Therefore, there's not much point in quoting it properly other than for performance and to confine the arguments to a single word to reduce the likelihood of obscure side-effects possible in some shells.

If the input to any arithmetic context (including (( or let), or [ test expression involving numeric comparisons can't be guaranteed then you must always validate your input before evaluating the expression.

# POSIX
case $foo in
    *[^[:digit:]]*)
        printf '$foo expanded to a non-digit: %s\n' "$foo" >&2
        exit 1
        ;;
    *)
        [ $foo -gt 7 ]
esac

8. grep foo bar | while read -r; do ((count++)); done

The code above looks OK at first glance, doesn't it? Sure, it's just a poor implementation of grep -c, but it's intended as a simplistic example. Changes to count won't propagate outside the while loop because each command in a pipeline is executed in a separate SubShell. This surprises almost every Bash beginner at some point.

POSIX doesn't specify whether or not the last element of a pipeline is evaluated in a subshell. Some shells such as ksh93 and Bash >= 4.2 with shopt -s lastpipe enabled will run the while loop in this example in the original shell process, allowing any side-effects within to take effect. Therefore, portable scripts must be written in such a way as to not depend upon either behavior.

For workarounds for this and similar issues, please see Bash FAQ #24. It's a bit too long to fit here.

9. if [grep foo myfile]

Many beginners have an incorrect intuition about if statements brought about by seeing the very common pattern of an if keyword followed immediately by a [ or [[. This convinces people that the [ is somehow part of the if statement's syntax, just like parentheses used in C's if statement.

This is not the case! if takes a command. [ is a command, not a syntax marker for the if statement. It's equivalent to the test command, except that the final argument must be a ]. For example:

# POSIX
if [ false ]; then echo "HELP"; fi
if test false; then echo "HELP"; fi

are equivalent -- both checking that the argument "false" is non-empty. In both cases HELP will always be printed, to the surprise of programmers from other languages guessing about shell syntax.

The syntax of an if statement is:

if COMMANDS
then <COMMANDS>
elif <COMMANDS> # optional
then <COMMANDS>
else <COMMANDS> # optional
fi # required

Once again, [ is a command. It takes arguments like any other regular simple command. if is a compound command which contains other commands -- and there is no [ in its syntax!

While bash has a builtin command [ and thus knows about [ it has nothing special to do with ]. Bash only passes ] as argument to the [ command, which requires ] to be the last argument only to make scripts look better.

There may be zero or more optional elif sections, and one optional else section.

The if compound command is made up of two or more sections containing lists of commands, each delimited by a then, elif, or else keyword, and is terminated by the fi keyword. The exit status of the final command of the first section and each subsequent elif section determines whether each corresponding then section is evaluated. Another elif is evaluated until one of the then sections is executed. If no then section is evaluated, then the else branch is taken, or if no else is given, the if block is complete and the overall if command returns 0 (true).

If you want to make a decision based on the output of a grep command, you do not want to enclose it in parentheses, brackets, backticks, or any other syntax! Just use grep as the COMMANDS after the if, like this:

if grep -q fooregex myfile; then
...
fi

If the grep matches a line from myfile, then the exit code will be 0 (true), and the then part will be executed. Otherwise, if there are no matches, grep will return non-zero and the overall if command will be zero.

See also:

• BashGuide/TestsAndConditionals

• http://wiki.bash-hackers.org/syntax/ccmd/if_clause

10. if [bar="$foo"]; then ...

[bar="$foo"]   # Wrong!
[ bar="$foo" ] # Still wrong!

As explained in the previous example, [ is a command (which can be proven with type -t [ or whence -v [). Just like with any other simple command, Bash expects the command to be followed by a space, then the first argument, then another space, etc. You can't just run things all together without putting the spaces in! Here is the correct way:

if [ bar = "$foo" ]; then ...

Each of bar, =, the expansion of "$foo", and ] is a separate argument to the [ command. There must be whitespace between each pair of arguments, so the shell knows where each argument begins and ends.

11. if [ [ a = b ] && [ c = d ] ]; then ...

Here we go again. [ is a command. It is not a syntactic marker that sits between if and some sort of C-like "condition". Nor is it used for grouping. You cannot take C-like if commands and translate them into Bash commands just by replacing parentheses with square brackets!

If you want to express a compound conditional, do this:

if [ a = b ] && [ c = d ]; then ...

Note that here we have two commands after the if, joined by an && (logical AND, shortcut evaluation) operator. It's precisely the same as:

if test a = b && test c = d; then ...

If the first test command returns false, the body of the if statement is not entered. If it returns true, then the second test command is run; and if that also one returns true, then the body of the if statement will be entered. (C programmers are already familiar with &&. Bash uses the same short-circuit evaluation. Likewise || does short-circuit evaluation for the OR operation.)

The [[ keyword does permit the use of &&, so it could also be written this way:

if [[ a = b && c = d ]]; then ...

See pitfall #6 for a pitfall related to tests combined with conditional operators.

12. read $foo

You don't use a $ before the variable name in a read command. If you want to put data into the variable named foo, you do it like this:

•  read foo

Or more safely:

•  IFS= read -r foo

read $foo would read a line of input and put it in the variable(s) whose name(s) are in $foo. This might be useful if you actually intended foo to be a reference to some other variable; but in the majority of cases, this is simply a bug.

13. cat file | sed s/foo/bar/ > file

You cannot read from a file and write to it in the same pipeline. Depending on what your pipeline does, the file may be clobbered (to 0 bytes, or possibly to a number of bytes equal to the size of your operating system's pipeline buffer), or it may grow until it fills the available disk space, or reaches your operating system's file size limitation, or your quota, etc.

If you want to make a change to a file safely, other than appending to the end of it, use a text editor.

•  printf %s\\n ',s/foo/bar/g' w q | ed -s file

If you are doing something that cannot be done with a text editor there must be a temporary file created at some point(*). For example, the following is completely portable:

•  sed 's/foo/bar/g' file > tmpfile && mv tmpfile file

The following will only work on GNU sed 4.x:

•  sed -i 's/foo/bar/g' file(s)

Note that this also creates a temporary file, and does the same sort of renaming trickery -- it just handles it transparently.

And the following equivalent command requires perl 5.x (which is probably more widely available than GNU sed 4.x):

•  perl -pi -e 's/foo/bar/g' file(s)

For more details on replacing contents of files, please see Bash FAQ #21.

(*) sponge from moreutils uses this example in its manual:

•  sed '...' file | grep '...' | sponge file

Rather than using a temporary file plus an atomic mv, this version "soaks up" (the actual description in the manual!) all the data, before opening and writing to the file. This version will cause data loss if the program or system crashes during the write operation, because there's no copy of the original file on disk at that point.

Using a temporary file + mv still incurs a slight risk of data loss in case of a system crash / power loss; to be 100% certain that either the old or the new file will survive a power loss, you must use sync before the mv.

14. echo $foo

This relatively innocent-looking command causes massive confusion. Because the $foo isn't quoted, it will not only be subject to WordSplitting, but also file globbing. This misleads Bash programmers into thinking their variables contain the wrong values, when in fact the variables are OK -- it's just the word splitting or filename expansion that's messing up their view of what's happening.

•  msg="Please enter a file name of the form *.zip"
   echo $msg

This message is split into words and any globs are expanded, such as the *.zip. What will your users think when they see this message:

•  Please enter a file name of the form freenfss.zip lw35nfss.zip

To demonstrate:

•  var="*.zip"   # var contains an asterisk, a period, and the word "zip"
   echo "$var"   # writes *.zip
   echo $var     # writes the list of files which end with .zip

In fact, the echo command cannot be used with absolute safety here. If the variable contains -n for example, echo will consider that an option, rather than data to be printed. The only absolutely sure way to print the value of a variable is using printf:

•  printf "%s\n" "$foo"

15. $foo=bar

No, you don't assign a variable by putting a $ in front of the variable name. This isn't perl.

16. foo = bar

No, you can't put spaces around the = when assigning to a variable. This isn't C. When you write foo = bar the shell splits it into three words. The first word, foo, is taken as the command name. The second and third become the arguments to that command.

Likewise, the following are also wrong:

•  foo= bar    # WRONG!
   foo =bar    # WRONG!
   $foo = bar; # COMPLETELY WRONG!
  
   foo=bar     # Right.
   foo="bar"   # More Right.

17. echo <<EOF

A here document is a useful tool for embedding large blocks of textual data in a script. It causes a redirection of the lines of text in the script to the standard input of a command. Unfortunately, echo is not a command which reads from stdin.

•   # This is wrong:
    echo <<EOF
    Hello world
    How's it going?
    EOF
  
    # This is what you were trying to do:
    cat <<EOF
    Hello world
    How's it going?
    EOF
  
    # Or, use quotes which can span multiple lines (efficient, echo is built-in):
    echo "Hello world
    How's it going?"

Using quotes like that is fine -- it works great, in all shells -- but it doesn't let you just drop a block of lines into the script. There's syntactic markup on the first and last line. If you want to have your lines untouched by shell syntax, and don't want to spawn a cat command, here's another alternative:

•   # Or use printf (also efficient, printf is built-in):
    printf %s "\
    Hello world
    How's it going?
    "

In the printf example, the \ on the first line prevents an extra newline at the beginning of the text block. There's a literal newline at the end (because the final quote is on a new line). The lack of \n in the printf format argument prevents printf adding an extra newline at the end. The \ trick won't work in single quotes. If you need/want single quotes around the block of text, you have two choices, both of which necessitate shell syntax "contaminating" your data:

•   printf %s \
    'Hello world
    '
  
    printf %s 'Hello world
    '

18. su -c 'some command'

This syntax is almost correct. The problem is, on many platforms, su takes a -c argument, but it's not the one you want. For example, on OpenBSD:

•  $ su -c 'echo hello'
   su: only the superuser may specify a login class

You want to pass -c 'some command' to a shell, which means you need a username before the -c.

•  su root -c 'some command' # Now it's right.

su assumes a username of root when you omit one, but this falls on its face when you want to pass a command to the shell afterward. You must supply the username in this case.

19. cd /foo; bar

If you don't check for errors from the cd command, you might end up executing bar in the wrong place. This could be a major disaster, if for example bar happens to be rm -f *.

You must always check for errors from a cd command. The simplest way to do that is:

•  cd /foo && bar

If there's more than just one command after the cd, you might prefer this:

•  cd /foo || exit 1
   bar
   baz
   bat ... # Lots of commands.

cd will report the failure to change directories, with a stderr message such as "bash: cd: /foo: No such file or directory". If you want to add your own message in stdout, however, you could use command grouping:

•  cd /net || { echo "Can't read /net. Make sure you've logged in to the Samba network, and try again."; exit 1; }
   do_stuff
   more_stuff

Note there's a required space between { and echo, and a required ; before the closing }.

Some people also like to enable set -e to make their scripts abort on any command that returns non-zero, but this can be rather tricky to use correctly (since many common commands may return a non-zero for a warning condition, which you may not want to treat as fatal).

By the way, if you're changing directories a lot in a Bash script, be sure to read the Bash help on pushd, popd, and dirs. Perhaps all that code you wrote to manage cd's and pwd's is completely unnecessary.

Speaking of which, compare this:

•  find ... -type d -print0 | while IFS= read -r -d '' subdir; do
     here=$PWD
     cd "$subdir" && whatever && ...
     cd "$here"
   done

With this:

•  find ... -type d -print0 | while IFS= read -r -d '' subdir; do
     (cd "$subdir" || exit; whatever; ...)
   done

Forcing a SubShell here causes the cd to occur only in the subshell; for the next iteration of the loop, we're back to our normal location, regardless of whether the cd succeeded or failed. We don't have to change back manually, and we aren't stuck in a neverending string of ... && ... logic preventing the use of other conditionals. The subshell version is simpler and cleaner (albeit a tiny bit slower).

20. [ bar == "$foo" ]

The == operator is not valid for the [ command. Use = or the [[ keyword instead.

•  [ bar = "$foo" ] && echo yes
   [[ bar == $foo ]] && echo yes

21. for i in {1..10}; do ./something &; done

You cannot put a ; immediately after an &. Just remove the extraneous ; entirely.

•  for i in {1..10}; do ./something & done

Or:

•  for i in {1..10}; do
     ./something &
   done

& already functions as a command terminator, just like ; does. And you cannot mix the two.

In general, a ; can be replaced by a newline, but not all newlines can be replaced by ;.

22. cmd1 && cmd2 || cmd3

Some people like to use && and || as a shortcut syntax for if ... then ... else ... fi. In many cases, this is perfectly safe:

•  [[ -s $errorlog ]] && echo "Uh oh, there were some errors." || echo "Successful."

However, this construct is not completely equivalent to if ... fi in the general case, because the command that comes after the && also generates an exit status. And if that exit status isn't "true" (0), then the command that comes after the || will also be invoked. For example:

•  i=0
   true && ((i++)) || ((i--))
   echo $i # Prints 0

What happened here? It looks like i should be 1, but it ends up 0. Why? Because both the i++ and the i-- were executed. The ((i++)) command has an exit status, and that exit status is derived from a C-like evaluation of the expression inside the parentheses. That expression's value happens to be 0 (the initial value of i), and in C, an expression with an integer value of 0 is considered false. So ((i++)) (when i is 0) has an exit status of 1 (false), and therefore the ((i--)) command is executed as well.

This does not occur if we use the pre-increment operator, since the exit status from ++i is true:

•  i=0
   true && (( ++i )) || (( --i ))
   echo $i # Prints 1

But that's missing the point of the example. It just happens to work by coincidence, and you cannot rely on x && y || z if y has any chance of failure! (This example fails if we initialize i to -1 instead of 0.)

If you need safety, or if you simply aren't sure how this works, or if anything in the preceding paragraphs wasn't completely clear, please just use the simple if ... fi syntax in your programs.

•  i=0
   if true; then
     ((i++))
   else
     ((i--))
   fi
   echo $i # Prints 1

This section also applies to Bourne shell, here is the code that illustrates it:

•  true && { echo true; false; } || { echo false; true; }

Output is two lines "true" and "false", instead the single line "true".

23. echo "Hello World!"

The problem here is that, in an interactive Bash shell, you'll see an error like:

•  bash: !": event not found

This is because, in the default settings for an interactive shell, Bash performs csh-style history expansion using the exclamation point. This is not a problem in shell scripts; only in interactive shells.

Unfortunately, the obvious attempt to "fix" this won't work:

•  $ echo "hi\!"
   hi\!

The easiest solution is unsetting the histexpand option: this can be done with set +H or set +o histexpand

• Question: Why is playing with histexpand more apropriate than single quotes?

  • I personally ran into this issue when I was manipulating song files, using commands like

    mp3info -t "Don't Let It Show" ...
    mp3info -t "Ah! Leah!" ...

    Using single quotes is extremely inconvenient because of all the songs with apostrophes in their titles. Using double quotes ran into the history expansion issue. (And imagine a file that has both in its name. The quoting would be atrocious.) Since I never actually use history expansion, my personal preference was to turn it off in ~/.bashrc. -- GreyCat

These solutions will work:

•  echo 'Hello World!'

or

•  set +H
   echo "Hello World!"

or

•  histchars=

Many people simply choose to put set +H or set +o histexpand in their ~/.bashrc to deactivate history expansion permanently. This is a personal preference, though, and you should choose whatever works best for you.

Another solution is:

•  exmark='!'
   echo "Hello, world$exmark"

24. for arg in $*

Bash (like all Bourne shells) has a special syntax for referring to the list of positional parameters one at a time, and $* isn't it. Neither is $@. Both of those expand to the list of words in your script's parameters, not to each parameter as a separate word.

The correct syntax is:

•  for arg in "$@"
  
   # Or simply:
   for arg

Since looping over the positional parameters is such a common thing to do in scripts, for arg defaults to for arg in "$@". The double-quoted "$@" is special magic that causes each parameter to be used as a single word (or a single loop iteration). It's what you should be using at least 99% of the time.

Here's an example:

•  # Incorrect version
   for x in $*; do
     echo "parameter: '$x'"
   done
  
   $ ./myscript 'arg 1' arg2 arg3
   parameter: 'arg'
   parameter: '1'
   parameter: 'arg2'
   parameter: 'arg3'

It should have been written:

•  # Correct version
   for x in "$@"; do
     echo "parameter: '$x'"
   done
  
   $ ./myscript 'arg 1' arg2 arg3
   parameter: 'arg 1'
   parameter: 'arg2'
   parameter: 'arg3'

25. function foo()

This works in some shells, but not in others. You should never combine the keyword function with the parentheses () when defining a function.

Bash (at least some versions) will allow you to mix the two. Most of the shells won't accept that (zsh 4.x and perhaps above will - for example). Some shells will accept function foo, but for maximum portability, you should always use:

•  foo() {
    ...
   }

26. echo "~"

Tilde expansion only applies when '~' is unquoted. In this example echo writes '~' to stdout, rather than the path of the user's home directory.

Quoting path parameters that are expressed relative to a user's home directory should be done using $HOME rather than '~'. For instance consider the situation where $HOME is "/home/my photos".

•  "~/dir with spaces" # expands to "~/dir with spaces"
   ~"/dir with spaces" # expands to "~/dir with spaces"
   ~/"dir with spaces" # expands to "/home/my photos/dir with spaces"
   "$HOME/dir with spaces" # expands to "/home/my photos/dir with spaces"

27. local varname=$(command)

When declaring a local variable in a function, the local acts as a command in its own right. This can sometimes interact oddly with the rest of the line -- for example, if you wanted to capture the exit status ($?) of the CommandSubstitution, you can't do it. local's exit status masks it.

It's best to use separate commands for this:

•  local varname
   varname=$(command)
   rc=$?

The next pitfall describes another issue with this syntax:

28. export foo=~/bar

Tilde expansion (with or without a username) is only guaranteed to occur when the tilde appears at the beginning of a word, either by itself or followed by a slash. It is also guaranteed to occur when the tilde appears immediately after the = in an assignment.

However, the export and local commands do not constitute an assignment. So, in some shells (like Bash), export foo=~/bar will undergo tilde expansion; in others (like dash), it will not.

•  foo=~/bar; export foo    # Right!
   export foo="$HOME/bar"   # Right!

29. sed 's/$foo/good bye/'

In single quotes, bash parameter expansions like $foo do not get expanded. That is the purpose of single quotes, to protect characters like $ from the shell.

Change the quotes to double quotes:

•  foo="hello"; sed "s/$foo/good bye/"

But keep in mind, if you use double quotes you might need to use more escapes. See the Quotes page.

30. tr [A-Z] [a-z]

There are (at least) three things wrong here. The first problem is that [A-Z] and [a-z] are seen as globs by the shell. If you don't have any single-lettered filenames in your current directory, it'll seem like the command is correct; but if you do, things will go wrong. Probably at 0300 hours on a weekend.

The second problem is that this is not really the correct notation for tr. What this actually does is translate '[' into '['; anything in the range A-Z into a-z; and ']' into ']'. So you don't even need those brackets, and the first problem goes away.

The third problem is that depending on the locale, A-Z or a-z may not give you the 26 ASCII characters you were expecting. In fact, in some locales z is in the middle of the alphabet! The solution to this depends on what you want to happen:

•  # Use this if you want to change the case of the 26 latin letters
   LC_COLLATE=C tr A-Z a-z
  
   # Use this if you want the case conversion to depend upon the locale, which might be more like what a user is expecting
   tr '[:upper:]' '[:lower:]'

The quotes are required on the second command, to avoid globbing.

31. ps ax | grep gedit

The fundamental problem here is that the name of a running process is inherently unreliable. There could be more than one legitimate gedit process. There could be something else disguising itself as gedit (changing the reported name of an executed command is trivial). For real answers to this, see ProcessManagement.

The following is the quick and dirty stuff.

Searching for the PID of (for example) gedit, many people start with

$ ps ax | grep gedit
10530 ?        S      6:23 gedit
32118 pts/0    R+     0:00 grep gedit

which, depending on a RaceCondition, often yields grep itself as a result. To filter grep out:

ps ax | grep -v grep | grep gedit   # will work, but ugly

An alternative to this is to use:

ps ax | grep '[g]edit'              # quote to avoid shell GLOB

This will ignore the grep itself in the process table as that is [g]edit and grep is looking for gedit once evaluated.

On GNU/Linux, the parameter -C can be used instead to filter by commandname:

$ ps -C gedit
  PID TTY          TIME CMD
10530 ?        00:06:23 gedit

But why bother when you could just use pgrep instead?

$ pgrep gedit
10530

Now in a second step the PID is often extracted by awk or cut:

$ ps -C gedit | awk '{print $1}' | tail -n1

but even that can be handled by some of the trillions of parameters for ps:

$ ps -C gedit -opid=
10530

If you're stuck in 1992 and aren't using pgrep, you could use the ancient, obsolete, deprecated pidof (GNU/Linux only) instead:

$ pidof gedit
10530

and if you need the PID to kill the process, pkill might be interesting for you. Note however that, for example, pgrep/pkill ssh would also find processes named sshd, and you wouldn't want to kill those.

Unfortunately some programs aren't started with their name, for example firefox is often started as firefox-bin, which you would need to find out with - well - ps ax | grep firefox. Or, you can stick with pgrep by adding some parameters:

$ pgrep -fl firefox
3128 /usr/lib/firefox/firefox
7120 /usr/lib/firefox/plugin-container /usr/lib/flashplugin-installer/libflashplayer.so -greomni /usr/lib/firefox/omni.ja 3128 true plugin

Please read ProcessManagement. Seriously.

32. printf "$foo"

This isn't wrong because of quotes, but because of a format string exploit. If $foo is not strictly under your control, then any \ or % characters in the variable may cause undesired behavior.

Always supply your own format string:

printf %s "$foo"
printf '%s\n' "$foo"

33. for i in {1..$n}

The BashParser performs BraceExpansion before any other expansions or substitutions. So the brace expansion code sees the literal $n, which is not numeric, and therefore it doesn't expand the curly braces into a list of numbers. This makes it nearly impossible to use brace expansion to create lists whose size is only known at run-time.

Do this instead:

for ((i=1; i<=n; i++)); do
...
done

In the case of simple iteration over integers, an arithmetic for loop should almost always be preferred over brace expansion to begin with, because brace expansion pre-expands every argument which can be slower and unnecessarily consumes memory.

34. if [[ $foo = $bar ]] (depending on intent)

When the right-hand side of an = operator inside [[ is not quoted, bash does pattern matching against it, instead of treating it as a string. So, in the code above, if bar contains *, the result will always be true. If you want to check for equality of strings, the right-hand side should be quoted:

if [[ $foo = "$bar" ]]

If you want to do pattern matching, it might be wise to choose variable names that indicate the right-hand side contains a pattern. Or use comments.

It's also worth pointing out that if you quote the right-hand side of =~ it also forces a simple string comparison, rather than a regular expression matching. This leads us to:

35. if [[ $foo =~ 'some RE' ]]

The quotes around the right-hand side of the =~ operator cause it to become a string, rather than a RegularExpression. If you want to use a long or complicated regular expression and avoid lots of backslash escaping, put it in a variable:

re='some RE'
if [[ $foo =~ $re ]]

This also works around the difference in how =~ works across different versions of bash. Using a variable avoids some nasty and subtle problems.

The same problem occurs with pattern matching inside [[:

[[ $foo = "*.glob" ]]      # Wrong! *.glob is treated as a literal string.
[[ $foo = *.glob ]]        # Correct. *.glob is treated as a glob-style pattern.

36. [ -n $foo ] or [ -z $foo ]

When using the [ command, you must quote each substitution that you give it. Otherwise, $foo could expand to 0 words, or 42 words, or any number of words that isn't 1, which breaks the syntax.

[ -n "$foo" ]
[ -z "$foo" ]
[ -n "$(some command with a "$file" in it)" ]

# [[ doesn't perform word-splitting or glob expansion, so you could also use:
[[ -n $foo ]]
[[ -z $foo ]]

37. [[ -e "$broken_symlink" ]] returns 1 even though $broken_symlink exists

Test follows symlinks, therefore if a symlink is broken, i.e. it points to a file that doesn't exists, test -e returns 1 for it even though it exists.

In order to work around it (and prepare against it) you should use:

[[ -e "$broken_symlink" || -L "$broken_symlink" ]]

38. ed file <<<"g/d\{0,3\}/s//e/g" fails

The problem caused because ed doesn't accept 0 for \{0,3\}.

You can check that the following do work:

ed file <<<"g/d\{1,3\}/s//e/g"

Note that this happens even though POSIX states that BRE (which is the Regular Expression flavor used by ed) should accept 0 as the minimum number of occurrences (see section 5).

39. expr sub-string fails for "match"

This works reasonably well

---

most of the time

word=abcde
expr "$word" : ".\(.*\)"
bcde

But WILL fail for the word "match"

word=match
expr "$word" : ".\(.*\)"

The problem is "match" is a keyword. Solution (GNU only) is prefix with a '+'

word=match
expr + "$word" : ".\(.*\)"
atch

Or, y'know, stop using expr. You can do everything expr does by using Parameter Expansion. What's that thing up there trying to do? Remove the first letter of a word? That can be done in POSIX shells using PE or Substring Expansion:

$ word=match
$ echo "${word#?}"    # PE
atch
$ echo "${word:1}"    # SE
atch

Seriously, there's no excuse for using expr unless you're on Solaris with its non-POSIX-conforming /bin/sh. It's an external process, so it's much slower than in-process string manipulation. And since nobody uses it, nobody understands what it's doing, so your code is obfuscated and hard to maintain.

40. On UTF-8 and Byte-Order Marks (BOM)

In general: Unix UTF-8 text does not use BOM. The encoding of plain text is determined by the locale or by mime types or other metadata. While the presence of a BOM would not normally damage a UTF-8 document meant only for reading by humans, it is problematic (often syntactically illegal) in any text file meant to be interpreted by automated processes such as scripts, source code, configuration files, and so on. Files starting with BOM should be considered equally foreign as those with MS-DOS linebreaks.

In shell scripting: 'Where UTF-8 is used transparently in 8-bit environments, the use of a BOM will interfere with any protocol or file format that expects specific ASCII characters at the beginning, such as the use of "#!" of at the beginning of Unix shell scripts.' http://unicode.org/faq/utf_bom.html#bom5

41. content=$(<file)

There isn't anything wrong with this expression, but you should be aware that command substitutions (all forms: `...`, $(...), $(<file), `<file`, and ${ ...; } (ksh)) remove any trailing newlines. This is often inconsequential or even desirable, but if you must preserve the literal output including any possible trailing newlines, it gets tricky because you have no way of knowing whether the output had them or how many. One ugly but usable workaround is to add a postfix inside the command substitution and remove it on the outside:

absolute_dir_path_x=$(readlink -fn -- "$dir_path"; printf x)
absolute_dir_path=${absolute_dir_path_x%x}

A less portable but arguably prettier solution is to use read with an empty delimiter.

# Ksh (or bash 4.2+ with lastpipe enabled)
readlink -fn -- "$dir_path" | IFS= read -rd '' absolute_dir_path

The downside to this method is that the read will always return false unless the command outputs a NUL byte causing only part of the stream to be read. The only way to get the exit status of the command is through PIPESTATUS. You could also intentionally output a NUL byte to force read to return true, and use pipefail.

set -o pipefail
{ readlink -fn -- "$dir_path"; printf '\0x'; } | IFS= read -rd '' absolute_dir_path

This is somewhat of a portability mess, as Bash supports both pipefail and PIPESTATUS, ksh93 supports pipefail only, and only recent versions of mksh support pipefail, while earlier versions supported PIPESTATUS only. Additionally, a bleeding-edge ksh93 version is required in order for read to stop at the NUL byte.

42. for file in ./* ; do if [[ $file != *.* ]]

One way to prevent programs from interpreting filenames passed to them as options is to use pathnames (see pitfall #3 above). For files under the current directory, names may be prefixed with a relative pathname ./.

In the case of a pattern like *.* however, problems can arise because it matches a string of the form ./filename. In a simple case, you can just use the glob directly to generate the desired matches. If however a separate pattern-matching step is required (e.g. the results have been preprocessed and stored in an array, and need to be filtered), it could be solved by taking the prefix into account in the pattern: [[ $file != ./*.* ]], or by stripping the pattern from the match.

# Bash
shopt -s nullglob
for path in ./*; do
    [[ ${path##*/} != *.* ]] && rm "$path"
done

# Or even better
for file in *; do
    [[ $file != *.* ]] && rm "./${file}"
done

# Or better still
for file in *.*; do
    rm "./${file}"
done

Another possibility is to signal the end of options with a -- argument. (Again, covered in #pf3).

shopt -s nullglob
for file in *; do
    [[ $file != *.* ]] && rm -- "$file"
done

43. somecmd 2>&1 >>logfile

This is by far the most common mistake involving redirections, typically performed by someone wanting to direct both stdout and stderr to a file or pipe will try this and not understand why stderr is still showing up on their terminal. If you're perplexed by this, you probably don't understand how redirections or possibly file descriptors work to begin with. Redirections are evaluated left-to-right before the command is executed. This semantically incorrect code essentially means: "first redirect standard error to where standard out is currently pointing (the tty), then redirect standard out to logfile". This is backwards. Standard error is already going to the tty. Use the following instead:

somecmd >>logfile 2>&1

See a more in-depth explanation, Copy descriptor explained, and BashGuide - redirection.

44. cmd; (( ! $? )) || die

$? is only required if you need to retrieve the exact status of the previous command. If you only need to test for success or failure (any non-zero status), just test the command directly. e.g.:

if cmd; then
    ...
fi

Checking an exit status against a list of alternatives might follow a pattern like this:

cmd
status=$?
case $status in
    0)
        echo success >&2
        ;;
    1)
        echo 'Must supply a parameter, exiting.' >&2
        exit 1
        ;;
    *)
        echo "Unknown error $status, exiting." >&2
        exit "$status"
esac

45. y=$(( array[$x] ))

Due to the POSIX wording of arithmetic expansion (which calls for expansion of command substitutions after parameter expansion), expansion of an array subscript inside an arithmetic expansion can lead to code injection exploits.

Yeah, that's a lot of big, confusing words. Here's how it breaks:

$ x='$(date >&2)'        # redirection is just so we can see everything happen
$ y=$((array[$x]))       # array doesn't even have to exist
Mon Jun  2 10:49:08 EDT 2014

Quoting "$x" won't help, either:

$ y=$((array["$x"]))
Mon Jun  2 10:51:03 EDT 2014

The two tricks that do work are:

# 1. Escape the $x so it isn't expanded prematurely.
$ y=$((array[\$x]))

# 2. Use the full ${array[$x]} syntax.
$ y=$((${array[$x]}))

46. read num; echo $((num+1))

Always validate your input (see BashFAQ/054) before using num in an arithmetic context as it allows code injection.

$ echo 'a[$(echo injection >&2)]' | bash -c 'read num; echo $((num+1))'
injection
1

47. IFS=, read -ra fields <<< "$csv_line"

Unbelievable as it may seem, POSIX requires the treatment of IFS as a field terminator, rather than a field separator. What this means in our example is that if there's an empty field at the end of the input line, it will be discarded:

$ IFS=, read -ra fields <<< "a,b,"
$ declare -p fields
declare -a fields='([0]="a" [1]="b")'

Where did the empty field go? It was eaten for historical reasons ("because it's always been that way"). This behavior is not unique to bash; all conformant shells do it. A non-empty field is properly scanned:

$ IFS=, read -ra fields <<< "a,b,c"
$ declare -p fields
declare -a fields='([0]="a" [1]="b" [2]="c")'

So, how do we work around this nonsense? As it turns out, appending an IFS character to the end of the input string will force the scanning to work. If there was a trailing empty field, the extra IFS character "terminates" it so that it gets scanned. If there was a trailing non-empty field, the IFS character creates a new, empty field that gets dropped.

$ input="a,b,"
$ IFS=, read -ra fields <<< "$input,"
$ declare -p fields
declare -a fields='([0]="a" [1]="b" [2]="")'

---

CategoryShell