Shell from vi

A good sign of a philosophically sound interactive Unix tool is the facilities it offers for interacting with the filesystem and the shell: specifically, how easily can you run file operations and/or shell commands with reference to data within the tool? The more straightforward this is, the more likely the tool will fit neatly into a terminal-driven Unix workflow.

If all else fails, you could always suspend the task with Ctrl+Z to drop to a shell, but it’s helpful if the tool shows more deference to the shell than that; it means you can use and (even more importantly) write tools to manipulate the data in the program in whatever languages you choose, rather than being forced to use any kind of heretical internal scripting language, or worse, an over-engineered API.

vi is a good example of a tool that interacts openly and easily with the Unix shell, allowing you to pass open buffers as streams of text transparently to classic filter and text processing tools. In the case of Vim, it’s particularly useful to get to know these, because in many cases they allow you to avoid painful Vimscript, and to do things your way, without having to learn an ad-hoc language or to rely on plugins. This was touched on briefly in the “Editing” article of the Unix as IDE series.

Choosing your shell

By default, vi will use the value of your SHELL environment variable as the shell in which your commands will be run. In most cases, this is probably what you want, but it might pay to check before you start:

:set shell?

If you’re using Bash, and this prints /bin/bash, you’re good to go, and you’ll be able to use Bash-specific features or builtins such as [[ comfortably in your command lines if you wish.

Running commands

You can run a shell command from vi with the ! ex command. This is inherited from the same behaviour in ed. A good example would be to read a manual page in the same terminal window without exiting or suspending vi:

:!man grep

Or to build your project:


You’ll find that exclamation point prefix ! shows up in the context of running external commands pretty consistently in vi.

You will probably need to press Enter afterwards to return to vi. This is to allow you to read any output remaining on your screen.

Of course, that’s not the only way to do it; you may prefer to drop to a forked shell with :sh, or suspend vi with ^Z to get back to the original shell, resuming it later with fg.

You can refer to the current buffer’s filename in the command with %, but be aware that this may cause escaping problems for files with special characters in their names:

:!gcc % -o foo

If you want a literal %, you will need to escape it with a backslash:

:!grep \% .vimrc

The same applies for the # character, for the alternate buffer.

:!gcc # -o bar
:!grep \# .vimrc

And for the ! character, which expands to the previous command:

:!echo !
:!echo \!

You can try to work around special characters for these expansions by single-quoting them:

:!gcc '%' -o foo
:!gcc '#' -o bar

But that’s still imperfect for files with apostrophes in their names. In Vim (but not vi) you can do this:

:exe "!gcc " . shellescape(expand("%")) . " -o foo"

The Vim help for this is at :help :!.

Reading the output of commands into a buffer

Also inherited from ed is reading the output of commands into a buffer, which is done by giving a command starting with ! as the argument to :r:

:r !grep vim .vimrc

This will insert the output of the command after the current line position in the buffer; it works in the same way as reading in a file directly.

You can add a line number prefix to :r to place the output after that line number:

:5r !grep vim .vimrc

To put the output at the very start of the file, a line number of 0 works:

:0r !grep vim .vimrc

And for the very end of the file, you’d use $:

:$r !grep vim .vimrc

Note that redirections work fine, too, if you want to prevent stderr from being written to your buffer in the case of errors:

:$r !grep vim .vimrc 2>>vim_errorlog

Writing buffer text into a command

To run a command with standard input coming from text in your buffer, but without deleting it or writing the output back into your buffer, you can provide a ! command as an argument to :w. Again, this behaviour is inherited from ed.

By default, the whole buffer is written to the command; you might initially expect that only the current line would be written, but this makes sense if you consider the usual behaviour of w when writing directly to a file.

Given a file with a first column full of numbers:

304 Donald Trump
227 Hillary Clinton
3   Colin Powell
1   Spotted Eagle
1   Ron Paul
1   John Kasich
1   Bernie Sanders

We could calculate and view (but not save) the sum of the first column with awk(1), to see the expected value of 538 printed to the terminal:

:w !awk '{sum+=$1}END{print sum}'

We could limit the operation to the faithless electoral votes by specifying a line range:

:3,$w !awk '{sum+=$1}END{print sum}'

You can also give a range of just ., if you only want to write out the current line.

In Vim, if you’re using visual mode, pressing : while you have some text selected will automatically add the '<,'> range marks for you, and you can write out the rest of the command:

:'<,'>w !grep Bernie

Note that this writes every line of your selection to the command, not merely the characters you have selected. It’s more intuitive to use visual line mode (Shift+V) if you take this approach.

Filtering text

If you want to replace text in your buffer by filtering it through a command, you can do this by providing a range to the ! command:

:1,2!tr '[:lower:]' '[:upper:]'

This example would capitalise the letters in the first two lines of the buffer, passing them as input to the command and replacing them with the command’s output.

3 Colin Powell
1 Spotted Eagle
1 Ron Paul
1 John Kasich
1 Bernie Sanders

Note that the number of lines passed as input need not match the number of lines of output. The length of the buffer can change. Note also that by default any stderr is included; you may want to redirect that away.

You can specify the entire file for such a filter with %:

:%!tr '[:lower:]' '[:upper:]'

As before, the current line must be explicitly specified with . if you want to use only that as input, otherwise you’ll just be running the command with no buffer interaction at all, per the first heading of this article:

:.!tr '[:lower:]' '[:upper:]'

You can also use ! as a motion rather than an ex command on a range of lines, by pressing ! in normal mode and then a motion (w, 3w, }, etc) to select all the lines you want to pass through the filter. Doubling it (!!) filters the current line, in a similar way to the yy and dd shortcuts, and you can provide a numeric prefix (e.g. 3!!) to specify a number of lines from the current line.

This is an example of a general approach that will work with any POSIX-compliant version of vi. In Vim, you have the gU command available to coerce text to uppercase, but this is not available in vanilla vi; the best you have is the tilde command ~ to toggle the case of the character under the cursor. tr(1), however, is specified by POSIX–including the locale-aware transformation–so you are much more likely to find it works on any modern Unix system.

If you end up needing such a command during editing a lot, you could make a generic command for your private bindir, say named upp for uppercase, that forces all of its standard input to uppercase:

tr '[:lower:]' '[:upper:]'

Once saved somewhere in $PATH and made executable, this would allow you simply to write the following to apply the filter to the entire buffer:


The main takeaway from this is that the scripts you use with your editor don’t have to be in shell. You might prefer Awk:

#!/usr/bin/awk -f
{ print toupper($0) }

Or Perl:

#!/usr/bin/env perl
print uc while <>;

Or Python, or Ruby, or Rust, or …

Incidentally, this “filtering” feature is where vi‘s heritage from ed ends as far as external commands are concerned. In POSIX ed, there isn’t a way to filter buffer text through a command in one hit. It’s not too hard to emulate it with a temporary file, though, using all the syntax learned above:

*1,2w !upp > tmp
*0r tmp
*!rm tmp

Bash hostname completion

As part of its programmable completion suite, Bash includes hostname completion. This completion mode reads hostnames from a file in hosts(5) format to find possible completions matching the current word. On Unix-like operating systems, it defaults to reading the file in its usual path at /etc/hosts.

For example, given the following hosts(5) file in place at /etc/hosts:      localhost www mx rad

An appropriate call to compgen would yield this output:

$ compgen -A hostname

We could then use this to complete hostnames for network diagnostic tools like ping(8):

$ complete -A hostname ping

Typing ping we and then pressing Tab would then complete to ping If the shopt option hostcomplete is on, which it is by default, Bash will also attempt host completion if completing any word with an @ character in it. This can be useful for email address completion or for SSH username@hostname completion.

We could also trigger hostname completion in any other Bash command line (regardless of complete settings) with the Readline shortcut Alt+@ (i.e. Alt+Shift+2). This works even if hostcomplete is turned off.

However, with DNS so widely deployed, and with system /etc/hosts files normally so brief on internet-connected systems, this may not seem terribly useful; you’d just end up completing localhost, and (somewhat erroneously) a few IPv6 addresses that don’t begin with a digit. It may seem even less useful if you have your own set of hosts in which you’re interested, since they may not correspond to the hosts in the system’s /etc/hosts file, and you probably really do want them looked up via DNS each time, rather than maintaining static addresses for them.

There’s a simple way to make host completion much more useful by defining the HOSTFILE variable in ~/.bashrc to point to any other file containing a list of hostnames. You could, for example, create a simple file ~/.hosts in your home directory, and then include this in your ~/.bashrc:

# Use a private mock hosts(5) file for completion

You could then populate the ~/.hosts file with a list of hostnames in which you’re interested, which will allow you to influence hostname completion usefully without messing with your system’s DNS resolution process at all. Because of the way the Bash HOSTFILE parsing works, you don’t even have to fake an IP address as the first field; it simply scans the file for any word that doesn’t start with a digit:

# Comments with leading hashes will be excluded router

You can even include other files from it with an $include directive!

$include /home/tom/.hosts.home
$include /home/tom/

Author’s note: This really surprised me when reading the source, because I don’t think /etc/hosts files generally support that for their usual name resolution function. I would love to know if any systems out there actually do support this.

The behaviour of the HOSTFILE variable is a bit weird; all of the hosts from the HOSTFILE are appended to the in-memory list of completion hosts each time the HOSTFILE variable is set (not even just changed), and host completion is attempted, even if the hostnames were already in the list. It’s probably sufficient just to set the file once in ~/.bashrc.

This setup allows you to set hostname completion as the default method for all sorts of network-poking tools, falling back on the usual filename completion if nothing matches with -o default:

$ complete -A hostname -o default curl dig host netcat ping telnet

You could also use hostname completions for ssh(1), but to account for hostname aliases and other ssh_config(5) tricks, I prefer to read Host directives values from ~/.ssh/config for that.

If you have machine-readable access to the complete zone data for your home or work domain, it may even be worth periodically enumerating all of the hostnames into that file, perhaps using rndc dumpdb -zones for a BIND9 setup, or using an AXFR request. If you have a locally caching recursive nameserver, you could even periodically examine the contents of its cache for new and interesting hosts to add to the file.

Custom commands

As users grow more familiar with the feature set available to them on UNIX-like operating systems, and grow more comfortable using the command line, they will find more often that they develop their own routines for solving problems using their preferred tools, often repeatedly solving the same problem in the same way. You can usually tell if you’ve entered this stage if one or more of the below applies:

  • You repeatedly search the web for the same long commands to copy-paste.
  • You type a particular long command so often it’s gone into muscle memory, and you type it without thinking.
  • You have a text file somewhere with a list of useful commands to solve some frequently recurring problem or task, and you copy-paste from it a lot.
  • You’re keeping large amounts of history so you can search back through commands you ran weeks or months ago with ^R, to find the last time an instance of a problem came up, and getting angry when you realize it’s fallen away off the end of your history file.
  • You’ve found that you prefer to run a tool like ls(1) more often with a non-default flag than without it; -l is a common example.

You can definitely accomplish a lot of work quickly with shoving the output of some monolithic program through a terse one-liner to get the information you want, or by developing muscle memory for your chosen toolbox and oft-repeated commands, but if you want to apply more discipline and automation to managing these sorts of tasks, it may be useful for you to explore more rigorously defining your own commands for use during your shell sessions, or for automation purposes.

This is consistent with the original idea of the Unix shell as a programming environment; the tools provided by the base system are intentionally very general, not prescribing how they’re used, an approach which allows the user to build and customize their own command set as appropriate for their system’s needs, even on a per-user basis.

What this all means is that you need not treat the tools available to you as holy writ. To leverage the Unix philosophy’s real power, you should consider customizing and extending the command set in ways that are useful to you, refining them as you go, and sharing those extensions and tweaks if they may be useful to others. We’ll discuss here a few methods for implementing custom commands, and where and how to apply them.


The first step users take toward customizing the behaviour of their shell tools is often to define shell aliases in their shell’s startup file, usually specifically for interactive sessions; for Bash, this is usually ~/.bashrc.

Some aliases are so common that they’re included as commented-out suggestions in the default ~/.bashrc file for new users. For example, on Debian systems, the following alias is defined by default if the dircolors(1) tool is available for coloring ls(1) output by filetype:

alias ls='ls --color=auto'

With this defined at startup, invoking ls, with or without other arguments, will expand to run ls --color=auto, including any given arguments on the end as well.

In the same block of that file, but commented out, are suggestions for other aliases to enable coloured output for GNU versions of the dir and grep tools:

#alias dir='dir --color=auto'
#alias vdir='vdir --color=auto'

#alias grep='grep --color=auto'
#alias fgrep='fgrep --color=auto'
#alias egrep='egrep --color=auto'

Further down still, there are some suggestions for different methods of invoking ls:

#alias ll='ls -l'
#alias la='ls -A'
#alias l='ls -CF'

Commenting these out would make ll, la, and l work as commands during an interactive session, with the appropriate options added to the call.

You can check the aliases defined in your current shell session by typing alias with no arguments:

$ alias
alias ls='ls --color=auto'

Aliases are convenient ways to add options to commands, and are very common features of ~/.bashrc files shared on the web. They also work in POSIX-conforming shells besides Bash. However, for general use, they aren’t very sophisticated. For one thing, you can’t process arguments with them:

# An attempt to write an alias that searches for a given pattern in a fixed
# file; doesn't work because aliases don't expand parameters
alias grepvim='grep "$1" ~/.vimrc'

They also don’t work for defining new commands within scripts for certain shells:

alias ll='ls -l'

When saved in a file as test, made executable, and run, this script fails:

./test: line 3: ll: command not found

So, once you understand how aliases work so you can read them when others define them in startup files, my suggestion is there’s no point writing any. Aside from some very niche evaluation tricks, they have no functional advantages over shell functions and scripts.


A more flexible method for defining custom commands for an interactive shell (or within a script) is to use a shell function. We could declare our ll function in a Bash startup file as a function instead of an alias like so:

# Shortcut to call ls(1) with the -l flag
ll() {
    command ls -l "$@"

Note the use of the command builtin here to specify that the ll function should invoke the program named ls, and not any function named ls. This is particularly important when writing a function wrapper around a command, to stop an infinite loop where the function calls itself indefinitely:

# Always add -q to invocations of gdb(1)
gdb() {
    command gdb -q "$@"

In both examples, note also the use of the "$@" expansion, to add to the final command line any arguments given to the function. We wrap it in double quotes to stop spaces and other shell metacharacters in the arguments causing problems. This means that the ll command will work correctly if you were to pass it further options and/or one or more directories as arguments:

$ ll -a
$ ll ~/.config

Shell functions declared in this way are specified by POSIX for Bourne-style shells, so they should work in your shell of choice, including Bash, dash, Korn shell, and Zsh. They can also be used within scripts, allowing you to abstract away multiple instances of similar commands to improve the clarity of your script, in much the same way the basics of functions work in general-purpose programming languages.

Functions are a good and portable way to approach adding features to your interactive shell; written carefully, they even allow you to port features you might like from other shells into your shell of choice. I’m fond of taking commands I like from Korn shell or Zsh and implementing them in Bash or POSIX shell functions, such as Zsh’s vared or its two-argument cd features.

If you end up writing a lot of shell functions, you should consider putting them into separate configuration subfiles to keep your shell’s primary startup file from becoming unmanageably large.

Examples from the author

You can take a look at some of the shell functions I have defined here that are useful to me in general shell usage; a lot of these amount to implementing convenience features that I wish my shell had, especially for quick directory navigation, or adding options to commands:

Other examples

Variables in shell functions

You can manipulate variables within shell functions, too:

# Print the filename of a path, stripping off its leading path and
# extension
fn() {
    printf '%s\n' "$name"

This works fine, but the catch is that after the function is done, the value for name will still be defined in the shell, and will overwrite whatever was in there previously:

$ printf '%s\n' "$name"
$ fn /home/you/Task_List.doc
$ printf '%s\n' "$name"

This may be desirable if you actually want the function to change some aspect of your current shell session, such as managing variables or changing the working directory. If you don’t want that, you will probably want to find some means of avoiding name collisions in your variables.

If your function is only for use with a shell that provides the local (Bash) or typeset (Ksh) features, you can declare the variable as local to the function to remove its global scope, to prevent this happening:

# Bash-like
fn() {
    local name
    printf '%s\n' "$name"

# Ksh-like
# Note different syntax for first line
function fn {
    typeset name
    printf '%s\n' "$name"

If you’re using a shell that lacks these features, or you want to aim for POSIX compatibility, things are a little trickier, since local function variables aren’t specified by the standard. One option is to use a subshell, so that the variables are only defined for the duration of the function:

# POSIX; note we're using plain parentheses rather than curly brackets, for
# a subshell
fn() (
    printf '%s\n' "$name"

# POSIX; alternative approach using command substitution:
fn() {
    printf '%s\n' "$(
        printf %s "$name"

This subshell method also allows you to change directory with cd within a function without changing the working directory of the user’s interactive shell, or to change shell options with set or Bash options with shopt only temporarily for the purposes of the function.

Another method to deal with variables is to manipulate the positional parameters directly ($1, $2 … ) with set, since they are local to the function call too:

# POSIX; using positional parameters
fn() {
    set -- "${1##*/}"
    set -- "${1%.*}"
    printf '%s\n' "$1"

These methods work well, and can sometimes even be combined, but they’re awkward to write, and harder to read than the modern shell versions. If you only need your functions to work with your modern shell, I recommend just using local or typeset. The Bash Guide on Greg’s Wiki has a very thorough breakdown of functions in Bash, if you want to read about this and other aspects of functions in more detail.

Keeping functions for later

As you get comfortable with defining and using functions during an interactive session, you might define them in ad-hoc ways on the command line for calling in a loop or some other similar circumstance, just to solve a task in that moment.

As an example, I recently made an ad-hoc function called monit to run a set of commands for its hostname argument that together established different types of monitoring system checks, using an existing script called nmfs:

$ monit() { nmfs "$1" Ping Y ; nmfs "$1" HTTP Y ; nmfs "$1" SNMP Y ; }
$ for host in webhost{1..10} ; do
> monit "$host"
> done

After that task was done, I realized I was likely to use the monit command interactively again, so I decided to keep it. Shell functions only last as long as the current shell, so if you want to make them permanent, you need to store their definitions somewhere in your startup files. If you’re using Bash, and you’re content to just add things to the end of your ~/.bashrc file, you could just do something like this:

$ declare -f monit >> ~/.bashrc

That would append the existing definition of monit in parseable form to your ~/.bashrc file, and the monit function would then be loaded and available to you for future interactive sessions. Later on, I ended up converting monit into a shell script, as its use wasn’t limited to just an interactive shell.

If you want a more robust approach to keeping functions like this for Bash permanently, I wrote a tool called Bashkeep, which allows you to quickly store functions and variables defined in your current shell into separate and appropriately-named files, including viewing and managing the list of names conveniently:

$ keep monit
$ keep
$ ls ~/.bashkeep.d
$ keep -d monit


Shell functions are a great way to portably customize behaviour you want for your interactive shell, but if a task isn’t specific only to an interactive shell context, you should instead consider putting it into its own script whether written in shell or not, to be invoked somewhere from your PATH. This makes the script useable in contexts besides an interactive shell with your personal configuration loaded, for example from within another script, by another user, or by an X11 session called by something like dmenu.

Even if your set of commands is only a few lines long, if you need to call it often–especially with reference to other scripts and in varying contexts– making it into a generally-available shell script has many advantages.


Users making their own scripts often start by putting them in /usr/local/bin and making them executable with sudo chmod +x, since many Unix systems include this directory in the system PATH. If you want a script to be generally available to all users on a system, this is a reasonable approach. However, if the script is just something for your own personal use, or if you don’t have the permissions necessary to write to this system path, it may be preferable to have your own directory for logical binaries, including scripts.

Private bindir

Unix-like users who do this seem to vary in where they choose to put their private logical binaries directory. I’ve seen each of the below used or recommended:

  • ~/bin
  • ~/.bin
  • ~/.local/bin
  • ~/Scripts

I personally favour ~/.local/bin, but you can put your scripts wherever they best fit into your HOME directory layout. You may want to choose something that fits in well with the XDG standard, or whatever existing standard or system your distribution chooses for filesystem layout in $HOME.

In order to make this work, you will want to customize your login shell startup to include the directory in your PATH environment variable. It’s better to put this into ~/.profile or whichever file your shell runs on login, so that it’s only run once. That should be all that’s necessary, as PATH is typically exported as an environment variable for all the shell’s child processes. A line like this at the end of one of those scripts works well to extend the system PATH for our login shell:


Note that we specifically put our new path at the front of the PATH variable’s value, so that it’s the first directory searched for programs. This allows you to implement or install your own versions of programs with the same name as those in the system; this is useful, for example, if you like to experiment with building software in $HOME.

If you’re using a systemd-based GNU/Linux, and particularly if you’re using a display manager like GDM rather than a TTY login and startx for your X11 environment, you may find it more robust to instead set this variable with the appropriate systemd configuration file. Another option you may prefer on systems using PAM is to set it with pam_env(8).

After logging in, we first verify the directory is in place in the PATH variable:

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

We can test this is working correctly by placing a test script into the directory, including the #!/bin/sh shebang, and making it executable by the current user with chmod(1):

$ cat >~/.local/bin/test-private-bindir
printf 'Working!\n'
$ chmod u+x ~./local/bin/test-private-bindir
$ test-private-bindir

Examples from the author

I publish the more generic scripts I keep in ~/.local/bin, which I keep up-to-date on my personal systems in version control using Git, along with my configuration files. Many of the scripts are very short, and are intended mostly as building blocks for other scripts in the same directory. A few examples:

  • gscr(1df): Run a set of commands on a Git repository to minimize its size.
  • fgscr(1df): Find all Git repositories in a directory tree and run gscr(1df) over them.
  • hurl(1df): Extract URLs from links in an HTML document.
  • maybe(1df): Exit with success or failure with a given probability.
  • rfcr(1df): Download and read a given Request for Comments document.
  • tot(1df): Add up a list of numbers.

For such scripts, I try to write them as much as possible to use tools specified by POSIX, so that there’s a decent chance of them working on whatever Unix-like system I need them to.

On systems I use or manage, I might specify commands to do things relevant specifically to that system, such as:

  • Filter out uninteresting lines in an Apache HTTPD logfile with awk.
  • Check whether mail has been delivered to system users in /var/mail.
  • Upgrade the Adobe Flash player in a private Firefox instance.

The tasks you need to solve both generally and specifically will almost certainly be different; this is where you can get creative with your automation and abstraction.

X windows scripts

An additional advantage worth mentioning of using scripts rather than shell functions where possible is that they can be called from environments besides shells, such as in X11 or by other scripts. You can combine this method with X11-based utilities such as dmenu(1), libnotify’s notify-send(1), or ImageMagick’s import(1) to implement custom interactive behaviour for your X windows session, without having to write your own X11-interfacing code.

Other languages

Of course, you’re not limited to just shell scripts with this system; it might suit you to write a script completely in a language like awk(1), or even sed(1). If portability isn’t a concern for the particular script, you should use your favourite scripting language. Notably, don’t fall into the trap of implementing a script in shell for no reason …

awk 'NF>2 && /foobar/ {print $1}' "$@"

… when you can instead write the whole script in the main language used, and save a fork(2) syscall and a layer of quoting:

#!/usr/bin/awk -f
NF>2 && /foobar/ {print $1}

Versioning and sharing

Finally, if you end up writing more than a couple of useful shell functions and scripts, you should consider versioning them with Git or a similar version control system. This also eases implementing your shell setup and scripts on other systems, and sharing them with others via publishing on GitHub. You might even go so far as to write a Makefile to install them, or manual pages for quick reference as documentation … if you’re just a little bit crazy …

Cron best practices

The time-based job scheduler cron(8) has been around since Version 7 Unix, and its crontab(5) syntax is familiar even for people who don’t do much Unix system administration. It’s standardised, reasonably flexible, simple to configure, and works reliably, and so it’s trusted by both system packages and users to manage many important tasks.

However, like many older Unix tools, cron(8)‘s simplicity has a drawback: it relies upon the user to know some detail of how it works, and to correctly implement any other safety checking behaviour around it. Specifically, all it does is try and run the job at an appropriate time, and email the output. For simple and unimportant per-user jobs, that may be just fine, but for more crucial system tasks it’s worthwhile to wrap a little extra infrastructure around it and the tasks it calls.

There are a few ways to make the way you use cron(8) more robust if you’re in a situation where keeping track of the running job is desirable.

Apply the principle of least privilege

The sixth column of a system crontab(5) file is the username of the user as which the task should run:

0 * * * *  root  cron-task

To the extent that is practical, you should run the task as a user with only the privileges it needs to run, and nothing else. This can sometimes make it worthwhile to create a dedicated system user purely for running scheduled tasks relevant to your application.

0 * * * *  myappcron  cron-task

This is not just for security reasons, although those are good ones; it helps protect you against nasties like scripting errors attempting to remove entire system directories.

Similarly, for tasks with database systems such as MySQL, don’t use the administrative root user if you can avoid it; instead, use or even create a dedicated user with a unique random password stored in a locked-down ~/.my.cnf file, with only the needed permissions. For a MySQL backup task, for example, only a few permissions should be required, including SELECT, SHOW VIEW, and LOCK TABLES.

In some cases, of course, you really will need to be root. In particularly sensitive contexts you might even consider using sudo(8) with appropriate NOPASSWD options, to allow the dedicated user to run only the appropriate tasks as root, and nothing else.

Test the tasks

Before placing a task in a crontab(5) file, you should test it on the command line, as the user configured to run the task and with the appropriate environment set. If you’re going to run the task as root, use something like su or sudo -i to get a root shell with the user’s expected environment first:

$ sudo -i -u cronuser
$ cron-task

Once the task works on the command line, place it in the crontab(5) file with the timing settings modified to run the task a few minutes later, and then watch /var/log/syslog with tail -f to check that the task actually runs without errors, and that the task itself completes properly:

May  7 13:30:01 yourhost CRON[20249]: (you) CMD (cron-task)

This may seem pedantic at first, but it becomes routine very quickly, and it saves a lot of hassles down the line as it’s very easy to make an assumption about something in your environment that doesn’t actually hold in the one that cron(8) will use. It’s also a necessary acid test to make sure that your crontab(5) file is well-formed, as some implementations of cron(8) will refuse to load the entire file if one of the lines is malformed.

If necessary, you can set arbitrary environment variables for the tasks at the top of the file:


0 * * * *  you  cron-task

Don’t throw away errors or useful output

You’ve probably seen tutorials on the web where in order to keep the crontab(5) job from sending standard output and/or standard error emails every five minutes, shell redirection operators are included at the end of the job specification to discard both the standard output and standard error. This kluge is particularly common for running web development tasks by automating a request to a URL with curl(1) or wget(1):

*/5 * * *  root  curl >/dev/null 2>&1

Ignoring the output completely is generally not a good idea, because unless you have other tasks or monitoring ensuring the job does its work, you won’t notice problems (or know what they are), when the job emits output or errors that you actually care about.

In the case of curl(1), there are just way too many things that could go wrong, that you might notice far too late:

  • The script could get broken and return 500 errors.
  • The URL of the cron.php task could change, and someone could forget to add a HTTP 301 redirect.
  • Even if a HTTP 301 redirect is added, if you don’t use -L or --location for curl(1), it won’t follow it.
  • The client could get blacklisted, firewalled, or otherwise impeded by automatic or manual processes that falsely flag the request as spam.
  • If using HTTPS, connectivity could break due to cipher or protocol mismatch.

The author has seen all of the above happen, in some cases very frequently.

As a general policy, it’s worth taking the time to read the manual page of the task you’re calling, and to look for ways to correctly control its output so that it emits only the output you actually want. In the case of curl(1), for example, I’ve found the following formula works well:

curl -fLsS -o /dev/null
  • -f: If the HTTP response code is an error, emit an error message rather than the 404 page.
  • -L: If there’s an HTTP 301 redirect given, try to follow it.
  • -sS: Don’t show progress meter (-S stops -s from also blocking error messages).
  • -o /dev/null: Send the standard output (the actual page returned) to /dev/null.

This way, the curl(1) request should stay silent if everything is well, per the old Unix philosophy Rule of Silence.

You may not agree with some of the choices above; you might think it important to e.g. log the complete output of the returned page, or to fail rather than silently accept a 301 redirect, or you might prefer to use wget(1). The point is that you take the time to understand in more depth what the called program will actually emit under what circumstances, and make it match your requirements as closely as possible, rather than blindly discarding all the output and (worse) the errors. Work with Murphy’s law; assume that anything that can go wrong eventually will.

Send the output somewhere useful

Another common mistake is failing to set a useful MAILTO at the top of the crontab(5) file, as the specified destination for any output and errors from the tasks. cron(8) uses the system mail implementation to send its messages, and typically, default configurations for mail agents will simply send the message to an mbox file in /var/mail/$USER, that they may not ever read. This defeats much of the point of mailing output and errors.

This is easily dealt with, though; ensure that you can send a message to an address you actually do check from the server, perhaps using mail(1):

$ printf '%s\n' 'Test message' | mail -s 'Test subject'

Once you’ve verified that your mail agent is correctly configured and that the mail arrives in your inbox, set the address in a MAILTO variable at the top of your file:

0 * * * *    you  cron-task-1
*/5 * * * *  you  cron-task-2

If you don’t want to use email for routine output, another method that works is sending the output to syslog with a tool like logger(1):

0 * * * *   you  cron-task | logger -it cron-task

Alternatively, you can configure aliases on your system to forward system mail destined for you on to an address you check. For Postfix, you’d use an aliases(5) file.

I sometimes use this setup in cases where the task is expected to emit a few lines of output which might be useful for later review, but send stderr output via MAILTO as normal. If you’d rather not use syslog, perhaps because the output is high in volume and/or frequency, you can always set up a log file /var/log/cron-task.log … but don’t forget to add a logrotate(8) rule for it!

Put the tasks in their own shell script file

Ideally, the commands in your crontab(5) definitions should only be a few words, in one or two commands. If the command is running off the screen, it’s likely too long to be in the crontab(5) file, and you should instead put it into its own script. This is a particularly good idea if you want to reliably use features of bash or some other shell besides POSIX/Bourne /bin/sh for your commands, or even a scripting language like Awk or Perl; by default, cron(8) uses the system’s /bin/sh implementation for parsing the commands.

Because crontab(5) files don’t allow multi-line commands, and have other gotchas like the need to escape percent signs % with backslashes, keeping as much configuration out of the actual crontab(5) file as you can is generally a good idea.

If you’re running cron(8) tasks as a non-system user, and can’t add scripts into a system bindir like /usr/local/bin, a tidy method is to start your own, and include a reference to it as part of your PATH. I favour ~/.local/bin, and have seen references to ~/bin as well. Save the script in ~/.local/bin/cron-task, make it executable with chmod +x, and include the directory in the PATH environment definition at the top of the file:


0 * * * *  you  cron-task

Having your own directory with custom scripts for your own purposes has a host of other benefits, but that’s another article…

Avoid /etc/crontab

If your implementation of cron(8) supports it, rather than having an /etc/crontab file a mile long, you can put tasks into separate files in /etc/cron.d:

$ ls /etc/cron.d

This approach allows you to group the configuration files meaningfully, so that you and other administrators can find the appropriate tasks more easily; it also allows you to make some files editable by some users and not others, and reduces the chance of edit conflicts. Using sudoedit(8) helps here too. Another advantage is that it works better with version control; if I start collecting more than a few of these task files or to update them more often than every few months, I start a Git repository to track them:

$ cd /etc/cron.d
$ sudo git init
$ sudo git add --all
$ sudo git commit -m "First commit"

If you’re editing a crontab(5) file for tasks related only to the individual user, use the crontab(1) tool; you can edit your own crontab(5) by typing crontab -e, which will open your $EDITOR to edit a temporary file that will be installed on exit. This will save the files into a dedicated directory, which on my system is /var/spool/cron/crontabs.

On the systems maintained by the author, it’s quite normal for /etc/crontab never to change from its packaged template.

Include a timeout

cron(8) will normally allow a task to run indefinitely, so if this is not desirable, you should consider either using options of the program you’re calling to implement a timeout, or including one in the script. If there’s no option for the command itself, the timeout(1) command wrapper in coreutils is one possible way of implementing this:

0 * * * *  you  timeout 10s cron-task

Greg’s wiki has some further suggestions on ways to implement timeouts.

Include file locking to prevent overruns

cron(8) will start a new process regardless of whether its previous runs have completed, so if you wish to avoid locking for long-running task, on GNU/Linux you could use the flock(1) wrapper for the flock(2) system call to set an exclusive lockfile, in order to prevent the task from running more than one instance in parallel.

0 * * * *  you  flock -nx /var/lock/cron-task cron-task

Greg’s wiki has some more in-depth discussion of the file locking problem for scripts in a general sense, including important information about the caveats of “rolling your own” when flock(1) is not available.

If it’s important that your tasks run in a certain order, consider whether it’s necessary to have them in separate tasks at all; it may be easier to guarantee they’re run sequentially by collecting them in a single shell script.

Do something useful with exit statuses

If your cron(8) task or commands within its script exit non-zero, it can be useful to run commands that handle the failure appropriately, including cleanup of appropriate resources, and sending information to monitoring tools about the current status of the job. If you’re using Nagios Core or one of its derivatives, you could consider using send_nsca to send passive checks reporting the status of jobs to your monitoring server. I’ve written a simple script called nscaw to do this for me:

0 * * * *  you  nscaw CRON_TASK -- cron-task

Consider alternatives to cron(8)

If your machine isn’t always on and your task doesn’t need to run at a specific time, but rather needs to run once daily or weekly, you can install anacron and drop scripts into the cron.hourly, cron.daily, cron.monthly, and cron.weekly directories in /etc, as appropriate. Note that on Debian and Ubuntu GNU/Linux systems, the default /etc/crontab contains hooks that run these, but they run only if anacron(8) is not installed.

If you’re using cron(8) to poll a directory for changes and run a script if there are such changes, on GNU/Linux you could consider using a daemon based on inotifywait(1) instead.

Finally, if you require more advanced control over when and how your task runs than cron(8) can provide, you could perhaps consider writing a daemon to run on the server consistently and fork processes for its task. This would allow running a task more often than once a minute, as an example. Don’t get too bogged down into thinking that cron(8) is your only option for any kind of asynchronous task management!

Shell config subfiles

Large shell startup scripts (.bashrc, .profile) over about fifty lines or so with a lot of options, aliases, custom functions, and similar tweaks can get cumbersome to manage over time, and if you keep your dotfiles under version control it’s not terribly helpful to see large sets of commits just editing the one file when it could be more instructive if broken up into files by section.

Given that shell configuration is just shell code, we can apply the source builtin (or the . builtin for POSIX sh) to load several files at the end of a .bashrc, for example:

source ~/.bashrc.options
source ~/.bashrc.aliases
source ~/.bashrc.functions

This is a better approach, but it still binds us into using those filenames; we still have to edit the ~/.bashrc file if we want to rename them, or remove them, or add new ones.

Fortunately, UNIX-like systems have a common convention for this, the .d directory suffix, in which sections of configuration can be stored to be read by a main configuration file dynamically. In our case, we can create a new directory ~/.bashrc.d:

$ ls ~/.bashrc.d

With a slightly more advanced snippet at the end of ~/.bashrc, we can then load every file with the suffix .bash in this directory:

# Load any supplementary scripts
for config in "$HOME"/.bashrc.d/*.bash ; do
    source "$config"
unset -v config

Note that we unset the config variable after we’re done, otherwise it’ll be in the namespace of our shell where we don’t need it. You may also wish to check for the existence of the ~/.bashrc.d directory, check there’s at least one matching file inside it, or check that the file is readable before attempting to source it, depending on your preference.

The same method can be applied with .profile to load all scripts with the suffix .sh in ~/.profile.d, if we want to write in POSIX sh, with some slightly different syntax:

# Load any supplementary scripts
for config in "$HOME"/.profile.d/*.sh ; do
    . "$config"
unset -v config

Another advantage of this method is that if you have your dotfiles under version control, you can arrange to add extra snippets on a per-machine basis unversioned, without having to update your .bashrc file.

Here’s my implementation of the above method, for both .bashrc and .profile:

Thanks to commenter oylenshpeegul for correcting the syntax of the loops.

Prompt directory shortening

The common default of some variant of \h:\w\$ for a Bash prompt PS1 string includes the \w escape character, so that the user’s current working directory appears in the prompt, but with $HOME shortened to a tilde:


This is normally very helpful, particularly if you leave your shell for a time and forget where you are, though of course you can always call the pwd shell builtin. However it can get annoying for very deep directory hierarchies, particularly if you’re using a smaller terminal window:


If you’re using Bash version 4.0 or above (bash --version), you can save a bit of terminal space by setting the PROMPT_DIRTRIM variable for the shell. This limits the length of the tail end of the \w and \W expansions to that number of path elements:

tom@sanctum:/chroot/apache/usr/local/app-library/lib/App/Library/Class$ PROMPT_DIRTRIM=3

This is a good thing to include in your ~/.bashrc file if you often find yourself deep in directory trees where the upper end of the hierarchy isn’t of immediate interest to you. You can remove the effect again by unsetting the variable:

tom@sanctum:.../App/Library/Class$ unset PROMPT_DIRTRIM

Testing exit values in Bash

In Bash scripting (and shell scripting in general), we often want to check the exit value of a command to decide an action to take after it completes, likely for the purpose of error handling. For example, to determine whether a particular regular expression regex was present somewhere in a file options, we might apply grep(1) with its POSIX -q option to suppress output and just use the exit value:

grep -q regex options

An approach sometimes taken is then to test the exit value with the $? parameter, using if to check if it’s non-zero, which is not very elegant and a bit hard to read:

# Bad practice
grep -q regex options
if (($? > 0)); then
    printf '%s\n' 'myscript: Pattern not found!' >&2
    exit 1

Because the if construct by design tests the exit value of commands, it’s better to test the command directly, making the expansion of $? unnecessary:

# Better
if grep -q regex options; then
    # Do nothing
    printf '%s\n' 'myscript: Pattern not found!\n' >&2
    exit 1

We can precede the command to be tested with ! to negate the test as well, to prevent us having to use else as well:

# Best
if ! grep -q regex options; then
    printf '%s\n' 'myscript: Pattern not found!' >&2
    exit 1

An alternative syntax is to use && and || to perform if and else tests with grouped commands between braces, but these tend to be harder to read:

# Alternative
grep -q regex options || {
    printf '%s\n' 'myscript: Pattern not found!' >&2
    exit 1

With this syntax, the two commands in the block are only executed if the grep(1) call exits with a non-zero status. We can apply && instead to execute commands if it does exit with zero.

That syntax can be convenient for quickly short-circuiting failures in scripts, for example due to nonexistent commands, particularly if the command being tested already outputs its own error message. This therefore cuts the script off if the given command fails, likely due to ffmpeg(1) being unavailable on the system:

hash ffmpeg || exit 1

Note that the braces for a grouped command are not needed here, as there’s only one command to be run in case of failure, the exit call.

Calls to cd are another good use case here, as running a script in the wrong directory if a call to cd fails could have really nasty effects:

cd wherever || exit 1

In general, you’ll probably only want to test $? when you have specific non-zero error conditions to catch. For example, if we were using the --max-delete option for rsync(1), we could check a call’s return value to see whether rsync(1) hit the threshold for deleted file count and write a message to a logfile appropriately:

rsync --archive --delete --max-delete=5 source destination
if (($? == 25)); then
    printf '%s\n' 'Deletion limit was reached' >"$logfile"

It may be tempting to use the errexit feature in the hopes of stopping a script as soon as it encounters any error, but there are some problems with its usage that make it a bit error-prone. It’s generally more straightforward to simply write your own error handling using the methods above.

For a really thorough breakdown of dealing with conditionals in Bash, take a look at the relevant chapter of the Bash Guide.

GNU/Linux Crypto: Importance

While this series was being written, from June 2013, Edward Snowden began leaking top-secret documents from the United States National Security Agency, showing that the agency was capable of Internet surveillance on a massive scale with the PRISM surveillance system and with the XKeyscore interface into their amassed data. The fact that covert government surveillance was possible and was taking place does not come as particularly surprising news to network engineers and conspiracy theorists, but the revelations have finally given the general, non-technical public an idea of how badly the proprietary systems around which they have built much of their digital lives can be used to harm them and compromise their privacy.

Concerned people in the United States will be only too aware of how the secret abuse of power to exercise this surveillance and the failed motions to curtail it by the United States Congress has dented their trust in their own government. However, the leaks’ implications are international as well. The foreign intelligence agency in my own country of New Zealand, the Government Communications Security Bureau, was earlier this year accused of illegally spying on New Zealand citizens, and diplomatic cables from WikiLeaks show the GCSB is potentially already cooperating with the NSA. In spite of this, new legislation is set to extend the GCSB’s powers, despite independent reviews condemning the bill from both a legal and human rights perspective, even after amendments. The scandal and the anger over surveillance abuse extends to the United Kingdom, Germany, Sweden, and many other countries.

I do hold out some hope for the efforts such as the Electronic Frontier Foundation’s class action suit to curtail the surveillance or at the very least to register the public’s anger about this unwarranted intrusion into private lives. However I am concerned not just by the possibility of the rise of a global surveillance state, but by the implications this has for the right to secure communications using cryptography for authentication and encryption.

It’s no secret that cryptography and encryption presents a problem to the NSA’s surveillance systems, and that they expend a great deal of effort in attempting to circumvent it, including demanding private keys from businesses for applications like HTTPS. My concern is this: If it becomes publically accepted that governments spy warrantlessly on international networks and that this is justified or necessary, then we may reach a point where the legality of the general public’s use of cryptography itself may again be called into question.

Computing professionals of my generation likely did not begin their careers until after the United States’ cryptographic export controls were relaxed in 1999, perhaps prompting us to take for granted the availability of algorithms like RSA and AES with high key sizes for cryptographic purposes. A world where a government agency would actively attempt to curtail the use of such technology may seem very far-fetched to us — perhaps less so to those who remember that Pretty Good Privacy was a radical new idea that caused its activist creator Phil Zimmermann real legal trouble.

I believe that computing enthusiasts and users of free software operating systems, not just cryptographic experts, are in a special position to assist their concerned friends and family with defending their online privacy and securing their communications, and that if we value both freedom and security of information, then we in fact have a responsibility to do so. I believe that people need to be aware of not just the implications of massive surveillance on a global scale, but also how to exercise their rights to fight against it. If the legality of cryptography is ever called into question again as the result of its impeding warrantless surveillance, then its pervasiveness and the public’s insistence on its free availability should make restricting its use not just impractical, but unthinkable.

GNU/Linux Crypto: Disks

GnuPG provides us with a means to securely encrypt individual files on a filesystem, but for really high-security information or environments, it may be appropriate to encrypt an entire disk, to mitigate problems such as caching sensitive files in plaintext. The GNU/Linux kernel includes its own disk encryption solution in the kernel, dm-crypt. This can be leveraged with a low-end tool called cryptsetup, or more easily with LUKS, the Linux Unified Key Setup, implementing strong cryptography with passphrases or keyfiles.

In this example, we’ll demonstrate how this can work to encrypt a USB drive, which is a good method for securely storing really sensitive data such as PGP master keys that’s only needed occasionally, rather than leaving it always mounted on a networked device. Be aware that this erases any existing files on the drive.


The cryptographic tools used by dm-crypt and LUKS are built-in to Linux kernels after 2.6, but you may have to install a package to get access to the cryptsetup frontend. On Debian-derived systems, it’s available in cryptsetup:

# apt-get install cryptsetup

On RPM-based systems like Fedora or CentOS, the package has the same name, cryptsetup:

# yum install cryptsetup

Creating the volume

After identifying the block device on which we want the encrypted filesystem, for example /dev/sdc1, we can erase any existing contents using a call to wipefs:

# wipefs -a /dev/sdc1

Alternatively, we can zero out the whole disk, if we want to completely overwrite any trace of the data previously on the disk; this can take a long time for large volumes:

# cat /dev/zero >/dev/sdc1

If you don’t have a USB drive to hand, but would still like to try this out, you can use a loop block device in a file. For example, to create a 100MB loop device:

# dd if=/dev/zero of=/loopdev bs=1k count=102400
102400+0 records in
102400+0 records out
104857600 bytes (105 MB) copied, 0.331452 s, 316 MB/s
# losetup -f
# losetup /dev/loop0 /loopdev

You can then follow the rest of this guide using /dev/loop0 as the raw block device in place of /dev/sdc1. In the above output, losetup -f returns the first available loop device for use.

Setting up a LUKS container on the block device is then done as follows, providing a passphrase of decent strength; as always, the longer the better. Ideally, you should not use the same passphrase as your GnuPG or SSH keys.

# cryptsetup luksFormat /dev/sdc1

This will overwrite data on /dev/sdc1 irrevocably.

Are you sure? (Type uppercase yes): YES
Enter passphrase:
Verify passphrase:

This creates an abstracted encryption container on the disk, which can be opened by providing the appropriate passphrase. A virtual mapped device is then provided that encrypts all data written to it transparently, with the encrypted data written to the disk.

Using the mapped device

We can open the mapped device using cryptsetup luksOpen, which will prompt us for the passphrase:

# cryptsetup luksOpen /dev/sdc1 secret

The last argument here is the filename for the block device to appear under /dev/mapper; this example provides /dev/mapper/secret.

With this done, the block device on /dev/mapper/secret can now be used in the same way as any other block device; all of the disk operations are abstracted through encryption operations. You’ll probably want to create a filesystem on it; in this case, I’m creating an ext4 filesystem:

# mkfs.ext4 /dev/mapper/secret
mke2fs 1.42.8 (20-Jun-2013)
Filesystem label=
OS type: Linux
Block size=1024 (log=0)
Fragment size=1024 (log=0)
Stride=0 blocks, Stripe width=0 blocks
25168 inodes, 100352 blocks
5017 blocks (5.00%) reserved for the super user
First data block=1
Maximum filesystem blocks=67371008
13 block groups
8192 blocks per group, 8192 fragments per group
1936 inodes per group
Superblock backups stored on blocks:
        8193, 24577, 40961, 57345, 73729

Allocating group tables: done
Writing inode tables: done
Creating journal (4096 blocks): done
Writing superblocks and filesystem accounting information: done

We can then mount the device as normal, and data put into the newly created filesytem will be transparently encrypted:

# mkdir -p /mnt/secret
# mount /dev/mapper/secret /mnt/secret

For example, we could store a private GnuPG key on it:

# cp -prv /home/tom/.gnupg/secring.gpg /mnt/secret

Information about the device

We can get some information about the LUKS container and the specifics of its encryption using luksDump on the underlying block device. This shows us the encryption method used, in this case aes-xts-plain64.

# cryptsetup luksDump /dev/sdc1
LUKS header information for /dev/sdc1

Version:        1
Cipher name:    aes
Cipher mode:    xts-plain64
Hash spec:      sha1
Payload offset: 4096
MK bits:        256
MK digest:      87 6d 08 59 b2 f0 c6 6e ca ec 5f 72 2c e0 35 33 c2 9e cb 8e
MK salt:        7f a5 38 4c 14 85 61 cb 6c 22 65 48 87 21 60 8f
                fa 40 2a ab ae 7d cc df c9 9b a4 e3 3c 64 b6 bb
MK iterations:  49375
UUID:           f4e5f28c-3b34-4003-9bcd-dbb2352042ba

Key Slot 0: ENABLED
        Iterations:             197530
        Salt:                   2d 57 f6 2b 44 a6 61 ee d6 ee e4 7d 64 f0 71 d6
                                55 16 09 83 b4 f0 94 ca 19 17 11 a9 34 84 02 96
        Key material offset:    8
        AF stripes:             4000
Key Slot 1: DISABLED
Key Slot 2: DISABLED
Key Slot 3: DISABLED
Key Slot 4: DISABLED
Key Slot 5: DISABLED
Key Slot 6: DISABLED
Key Slot 7: DISABLED

Unmounting the device

When finished with the data on the device, we should both unmount any filesystem on it, and also close the mapped device so that the passphrase is required to re-open it:

# umount /mnt/secret
# cryptsetup luksClose /dev/mapper/secret

If the data is a removable device, you should also consider physically removing the media from the machine and placing it in some secure location.

This post only scratches the surface of LUKS functionality; many more things are possible with the system, including automatic mounting of encrypted filesystems and the use of stored keyfiles instead of typed passphrases. The FAQ for cryptsetup contains a great deal of information, including some treatment of data recovery, and the Arch Wiki has an exhaustive page on various ways of using LUKS securely.

GNU/Linux Crypto: Backups

While having local backups for quick restores is important, such as on a USB disk or spare hard drive, it’s equally important to have a backup offsite from which you can restore your important documents if, for example, your office was burgled or burned down, losing both your workstation and backup media.

The easiest way to do this for most people is with a storage provider, offering convenient access to bulk storage of suitable size maintained on another company’s systems for a relatively modest price or even for free, such as the Ubuntu One service, or Microsoft’s offering, Skydrive. The best storage providers will also encrypt the data on their own servers, whether or not they have access.

Trusting a company with all your data and the encryption thereof is risky, particularly given recent revelations of corporate collusion with the NSA, and privacy-conscious users should prefer the security of encrypting the backups before they go up onto the provider’s servers. The provider may implement closed and/or symmetric encryption mechanisms of their own, which may or may not be trustworthy. For very strong personal encryption, as established, we can use our GnuPG setup to encrypt files before we put them up there:

$ tar -cf docsbackup-"$(date +%Y-%m-%d)".tar $HOME/Documents
$ gpg --encrypt docsbackup-2013-07-27.tar
$ scp docsbackup-2013-07-27.tar.gpg

The problem with encrypting whole files before we put them up for storage is that for even modestly sized data, performing entire backups and uploading all of the files together every time can cost a lot of bandwidth. Similarly, we’d like to be able to restore our personal files as they were on a specific date, in case of bad backups or accidental deletion, but without storing every file on every backup day, which may end up requiring far too much space.

Incremental backups

Normally, the solution is to use an incremental backup system, meaning after first uploading your files in their entirety to the backup system, successive backups upload only the changes, storing them in a retrievable and space-efficient format. Systems like Dirvish, a free Perl frontend to rsync(1), allow this.

Unfortunately, Dirvish doesn’t encrypt the files or changesets it stores. What’s needed is an incremental backup solution that efficiently calculates and stores changes in files on a remote server, and also encrypts them. Duplicity, a Python tool built around librsync, excels at this, and can use our GnuPG asymmetric key setup for the file encryption. It’s available in Debian-derived systems in the duplicity package. Note that, as before, a GnuPG key setup with an agent is required for this to work.


We can get an idea of how duplicity(1) works by asking it to start a backup vault on our local machine. It uses much the same source destination argument as tools like rsync or scp:

$ cd
$ duplicity --encrypt-key Documents file://docsbackup

It’s important to specify --encrypt-key, because otherwise duplicity(1) will use symmetric encryption with a passphrase rather than a public key, which is considerably less secure. Specify the email address corresponding to the public keypair you would like to use for the encryption.

This performs a full encrypted backup of the directory, returning the following output:

Local and Remote metadata are synchronized, no sync needed.
Last full backup date: none
No signatures found, switching to full backup.
--------------[ Backup Statistics ]--------------
StartTime 1374903081.74 (Sat Jul 27 17:31:21 2013)
EndTime 1374903081.75 (Sat Jul 27 17:31:21 2013)
ElapsedTime 0.01 (0.01 seconds)
SourceFiles 4
SourceFileSize 142251 (139 KB)
NewFiles 4
NewFileSize 142251 (139 KB)
DeletedFiles 0
ChangedFiles 0
ChangedFileSize 0 (0 bytes)
ChangedDeltaSize 0 (0 bytes)
DeltaEntries 4
RawDeltaSize 138155 (135 KB)
TotalDestinationSizeChange 138461 (135 KB)
Errors 0

You’ll note you were not prompted for your passphrase to do this. Remember, encrypting files with your public key does not require a passphrase; the whole idea is that anyone can encrypt using your key without needing your permission.

Checking the created directory docsbackup, we find three new files within it, all three of them encrypted:

$ ls -1 docsbackup

The vol1.difftar.gpg file contains the actual data stored; the other two files contain metadata about the backup’s contents, for use to calculate differences the next time the backup runs.

If we make a small change to a file in the directory being backed up, and run the same command again, we note that the backup has been performed incrementally, and only the changes (the new file) have been saved:

$ duplicity --encrypt-key Documents file://docsbackup
Local and Remote metadata are synchronized, no sync needed.
Last full backup date: Sat Jul 27 17:34:33 2013
--------------[ Backup Statistics ]--------------
StartTime 1374903396.52 (Sat Jul 27 17:36:36 2013)
EndTime 1374903396.52 (Sat Jul 27 17:36:36 2013)
ElapsedTime 0.01 (0.01 seconds)
SourceFiles 5
SourceFileSize 142255 (139 KB)
NewFiles 2
NewFileSize 4100 (4.00 KB)
DeletedFiles 0
ChangedFiles 0
ChangedFileSize 0 (0 bytes)
ChangedDeltaSize 0 (0 bytes)
DeltaEntries 2
RawDeltaSize 4 (4 bytes)
TotalDestinationSizeChange 753 (753 bytes)
Errors 0

We also find three new files in the docsbackup directory containing the new data:

$ ls -1 docsbackup

Note that the new files have the prefix duplicity-inc- or duplicity-new-, denoting them as incremental backups and not full ones.

Note that in order to keep track of what files have already been backed up, duplicity(1) stores metadata in ~/.cache/duplicity, as well as storing them along with the backup. This allows us to let our backup processes run unattended, rather than having to put in our passphrase to read the metadata on the remote server before performing an incremental backup. Of course, if we lose our cached files, that’s OK; we can read the ones out of the backup vault by supplying our passphrase on request for decryption.

Remote backups

If you have SSH or even just SCP/SFTP access to your storage provider’s servers, not much has to change to make duplicity(1) store the files up there instead:

$ duplicity --encrypt-key Documents s

Your backups will then be sent over an SSH link to the directory docsbackup on the system, with username user. In this way, not only is all the data protected in transmission, it’s stored encrypted on the remote server; it never sees your plaintext data. All anyone with access to your backups can see is their approximate size, the dates they were made, and (if you publish your public key) the user ID on the GnuPG key used to encrypt them.

If you’re using the ssh-agent(1) program to store your decrypted private keys, you won’t even have to enter a passphrase for that.

The duplicity(1) frontend supports other methods of uploading to different servers, too, including the boto backend for S3 Amazon Web Services, the gdocs backend for Google Docs, and httplib2 or oauthlib for Ubuntu One.

If you like, you can also sign your backups to make sure they haven’t been tampered with at the time of restoration, by changing --encrypt-key to --encrypt-sign-key. Note that this will require your passphrase.


Restoring from a duplicity(1) backup volume is much the same, but with the arguments reversed:

$ duplicity s docsrestore
Synchronizing remote metadata to local cache...
GnuPG passphrase:
Copying duplicity-full-signatures.20130727T053433Z.sigtar.gpg to local cache.
Copying duplicity-full.20130727T053433Z.manifest.gpg to local cache.
Copying to local cache.
Copying to local cache.
Last full backup date: Sat Jul 27 17:34:33 2013

Note that this time you are asked for your passphrase. This is because restoring the backup requires decrypting the data and possibly the signatures in the backup vault. After doing this, the complete set of documents from the time of your most recent incremental backup will be available in docsrestore.

Using this incremental system also allows you to restore your data in the state in the last backup before a given time. For example, to retrieve my ~/Documents directory as it was three days ago, I might run this:

$ duplicity --time 3D \
    s \

You can extend this to only restore particular files for large vaults, if you only need a particular file from the vault:

$ duplicity --time 3D \
    --file-to-restore private/eff.txt \
    s \


You should run your first full backup interactively to make sure it’s doing exactly what you need, but once you’re confident that everything is working correctly, you can set up a simple Bash script to run incremental backups for you. Here’s an example script, saved in $HOME/.local/bin/backup-remote:

#!/usr/bin/env bash

# Run keychain to recognise any agents holding decrypted keys we might need
# (optional, depending on your SSH key setup)
eval "$(keychain --eval --quiet)"

# Specify directory to back up, GnuPG key ID, and remote username and
# hostname

# Run backup with duplicity
/usr/bin/duplicity --encrypt-key "$keyid" -- "$local" "$remote"

The line with keychain is optional, but will be necessary if you’re using an SSH key with a passphrase on it; you’ll also need to have authenticated with ssh-agent at least once. See the earlier article on SSH/GPG agents for details on this setup.

Don’t forget to make the script executable:

$ chmod +x ~/.local/bin/backup-remote

You can then have cron(8) call this for you every week, running it as your user, by editing your user crontab(5) file:

$ crontab -e

The following line would run this script every morning, beginning at 6.00am:

0 6 * * *   ~/.local/bin/backup-remote


A few general best practices apply to this, consistent with the Tao of Backup:

  • Check that your backups completed; either have the output of the cron script mailed to you, or log it to a file that you check at least occasionally to make sure your backups are working. I highly recommend using an email message, and including error output:
    0 6 * * *   ~/.local/bin/backup-remote 2>&1
  • Run backups to your local servers too; this might prevent your backup provider from reading your files, but it won’t save them from being accidentally deleted.

  • Don’t forget to occasionally test-restore your backups to make sure they’re working correctly. It’s also wise to use duplicity verify on them occasionally, particularly if you don’t back up every day:

    $ duplicity verify s Documents
    Local and Remote metadata are synchronized, no sync needed.
    Last full backup date: Sat Jul 27 17:34:33 2013
    GnuPG passphrase:
    Verify complete: 2195 files compared, 0 differences found.
  • This incremental system means that you’ll likely only have to make full backups once, so you should back up too much data rather than too little; if you can spare the bandwidth and have the space, backing up your entire computer isn’t really that extreme.

  • Try not to depend too much on your remote backups; see them as a last resort, and work securely and with local backups as much as you can. Certainly, never rely on backups as a version control system; use Git for that.