Most of us use a shell, at least once in a while. Some spend most of their time in one. Often we take it for granted, not knowing or caring how it does what it does. In this workshop, we'll delve a little into that by building our very own shell!
Shells have many different functions. This workshop will be focused on how processes are launched, and how their IO can be controlled with pipes. We'll be leaving things like environment variable expansion, signal handling and other things out, but I will give some examples of other directions to go at the end.
Rule 0. In particular, we have zero interest in dealing with a whole bunch of messy details, including but not limited to:
- memory management
- parsing and tokenising and other text processing
- ‘proper’ error handling
Download the workshop skeleton code and utility functions. TODO link
I highly recommend using a version control system and making frequent commits.
The most important thing to start with is picking a name for your shell,
and its prompt. I went with heee>
while I was putting this workshop together,
so a typical interaction looks like this:
heeee> ls
shell shell.c
So, let's get started!
Exercise 1: Set up the main function of your shell to repeatedly prompt the user for a line. For now, we'll just print that line back out. Our shell is a slightly glorified cat
, just more personalised!
Here's a little snippet to read a line in C into a variable line
:
char *line = NULL;
int capacity = 0;
getline(&line, &capacity, stdin);
So now we can read in a user's command. Next up is to actually run it. To do this, we need some help from the kernel, which is in charge of most things related to processes—among many other things.
The way our programs talk to the kernel is through system calls. These are operations we can tell the kernel to do on our behalf. Examples include:
- opening files (
open
) - reading and writing to them (
read
,write
) - sending and receiving data over the network. (
sendto
,recvfrom
) - starting programs (
execve
)
In fact, more or less everything that involves our programs interacting with the ‘real world’ involves system calls.
Right now, we're interested in executing a program. That uses the execve
system call, and we'll use a wrapper function called execvp
. Not confusing at all, right?
Before we run a program, we need to understand how program arguments get passed.
When a program is run, it gets an array of arguments passed in by the user, traditionally called argv
. By convention, the first of these is the name of the program. So, when we run something like ls -l /tmp
, the argv
that ls
receives consists of { "ls", "-l", "/tmp" }
.
So the call looks like execvp('ls', ['ls', '-l', '/tmp'])
, except that C doesn't let you write the array directly like that.
Now we're ready to run a program. Parsing ls -l /tmp
into an array of arguments is annoying in C (and not very instructive), so I've written a parse_command
function that you can use.
Exercise 2. Modify your shell so that after it has read a line from stdin
, it executes the command with its arguments.
Try that out. Anything unexpected happen?
In the last exercise, once the program you ran terminated, you didn't go back to your prompt, but got dumped right back to your regular shell. The exec
call transforms the current process into the command we specify. So our shell's process is replaced by ls
or whatever we chose to run. Once that program is done, the process exits—no more shell.
To avoid this happening, we want the command to run as a new process. We create processes with the fork
system call. This creates a new process that is a clone of the existing one, right down to the point in the program that's executing. We can then call exec
in the newly created child process, leaving our shell's process around for future interaction.
The return value of fork
tells us if we're in the child or the parent: it returns 0
for the child, and the id of the new process in the parent.
Exercies 3. Modify your shell to call fork
, and then exec
in the child.
Congratulations! Your shell is now mildly useful. There's still a niggle though: when you run a command, your prompt is probably printing out before the command's output.
Looking at your code, you'll see that the parent goes right back to the prompt. To wait for the child process, there's a well-named wait
system call. For our purposes, we can just call it with a NULL
argument: wait(NULL);
.
Exercise 4. Make your shell wait
on the child process before going back to the prompt.
Now it's starting to act like a real shell! Try out a few of your favourite commands. I'm a fan of cowsay
, but don't let me stop you from running ls
or cat
!
What happens if we try to run cd
? There's a program at /bin/cd
that should let us change directories, so try out this sequence:
heeee> pwd
/Users/kalmar/projects/conferences/2014/strangeloop/shell-workshop
heeee> cd ..
heeee> pwd
[something... unexpected?]
heeee> cd /tmp
heeee> pwd
[something... unexpected?]
heeee> echo HEEEELLLPPPP!!!!!
HEEEELLLPPPP!!!!!
That last little sequence of commands probably surprised you a bit. To understand just why we couldn't change directory, we need to find out a bit about how processes work.
Every process has
- its own memory (where all the global and local variables, the stack and heap)
- extra metadata that the kernel manages, but that the process can't change directly. It can change these with system calls.
- user id (
setuid
) - group id (
setgid
) - priority (
setpriority
) - working directory (
chdir
)
- user id (
So let's imagine that we fork
and exec
the /bin/cd
process.
1) parent: pid 1000, working directory /tmp
---- fork ----
2) parent: pid 1000, working directory /tmp
child: pid 2000, working directory /tmp
---- exec `/bin/cd /home/awesome` in child ----
3) parent: pid 1000, working directory /tmp
child: pid 2000, working directory /tmp/awesome
---- child exits ----
4) parent: pid 1000, working directory /tmp
A process can only use system calls to modify its own state, so the parent shell process still has the same working directory it started with: /tmp
.
To change the shell's working directory, the shell itself needs to call the chdir
system call. So the shell has to treat a cd
command specially: if the command is cd
, then call chdir
instead of the usual fork
/ exec
. The handling of cd
is then built in to our shell—it's a builtin!
Exercise 6. Add a cd
builtin to your shell.
cd
isn't the only builtin in your regular shell. Take a look at man builtin
to see what else is there. Note that not all of them need to be builtins. For example echo
works just fine as an external command, but many shells have it as a builtin. Some of these are builtin for efficiency, since creating a new process has some performance overhead.
Next up we're going to implement pipelines. Try running ls | wc
in your shell. What happens?
heeee> ls | wc
ls: cannot access |: No such file or directory
ls: cannot access wc: No such file or directory
Right now it doesn't know that |
has any special meaning, and so tries to list the files |
and wc
. Before we fix that, we need to understand better how pipes work.
Files on Unix-like systems are managed using file descriptors. If you're using Linux, you can see all the file descriptors for a process using ls -l /proc/<pid>/fd
. File descriptors are positive integers (0
is stdin, 1
is stdout, 2
is stderr, and 3 and up are any other files the process has open).
The pipe
system call creates two new file descriptors with a special property.
int pipe_fds[2];
pipe(pipe_fds); // puts returns [4, 5] in pipe_fds
If pipe
returns [4, 5]
, then any data written to 5
can be read out of 4
. What this means is that we need to set ls
's stdout to be 5
and wc
's stdin to be 4
. The way to change which file descriptor a file points to is with the dup2
system call. In this example, we'd need to run dup2(5, 1)
for ls
and dup2(4, 0)
for wc
.
To recap, to create a pipe ls | wc
, we need to:
- create a pipe with
pipe
. Imagine that this creates file descriptors4
and5
. - fork twice (once for
ls
and once forwc
).
parent: pid 1000, stdin: terminal, stdout, terminal
---- fork twice ----
parent: pid 1000, stdin: terminal, stdout: terminal
child 1: pid 2000, stdin: terminal, stdout: terminal (for `ls`)
child 2: pid 3000, stdin: terminak, stdout: terminal (for `wc`)
- In child 1, run
dup2(pipe_fds[1], 1)
and execls
- In child 2, run
dup2(pipe_fds[0], 0)
and execwc
- Close both ends of the pipe in the parent process.
That's it!
TODO ps | grep shows up grep
Exercise 7: Write a fork_and_exec_with_io(cmd* cmd, int stdout_fd, int stdin_fd)
that forks and execs the specified command and changes its stdin and stdout as needed
Exercise 8: Use fork_and_exec_with_io
to implement pipes with 2 elements.
There are a few directions to go in once you've completed the exercises here. I've split these into things that are fairly clear progression from what we've done, and things that are exploring other areas of shell features.
We got a single pipe set up, but longer pipelines are much more useful. The parse_line
function can return any number of piped commands, it's up to you to make it go!
If you run grep shel /usr/share/dict/words > /tmp/shel_words
in your regular shell, the list of words containing shel
will end up in /tmp/shel_words
. To do this, you'll have to extend the parsing code to pick up these redirections. Once done, the redirection itself won't be too different from how pipe was handled. You'll just have to use open
instead of pipe
to get the file descriptor to read from or write to.
If you run some-command 2>/tmp/errors
in your regular shell, you are instructing it to redirect stderr
(fd
2) to /tmp/errors
. This is useful to capture log output when a command writes it to stderr
. This is a generalisation of the previous extension.
If you run some-command 2>&1
in your regular shell, data written by some-command
to its stderr
(fd
2) will go to the same place as its stdout
(fd
1). This is a further extension of redirection.
Compare the output of echo $HOME
in your shell to your regular shell. Your regular shell treats tokens beginning with $
specially: it looks them up by name in the environment, and substitutes them in the command. The getenv
function looks up a variable in the environment. This can be handled during or after parsing a command.
In your regular shell, compare the output of date
and TZ=Pacific/Samoa date
. The date
command uses the TZ
environment variable to decide how to display the time. Prefixing a command with a string like VAR=value
sets VAR
to value
in the child's environment. The setenv
function sets the value of an environment variable. This will require modifying the parser to find these prefixes.
In bash
, export
lets us change an environment variable's value in the shell itself. As with cd
, because a child cannot modify its parent's environment, export
must be a builtin.
Compare what happens when you hit ^C
(control-C) between your shell and your regular shell. ^C
sends a signal to your process, in this case SIGINT
. Signals interrupt the normal execution of your program. By default, this aborts the process. To prevent that, we can register a signal handler function to be called when SIGINT
is received, or decide to ignore it.
Compare the output of wc -c *
between your shell and your regular shell. Expanding the wildcard *
is done by the shell before passing arguments to the child. The glob
function in the <glob.h>
standard header will be useful.
In your regular shell, if you end a command with &
, it runs in the background. The command will run, but the shell will not block while it does. Ie, it does not wait
on the child process.
TODO