Anatomy of a shell December 28, 2018 on Drew DeVault's blog

I’ve been contributing where I can to Simon Ser’s mrsh project, a work-in-progress strictly POSIX shell implementation. I worked on some small mrsh features during my holiday travels and it’s in the forefront of my mind, so I’d like to share some of its design details with you.

There are two main components to a shell: parsing and execution. mrsh uses a simple recursive descent parser to generate an AST (Abstract Syntax Tree, or an in-memory model of the structure of the parsed source). This design was chosen to simplify the code and avoid dependencies like flex/bison, and is a good choice given that performance isn’t critical for parsing shell scripts. Here’s an example of the input source and output AST:

say_hello() {
	echo "hello $1!"

say_hello "$who"

This script is parsed into this AST (this is the output of mrsh -n

└─command_list ─ pipeline
  └─function_definition say_hello ─ brace_group
    └─command_list ─ pipeline
        ├─name ─ word_string [3:2 → 3:6] echo
        └─argument 1 ─ word_list (quoted)
          ├─word_string [3:8 → 3:14] hello
          │ └─name 1
          └─word_string [3:16 → 3:17] !
└─command_list ─ pipeline
      ├─name who
      └─value ─ word_command ─ program
        └─command_list ─ pipeline
            └─name ─ word_string [6:7 → 6:13] whoami
└─command_list ─ pipeline
    ├─name ─ word_string [7:1 → 7:10] say_hello
    └─argument 1 ─ word_list (quoted)
        └─name who

Most of these names come directly from the POSIX shell specification. The parser and AST is made available as a standalone public interface of libmrsh, which can be used for a variety of use-cases like syntax-aware text editors, syntax highlighting (see highlight.c), linters, etc. The most important use-case is, of course, task planning and execution.

Most of these AST nodes becomes a task. A task defines an implementation of the following interface:

struct task_interface {
	 * Request a status update from the task. This starts or continues it.
	 * `poll` must return without blocking with the current task's status:
	 * - TASK_STATUS_WAIT in case the task is pending
	 * - TASK_STATUS_ERROR in case a fatal error occured
	 * - A positive (or null) code in case the task finished
	 * `poll` will be called over and over until the task goes out of the
	 * TASK_STATUS_WAIT state. Once the task is no longer in progress, the
	 * returned state is cached and `poll` won't be called anymore.
	int (*poll)(struct task *task, struct context *ctx);
	void (*destroy)(struct task *task);

Most of the time the task will just do its thing. Many tasks will have sub-tasks as well, such as a command list executing a list of commands, or each branch of an if statement, which it can defer to with task_poll. Many tasks will wait on an external process, in which case it can return TASK_STATUS_WAIT to have the process waited on. Feel free to browse the full list of tasks to get an idea.

One concern more specific to POSIX shells is built-in commands. Some commands have to be built-in because they manipulate the shell’s state, such as . and cd. Others, like true & false, are there for performance reasons, since they’re simple and easily implemented internally. POSIX specifies a list of special builtins which are necessary to implement in the shell itself. There’s a second list that must be present for the shell environment to be considered POSIX compatible (plus some reserved names like local and pushd that invoke undefined behavior - mrsh aborts on these).

Here are some links to more interesting parts of the code so you can explore on your own:

I might write more articles in the future diving into specific concepts, feel free to shoot me an email if you have suggestions. Shoutout to Simon for building such a cool project! I’m looking forward to contributing more until we have a really nice strictly POSIX shell.

Have a comment on one of my posts? Start a discussion in my public inbox by sending an email to ~sircmpwn/ [mailing list etiquette]

Articles from blogs I read Generated by openring

SourceHut API 2.0 dev log

Completing the project now known as “API 2.0” is one of the most important steps in finalizing the beta, and I’ve been trying to keep the community abreast of developments, especially on the mailing list. I’m summarizing for the blog as well t…

via Blogs on Sourcehut September 25, 2020

People expect technology to suck because it actually sucks

I get at least 27 tech annoyances on a single day

via September 25, 2020

Rewriting the Technical Interview

Frigitt, vi danser For et øyeblikk, vi leker Vi tusen små bladskiper Gleder oss, på det klare morgenlys The mothercone unfurls her softening scales, and in the time of days I am emancipated. Adrift in the chill dawn light, my siblings flutter, slender forms c…

via Aphyr: Posts September 18, 2020