A Review of Shell Features

What’s to like about the shell?

Now, just to be clear, we’re not asking if the ability to join two commands together in a pipeline is something to be admired – I like to think that we can stumble through the pipe(2), fork(2) and execve(2) man pages with qualified success – but rather whether the syntactic abstraction the shell uses, here, command | command, or something closely resembling it, is what we want in our shell.

If it is then we need to start thinking about how we are going to implement it. Many of these abstractions are infix, arithmetic is another example, 1 + 2. I, for one, don’t read this as add (1, 1) or pipe ("zcat file.tgz", "tar tf -"), I’m reading it as a sequence of operations, one I know I can arbitrarily extend.

We can add more arithmetic, 1 + 2 * 3 (let’s hope we agree on operator precedence!), or more filtering, zcat file.tgz | tar tf - | grep foo, casually. It’s more complicated with function calls:

add (1,
     mul (2, 3));

pipe ("zcat file.tgz",
      pipe ("tar xf -",
            "grep foo"));

They look less straightforward and elegant and instead clumsy and forced. We can’t see the wood for the trees.

Before we casually toss the problem of inline operators at yacc/bison or go for it with ANTLR we need to know if we are even going to use everyone’s favourite language parsers. Hint: no. Not because they are easy but because they are hard.

Syntactic Structure

Without getting lost in the detail the syntactic structure of shell commands is very clean:

ls -al *.txt

There’s no pre-amble, superfluous syntactic clutter and no trailing end-of-statement marker, the end of the line terminates the statement (usually). In fact, that last point, that the shell is, by and large, line-oriented is something I’m quite keen to keep – although it has some unpleasant side-effects.

The first word is the name of the command to run and the remaining words, separated by one or more whitespace characters are the command’s arguments.

Compare that with most languages which, for example, in C:

func (val1, val2);

Here we have the same word ordering (command then arguments) but also have parenthesis separating the function name from the arguments and, indeed, commas separating the arguments. Now it doesn’t take long to realise why there’s all that punctuation as languages like C allow you to recursively call other functions in place of arguments:

func (sub1 (a1, a2), sub2 (b1, b2));

Which now has an impressive 10 pieces of punctuation and is getting hard to read. I, depending on whim, might have rewritten it as:

func (sub1 (a1, a2),
      sub2 (b1, b2));

(which is either more pleasant or is a calculated affront to public decency and has you caused spontaneous apoplectic rage. Such is the modern way.)

Of interest, Scheme with its notorious superfluity of parenthesis, looks like:

(func (sub1 a1 a2)
      (sub2 b1 b2))

Go figure.

You can do the subroutine calling of a sorts with the shell’s command substitution operator $():

ls -al $(generate_txt_file_names)

but you can’t do that multiple times and distinguish the results:

func $(sub1 $a1 $a2) $(sub2 $b1 $b2)

Here, func is going to see one long list of arguments, not two (sets of arguments) as in C.

Variables Syntax

If I was going to drop one thing from shell syntax it would be its use of sigils. For the shell, just the one, $, used to introduce variables.

I think it’s visual clutter and we should be more C- or Python-like:

a = func (b, c);

rather than the mish-mash of:

a=${b[i]}
func $b $c

When do I use a sigil and when not? Get rid of the lot.

(Except when we need them.)

Having said that, in the shell it is very common to use a variable embedded in the assignment of another or in a string (string interpolation).

So we might do:

PATH=/usr/local/bin:$PATH

echo "PATH=$PATH"

Both of those are very convenient, I have to say. Are they required, though? What would we do elsewhere? Actually, for the former, I’m sure I’m not the only one who wrote a bunch of shell functions to manipulate Unix paths (PATH, LD_LIBRARY_PATH, PERL5LIB etc.) resulting in something like:

path_prepend PATH /usr/local/bin

and you can imagine variants for modifying multiple related paths simultaneously and various path normalisation functions (removing duplicates etc.). The style is more programming language-like, in Python you might say:

sys.path.append ("/usr/local/lib/python")

As for string interpolation, I’m less sure I’d miss it in the above for as I might use a format string of some kind as in Perl:

printf "PATH=%s\n", $ENV{PATH};

or Python

print ("PATH={0}\n".format (os.environ['PATH']))

Clearly, there’s a lot more syntactic clutter, which I’m nominally against, but I usually end up trying to control the output anyway.

Here-document

There’s another embedded-variable situation which I use often enough where I’m generating a block of output (often a usage statement) or a code-snippet for another language. I’ll be using a here-document (a terrific bit of left-field thinking):

perl << EOF
printf "PATH=$PATH\n";
EOF

where I’m ostensibly creating a (multi-line) string but substituting in some variables.

You can get quite quickly annoyed, though, when Bash’s sigil, $, conflicts with, in this case, one of Perl’s sigils leaving you to judiciously escape some of them:

perl << EOF
my \$path = "$PATH";
printf "PATH=%s\n", \$path;
EOF

Here-documents in general, and code-snippets for other languages, in particular, look at though they need a bit more thought. It seems like the here-document is some sort of template in which we perform variable substitution when we see a $ sigil introduce a variable, $var.

The “other languages” variant is also crying out for some means to choose the sigil itself. Wouldn’t it be great if we could have written that as:

perl << EOF
my $path = "!PATH";
printf "PATH=%s\n", $path;
EOF

where ! – or some other sigil that you get to choose, something appropriate to your template – means we can write as-native-as-possible Perl, in this case, with considerably less hassle?

I have a cunning plan.

Environment Variables

Another simple yet clever trick the shell plays is that all environment variables are exposed as shell variables. We don’t need to call getenv and putenv/setenv or indirect through a named hash table like Perl’s $ENV{} or Python’s os.environ[]. They’re right there as primary variables in the script.

I like that. Remember, we’re operating at the level of orchestrating programs and it seems right that we should have direct access to those things we manipulate often.

Clearly there’s a little bit of magic floating about as some shell variables are marked for export and some not.

Shell Commands

Bash, at least, has simple commands, pipelines, lists, compound commands, coprocesses and function definitions. Let’s take a look at those.

Simple Commands

From bash(1):

A simple command is a sequence of optional variable assignments followed by blank-separated words and redirections, and terminated by a control operator. The first word specifies the command to be executed, and is passed as argument zero. The remaining words are passed as arguments to the invoked command.

A control operator is one of || & && ; ;; ;& ;;& ( ) | |& <newline>.

Variable Assignments

Variable assignment is something I do use:

TZ=GMT+0 date

PATH=/somewhere/else/first:$PATH cmd args

This is a really neat trick, we make a change to the pending command’s environment (but, crucially, not our own).

It also looks tricky to parse, we need to be able to figure out the following:

CFLAGS=one-thing make CFLAGS=another-thing

I don’t like that, it looks too hard. We could have achieved the same with a subshell:

(
 PATH=/somewhere/else:$PATH

 cmd args
)

Where, if we forget that the parenthesis have introduced a subshell and think of it as a code block, I’m getting a sense of a transient assignment, that PATH is a dynamic variable, we modify it for the duration of the code block and that come the time we need to use it, we figure out its value then.

Interestingly, and I said I learned something new every time I read the man page, if Bash determines there is no command then the variable assignments do affect the current shell’s environment. So, changing the current shell’s PATH:

PATH=/somewhere/else:$PATH

is a side-effect of the absence of a command rather than an explicit shell-modifying statement in its own right. Who knew? (The guy that wrote the man page did, for a start.)

Redirections

Bash, at least, is fairly free with redirections, in the sense that you can have lots of them and they are processed left to right. So:

ls -al > foo > bar

will create both files foo and bar but foo will be empty and only bar will have any contents.

I guess you’re not meant to be doing that, as it looks like a mistake, but rather more something like a sequence of pseudo-dependent redirects:

exec >log-file 2>&1

Here, of course, the order is critical as we are redirecting stdout to log-file and then redirecting stderr to wherever stdout is currently (ie. log-file not wherever it was when we started processing the line).

Another example is stashing the current whereabouts of a file descriptor:

exec 3>&1 >log-file
...
exec >&3

where we redirect file descriptor 3 to wherever stdout is currently pointing and then stdout to log-file. We do our thing with stdout going to log-file and then redirect stdout to wherever file descriptor 3 is pointing – handily, where stdout was originally pointing.

It’s a neat trick and handy in a script for directing tedious output to a log file whilst simultaneously retaining the ability to print to the original stdout with the likes of echo "don't stop believing..." >&3 to keep the user’s hopes up but it has a terrible failing. We, the punter, are somehow supposed to know when file descriptors are free to use. How do we know that? Ans: we don’t, we just stomp over, in this case, file descriptor 3 regardless.

In other languages, such as Python, you might see an expression of the form:

with open ("log-file", "w") as f:
    f.write ()

which is nearly what we want – we actually want to transiently replace existing file descriptors and in such a way that they can be inherited by any commands we run. Something more like:

with open ("log-file", "w") as stdout:
    ...

(albeit we’ve skipped our provision to keep the user up to date.)

Obviously IO redirection is a requirement but I sense that this carefree way with file descriptors is really because the shell can’t maintain a reference to a file descriptor outside of command invocation. In a programming language with access to the usual systems programming libraries we’d be calling dup(2) in some form and be able to pass the return value around as you would hope.

One problem, though, with our programming language design hats on, is that this IO redirection is inline – in the sense that the IO redirection is textually mixed up with the command and its arguments – and more particularly is an infix operation. It’s very convenient to have it parsed out of the line for us but that, to us doing language design, is a problem. We now have to parse it out. I’m not getting good vibes about that.

Pipelines

bzcat logfile | grep pattern | sort -k 2n > file

There’s a certain elegant simplicity in writing a shell pipeline, the output of a command piped into another command, the output of which, in turn, is piped into a further command… and with the final output redirected into a file.

If you’ve not had the pleasure of implementing a pipeline of commands then that joy awaits us later on. Probably more than once (because we are/should be committed).

However, what this pipeline (and its minimalist variant friend, the simple command, above) overlooks is that the shell is manipulating a secondary characteristic of Unix commands one we need to be in control of.

By and large, when we construct a pipeline, in particular, or even a simple command we, the user, are looking for some side-effect of the output of that command. The pipeline is, well, can be, affected by the exit status of any component of the pipeline but that’s not the effect we’re looking for. We want the output stream of one command to be filtered by the next and so on but we are agnostic to the command status along the way.

The shell, however, is agnostic to the output and predicates success or failure of the pipeline on the command status alone – well, the command status of the last component of the pipeline – and only if you ask it to.

That is to say that:

something | grep foo

fails, not because there was no output but because grep exits non-zero if it cannot match the regular expression (foo) in its input stream.

Importantly, the pipeline succeeds if grep does match foo even if something crashed and burned spewing errors left right and center. So long as it managed to splutter foo, somehow, before it died, then grep is happy and therefore the pipeline is happy.

The canonical example is to have everything fail except the final component of the pipeline:

false | false | false | true
echo $?
0

is all good. Bash’s PIPESTATUS variable is a little more honest:

false | false | false | true
echo ${PIPESTATUS[*]}
1 1 1 0

If the command output vs. exit status is not a familiar distinction then 1) we’re not going to be best friends and 2) try putting:

set -e

at the top of your script and sit back and watch the fireworks. Not in production, though. That might be bad, fix your script first. (In fact, try set -eu and patch up the mess.)

In most programming languages, when you invoke a command or function call you pass some arguments and you expect a result. With the shell we do get a result, an exit status, albeit commonly overlooked. We’re not going to be able to overlook it with one particularly good example in if, below.

All these commands, and, rather consistently, the builtins and user-defined functions, return a status with zero indicating success and any other number being a command-specific failure. (There is a set of common exit statuses relating to whether the command was killed by a signal but either way the result is just a simple 8-bit number.)

Pipelines have the same problem as IO redirection, though, they are quite obviously inline and infix again.

We need to put our thinking caps on. Not our sleeping caps, those nodding off at the back, our thinking caps. How do we handle inline operators?

Lists

Lists, here in Bash at least, are pipelines separated by a subtle combination of statement terminators and logical operands.

For logical operands, Bash uses && and || although I personally prefer Perl’s and and or – which, coincidentally, match Scheme’s and and or – but also free up && and || for other uses.

; terminates a statement/pipeline as does a newline. I’m not so keen on ; as I don’t use it anywhere other than mandated syntax (in if or while etc.) or one-liners.

Bah! I do a lot of one-liners interactively, usually to finesse some complicated filter whereon I then I reuse it with $(!!):

...fiddles...
...more fiddling...
...perfecto!...
for x in $(!!) ; do thing with $x ; done

So maybe it’s a thing.

& is a weird, cross-over, end-of-statement marker and signal to run the pipeline in the background. Putting stuff in the background is something, I think, people are fairly used to:

sleep 10 &

which is a slightly pointless example but easily recognisable as putting the command “in the background” (whatever that means).

& is probably less used as a statement separator because things can get a bit wild:

for x in {1..10} ; do sleep $x & done

Notice I have & instead of ; before the done keyword. The shell will immediately kick off ten background processes. Hitting RETURN a few times over the next ten seconds should get a staggered notification that ten sleep $xs have completed.

Back to our syntax considerations. The logical operators are clearly infix and the statement terminators are some strange infix or postfix operation.

Compound Commands

Compound commands are interesting because several of them operate on Lists which are built from Pipelines which are built from Shell Commands which are built from, er, compound commands. I’m pretty sure I’ve done an if case ... combo but the therapy is helping a lot.

Let’s take a look at them.

Subshells

I use subshells quite a lot but I think I largely use them so I can cd somewhere without affecting the current shell. Even better when backgrounded:

for x in ... ; do
    (
        cd some/where

        mess with the environment

        do some thing with $x
    ) &
done
wait

As we know – or will know if we don’t – every command we run is initially in a subshell because we have forked and are about to exec. So is ( ... ) syntactic sugar for fork and run this block of code?

Group Command

I’m scratching my brain here but I can’t think of anywhere where I’ve used { ... } other than as a function body – and there to such an extent that I think I’ve only once written a function that didn’t use { ... }.

It’s unclear to me what other semantic behaviour a group command in Bash has. Clearly(?), a device that runs a sequence of commands and returns the result of the last one run is a useful programming construct. Is there anything more to it? I haven’t followed the code through.

Let Expression

I have to assume that (( ... )) is only meant to be used in an if or while conditional expression as it explicitly returns a non-zero status if the arithmetic result is zero. So that’s of no use whatsoever in general flow with any kind of error handling (set -e or trap ... ERR).

If I want to do sums I use Arithmetic Expansion.

Conditional Expression

Hopefully most people use [[ ... ]] rather than the anachronistic [ ... ] which, is a synonym for test(1) and, depending on the operating system, has some weird rules on the number of arguments affecting how it behaves. Yep, not what they are but the number of arguments. Just use [[ ... ]].

[[ does have a heap of command-specific operators including && and || leaving the possibility of:

[[ $a || $b ]] || [[ $c || $d ]]

where the middle || is managing pipelines and the outer ||s are managing conditional expressions.

It’s inside [[ that we get regular expression matching (as well as “regular” Pattern Matching).

There’s a lot of behaviour loaded in [[ that feels like it’s bundled in because that’s the only place it could fit – or [[ was designed to be extended arbitrarily.

Certainly regular expressions are just normal function calls in other programming languages and you feel that much of the rest of it should be as well.

For

There’s two variants of for: the common iterator, for x in ...; and the C-like for (init; condition; step) ....

The former is used all the time. People like iterating over things and programmers get quite angry when they can’t. The latter, I don’t think I’ve used in a shell. I must have…surely?

Select

select is a peculiar beastie. I don’t use it. Maybe it’s more useful than I think. Who wants to interact with users anyway? They’ll only type the wrong thing.

Case

I use case a lot as my GoTo means for doing conditional Pattern Matching:

HOSTNAME=$(uname -n)
case "${HOSTNAME}" in
    '') echo "no hostname?" ;;
    *.*) ;;
    *) echo "need a FQDN!" ;;
esac

Pattern matching is great!

If

We’ll take a moment with if as it illustrates something quite ingenious about the shell. We’re back to this exit status business.

The basic syntax is:

if list ; then list ; [ else list ; ] fi

(I’ve skipped the elif bits.)

You’ll probably have called if two ways:

if [[ ... ]] ; then ... ; fi

if cmd args ; then ... ; fi

The first we always look at as a test, like in most languages, if is:

if condition then consequent else alternative

So if [[ ... ]] ; then ... looks like the regular programming language version. Except it’s not, it is exactly the other form of if:

if cmd args ; then ... ; fi

because [[ ... ]] is a builtin command which returns a status code of 0 or 1. Which brings us neatly back round to the exit status.

if will run the condition as a command and irrespective of the output or other side-effects will determine the truthiness (stop me if I’m getting too technical) based on the exit status of the command (technically, the exit status of the condition list).

That if is conditional on the exit status of the command is also evident when it is masked by the syntactic sugar of command substitution:

if output=$(cmd args) ; then ... ; fi
Trapped If

Of interest is that if will not trigger an error trap. That’s obviously what you want, at least I think it is obvious:

if something | grep foo ; then
    ...
fi

You don’t want your shell to exit (set -e) if grep doesn’t get a match. That’s the whole point of it being in a conditional test. Compare that with:

# helpful debug!
something | grep foo

if something | grep foo ; then
    ...
fi

where you’ll not reach the if statement because the failure to match foo will cause grep to exit non-zero, set -e will then exit your script and you’ll be none-the-wiser having seen nothing printed out.

Similarly, while and the logical operators && and || also mask any error trap.

While

Interestingly, I’ve done OK without having while in my arsenal. As we know (cue recovered memories from computer science classes) iterative control flow operators (eg., while) can be re-written as recursive function calls (and vice versa) and Scheme has a big thing about being able to do tail-call recursion so, uh, that’s what you do.

Like many things while is syntactic sugar for what’s really happening underneath the hood.

Co-processes

These are relatively new to Bash though they’re been in other shells, eg. Ksh.

I’ve not used them. Dunno.

In a programming language we’d simply have several file descriptors floating about we can read from or write to at our leisure. No need for specific co-processes.

Function Definitions

A given.

Technically, the body of a shell function is a compound command hence why most function bodies look like { ... }, the group command, but it could be a single if or case statement.

I’m not sure I’ve ever used the IO redirection for a shell function.

One weirdness, in Ksh, at least, regards whether you declare the function with the keyword function:

foo() { ... }

function bar { ... }

and therefore whether a trap on EXIT is executed.

Expansion

Slightly out of order from the man page but expansion is easily the shell’s most distinguishing feature and, likely therefore, its most misunderstood.

The real bugbears are brace expansion, word splitting and pathname expansion (and parameter/array expansion) because they change the number of words in the command expression.

That’s bonkers! Are there any other languages which actively change the number of words they are processing? (There must be, I can’t think of any.) It defies any form of programmatic rigour when you can’t determine the arguments you have to hand:

${TAR} ${TAR_FLAGS}

Is that potentially erroneous because we did or did not pass the tar archive (and optional files) in ${TAR_FLAGS} or are they passed in ${TAR_FLAGS} and all is well.

Is either variable even set?

I have written code like:

${DEBUG} cmd args

where ${DEBUG} can optionally be set in the environment and is therefore either nothing, in which case ${DEBUG} cmd args is expanded to just cmd args and is run as you would expect, or it is set to, say, echo, in which case the expansion is echo cmd args and cmd args is echoed to stdout (and not run).

Programmatically, then, shell commands can be worryingly non-deterministic – and we haven’t even started on what cmd is anyway, a shell function, a shell builtin, an executable expected to be found on your PATH?

Expansion isn’t quite performed all at once, brace, tilde, parameter, arithmetic, command (and optionally, process) substitution are performed in that order, left-to-right and then word splitting and pathname expansion are applied before final quote removal.

There’s a lot going on!

Brace Expansion

Brace expansion comes in two forms: a comma variant and a sequence variant.

ls /usr/{bin,lib}

echo {01..16}

The former saves us writing a loop and the latter saves us calling seq (not available on all platforms) – although brace expansion does implicitly include the leading-zeroes formatting, see the -w flag to seq).

Is this syntax something we need, though? I’m not sure. There’s clearly a function returning a list of (formatted) strings which we could just call, very much like seq:

for x in {01..16} ; do ...

for x in $(seq -w 1 16) ; do ...

We can probably live without this syntax.

Tilde Expansion

~me, ~you

Of interest is that the shell also checks not just simple variable assignments, HERE=~me but also after :s in variable assignments so that EVERYWHERE=~me:~you does the right thing.

It’s nice enough though I fancy I’ve only really used it interactively. I’m happy to be manipulating paths in a more long-winded fashion so maybe I’ve happy enough to make an extra call where I thought it was required:

path_prepend EVERYWHERE $(tilde_expand ~me)

although, as we know tilde expansion is the pw_dir field from a getpwnam(3) call, then in some putative language it might look more like:

me_dir = (getpwnam "me").pw_dir
path_prepend EVERYWHERE me_dir

which is clearly more “work” but suggests we’ve more (systems programming) access to the data sources. We need to handle failures, of course – we might need to have an existential crisis if me doesn’t exist, for example.

So I’m thinking that the syntax for tilde expansion isn’t required.

Parameter Expansion

I use parameter expansion a lot. I particularly use it for manipulating pathnames where dirname and basename can be replaced with ${FOO%/*} and ${FOO##*/} respectively. Who doesn’t want to do array pattern substitution, ${array[*]/%\/bin/\/lib}? Who, who?

Others might think this mix of terse syntax and pattern-matching gives Perl a good name.

I can understand that, even when you point at %/* and say it’s two parts, a remove-shortest-match-at-the-end, %, for the pattern /* (which is a loaded concept in its own right), it still befuddles non-shell programmers.

It’s hook has probably been slung – which is a good thing, as I don’t fancy trying to replicate any of that terseness. We’ll just have to plod along manipulating our strings bit by bit like everyone else.

Command Substitution

I use command substitution, $( ... ), all the time as well. I’m usually collecting some fact from a command, often a pipeline:

HOSTNAME=$(uname -n)

CIDR=$(ip addr show dev lo | awk '$1 ~ /^inet$/ {print $2}')

(no second guessing on the loopback device’s IPv4 address – although I’m assuming there’s only one, here!)

Mechanically, of course, the command we run knows nothing about us and our attempts to capture its output. It’ll be printing to its stdout, nothing more, nothing less. The trick is, of course, to:

  • create a temporary file

  • redirect the command’s stdout to that file

  • run the command (duh!)

  • read the contents of the temporary file

  • delete the temporary file

Wrap all that up in the syntactic sugar of $( ... ) and we’re done. Very neat!

A requirement, surely!

That said, command expansion is one of the most guilty parties for introducing unexpected whitespace (whitespace as the lesser of such evils). Were we to be blessed with the directory My Documents in the current directory:

ls $(ls)

results in:

ls: My: No such file or directory
ls: Documents: No such file or directory

Why? Performing the expansion by hand we see:

ls My Documents

and we probably wanted:

ls "My Documents"

There is no general solution to unexpected whitespace, newlines, etc. introduced by command expansion! The worst of which will be a shell-ish Little Bobby Tables

Arithmetic Expansion

Attention

Please, no more expr!

We can do sums in the shell:

echo $(( 1 + 1 ))

Performs the arithmetic and replaces the expression with:

echo 2

Usefully, during arithmetic expansion you don’t need to perform parameter expansion, that is, you can use variables without the $ sigil:

p=2
echo $(( $p + 2 )) $(( p + 3 ))

will become:

echo 4 5

Notice, however, that parameter expansion can be your undoing:

echo $(( p++ ))

will become:

echo 2

and p will now have the value 3. However, were we to have typed:

echo $(( $p++ ))

parameter expansion will have gotten there first:

echo $(( 2++ ))

which is, obviously(?), an error.

Arithmetic expansion occurs in array index calculations such as ${array[base+offset]} and ${array[i++]}.

Arithmetic is assumed, I suppose, for a programming language, though several of the C-like operators will be hived off as functions if available at all (think bitwise operators).

However, one thing to note is that the C-like operators are, largely, infix binary operators, that is, they take two arguments, one before the operator and one after:

1 + 2

Everyone does that, right? No, not really. They’re in the camp of yet another (set of) infix operator(s).

Process Substitution

On systems supporting named pipes we can substitute a filename for a dynamic stream:

diff expected-result <(cmd args)

results in something like:

diff expected-result /dev/fd/M

where /dev/fd/M is the filename for the file descriptor representing the output of the pipeline from the invocation of cmd args.

This is really useful for commands like diff which only operate on files.

Another use case is where you have a requirement to iterate over the output of a command and to modify a local variable:

cmd args | while read line ; do
    local_var=$(process ${line})
done

doesn’t work because the while loop, as part of a command pipeline, is in a subshell so modifications to local_var have no effect in us, the parent shell, where we want them. You need to rewrite this to, say:

while read line ; do
    local_var=$(process ${line})
done < <(cmd args)

which will be expanded to something like:

while read line ; do
    local_var=$(process ${line})
done < /dev/fd/M

It’s useful functionality for the shell where we can’t hold a file descriptor open to a sub-process (although see co-process, above) in the latter case.

In the former case does it justify a special syntax? Maybe. It is, I think, more or less a function call, something along the lines of:

diff expected-result $(named-pipe cmd args)

Word Splitting

This is where the problems usually start!

Quoth bash(1):

The shell scans the results of parameter expansion, command substitution, and arithmetic expansion that did not occur within double quotes for word splitting.

The shell splits the result of the expansion based on the contents of the IFS variable (usually the ASCII characters SPACE, TAB and NEWLINE).

Note

This word splitting isn’t a rigorous split on every occurrence of a delimiter in IFS in that a sequence of IFS characters only generate a single split.

Naturally, it’s more complicated than that. RTFM!

So, casual use of space-containing variables:

dir="My Documents"
ls ${dir}

expands to:

ls My Documents

and fails because word splitting thinks you’ve passed two separate arguments, My and Documents to the command, just like we we might pass two arguments, -a and -l to ls if we typed ls -a -l. We should have written:

ls "${dir}"

which expands to:

ls "My Documents"

and works as expected.

A general rule of thumb is to double quote everything unless otherwise advised.

This whole word splitting thing is a bit of a nightmare. If we’re going to have complex data structures then you feel that these whitespace-including entities should be being passed as proper parameters and that:

dir="My Documents"
ls ${dir}

should work as expected without word splitting.

I’m not a splitter.

Pathname Expansion

Pathname expansion should be a requirement but, it transpires, it’s going to be tricky for some other reasons that we’ll get onto in due course. In the meanwhile, pattern matching must be one of the finest examples of abstraction and utility in computing!

Famously, or not:

ls *

does not pass * to ls (most of the time). Rather, the shell has been looking for meta-characters, in particular, *, ? and [. In Bash the extglob shell option adds some more matching operators (some of which are available by default in other shells).

When it identifies a meta-character in a word then the whole word is treated as a pattern and filename Pattern Matching, aka globbing, begins. Globbing began life at the very beginning of Unix as a standalone program, glob (authored by one Dennis Ritchie), so it’s “got some previous” but is now more readily available as a library call, glob(3). I think Bash still rolls its own version, it certainly has the code in a subdirectory (.../lib/glob).

More importantly to us are the results of pathname expansion. We get back a list of filenames. I would suggest that we subsequently preserve that list and pass it around as you might a list in a regular programming language.

When we finally get around to running a command (remind me, that’s why we’re here, right?) we can expand the list, preserving whitespace quite happily.

Sorted glob

There’s a more general irritation with pathname expansion: you can’t sort the results based on attributes of the files your pattern matches or, indeed, any other arbitrary sorting scheme. The shell appears to do a lexicographical sort on its results (possibly locale(1)-specific).

There are any number of situations where a list of files could be bettered by being sorted based on modification date or size or qualified by ownership for which you have to break out to another command (and suffer the problems of managing the quoting correctly) to gather the results back. Or you rewrite everything in another programming language…

There’s room here, I think, for the results of glob(3) (or whatever) to be passed to something that can sort and/or filter the results based on rules of its own choosing.

Hint: sorted.

Pattern Matching

Pattern matching is a bit like regular expression matching (if * and ? became .* and .?) until all the subtleties kick in. An obvious one is that ls * will not report any dot files (unless you explicitly match the leading dot with .*, or the dotglob shell option is set and even then you must match . and .. explicitly).

Luckily for us, someone has written glob(3) and we can claim uniformity with “everyone else” (noting that many others will be rolling their own variation, standards, eh?). We could go there but it’s certainly not necessary right now.

Quote Removal

If we don’t do Word Splitting we don’t need to do any quote removal. That’s the plan!

Quoting

I’m hoping that (what feels like) the normal use for quoting things, to avoid Word Splitting, is nullified because we’re not going to do any word splitting.

However, it is convenient to build strings using variables:

echo "PATH=$PATH"

which would require something like the templating mechanism I hinted at for here-documents.

Parameters

Parameters being variously positional parameters, special parameters and variables.

First though some parameter attributes.

Parameter Attributes

Variables are, by default, shared between the main program and shell functions. You can restrict the scope of a variable to be local to the function it is declared in. I suppose, technically, the group command it is declared in (not sure).

We previously mentioned that some variables are tagged for export and I’m thinking about them in terms of variables with dynamic scope, whose values come and go based on the run-time path the code takes.

You can also mark a variable as readonly.

A final classification is that of being an alias. I keep thinking that having synonyms would be handy but then keep being reminded that even in Bash the suggestion is to use a function instead. Your synonym function simply calls the aliased function with all of its arguments.

Shell Variables

(Of interest to us with design in mind!)

$!

The PID of the last backgound(ed) command.

One shot or you’ve missed it. I envisage the ability to go back and query interesting things about your child processes at any time although it is possible, like the backgrounded sleeps above, there’s no easy way to distinguish between them.

$?

In Bash it is the exit status of the most recently executed foreground pipeline.

I’ve emphasised foreground, there, as I confess it hadn’t occurred to me. It’s only really useful if you’re not running any error handling (set -e or trap ... ERR) otherwise the value will be 0 or your script has errored.

If you run a bunch of processes in the background (remembering their PIDs with $!) then wait for each PID in turn then each wait (being run as a foreground command) returns as its own exit status the exit status of the PID it was waiting on. A little bit of digital legerdemain.

Being able to reference a (child) process’ status is a useful thing. There’s a not unreasonable argument that says being able to access its status at any time is a good thing.

PWD

PWD is the current working directory as set by cd – not the result of getcwd(3).

Here, you are maintaining the logical path in your rat’s nest of symlinks in the underlying filesystem.

SECONDS (and RANDOM)

I like SECONDS and use it to report the elapsed time of a script (genius!) but think about what it is. When you reference it you get back a volatile value. It is what we might describe as a computed variable in the sense that when it is queried some function is called and the value returned by the function is the value of the variable.

Signals and traps

Signals in the shell are as complicated as anywhere else compounded by the “rules” for job control.

By and large, I avoid getting involved as it’s hard and prone to hard-to-repeat errors. Which is a shame as I’m now trying to write shell which needs to deal with signals.

In addition to regular Unix signals there are a few fake signals: DEBUG, RETURN, EXIT and ERR. I only use EXIT and ERR and I use them all the time.

A trap on EXIT is executed before the shell terminates. Quite when is less clear but it seems close enough to the end to do any clearing up. If you were of a sort to create a temporary directory and do all your processing in there then an EXIT handler can easily rm -rf the temporary data.

A trap on ERR is my GoTo replacement for set -e. The problem with set -e is that your script just dies. I’d like to know a little more so I tend to have something like:

handle_ERR ()
{
    echo "ERROR at line $1: exit ($2)" >&2
    exit $2
}

trap 'handle_ERR $LINENO $?' ERR

which gives me a few more clues.

It’s not foolproof as the line number is often reported as the end of a compound command which could be anything within.

Until recently, functions and subshells did not inherit this trap which was spectacularly annoying.

So, if there’s one thing to do, we must be able to handle errors decently.

Job Control

Job control is the idea of selectively stopping and resuming processes that you have backgrounded. A job, here, means a pipeline. When a pipeline is launched all the processes in the pipeline share a process group. If the pipeline/job is in the foreground then that process group is associated with the terminal such that any keyboard signals raised go to that process group – and not to the shell nor any of the backgrounded or stopped jobs.

You can’t have a foreground job in the sense that anything running in the foreground is receiving input from the controlling terminal and the shell (and any backgrounded jobs) are not. If the shell isn’t getting any input then it’s not controlling anything, let alone a job.

You can’t, therefore, have the shell do anything until the foreground command completes or is stopped. The shell’s very purpose in life is to hang about waiting for its children, in this case the foreground process, to complete/stop, so it will then re-arrange signals and the terminal and will continue doing shell-like things.

When you have a foreground pipeline running and you hit Ctrl-C (or other signal-raising keystroke) the SIGINT is sent to the process group associated with the terminal and the processes within that process group will act as they see fit.

The shell, for example, has a SIGINT handler – primarily to interrupt wait – but it is functionally ignored. Thus, when the shell itself is “foreground”, Ctrl-C doesn’t do much.

If you did run a pipeline in the foreground you can raise a “terminal stop” signal, SIGTSTP, slightly confusingly called the terminal suspend character, usually, Ctrl-Z. Assuming it is not ignored (a shell usually ignores it!) then the default signal disposition means the pipeline will stop, the shell is signalled that a child process has changed state (SIGCHLD) and can handle the pipeline as a job.

Note that the pipeline/job you just stopped is still stopped. In most shells you can immediately bg the pipeline/job to let it carry on processing.

Foregrounding and backgrounding involves careful manipulation of process groups, signals and the state of the controlling terminal. It’s quite complicated but, to the relief of all, there’s a handy description of most requirements in the info(1) pages for libc under the menu item Job Control or try the equivalent online Job Control web pages.

Much of the complexity of job control is for interactive sessions. Non-interactive shell scripts can still background jobs and the signalling and terminal management differ.

Last built at 2024-12-21T07:11:18Z+0000 from 463152b (dev)