Do It Yourself with the Shell

Ed Schaefer and Michael Wang

Handy Unix tools such as Perl, awk, sed, and grep are frequently used to solve programming problems, while the shell may be overlooked as a problem-solving tool. Most of us garner Unix knowledge by learning Unix commands and, to the uninitiated, shell programming is just chaining a series of commands together. As a command interpreter, the shell does not distinguish its built-in commands from external commands; you do not call external commands using a function as Perl's system() function. Using the shell over an external tool is analogous to performing a task yourself versus hiring somebody else to do it. Provided you have the knowledge and experience, it's generally better to do it yourself.

This article presents "do it yourself" techniques in shell using real-world examples -- examples gathered from printed material, USENET postings, and our own work. In this article, we contrast standard Unix tool solutions with corresponding shell solutions and compare the different approaches. We present 13 specific sample problems with corresponding solutions within 5 broader technique categories. The techniques considered are: parameter expansion and positional parameters, pattern matching, while-loop, shell arithmetic, and file and process operations.

All of the shell solutions were tested using bash 2.x and ksh93. Most of them work under ksh88 unaltered or with minor modification; however, we did not purposefully twist the code to work under ksh88. Major commercial Unixes (e.g., Solaris, AIX, HP-UX) ship with ksh93. Recent versions of ksh93 and bash are available via open source licenses. When we refer to portability, we mean portability across different platforms, not across different shells.

Technique: Parameter Expansion and Positional Parameters

Problem 1: Find Year, Month, Date from String

Consider the string i="2003-07-04" in the format of YYYY-MM-DD. Find the year, month, and day. Here are two possible awk solutions (Examples 1a and 1b):

set -- $(echo $i | awk '{gsub(/-/, " "); print}')
set -- $(echo $i | awk -F- '{print $1" "$2" "$3}')
# Your Unix variant might require nawk

set -- $(echo $i | sed "s/-/ /g")
set -- $(echo $i | sed "s/\([0-9]*\)-\([0-9]*\)-\([0-9]*\)/\1 \2 \3/")

set -- $(echo $i | perl -pe "s/-/ /g")
set -- $(echo $i | perl -ne "print(STDOUT join(' ', split(/-/)))")

set -- $(IFS=-; echo $i)

The use of "subshell $(...)" localizes IFS to the subshell environment, for the purpose of $i expansion. IFS is not affected in the parent shell environment.

echo works in our example because the string in $i is known. In general, it may contain a space or other shell special character. The following shell command takes care of these (Example 1h):

eval set -- $(IFS=-; printf '"%s" ' $i)

oIFS=$IFS
IFS=-
set -- $i
IFS=$oIFS

set -- ${i//-/ }

eval set -- \"${i//-/\" \"}\"

The third shell solution (Example 1j) looks like this:

year=${i%%-*}
month=${i#*-}
month=${month%-*}
day=${i##*-}

year=$(echo $i | sed "s:-.*$::")
month=$(echo $i | sed "s:^[^-]*-::")
month=$(echo $month | sed "s:-[^-]*$::")
day=$(echo $i | sed "s:^.*-::")

year=${i%%????}
month=${i#????}
month=${month%??}
day=${i##??????}

Using the original format, YYYY-MM-DD, the fourth shell solution (Example 1m) is probably the most straightforward solution, if the exact position of the data is known:

year=${i:0:4}
month=${i:5:2}
day=${i:8}

year=$(echo $i | cut -c1-4)
month=$(echo $i | cut -c6-7)
day=$(echo $i | cut -c9-)

As shown in these examples, the shell provides multiple ways to do the string manipulation. There is no reason to use sed, awk, or Perl to perform this function. The shell solution is preferred for independence and efficiency. We do not need to worry about PATH, tool availability, and tool behavior on different platforms.

Performance can be an issue when using external commands -- especially in a loop. Suppose you need to process a large file, editing data in each line. The time saved using the shell over other tools may be significant.

Problem 2: File System Usage

Parse df -Pk /usr output, and find the usage percentage. The -P implies the POSIX output format. The POSIX-compliant command under Solaris is /usr/xpg4/bin/df. Example 2a shows an awk solution:

# may require nawk under Solaris 7 & SCO Open Server V
df -Pk /usr | awk '!/^Filesystem/ {sub(/%/, ""); print $5}'

df -Pk /usr | sed -n '/^Filesystem/!s/.* \ \([0-9]\{1,\}\)%.*/\1/p'

df -Pk /usr | perl -pe "BEGIN { $/ = '' } s/.*\s([0-9]+)%\s.*/\$1\n/s"

df -Pk /usr | grep -v "^Filesystem" | awk '{print $5}' | sed "s:%::"

set -- $(df -Pk /usr); echo ${12%\%}

$ df -Pk
Filesystem             Kbytes     Used    Avail Capacity Mounted on
C:\Program Files\UWIN  4000153  3868891   131262     97% /

A second shell solution (Example 2f) is as follows:

(IFS=%; set -- $(df -Pk /usr); echo ${1##* })

Problem 2: Discussion

Example 2f demonstrates simple techniques on this not-so-obvious case, providing the shortest, quickest, and most elegant solution.

Technique: Pattern Matching

Problem 3: Find Whether $i Is an Integer

To solve this problem, we need to define an integer. If integer definition is 0 or 1 plus or minus signs, followed by a series of digits, the phrase can be directly translated to a shell function (Example 3a):

function is_integer {
  [[ $1 = ?([+-])+([0-9]) ]]
}

is_integer $i && echo "yes" || echo "no"

However, if decimal and scientific notation are accepted, Bolsky and Korn provide a more generic solution in their book The New KornShell Command and Programming Language (Example 3c):

isnum()
{
    case $1 in
    ( ?([-+])+([0-9])?(.)*([0-9])?([Ee]?([-+])+([0-9])) )
        return 0;;
    ( ?([-+])*([0-9])?(.)+([0-9])?([Ee]?([-+])+([0-9])) )
        return 0;;
    *)  return 1;;
    esac
}

shopt -s extglob

It happens that the external command expr, which evaluates arguments as an expression, only accepts integers as a series of digits. This is a tool solution (Example 3d):

function is_integer {
  expr "$1" + 0 > /dev/null 2>&1
}

We can use regular expressions used by grep -E (Example 3e):

function is_integer {
  printf "%s\n" $1 | grep -E "^[+-]?[0-9]+$"  >/dev/null
}

function is_integer {
  perl -e "\$_ = \"$1\"; /^[+-]?\d+$/ || exit 1"
}

function is_integer {
  printf "%s\n" $1 | perl -e "<> =~ /^[+-]?\d+$/ || exit 1"
}

The shell provides a more direct pattern-matching solution than external tools. You must either feed the tool's standard input, or expand it to fit the tool's syntax (Example 3f). Using the shell eliminates the extra process.

Problem 4: Add a Directory in the PATH

In this problem, we need to add a directory in the PATH if it's not already in the PATH (path munge). The /etc/profile that comes with a certain Linux distribution contains this function (Example 4a):

pathmunge () {
        if ! echo $PATH | /bin/egrep -q "(^|:)$1($|:)" ; then
           if [ "$2" = "after" ] ; then
              PATH=$PATH:$1
           else
              PATH=$1:$PATH
           fi
        fi
}

shopt -s extglob 2>/dev/null
pathmunge() {
  [[ $PATH = ?(*:)$1?(:*) ]] && return 0
  [[ "$2" = "after" ]] && PATH=$PATH:$1 || PATH=$1:$PATH
}

When running each version of the function 10 times in a ksh93 environment, and using $SECONDS for timing (bash does not support floating-point arithmetic), the results are:

pattern version: 0.016 seconds
egrep   version: 0.542 seconds

#!/bin/sh
exec grep -E ${1+"$@"}

PATH=
pathmunge /bin

You may think that an easier solution is to use /bin/grep -E directly instead of /bin/egrep. While this works on Linux, it produces an error on Solaris:

/bin/grep: illegal option -- E
Usage: grep -hblcnsviw pattern file . . .

In ksh-style functions that support localized variables, you can do this:

function name {
  typeset PATH=$(PATH=/usr/bin:... getconf PATH)
  ...
}

function_name() {
  OPATH=$PATH
  PATH=$(PATH=/usr/bin:... getconf PATH)
  ...
  PATH=$OPATH
}

Problem 4: Discussion

A shell program is more portable because it does not depend on external commands; therefore, we need not worry about the PATH definition. Besides portability, there is a cost associated with spawning a process, especially when the process executes often (as in a loop). While it's not always possible to totally avoid external commands, make an effort to use available shell resources.

Technique: While-Loop

Problem 5: Get infile

An Oracle SQL*Loader control file looks like the following:

-- This is comment line: Oracle Sql*Loader reads records in infile
-- ANNOTATIONS.dat, separated by <er>, and loads them into table
-- ANNOTATIONS.
load data
infile 'ANNOTATIONS.dat' "str '<er>'"
into table ANNOTATIONS
...

cat $file | grep -v -- "--" | grep "infile" | awk "{print \$2}" | \
  sed "s:'::g" | head -1

grep -v -- "--" $file | grep "infile" | awk "{print \$2}" | sed \
  "s:'::g" | head -1

awk -F\' "/^infile/ {print \$2; exit}" $control_file

sed -n "1,/^infile/s/^infile[^']*'\([^']*\)'.*$/\1/p" $control_file

perl -ne "s/^infile[^']*'([^']+)'.*/\$1/ && print && last" $control_file

while read i; do
  [[ $i == infile* ]] && { (IFS="'"; set -- $i; echo $2); break; }
done < $control_file

while read i; do
  [[ $i == infile* ]] && { i=${i#*\'}; echo ${i%%\'*}; break; }
done < $control_file

This example represents a generic problem. You often must read a file line by line, retrieving some data and making changes to the current line (or the previous or next line). When many files need to be processed, we prefer a shell solution. We processed 450 such files using a for loop:

for ctl in *.ctl
do
  ...
done

real        0.190
user        0.112
sys         0.073

real        6.649
user        0.290
sys         1.022

Problem 6: Skip head and tail

In this problem, we want to print a file ($file), but skip the first "$first" lines and last "$last" lines. In these examples, assume that shell variables "file", "first", and "last" are predefined such as file="/etc/passwd", first=2, and last=3.

Many shell programmers will immediately think of the head and tail commands. Yes indeed, head and tail can do the job. Here is a three-tool solution (Example 6a):

total=$(wc -l < $file)
(( head = total - last ))
(( tail = head - first ))
(( tail > 0 )) && head -$head $file | tail -$tail

total=$(wc -l < $file)
(( start = first + 1 ))
(( stop = total - last ))
(( total > first + last )) && sed -n "${start},${stop}p" $file

sed -ne :a -e "1,$((first + last))!{P;N;D;}" -e "${first}d;N;ba" $file

(( a = 0 ))
set --
while IFS= read -r i; do
  (( a++ ))
  (( a <  first )) && set -- "$@" "$i"
  (( a == first )) && set --
  (( a >  first )) && set -- "$@" "$i"
  (( a >  first + last )) && { printf "%s\n" "$1"; shift; }
done < $file

However, the shell has its own idioms, and it does not follow sed's idiosyncrasies. Here is a simplified version (Example 6e):

(( a = 0 ))
set --
while IFS= read -r i; do
  (( a++ ))
  (( a >  first )) && set -- "$@" "$i"
  (( a >  first + last )) && { printf "%s\n" "$1"; shift; }
done < $file

(( a = 0 ))
while IFS= read -r i; do     
  (( a++ ))
  (( b = a % last ))
  (( a > first + last )) && printf "%s\n" "${L[b]}"
  (( a > first )) && L[b]="$i"
done < $file

The advantage of this example is its flexibility. The difference between the tool solution and the shell solution is analogous to the SQL language (the tool) and the PL/SQL environment (the shell). With the tool solution, you generally tell the tool what you want to do, but with the shell solution, you build a step-by-step procedure for a solution. Because the work is done step-by-step, it is easy to add new functionality. For example, you can skip certain lines or calculate line lengths with ease.

Example 6c is the fastest when processing a few larger files, and the shell solutions are the fastest when processing many smaller files. If maximum clarity is your goal, then Examples 6a and 6b are easiest to read.

Technique: Shell Arithmetic

Problem 7: What is Next Month

Given a year and month in YYYYMM format, we need to return the next month in the same format. Here is a shell solution (Example 7a):

function next_month {
  typeset ym=$1 y m
  (( y = ym / 100 ))
  (( m = ym - y*100 ))
  (( y += m / 12 ))
  (( m = m % 12 + 1 ))
  printf "%.4d%.2d\n" $y $m
}

(( m = 200307 - 2003*100 )) calculates the month, which is 7. Since (( 7/12 )) is 0, the year is 2003 + 0, which is still 2003. Only when the current month equals 12 is the year incremented by 1.

The modulo "%" operator delivers the remainder. This is similar to the mod() function in SQL. (( 7 % 12 )) is 7, so next month is 7 + 1, which equals 8. When the current month is 12, (( 12 % 12 )) is 0, so next month is 1. The printf command delivers the results in the required format.

If the math makes you queasy, use any combination of shell techniques for "next_month". Without using any external commands, the function can be rewritten as such (Example 7b):

function next_month {
  typeset y=${1%??} m=${1#????}
  (( m += 1 ))
  (( m == 13 )) && { (( m = 1 )); (( y += 1 )); }
  printf "%.4d%.2d\n" $y $m
}

function next_month {
  typeset ym=$1 y m
  y=$(expr $ym / 100)
  m=$(expr $ym - $y \* 100)
  y=$(expr $y + $m / 12)
  m=$(expr $m % 12 + 1)
  printf "%.4d%.2d\n" $y $m
}

function next_month {
  typeset ym=$1 y m
  y=$(print "$ym / 100" | bc)
  m=$(print "$ym - $y*100" | bc)
  y=$(print "$y + $m / 12" | bc)
  m=$(print "$m % 12 + 1" | bc)
  printf "%.4d%.2d\n" $y $m
}

GNU_DATE=/path/to/GNU/date
function next_month {
  ${GNU_DATE} -d "${1%??}-${1#????}-01 + 1 month" +%Y%m
}

Generally, in the modern shells, there is no reason to use external commands such as expr or bc to perform arithmetic. For complicated date arithmetic, using GNU date or Perl is better than building the logic yourself using shell. Do what you can, but don't hurt yourself.

Problem 8: Sum across Rows

Harry Potter, taking a shell class at Hogwarts School of Witchcraft and Wizardry, posted this problem on comp.unix.shell. He has an input file like the following:

Date,2003-05-01,2003-05-02,2003-05-03,2003-05-04,2003-05-05
item-a,1,1,1,0,3
item-b,2,2,0,0,2
...

awk -F"," '/^Date/ {print $0 ",total"}
           /^item/ {sum=0; for(i=2;i<=NF;i++){sum+=$i};
                    print $0 "," sum}' input.txt

while IFS= read -r; do
  if [[ $REPLY = "Date"* ]]; then
    echo "$REPLY,total"
  elif [[ $REPLY = "item"* ]]; then
    (( sum = 0 ))
    IFS=","; set -- $REPLY; shift
    for i; do (( sum += i )); done
    echo "$REPLY,$sum"
  fi
done < input.txt

while IFS= read -r a; do
  [[ $a = "Date"* ]] && { echo "$a,total"; continue; }
  [[ $a = "item"* ]] && { 
    b=${a#*,}; eval echo "$a,$(( ${b//,/+} ))"
  }
done < input.txt

Problem 8: Discussion

We compared the performance of Example 3c with that of Example 3a on a 300-MHz Intel box running UWIN. We found that ksh93 performs better for smaller data sets, while awk performs better for larger datasets. At 200 records, ksh93 and awk break even (0.25 seconds). Even for 1000 records, the ksh93 program finishes within a second (0.75 seconds) while awk finishes in 0.35 seconds. Having a pure shell solution -- especially inside a shell program -- may outweigh the sub-second loss for moderate-sized files.

For files with millions of records, however, neither shell nor Unix tools are good solutions. You probably need to load the data into a database, and process it with SQL. Use the right tools for the job.

Technique: File and Process Operations

Problem 9: Move All Files Except "old" to "old" Directory

When applied to this problem, the simple command (Example 9a):

mv * old

mv: old: cannot rename to old/old

ls -1 | grep -v "^old$" | xargs -I {} -t mv {} old

mv !(old) old

-ksh: /usr/xpg4/bin/mv: cannot execute [Arg list too long]

for i in !(old); do mv $i old; done

Shell patterns can be used in more sophisticated situations. For example, to uncompress all *.gz files except the following:

FRAMEWORK.dat1.gz
VARHISTORY.dat1.gz
VARHISTORY.dat2.gz
SCENARIOS.dat1.gz
SCENARIOS.dat2.gz
SCENARIOS.dat3.gz

gzip -d !(FRAMEWORK.dat?|VARHISTORY.dat?|SCENARIOS.dat?).gz

Problem 10: Find Files Created Today

In this problem, we want to find all the files in the current directory created today. (For simplicity, assume there are no subdirectories.) For all solutions, we need to create a reference file (Example 10a):

touch -t $(date +%Y%m%d)0000.00 /tmp/timestamp$$

find * -type f -newer /tmp/timestamp$$ -print

ls -1rt * /tmp/timestamp11354 | sed "1,/timestamp11354/d"

for i in *; do [[ $i -nt /tmp/timestamp$$ ]] && echo $i; done

We can use a combination of tool and shell solutions to solve problems in more complicated situations. For example:

find . -type f | while read file; do ...; done

[[ /proc/$PID1 -nt /proc/$PID2 ]] && { echo "$PID1 is started after $PID2"; }

The following example is taken from a magazine published by a commercial database vendor (Example 11a):

#!/bin/ksh
...
v_corruptions="'cat $TMP_FILE | wc -l'"
if [ $v_corruptions -ne 0 ]; then
  echo "Data Block Corruptions Occurred."
fi

[[ -s $TMP_FILE ]] && { echo "Data Block Corruptions Occurred."; }

Problem 11: Discussion

This technique may not make a big difference when the $TMP_FILE file is small, but we think it is important to habitually use more efficient solutions.

Problem 12: Check the Existence of a Process

Quite often, we see shell programs like this (Example 12a):

ps -ef | grep ora_smon_$SID | grep -v grep >/dev/null && {
  echo "Instance $SID is up."
}

ps -u$OWNER -f | grep "[o]ra_smon_$SID"

grep [o]ra_smon_$SID

Use a more strict expression:

ps -u$OWNER -f | grep "[0-9] ora_smon_$SID"

Problem 12: Discussion

Some systems (Solaris, Red Hat Linux) provide a pgrep command that "will never consider itself a potential match", according to the man page. Use pgrep if portability is not a concern.

Problem 13: Shut Down the Database

We want to gracefully shut down the Oracle database (shutdown_immediate function). If this does not shut down the database within, say, five minutes, it kills the shutdown process and issues the forceful shutdown command (shutdown_abort):

shutdown_immediate & (( BPID = $! ))
(( SECONDS = 0 ))
while (( SECONDS <= ${MAXTIME:-300} )); do
  [[ -d /proc/$BPID ]] || break
  sleep 1
done
[[ -d /proc/$BPID ]] && kill -TERM $BPID
shutdown_abort

Additionally, [[ -d /proc/$BPID ]] && kill -TERM $BPID is unlikely to produce an error "kill: $BPID: no such process" than ps -p $BPID && kill -TERM $BPID.

Problem 13: Discussion

In theory, this solution might force a race condition when shutdown_immediate exists, if its PID is reused by another process within a second. In practice, the chance of this happening is nil; however, if it is a concern, check the timestamp and inode number of /proc/$BPID.

Conclusion

Through the problems and solutions presented in this article, we have shown that shell solutions are often more portable, using only the shell, you don't need to worry about various tool availability and behavior. Shell solutions are more efficient -- especially when used within a loop -- and can provide greater flexibility in solving complex problems when no existing tool fits the purpose.

We do not advocate replacing Unix tools with the shell, but encourage you to use tools wisely.

We often see tools overly used when more efficient shell solutions exist. Although it is a path less traveled, the shell offers power and advantages over Unix tools.

Acknowledgements

The comp.unix.shell code snippets from the following individuals are used in our examples: Al Sharka and Dan Mercer (Example 2e), Stephane Chazelas (Example 6c), Charles Demas (Example 8a), Chris F.A. Johnson (Example 8c). We are responsible for any errors in the interpretations and modifications.

References

Bolsky, Morris, David Korn. The New KornShell Command and Programming Language, 1995. Upper Saddle River, NJ: Prentice Hall PTR.

Dougherty, Dale, Arnold Robbins. Sed & Awk, March 1997. Sebastopol, CA: O'Reilly & Associates, Inc.

Wall, Larry, Tom Christiansen, and Jon Orwant. Programming Perl, July 2000. Sebastopol, CA: O'Reilly & Associates, Inc.

Michael Wang earned Master Degrees in Physics (Peking) and Statistics (Columbia). Currently, he is applying his knowledge to Unix systems administration, database management, and programming. He can be reached at: [email protected].

Ed Schaefer is a frequent contributor to Sys Admin. He is a Software Developer and DBA for Intel's Factory Integrated Information Systems, FIIS, in Aloha, OR. Ed also edits the UnixReview.com monthly Shell Corner column. He can be reached at: [email protected].