6.2 KiB
Process substitution
![](keywords>bash shell scripting expansion substitution text stdin stdout save capture)
Process substitution is a form of redirection where the input or output of a process (some sequence of commands) appear as a temporary file.
<( <LIST> )
>( <LIST> )
Process substitution is performed simultaneously with parameter expansion, command substitution and arithmetic expansion.
The command list <LIST>
is executed and
its
- standard output filedescriptor in the
<( ... )
form or - standard input filedescriptor in the
>( ... )
form
is connected to a FIFO or a file in /dev/fd/
. The filename (where the
filedescriptor is connected) is then used as a substitution for the
<(...)
-construct.
That, for example, allows to give data to a command that can't be
reached by pipelining (that doesn't expect its data from stdin
but
from a file).
Scope
<note important> Note: According to multiple comments and sources, the scope of process substitution file descriptors is not stable, guaranteed, or specified by bash. Newer versions of bash (5.0+) seem to have shorter scope, and substitutions scope seems to be shorter than function scope. See stackexchange and stackoverflow; the latter discussion contains a script that can test the scoping behavior case-by-case </note>
If a process substitution is expanded as an argument to a function, expanded to an environment variable during calling of a function, or expanded to any assignment within a function, the process substitution will be "held open" for use by any command within the function or its callees, until the function in which it was set returns. If the same variable is set again within a callee, unless the new variable is local, the previous process substitution is closed and will be unavailable to the caller when the callee returns.
In essence, process substitutions expanded to variables within functions remain open until the function in which the process substitution occured returns - even when assigned to locals that were set by a function's caller. Dynamic scope doesn't protect them from closing.
Examples
This code is useless, but it demonstrates how it works:
$ echo <(ls)
/dev/fd/63
The output of the ls
-program can then be accessed by reading the
file /dev/fd/63
.
Consider the following:
diff <(ls "$first_directory") <(ls "$second_directory")
This will compare the contents of each directory. In this command, each process is substituted for a file, and diff doesn't see <(bla), it sees two files, so the effective command is something like
diff /dev/fd/63 /dev/fd/64
where those files are written to and destroyed automatically.
Avoiding subshells
<WRAP center round info 60%> See Also: BashFAQ/024 -- I set variables in a loop that's in a pipeline. Why do they disappear after the loop terminates? Or, why can't I pipe data to read? </WRAP>
One of the most common uses for process substitutions is to avoid the
final subshell that results from executing a pipeline. The following is
a wrong piece of code to count all files in /etc
is:
counter=0
find /etc -print0 | while IFS= read -rd '' _; do
((counter++))
done
echo "$counter files" # prints "0 files"
Due to the pipe, the while read; do ... done
part is executed in a
subshell (in Bash, by default), which means counter
is only
incremented within the subshell. When the pipeline finishes, the
subshell is terminated, and the counter
visible to echo
is still at
"0"!
Process substitution helps us avoid the pipe operator (the reason for the subshell):
counter=0
while IFS= read -rN1 _; do
((counter++))
done < <(find /etc -printf ' ')
echo "$counter files"
This is the normal input file redirection < FILE
, just that the FILE
in this case is the result of process substitution. It's important to
note that the space is required in order to disambiguate the syntax from
here documents.
: < <(COMMAND) # Good.
: <<(...) # Wrong. Will be parsed as a heredoc. Bash fails when it comes across the unquoted metacharacter ''(''
: ><(...) # Technically valid but pointless syntax. Bash opens the pipe for writing, while the commands within the process substitution have their stdout connected to the pipe.
Process substitution assigned to a parameter
This example demonstrates how process substitutions can be made to
resemble "passable" objects. This results in converting the output of
f
's argument to uppercase.
f() {
cat "$1" >"$x"
}
x=>(tr '[:lower:]' '[:upper:]') f <(echo 'hi there')
See the above section on #scope
Bugs and Portability Considerations
- Process substitution is not specified by POSIX.
- Process substitution is disabled completely in Bash POSIX mode.
- Process substitution is implemented by Bash, Zsh, Ksh{88,93}, but not (yet) pdksh derivatives (mksh). Coprocesses may be used instead.
- Process substitution is supported only on systems that support
either named pipes (FIFO - a special
file) or the
/dev/fd/*
method for accessing open files. If the system doesn't support/dev/fd/*
, Bash falls back to creating named pipes. Note that not all shells that support process substitution have that fallback. - Bash evaluates process substitutions within array indices, but not other arithmetic contexts. Ksh and Zsh do not. (Possible Bug)
# print "moo"
dev=fd=1 _[1<(echo moo >&2)]=
# fork bomb
${dev[${dev='dev[1>(${dev[dev]})]'}]}
- Issues with wait, race conditions, etc: https://groups.google.com/forum/?fromgroups=#!topic/comp.unix.shell/GqLNzUA4ulA
See also
- Internal: Introduction to expansion and substitution
- Internal: Bash in the process tree (subshells)
- Internal: Redirection