Is there any way for a child process to set an environment variable visible to its parent?
The answer is both yes and no. Let's get the "no" out of the way first: no matter what you do, there is no way for a child process to affect the value of a variable in its parent. That said, it is possible for a parent process and a child process to cooperate and share whatever you need them to share.
So, if you have a script that says MYVAR="foo" and it calls another script that sets MYVAR="bar", $MYVAR remains "foo" in the parent script. The same idea is true if you change directories in the called script: when it exits, you won't be in the directory it changed to.
There are always ways to get what you need. One of the simplest is to NOT call the other script, but rather to execute it in-line, in the context of the current script. In most Unixy shells, you can use what's called "dot syntax". That is, instead of doing:
you do this:
This way (with the ". otherscript") there won't be another process created. You can think of it like an "include". If "otherscript" changes MYVAR, that last echo will show the change.
A couple of other interesting things: "otherscript" does not need to be executable and if called as shown above, it would have to be in the current directory.
The C shells, /bin/csh, tcsh and others use the ``source'' command:
If the "variable" can be handled with an integer, the exit
status of the child is available to the parent. That's the
usual way of learning why something terminated. So, if your
child process wants to communicate "128" to its parent, you can do
#whatever stuff it does
If not (more data needs to be passed), a shared named pipe
is a little better than a file (and easier to handle
concurrency with) , shared memory is the fastest (but also
more overhead to setup). You can also just setup a
bidirectional pipe between two programs; this can even be
across a network if necessary.
*Exactly* what happens with pipes depends on the shell. But the differences are very minor- a job control shell vs. a non job control shell will assign
process groups differently. Under ordinary circumstances that is so
unimportant to you that I probably shouldn't even mention it.
Pipes block when full. How much data you can write is (for example) examined for bash with "ulimit -p" , which is the size of the pipe in 512 byte blocks and the bash man page plainly tells you that it is a read-only value (you can't make it larger).
However, a named pipe provides communication between two otherwise unrelated proceesses. One doesn't have to call the other, or even know its name, and the one side of the pipe doesn't even have to be running when the other side runs. As a
practical example, I often run across stupid programs that think they
have to print to a device and don't know how to use the spooler. That's
fine until we want to make the printer into a network printer. The
solution is to make a named pipe and tell the dumb program that the pipe
is its print device. You start another process that does this:
lpr -P myrealnetprinter
Then when the program writes to the pipe /dev/myfakenetprint, it
actually goes to the spooler.