Linux pipes and open3Three articles on how open3 is implemented in Linux, and how it is properly used
Open3 (or popen3) is a common name for a library function to spawn a child process with its standard streams (input, output and error) both connected to newly created filehandlers in the original process. You may encounter that function in such languages as Perl, Ruby or Python.
Standard Linux C library, unfortunately, does not contain such a function. Only its less capable counterpart, popen, is a member of C library. The popen function, which is more common, opens only one pipe, either to the input or to the output of the process being spawned. This means that if you program in C you'll have to implement open3 on your own.
Another reason why you would have to re-implement open3, even if your standard (or non-standard) library provides it, is when you need more control over what happens during its call. For instance, Perl and Ruby 1.8 versions of open3 lack the ability to adjust the environment of the process being invoked (see my older rant). This functionality is why I re-implemented open3 in the project I work on.
The choice of Ruby was driven by the fact that I used it in the project I work on, and that it has a killer feature, unhandled exceptions (so that you don't have to check return code of each library call invoked, as in C). Besides, it's a lot less verbose.
The subsequent sections of this article depict sample implementation of open3 in Ruby. However, it relies on the certain system and C library calls I will also outline, so you'll be able to do the same in C or in any other language with access to such primitives.
Foundation for open3
The open3 implementation contains no magic. It relies on two powerful features of Linux system.
Good old forking comes to help in this case. The process spawning in Linux is done via fork call, cloning the current process, followed by exec call, replacing the newly spawned child with the command we wanted to invoke in the first place. It may sound unnecessarily complex, but in our case it has a feature that makes open3 as a library function possible. After fork, but prior to exec we can invoke commands in the context of the new process, while retaining the knowledge we had in the old one. How does it help us?
First, we may alter the environment of the child being invoked, which was my original motivation for reimplementing open3. Many scripting languages have a global, process-wide ENV hash with read/write access, which is used as the environment at execve system calls. If we need to alter the environment in the child process in a thread-safe way, we may modify ENV between fork and exec, so that it will be executed at the context of child process without interfering with the parent.
Second, some filehandlers (internal OS "pointers" to open files, pipes and sockets) preserve across exec! Namely, those that have the flag FD_CLOEXEC cleared. The flag is set by default (I guess, for all except standard 0, 1, and 2 descriptors), but we can change it via fcntl call.
So what we could do is to open pipes in parent process, and inculcate the knowledge about them into the child process before the target command is executed. How should we do the inculcation?
File handler renaming
The child process should spawn with the pipes instead of its STDIN, STDOUT, and STDERR. However, after forking, it already has some standard streams assigned by the OS or the language runtime. What should we do now? The answer is obvious: rename the relevant file handlers.
Closing and opening a file IO object in a high-level language (such as in C++) is not the same as "reopening" it, as the OS filehandler will not be retained. Which is extremely important to altering standard streams.
Language runtime libraries usually have a "reopen" functions (freopen in C, or IO.reopen in Ruby). The thing is that it substitutes the file pointed to by OS file handler, as well as the underlying file of the "language" file handler (this is an integer value in C, or IO object in Ruby or, say, C++).
Changing IO objects intrinsics left to the specific languages, OS filehandler swapping is performed via C library function dup2. A usual dup function just duplicates a filehandler. Called as dup2(keep,replace), it will close(replace), and re-name keep to replace. This renaming is done via fcntl system call, and should be relatively cheap. Additionally, this fcntl call will reset FD_CLOEXEC flag.
Forking and threading in POSIX is an interesting subject itself. I one had a lot of fun, reading this thread, then the FAQ of the newsgroup.
Note that fcntl and dup2 are async-signal-safe function, and therefore it can be called between the fork and exec safely in a multithreaded process.
Okay, let's assemble all the tools together. Opening pipes is easy, as described before, then comes the fork, and renaming the file descriptors.
Instead of just assigning an environment variable, we may allow user to specify the actions they desire to be executed in child's context as a function pointer (or as a block, in Ruby).
Process termination control
From this simple implementation we have completely omitted an important way to control what's happening with the process. Controlling when it's terminated (or forcing it to be).
The thing is that whether the process is terminated does not depend on the status of its standard pipes. First, closing the standard input doesn't necessarily make the process terminate (as with the standard grep command). Actually, very few of them do. So if we've got all the results we wanted from our forked creature, we should have a means to kill it before it gets too late...
Second, the filehandlers we read from do not tell us whether the process is alive either. Remember that filehandlers may survive the exec if FD_CLOEXEC is not set? The process we spawned might, in turn, have given birth to a daemon, which we don't want to mess with. It could inherit the filehandlers, and just not use them. This actually happened with SSHFS mounter: the process wouldn't close the output pipes.
Therefore, pipes won't help us to control the process spawned. What could? Of course, its PID. The fork returns to us the process identifier of the child spawned, and our process, as a parent, has a full access to its status and to its return code. So, along with the pipe filehandlers, we may return child's PID to the caller for it to control the process in a more fine-grained way.
We've learned how to implement open3 itself. However, the bigger troubles are coming when we try to use it.
The thing is that open3 is a powerful tool, but with great power comes great responsibility. As outlined in the previous post, reads and writes to pipes may block at each endpoint, and this renders the usage of open3 error-prone. You should not use it for each process spawn. If a simpler system, fork, or popen would suffice, use them. But if it's not, you're welcome to read the next article on how to actually use open3 in an efficient and correct way.
Proceed to "How to Use Open3 and Avoid Dead Locks".