How to on starting processes (mostly in Linux)


Do you want to run an executable file from your program? Or execute a shell command programmatically? Or maybe just parallelize your code? Have you read a lot of information regarding execve()) functions family and fork()) but still have a mess in your head? Then this article is for you.


System calls

Let's keep it simple and start from the beginning. We are developing a program for Linux. Let's have a look on so called system calls - the interface Linux provides us to request kernel functionalities.

Linux has the next system calls to work with processes:

  • fork(void) (man 2 fork) - creates a full copy of the calling process. Sounds ineffecient because of need of copying the enterie process's address space, but uses copy-on-write optimization. This is the only (ideological) way to create a process in Linux. However, in fresh versions of the kernel fork() is implemented on top of tricky clone() system call and now it's possible to use clone() directly to create processes, but for simplicity we are going to skip these details.
  • execve(path, args, env) (man 2 execve) - transforms the calling process into a new process by executing a file under the specified path. In effect, it replaces the current process image with a new process image and doesn't create any new processes.
  • pipe(fildes[2] __OUT) (man 2 pipe) - creates a pipe which is an inter-process communication primitive. Usually pipes are unidirectional data flows. The first element of the array connects to the read end of the pipe, and the second element connects to the write end. The data written to fildes[1] can be read from the fildes[0].

We are not going to have a look at the aforementioned system calls source code because it's a part of the kernel and could be hardly understandable.

Also an important part of our consideration is the Linux shell - a command interpreter utility (i.e. regular program). The shell process constantly reads from the stdin. A user usually interacts with the shell by typing some commands and pressing enter key. The shell process then executes provided commands. Standard outputs of these processes are connected to the stdout of the shell process. However, the shell process can be launched as a subprocess by itself and the command to execute can be specified via -c argument. Eg. bash -c "date".

C standard library

Of course we are developing our program in C to be as close to the OS-level primitives as possible. C has a so called standard library libc - a broad set of functions to simplify writing programs in this language. Also it provides wrapping around syscalls.

C standard library has the next functions (on Debian-based distros apt-get download glibc-source):

  • system(command) (man 3 system) - launches a shell process to execute provided command. The calling process is blocked till the end of the execution of the underlying shell process. system() returns an exit code of the shell process. Let's have a look on the implementation of this function in the stdlib:

    int system(char *command)
     // ... skip signals tricks for simplicity ...
     switch(pid = vfork()) {
         case -1:            // error
            // ...
         case 0:             // child
             execl("/bin/sh", "sh", "-c", command, (char *)NULL);
             _exit(127);  // will be called only if execl() returns, i.e. a syscall faield.
     // ... skip signals tricks for simplicity ...
     waitpid(pid, (int *)&pstat, 0);  // waiting for the child process, i.e. shell.
     return pstat.w_status;

    So in effect, system() just uses the combination of the fork() + exec() + waitpid().

  • popen(command, mode = 'r|w') (man 3 popen) - forks and replaces the forked process with a shell instance executing provided command. Sounds pretty the same like the system()? The difference is an abilitiy to communicate with the child process via its stdin or stdout. But usually in the unidirectional way. To communicate with this process a pipe is used. Real implementations can be found here and here but the main idea is the following:

    FILE * popen(char *program, char *type)
      int pdes[2], fds, pid;
      pipe(pdes);  // create a pipe
      switch (pid = vfork()) { // fork the current process
      case -1:            // error
          // ...
      case 0:             // child
          if (*type == 'r') {
              dup2(pdes[1], fileno(stdout));  // bind stdout of the child process to the writing end of the pipe
              close(pdes[0]);                 // close reading end of the pipe on the child side
          } else {
              dup2(pdes[0], fileno(stdin));  // bind stdin of the child process to the reading end of the pipe
              close(pdes[1]);                // close writing end of the pipe on the child side
          execl("/bin/sh", "sh", "-c", program, NULL);  // replace the child process with the shell running our command
          _exit(127);  // will be called only if execl() returns, i.e. a syscall faield.
      // parent
      if (*type == 'r') {
          result = pdes[0];
      } else {
          result = pdes[1];
      return result;

Congratulations, that is!

NB1: shell implementation of the subprocess launching is pretty the same. i.e. fork() + execve().

NB2: it's good to mention that other programming languages usually implement bindings to the OS's libc (and do some wrappings for convenience) to provide OS-specific funtionality.

Use cases

Parallelize execution

The simplest one. We need only fork(). Call of the fork() in effect duplicates your program process. But since this process uses completely separate address space to communicate with it we anyway need inter-process communication primitives. Even the instructions set of the forked process is the same as the parent's one, it's a different instance of the program.

Just run a program from your code

If you need just to run a program, without communicating with its stdin/stdout the libc system() function is the simplest solution. Yep, you also can fork() your process and then run exec() in the child process, but since it's a quite common scenario there is system() function.

Run a process and read its stdout (or write to its stdin)

We need popen() libc function. Yep, you still can achieve the goal just by combining pipe() + fork() + exec() as shown above, but popen() is here to reduce the amount of boilerplate code.

Run a process, write to its stdin and read from its stdout

The most interesting one. For some reasons default popen() implementation is usually unidirectional. But looks like we can easily come up with the bidirectional solution: we need two pipes, first will be attached to child's stdin and the second one to the child's stdout. The remaining part is to fork() a child process, connect pipes via dup2() to IO descriptors and execve() the command. One of the potential implementations can be found on my GitHub popen2() project. An extra thing you should be aware while developing such function is leaking of open file descriptors of pipes from previously opened via popen() processes. If we forget to close explicitly foreign file descriptors in each child fork, there will be a possibility to do IO operations with the siblings' stdins and stdouts. Sounds like a vulnerability. To be able to close all those file descriptors we have to track them. I used a static variable with the linked list of such descriptors:

static files_chain_t *files_chain;

file_t *popen2(const char *command) {
    file_t *fp = malloc();  // allocate new element of the chain

    _do_popen(fp, command);

    // add the current result to the chain
    fp->next = files_chain;
    files_chain = fp;

_do_popen() {
    // open pipes
    // fork()
    // if is_child:
    //     for (fp in files_chain):
    //         close(fp->in); close(fp->out);

int pclose2(file_t *fp) {
    // if (fp in files_chain):
    //     ... do payload ...
    //     remove fp from the chain


A few words about Windows

Windows OS family have a slightly different paradigm for working with processes. If we skip neoteric Unix compatibility layer introduced on Windows 10 and attempts to port POSIX API support for Windows we will have only two functions from the oldschool WinAPI:

So, no forks and execves. However, to communicate with the started processes pipes also can be used.

Instead of conclusions

Make code not war!