Like, I suspect, many programmers, there are many software tools I have used on a regular basis for many years but remain woefully ignorant of their inner workings and true potential. One of these is the Unix command line.

It’s a common, for example, to want to make the error output of a program to appear as the normal output and trial and error, or Googling or looking at Stack Overflow leads to:

strace ls 2>&1 >/dev/null

which works fine but seems puzzling – we’ve told the program to send error output to normal output, then normal output to /dev/null so why doesn’t that discard everything, similar to:

strace ls >/dev/null 2>&1

This is because we don’t understand what is going on.

An indirection is actually a call to the dup2 system call. From the man page:

dup2() makes newfd be the copy of oldfd, closing newfd first if necessary'

So: n>&m does a dup2(m,n): close fd n if necessary, then make n be a copy of fd m, and n>file means: close n if necessary, open file as fd m, then do dup2(m,n).

Now it all makes sense:

strace ls 2>&1 1>/dev/null 

first of all makes 2 be a copy of 1, then changes 1 to point to /dev/null – the copying is done ‘by value’ as it were (despite the confusing, for C++ programmers anyway, use of ‘&’).

Using strace here is not an accident, but used like this doesn’t tell us much: indirection is handled by the shell, not by the program, so we need to do something like this for further insight:

$ strace -f -etrace=clone,execve,open,dup2 bash -c 'ls >/dev/null 2>&1'
execve("/bin/bash", ["bash", "-c", "ls >/dev/null 2>&1"], [/* 46 vars */]) = 0
clone(Process 20454 attached
child_stack=0, flags=CLONE_CHILD_CLEARTID|CLONE_CHILD_SETTID|SIGCHLD, child_tidptr=0x7f643f4c09d0) = 20454
Process 20453 suspended
[pid 20454] open("/dev/null", O_WRONLY|O_CREAT|O_TRUNC, 0666) = 3
[pid 20454] dup2(3, 1)                  = 1
[pid 20454] dup2(1, 2)                  = 2
[pid 20454] execve("/bin/ls", ["ls"], [/* 45 vars */]) = 0
Process 20453 resumed
Process 20454 detached
--- SIGCHLD (Child exited) @ 0 (0) ---

bash forks off a subprocess (a clone syscall these days rather than fork), which then sets up the input and output before calling execve to actually run the command.

We don’t have to limit ourselves to fds 0,1 and 2 that we get to start with:

$ runit () { echo stderr 1>&2; echo stdout; }
$ runit 1>/dev/null
$ runit 2>/dev/null
$ runit 3>&2 2>&1 1>&3
$ (runit 3>&2 2>&1 1>&3) 1> /dev/null
$ (runit 3>&2 2>&1 1>&3) 2> /dev/null

We duplicate 2 to 3, then 1 to 2, then 3 to 1, and we have swapped stdin and stderr.

We can also pass non-standard file descriptors in to programs, though this doesn’t seem to be a technique used much:

#include <unistd.h>
int main()
  char buffer[256];
  ssize_t n;
  while ((n = read(3,buffer,sizeof(buffer))) > 0) {

and do:

$ g++ -Wall cat34.cpp -o cat34
$ echo Hello World | ./cat34 3<&0 4>&1
Hello World

It’s interesting that this also works:

$ echo Hello World | ./cat34 3>&0 4<&1
Hello World

While we are this part of town, let’s talk briefly about named pipes, another feature that has been around for ever, but doesn’t seem to get used as much as it deserves. We can run in to problems though:

Suppose I want to capture an HTTP response from a web server, I can do this:

$ mkfifo f1
$ mkfifo f2
$ mkfifo f3
$ netcat -l 8888 <f1 >f2 &
$ netcat 80 <f2 >f3 &
$ tee foo.txt <f3 >f1 &

and try a download, but alas:

$ GET http://localhost:8888/ >/dev/null
Can't connect to localhost:8888 (Connection refused)

This doesn’t seem right, netcat should be listening on port 8888, I told it so myself! And checking with netstat shows no listener on 8888 and finally ps -aux shows no sign of any netcat processes – what is going on?

Once again, strace helps us see the true reality of things:

$ kill %1
$ strace netcat -l 8888 < f1 > f2

But strace tells us nothing – there is no output! Like the dog that didn’t bark in the night though, this is an important clue, and widening our area of investigation, we find:

$ strace -f -etrace=open,execve bash -c 'netcat -l 8888 < f1 > f2'
execve("/bin/bash", ["bash", "-c", "netcat -l 8888 < f1 > f2"], [/* 46 vars */]) = 0
Process 20621 attached
Process 20620 suspended
[pid 20621] open("f1", O_RDONLY ...

The shell is stalled trying to open the “f1” fifo, before it even gets around to starting the netcat program, which is why the first strace didn’t show anything. What we have forgotten is that opening a pipe blocks if there is no process with the other end open (it doesn’t have to actively reading or writing, it just has to be there). The shell handles redirections in the order they appear, so since our 3 processes are all opening their read fifo first, none have got around to opening their write fifo – we have deadlock, in fact, the classic dining philosophers problem, and a simple solution is for one philosopher to pick up the forks in a different order:

$ netcat -l 8888 >f2 <f1 &
$ netcat 80 <f2 >f3 &
$ tee foo.txt <f3 >f1 &
$ GET http://localhost:8888/ >/dev/null
$ cat foo.txt
HTTP/1.1 301 Moved Permanently
Server: Apache

We can, it should be noted, do this more easily with a normal pipeline and a single fifo, and avoid all these problems:

$ netcat 80 <f1 | tee foo.txt | netcat -l 8888 >f1

but that would be less fun and possibly less instructive.

Embedded Python Interpreter

And now for something completely different…

Often, I’d like to embed a reasonably capable command interpreter in a C++ application. Python seems a likely candidate, so here’s some investigative code using separate processes (the next step will be to use threads, if that’s possible, so the interpreter can live in the same memory space as our application, that can wait for part II though). As well as the mechanics of embedding Python, we have a pleasant excursion through the sometimes murky worlds of signal handling and pseudo-terminals.

The server structure is conventional (though not necessarily suitable for a serious production server), on each incoming connection we fork a handler process, this in turn splits into two processes, which form their own process group under the control of a pseudo-terminal (pty). One forwarding process copies data between the socket and the master side of the pty, the other process runs the interpreter itself on the slave side. Simple enough, with a few subtleties. To get signal handling right, we have to ignore SIGINT in the forwarding process (otherwise it will terminate on interrupt, taking the interpreter with it), but leave the default handler in the interpreter process – Python sets up its own signal handler, but it only seems to do this if the handler hasn’t been redefined already. Also, Python seems to insist that it uses fds 0,1 and 2 so we need to rebind them, and, finally, to get Python to do line editing, we need to import readline in the interpreter.

My main interest here is in getting external access to the interpreter, rather than the mechanics of calling between C and Python, so we just have a couple of simple functions init() and func() defined in the embedded interpreter as examples. At this simple level I don’t think we need to worry about reference counts etc.

#include <Python.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <fcntl.h>
#include <signal.h>
#include <time.h>
#include <errno.h>
#include <netinet/ip.h>
#include <sys/epoll.h>

// Some handy macros to help with error checking
// When prototyping, it's a good idea to check every
// system call for errors, these macros help to keep
// the code uncluttered.

#define CHECK(e) \
 ((e)? \
  (void)0: \
  (fprintf(stderr, "'%s' failed at %s:%d\n - %s\n", \
           #e, __FILE__, __LINE__,strerror(errno)), \

#define CHECKSYS(e) (CHECK((e)==0))
#define CHECKFD(e) (CHECK((e)>=0))

// We are told not to use signal, due to portability problems
// so we will define a similar function ourselves with sigaction
void setsignal(int signal, sighandler_t handler)
  struct sigaction sa;
  sa.sa_handler = handler;

// Make a suitable server socket, as a small concession to
// security, we will hardwire the loopback address as the
// bind address. People elsewhere can come in through an SSH
// tunnel.
int makeserversock(int port)
  int serversock = socket(AF_INET,SOCK_STREAM,0);
  sockaddr_in saddr;
  saddr.sin_family = PF_INET;
  saddr.sin_port = htons(port);
  saddr.sin_addr.s_addr = htonl(INADDR_LOOPBACK);

  int optval = 1;
  CHECKSYS(setsockopt(serversock, SOL_SOCKET, SO_REUSEADDR, 
                      &optval, sizeof optval));
  return serversock;

// Copy data between our socket fd and the master
// side of the pty. A simple epoll loop.
int runforwarder(int mpty, int sockfd)
  static const int MAX_EVENTS = 10;
  int epollfd = epoll_create(MAX_EVENTS);
  epoll_event event;
  memset (&event, 0, sizeof(event)); = EPOLLIN; = sockfd;
  CHECKSYS(epoll_ctl(epollfd, EPOLL_CTL_ADD, sockfd, &event)); = mpty;
  CHECKSYS(epoll_ctl(epollfd, EPOLL_CTL_ADD, mpty, &event));
  char ibuff[256];
  while (true) {
    struct epoll_event events[MAX_EVENTS];
    int nfds = epoll_wait(epollfd, events, MAX_EVENTS, -1);
    // Maybe treat EINTR specially here.
    CHECK(nfds >= 0);
    for (int i = 0; i < nfds; ++i) {
      int fd = events[i].data.fd;
      if (events[i].events & EPOLLIN) {
        ssize_t nread = read(fd,ibuff,sizeof(ibuff));
        CHECK(nread >= 0);
        if (nread == 0) {
          goto finish;
        } else {
      } else if (events[i].events & (EPOLLERR|EPOLLHUP)) {
        goto finish;
      } else {
        fprintf(stderr, "Unexpected event for %d: 0x%x\n", 
                fd, events[i].events);
        goto finish;
  return 0;

// The "application" functions to be accessible from
// the embedded interpreter
int myinit()
  return 0;

int myfunc()
  return rand();

// Python wrappers around our application functions
static PyObject*
emb_init(PyObject *self, PyObject *args)
    if (!PyArg_ParseTuple(args, ":init")) return NULL;
    return Py_BuildValue("i", myinit());

static PyObject*
emb_func(PyObject *self, PyObject *args)
    if (!PyArg_ParseTuple(args, ":func")) return NULL;
    return Py_BuildValue("i", myfunc());

static PyMethodDef EmbMethods[] = {
    {"init", emb_init, METH_VARARGS,
     "(Re)initialize the application."},
    {"func", emb_func, METH_VARARGS,
     "Run the application"},
    {NULL, NULL, 0, NULL}

int runinterpreter(char *argname, int fd)

  Py_InitModule("emb", EmbMethods);
  PyRun_SimpleString("from time import time,ctime\n");
  PyRun_SimpleString("from emb import init,func\n");
  PyRun_SimpleString("print('Today is',ctime(time()))\n");
  PyRun_SimpleString("import readline\n");
  PyRun_InteractiveLoop(stdin, "-");

  return 0;

int main(int argc, char *argv[])
  int port = -1;
  if (argc > 1) {
    port = atoi(argv[1]);
  } else {
    fprintf(stderr, "Usage: %s <port>\n", argv[0]);
  setsignal(SIGCHLD, SIG_IGN);
  int serversock = makeserversock(port);
  while (true) {
    int sockfd = accept(serversock,NULL,NULL);
    if (fork() != 0) {
      // Server side, close new connection and continue
    } else {
      // Client side, close server socket
      CHECKSYS(close(serversock)); serversock = -1;
       // Create a pseudo-terminal
      int mpty = posix_openpt(O_RDWR);
      CHECKSYS(grantpt(mpty)); // pty magic
      // Start our own session
      int spty = open(ptsname(mpty),O_RDWR);
      // spty is now our controlling terminal
      // Now split into two processes, one copying data
      // between socket and pty; the other running the
      // actual interpreter.
      if (fork() != 0) {
        // Ignore sigint here
        setsignal(SIGINT, SIG_IGN);
        return runforwarder(sockfd,mpty);
      } else {
        // Default sigint here - will be replace by interpreter
        setsignal(SIGINT, SIG_DFL);
        return runinterpreter(argv[0],spty);

Compilation needs something like:

g++ -g -L/usr/lib/python2.6/config -lpython2.6 -I/usr/include/python2.6 -Wall embed.cpp -o embed

Suitable flags can be obtained by doing:

	/usr/bin/python2.6-config --cflags
	/usr/bin/python2.6-config --ldflags

Of course, all this will depend on your exact Python version and where it is installed. Embedding has changed somewhat in Python 3, but most of this will still apply.

To connect to the interpreter, we can use our good friend netcat, with some extra tty mangling (we want eg. control-C to be handled by the pty defined above in the server code, not the user terminal, so we put that into raw mode).

ttystate=`stty --save`
stty raw -echo
netcat $*
stty $ttystate

We set up the server socket to only listen on the loopback interface, so in order to have secure remote access, we can set up an SSH tunnel by running something like:

$ ssh -N -L 9998:localhost:9999 <serverhost>

on the client host.

Finally, we can run some Python:

$ connect localhost 9998
('Today is', 'Sun Nov  4 21:09:09 2012')
>>> print 1
>>> init()
>>> func()
>>> ^C
>>> ^D