Using Processes

Along with connection to servers across the internet, Twisted also connects to local processes with much the same API. The API is described in more detail in the documentation of:

Processes are run through the reactor, using reactor.spawnProcess. Pipes are created to the child process, and added to the reactor core so that the application will not block while sending data into or pulling data out of the new process. reactor.spawnProcess requires two arguments, processProtocol and executable, and optionally takes several more: args, environment, path, userID, groupID, usePTY, and childFDs. Not all of these are available on Windows.

1
2
3
4
5
6
from twisted.internet import reactor

processProtocol = MyProcessProtocol()
reactor.spawnProcess(processProtocol, executable, args=[program, arg1, arg2],
                     env={'HOME': os.environ['HOME']}, path,
                     uid, gid, usePTY, childFDs)

The ProcessProtocol you pass to spawnProcess is your interaction with the process. It has a very similar signature to a regular Protocol, but it has several extra methods to deal with events specific to a process. In our example, we will interface with 'wc' to create a word count of user-given text. First, we'll start by importing the required modules, and writing the initialization for our ProcessProtocol.

1
2
3
4
5
from twisted.internet import protocol
class WCProcessProtocol(protocol.ProcessProtocol):

    def __init__(self, text):
        self.text = text

When the ProcessProtocol is connected to the protocol, it has the connectionMade method called. In our protocol, we will write our text to the standard input of our process and then close standard input, to let the process know we are done writing to it.

1
2
3
4
...
    def connectionMade(self):
        self.transport.write(self.text)
        self.transport.closeStdin()

At this point, the process has receieved the data, and it's time for us to read the results. Instead of being received in dataReceived, data from standard output is received in outReceived. This is to distinguish it from data on standard error.

1
2
3
4
5
6
7
8
...
    def outReceived(self, data):
        fieldLength = len(data) / 3
        lines = int(data[:fieldLength])
        words = int(data[fieldLength:fieldLength*2])
        chars = int(data[fieldLength*2:])
        self.transport.loseConnection()
        self.receiveCounts(lines, words, chars)

Now, the process has parsed the output, and ended the connection to the process. Then it sends the results on to the final method, receiveCounts. This is for users of the class to override, so as to do other things with the data. For our demonstration, we will just print the results.

1
2
3
4
5
6
...
    def receiveCounts(self, lines, words, chars):
        print 'Received counts from wc.'
        print 'Lines:', lines
        print 'Words:', words
        print 'Characters:', chars

We're done! To use our WCProcessProtocol, we create an instance, and pass it to spawnProcess.

1
2
3
4
from twisted.internet import reactor
wcProcess = WCProcessProtocol("accessing protocols through Twisted is fun!\n")
reactor.spawnProcess(wcProcess, 'wc', ['wc'])
reactor.run()

These are the methods that you can usefully override in your subclass of ProcessProtocol:

The following are the basic ways to control the child process:

• self.transport.write(data): Stuff some data in the stdin pipe. Note that this write method will queue any data that can't be written immediately. Writing will resume in the future when the pipe becomes writable again.
• self.transport.closeStdin: Close the stdin pipe. Programs which act as filters (reading from stdin, modifying the data, writing to stdout) usually take this as a sign that they should finish their job and terminate. For these programs, it is important to close stdin when you're done with it, otherwise the child process will never quit.
• self.transport.closeStdout: Not usually called, since you're putting the process into a state where any attempt to write to stdout will cause a SIGPIPE error. This isn't a nice thing to do to the poor process.
• self.transport.closeStderr: Not usually called, same reason as closeStdout.
• self.transport.loseConnection: Close all three pipes.
• self.transport.signalProcess('KILL'): Kill the child process. This will eventually result in processEnded being called.

Here is an example that is rather verbose about exactly when all the methods are called. It writes a number of lines into the wc program and then parses the output.

stdin, stdout, and stderr are just conventions. Programs which operate as filters generally accept input on fd0, write their output on fd1, and emit error messages on fd2. This is common enough that the standard C library provides macros like stdin to mean fd0, and shells interpret the pipe character | to mean redirect fd1 from one command into fd0 of the next command.

But these are just conventions, and programs are free to use additional file descriptors or even ignore the standard three entirely. The childFDs argument allows you to specify exactly what kind of files descriptors the child process should be given.

Each child FD can be put into one of three states:

• Mapped to a parent FD: this causes the child's reads and writes to come from or go to the same source/destination as the parent.
• Feeding into a pipe which can be read by the parent.
• Feeding from a pipe which the parent writes into.

Mapping the child FDs to the parent's is very commonly used to send the child's stderr output to the same place as the parent's. When you run a program from the shell, it will typically leave fds 0, 1, and 2 mapped to the shell's 0, 1, and 2, allowing you to see the child program's output on the same terminal you used to launch the child. Likewise, inetd will typically map both stdin and stdout to the network socket, and may map stderr to the same socket or to some kind of logging mechanism. This allows the child program to be implemented with no knowledge of the network: it merely speaks its protocol by doing reads on fd0 and writes on fd1.

Feeding into a parent's read pipe is used to gather output from the child, and is by far the most common way of interacting with child processes.

Feeding from a parent's write pipe allows the parent to control the child. Programs like bc or ftp can be controlled this way, by writing commands into their stdin stream.

The childFDs dictionary maps file descriptor numbers (as will be seen by the child process) to one of these three states. To map the fd to one of the parent's fds, simply provide the fd number as the value. To map it to a read pipe, use the string r as the value. To map it to a write pipe, use the string w.

For example, the default mapping sets up the standard stdin/stdout/stderr pipes. It is implemented with the following dictionary:

1
childFDs = { 0: "w", 1: "r", 2: "r" }

To launch a process which reads and writes to the same places that the parent python program does, use this:

1
childFDs = { 0: 0, 1: 1, 2: 2}

To write into an additional fd (say it is fd number 4), use this:

1
childFDs = { 0: "w", 1: "r", 2: "r" , 4: "w"}

When you provide a childFDs dictionary with more than the normal three fds, you need addtional methods to access those pipes. These methods are more generalized than the .outReceived ones described above. In fact, those methods (outReceived and errReceived) are actually just wrappers left in for compatibility with older code, written before this generalized fd mapping was implemented. The new list of things that can happen to your ProcessProtocol is as follows:

• .connectionMade: This is called when the program is started.
• .childDataReceived(childFD, data): This is called with data that was received from one of the process' output pipes (i.e. where the childFDs value was r. The actual file number (from the point of view of the child process) is in childFD. For compatibility, the default implementation of .childDataReceived dispatches to .outReceived or .errReceived when childFD is 1 or 2.
• .childConnectionLost(childFD): This is called when the reactor notices that one of the process' pipes has been closed. This either means you have just closed down the parent's end of the pipe (with .transport.closeChildFD), the child closed the pipe explicitly (sometimes to indicate EOF), or the child process has terminated and the kernel has closed all of its pipes. The childFD argument tells you which pipe was closed. Note that you can only find out about file descriptors which were mapped to pipes: when they are mapped to existing fds the parent has no way to notice when they've been closed. For compatibility, the default implementation dispatches to .inConnectionLost, .outConnectionLost, or .errConnectionLost.
• .processEnded(status): This is called when the child process has been reaped, and all pipes have been closed. This insures that all data written by the child prior to its death will be received before .processEnded is invoked.

In addition to those methods, there are other methods available to influence the child process:

• self.transport.writeToChild(childFD, data): Stuff some data into an input pipe. .write simply writes to childFD=0.
• self.transport.closeChildFD(childFD): Close one of the child's pipes. Closing an input pipe is a common way to indicate EOF to the child process. Closing an output pipe is neither very friendly nor very useful.

GnuPG, the encryption program, can use additional file descriptors to accept a passphrase and emit status output. These are distinct from stdin (used to accept the crypttext), stdout (used to emit the plaintext), and stderr (used to emit human-readable status/warning messages). The passphrase FD reads until the pipe is closed and uses the resulting string to unlock the secret key that performs the actual decryption. The status FD emits machine-parseable status messages to indicate the validity of the signature, which key the message was encrypted to, etc.

gpg accepts command-line arguments to specify what these fds are, and then assumes that they have been opened by the parent before the gpg process is started. It simply performs reads and writes to these fd numbers.

To invoke gpg in decryption/verification mode, you would do something like the following:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
class GPGProtocol(ProcessProtocol):
    def __init__(self, crypttext):
        self.crypttext = crypttext
        self.plaintext = ""
        self.status = ""
    def connectionMade(self):
        self.transport.writeToChild(3, self.passphrase)
        self.transport.closeChildFD(3)
        self.transport.writeToChild(0, self.crypttext)
        self.transport.closeChildFD(0)
    def childDataReceived(self, childFD, data):
        if childFD == 1: self.plaintext += data
        if childFD == 4: self.status += data
    def processEnded(self, status):
        rc = status.value.exitCode
        if rc == 0:
            self.deferred.callback(self)
        else:
            self.deferred.errback(rc)

def decrypt(crypttext):
    gp = GPGProtocol(crypttext)
    gp.deferred = Deferred()
    cmd = ["gpg", "--decrypt", "--passphrase-fd", "3", "--status-fd", "4",
           "--batch"]
    p = reactor.spawnProcess(gp, cmd[0], cmd, env=None,
                             childFDs={0:"w", 1:"r", 2:2, 3:"w", 4:"r"})
    return gp.deferred

In this example, the status output could be parsed after the fact. It could, of course, be parsed on the fly, as it is a simple line-oriented protocol. Methods from LineReceiver could be mixed in to make this parsing more convenient.

The stderr mapping (2:2) used will cause any GPG errors to be emitted by the parent program, just as if those errors had caused in the parent itself. This is sometimes desireable (it roughly corresponds to letting exceptions propagate upwards), especially if you do not expect to encounter errors in the child process and want them to be more visible to the end user. The alternative is to map stderr to a read-pipe and handle any such output from within the ProcessProtocol (roughly corresponding to catching the exception locally).