Skip to content

d11wtq/dockerpty

Repository files navigation

Docker PTY

Provides the functionality needed to operate the pseudo-tty (PTY) allocated to a docker container, using the Python client.

[Build Status] (https://travis-ci.org/d11wtq/dockerpty)

Installation

Via pip:

pip install dockerpty

Dependencies:

  • docker-py>=0.3.2

However, this library does not explicitly declare this dependency in PyPi for a number of reasons. It is assumed you have it installed.

Usage

The following example will run busybox in a docker container and place the user at the shell prompt via Python.

This obviously only works when run in a terminal.

import docker
import dockerpty

client = docker.Client()
container = client.create_container(
    image='busybox:latest',
    stdin_open=True,
    tty=True,
    command='/bin/sh',
)

dockerpty.start(client, container)

Keyword arguments passed to start() will be forwarded onto the client to start the container.

When the dockerpty is started, control is yielded to the container's PTY until the container exits, or the container's PTY is closed.

This is a safe operation and all resources are restored back to their original states.

Note: dockerpty does support attaching to non-tty containers to stream container output, though it is obviously not possible to 'control' the container if you do not allocate a pseudo-tty.

If you press C-p C-q, the container's PTY will be closed, but the container will keep running. In other words, you will have detached from the container and can re-attach with another dockerpty.start() call.

Tests

If you want to hack on dockerpty and send a PR, you'll need to run the tests. In the features/ directory, are features/user stories for how dockerpty is supposed to work. To run them:

-bash$ pip install -r requirements-dev.txt
-bash$ behave features/

You'll need to have docker installed and running locally. The tests use busybox container as a test fixture, so are not too heavy.

Step definitions are defined in features/steps/.

There are also unit tests for the parts of the code that are not inherently dependent on controlling a TTY. To run those:

-bash$ pip install -r requirements-dev.txt
-bash$ py.test tests/

Travis CI runs this build inside a UML kernel that is new enough to run docker. Your PR will need to pass the build before I can merge it.

How it works

In a terminal, the three file descriptors stdin, stdout and stderr are all connected to the controlling terminal (TTY). When you pass the tty=True flag to docker's create_container(), docker allocates a fake TTY inside the container (a PTY) to which the container's stdin, stdout and stderr are all connected.

The docker API provides a way to access the three sockets connected to the PTY. If with access to the host system's TTY file descriptors and the container's PTY file descriptors, it is trivial to simply 'pipe' data written to these file descriptors between the host and the container. Doing this makes the user's terminal effectively become the pseudo-terminal from inside the container.

In reality it's a bit more complicated than this, since care must be taken to put the host terminal into raw mode (where keys such as enter are not interpreted with any special meaning) and restore it on exit. Additionally, the container's stdout and stderr streams along with sys.stdin must be made non-blocking so that they can be used with select() without blocking the main process. These attributes are restored on exit.

The size of a terminal cannot be controlled by sending data to stdin and can only be controlled by the terminal program itself. Since the pseudo-terminal is running inside a real terminal, it is import that the size of the PTY be kept the same as that of the presenting TTY. For this reason, docker provides an API call to resize the allocated PTY. A SIGWINCH handler is used to detect window size changes and resize the pseudo-terminal as needed.

Contributors

Copyright & Licensing

Copyright © 2014 Chris Corbyn. See the LICENSE.txt file for details.