Overview

Features

• Uses unix domain sockets, only root or same effective user can connect.

• Can run the connection in a thread or in a signal handler (see oneshot_on option).

• Can start the thread listening for connections from a signal handler (see activate_on option)

• Compatible with apps that fork, reinstalls the Manhole thread after fork - had to monkeypatch os.fork/os.forkpty for this.

• Compatible with gevent and eventlet with some limitations - you need to either:

  • Use oneshot_on, or
  • Disable thread monkeypatching (eg: gevent.monkey.patch_all(thread=False), eventlet.monkey_patch(thread=False)

  Note: on eventlet you might need to setup the hub first to prevent circular import problems:

  import eventlet
  eventlet.hubs.get_hub()  # do this first
  eventlet.monkey_patch(thread=False)
  

• The thread is compatible with apps that use signalfd (will mask all signals for the Manhole threads).

Options

manhole.install(
    verbose=True,
    verbose_destination=2,
    patch_fork=True,
    activate_on=None,
    oneshot_on=None,
    sigmask=manhole.ALL_SIGNALS,
    socket_path=None,
    reinstall_delay=0.5,
    locals=None,
    strict=True,
)

• verbose - Set it to False to squelch the logging.
• verbose_destination - Destination for verbose messages. Set it to a file descriptor or handle. Default is unbuffered stderr (stderr 2 file descriptor).
• patch_fork - Set it to False if you don't want your os.fork and os.forkpy monkeypatched
• activate_on - Set to "USR1", "USR2" or some other signal name, or a number if you want the Manhole thread to start when this signal is sent. This is desirable in case you don't want the thread active all the time.
• thread - Set to True to start the always-on ManholeThread. Default: True. Automatically switched to False if oneshot_on or activate_on are used.
• oneshot_on - Set to "USR1", "USR2" or some other signal name, or a number if you want the Manhole to listen for connection in the signal handler. This is desireable in case you don't want threads at all.
• sigmask - Will set the signal mask to the given list (using signalfd.sigprocmask). No action is done if signalfd is not importable. NOTE: This is done so that the Manhole thread doesn't steal any signals; Normally that is fine because Python will force all the signal handling to be run in the main thread but signalfd doesn't.
• socket_path - Use a specific path for the unix domain socket (instead of /tmp/manhole-). This disables patch_fork as children cannot reuse the same path.
• reinstall_delay - Delay the unix domain socket creation reinstall_delay seconds. This alleviates cleanup failures when using fork+exec patterns.
• locals - Names to add to manhole interactive shell locals.
• daemon_connection - The connection thread is daemonic (dies on app exit). Default: False.
• redirect_stderr - Redirect output from stderr to manhole console. Default: True.
• strict - If True then AlreadyInstalled will be raised when attempting to install manhole twice. Default: True.

Environment variable installation

Manhole can be installed via the PYTHONMANHOLE environment variable.

This:

PYTHONMANHOLE='' python yourapp.py

Is equivalent to having this in yourapp.py:

import manhole
manhole.install()

Any extra text in the environment variable is passed to manhole.install(). Example:

PYTHONMANHOLE='oneshot_on="USR2"' python yourapp.py

What happens when you actually connect to the socket

1. Credentials are checked (if it's same user or root)
2. sys.__std*__/sys.std* are redirected to the UDS
3. Stacktraces for each thread are written to the UDS
4. REPL is started so you can fiddle with the process

Known issues

• Using threads and file handle (not raw file descriptor) verbose_destination can cause deadlocks. See bug reports: PyPy and Python 3.4.

SIGTERM and socket cleanup

By default Python doesn't call the atexit callbacks with the default SIGTERM handling. This makes manhole leave stray socket files around. If this is undesirable you should install a custom SIGTERM handler so atexit is properly invoked.

Example:

import signal
import sys

def handle_sigterm(signo, frame):
    sys.exit(128 + signo)  # this will raise SystemExit and cause atexit to be called

signal.signal(signal.SIGTERM, handle_sigterm)

Using Manhole with uWSGI

Because uWSGI overrides signal handling Manhole is a bit more tricky to setup. One way is to use "uWSGI signals" (not the POSIX signals) and have the workers check a file for the pid you want to open the Manhole in.

Stick something this in your WSGI application file:

from __future__ import print_function
import sys
import os
import manhole

stack_dump_file = '/tmp/manhole-pid'
uwsgi_signal_number = 17

try:
    import uwsgi

    if not os.path.exists(stack_dump_file):
        open(stack_dump_file, 'w')

    def open_manhole(dummy_signum):
        with open(stack_dump_file, 'r') as fh:
            pid = fh.read().strip()
            if pid == str(os.getpid()):
                inst = manhole.install(strict=False, thread=False)
                inst.handle_oneshot(dummy_signum, dummy_signum)

    uwsgi.register_signal(uwsgi_signal_number, 'workers', open_manhole)
    uwsgi.add_file_monitor(uwsgi_signal_number, stack_dump_file)

    print("Listening for stack mahole requests via %r" % (stack_dump_file,), file=sys.stderr)
except ImportError:
    print("Not running under uwsgi; unable to configure manhole trigger", file=sys.stderr)
except IOError:
    print("IOError creating manhole trigger %r" % (stack_dump_file,), file=sys.stderr)


# somewhere bellow you'd have something like
from django.core.wsgi import get_wsgi_application
application = get_wsgi_application()
# or
def application(environ, start_response):
    start_response('200 OK', [('Content-Type', 'text/plain'), ('Content-Length', '2')])
    yield b'OK'

To open the Manhole just run echo 1234 > /tmp/manhole-pid and then manhole-cli 1234.

Requirements

OS
Linux, OS X

Runtime
Python 2.7, 3.4, 3.5, 3.6 or PyPy

Similar projects

• Twisted's manhole - it has colors and server-side history.
• wsgi-shell - spawns a thread.
• pyrasite - uses gdb to inject code.
• pydbattach - uses gdb to inject code.
• pystuck - very similar, uses rpyc for communication.
• pyringe - uses gdb to inject code, more reliable, but relies on dbg python builds unfortunatelly.
• pdb-clone - uses gdb to inject code, with a different strategy.

Changelog

1.8.1 (2024-07-24)

• Fixed buffering issue on Python 3.11. See 66.
• Cleaned up some packaging/test problems.
• Removed more leftover Python 2 code.
• Fixed license metadata. See: 68.

1.8.0 (2021-04-08)

• Simplified connection closing code. Contributed by Anton Ryzhov in 62.
• Made connection shutdown in manhole-cli more graceful. Contributed by Anton Ryzhov in 63.

1.7.0 (2021-03-22)

• Fixed memory leak via sys.last_type, sys.last_value, sys.last_traceback. Contributed by Anton Ryzhov in 59.
• Fixed a bunch of double-close bugs and simplified stream handler code. Contributed by Anton Ryzhov in 58.
• Loosen up pid argument parsing in manhole-cli to allow using paths with any prefix (not just /tmp).

1.6.0 (2019-01-19)

• Testing improvements (changed some skips to xfail, added osx in Travis).
• Fixed long standing Python 2.7 bug where sys.getfilesystemencoding() would be broken after installing a threaded manhole. See 51.
• Dropped support for Python 2.6, 3.3 and 3.4.
• Fixed handling when socket.setdefaulttimeout() is used. Contributed by "honnix" in 53.
• Fixed some typos. Contributed by Jesús Cea in 43.
• Fixed handling in manhole-cli so that timeout is actually seconds and not milliseconds. Contributed by Nir Soffer in 45.
• Cleaned up useless polling options in manhole-cli. Contributed by Nir Soffer in 46.
• Documented and implemented a solution for using Manhole with Eventlet. See 49.

1.5.0 (2017-08-31)

• Added two string aliases for connection_handler option. Now you can conveniently use connection_handler="exec".
• Improved handle_connection_exec. It now has a clean way to exit (exit()) and properly closes the socket.

1.4.0 (2017-08-29)

• Added the connection_handler install option. Default value is manhole.handle_connection_repl, and alternate manhole.handle_connection_exec is provided (very simple: no output redirection, no stacktrace dumping).
• Dropped Python 3.2 from the test grid. It may work but it's a huge pain to support (pip/pytest don't support it anymore).
• Added Python 3.5 and 3.6 in the test grid.
• Fixed issues with piping to manhole-cli. Now echo foobar | manhole-cli will wait 1 second for output from manhole (you can customize this with the --timeout option).
• Fixed issues with newer PyPy (caused by gevent/eventlet socket unwrapping).

1.3.0 (2015-09-03)

• Allowed Manhole to be configured without any thread or activation (in case you want to manually activate).
• Added an example and tests for using Manhole with uWSGi.
• Fixed error handling in manhole-cli on Python 3 (exc vars don't leak anymore).
• Fixed support for running in gevent/eventlet-using apps on Python 3 (now that they support Python 3).
• Allowed reinstalling the manhole (in non-strict mode). Previous install is undone.

1.2.0 (2015-07-06)

• Changed manhole-cli:
  • Won't spam the terminal with errors if socket file doesn't exist.
  • Allowed sending any signal (new --signal argument).
  • Fixed some validation issues for the PID argument.

1.1.0 (2015-06-06)

• Added support for installing the manhole via the PYTHONMANHOLE environment variable.
• Added a strict install option. Set it to false to avoid getting the AlreadyInstalled exception.
• Added a manhole-cli script that emulates socat readline unix-connect:/tmp/manhole-1234.

1.0.0 (2014-10-13)

• Added socket_path install option (contributed by Nir Soffer).
• Added reinstall_delay install option.
• Added locals install option (contributed by Nir Soffer).
• Added redirect_stderr install option (contributed by Nir Soffer).
• Lots of internals cleanup (contributed by Nir Soffer).

0.6.2 (2014-04-28)

• Fix OS X regression.

0.6.1 (2014-04-28)

• Support for OS X (contributed by Saulius Menkevičius).