Using mysqld_multi via sh kills mysqladmin grandchildren with SIGHUP before they have a chance to do anything

[hoelzro]

I found this issue when writing a MySQL helper script that uses mysqld_multi. The problematic line looks like this:

sh.mysqld_multi('stop', '1')

mysqld_multi is an interesting script that forks off a mysqladmin process for each start/stop action, but doesn't wait for them to finish before exiting itself. As a result, the controlling terminal that sh creates for mysqld_multi (which is the controlling terminal for its mysqladmin children) is closed when sh cleans up after invoking mysqld_multi, and the mysqladmin grandchildren processes are sent a SIGHUP by the kernel.

Here is a script that reproduces the behavior without involving MySQL scripts:

import sh
from tempfile import NamedTemporaryFile
from textwrap import dedent
import time

with \
    NamedTemporaryFile() as output_file, \
    NamedTemporaryFile() as child_file, \
    NamedTemporaryFile() as parent_file:

    child_file.write(dedent('''
    import signal
    import sys
    import time

    with open({output_file}, 'w') as f:
        def handle_sighup(signum, frame):
            f.write('got signal %d' % signum)
            sys.exit(signum)
        signal.signal(signal.SIGHUP, handle_sighup)
        time.sleep(5)
        f.write('made it!\\n')
    '''.format(output_file=repr(output_file.name))))
    child_file.flush()

    parent_file.write(dedent('''
    import os
    import time

    os.spawnlp(os.P_NOWAIT, 'python', 'python', {child_file})
    time.sleep(2) # give child a chance to set up
    '''.format(child_file=repr(child_file.name))))

    parent_file.flush()

    sh.python(parent_file.name)
    time.sleep(10)

    print output_file.readlines()[0] # should print 'made it'

[amoffat]

After a lot of reading, I think the right move here is to only give an sh subprocess a controlling terminal IFF we are using _tty_in=True (which is False by default). My rationale is that if a process doesn't have a tty for stdin, then it doesn't make sense for it to have a controlling terminal (because, how could you send signals from the controlling terminal if you cannot input through a tty). This solves grandchild processes from receiving SIGHUP (because there is no ctty that hangs up). Additonally, if you do choose to launch a process with _tty_in=True, then you're launching the process as if it were a shell, so any subprocesses launched within that shell should receive SIGHUP, just like if they were launched in the foreground from a shell, and the shell hangs up.

e1d2a2c

this will go out with the release-1.2 release. interested if you have any additional thoughts