Commit 7bb2fef0e2 attempted to fix
the "missing NUL-terminator in VERSION" issue by replacing the "%s"
conversion specifier with "%*s". However, this fix itself was wrong,
as "%*s" specifies the field width, i.e. _minimum_ number of characters
to print. The proper solution is to use "%.*s", which specifies the
precision, i.e. the _maximum_ number of printed characters.
The "%s" conversion specifier expects a NUL-terminated string.
However, the VERSION variable does not contain a NUL-terminator,
so formatting it using "%s" may lead to printing whatever happens
to be in memory next to VERSION.
Using "%*s" allows to specify how many characters to print,
thus making sure we don't go off the array.
There's no reason the dumb-init process should hold on to a reference
to the original cwd (but of course the child must be spawned with that
cwd). One example where that can be problematic:
# Terminal 1
$ mkdir -p /tmp/d/e
$ sudo mount --make-shared -t tmpfs tmpfs /tmp/d/e
# Terminal 2
$ docker run -v /tmp/d:/tmp/d:rshared --rm -ti --workdir /tmp/d/e some-image-with-dumb-init bash
root@922ef180b49a:/tmp/d/e# cd ..
root@922ef180b49a:/tmp/d# ls -l /proc/*/cwd
lrwxrwxrwx 1 root root 0 Oct 1 11:25 /proc/1/cwd -> /tmp/d/e
lrwxrwxrwx 1 root root 0 Oct 1 11:25 /proc/6/cwd -> /tmp/d
lrwxrwxrwx 1 root root 0 Oct 1 11:25 /proc/self/cwd -> /tmp/d
lrwxrwxrwx 1 root root 0 Oct 1 11:25 /proc/thread-self/cwd -> /tmp/d
# Terminal 1
$ sudo umount /tmp/d/e
umount: /tmp/d/e: target is busy.
i.e., the application inside the container has already let go of its
reference, but the host is not able to unmount /tmp/d/e because the
container's pid 1 still has one.
As for a test case, the 'docker run --workdir' could be emulated by
something like
sh -c "cd /tmp/d/e; exec $PWD/dumb-init bash -i"
but one would still need root to do the mount and umount.
We've decided to allow arbitrary signal rewriting, not just SIGTERM
rewriting. To use, call dumb-init with the '-r s:r' option, which will
rewrite signum s to signum r. Only signals 1-31 are allowed to be
rewritten, which should cover all the signals we need to cover.
Note that this commit does not add new tests, it only fixes the existing
broken test.
Many servers respond to other signals than SIGTERM for their "soft
shutdown" option, such as Unicorn which requires SIGQUIT to wait on
outstanding connections. The 'docker stop' command sends the SIGTERM
signal to the container, and provides no option for modifying this
behavior. The 'docker kill' command has an '-s' option which allows one
to modify the signal sent to the container, but orchestration frameworks
such as Mesos don't provide a way to use this functionality.
This commit adds the '-s/--signal' option to dumb-init, which provides a
replacement signal for SIGTERM. For instance, running dumb-init like so:
dumb-init -s 3 <command>
Will send SIGQUIT (3) to the <command> process it spawns when it
receives a SIGTERM. This allows Docker image writers the freedom to
specify how SIGTERM will behave in their containers.
A further improvement to this option could be to provide an arbirary
mapping from signal to signal, but that would greatly complicate the
code for a probably minor use case.
