This replaces the constructor implementation of cgroup handling with a simpler,
thread-safe on-demand model of cgroup driver initialization.
Making the cgroup initialization code run in a constructor means that each time
the shared library gets mapped the cgroup parsing code gets run. That's
unnecessary overhead.
It also feels to me that this is only accidently thread-safe because
constructors are only run once. But should threads actually end up manipulating
or freeing memory that is file-global to cgfsng.c we'd be screwed. Now, I might
be wrong here but the cleaner implementation is to allocate a cgroup driver on
demand whenever we need it.
Take the chance and rework the cgroup_ops interface to make the functions it
wants to have implemented a lot cleaner.
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Unprivileged containers can safely mount /sys as read-write. This also allows
systemd-udevd to be started in unprivileged containers.
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
This reverts commit 2ec47d5149.
First, I forgot to actually replace strncpy() with strlcpy(). Second, we don't
want to \0-terminate since this is an abstract unix socket and this is not
required. Instead, let's simply use memcpy() which is more correct and also
silences gcc-8.
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
The additional \0-byte space added is not needed since IFNAMSIZ needs to
include the \0-byte.
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Apparently -Werror=stringop-overflow will trigger an error here even though
this is completely valid since we now that we're definitely copying a \0-byte.
Work around this gcc-8 quirk by using memcpy(). This shouldn't trigger the
warning.
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Dereference after null check
userns_exec_{1,full} are called from functions that might not have a conf.
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
These files have never been used and as such have no dependencies in the
codebase whatsoever. So remove them. If we need them we can simply pull them
out of the git history.
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
I'm not really sure we should be looking in the rootfs for an existing
init, but I'll send a much more invasive patch to correct that. For now,
let's just make sure we set init_path when we find one, so that later in
execute_start() we don't bail.
Signed-off-by: Tycho Andersen <tycho@tycho.ws>
This always works fine... until your exec() fails and you try to go and
free it, you've overwritten the allocator's metadata (and potentially other
stuff) and it fails.
Signed-off-by: Tycho Andersen <tycho@tycho.ws>
The errors in execute_start are important because nothing actually prints
out what error if any there was in these cases, so you're left with an
empty log.
The TRACE logs are simply to tell you which version of start lxc chose to
invoke: exec or start.
Signed-off-by: Tycho Andersen <tycho@tycho.ws>
