This is the default thread size for glibc, so it is reasonable to match
that when we clone().
Mostly this is a science experiment suggested by brauner, and who doesn't
love science?
Signed-off-by: Tycho Andersen <tycho@tycho.ws>
Handle offline cpus in v1 hierarchy.
In addition to isolated cpus we also need to account for offline cpus when our
ancestor cgroup is the root cgroup and we have not been initialized yet.
Closes#2953.
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Let lxc_attach() reuse the already initialized container.
Closes https://github.com/lxc/lxd/issues/5755.
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Updates lxc_restore_phys_nics_to_netns() to move phys netdevs back to the monitor's network namespace rather than the previously hardcoded PID 1 net ns.
This is to fix instances where LXC is started inside a net ns different from PID 1 and physical devices are moved back to a different net ns when the container is shutdown than the net ns than where the container was started from.
Signed-off-by: Thomas Parrott <thomas.parrott@canonical.com>
We have a do_clone(), which just calls a void f(void *) that it gets
passed. We build up a struct consisting of two args that are just the
actual arg and actual function. Let's just have the syscall do this for us.
Signed-off-by: Tycho Andersen <tycho@tycho.ws>
There are two problems with this code:
1. The math is wrong. We allocate a char *foo[__LXC_STACK_SIZE]; which
means it's really sizeof(char *) * __LXC_STACK_SIZE, instead of just
__LXC_STACK SIZE.
2. We can't actually allocate it on our stack. When we use CLONE_VM (which
we do in the shared ns case) that means that the new thread is just
running one page lower on the stack, but anything that allocates a page
on the stack may clobber data. This is a pretty short race window since
we just do the shared ns stuff and then do a clone without CLONE_VM.
However, it does point out an interesting possible privilege escalation if
things aren't configured correctly: do_share_ns() sets up namespaces while
it shares the address space of the task that spawned it; once it enters the
pid ns of the thing it's sharing with, the thing it's sharing with can
ptrace it and write stuff into the host's address space. Since the function
that does the clone() is lxc_spawn(), it has a struct cgroup_ops* on the
stack, which itself has function pointers called later in the function, so
it's possible to allocate shellcode in the address space of the host and
run it fairly easily.
ASLR doesn't mitigate this since we know exactly the stack offsets; however
this patch has the kernel allocate a new stack, which will help. Of course,
the attacker could just check /proc/pid/maps to find the location of the
stack, but they'd still have to guess where to write stuff in.
The thing that does prevent this is the default configuration of apparmor.
Since the apparmor profile is set in the second clone, and apparmor
prevents ptracing things under a different profile, attackers confined by
apparmor can't do this. However, if users are using a custom configuration
with shared namespaces, care must be taken to avoid this race.
Shared namespaces aren't widely used now, so perhaps this isn't a problem,
but with the advent of crio-lxc for k8s, this functionality will be used
more.
Signed-off-by: Tycho Andersen <tycho@tycho.ws>
