This commit follows further on the deprecation path for --share-system,
by splitting and gating each share-able namespace behind its own
environment flag.
Currently, systemd uses either the legacy hierarchies or the unified hierarchy.
When the legacy hierarchies are used, systemd uses a named legacy hierarchy
mounted on /sys/fs/cgroup/systemd without any kernel controllers for process
management. Due to the shortcomings in the legacy hierarchy, this involves a
lot of workarounds and complexities.
Because the unified hierarchy can be mounted and used in parallel to legacy
hierarchies, there's no reason for systemd to use a legacy hierarchy for
management even if the kernel resource controllers need to be mounted on legacy
hierarchies. It can simply mount the unified hierarchy under
/sys/fs/cgroup/systemd and use it without affecting other legacy hierarchies.
This disables a significant amount of fragile workaround logics and would allow
using features which depend on the unified hierarchy membership such bpf cgroup
v2 membership test. In time, this would also allow deleting the said
complexities.
This patch updates systemd so that it prefers the unified hierarchy for the
systemd cgroup controller hierarchy when legacy hierarchies are used for kernel
resource controllers.
* cg_unified(@controller) is introduced which tests whether the specific
controller in on unified hierarchy and used to choose the unified hierarchy
code path for process and service management when available. Kernel
controller specific operations remain gated by cg_all_unified().
* "systemd.legacy_systemd_cgroup_controller" kernel argument can be used to
force the use of legacy hierarchy for systemd cgroup controller.
* nspawn: By default nspawn uses the same hierarchies as the host. If
UNIFIED_CGROUP_HIERARCHY is set to 1, unified hierarchy is used for all. If
0, legacy for all.
* nspawn: arg_unified_cgroup_hierarchy is made an enum and now encodes one of
three options - legacy, only systemd controller on unified, and unified. The
value is passed into mount setup functions and controls cgroup configuration.
* nspawn: Interpretation of SYSTEMD_CGROUP_CONTROLLER to the actual mount
option is moved to mount_legacy_cgroup_hierarchy() so that it can take an
appropriate action depending on the configuration of the host.
v2: - CGroupUnified enum replaces open coded integer values to indicate the
cgroup operation mode.
- Various style updates.
v3: Fixed a bug in detect_unified_cgroup_hierarchy() introduced during v2.
v4: Restored legacy container on unified host support and fixed another bug in
detect_unified_cgroup_hierarchy().
A following patch will update cgroup handling so that the systemd controller
(/sys/fs/cgroup/systemd) can use the unified hierarchy even if the kernel
resource controllers are on the legacy hierarchies. This would require
distinguishing whether all controllers are on cgroup v2 or only the systemd
controller is. In preparation, this patch renames cg_unified() to
cg_all_unified().
This patch doesn't cause any functional changes.
This removes the --share-system switch: from the documentation, the --help text
as well as the command line parsing. It's an ugly option, given that it kinda
contradicts the whole concept of PID namespaces that nspawn implements. Since
it's barely ever used, let's just deprecate it and remove it from the options.
It might be useful as a debugging option, hence the functionality is kept
around for now, exposed via an undocumented $SYSTEMD_NSPAWN_SHARE_SYSTEM
environment variable.
This has the benefit that the container can follow the host's DNS server
changes without us having to constantly update the container's resolv.conf
settings.
With this change we'll no longer write to /etc/machine-id from nspawn, as that
breaks the --volatile= operation, as it ensures the image is never considered
in "first boot", since that's bound to the pre-existance of /etc/machine-id.
The new logic works like this:
- If /etc/machine-id already exists in the container, it is read by nspawn and
exposed in "machinectl status" and friends.
- If the file doesn't exist yet, but --uuid= is passed on the nspawn cmdline,
this UUID is passed in $container_uuid to PID 1, and PID 1 is then expected
to persist this to /etc/machine-id for future boots (which systemd already
does).
- If the file doesn#t exist yet, and no --uuid= is passed a random UUID is
generated and passed via $container_uuid.
The result is that /etc/machine-id is never initialized by nspawn itself, thus
unbreaking the volatile mode. However still the machine ID configured in the
machine always matches nspawn's and thus machined's idea of it.
Fixes: #3611
We currently have code to read and write files containing UUIDs at various
places. Unify this in id128-util.[ch], and move some other stuff there too.
The new files are located in src/libsystemd/sd-id128/ (instead of src/shared/),
because they are actually the backend of sd_id128_get_machine() and
sd_id128_get_boot().
In follow-up patches we can use this reduce the code in nspawn and
machine-id-setup by adopted the common implementation.
Matching the behaviour of gpt-auto-generator, if we find an ESP while
dissecting a container image, mount it to /efi or /boot if those dirs exist and
are empty.
This should enable us to run "bootctl" inside a container and do the right
thing.
Normally we make all of /proc/sys read-only in a container, but if we do have
netns enabled we can make /proc/sys/net writable, as things are virtualized
then.
(NOTE: Cgroup namespaces work with legacy and unified hierarchies: "This is
completely backward compatible and will be completely invisible to any existing
cgroup users (except for those running inside a cgroup namespace and looking at
/proc/pid/cgroup of tasks outside their namespace.)"
(https://lists.linuxfoundation.org/pipermail/containers/2016-January/036582.html)
So there is no need to special case unified.)
If cgroup namespaces are supported we skip mount_cgroups() in the
outer_child(). Instead, we unshare(CLONE_NEWCGROUP) in the inner_child() and
only then do we call mount_cgroups().
The clean way to handle cgroup namespaces would be to delegate mounting of
cgroups completely to the init system in the container. However, this would
likely break backward compatibility with the UNIFIED_CGROUP_HIERARCHY flag of
systemd-nspawn. Also no cgroupfs would be mounted whenever the user simply
requests a shell and no init is available to mount cgroups. Hence, we introduce
mount_legacy_cgns_supported(). After calling unshare(CLONE_NEWCGROUP) it parses
/proc/self/cgroup to find the mounted controllers and mounts them inside the
new cgroup namespace. This should preserve backward compatibility with the
UNIFIED_CGROUP_HIERARCHY flag and mount a cgroupfs when no init in the
container is running.
This the patch implements a notificaiton mechanism from the init process
in the container to systemd-nspawn.
The switch --notify-ready=yes configures systemd-nspawn to wait the "READY=1"
message from the init process in the container to send its own to systemd.
--notify-ready=no is equivalent to the previous behavior before this patch,
systemd-nspawn notifies systemd with a "READY=1" message when the container is
created. This notificaiton mechanism uses socket file with path relative to the contanier
"/run/systemd/nspawn/notify". The default values it --notify-ready=no.
It is also possible to configure this mechanism from the .nspawn files using
NotifyReady. This parameter takes the same options of the command line switch.
Before this patch, systemd-nspawn notifies "ready" after the inner child was created,
regardless the status of the service running inside it. Now, with --notify-ready=yes,
systemd-nspawn notifies when the service is ready. This is really useful when
there are dependencies between different contaniers.
Fixes https://github.com/systemd/systemd/issues/1369
Based on the work from https://github.com/systemd/systemd/pull/3022
Testing:
Boot a OS inside a container with systemd-nspawn.
Note: modify the commands accordingly with your filesystem.
1. Create a filesystem where you can boot an OS.
2. sudo systemd-nspawn -D ${HOME}/distros/fedora-23/ sh
2.1. Create the unit file /etc/systemd/system/sleep.service inside the container
(You can use the example below)
2.2. systemdctl enable sleep
2.3 exit
3. sudo systemd-run --service-type=notify --unit=notify-test
${HOME}/systemd/systemd-nspawn --notify-ready=yes
-D ${HOME}/distros/fedora-23/ -b
4. In a different shell run "systemctl status notify-test"
When using --notify-ready=yes the service status is "activating" for 20 seconds
before being set to "active (running)". Instead, using --notify-ready=no
the service status is marked "active (running)" quickly, without waiting for
the 20 seconds.
This patch was also test with --private-users=yes, you can test it just adding it
at the end of the command at point 3.
------ sleep.service ------
[Unit]
Description=sleep
After=network.target
[Service]
Type=oneshot
ExecStart=/bin/sleep 20
[Install]
WantedBy=multi-user.target
------------ end ------------
The current raw_clone function takes two arguments, the cloning flags and
a pointer to the stack for the cloned child. The raw cloning without
passing a "thread main" function does not make sense if a new stack is
specified, as it returns in both the parent and the child, which will fail
in the child as the stack is virgin. All uses of raw_clone indeed pass NULL
for the stack pointer which indicates that both processes should share the
stack address (so you better don't pass CLONE_VM).
This commit refactors the code to not require the caller to pass the stack
address, as NULL is the only sensible option. It also adds the magic code
needed to make raw_clone work on sparc64, which does not return 0 in %o0
for the child, but indicates the child process by setting %o1 to non-zero.
This refactoring is not plain aesthetic, because non-NULL stack addresses
need to get mangled before being passed to the clone syscall (you have to
apply STACK_BIAS), whereas NULL must not be mangled. Implementing the
conditional mangling of the stack address would needlessly complicate the
code.
raw_clone is moved to a separete header, because the burden of including
the assert machinery and sched.h shouldn't be applied to every user of
missing_syscalls.h
Split seccomp into nspawn-seccomp.[ch]. Currently there are no changes,
but this will make it easy in the future to share or use the seccomp logic
from systemd core.
This adds a new concept of network "zones", which are little more than bridge
devices that are automatically managed by nspawn: when the first container
referencing a bridge is started, the bridge device is created, when the last
container referencing it is removed the bridge device is removed again. Besides
this logic --network-zone= is pretty much identical to --network-bridge=.
The usecase for this is to make it easy to run multiple related containers
(think MySQL in one and Apache in another) in a common, named virtual Ethernet
broadcast zone, that only exists as long as one of them is running, and fully
automatically managed otherwise.
* sd-netlink: permit RTM_DELLINK messages with no ifindex
This is useful for removing network interfaces by name.
* nspawn: explicitly remove veth links we created after use
Sometimes the kernel keeps veth links pinned after the namespace they have been
joined to died. Let's hence explicitly remove veth links after use.
Fixes: #2173
Sometimes the kernel keeps veth links pinned after the namespace they have been
joined to died. Let's hence explicitly remove veth links after use.
Fixes: #2173
In order to implement this we change the bool arg_userns into an enum
UserNamespaceMode, which can take one of NO, PICK or FIXED, and replace the
arg_uid_range_pick bool with it.
This adds the new value "pick" to --private-users=. When specified a new
UID/GID range of 65536 users is automatically and randomly allocated from the
host range 0x00080000-0xDFFF0000 and used for the container. The setting
implies --private-users-chown, so that container directory is recursively
chown()ed to the newly allocated UID/GID range, if that's necessary. As an
optimization before picking a randomized UID/GID the UID of the container's
root directory is used as starting point and used if currently not used
otherwise.
To protect against using the same UID/GID range multiple times a few mechanisms
are in place:
- The first and the last UID and GID of the range are checked with getpwuid()
and getgrgid(). If an entry already exists a different range is picked. Note
that by "last" UID the user 65534 is used, as 65535 is the 16bit (uid_t) -1.
- A lock file for the range is taken in /run/systemd/nspawn-uid/. Since the
ranges are taken in a non-overlapping fashion, and always start on 64K
boundaries this allows us to maintain a single lock file for each range that
can be randomly picked. This protects nspawn from picking the same range in
two parallel instances.
- If possible the /etc/passwd lock file is taken while a new range is selected
until the container is up. This means adduser/addgroup should safely avoid
the range as long as nss-mymachines is used, since the allocated range will
then show up in the user database.
The UID/GID range nspawn picks from is compiled in and not configurable at the
moment. That should probably stay that way, since we already provide ways how
users can pick their own ranges manually if they don't like the automatic
logic.
The new --private-users=pick logic makes user namespacing pretty useful now, as
it relieves the user from managing UID/GID ranges.
This adds a new --private-userns-chown switch that may be used in combination
with --private-userns. If it is passed a recursive chmod() operation is run on
the OS tree, fixing all file owner UID/GIDs to the right ranges. This should
make user namespacing pretty workable, as the OS trees don't need to be
prepared manually anymore.
Luca Bruno
Tejun Heo
Lennart Poettering
Zbigniew Jędrzejewski-Szmek
Christian Brauner
Thomas Hindoe Paaboel Andersen
Michael Biebl
Alessandro Puccetti
Michael Karcher
Djalal Harouni
Evgeny Vereshchagin