The OCI changes in #9762 broke a use case in which we use nspawn from
inside a container that has dropped capabilities from the bounding set
that nspawn expected to retain. In an attempt to keep OCI compliance
and support our use case, I made hard failing on setting capabilities
not in the bounding set optional (hard fail if using OCI and log only
if using nspawn cmdline).
Fixes#12539
The console tty is now allocated from within the container so it's not
necessary anymore to allocate it from the host and bind mount the pty slave
into the container. The pty master is sent to the host.
/dev/console is now a symlink pointing to the pty slave.
This might also be less confusing for applications running inside the container
and the overall result looks cleaner (we don't need to apply manually the
passed selinux context, if any, to the allocated pty for instance).
The CPU_SET_S api is pretty bad. In particular, it has a parameter for the size
of the array, but operations which take two (CPU_EQUAL_S) or even three arrays
(CPU_{AND,OR,XOR}_S) still take just one size. This means that all arrays must
be of the same size, or buffer overruns will occur. This is exactly what our
code would do, if it received an array of unexpected size over the network.
("Unexpected" here means anything different from what cpu_set_malloc() detects
as the "right" size.)
Let's rework this, and store the size in bytes of the allocated storage area.
The code will now parse any number up to 8191, independently of what the current
kernel supports. This matches the kernel maximum setting for any architecture,
to make things more portable.
Fixes#12605.
We noticed in our tests that occasionally SystemCallFilter= would
fail to set and the service would run with no syscall filtering.
Most of the time the same tests would apply the filter and fail
the service as expected. While it's not totally clear why this happens,
we noticed seccomp_load() in the systemd code base would fail open for
all errors except EPERM and EACCES.
ENOMEM, EINVAL, and EFAULT seem like reasonable values to add to the
error set based on what I gather from libseccomp code and man pages:
-ENOMEM: out of memory, failed to allocate space for a libseccomp structure, or would exceed a defined constant
-EINVAL: kernel isn't configured to support the operations, args are invalid (to seccomp_load(), seccomp(), or prctl())
-EFAULT: addresses passed as args are invalid
/tmp might not be mounted at all yet (given that we support
SYSTEMD_NSPAWN_TMPFS_TMP=0 to turn this off), and /tmp is a dir systemd
usually tries to unmount during shutdown (unlike /run), and we shouldn't
keep it busy. Hence let's just move these deleted files to /run so that
we don't keep /tmp needlessly busy.
Some chattrs only work sensible if you set them right after opening a
file for create (think: FS_NOCOW_FL). Others only work when they are
applied when the file is fully written (think: FS_IMMUTABLE_FL). Let's
take that into account when copying files and applying a chattr to them.
The function is otherwise generic enough to toggle other bind mount
flags beyond MS_RDONLY (for example: MS_NOSUID or MS_NODEV), hence let's
beef it up slightly to support that too.
Previously both run() and run_container() would free 'fds'. Let's fix
that, and let run() free it but make run_container() already remove all
fds from it, because that's what we actually want to do.
Fixes: #12073
This is a pretty large patch, and adds support for OCI runtime bundles
to nspawn. A new switch --oci-bundle= is added that takes a path to an
OCI bundle. The JSON file included therein is read similar to a .nspawn
settings files, however with a different feature set.
Implementation-wise this mostly extends the pre-existing Settings object
to carry additional properties for OCI. However, OCI supports some
concepts .nspawn files did not support yet, which this patch also adds:
1. Support for "masking" files and directories. This functionatly is now
also available via the new --inaccesible= cmdline command, and
Inaccessible= in .nspawn files.
2. Support for mounting arbitrary file systems. (not exposed through
nspawn cmdline nor .nspawn files, because probably not a good idea)
3. Ability to configure the console settings for a container. This
functionality is now also available on the nspawn cmdline in the new
--console= switch (not added to .nspawn for now, as it is something
specific to the invocation really, not a property of the container)
4. Console width/height configuration. Not exposed through
.nspawn/cmdline, but this may be controlled through $COLUMNS and
$LINES like in most other UNIX tools.
5. UID/GID configuration by raw numbers. (not exposed in .nspawn and on
the cmdline, since containers likely have different user tables, and
the existing --user= switch appears to be the better option)
6. OCI hook commands (no exposed in .nspawn/cmdline, as very specific to
OCI)
7. Creation of additional devices nodes in /dev. Most likely not a good
idea, hence not exposed in .nspawn/cmdline. There's already --bind=
to achieve the same, which is the better alternative.
8. Explicit syscall filters. This is not a good idea, due to the skewed
arch support, hence not exposed through .nspawn/cmdline.
9. Configuration of some sysctls on a whitelist. Questionnable, not
supported in .nspawn/cmdline for now.
10. Configuration of all 5 types of capabilities. Not a useful concept,
since the kernel will reduce the caps on execve() anyway. Not
exposed through .nspawn/cmdline as this is not very useful hence.
Note that this only implements the OCI runtime logic itself. It does not
provide a runc-compatible command line tool. This is left for a later
PR. Only with that in place tools such as "buildah" can use the OCI
support in nspawn as drop-in replacement.
Currently still missing is OCI hook support, but it's already parsed and
everything, and should be easy to add. Other than that it's OCI is
implemented pretty comprehensively.
There's a list of incompatibilities in the nspawn-oci.c file. In a later
PR I'd like to convert this into proper markdown and add it to the
documentation directory.
if we sync the legacy and unified trees before moving to the right
subcgroup then ultimately the cgroup paths in the hierarchies will be
out-of-sync... Hence, let's move the payload first, and sync then.
Addresses: https://github.com/systemd/systemd/pull/9762#issuecomment-441187979
Previously, when we'd copy an individual file we'd synthesize a
user.crtime_usec xattr with the source's creation time if we can
determine it. As the creation/birth time was until recently not
queriable form userspace this effectively just propagated the same xattr
on the source to the same xattr on the destination. However, current
kernels now allow to query the birthtime using statx() and we do make
use of that now. Which means that suddenly we started synthesizing these
xattrs much more regularly.
Doing this actually does make sense, but only in very few cases:
not for the typical regular files we copy, but certainly when dealing
with disk images. Hence, let's keep this kind of propagation, but let's
make it a flag and default to off. Then turn it on whenever we deal with
disk images, and leave it off otherwise.
This is particularly relevant as overlayfs combining a real fs, and a
tmpfs on top will result in EOPNOTSUPP when it is attempted to open a
file with xattrs for writing, as tmpfs does not support xattrs, and
hence the copy-up cannot work. Hence, let's avoid synthesizing this
needlessly, to increase compat with overlayfs.
Previously, we'd refuse the combination, and claimed we'd imply it, but
actually didn't. Let's allow the combination and imply read-only from
--volatile=, because that's what's documented, what we claim we do, and
what makes sense.
cg_ns_supported() caches, so the condition was really checked just once, but
it looks weird to assign the return value to arg_use_cgns (if the variable is not present),
because then the other checks are effectively equivalent to
if (cg_ns_supported() && cg_ns_supported()) { ...
and later
if (!cg_ns_supported() || !cg_ns_supported()) { ...
That's what the function is for after all, and only if it's done there
we can verify the effect of .nspawn files correctly too: after all we
should not just validate that everything configured on the command line
makes sense, but the stuff configured in the .nspawn files, too.
This then let's us to ensure it's called after we parsed the cmdline,
and after we loaded the settings file, so that it these env var settings
override everything loaded from there.
This splits out a bunch of functions from fileio.c that have to do with
temporary files. Simply to make the header files a bit shorter, and to
group things more nicely.
No code changes, just some rearranging of source files.
Ideally, coccinelle would strip unnecessary braces too. But I do not see any
option in coccinelle for this, so instead, I edited the patch text using
search&replace to remove the braces. Unfortunately this is not fully automatic,
in particular it didn't deal well with if-else-if-else blocks and ifdefs, so
there is an increased likelikehood be some bugs in such spots.
I also removed part of the patch that coccinelle generated for udev, where we
returns -1 for failure. This should be fixed independently.
Pretty much everything uses just the first argument, and this doesn't make this
common pattern more complicated, but makes it simpler to pass multiple options.
From #10526:
$ sudo systemd-nspawn -i image
Spawning container image on /home/zbyszek/src/mkosi/image.
Press ^] three times within 1s to kill container.
Short read while reading cgroup mode.
We have the machine name anyway, let's use TerminateMachine() on
machined's Manager object directly with it. That way it's a single
method call only, instead of two, to terminate the machine.
When running a read-only file system, we might not be able to create
/var/log/journal. Do not fail on this, unless actually requested by the
--link-journal options.
$ systemd-nspawn --image=image.squashfs ...
All over the place we define local variables for the various sockopts
that take a bool-like "int" value. Sometimes they are const, sometimes
static, sometimes both, sometimes neither.
Let's clean this up, introduce a common const variable "const_int_one"
(as well as one matching "const_int_zero") and use it everywhere, all
acorss the codebase.
With this change almost all log messages that are suppressed through
--quiet are not actually suppressed anymore, but simply downgraded to
LOG_DEBUG. Previously we did it this way for some log messages and fully
suppressed them for others. With this it's pretty much systematic.
Inspired by #10122.
This is a bit like the info link in most of GNU's --help texts, but we
don't do info but man pages, and we make them properly clickable on
terminal supporting that, because awesome.
I think it's generally advisable to link up our (brief) --help texts and
our (more comprehensive) man pages a bit, so this should be an easy and
straight-forward way to do it.
One of the things that tmpfs_patch_options does is take an (optional) UID,
and insert "uid=${UID},gid=${UID}" into the options string. So we need a
uid_t argument, and a way of telling if we should use it. Fortunately,
that is built in to the uid_t type by having UID_INVALID as a possible
value.
So this is really a feature that requires one argument. Yet, it is somehow
taking 4! That is absurd. Simplify it to only take one argument, and have
that trickle all the way up to mount_all()'s usage.
Now, in may of the uses, the argument becomes
uid_shift == 0 ? UID_INVALID : uid_shift
because it used to treat uid_shift=0 as invalid unless the patch_ids flag
was also set. This keeps the behavior the same. Note that in all cases
where it is invoked, if !use_userns (sometimes called !userns), then
uid_shift is 0; we don't have to add any checks for that.
That said, I'm pretty sure that "uid=0" and not setting "uid=" are the
same, but Christian Brauner seemed to not think so when implementing the
cgns support. https://github.com/systemd/systemd/pull/3589
nspawn as it is now is a generally useful tool, hence let's drop the
comments about it being useful for debug and so on only.
The new wording just makes the first sentence of the main page also the
summary.
These lines are generally out-of-date, incomplete and unnecessary. With
SPDX and git repository much more accurate and fine grained information
about licensing and authorship is available, hence let's drop the
per-file copyright notice. Of course, removing copyright lines of others
is problematic, hence this commit only removes my own lines and leaves
all others untouched. It might be nicer if sooner or later those could
go away too, making git the only and accurate source of authorship
information.
This part of the copyright blurb stems from the GPL use recommendations:
https://www.gnu.org/licenses/gpl-howto.en.html
The concept appears to originate in times where version control was per
file, instead of per tree, and was a way to glue the files together.
Ultimately, we nowadays don't live in that world anymore, and this
information is entirely useless anyway, as people are very welcome to
copy these files into any projects they like, and they shouldn't have to
change bits that are part of our copyright header for that.
hence, let's just get rid of this old cruft, and shorten our codebase a
bit.
This is mostly fall-out from d1a1f0aaf0,
however some cases are older bugs.
There might be more issues lurking, this was a simple grep for "%m"
across the tree, with all lines removed that mention "errno" at all.
Let's rework error handling a bit in image_find() and friends: when we
can't find an image, return -ENOENT rather than 0. That's better as
before we violated the usual rule in our codebase that return parameters
are initialized when the return value is >= 0 and otherwise not touched.
This also makes enumeration and validation a bit more strict: we'll only
accept ".raw" as suffix for regular files, and filter out this suffix
handling on directories/subvolumes, where it makes no sense.
This distuingishes two different classes of images, one for the purpose
of npsawn-like containers, i.e. "machines", and one for portable
services.
This distinction is mostly about search paths. We look for machine
images in /var/lib/machines and for portable images in
/var/lib/portables.
We already do this kind of validation in nspawn when we operate on a
plain directory, let's also do this on raw images under the same
condition: that we are about too boot the image. Also, do this when we
are about to read OS metadata from it.
If we log to the pty that is configured as stdin/stdout/stderr of the
container too early we risk filling it up in full before we start
processing the pty from the parent process, resulting in deadlocks.
Let's hence keep a copy of the original tty we were started on before
setting up stdin/stdout/stderr, so that we can log to it, and keep using
it as long as we can.
Since the kernel's pty internal buffer is pretty small this actually
triggered deadlocks when we debug logged at lot from nspawn's child
processes, see: https://github.com/systemd/systemd/pull/9024#issuecomment-390403674
With this change we won't use the pty at all, only the actual payload we
start will, and hence we won't deadlock on it, ever.
Let's make use of log_set_open_when_needed() in nspawn too, i.e. at the
point where we close logging because we are about to rearrange fds,
let's automatically reopen the logging fds when we need them, the same
way as we do that in the service manager. This makes things simpler and
more robust.
Let's separate the loading of the settings object and the merging into
our arg_xyz fields into two.
This will become particularly useful when we eventually are able to load
settings from OCI runtime files in addition to .nspawn files.
Similar as the other options added before, this is primarily useful to
provide comprehensive OCI runtime compatbility, but might be useful
otherwise, too.
This simply controls the PR_SET_NO_NEW_PRIVS flag for the container.
This too is primarily relevant to provide OCI runtime compaitiblity, but
might have other uses too, in particular as it nicely complements the
existing --capability= and --drop-capability= flags.
Previously, the container's hostname was exclusively initialized from
the machine name configured with --machine=, i.e. the internal name and
the external name used for and by the container was synchronized. This
adds a new option --hostname= that optionally allows the internal name
to deviate from the external name.
This new option is mainly useful to ultimately implement the OCI runtime
spec directly in nspawn, but it might be useful on its own for some
other usecases too.
This ensures we set the various resource limits of our container
explicitly on each invocation so that we inherit less from our callers
into the payload.
By default resource limits are now set to the same values Linux
generally passes to the host PID 1, thus minimizing needless differences
between host and container environments.
The limits are now also configurable using a new --rlimit= switch. This
is preparation for teaching nspawn native OCI runtime support as OCI
permits setting resource limits for container payloads, and it hence
probably makes sense if we do too.
We need to do this in all cases, including on cgroupsv1 in order to
ensure the host systemd and any systemd in the payload won't fight for
the cgroup attributes of the top-level cgroup of the payload.
This is because systemd for Delegate=yes units will only delegate the
right to create children as well as their attributes. However, nspawn
expects that the cgroup delegated covers both the right to create
children and the attributes of the cgroup itself. Hence, to clear this
up, let's unconditionally insert a intermediary cgroup, on cgroupsv1 as
well as cgroupsv2, unconditionally.
This is also nice as it reduces the differences in the various setups
and exposes very close behaviour everywhere.
Let's not make /run too special and let's make sure the source file is
not guessable: let's use our regular temporary file helper calls to
create the source node.
This drops a good number of type-specific _cleanup_ macros, and patches
all users to just use the generic ones.
In most recent code we abstained from defining type-specific macros, and
this basically removes all those added already, with the exception of
the really low-level ones.
Having explicit macros for this is not too useful, as the expression
without the extra macro is generally just 2ch wider. We should generally
emphesize generic code, unless there are really good reasons for
specific code, hence let's follow this in this case too.
Note that _cleanup_free_ and similar really low-level, libc'ish, Linux
API'ish macros continue to be defined, only the really high-level OO
ones are dropped. From now on this should really be the rule: for really
low-level stuff, such as memory allocation, fd handling and so one, go
ahead and define explicit per-type macros, but for high-level, specific
program code, just use the generic _cleanup_() macro directly, in order
to keep things simple and as readable as possible for the uninitiated.
Note that before this patch some of the APIs (notable libudev ones) were
already used with the high-level macros at some places and with the
generic _cleanup_ macro at others. With this patch we hence unify on the
latter.
We don't know what the container payload needs, hence default to a PATH
with both bin and sbin included, as well as / and /usr.
Follow-up for #8324Fixes: #8698
Double newlines (i.e. one empty lines) are great to structure code. But
let's avoid triple newlines (i.e. two empty lines), quadruple newlines,
quintuple newlines, …, that's just spurious whitespace.
It's an easy way to drop 121 lines of code, and keeps the coding style
of our sources a bit tigther.
Files which are installed as-is (any .service and other unit files, .conf
files, .policy files, etc), are left as is. My assumption is that SPDX
identifiers are not yet that well known, so it's better to retain the
extended header to avoid any doubt.
I also kept any copyright lines. We can probably remove them, but it'd nice to
obtain explicit acks from all involved authors before doing that.
Otherwise, network interfaces can be "moved" into the container's
namespace while it's still the same as the host namespace, in which case
e.g. host0 for a veth ends up on the host side instead of inside the
container.
Regression introduced in 0441378080.
Fixes#8599.
This macro will read a pointer of any type, return it, and set the
pointer to NULL. This is useful as an explicit concept of passing
ownership of a memory area between pointers.
This takes inspiration from Rust:
https://doc.rust-lang.org/std/option/enum.Option.html#method.take
and was suggested by Alan Jenkins (@sourcejedi).
It drops ~160 lines of code from our codebase, which makes me like it.
Also, I think it clarifies passing of ownership, and thus helps
readability a bit (at least for the initiated who know the new macro)
Let's remove a number of synchronization points from our service
startups: let's drop synchronous match installation, and let's opt for
asynchronous instead.
Also, let's use sd_bus_match_signal() instead of sd_bus_add_match()
where we can.
This renames wait_for_terminate_and_warn() to
wait_for_terminate_and_check(), and adds a flags parameter, that
controls how much to log: there's one flag that means we log about
abnormal stuff, and another one that controls whether we log about
non-zero exit codes. Finally, there's a shortcut flag value for logging
in both cases, as that's what we usually use.
All callers are accordingly updated. At three occasions duplicate logging
is removed, i.e. where the old function was called but logged in the
caller, too.
Add a new option `--network-namespace-path` to systemd-nspawn to allow
users to specify an arbitrary network namespace, e.g. `/run/netns/foo`.
Then systemd-nspawn will open the netns file, pass the fd to
outer_child, and enter the namespace represented by the fd before
running inner_child.
```
$ sudo ip netns add foo
$ mount | grep /run/netns/foo
nsfs on /run/netns/foo type nsfs (rw)
...
$ sudo systemd-nspawn -D /srv/fc27 --network-namespace-path=/run/netns/foo \
/bin/readlink -f /proc/self/ns/net
/proc/1/ns/net:[4026532009]
```
Note that the option `--network-namespace-path=` cannot be used together
with other network-related options such as `--private-network` so that
the options do not conflict with each other.
Fixes https://github.com/systemd/systemd/issues/7361
If user namespacing is used, let's make sure that the root user in the
container gets access to both /sys/fs/cgroup/systemd and
/sys/fs/cgroup/unified.
This matches similar logic in cg_set_access().
In -U mode we might need to re-chown() all files and directories to
match the UID shift we want for the image. That's problematic on fat
partitions, such as the ESP (and which is generated by mkosi's
--bootable switch), because fat of course knows no UID/GID file
ownership natively.
With this change we take benefit of the uid= and gid= mount options FAT
knows: instead of chown()ing all files and directories we can just
specify the right UID/GID to use at mount time.
This beefs up the image dissection logic in two ways:
1. First of all support for mounting relevant file systems with
uid=/gid= is added: when a UID is specified during mount it is used for
all applicable file systems.
2. Secondly, two new mount flags are added:
DISSECT_IMAGE_MOUNT_ROOT_ONLY and DISSECT_IMAGE_MOUNT_NON_ROOT_ONLY.
If one is specified the mount routine will either only mount the root
partition of an image, or all partitions except the root partition.
This is used by nspawn: first the root partition is mounted, so that
we can determine the UID shift in use so far, based on ownership of
the image's root directory. Then, we mount the remaining partitions
in a second go, this time with the right UID/GID information.
If we operate on a disk image (i.e. --image=) then it's pointless to
look into the mount directory before it is actually mounted to see which
systemd version is running inside...
Unfortunately we only mount the disk image in the child process, but the
parent needs to know the cgroup mode, hence add some IPC for this
purpose and communicate the cgroup mode determined from the image back
to the parent.
When "-U" is used we look for a UID range we can use for our container.
We start with the UID the tree is already assigned to, and if that
didn't work we'd pick random ranges so far. With this change we'll first
try to hash a suitable range from the container name, and use that if it
works, in order to make UID assignments more likely to be stable.
This follows a similar logic PID 1 follows when using DynamicUser=1.
