vmspawn systems might take quite a while to boot in particular if they
go through uefi and wait for a network lease. Hence let's increase the
start timeout to 2min (from 45s). We'll do that for both nspawn and
vmspawn, even though the UEFI thing certainly doesn't apply there (but
the DHCP thing still does).
This mimics the switch of the same name from nspawn: it controls whether
we expect a READY=1 message from the payload or not. Previously we'd
always expect that. This makes it configurable, just like it is in
nspawn.
There's one fundamental difference in behaviour though: in nspawn it
defaults to off, in vmspawn it defaults to on. (for historical reasons,
ideally we'd default to on in both cases, but changing is quite a compat
break both directly and indirectly: since timeouts might get triggered).
This beefs up the cgroup logic, adding --slice=, --property= to vmspawn
the same way it already exists in nspawn.
There are a bunch of differences though: we don't delegate the cgroup
access in the allocated unit (since qemu wouldn't need that), and we do
registration via varlink not dbus. Hence, while this follows a similar
logic now, it differs in a lot of details.
This makes in particular one change: when invoked on the command line
we'll only add the qemu instance to the allocated scope, not the vmspawn
process itself (this follows more closely how nspawn does this where
only the container payload has its scope, not nspawn itself). This is
quite tricky to implement: unlike in nspawn we have auxiliary services
to start, with depencies to the scope. This means we need to start the
scope early, so that we know the scope's name. But the command line to
invoke is only assembled from the data we learn about the auxiliary
services, hence much later. To addres we'll now fork off the child that
eventually will become early, then move it to a scope, prepare the
cmdline and then very late send the cmdline (and the fds we want to
pass) to the prepared child, which then execs it.
Just like in nspawn, there's a chance we need to PK authenticate the
registration, hence let's spawn off the agent for that during that
phase, and terminate it once we don't need it anymore.
Let's also send out a STATUS= message when we get READY=1 if it didn't
come with a STATUS= message itself.
Also, let's initially say the container is "started", and only once the
READY=1 is seen claim it was "running".
This cleans up allocation of a scope unit for the container: when
invoked in user context we'll now allocate a scope through the per-user
service manager instead of the per-system manager. This makes a ton more
sense, since it's the user that invokes things after all. And given that
machined now can register containers in the user manager there's nothing
stopping us to clean this up.
Note that this means we'll connect to two busses if run unpriv: once to
the per-user bus to allocate the scope unit, and once to the per-system
bus to register it with machined.
So far, machined strictly tracked the "leader" process of a machine,
i.e. the topmost process that is actually the payload of the machine.
Its runtime also defines the runtime of the machine, and we can directly
interact with it if we need to, for example for containers to join the
namespaces, or kill it.
Let's optionally also track the "supervisor" process of a machine, i.e.
the host process that manages the payload if there is one. This is
generally useful info, but in particular is useful because we might need
to communicate with it to shutdown a machine without cooperation of the
payload. Traditionally we did this by simply stopping the unit of the
machine, but this is not doable now that the host machined can be used
to track per-user machines.
In the long run we probably want a more bespoke protocol between
machined and supervisors (so that we can execute other commands too,
such as request cooperative reboots/shutdowns), but that's for later.
Some environments call the concept "monitor" rather than "supervisor" or
use some other term. I stuck to "supervisor" because nspawn uses this,
and ultimately one name is as good as another.
And of course, in other implementations of VM managers of containers
there might not be a single process tracking each VM/container. Because
of this, the concept of a supervisor is optional.
The difference between these two operations are large: one is relatively
superficial: for "registration" all resources remain associated with the
invoking user, only the cgroup is reported to machined which then keeps
track of the machine, too. OTOH "creation" a scope is allocated in
system context, hence the invoked code will be owned by the system, and
its resource usage charged against the system.
Hence, use two distinct polkit actions for this, so that we can relax
access to registration, but keep access to creation tough.
Now that unpriv clients can register machines, let's register their UID
too. This allows us to do two things:
1. make sure the scope delegation is assigned to the right UID (so that
the unpriv user can actually create cgroups below the delegated
scope)
2. permit certain types of access (i.e. killing, or pty access) to the
client without auth if it owns the machine.
This new helper takes both a PID and and a pidfd ID, and initializes a
PidRef from it. It ensures they actually belong together and returns an
error if not.
Continuation of #37960.
The same concern as expalined in #37960 exists also in
missing_syscall.h. If we use enough new glibc, a function we want to use
may be already provided by glibc, but our baseline glibc may not. And it
is hard to detect in our daily development.
This moves all prototypes of syscalls to relevant headers, and missing
syscall functions are defined in relevant .c files of libc wrapper. This
way, we can use usual header as is, e.g. when we want to write code with
`move_mount()`, we can simply use sys/mount.h without checking if it is
supported by our baseline glibc.
This introduces `struct ConfFile` that stores detailed information of an
enumerated file, and introduces `conf_files_list_full()` and friends
that provide results in `ConfFile`.
Then make udev, hwdb, catalog, and cat-files use the new function and
struct to make them not read files outside of specified root directory.
Then, move syscall definitions to the wrapper, and prototypes are moved
to relevant headers.
This also adds checks for add_key() and request_key(), as one day
glibc may be going to add some of them separatedly.
The check for fspick in meson.build is dropped, as it is currently
unused in our code.
This also moves
- basic/missing_bpf.h -> include/override/linux/bpf.h,
- basic/missing_keyctl.h -> include/override/linux/keyctl.h.
This also moves syscall tables and generators to the same directory.
Note, inclusion of asm/sgidefs.h is dropped, as it is already included
by unistd.h and sys/syscall.h.
- pass our system include directories to make generators use our libc
wrappers and latest kernel headers,
- include relevant headers in generated gperf file,
- use files() rather than find_program(), as the result of
find_program() cannot be passed to 'input' of custom_target(),
- move generate-bpf-delegate-configs.py to src/core/, as it is only used
by libcore.
Previously, the logger is only used in error paths, but since
fe3f2ac073, the logger is also used in a
success path. Let's not log loudly on success.
This also drops unused log_selinux_enforcing().
Then, make the function show the original and resolved path if they are
different.
With this change, procfs needs to be mounted on /proc/, hence the test
code is slightly updated.
- drop redundant error messages in cat_files(), as cat_file() internally
logs errors,
- show an empty line and filename before opening file, to make not mix
any error messages with the previous file,
- drop unnecessary fflush(),
- use RET_GATHER() and continue to show files even if some files cannot
be shown.
Lennart Poettering
Daan De Meyer
Yu Watanabe