The code in service_spawn() was written as if exec_fd_event_source
was always unset. (We would either fail the assertion that is moved in the
patch, or leak the event source object if it was set.)
To make this work, let's always assert that exec_fd_event_source is unset,
and actually unset it service_sigchld_event(). I think this is the most
elegant approach. The problem is that we don't have the same information
about execution flags as in service_spawn(), so we need to conditionalize
on pid==main_pid to know if we should disable exec_fd_event_source.
I think this matches all cases where we may set exec_fd_event_source:
service_enter_start() and service_run_next_main().
service_enter_stop_post() calls service_set_state(), which will also destroy
the source. But that happens too late, because from service_enter_stop_post()
we call service_spawn() first, and then service_set_state() second.
(An alternative approach would be to deallocate the existing
exec_fd_event_source in service_spawn(). But this would mean that we would
temporarily have an event source attached to a process that we already know is
dead, which seems less than ideal.)
Original report from Dimitri John Ledkov <dimitri.ledkov@canonical.com>:
> Ubuntu private bug reference for this issue at the moment is
> https://bugs.launchpad.net/ubuntu/+source/systemd/+bug/1921145
> Michael's and Ian's team run into an issue when using systemd in the
> initrd, without dbus daemon running, and launching a unit in a
> particular way that appears to lock up systemd (pid 1) it self.
> michael vogt: "The attached script works for me to reproduce this on
> classic. I tested 20.04 (245) and 21.04 (247) in a qemu VM. Sometimes
> I need to run it multiple times but usually it crashes after at most 2
> runs. Use "journalctl | tail" to see the messages, it's the same that
> Ian reported. There is also a /var/crash/_usr_lib_systemd_systemd
> crash file created."
> I understand that the particular way to run a unit is very odd,
> however, it is currently possible to invoke, and it would be expected
> for pid1 to not lock up and crash.
> The Assertion that systemd hits is along the lines of:
> [ 10.182627] systemd[1]: Assertion 's' failed at
> src/core/service.c:3204, function service_dispatch_exec_io().
> Aborting.
> [ 10.195458] systemd[1]: Caught <ABRT>, dumped core as pid 449.
> [ 10.204446] systemd[1]: Freezing execution.
Apparently CAN links will show up in rtnetlink with very low MTUs. We
shouldn't consider them relevant if no IP is spoken over them, since
these MTUs are irrelevant for us then.
Hence, let's check if there's an address assigned to the link before
considering its MTU.
As additional safety net filter out MTUs smaller than the minimum DNS
packet size, too.
Finally, in case we don't find any suitable interface MTU, let's default
to 1500 as the generic Ethernet MTU.
Fixes: #19396
This was a copy/paste mistae apparently, there's not "try_authtok" and
this was supposed to copy what Fedora uses, which uses "use_authtok"
correctly. Hence adjust this.
Fixes: #19369
This mirrors the change done for systemd-resolved in
9793530228. Quoting that patch:
> We generally operate on the assumption that a source is "gone" as soon as we
> unref it. This is generally true because we have the only reference. But if
> something else holds the reference, our unref doesn't really stop the source
> and it could fire again.
In particular, we take temporary references from sd-event code, and when called
from an sd-event callback, we could temporarily see this elevated reference
count. This patch doesn't seem to change anything, but I think it's nicer to do
the same change as in other places and not rely on _unref() immediately
disabling the source.
This drops the "const" specifier from the opaque object parameters to
various functions in our API.
This effectively reverts #19292 and more.
Why drop this? Our public APIs should not leak too much information
about how stuff is implemented internally. In our public APIs we
shouldn't give too many guarantees we don#t want to necessarily keep.
Specifically: in many cases it makes sense that getters actually
generate/parse/allocate data on the fly, storing/caching the result
internally, to speed things up, do things lazily or to track memory
allocations so that they can be freed later. Doing this means we need to
change the objects, even though the getters are semantically a read
operation.
We want to retain the freedom that we can change things around
internally. By exposing the objects as "const" we remove a good chunk of
that, for little gain.
See sd_bus_creds_get_description() for a real example of a getter that
implicitly caches and thus modifies the relevant object.
This removes the "const" decorators from sd-dhcp and sd-netlink, two
APIs that we intend to make public eventually even though they still are
not, leaving us the chance to still fix this before it becomes set in
stone.
Why is this necessary? Several examples below.
- When a route sets prefsrc, then the address must be already assigned
(see issue #19285), and also it must be ready if IPv6.
- When a route or nexthop sets gateway, then the address must be reachable.
- When a route sets nexthop ID, then the corresponding nexthop must be
assigned.
- When a route sets multipath routes on another interface, then the
interface must exist and be ready to configure.
- When configuring address, the same address must not be under removing
(see issue #18108).
Etc,. etc,...
So, this makes all requests about addresses, routes, and nethops are once
stored in the queue, and will be processed when they are ready to configure.
Fixes#18108 and #19285.
We usually call specifier_printf() and then check the validity of
the result. In many cases, validity checkers, e.g. path_is_valid(),
refuse too long strings. This makes specifier_printf() refuse such
long results earlier.
Moreover, unit_full_string() and description field in sysuser now
refuse results longer than LONG_LINE_MAX. config_parse() already
refuses the line longer than LONG_LINE_MAX. Hence, it should be ok
to set the same value as the maximum length of the resolved string.
Zbigniew Jędrzejewski-Szmek
Michael Catanzaro
Lennart Poettering
Yu Watanabe