This reworks how we reply to D-Bus messages that come in for resolution
requests. Previously, we'd store them in the .bus_request field of
the main DnsQuery (but not any auxiliary one), and reply to it whenever
we had something to reply. In error paths this could mean we'd
accidentally reply twice.
This cleans this logic up: whenever we reply to a message we'll now go up
the tree of auxiliary queries, to find the primary query, i.e. the one
we actually want to reply to. Once we found it, we take out the bus
message object, resetting it to NULL. This way we can be sure we'll
reply at most once to each message.
Fixes: #22477
There's currently a deadlock between PID 1 and dbus-daemon: in some
cases dbus-daemon will do NSS lookups (which are blocking) at the same
time PID 1 synchronously blocks on some call to dbus-daemon. Let's break
that by setting SYSTEMD_NSS_DYNAMIC_BYPASS=1 env var for dbus-daemon,
which will disable synchronously blocking varlink calls from nss-systemd
to PID 1.
In the long run we should fix this differently: remove all synchronous
calls to dbus-daemon from PID 1. This is not trivial however: so far we
had the rule that synchronous calls from PID 1 to the dbus broker are OK
as long as they only go to interfaces implemented by the broke itself
rather than services reachable through it. Given that the relationship
between PID 1 and dbus is kinda special anyway, this was considered
acceptable for the sake of simplicity, since we quite often need
metadata about bus peers from the broker, and the asynchronous logic
would substantially complicate even the simplest method handlers.
This mostly reworks the existing code that sets SYSTEMD_NSS_BYPASS_BUS=
(which is a similar hack to deal with deadlocks between nss-systemd and
dbus-daemon itself) to set SYSTEMD_NSS_DYNAMIC_BYPASS=1 instead. No code
was checking SYSTEMD_NSS_BYPASS_BUS= anymore anyway, and it used to
solve a similar problem, hence it's an obvious piece of code to rework
like this.
Issue originally tracked down by Lukas Märdian. This patch is inspired
and closely based on his patch:
https://github.com/systemd/systemd/pull/22038Fixes: #15316
Co-authored-by: Lukas Märdian <slyon@ubuntu.com>
A first step of removing blocking calls to the D-Bus broker from PID 1.
There's a lot more to got (i.e. grep src/core/ for sd_bus_creds
basically), but it's a start.
Removing blocking calls to D-Bus broker deals systematicallly with
deadlocks caused by dbus-daemon blocking on synchronous IPC calls back
to PID1 (e.g. Varlink calls through nss-systemd). Bugs such as #15316.
Also-see: https://github.com/systemd/systemd/pull/22038#issuecomment-1042958390
Previously we'd only watch configured service bus names if Type=dbus was
set. Let's also watch it for other types. This is useful to pick up the
main PID of such a service. In fact the code to pick it up was already
in place, alas it didn't do anything given the signal was never received
for it. Fix that.
(It's also useful for debugging)
sd-bus for compat with old C expects "int" for booleans. Internally we
prefer to use C99 "bool" however. When pointing a pointer to it, we need
to handle the fact that they are typically differently sized, hence use
te right accessors for it.
The value is used as a boolean, but stored in an unsigned. It's exposed
with sd-bus default boolean handling which however expects a (signed)
int. Let's clear this up, and use "bool" for this everywhere in
non-local scope, and make sure that when we pass it to sd-bus we
properly pass an "int".
This reverts commit 0ad4efb14b.
See
https://github.com/systemd/systemd/pull/20559#issuecomment-1028011030
for reasoning. Quoting:
> I think it should be OK to advertise extra stub listeners to local
> clients, but you prohibit this now. i.e. there are two different
> concepts here, and we shouldn't mix them up:
>
> 1. tracking configured dns servers and advertise them to local programs
> 2. actually using them ourselves
>
> I am pretty sure that our own stubs are OK for 1 but not OK for 2,
> hence we should filter at the time of use not at the time of parse.
Fixes regression introduced in 599be274c1
Moving the Xen check before the CPUID check, in order to handle the case where
a Xen domain is nested within a hypervisor which can be detected by via the
CPUID check, had an unintended consequence of causing Dom0 to report as a Xen
VM when it is not nested.
This patch stops further checks once it has been determined that Dom0 is not
nested within another hypervisor, meaning that the non-nested case matches its
previous logic (where it does not report as a VM).
Also, tidy the conditionals for the Xen and UML checks by removing handling of
a VIRTUALIZATION_VM_OTHER result, which has no code path.
Fixes#22511
We would print "Setting NUMA policy to bind, with nodes .".
This is not very clear, change it to "… with nodes {}.".
Also use range formatting for masks to make output shorter.
For https://bugzilla.redhat.com/show_bug.cgi?id=1986176:
if we are trying to reexecute, and this fails for any reason, we shouldn't
try to execute /sbin/init or /bin/sh. It is better to just freeze.
If we freeze it is easier to diagnose what happened, but if we execute
one of the fallbacks, we don't really know what will happen. In particular
the new init might just return, causing the machine to shut down. Or we
may successfully spawn /bin/sh, which could leave the machine open.
If manager_loop() fails, we would print an error message, but then actually
ignore the error in main(), and potentially execute the shutdown binary.
I'm not sure how likely this is to happen in practice, but it seems sloppy.
So let's do the cleanup, but actually freeze() if manager_loop() returned
an error.
invoke_main_loop() is refactored to return the manager objective. This way
we don't need to pass a separate parameter to specify whether we are
reexecuting. Subsequent patch will make further use of the returned objective.
Jan Janssen
Lennart Poettering
Yu Watanabe
Luca Boccassi
Richard Neill
Zbigniew Jędrzejewski-Szmek