When we send a batch of nfnl messages, but e.g. without sufficient
privilege, the kernel may only return an error message for
NFNL_MSG_BATCH_BEGIN and ignore all later messages.
So, we need to read the response for the NFNL_MSG_BATCH_BEGIN,
and if it is an error ignore the replies for the rest.
When we send a message with NLM_F_ACK, but if later we are not
interested in the reply and do not want to call sd_netlink_read(),
the reply will be stored in the rqueue forever.
Let's introduce a way to ignore received message without waiting reply.
When sd-event is attached to the sd-netlink object, use the timestamp of
the current event, otherwise call now().
This mostly change nothing. Preparation for later change.
… and also when it fails, in the few cases where this wasn't done already.
When I was debugging https://github.com/systemd/systemd/issues/39048, it
was quite confusing that we print copious information about D-Bus messages
and all other steps we're undertaking, but nothing about the sd-notify
messages which are much more important. Add some debug messages for the
cases where we didn't print anything. (The case where $NOTIFY_SOCKET is
not set in left unchanged. The variable is often left unset and we don't
need to spam logs in that trivial case.)
gcry_mpi_t is defined as "typedef struct gcry_mpi *gcry_mpi_t;".
When const is applied to this type, it resolves to
"struct gcry_mpi *const" instead of what we expect ("const struct gcry_mpi *").
So we end up with a const pointer to a mutable object instead of a mutable
pointer to a const object. Since the pointer passed to the function
is copied regardless, making it const has zero benefit.
You'd think we could instead stop using gcry_mpi_t and replace it with
"const struct gcry_mpi *", except that gcrypt leaked this mess into its
api, so it expects const pointers to mutable objects as well, which means
we can't take pointers to const objects as arguments, as we'd discard the qualifier
when calling a gcrypt function.
To avoid confusion, let's drop the const qualifiers from the gcry_mpi_t arguments.
Latest glibc uses _Generic to have strstr() and other functions return
const char* or char* based on whether the input is a const char* or a
char*. This causes build failures as we previously always expected a char*.
Let's fix the compilation failures and add our own macros similar to glibc's
to have string functions that return a mutable or const pointer depending on
the input.
Instead of failing the event loop with a generic EXIT_FAILURE
error code when exit-on-disconnect is used, let's propagate the
error code instead of swallowing it.
Whereas previously sd_event_loop() would always fail with exit code
'1' when exit-on-disconnect is used with an sd-bus instance registered
with the event loop that encounters a failure, now we'll correctly
propagate the error to sd_event_loop() that caused sd-bus to fail and
exit the event loop. Additionally, the error is now also properly
propagated to outstanding reply callbacks for async dbus calls started
with sd_bus_call_async() and friends, whereas before we always used
ETIMEDOUT for these calls which is extremely confusing for users.
Why is this confusing? We always start sd-bus instances asynchronously,
in other words, sd_bus_start() will not actually wait until the bus instance
is connected, but it'll happen in the background, either driven by the first
sd_bus_call() when there is no event loop or by sd-event when there is an
event loop attached to the sd-bus instance. Assuming an event loop is attached,
when we fail to connect to the bus, the sd-bus instance will close down and the
first async method call we queued will fail with ETIMEDOUT. Nowhere in this process
do we inform the user that we failed to connect to the bus because of e.g. a permission
error, except for a debug log message.
By propagating the error to sd_event_exit() if exit-on-disconnect is enabled
and always propagating it to outstanding reply callbacks, debugging failures
becomes much easier as users will now get the actual error code causing the
bus instance to close down instead of ETIMEDOUT and 1 respectively.
We have dropped cgroup v1 support in v258. Let's drop legacy code.
Then, we can drop 'controller' argument from cg_pid_get_path() and
cg_pidref_get_path().
Reverts systemd/systemd#38680
After taking a closer look I'm not convinced by the approach, see below.
First of all, all other SD_PATH_SEARCH_* are either somewhat generic,
i.e. encode the common prefix for configurations, binaries, etc., or are
subdirectories under systemd/ hence in our own "domain". The
tmpfiles/sysctl/binfmt we don't prefix with "systemd" precisely because
the concept is generic and there're actually other impls of them. A
specific SD_PATH_SEARCH_SYSCTL doesn't fit into our existing scheme.
Instead something along the lines of "SEARCH_SYSTEM_CONFIGURATION" shall
be introduced, and consumers will just suffix
sysctl.d/tmpfiles.d/binfmt.d for the final result.
And secondly, I don't grok why systemd-sysctl now unnecessarily calls
into sd-path to obtain the fixed search path. None of our other tools do
that.
-----------
An alternate approach, SD_PATH_SYSTEM_SEARCH_CONFIGURATION, which does
exactly above, will be introduced instead. It provides a universal
interface for querying any system config with our idiomatic
/etc/:/run/:/usr/local/lib/:/usr/lib/ hierarchy.
Its value depends on how strerror_r() implemented, and the
implementations of the function in glibc and musl are actually
different. Let's drop the checks.
Defer and exit event sources are marked pending once when they are added
and never again afterwards. This means their pending_iteration is never
incremented after they are initially added, which breaks fairness among
event sources with equal priority which depend on the pending_iteration
variable getting updated in source_set_pending(). To fix this, let's assign
iterations for defer and exit sources in source_dispatch() instead so that
those get their pending_iteration updated as well.
Sometimes it's hard to assign responsibility to a specific event source
for exiting when there's no more work to be done. So let's add exit-on-idle
support where we exit when there are no more event sources.
As documented in sd_event_add_memory_pressure(), we can only add
the memory pressure fd to epoll once we've written the watch string,
so make sure we don't register the memory pressure in
event_source_online() until we've written the watch string.
More post event sources might get added during dispatching, we want
to make sure those become pending as well if we're dispatching a non-post
event source.
If we're switching users but not entering a container, then we can
assume that new switches for systemd-stdio-bridge are available, so
make use of them in that case.
This provides functionality to replace what was provided by the preceding
revert:
$ build/systemd-path system-search-configuration --suffix=sysctl.d
/etc/sysctl.d:/run/sysctl.d:/usr/local/lib/sysctl.d:/usr/lib/sysctl.d
The result is identical, but more generic, since by changing suffix we can also
get the answer for sysusers.d, tmpfiles.d, and any other of the directories
which follow the same general rule.
David Tardon
Zbigniew Jędrzejewski-Szmek
Yu Watanabe
Luca Boccassi
Daan De Meyer
Frantisek Sumsal
Jelle van der Waa
Lennart Poettering
Mike Yuan