Talk to the journal ourselves using sendmsg() instead of linking
against libsystemd for sd_journal_sendv(). At the same time, we
can also avoid excessive copying.
The motivation for dropping the dependency is that we can
then use structured logging e.g. in a flatpak sandbox where
libsystemd may not be present in the runtime.
The code here is inspired by similar code in libvirt.
Currently the docs for GChecksumType are simpy annotated
with 'Since 2.16' which is when GChecksumType was first
introduced. No mention is made of the fact that the
G_CHECKSUM_SHA512 constant was only added much later
in 2.36.
Signed-off-by: Daniel P. Berrange <berrange@redhat.com>
https://bugzilla.gnome.org/show_bug.cgi?id=769027
It turns out that the current approach of parsing g_log_structured
varargs is unworkable, because vprintf is not guaranteed to advance
the passed-in va_list. So, we have to reshuffle the argument list
a bit; I've come up with this approach:
g_log_structured (domain, level,
key-value pairs...
"MESSAGE", format,
printf arguments);
This requires a "MESSAGE" key to always be present, and it requires
the "MESSAGE"-format pair to be last, but it avoids an extra NULL
as marker after the key-value pairs. And it can be parsed with a
single pass over the va_list, without any va_copy.
Since we have G_LOG_USE_STRUCTURED, the separate ...structured()
convenience macros are pretty pointless, and I have dropped them
for now.
Look for a macro G_LOG_USE_STRUCTURED, and if it is defined, use
g_log_structured instead of g_log when defining g_warning and friends.
This avoids the extra complication of going through g_logv _and_
g_log_structured to get a message logged; it also lets us pass
the code-related fields.
We don't do this unconditionally (yet), since some users might
rely on the more fine-grained fatality support in g_logv. It has
also been proven problematic in the past to inject a dependency
on bleeding-edge API via a widely-used macro.
https://bugzilla.gnome.org/show_bug.cgi?id=744456
g_socket_listener_add_address() is synchronous; all of the events will
have been emitted before it returns and it doesn't queue any sources.
The test was unintentionally depending on the fact that
g_main_context_iterate(NULL, TRUE) would return anyway (at least the
first time it was called), but that's no longer true after e4ee307.
https://bugzilla.gnome.org/show_bug.cgi?id=768968
GCC fails to build because of the trailing arguments, not part of the
format:
../../glib/gmessages.c: In function 'g_log_default_handler':
../../glib/gmessages.c:2385:21: error: too many arguments for format
[-Werror=format-extra-args]
NULL);
^
The documentation for `__attribute__((format(...)))` in GCC
https://gcc.gnu.org/onlinedocs/gcc/Common-Function-Attributes.html#Common-Function-Attributes
States that the second index must be 0 for functions that are not
available to be checked, like for vprintf-style functions. In this case
it's also appropriate because of the trailing arguments.
The sd-journal API in systemd, upon which the structured logging API is
modelled, also uses 0 as the second argument for the format attribute.
https://bugzilla.gnome.org/show_bug.cgi?id=744456
It turns out that g_info_structured (format, ...) makes
g_info_structured ("Hey!") not work, since it requires at
least one argument after the format string. So shorten
the argument list to just ...
