Unless the systemd devs are immune to the same problems that plague literally every other software developer
Here's a couple of fun problems from years past.
Docs say NoNewPrivileges as yes "prohibits UID changes of any kind", but that's not exactly true.
It's spammed so much debug info the system fails to boot.
"systemd, when updating file permissions, allows local users to change the permissions and SELinux security contexts for arbitrary files via a symlink attack on unspecified files."
"The rm_rf_children function in util.c in the systemd-logind login manager in systemd before 44, when logging out, allows local users to delete arbitrary files via a symlink attack on unspecified files, related to "particular records related with user session."
It's complex software running on many different kernels, and many different distributions, doing many different things with lots of code written in C. There were over 80 people mentioned as contributing to the last release. Of course there are going to be problems!
Anything that requires DBUS is (at minimum, because it now depends on DBUS, which is -itself- rather complicated) substantially more complicated than a script that does
exec "$@"
Surely you can't argue with that?It may not appear to be more complex to the user, but when your software has a firm dependency on a complex system, it becomes more complex because it is affected by the failure modes, quirks, and restrictions of the software on which it depends. [0]
> Lots of init scripts used to be quite different across distributions and very long.
I've talked elsewhere in these threads about how systemd's unit files simply shift the complexity for services that need (for whatever reason) complex pre-start logic out of the unit file and into hand-maintained tools. See [1] for my most recent comment on the topic. Make sure to check the footnotes in that comment, too. :)
[0] Entirely valid anecdotal evidence on the relative stability of dbus doesn't invalidate my point. dbus is one of many complex software projects that others depend on. :)
But fine, you wanted numbers. First, systemd's LoC count is 699,342 across 1,912 files as of a git clone of their SCM[1] and gitstats five minutes ago.
That number was around 550,000 back in May[2], so kindly update your knowledge before regurgitating it mindlessly.
Debian's (sid) sysvinit package[3], by the same methodology, is 32,151 lines across 235 files.
Systemd is therefore about 22x (21.75179621162639) more complicated than sysv.
I'd go run the numbers for Upstart, but I'm not going to go download bzr and appropriate profiling tool to conclusively prove something that's common knowledge to anyone not in an internet argument.
And kindly stop complaining about downvotes - it's against the site rules.
[1]:http://anongit.freedesktop.org/git/systemd/systemd.git
[2]:http://www.phoronix.com/scan.php?page=news_item&px=MTY5NjM
[3]:https://anonscm.debian.org/git/collab-maint/sysvinit.git
Single out the init management portion of systemd (and journald since it's a dependency), and compare those numbers.
Everything else is an optional module.
The init system should be an init system and nothing else, IMO, so targeting all that extra complexity (which was my original point) is fair game.
Whether a distribution chooses to include timesyncd or another ntpd doesn't magically make timesyncd an init system.
If sysvinit also included an ntpd client, _then_ it would be appropriate to include that module as well for a fair comparison of SLOC.
Another example would be to take vim, counting the SLOC including a bunch of downloaded plugins, _then_ comparing that number against emacs without any plugins.
It's comparing apples to apples and oranges simply because they're in the same basket.
If your system gets pwned, does it matter to the user if it was via the "systemd" binary itself or by one of the official "systemd-somethingd" binaries (that have no other purpose other than interoperation with the main app, developed by the same people) that come with every OS using it? Had the distro maintainers not shoveled an extra ~670,000LoC down their throat, they wouldn't have been pwned.
If systemd wants to replace my freaking cron, DNS, ntp, and logging daemons, and the majority of in-production deployments use the application like that, I will be evaluating the application's complexity against how it's used in the real world.
sysvinit doesn't attempt to replace resolvconf, ntpd, crond, and syslog-ng. systemd does.
They have no access to systemd-init, only to an API presented by it (as does every other application if they so choose to use it), and thus this statement is incorrect. A bug in one of those applications cannot magically get access to systemd-init (which I'm assuming is what you meant by "main" here).
> If your system gets pwned, does it matter to the user if it was via the "systemd" binary itself or by one of the official "systemd-somethingd" binaries (that have no other purpose other than interoperation with the main app, developed by the same people) that come with every OS using it? Had the distro maintainers not shoveled an extra ~670,000LoC down their throat, they wouldn't have been pwned.
It absolutely matters! A bug in systemd-init can do much more damage than a bug in timesyncd could. Also, I really don't care about your anger towards an open source project, please refrain from expressing it to me if you'd like to continue this conversation.
> If systemd wants to replace my freaking cron, DNS, ntp, and logging daemons, and the majority of in-production deployments use the application like that, I will be evaluating the application's complexity against how it's used in the real world.
You may want to create a separate post somewhere which compares the SLOC of all the applications systemd attempts to re-create (choosing perhaps the most popular from each) along with the complete systemd package, but I personally wouldn't have any interest in it since it would be a factually correct comparison.
> sysvinit doesn't attempt to replace resolvconf, ntpd, crond, and syslog-ng. systemd does.
That's exactly right, but systemd-init doesn't attempt to replace those either (besides syslog-ng (journald), which I already mentioned as being fair to include with the SLOC count)*.
I hereby coin Vacri's Law: When a comment-maker announces incoming downvoting, they're aware their argument is weak and is instead trying deflect criticism through shaming others.
In any case, you're being a hypocrite. From your very first comment in this thread ("clearly done a full audit") you've been snarky, and now you're crying foul that people are being snarky back. I actually laughed out loud when you invoked the "good faith" HN guideline against Karunamon.
It's not in question that systemd is infinitely more complicated than equivalent sysv or other script-based init system. Complexity means attack surface.