[WIP] Modernize start method

[wenningerk]

As the name says focus is bringing sbd from a forking model to either 'simple' or 'notice' start-model in systemd. Discussion started off in #134. PR atm still contains changes from #134 - will be rebased once #134 is merged.

[wenningerk]

Continuing discussion started in #134:

I wasn't thinking of putting anything pacemaker-related in sbd's public API, since that's just for sbd's internal use. The public API would just have low-level stuff for external code to check on sbd. It should be segregated in a separate directory and not use any pacemaker APIs.

That wasn't my point. sbd already has pacemaker as build-time requirement. So building a library as part of sbd-build that pacemaker is supposed to build against creates some kind of build-order-issue.

Ah right, of course. Hmm.

Isolating the pacemaker-related stuff into the pacemaker project or even its own project could be worthwhile. I believe at one point we had discussed making the interface more generic so the corosync/pacemaker stuff wasn't so tightly integrated. I'm not sure if that would open any demand for other sbd "plugins" but it could be interesting. The question is whether it's worth the effort ...

Discussion regarding integration of the components was controversal as well suggesting tight integration of sbd into corosync. We do have 3 types of watchers in sbd - disk, cluster & pacemaker How sbd is handling those is quite different and interwoven. But maybe trying to find some generic abstraction could as well be helpful to make sbd-code more understandable. Finding an abstraction for the watchdog that can be implemented as a plugin maybe interesting as well although it isn't relevant for the build-order thing.

[gao-yan]

Type=forking is a so-called traditional way, but it doesn't necessarily mean it's legacy or bad? Especially for handling ordered startups of sbd and pacemaker in our case, it's probably one of the most suitable ways...

Although we now have startup synchronization between sbd and pacemaker, but if to switch to Type=simple, AFAICS SBD_SYNC_RESOURCE_STARTUP=yes being enabled/defaulted would have to be required to get the correct startup behavior, not to mention that would be another trap for upgrade of existing setups where there might be the explicit setting SBD_SYNC_RESOURCE_STARTUP=no in sysconfig ...

Type=notify might be another option, but that'd require to introduce dependencies on systemd/dbus for sbd. So eventually it'd be more of a question whether that'd have real benefits so that it'd be worth it ...

[wenningerk]

Type=simple was just a first try and it became obvious quite quickly that Type=notify would probably be the way to go. Guess we should still be able to build without systemd or dbus but on the other hand on a system under systemd control using the benefits of it wouldn't be bad. If sbdy insists on building sbd without library-dependency to dbus or systemd we could still fall back to calling sd_notify-tool from the C-code. And fiddling around with pid-files and where the system prefers them to currently be isn't that elegant either. Furthermore checking for a pid-file from pacemaker is in my eyes anyway a weak point of the whole synchronization. I'd definitely prefer asking systemd if sbd is enabled or running and leave the current code just as a fallback - weather to be controlled via cmdline and/or compile-switch is open for discussion. Anyway I guess there are a couple of points her to discuss and we shouldn't rush it.

[gao-yan]

Type=simple was just a first try and it became obvious quite quickly that Type=notify would probably be the way to go. Guess we should still be able to build without systemd or dbus but on the other hand on a system under systemd control using the benefits of it wouldn't be bad. If sbdy insists on building sbd without library-dependency to dbus or systemd we could still fall back to calling sd_notify-tool from the C-code.

Yes, it'd for sure make sense to have a build detection/option.

And fiddling around with pid-files and where the system prefers them to currently be isn't that elegant either. Furthermore checking for a pid-file from pacemaker is in my eyes anyway a weak point of the whole synchronization. I'd definitely prefer asking systemd if sbd is enabled or running and leave the current code just as a fallback

Good point.