On Mon, 02.03.15 20:32, Hoyer, Marko (ADITG/SW2) (mho...@de.adit-jv.com) wrote:
> > Why would you need this? Watchdog is to prevent system being stuck > > somewhere. If activation fails within TimeoutStartSec=, systemd will > > put the service in "failed to activate" state anyways. > > > > Is waiting 20 seconds to detect the freeze is too much for your case? > > Is it not possible to lower the activation time? > > Yes it is too long. The process is a kind of application starter and observer > processing a startup phase. It is responsible for: > - observing the application internal states of upcoming applications > - it kicks off the startup of applications depending on the internal state of > other once > - it delays the startup of late services started the normal way by systemd > once the startup phase is done > > And yes, one could say that we are implementing a kind of second > systemd started by systemd. The difference is that our starter knows > about some more states than just ACTIVATING and RUNING which is not > really realizable with systemd especially when more than one > "application" is contained in one process. > > So the idea was to bring up a set of applications with our starter > application and stay with our starter in state ACTIVATING until it > is done with bringing up the set of applications. > > Depending on the product, bringing up our set of applications can > take 10-20secs. Since the starter is a kind of sensible application, > it needs to be supervised by a watchdog with a timeout far less than > 20secs. > > Hope this gives a rough picture of our use case. So, I can see that having watchdog support during the activating phase might make sense in this case, but I am not sure this case is strong enough to add it to systemd proper, since it would complicate things for most: it's not a behaviour we can just enable for everybody, it would have to be something we'd have to add an explicit opt-in option for, since most daemons don't work like this. I mean, in most cases the start-up phase of a daemon is relatively quick, and the daemon will not enter its event loop before the start-up is complete. But it is the event loop that normally needs to process watchdog events. Anyway, I'd prefer not adding support for this now. We can revisit this if more folks are asking for this, and this turns out to be a more common setup. I hope that makes sense, Lennart -- Lennart Poettering, Red Hat _______________________________________________ systemd-devel mailing list systemd-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/systemd-devel
