David Teigland wrote:

On Thu, Jan 26, 2012 at 08:44:00AM -0700, Jim Fehlig wrote:

...

[ resend after subscribing to list ]

Hi,

I've been playing with sanlock on SUSE distros and noticed that running vms with disk leases controlled by sanlock are killed (with unclean state on filesystems, databases, and block devices) when the sanlock daemon is terminated. Actually, I noticed this when simply updating the package, which caused a restart of the daemon.

Is there a way to control this behavior? I couldn't find anything with a quick peek at the docs and code.

Hi Jim, the init script currently uses SIGTERM to stop the daemon, which is equivalent to sanlock shutdown -f1 (forces shutdown even if the daemon is holding leases.) I think what the init script should probably do instead is shutdown -f0, so the shutdown is ignored if the daemon is holding leases (we could also change SIGTERM to be the equivalent of -f0 instead of -f1.) I think think -f0 (ignore shutdown when leases are held) is the behavior everyone would prefer from the init script stop/restart functions.

Perhaps the init script can check if the daemon is holding leases and emit a message without stopping it, otherwise shutdown as it currently does? Looking at the docs, I'm not quite sure how to do this though. I don't think it is as simple as

if $prog client status | egrep ^r > /dev/null; then echo -n "Resources controlled by sanlock, refusing to shutdown " else echo -n $"Stopping $prog: " killproc -p $lockfile $prog -TERM fi ...

We don't have the option of leaving pids/vms running after the sanlock daemon exits. Because the daemon holds their leases, the resources they are using are no longer protected when the daemon is gone. This is why the sanlock daemon kills the pids/vms before it exits. And if the daemon is SIGKILL'ed, then the watchdog comes to the rescue and resets the entire host so the pids/vms will be killed that way.

Ah, understood. Thanks for the explanation David!

Regards, Jim