Low: libpacemaker: Don't assert on injecting nonexistent nodes.

[clumens]

If crm_simulate tries to inject a node up or node fail action for a nonexistent node, it'll just assert. Instead, log an error message and return NULL.

That NULL then needs to be propagated up a couple layers to functions that actually expect and know how to handle errors.

Fixes T945

[nrwahl2]

My inclination is to fix this in an even simpler way. Currently, for --node-up, if the node doesn't exist then we simply inject it. We could do the same thing for --node-down (which also asserts) and --node-fail.

We haven't done that in the past, and there may be only little or no legitimate use case for it. But it seems at least as reasonable as ignoring the nonexistent node for --node-down and --node-up. And it seems unlikely that anyone is relying on not injecting the nonexistent node, because it's probably been asserting for a long time/forever.

To do the injection, all we have to do is set pcmk__simulate_node_config = true to node_down_cb() and node_fail_cb().

What do you think?

[nrwahl2]

To do the injection, all we have to do is set pcmk__simulate_node_config = true to node_down_cb() and node_fail_cb().

What do you think?

Actually, injecting an action that refers to a nonexistent node would run into a similar issue. There may be other commands that do too. So, two options:

1. Do it the way you're doing it currently in this PR (returning NULL and letting the caller handle it as appropriate).
2. Get rid of pcmk__simulate_node_config and treat it as always true. Inject the node if it didn't already exist. Currently we assert in those situations, so no one should be relying on old behavior. I haven't yet looked deeply into whether this would have other unintended consequences. I've only glanced at this PR; still focused on the D-Bus stuff.