This affects attrd_updater and crm_attribute.

See https://github.com/ClusterLabs/pacemaker/pull/3277#issuecomment-1846224118.

pcmk__node_attr_target() looks up environment variables that don't exist:

• CRM_meta_container_attribute_target instead of OCF_RESKEY_CRM_meta_container_attribute_target
• CRM_meta_physical_host instead of OCF_RESKEY_CRM_meta_physical_host

OpenStack RAs work because they call ocf_attribute_target(). This avoids calling an attribute-update command with no node set.

ocf_attribute_target() {
    if [ x$1 = x ]; then
        if [ x$OCF_RESKEY_CRM_meta_container_attribute_target = xhost -a x$OCF_RESKEY_CRM_meta_physical_host != x ]; then
            echo $OCF_RESKEY_CRM_meta_physical_host
        else
            if [ x$OCF_RESKEY_CRM_meta_on_node != x ]; then
                echo $OCF_RESKEY_CRM_meta_on_node
            else
                ocf_local_nodename
            fi
        fi 
        return
...

That was added 6 years ago by https://github.com/ClusterLabs/resource-agents/commit/708e11c1.

That was very shortly after the equivalent of pcmk__node_attr_target() was added to Pacemaker:

• https://github.com/ClusterLabs/pacemaker/commit/cf34f4c95
• https://github.com/ClusterLabs/pacemaker/commit/ccbdb2a2

To me it seems likely that pcmk__node_attr_target() has never worked for container-attribute-target="host". It might not have been tested properly when those commits were added, and then ocf_attribute_target() masked the issue in resource agents.

We can either

1. Fix pcmk__node_attr_target() so that it uses the correct environment variables (as a failsafe in case an RA author forgets to use ocf_attribute_target()), or
2. Simply drop all or part of pcmk__node_attr_target().
   • This part has probably never worked.
   • ocf_attribute_target()'s logic is identical to what *the entirety of* pcmk__node_attr_target() is supposed to do, and more (OCF_RESKEY_CRM_meta_notify_all_uname).
   • The fix would have no effect on existing container images, since the call to pcmk__node_attr_target() happens on the container. It would only affect container images built with a new version of Pacemaker. So RAs will have to continue using ocf_attribute_target() for the foreseeable future, as a container might be running an older Pacemaker version.
   • Under normal circumstances, those environment variables should never be set anyway outside of an RA execution by Pacemaker. So they don't play a role in CLI execution of attrd_updater/crm_attribute on a container outside of an RA.
   • It's unusual for Pacemaker's C code to check OCF_RESKEY variables. Those are typically consumed only by OCF resource agents. Here, Pacemaker sets them for RAs and then checks them when RAs call back into Pacemaker. (There is one other case, when crm_node checks OCF_RESKEY_CRM_meta_on_node -- although that looks like it's used only as a small optimization and could be dropped if desired.)

In short, pcmk__node_attr_target() is basically redundant code at this point. It's only relevant within an RA on a container, and ocf_attribute_target() handles that case.