Bug 1162727

> Description of problem:


I'm recently running into a repeating problem when continuously adding/removing cluster nodes.

The issue as seen from pacemaker.log looks like this:

attrd:  warning: crm_find_peer:     Node 'virt-127' and 'virt-131' share the same cluster nodeid: 3 virt-131
attrd:    error: crm_abort:     crm_find_peer: Forked child 20878 to record non-fatal assert at membership.c:394 : member weirdness


I was able to reproduce this multiple times in various different sequences of steps, but here is how did I got to the latest one:

- i'm running a cluster of 7 nodes
- stopped virt-130
- removed node virt-126
- removed node virt-131
- removed node virt-127
- added back node virt-131

The initial node ID distribution:

1    virt-122
2    virt-126
3    virt-127
4    virt-128
5    virt-129
6    virt-130
7    virt-131

The final ID distribution after all additions/removals:

1    virt-122
3    virt-131
4    virt-128
5    virt-129
6    virt-130

As you can see virt-131 has been reassigned an ID previously used by node virt-127. I believe in theory this shouldn't be an issue as the node is long gone, but something in the logs disagrees with me. Here is more context to the log lines above, taken AFTER the final ID distribution:

crmd:     info: pcmk_quorum_notification:      Membership 92: quorum retained (4)
pacemakerd:     info: pcmk_quorum_notification:      Membership 92: quorum retained (4)
crmd:     info: crm_get_peer:  Created entry ad118388-45cd-41f0-9e30-e468433632ba/0x16f6e30 for node virt-131/3 (5 total)
crmd:     info: crm_get_peer:  Node 3 is now known as virt-131
crmd:     info: peer_update_callback:  virt-131 is now (null)
crmd:     info: crm_get_peer:  Node 3 has uuid 3
crmd:   notice: crm_update_peer_state:         pcmk_quorum_notification: Node virt-131[3] - state is now member (was (null))
crmd:     info: peer_update_callback:  virt-131 is now member (was (null))
pacemakerd:     info: crm_get_peer:  Created entry 95269e93-a643-4b65-b0e7-f1d6561e5990/0x1bbfcd0 for node virt-131/3 (5 total)
pacemakerd:     info: crm_get_peer:  Node 3 is now known as virt-131
pacemakerd:     info: crm_get_peer:  Node 3 has uuid 3
pacemakerd:   notice: crm_update_peer_state:         pcmk_quorum_notification: Node virt-131[3] - state is now member (was (null))
cib:     info: cib_process_ping:      Reporting our current digest to virt-129: a376e2e985dcd9996add4737f8416e41 for 0.15.22 (0x825930 0)
cib:     info: pcmk_cpg_membership:   Joined[14.0] cib.3
cib:     info: pcmk_cpg_membership:   Member[14.0] cib.1
cib:     info: pcmk_cpg_membership:   Member[14.1] cib.3
cib:     info: crm_update_peer_proc:  pcmk_cpg_membership: Node virt-127[3] - corosync-cpg is now online
cib:     info: pcmk_cpg_membership:   Member[14.2] cib.4
cib:     info: pcmk_cpg_membership:   Member[14.3] cib.5
attrd:     info: crm_dump_peer_hash:     crm_find_peer: Node 4/virt-128 = 0x2506230 - 7455b5f9-5dbf-4699-b29c-fc95c789dbbb
attrd:     info: crm_dump_peer_hash:     crm_find_peer: Node 6/virt-130 = 0x2506370 - 59912113-f30a-4b72-b615-809c7c7f97d0
attrd:     info: crm_dump_peer_hash:     crm_find_peer: Node 5/virt-129 = 0x25062d0 - d777a0c1-aac6-4998-bbc3-662541f2a22c
attrd:     info: crm_dump_peer_hash:     crm_find_peer: Node 7/virt-131 = 0x2506440 - d1aa5661-f9d0-4bdc-93d5-3062e0b44659
attrd:     info: crm_dump_peer_hash:     crm_find_peer: Node 1/virt-122 = 0x24830f0 - ad564e5c-7241-4f3f-8d2d-c3ac28fdcfcb
attrd:     info: crm_dump_peer_hash:     crm_find_peer: Node 2/virt-126 = 0x2506610 - dff4adfe-62f0-4e4d-a3c0-5e07b012b6b7
attrd:     info: crm_dump_peer_hash:     crm_find_peer: Node 3/virt-127 = 0x2508f40 - 45e96061-3b19-4a19-9837-5586e0ef1752
attrd:  warning: crm_find_peer:     Node 'virt-127' and 'virt-131' share the same cluster nodeid: 3 virt-131
attrd:    error: crm_abort:     crm_find_peer: Forked child 20878 to record non-fatal assert at membership.c:394 : member weirdness

I do not know how pacemaker's internal peer cache works, but this snippet of a log hints it is still somehow aware of the nodes after they've been removed.

When checking runtime CIB, a CIB stored in /var/lib/pacemaker, corosync quorum and corosync.conf file, the virt-127 node was mentioned i neither of those (besides a related fencing resource).



> Version-Release number of selected component (if applicable):

pacemaker-1.1.12-9.el7.x86_64
corosync-2.3.4-3.el7.x86_64
pcs-0.9.135-1.el7.x86_64 


> How reproducible:

Relatively Easy


> Steps to Reproduce:

1) Setup a cluster of i.e. 7 nodes, with nodes marked n1 to n7
2) Stop node n6 with pcs
3) Remove node n2 with pcs
4) Remove node n7 with pcs
5) Remove node n3 with pcs
6) Add node n7 with pcs
7) Compare node ID assignment before and after these steps - if a different ID was reused by n7 when added back to the cluster, it will most likely cause these errors.


> Actual results:

When the node is added back with a reused ID, membership errors appear in the logs and some pacemakr components will start crashing:

attrd:  warning: crm_find_peer:     Node 'virt-127' and 'virt-131' share the same cluster nodeid: 3 virt-131
attrd:    error: crm_abort:     crm_find_peer: Forked child 20878 to record non-fatal assert at membership.c:394 : member weirdness 


> Expected results:

Reusing of a node ID is not fatal and doesn't cause any membership errors.


> Additional info:

Feel free to reassign this bug to a different component shouldn't pacemaker be at fault here.