pacemaker/018-controller.patch
Ken Gaillot 103449b89e Backport a couple of fixes from upstream
- Fix watchdog agent version information
- Ensure transient attributes are cleared when multiple nodes are lost
- Resolves: rhbz1443666
- Resolves: rhbz1986998
2021-08-10 15:42:12 -05:00

123 lines
5.0 KiB
Diff

From ee7eba6a7a05bdf0a12d60ebabb334d8ee021101 Mon Sep 17 00:00:00 2001
From: Ken Gaillot <kgaillot@redhat.com>
Date: Mon, 9 Aug 2021 14:48:57 -0500
Subject: [PATCH] Fix: controller: ensure lost node's transient attributes are
cleared without DC
Previously, peer_update_callback() cleared a lost node's transient attributes
if either the local node is DC, or there is no DC.
However, that left the possibility of the DC being lost at the same time as
another node -- the local node would still have fsa_our_dc set while processing
the leave notifications, so no node would clear the attributes for the non-DC
node.
Now, the controller has its own CPG configuration change callback, which sets a
global boolean before calling the usual one, so that peer_update_callback() can
know when the DC has been lost.
---
daemons/controld/controld_callbacks.c | 4 +-
daemons/controld/controld_corosync.c | 57 ++++++++++++++++++++++++++-
2 files changed, 59 insertions(+), 2 deletions(-)
diff --git a/daemons/controld/controld_callbacks.c b/daemons/controld/controld_callbacks.c
index af24856ae..e564b3dcd 100644
--- a/daemons/controld/controld_callbacks.c
+++ b/daemons/controld/controld_callbacks.c
@@ -99,6 +99,8 @@ node_alive(const crm_node_t *node)
#define state_text(state) ((state)? (const char *)(state) : "in unknown state")
+bool controld_dc_left = false;
+
void
peer_update_callback(enum crm_status_type type, crm_node_t * node, const void *data)
{
@@ -217,7 +219,7 @@ peer_update_callback(enum crm_status_type type, crm_node_t * node, const void *d
cib_scope_local);
}
- } else if (AM_I_DC || (fsa_our_dc == NULL)) {
+ } else if (AM_I_DC || controld_dc_left || (fsa_our_dc == NULL)) {
/* This only needs to be done once, so normally the DC should do
* it. However if there is no DC, every node must do it, since
* there is no other way to ensure some one node does it.
diff --git a/daemons/controld/controld_corosync.c b/daemons/controld/controld_corosync.c
index db99630fb..c5ab6580a 100644
--- a/daemons/controld/controld_corosync.c
+++ b/daemons/controld/controld_corosync.c
@@ -87,6 +87,61 @@ crmd_cs_destroy(gpointer user_data)
}
}
+extern bool controld_dc_left;
+
+/*!
+ * \brief Handle a Corosync notification of a CPG configuration change
+ *
+ * \param[in] handle CPG connection
+ * \param[in] cpg_name CPG group name
+ * \param[in] member_list List of current CPG members
+ * \param[in] member_list_entries Number of entries in \p member_list
+ * \param[in] left_list List of CPG members that left
+ * \param[in] left_list_entries Number of entries in \p left_list
+ * \param[in] joined_list List of CPG members that joined
+ * \param[in] joined_list_entries Number of entries in \p joined_list
+ */
+static void
+cpg_membership_callback(cpg_handle_t handle, const struct cpg_name *cpg_name,
+ const struct cpg_address *member_list,
+ size_t member_list_entries,
+ const struct cpg_address *left_list,
+ size_t left_list_entries,
+ const struct cpg_address *joined_list,
+ size_t joined_list_entries)
+{
+ /* When nodes leave CPG, the DC clears their transient node attributes.
+ *
+ * However if there is no DC, or the DC is among the nodes that left, each
+ * remaining node needs to do the clearing, to ensure it gets done.
+ * Otherwise, the attributes would persist when the nodes rejoin, which
+ * could have serious consequences for unfencing, agents that use attributes
+ * for internal logic, etc.
+ *
+ * Here, we set a global boolean if the DC is among the nodes that left, for
+ * use by the peer callback.
+ */
+ if (fsa_our_dc != NULL) {
+ crm_node_t *peer = pcmk__search_cluster_node_cache(0, fsa_our_dc);
+
+ if (peer != NULL) {
+ for (int i = 0; i < left_list_entries; ++i) {
+ if (left_list[i].nodeid == peer->id) {
+ controld_dc_left = true;
+ break;
+ }
+ }
+ }
+ }
+
+ // Process the change normally, which will call the peer callback as needed
+ pcmk_cpg_membership(handle, cpg_name, member_list, member_list_entries,
+ left_list, left_list_entries,
+ joined_list, joined_list_entries);
+
+ controld_dc_left = false;
+}
+
extern gboolean crm_connect_corosync(crm_cluster_t * cluster);
gboolean
@@ -95,7 +150,7 @@ crm_connect_corosync(crm_cluster_t * cluster)
if (is_corosync_cluster()) {
crm_set_status_callback(&peer_update_callback);
cluster->cpg.cpg_deliver_fn = crmd_cs_dispatch;
- cluster->cpg.cpg_confchg_fn = pcmk_cpg_membership;
+ cluster->cpg.cpg_confchg_fn = cpg_membership_callback;
cluster->destroy = crmd_cs_destroy;
if (crm_cluster_connect(cluster)) {
--
2.27.0