diff --git a/lib/pacemaker/pcmk_sched_allocate.c b/lib/pacemaker/pcmk_sched_allocate.c
index b7d1b482a9..9f82c00251 100644
--- a/lib/pacemaker/pcmk_sched_allocate.c
+++ b/lib/pacemaker/pcmk_sched_allocate.c
@@ -1,2756 +1,2755 @@
/*
* Copyright 2004-2019 the Pacemaker project contributors
*
* The version control history for this file may have further details.
*
* This source code is licensed under the GNU General Public License version 2
* or later (GPLv2+) WITHOUT ANY WARRANTY.
*/
#include <crm_internal.h>
#include <sys/param.h>
#include <crm/crm.h>
#include <crm/cib.h>
#include <crm/msg_xml.h>
#include <crm/common/xml.h>
#include <glib.h>
#include <crm/pengine/status.h>
#include <pacemaker-internal.h>
CRM_TRACE_INIT_DATA(pe_allocate);
void set_alloc_actions(pe_working_set_t * data_set);
extern void ReloadRsc(resource_t * rsc, node_t *node, pe_working_set_t * data_set);
extern gboolean DeleteRsc(resource_t * rsc, node_t * node, gboolean optional, pe_working_set_t * data_set);
static void apply_remote_node_ordering(pe_working_set_t *data_set);
static enum remote_connection_state get_remote_node_state(pe_node_t *node);
enum remote_connection_state {
remote_state_unknown = 0,
remote_state_alive = 1,
remote_state_resting = 2,
remote_state_failed = 3,
remote_state_stopped = 4
};
static const char *
state2text(enum remote_connection_state state)
{
switch (state) {
case remote_state_unknown:
return "unknown";
case remote_state_alive:
return "alive";
case remote_state_resting:
return "resting";
case remote_state_failed:
return "failed";
case remote_state_stopped:
return "stopped";
}
return "impossible";
}
resource_alloc_functions_t resource_class_alloc_functions[] = {
{
native_merge_weights,
native_color,
native_create_actions,
native_create_probe,
native_internal_constraints,
native_rsc_colocation_lh,
native_rsc_colocation_rh,
native_rsc_location,
native_action_flags,
native_update_actions,
native_expand,
native_append_meta,
},
{
group_merge_weights,
group_color,
group_create_actions,
native_create_probe,
group_internal_constraints,
group_rsc_colocation_lh,
group_rsc_colocation_rh,
group_rsc_location,
group_action_flags,
group_update_actions,
group_expand,
group_append_meta,
},
{
clone_merge_weights,
clone_color,
clone_create_actions,
clone_create_probe,
clone_internal_constraints,
clone_rsc_colocation_lh,
clone_rsc_colocation_rh,
clone_rsc_location,
clone_action_flags,
pcmk__multi_update_actions,
clone_expand,
clone_append_meta,
},
{
pcmk__bundle_merge_weights,
pcmk__bundle_color,
pcmk__bundle_create_actions,
pcmk__bundle_create_probe,
pcmk__bundle_internal_constraints,
pcmk__bundle_rsc_colocation_lh,
pcmk__bundle_rsc_colocation_rh,
pcmk__bundle_rsc_location,
pcmk__bundle_action_flags,
pcmk__multi_update_actions,
pcmk__bundle_expand,
pcmk__bundle_append_meta,
}
};
gboolean
update_action_flags(action_t * action, enum pe_action_flags flags, const char *source, int line)
{
static unsigned long calls = 0;
gboolean changed = FALSE;
gboolean clear = is_set(flags, pe_action_clear);
enum pe_action_flags last = action->flags;
if (clear) {
action->flags = crm_clear_bit(source, line, action->uuid, action->flags, flags);
} else {
action->flags = crm_set_bit(source, line, action->uuid, action->flags, flags);
}
if (last != action->flags) {
calls++;
changed = TRUE;
/* Useful for tracking down _who_ changed a specific flag */
/* CRM_ASSERT(calls != 534); */
clear_bit(flags, pe_action_clear);
crm_trace("%s on %s: %sset flags 0x%.6x (was 0x%.6x, now 0x%.6x, %lu, %s)",
action->uuid, action->node ? action->node->details->uname : "[none]",
clear ? "un-" : "", flags, last, action->flags, calls, source);
}
return changed;
}
static gboolean
check_rsc_parameters(resource_t * rsc, node_t * node, xmlNode * rsc_entry,
gboolean active_here, pe_working_set_t * data_set)
{
int attr_lpc = 0;
gboolean force_restart = FALSE;
gboolean delete_resource = FALSE;
gboolean changed = FALSE;
const char *value = NULL;
const char *old_value = NULL;
const char *attr_list[] = {
XML_ATTR_TYPE,
XML_AGENT_ATTR_CLASS,
XML_AGENT_ATTR_PROVIDER
};
for (; attr_lpc < DIMOF(attr_list); attr_lpc++) {
value = crm_element_value(rsc->xml, attr_list[attr_lpc]);
old_value = crm_element_value(rsc_entry, attr_list[attr_lpc]);
if (value == old_value /* i.e. NULL */
|| crm_str_eq(value, old_value, TRUE)) {
continue;
}
changed = TRUE;
trigger_unfencing(rsc, node, "Device definition changed", NULL, data_set);
if (active_here) {
force_restart = TRUE;
crm_notice("Forcing restart of %s on %s, %s changed: %s -> %s",
rsc->id, node->details->uname, attr_list[attr_lpc],
crm_str(old_value), crm_str(value));
}
}
if (force_restart) {
/* make sure the restart happens */
stop_action(rsc, node, FALSE);
set_bit(rsc->flags, pe_rsc_start_pending);
delete_resource = TRUE;
} else if (changed) {
delete_resource = TRUE;
}
return delete_resource;
}
static void
CancelXmlOp(resource_t * rsc, xmlNode * xml_op, node_t * active_node,
const char *reason, pe_working_set_t * data_set)
{
guint interval_ms = 0;
action_t *cancel = NULL;
const char *task = NULL;
const char *call_id = NULL;
const char *interval_ms_s = NULL;
CRM_CHECK(xml_op != NULL, return);
CRM_CHECK(active_node != NULL, return);
task = crm_element_value(xml_op, XML_LRM_ATTR_TASK);
call_id = crm_element_value(xml_op, XML_LRM_ATTR_CALLID);
interval_ms_s = crm_element_value(xml_op, XML_LRM_ATTR_INTERVAL_MS);
interval_ms = crm_parse_ms(interval_ms_s);
crm_info("Action " CRM_OP_FMT " on %s will be stopped: %s",
rsc->id, task, interval_ms,
active_node->details->uname, (reason? reason : "unknown"));
cancel = pe_cancel_op(rsc, task, interval_ms, active_node, data_set);
add_hash_param(cancel->meta, XML_LRM_ATTR_CALLID, call_id);
custom_action_order(rsc, stop_key(rsc), NULL, rsc, NULL, cancel, pe_order_optional, data_set);
}
static gboolean
check_action_definition(resource_t * rsc, node_t * active_node, xmlNode * xml_op,
pe_working_set_t * data_set)
{
char *key = NULL;
guint interval_ms = 0;
const char *interval_ms_s = NULL;
const op_digest_cache_t *digest_data = NULL;
gboolean did_change = FALSE;
const char *task = crm_element_value(xml_op, XML_LRM_ATTR_TASK);
const char *digest_secure = NULL;
CRM_CHECK(active_node != NULL, return FALSE);
interval_ms_s = crm_element_value(xml_op, XML_LRM_ATTR_INTERVAL_MS);
interval_ms = crm_parse_ms(interval_ms_s);
if (interval_ms > 0) {
xmlNode *op_match = NULL;
/* we need to reconstruct the key because of the way we used to construct resource IDs */
key = generate_op_key(rsc->id, task, interval_ms);
pe_rsc_trace(rsc, "Checking parameters for %s", key);
op_match = find_rsc_op_entry(rsc, key);
if (op_match == NULL && is_set(data_set->flags, pe_flag_stop_action_orphans)) {
CancelXmlOp(rsc, xml_op, active_node, "orphan", data_set);
free(key);
return TRUE;
} else if (op_match == NULL) {
pe_rsc_debug(rsc, "Orphan action detected: %s on %s", key, active_node->details->uname);
free(key);
return TRUE;
}
free(key);
key = NULL;
}
crm_trace("Testing " CRM_OP_FMT " on %s",
rsc->id, task, interval_ms, active_node->details->uname);
if ((interval_ms == 0) && safe_str_eq(task, RSC_STATUS)) {
/* Reload based on the start action not a probe */
task = RSC_START;
} else if ((interval_ms == 0) && safe_str_eq(task, RSC_MIGRATED)) {
/* Reload based on the start action not a migrate */
task = RSC_START;
} else if ((interval_ms == 0) && safe_str_eq(task, RSC_PROMOTE)) {
/* Reload based on the start action not a promote */
task = RSC_START;
}
digest_data = rsc_action_digest_cmp(rsc, xml_op, active_node, data_set);
if(is_set(data_set->flags, pe_flag_sanitized)) {
digest_secure = crm_element_value(xml_op, XML_LRM_ATTR_SECURE_DIGEST);
}
if(digest_data->rc != RSC_DIGEST_MATCH
&& digest_secure
&& digest_data->digest_secure_calc
&& strcmp(digest_data->digest_secure_calc, digest_secure) == 0) {
if (is_set(data_set->flags, pe_flag_stdout)) {
printf("Only 'private' parameters to " CRM_OP_FMT " on %s changed: %s\n",
rsc->id, task, interval_ms, active_node->details->uname,
crm_element_value(xml_op, XML_ATTR_TRANSITION_MAGIC));
}
} else if (digest_data->rc == RSC_DIGEST_RESTART) {
/* Changes that force a restart */
pe_action_t *required = NULL;
did_change = TRUE;
key = generate_op_key(rsc->id, task, interval_ms);
crm_log_xml_info(digest_data->params_restart, "params:restart");
required = custom_action(rsc, key, task, NULL, TRUE, TRUE, data_set);
pe_action_set_flag_reason(__FUNCTION__, __LINE__, required, NULL,
"resource definition change", pe_action_optional, TRUE);
trigger_unfencing(rsc, active_node, "Device parameters changed", NULL, data_set);
} else if ((digest_data->rc == RSC_DIGEST_ALL) || (digest_data->rc == RSC_DIGEST_UNKNOWN)) {
/* Changes that can potentially be handled by a reload */
const char *digest_restart = crm_element_value(xml_op, XML_LRM_ATTR_RESTART_DIGEST);
did_change = TRUE;
trigger_unfencing(rsc, active_node, "Device parameters changed (reload)", NULL, data_set);
crm_log_xml_info(digest_data->params_all, "params:reload");
key = generate_op_key(rsc->id, task, interval_ms);
if (interval_ms > 0) {
action_t *op = NULL;
#if 0
/* Always reload/restart the entire resource */
ReloadRsc(rsc, active_node, data_set);
#else
/* Re-sending the recurring op is sufficient - the old one will be cancelled automatically */
op = custom_action(rsc, key, task, active_node, TRUE, TRUE, data_set);
set_bit(op->flags, pe_action_reschedule);
#endif
} else if (digest_restart) {
pe_rsc_trace(rsc, "Reloading '%s' action for resource %s", task, rsc->id);
/* Reload this resource */
ReloadRsc(rsc, active_node, data_set);
free(key);
} else {
pe_action_t *required = NULL;
pe_rsc_trace(rsc, "Resource %s doesn't know how to reload", rsc->id);
/* Re-send the start/demote/promote op
* Recurring ops will be detected independently
*/
required = custom_action(rsc, key, task, NULL, TRUE, TRUE, data_set);
pe_action_set_flag_reason(__FUNCTION__, __LINE__, required, NULL,
"resource definition change", pe_action_optional, TRUE);
}
}
return did_change;
}
/*!
* \internal
* \brief Do deferred action checks after allocation
*
* \param[in] data_set Working set for cluster
*/
static void
check_params(pe_resource_t *rsc, pe_node_t *node, xmlNode *rsc_op,
enum pe_check_parameters check, pe_working_set_t *data_set)
{
const char *reason = NULL;
op_digest_cache_t *digest_data = NULL;
switch (check) {
case pe_check_active:
if (check_action_definition(rsc, node, rsc_op, data_set)
&& pe_get_failcount(node, rsc, NULL, pe_fc_effective, NULL,
data_set)) {
reason = "action definition changed";
}
break;
case pe_check_last_failure:
digest_data = rsc_action_digest_cmp(rsc, rsc_op, node, data_set);
switch (digest_data->rc) {
case RSC_DIGEST_UNKNOWN:
crm_trace("Resource %s history entry %s on %s has no digest to compare",
rsc->id, ID(rsc_op), node->details->id);
break;
case RSC_DIGEST_MATCH:
break;
default:
reason = "resource parameters have changed";
break;
}
break;
}
if (reason) {
pe__clear_failcount(rsc, node, reason, data_set);
}
}
static void
check_actions_for(xmlNode * rsc_entry, resource_t * rsc, node_t * node, pe_working_set_t * data_set)
{
GListPtr gIter = NULL;
int offset = -1;
guint interval_ms = 0;
int stop_index = 0;
int start_index = 0;
const char *task = NULL;
const char *interval_ms_s = NULL;
xmlNode *rsc_op = NULL;
GListPtr op_list = NULL;
GListPtr sorted_op_list = NULL;
CRM_CHECK(node != NULL, return);
if (is_set(rsc->flags, pe_rsc_orphan)) {
resource_t *parent = uber_parent(rsc);
if(parent == NULL
|| pe_rsc_is_clone(parent) == FALSE
|| is_set(parent->flags, pe_rsc_unique)) {
pe_rsc_trace(rsc, "Skipping param check for %s and deleting: orphan", rsc->id);
DeleteRsc(rsc, node, FALSE, data_set);
} else {
pe_rsc_trace(rsc, "Skipping param check for %s (orphan clone)", rsc->id);
}
return;
} else if (pe_find_node_id(rsc->running_on, node->details->id) == NULL) {
if (check_rsc_parameters(rsc, node, rsc_entry, FALSE, data_set)) {
DeleteRsc(rsc, node, FALSE, data_set);
}
pe_rsc_trace(rsc, "Skipping param check for %s: no longer active on %s",
rsc->id, node->details->uname);
return;
}
pe_rsc_trace(rsc, "Processing %s on %s", rsc->id, node->details->uname);
if (check_rsc_parameters(rsc, node, rsc_entry, TRUE, data_set)) {
DeleteRsc(rsc, node, FALSE, data_set);
}
for (rsc_op = __xml_first_child(rsc_entry); rsc_op != NULL; rsc_op = __xml_next_element(rsc_op)) {
if (crm_str_eq((const char *)rsc_op->name, XML_LRM_TAG_RSC_OP, TRUE)) {
op_list = g_list_prepend(op_list, rsc_op);
}
}
sorted_op_list = g_list_sort(op_list, sort_op_by_callid);
calculate_active_ops(sorted_op_list, &start_index, &stop_index);
for (gIter = sorted_op_list; gIter != NULL; gIter = gIter->next) {
xmlNode *rsc_op = (xmlNode *) gIter->data;
offset++;
if (start_index < stop_index) {
/* stopped */
continue;
} else if (offset < start_index) {
/* action occurred prior to a start */
continue;
}
task = crm_element_value(rsc_op, XML_LRM_ATTR_TASK);
interval_ms_s = crm_element_value(rsc_op, XML_LRM_ATTR_INTERVAL_MS);
interval_ms = crm_parse_ms(interval_ms_s);
if ((interval_ms > 0) &&
(is_set(rsc->flags, pe_rsc_maintenance) || node->details->maintenance)) {
// Maintenance mode cancels recurring operations
CancelXmlOp(rsc, rsc_op, node, "maintenance mode", data_set);
} else if ((interval_ms > 0)
|| safe_str_eq(task, RSC_STATUS)
|| safe_str_eq(task, RSC_START)
|| safe_str_eq(task, RSC_PROMOTE)
|| safe_str_eq(task, RSC_MIGRATED)) {
/* If a resource operation failed, and the operation's definition
* has changed, clear any fail count so they can be retried fresh.
*/
if (pe__bundle_needs_remote_name(rsc)) {
/* We haven't allocated resources to nodes yet, so if the
* REMOTE_CONTAINER_HACK is used, we may calculate the digest
* based on the literal "#uname" value rather than the properly
* substituted value. That would mistakenly make the action
* definition appear to have been changed. Defer the check until
* later in this case.
*/
pe__add_param_check(rsc_op, rsc, node, pe_check_active,
data_set);
} else if (check_action_definition(rsc, node, rsc_op, data_set)
&& pe_get_failcount(node, rsc, NULL, pe_fc_effective, NULL,
data_set)) {
pe__clear_failcount(rsc, node, "action definition changed",
data_set);
}
}
}
g_list_free(sorted_op_list);
}
static GListPtr
find_rsc_list(GListPtr result, resource_t * rsc, const char *id, gboolean renamed_clones,
gboolean partial, pe_working_set_t * data_set)
{
GListPtr gIter = NULL;
gboolean match = FALSE;
if (id == NULL) {
return NULL;
} else if (rsc == NULL && data_set) {
for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) {
resource_t *child = (resource_t *) gIter->data;
result = find_rsc_list(result, child, id, renamed_clones, partial, NULL);
}
return result;
} else if (rsc == NULL) {
return NULL;
}
if (partial) {
if (strstr(rsc->id, id)) {
match = TRUE;
} else if (renamed_clones && rsc->clone_name && strstr(rsc->clone_name, id)) {
match = TRUE;
}
} else {
if (strcmp(rsc->id, id) == 0) {
match = TRUE;
} else if (renamed_clones && rsc->clone_name && strcmp(rsc->clone_name, id) == 0) {
match = TRUE;
}
}
if (match) {
result = g_list_prepend(result, rsc);
}
if (rsc->children) {
gIter = rsc->children;
for (; gIter != NULL; gIter = gIter->next) {
resource_t *child = (resource_t *) gIter->data;
result = find_rsc_list(result, child, id, renamed_clones, partial, NULL);
}
}
return result;
}
static void
check_actions(pe_working_set_t * data_set)
{
const char *id = NULL;
node_t *node = NULL;
xmlNode *lrm_rscs = NULL;
xmlNode *status = get_object_root(XML_CIB_TAG_STATUS, data_set->input);
xmlNode *node_state = NULL;
for (node_state = __xml_first_child(status); node_state != NULL;
node_state = __xml_next_element(node_state)) {
if (crm_str_eq((const char *)node_state->name, XML_CIB_TAG_STATE, TRUE)) {
id = crm_element_value(node_state, XML_ATTR_ID);
lrm_rscs = find_xml_node(node_state, XML_CIB_TAG_LRM, FALSE);
lrm_rscs = find_xml_node(lrm_rscs, XML_LRM_TAG_RESOURCES, FALSE);
node = pe_find_node_id(data_set->nodes, id);
if (node == NULL) {
continue;
/* Still need to check actions for a maintenance node to cancel existing monitor operations */
} else if (can_run_resources(node) == FALSE && node->details->maintenance == FALSE) {
crm_trace("Skipping param check for %s: can't run resources",
node->details->uname);
continue;
}
crm_trace("Processing node %s", node->details->uname);
if (node->details->online || is_set(data_set->flags, pe_flag_stonith_enabled)) {
xmlNode *rsc_entry = NULL;
for (rsc_entry = __xml_first_child(lrm_rscs); rsc_entry != NULL;
rsc_entry = __xml_next_element(rsc_entry)) {
if (crm_str_eq((const char *)rsc_entry->name, XML_LRM_TAG_RESOURCE, TRUE)) {
if (xml_has_children(rsc_entry)) {
GListPtr gIter = NULL;
GListPtr result = NULL;
const char *rsc_id = ID(rsc_entry);
CRM_CHECK(rsc_id != NULL, return);
result = find_rsc_list(NULL, NULL, rsc_id, TRUE, FALSE, data_set);
for (gIter = result; gIter != NULL; gIter = gIter->next) {
resource_t *rsc = (resource_t *) gIter->data;
if (rsc->variant != pe_native) {
continue;
}
check_actions_for(rsc_entry, rsc, node, data_set);
}
g_list_free(result);
}
}
}
}
}
}
}
static gboolean
apply_placement_constraints(pe_working_set_t * data_set)
{
GListPtr gIter = NULL;
crm_trace("Applying constraints...");
for (gIter = data_set->placement_constraints; gIter != NULL; gIter = gIter->next) {
pe__location_t *cons = gIter->data;
cons->rsc_lh->cmds->rsc_location(cons->rsc_lh, cons);
}
return TRUE;
}
static gboolean
failcount_clear_action_exists(node_t * node, resource_t * rsc)
{
gboolean rc = FALSE;
GList *list = pe__resource_actions(rsc, node, CRM_OP_CLEAR_FAILCOUNT, TRUE);
if (list) {
rc = TRUE;
}
g_list_free(list);
return rc;
}
/*!
* \internal
* \brief Force resource away if failures hit migration threshold
*
* \param[in,out] rsc Resource to check for failures
* \param[in,out] node Node to check for failures
* \param[in,out] data_set Cluster working set to update
*/
static void
check_migration_threshold(resource_t *rsc, node_t *node,
pe_working_set_t *data_set)
{
int fail_count, countdown;
resource_t *failed;
/* Migration threshold of 0 means never force away */
if (rsc->migration_threshold == 0) {
return;
}
// If we're ignoring failures, also ignore the migration threshold
if (is_set(rsc->flags, pe_rsc_failure_ignored)) {
return;
}
/* If there are no failures, there's no need to force away */
fail_count = pe_get_failcount(node, rsc, NULL,
pe_fc_effective|pe_fc_fillers, NULL,
data_set);
if (fail_count <= 0) {
return;
}
/* How many more times recovery will be tried on this node */
countdown = QB_MAX(rsc->migration_threshold - fail_count, 0);
/* If failed resource has a parent, we'll force the parent away */
failed = rsc;
if (is_not_set(rsc->flags, pe_rsc_unique)) {
failed = uber_parent(rsc);
}
if (countdown == 0) {
resource_location(failed, node, -INFINITY, "__fail_limit__", data_set);
crm_warn("Forcing %s away from %s after %d failures (max=%d)",
failed->id, node->details->uname, fail_count,
rsc->migration_threshold);
} else {
crm_info("%s can fail %d more times on %s before being forced off",
failed->id, countdown, node->details->uname);
}
}
static void
common_apply_stickiness(resource_t * rsc, node_t * node, pe_working_set_t * data_set)
{
if (rsc->children) {
GListPtr gIter = rsc->children;
for (; gIter != NULL; gIter = gIter->next) {
resource_t *child_rsc = (resource_t *) gIter->data;
common_apply_stickiness(child_rsc, node, data_set);
}
return;
}
if (is_set(rsc->flags, pe_rsc_managed)
&& rsc->stickiness != 0 && g_list_length(rsc->running_on) == 1) {
node_t *current = pe_find_node_id(rsc->running_on, node->details->id);
node_t *match = pe_hash_table_lookup(rsc->allowed_nodes, node->details->id);
if (current == NULL) {
} else if (match != NULL || is_set(data_set->flags, pe_flag_symmetric_cluster)) {
resource_t *sticky_rsc = rsc;
resource_location(sticky_rsc, node, rsc->stickiness, "stickiness", data_set);
pe_rsc_debug(sticky_rsc, "Resource %s: preferring current location"
" (node=%s, weight=%d)", sticky_rsc->id,
node->details->uname, rsc->stickiness);
} else {
GHashTableIter iter;
node_t *nIter = NULL;
pe_rsc_debug(rsc, "Ignoring stickiness for %s: the cluster is asymmetric"
" and node %s is not explicitly allowed", rsc->id, node->details->uname);
g_hash_table_iter_init(&iter, rsc->allowed_nodes);
while (g_hash_table_iter_next(&iter, NULL, (void **)&nIter)) {
crm_err("%s[%s] = %d", rsc->id, nIter->details->uname, nIter->weight);
}
}
}
/* Check the migration threshold only if a failcount clear action
* has not already been placed for this resource on the node.
* There is no sense in potentially forcing the resource from this
* node if the failcount is being reset anyway.
*
* @TODO A clear_failcount operation can be scheduled in stage4() via
* check_actions_for(), or in stage5() via check_params(). This runs in
* stage2(), so it cannot detect those, meaning we might check the migration
* threshold when we shouldn't -- worst case, we stop or move the resource,
* then move it back next transition.
*/
if (failcount_clear_action_exists(node, rsc) == FALSE) {
check_migration_threshold(rsc, node, data_set);
}
}
void
complex_set_cmds(resource_t * rsc)
{
GListPtr gIter = rsc->children;
rsc->cmds = &resource_class_alloc_functions[rsc->variant];
for (; gIter != NULL; gIter = gIter->next) {
resource_t *child_rsc = (resource_t *) gIter->data;
complex_set_cmds(child_rsc);
}
}
void
set_alloc_actions(pe_working_set_t * data_set)
{
GListPtr gIter = data_set->resources;
for (; gIter != NULL; gIter = gIter->next) {
resource_t *rsc = (resource_t *) gIter->data;
complex_set_cmds(rsc);
}
}
static void
calculate_system_health(gpointer gKey, gpointer gValue, gpointer user_data)
{
const char *key = (const char *)gKey;
const char *value = (const char *)gValue;
int *system_health = (int *)user_data;
if (!gKey || !gValue || !user_data) {
return;
}
if (crm_starts_with(key, "#health")) {
int score;
/* Convert the value into an integer */
score = char2score(value);
/* Add it to the running total */
*system_health = merge_weights(score, *system_health);
}
}
static gboolean
apply_system_health(pe_working_set_t * data_set)
{
GListPtr gIter = NULL;
const char *health_strategy = pe_pref(data_set->config_hash, "node-health-strategy");
int base_health = 0;
if (health_strategy == NULL || safe_str_eq(health_strategy, "none")) {
/* Prevent any accidental health -> score translation */
node_score_red = 0;
node_score_yellow = 0;
node_score_green = 0;
return TRUE;
} else if (safe_str_eq(health_strategy, "migrate-on-red")) {
/* Resources on nodes which have health values of red are
* weighted away from that node.
*/
node_score_red = -INFINITY;
node_score_yellow = 0;
node_score_green = 0;
} else if (safe_str_eq(health_strategy, "only-green")) {
/* Resources on nodes which have health values of red or yellow
* are forced away from that node.
*/
node_score_red = -INFINITY;
node_score_yellow = -INFINITY;
node_score_green = 0;
} else if (safe_str_eq(health_strategy, "progressive")) {
/* Same as the above, but use the r/y/g scores provided by the user
* Defaults are provided by the pe_prefs table
* Also, custom health "base score" can be used
*/
base_health = crm_parse_int(pe_pref(data_set->config_hash, "node-health-base"), "0");
} else if (safe_str_eq(health_strategy, "custom")) {
/* Requires the admin to configure the rsc_location constaints for
* processing the stored health scores
*/
/* TODO: Check for the existence of appropriate node health constraints */
return TRUE;
} else {
crm_err("Unknown node health strategy: %s", health_strategy);
return FALSE;
}
crm_info("Applying automated node health strategy: %s", health_strategy);
for (gIter = data_set->nodes; gIter != NULL; gIter = gIter->next) {
int system_health = base_health;
node_t *node = (node_t *) gIter->data;
/* Search through the node hash table for system health entries. */
g_hash_table_foreach(node->details->attrs, calculate_system_health, &system_health);
crm_info(" Node %s has an combined system health of %d",
node->details->uname, system_health);
/* If the health is non-zero, then create a new rsc2node so that the
* weight will be added later on.
*/
if (system_health != 0) {
GListPtr gIter2 = data_set->resources;
for (; gIter2 != NULL; gIter2 = gIter2->next) {
resource_t *rsc = (resource_t *) gIter2->data;
rsc2node_new(health_strategy, rsc, system_health, NULL, node, data_set);
}
}
}
return TRUE;
}
gboolean
stage0(pe_working_set_t * data_set)
{
xmlNode *cib_constraints = get_object_root(XML_CIB_TAG_CONSTRAINTS, data_set->input);
if (data_set->input == NULL) {
return FALSE;
}
if (is_set(data_set->flags, pe_flag_have_status) == FALSE) {
crm_trace("Calculating status");
cluster_status(data_set);
}
set_alloc_actions(data_set);
apply_system_health(data_set);
unpack_constraints(cib_constraints, data_set);
return TRUE;
}
/*
* Check nodes for resources started outside of the LRM
*/
gboolean
probe_resources(pe_working_set_t * data_set)
{
action_t *probe_node_complete = NULL;
for (GListPtr gIter = data_set->nodes; gIter != NULL; gIter = gIter->next) {
node_t *node = (node_t *) gIter->data;
const char *probed = pe_node_attribute_raw(node, CRM_OP_PROBED);
if (node->details->online == FALSE) {
if (pe__is_remote_node(node) && node->details->remote_rsc
&& (get_remote_node_state(node) == remote_state_failed)) {
pe_fence_node(data_set, node, "the connection is unrecoverable");
}
continue;
} else if (node->details->unclean) {
continue;
} else if (node->details->rsc_discovery_enabled == FALSE) {
/* resource discovery is disabled for this node */
continue;
}
if (probed != NULL && crm_is_true(probed) == FALSE) {
action_t *probe_op = custom_action(NULL, crm_strdup_printf("%s-%s", CRM_OP_REPROBE, node->details->uname),
CRM_OP_REPROBE, node, FALSE, TRUE, data_set);
add_hash_param(probe_op->meta, XML_ATTR_TE_NOWAIT, XML_BOOLEAN_TRUE);
continue;
}
for (GListPtr gIter2 = data_set->resources; gIter2 != NULL; gIter2 = gIter2->next) {
resource_t *rsc = (resource_t *) gIter2->data;
rsc->cmds->create_probe(rsc, node, probe_node_complete, FALSE, data_set);
}
}
return TRUE;
}
static void
rsc_discover_filter(resource_t *rsc, node_t *node)
{
GListPtr gIter = rsc->children;
resource_t *top = uber_parent(rsc);
node_t *match;
if (rsc->exclusive_discover == FALSE && top->exclusive_discover == FALSE) {
return;
}
for (; gIter != NULL; gIter = gIter->next) {
resource_t *child_rsc = (resource_t *) gIter->data;
rsc_discover_filter(child_rsc, node);
}
match = g_hash_table_lookup(rsc->allowed_nodes, node->details->id);
if (match && match->rsc_discover_mode != pe_discover_exclusive) {
match->weight = -INFINITY;
}
}
/*
* Count how many valid nodes we have (so we know the maximum number of
* colors we can resolve).
*
* Apply node constraints (i.e. filter the "allowed_nodes" part of resources)
*/
gboolean
stage2(pe_working_set_t * data_set)
{
GListPtr gIter = NULL;
crm_trace("Applying placement constraints");
gIter = data_set->nodes;
for (; gIter != NULL; gIter = gIter->next) {
node_t *node = (node_t *) gIter->data;
if (node == NULL) {
/* error */
} else if (node->weight >= 0.0 /* global weight */
&& node->details->online && node->details->type != node_ping) {
data_set->max_valid_nodes++;
}
}
apply_placement_constraints(data_set);
gIter = data_set->nodes;
for (; gIter != NULL; gIter = gIter->next) {
GListPtr gIter2 = NULL;
node_t *node = (node_t *) gIter->data;
gIter2 = data_set->resources;
for (; gIter2 != NULL; gIter2 = gIter2->next) {
resource_t *rsc = (resource_t *) gIter2->data;
common_apply_stickiness(rsc, node, data_set);
rsc_discover_filter(rsc, node);
}
}
return TRUE;
}
/*
* Create internal resource constraints before allocation
*/
gboolean
stage3(pe_working_set_t * data_set)
{
GListPtr gIter = data_set->resources;
for (; gIter != NULL; gIter = gIter->next) {
resource_t *rsc = (resource_t *) gIter->data;
rsc->cmds->internal_constraints(rsc, data_set);
}
return TRUE;
}
/*
* Check for orphaned or redefined actions
*/
gboolean
stage4(pe_working_set_t * data_set)
{
check_actions(data_set);
return TRUE;
}
static void *
convert_const_pointer(const void *ptr)
{
/* Worst function ever */
return (void *)ptr;
}
static gint
sort_rsc_process_order(gconstpointer a, gconstpointer b, gpointer data)
{
int rc = 0;
int r1_weight = -INFINITY;
int r2_weight = -INFINITY;
const char *reason = "existence";
const GListPtr nodes = (GListPtr) data;
const resource_t *resource1 = a;
const resource_t *resource2 = b;
node_t *r1_node = NULL;
node_t *r2_node = NULL;
GListPtr gIter = NULL;
GHashTable *r1_nodes = NULL;
GHashTable *r2_nodes = NULL;
if (a == NULL && b == NULL) {
goto done;
}
if (a == NULL) {
return 1;
}
if (b == NULL) {
return -1;
}
reason = "priority";
r1_weight = resource1->priority;
r2_weight = resource2->priority;
if (r1_weight > r2_weight) {
rc = -1;
goto done;
}
if (r1_weight < r2_weight) {
rc = 1;
goto done;
}
reason = "no node list";
if (nodes == NULL) {
goto done;
}
r1_nodes = rsc_merge_weights(convert_const_pointer(resource1),
resource1->id, NULL, NULL, 1,
pe_weights_forward | pe_weights_init);
dump_node_scores(LOG_TRACE, NULL, resource1->id, r1_nodes);
r2_nodes = rsc_merge_weights(convert_const_pointer(resource2),
resource2->id, NULL, NULL, 1,
pe_weights_forward | pe_weights_init);
dump_node_scores(LOG_TRACE, NULL, resource2->id, r2_nodes);
/* Current location score */
reason = "current location";
r1_weight = -INFINITY;
r2_weight = -INFINITY;
if (resource1->running_on) {
r1_node = pe__current_node(resource1);
r1_node = g_hash_table_lookup(r1_nodes, r1_node->details->id);
if (r1_node != NULL) {
r1_weight = r1_node->weight;
}
}
if (resource2->running_on) {
r2_node = pe__current_node(resource2);
r2_node = g_hash_table_lookup(r2_nodes, r2_node->details->id);
if (r2_node != NULL) {
r2_weight = r2_node->weight;
}
}
if (r1_weight > r2_weight) {
rc = -1;
goto done;
}
if (r1_weight < r2_weight) {
rc = 1;
goto done;
}
reason = "score";
for (gIter = nodes; gIter != NULL; gIter = gIter->next) {
node_t *node = (node_t *) gIter->data;
r1_node = NULL;
r2_node = NULL;
r1_weight = -INFINITY;
if (r1_nodes) {
r1_node = g_hash_table_lookup(r1_nodes, node->details->id);
}
if (r1_node) {
r1_weight = r1_node->weight;
}
r2_weight = -INFINITY;
if (r2_nodes) {
r2_node = g_hash_table_lookup(r2_nodes, node->details->id);
}
if (r2_node) {
r2_weight = r2_node->weight;
}
if (r1_weight > r2_weight) {
rc = -1;
goto done;
}
if (r1_weight < r2_weight) {
rc = 1;
goto done;
}
}
done:
crm_trace("%s (%d) on %s %c %s (%d) on %s: %s",
resource1->id, r1_weight, r1_node ? r1_node->details->id : "n/a",
rc < 0 ? '>' : rc > 0 ? '<' : '=',
resource2->id, r2_weight, r2_node ? r2_node->details->id : "n/a", reason);
if (r1_nodes) {
g_hash_table_destroy(r1_nodes);
}
if (r2_nodes) {
g_hash_table_destroy(r2_nodes);
}
return rc;
}
static void
allocate_resources(pe_working_set_t * data_set)
{
GListPtr gIter = NULL;
if (is_set(data_set->flags, pe_flag_have_remote_nodes)) {
/* Force remote connection resources to be allocated first. This
* also forces any colocation dependencies to be allocated as well */
for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) {
resource_t *rsc = (resource_t *) gIter->data;
if (rsc->is_remote_node == FALSE) {
continue;
}
pe_rsc_trace(rsc, "Allocating: %s", rsc->id);
/* For remote node connection resources, always prefer the partial
* migration target during resource allocation, if the rsc is in the
* middle of a migration.
*/
rsc->cmds->allocate(rsc, rsc->partial_migration_target, data_set);
}
}
/* now do the rest of the resources */
for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) {
resource_t *rsc = (resource_t *) gIter->data;
if (rsc->is_remote_node == TRUE) {
continue;
}
pe_rsc_trace(rsc, "Allocating: %s", rsc->id);
rsc->cmds->allocate(rsc, NULL, data_set);
}
}
/* We always use pe_order_preserve with these convenience functions to exempt
* internally generated constraints from the prohibition of user constraints
* involving remote connection resources.
*
* The start ordering additionally uses pe_order_runnable_left so that the
* specified action is not runnable if the start is not runnable.
*/
static inline void
order_start_then_action(resource_t *lh_rsc, action_t *rh_action,
enum pe_ordering extra, pe_working_set_t *data_set)
{
if (lh_rsc && rh_action && data_set) {
custom_action_order(lh_rsc, start_key(lh_rsc), NULL,
rh_action->rsc, NULL, rh_action,
pe_order_preserve | pe_order_runnable_left | extra,
data_set);
}
}
static inline void
order_action_then_stop(action_t *lh_action, resource_t *rh_rsc,
enum pe_ordering extra, pe_working_set_t *data_set)
{
if (lh_action && rh_rsc && data_set) {
custom_action_order(lh_action->rsc, NULL, lh_action,
rh_rsc, stop_key(rh_rsc), NULL,
pe_order_preserve | extra, data_set);
}
}
static void
cleanup_orphans(resource_t * rsc, pe_working_set_t * data_set)
{
GListPtr gIter = NULL;
if (is_set(data_set->flags, pe_flag_stop_rsc_orphans) == FALSE) {
return;
}
/* Don't recurse into ->children, those are just unallocated clone instances */
if(is_not_set(rsc->flags, pe_rsc_orphan)) {
return;
}
for (gIter = data_set->nodes; gIter != NULL; gIter = gIter->next) {
node_t *node = (node_t *) gIter->data;
if (node->details->online
&& pe_get_failcount(node, rsc, NULL, pe_fc_effective, NULL,
data_set)) {
pe_action_t *clear_op = NULL;
clear_op = pe__clear_failcount(rsc, node, "it is orphaned",
data_set);
/* We can't use order_action_then_stop() here because its
* pe_order_preserve breaks things
*/
custom_action_order(clear_op->rsc, NULL, clear_op,
rsc, stop_key(rsc), NULL,
pe_order_optional, data_set);
}
}
}
gboolean
stage5(pe_working_set_t * data_set)
{
GListPtr gIter = NULL;
if (safe_str_neq(data_set->placement_strategy, "default")) {
GListPtr nodes = g_list_copy(data_set->nodes);
nodes = sort_nodes_by_weight(nodes, NULL, data_set);
data_set->resources =
g_list_sort_with_data(data_set->resources, sort_rsc_process_order, nodes);
g_list_free(nodes);
}
gIter = data_set->nodes;
for (; gIter != NULL; gIter = gIter->next) {
node_t *node = (node_t *) gIter->data;
dump_node_capacity(show_utilization ? 0 : utilization_log_level, "Original", node);
}
crm_trace("Allocating services");
/* Take (next) highest resource, assign it and create its actions */
allocate_resources(data_set);
gIter = data_set->nodes;
for (; gIter != NULL; gIter = gIter->next) {
node_t *node = (node_t *) gIter->data;
dump_node_capacity(show_utilization ? 0 : utilization_log_level, "Remaining", node);
}
// Process deferred action checks
pe__foreach_param_check(data_set, check_params);
pe__free_param_checks(data_set);
if (is_set(data_set->flags, pe_flag_startup_probes)) {
crm_trace("Calculating needed probes");
/* This code probably needs optimization
* ptest -x with 100 nodes, 100 clones and clone-max=100:
With probes:
ptest[14781]: 2010/09/27_17:56:46 notice: TRACE: do_calculations: pengine.c:258 Calculate cluster status
ptest[14781]: 2010/09/27_17:56:46 notice: TRACE: do_calculations: pengine.c:278 Applying placement constraints
ptest[14781]: 2010/09/27_17:56:47 notice: TRACE: do_calculations: pengine.c:285 Create internal constraints
ptest[14781]: 2010/09/27_17:56:47 notice: TRACE: do_calculations: pengine.c:292 Check actions
ptest[14781]: 2010/09/27_17:56:48 notice: TRACE: do_calculations: pengine.c:299 Allocate resources
ptest[14781]: 2010/09/27_17:56:48 notice: TRACE: stage5: allocate.c:881 Allocating services
ptest[14781]: 2010/09/27_17:56:49 notice: TRACE: stage5: allocate.c:894 Calculating needed probes
ptest[14781]: 2010/09/27_17:56:51 notice: TRACE: stage5: allocate.c:899 Creating actions
ptest[14781]: 2010/09/27_17:56:52 notice: TRACE: stage5: allocate.c:905 Creating done
ptest[14781]: 2010/09/27_17:56:52 notice: TRACE: do_calculations: pengine.c:306 Processing fencing and shutdown cases
ptest[14781]: 2010/09/27_17:56:52 notice: TRACE: do_calculations: pengine.c:313 Applying ordering constraints
36s
ptest[14781]: 2010/09/27_17:57:28 notice: TRACE: do_calculations: pengine.c:320 Create transition graph
Without probes:
ptest[14637]: 2010/09/27_17:56:21 notice: TRACE: do_calculations: pengine.c:258 Calculate cluster status
ptest[14637]: 2010/09/27_17:56:22 notice: TRACE: do_calculations: pengine.c:278 Applying placement constraints
ptest[14637]: 2010/09/27_17:56:22 notice: TRACE: do_calculations: pengine.c:285 Create internal constraints
ptest[14637]: 2010/09/27_17:56:22 notice: TRACE: do_calculations: pengine.c:292 Check actions
ptest[14637]: 2010/09/27_17:56:23 notice: TRACE: do_calculations: pengine.c:299 Allocate resources
ptest[14637]: 2010/09/27_17:56:23 notice: TRACE: stage5: allocate.c:881 Allocating services
ptest[14637]: 2010/09/27_17:56:24 notice: TRACE: stage5: allocate.c:899 Creating actions
ptest[14637]: 2010/09/27_17:56:25 notice: TRACE: stage5: allocate.c:905 Creating done
ptest[14637]: 2010/09/27_17:56:25 notice: TRACE: do_calculations: pengine.c:306 Processing fencing and shutdown cases
ptest[14637]: 2010/09/27_17:56:25 notice: TRACE: do_calculations: pengine.c:313 Applying ordering constraints
ptest[14637]: 2010/09/27_17:56:25 notice: TRACE: do_calculations: pengine.c:320 Create transition graph
*/
probe_resources(data_set);
}
crm_trace("Handle orphans");
for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) {
resource_t *rsc = (resource_t *) gIter->data;
cleanup_orphans(rsc, data_set);
}
crm_trace("Creating actions");
for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) {
resource_t *rsc = (resource_t *) gIter->data;
rsc->cmds->create_actions(rsc, data_set);
}
crm_trace("Creating done");
return TRUE;
}
static gboolean
is_managed(const resource_t * rsc)
{
GListPtr gIter = rsc->children;
if (is_set(rsc->flags, pe_rsc_managed)) {
return TRUE;
}
for (; gIter != NULL; gIter = gIter->next) {
resource_t *child_rsc = (resource_t *) gIter->data;
if (is_managed(child_rsc)) {
return TRUE;
}
}
return FALSE;
}
static gboolean
any_managed_resources(pe_working_set_t * data_set)
{
GListPtr gIter = data_set->resources;
for (; gIter != NULL; gIter = gIter->next) {
resource_t *rsc = (resource_t *) gIter->data;
if (is_managed(rsc)) {
return TRUE;
}
}
return FALSE;
}
/*!
* \internal
* \brief Create pseudo-op for guest node fence, and order relative to it
*
* \param[in] node Guest node to fence
* \param[in] data_set Working set of CIB state
*/
static void
fence_guest(pe_node_t *node, pe_working_set_t *data_set)
{
resource_t *container = node->details->remote_rsc->container;
pe_action_t *stop = NULL;
pe_action_t *stonith_op = NULL;
/* The fence action is just a label; we don't do anything differently for
* off vs. reboot. We specify it explicitly, rather than let it default to
* cluster's default action, because we are not _initiating_ fencing -- we
* are creating a pseudo-event to describe fencing that is already occurring
* by other means (container recovery).
*/
const char *fence_action = "off";
/* Check whether guest's container resource has any explicit stop or
* start (the stop may be implied by fencing of the guest's host).
*/
if (container) {
stop = find_first_action(container->actions, NULL, CRMD_ACTION_STOP, NULL);
if (find_first_action(container->actions, NULL, CRMD_ACTION_START, NULL)) {
fence_action = "reboot";
}
}
/* Create a fence pseudo-event, so we have an event to order actions
* against, and the controller can always detect it.
*/
stonith_op = pe_fence_op(node, fence_action, FALSE, "guest is unclean", data_set);
update_action_flags(stonith_op, pe_action_pseudo | pe_action_runnable,
__FUNCTION__, __LINE__);
/* We want to imply stops/demotes after the guest is stopped, not wait until
* it is restarted, so we always order pseudo-fencing after stop, not start
* (even though start might be closer to what is done for a real reboot).
*/
if(stop && is_set(stop->flags, pe_action_pseudo)) {
pe_action_t *parent_stonith_op = pe_fence_op(stop->node, NULL, FALSE, NULL, data_set);
crm_info("Implying guest node %s is down (action %d) after %s fencing",
node->details->uname, stonith_op->id, stop->node->details->uname);
order_actions(parent_stonith_op, stonith_op,
pe_order_runnable_left|pe_order_implies_then);
} else if (stop) {
order_actions(stop, stonith_op,
pe_order_runnable_left|pe_order_implies_then);
crm_info("Implying guest node %s is down (action %d) "
"after container %s is stopped (action %d)",
node->details->uname, stonith_op->id,
container->id, stop->id);
} else {
/* If we're fencing the guest node but there's no stop for the guest
* resource, we must think the guest is already stopped. However, we may
* think so because its resource history was just cleaned. To avoid
* unnecessarily considering the guest node down if it's really up,
* order the pseudo-fencing after any stop of the connection resource,
* which will be ordered after any container (re-)probe.
*/
stop = find_first_action(node->details->remote_rsc->actions, NULL,
RSC_STOP, NULL);
if (stop) {
order_actions(stop, stonith_op, pe_order_optional);
crm_info("Implying guest node %s is down (action %d) "
"after connection is stopped (action %d)",
node->details->uname, stonith_op->id, stop->id);
} else {
/* Not sure why we're fencing, but everything must already be
* cleanly stopped.
*/
crm_info("Implying guest node %s is down (action %d) ",
node->details->uname, stonith_op->id);
}
}
/* Order/imply other actions relative to pseudo-fence as with real fence */
stonith_constraints(node, stonith_op, data_set);
}
/*
* Create dependencies for stonith and shutdown operations
*/
gboolean
stage6(pe_working_set_t * data_set)
{
action_t *dc_down = NULL;
action_t *stonith_op = NULL;
gboolean integrity_lost = FALSE;
gboolean need_stonith = TRUE;
GListPtr gIter;
GListPtr stonith_ops = NULL;
GList *shutdown_ops = NULL;
/* Remote ordering constraints need to happen prior to calculate
* fencing because it is one more place we will mark the node as
* dirty.
*
* A nice side-effect of doing it first is that we can remove a
* bunch of special logic from apply_*_ordering() because its
* already part of pe_fence_node()
*/
crm_trace("Creating remote ordering constraints");
apply_remote_node_ordering(data_set);
crm_trace("Processing fencing and shutdown cases");
if (any_managed_resources(data_set) == FALSE) {
crm_notice("Delaying fencing operations until there are resources to manage");
need_stonith = FALSE;
}
/* Check each node for stonith/shutdown */
for (gIter = data_set->nodes; gIter != NULL; gIter = gIter->next) {
node_t *node = (node_t *) gIter->data;
/* Guest nodes are "fenced" by recovering their container resource,
* so handle them separately.
*/
if (pe__is_guest_node(node)) {
if (node->details->remote_requires_reset && need_stonith) {
fence_guest(node, data_set);
}
continue;
}
stonith_op = NULL;
if (node->details->unclean
&& need_stonith && pe_can_fence(data_set, node)) {
stonith_op = pe_fence_op(node, NULL, FALSE, "node is unclean", data_set);
pe_warn("Scheduling Node %s for STONITH", node->details->uname);
stonith_constraints(node, stonith_op, data_set);
if (node->details->is_dc) {
// Remember if the DC is being fenced
dc_down = stonith_op;
} else {
if (is_not_set(data_set->flags, pe_flag_concurrent_fencing)
&& (stonith_ops != NULL)) {
/* Concurrent fencing is disabled, so order each non-DC
* fencing in a chain. If there is any DC fencing or
* shutdown, it will be ordered after the last action in the
* chain later.
*/
order_actions((pe_action_t *) stonith_ops->data,
stonith_op, pe_order_optional);
}
// Remember all non-DC fencing actions in a separate list
stonith_ops = g_list_prepend(stonith_ops, stonith_op);
}
} else if (node->details->online && node->details->shutdown &&
/* TODO define what a shutdown op means for a remote node.
* For now we do not send shutdown operations for remote nodes, but
* if we can come up with a good use for this in the future, we will. */
pe__is_guest_or_remote_node(node) == FALSE) {
action_t *down_op = sched_shutdown_op(node, data_set);
if (node->details->is_dc) {
// Remember if the DC is being shut down
dc_down = down_op;
} else {
// Remember non-DC shutdowns for later ordering
shutdown_ops = g_list_prepend(shutdown_ops, down_op);
}
}
if (node->details->unclean && stonith_op == NULL) {
integrity_lost = TRUE;
pe_warn("Node %s is unclean!", node->details->uname);
}
}
if (integrity_lost) {
if (is_set(data_set->flags, pe_flag_stonith_enabled) == FALSE) {
pe_warn("YOUR RESOURCES ARE NOW LIKELY COMPROMISED");
pe_err("ENABLE STONITH TO KEEP YOUR RESOURCES SAFE");
} else if (is_set(data_set->flags, pe_flag_have_quorum) == FALSE) {
crm_notice("Cannot fence unclean nodes until quorum is"
" attained (or no-quorum-policy is set to ignore)");
}
}
if (dc_down != NULL) {
/* Order any non-DC shutdowns before any DC shutdown, to avoid repeated
* DC elections. However, we don't want to order non-DC shutdowns before
* a DC *fencing*, because even though we don't want a node that's
* shutting down to become DC, the DC fencing could be ordered before a
* clone stop that's also ordered before the shutdowns, thus leading to
* a graph loop.
*/
if (safe_str_eq(dc_down->task, CRM_OP_SHUTDOWN)) {
for (gIter = shutdown_ops; gIter != NULL; gIter = gIter->next) {
action_t *node_stop = (action_t *) gIter->data;
crm_debug("Ordering shutdown on %s before %s on DC %s",
node_stop->node->details->uname,
dc_down->task, dc_down->node->details->uname);
order_actions(node_stop, dc_down, pe_order_optional);
}
}
// Order any non-DC fencing before any DC fencing or shutdown
if (is_set(data_set->flags, pe_flag_concurrent_fencing)) {
/* With concurrent fencing, order each non-DC fencing action
* separately before any DC fencing or shutdown.
*/
for (gIter = stonith_ops; gIter != NULL; gIter = gIter->next) {
order_actions((pe_action_t *) gIter->data, dc_down,
pe_order_optional);
}
} else if (stonith_ops) {
/* Without concurrent fencing, the non-DC fencing actions are
* already ordered relative to each other, so we just need to order
* the DC fencing after the last action in the chain (which is the
* first item in the list).
*/
order_actions((pe_action_t *) stonith_ops->data, dc_down,
pe_order_optional);
}
}
g_list_free(stonith_ops);
g_list_free(shutdown_ops);
return TRUE;
}
/*
* Determine the sets of independent actions and the correct order for the
* actions in each set.
*
* Mark dependencies of un-runnable actions un-runnable
*
*/
static GListPtr
find_actions_by_task(GListPtr actions, resource_t * rsc, const char *original_key)
{
GListPtr list = NULL;
list = find_actions(actions, original_key, NULL);
if (list == NULL) {
/* we're potentially searching a child of the original resource */
char *key = NULL;
char *task = NULL;
guint interval_ms = 0;
if (parse_op_key(original_key, NULL, &task, &interval_ms)) {
key = generate_op_key(rsc->id, task, interval_ms);
list = find_actions(actions, key, NULL);
} else {
crm_err("search key: %s", original_key);
}
free(key);
free(task);
}
return list;
}
static void
rsc_order_then(pe_action_t *lh_action, pe_resource_t *rsc,
pe__ordering_t *order)
{
GListPtr gIter = NULL;
GListPtr rh_actions = NULL;
action_t *rh_action = NULL;
enum pe_ordering type;
CRM_CHECK(rsc != NULL, return);
CRM_CHECK(order != NULL, return);
type = order->type;
rh_action = order->rh_action;
crm_trace("Processing RH of ordering constraint %d", order->id);
if (rh_action != NULL) {
rh_actions = g_list_prepend(NULL, rh_action);
} else if (rsc != NULL) {
rh_actions = find_actions_by_task(rsc->actions, rsc, order->rh_action_task);
}
if (rh_actions == NULL) {
pe_rsc_trace(rsc, "No RH-Side (%s/%s) found for constraint..."
" ignoring", rsc->id, order->rh_action_task);
if (lh_action) {
pe_rsc_trace(rsc, "LH-Side was: %s", lh_action->uuid);
}
return;
}
if (lh_action && lh_action->rsc == rsc && is_set(lh_action->flags, pe_action_dangle)) {
pe_rsc_trace(rsc, "Detected dangling operation %s -> %s", lh_action->uuid,
order->rh_action_task);
clear_bit(type, pe_order_implies_then);
}
gIter = rh_actions;
for (; gIter != NULL; gIter = gIter->next) {
action_t *rh_action_iter = (action_t *) gIter->data;
if (lh_action) {
order_actions(lh_action, rh_action_iter, type);
} else if (type & pe_order_implies_then) {
update_action_flags(rh_action_iter, pe_action_runnable | pe_action_clear, __FUNCTION__, __LINE__);
crm_warn("Unrunnable %s 0x%.6x", rh_action_iter->uuid, type);
} else {
crm_warn("neither %s 0x%.6x", rh_action_iter->uuid, type);
}
}
g_list_free(rh_actions);
}
static void
rsc_order_first(pe_resource_t *lh_rsc, pe__ordering_t *order,
pe_working_set_t *data_set)
{
GListPtr gIter = NULL;
GListPtr lh_actions = NULL;
action_t *lh_action = order->lh_action;
resource_t *rh_rsc = order->rh_rsc;
crm_trace("Processing LH of ordering constraint %d", order->id);
CRM_ASSERT(lh_rsc != NULL);
if (lh_action != NULL) {
lh_actions = g_list_prepend(NULL, lh_action);
} else {
lh_actions = find_actions_by_task(lh_rsc->actions, lh_rsc, order->lh_action_task);
}
if (lh_actions == NULL && lh_rsc != rh_rsc) {
char *key = NULL;
char *op_type = NULL;
guint interval_ms = 0;
parse_op_key(order->lh_action_task, NULL, &op_type, &interval_ms);
key = generate_op_key(lh_rsc->id, op_type, interval_ms);
if (lh_rsc->fns->state(lh_rsc, TRUE) == RSC_ROLE_STOPPED && safe_str_eq(op_type, RSC_STOP)) {
free(key);
pe_rsc_trace(lh_rsc, "No LH-Side (%s/%s) found for constraint %d with %s - ignoring",
lh_rsc->id, order->lh_action_task, order->id, order->rh_action_task);
} else if (lh_rsc->fns->state(lh_rsc, TRUE) == RSC_ROLE_SLAVE && safe_str_eq(op_type, RSC_DEMOTE)) {
free(key);
pe_rsc_trace(lh_rsc, "No LH-Side (%s/%s) found for constraint %d with %s - ignoring",
lh_rsc->id, order->lh_action_task, order->id, order->rh_action_task);
} else {
pe_rsc_trace(lh_rsc, "No LH-Side (%s/%s) found for constraint %d with %s - creating",
lh_rsc->id, order->lh_action_task, order->id, order->rh_action_task);
lh_action = custom_action(lh_rsc, key, op_type, NULL, TRUE, TRUE, data_set);
lh_actions = g_list_prepend(NULL, lh_action);
}
free(op_type);
}
gIter = lh_actions;
for (; gIter != NULL; gIter = gIter->next) {
action_t *lh_action_iter = (action_t *) gIter->data;
if (rh_rsc == NULL && order->rh_action) {
rh_rsc = order->rh_action->rsc;
}
if (rh_rsc) {
rsc_order_then(lh_action_iter, rh_rsc, order);
} else if (order->rh_action) {
order_actions(lh_action_iter, order->rh_action, order->type);
}
}
g_list_free(lh_actions);
}
extern void update_colo_start_chain(pe_action_t *action,
pe_working_set_t *data_set);
static int
is_recurring_action(action_t *action)
{
const char *interval_ms_s = g_hash_table_lookup(action->meta,
XML_LRM_ATTR_INTERVAL_MS);
guint interval_ms = crm_parse_ms(interval_ms_s);
return (interval_ms > 0);
}
static void
apply_container_ordering(action_t *action, pe_working_set_t *data_set)
{
/* VMs are also classified as containers for these purposes... in
* that they both involve a 'thing' running on a real or remote
* cluster node.
*
* This allows us to be smarter about the type and extent of
* recovery actions required in various scenarios
*/
resource_t *remote_rsc = NULL;
resource_t *container = NULL;
enum action_tasks task = text2task(action->task);
CRM_ASSERT(action->rsc);
CRM_ASSERT(action->node);
CRM_ASSERT(pe__is_guest_or_remote_node(action->node));
remote_rsc = action->node->details->remote_rsc;
CRM_ASSERT(remote_rsc);
container = remote_rsc->container;
CRM_ASSERT(container);
if(is_set(container->flags, pe_rsc_failed)) {
pe_fence_node(data_set, action->node, "container failed");
}
crm_trace("Order %s action %s relative to %s%s for %s%s",
action->task, action->uuid,
is_set(remote_rsc->flags, pe_rsc_failed)? "failed " : "",
remote_rsc->id,
is_set(container->flags, pe_rsc_failed)? "failed " : "",
container->id);
if (safe_str_eq(action->task, CRMD_ACTION_MIGRATE)
|| safe_str_eq(action->task, CRMD_ACTION_MIGRATED)) {
/* Migration ops map to "no_action", but we need to apply the same
* ordering as for stop or demote (see get_router_node()).
*/
task = stop_rsc;
}
switch (task) {
case start_rsc:
case action_promote:
/* Force resource recovery if the container is recovered */
order_start_then_action(container, action, pe_order_implies_then,
data_set);
/* Wait for the connection resource to be up too */
order_start_then_action(remote_rsc, action, pe_order_none,
data_set);
break;
case stop_rsc:
case action_demote:
if (is_set(container->flags, pe_rsc_failed)) {
/* When the container representing a guest node fails, any stop
* or demote actions for resources running on the guest node
* are implied by the container stopping. This is similar to
* how fencing operations work for cluster nodes and remote
* nodes.
*/
} else {
/* Ensure the operation happens before the connection is brought
* down.
*
* If we really wanted to, we could order these after the
* connection start, IFF the container's current role was
* stopped (otherwise we re-introduce an ordering loop when the
* connection is restarting).
*/
order_action_then_stop(action, remote_rsc, pe_order_none,
data_set);
}
break;
default:
/* Wait for the connection resource to be up */
if (is_recurring_action(action)) {
/* In case we ever get the recovery logic wrong, force
* recurring monitors to be restarted, even if just
* the connection was re-established
*/
if(task != no_action) {
order_start_then_action(remote_rsc, action,
pe_order_implies_then, data_set);
}
} else {
order_start_then_action(remote_rsc, action, pe_order_none,
data_set);
}
break;
}
}
static enum remote_connection_state
get_remote_node_state(pe_node_t *node)
{
resource_t *remote_rsc = NULL;
node_t *cluster_node = NULL;
CRM_ASSERT(node);
remote_rsc = node->details->remote_rsc;
CRM_ASSERT(remote_rsc);
cluster_node = pe__current_node(remote_rsc);
/* If the cluster node the remote connection resource resides on
* is unclean or went offline, we can't process any operations
* on that remote node until after it starts elsewhere.
*/
if(remote_rsc->next_role == RSC_ROLE_STOPPED || remote_rsc->allocated_to == NULL) {
/* The connection resource is not going to run anywhere */
if (cluster_node && cluster_node->details->unclean) {
/* The remote connection is failed because its resource is on a
* failed node and can't be recovered elsewhere, so we must fence.
*/
return remote_state_failed;
}
if (is_not_set(remote_rsc->flags, pe_rsc_failed)) {
/* Connection resource is cleanly stopped */
return remote_state_stopped;
}
/* Connection resource is failed */
if ((remote_rsc->next_role == RSC_ROLE_STOPPED)
&& remote_rsc->remote_reconnect_ms
&& node->details->remote_was_fenced
&& !pe__shutdown_requested(node)) {
/* We won't know whether the connection is recoverable until the
* reconnect interval expires and we reattempt connection.
*/
return remote_state_unknown;
}
/* The remote connection is in a failed state. If there are any
* resources known to be active on it (stop) or in an unknown state
* (probe), we must assume the worst and fence it.
*/
return remote_state_failed;
} else if (cluster_node == NULL) {
/* Connection is recoverable but not currently running anywhere, see if we can recover it first */
return remote_state_unknown;
} else if(cluster_node->details->unclean == TRUE
|| cluster_node->details->online == FALSE) {
/* Connection is running on a dead node, see if we can recover it first */
return remote_state_resting;
} else if (g_list_length(remote_rsc->running_on) > 1
&& remote_rsc->partial_migration_source
&& remote_rsc->partial_migration_target) {
/* We're in the middle of migrating a connection resource,
* wait until after the resource migrates before performing
* any actions.
*/
return remote_state_resting;
}
return remote_state_alive;
}
/*!
* \internal
* \brief Order actions on remote node relative to actions for the connection
*/
static void
apply_remote_ordering(action_t *action, pe_working_set_t *data_set)
{
resource_t *remote_rsc = NULL;
enum action_tasks task = text2task(action->task);
enum remote_connection_state state = get_remote_node_state(action->node);
enum pe_ordering order_opts = pe_order_none;
if (action->rsc == NULL) {
return;
}
CRM_ASSERT(action->node);
CRM_ASSERT(pe__is_guest_or_remote_node(action->node));
remote_rsc = action->node->details->remote_rsc;
CRM_ASSERT(remote_rsc);
crm_trace("Order %s action %s relative to %s%s (state: %s)",
action->task, action->uuid,
is_set(remote_rsc->flags, pe_rsc_failed)? "failed " : "",
remote_rsc->id, state2text(state));
if (safe_str_eq(action->task, CRMD_ACTION_MIGRATE)
|| safe_str_eq(action->task, CRMD_ACTION_MIGRATED)) {
/* Migration ops map to "no_action", but we need to apply the same
* ordering as for stop or demote (see get_router_node()).
*/
task = stop_rsc;
}
switch (task) {
case start_rsc:
case action_promote:
order_opts = pe_order_none;
if (state == remote_state_failed) {
/* Force recovery, by making this action required */
order_opts |= pe_order_implies_then;
}
/* Ensure connection is up before running this action */
order_start_then_action(remote_rsc, action, order_opts, data_set);
break;
case stop_rsc:
if(state == remote_state_alive) {
order_action_then_stop(action, remote_rsc,
pe_order_implies_first, data_set);
} else if(state == remote_state_failed) {
- /* We would only be here if the resource is
- * running on the remote node. Since we have no
- * way to stop it, it is necessary to fence the
- * node.
+ /* The resource is active on the node, but since we don't have a
+ * valid connection, the only way to stop the resource is by
+ * fencing the node. There is no need to order the stop relative
+ * to the remote connection, since the stop will become implied
+ * by the fencing.
*/
pe_fence_node(data_set, action->node, "resources are active and the connection is unrecoverable");
- order_action_then_stop(action, remote_rsc,
- pe_order_implies_first, data_set);
} else if(remote_rsc->next_role == RSC_ROLE_STOPPED) {
/* State must be remote_state_unknown or remote_state_stopped.
* Since the connection is not coming back up in this
* transition, stop this resource first.
*/
order_action_then_stop(action, remote_rsc,
pe_order_implies_first, data_set);
} else {
/* The connection is going to be started somewhere else, so
* stop this resource after that completes.
*/
order_start_then_action(remote_rsc, action, pe_order_none, data_set);
}
break;
case action_demote:
/* Only order this demote relative to the connection start if the
* connection isn't being torn down. Otherwise, the demote would be
* blocked because the connection start would not be allowed.
*/
if(state == remote_state_resting || state == remote_state_unknown) {
order_start_then_action(remote_rsc, action, pe_order_none,
data_set);
} /* Otherwise we can rely on the stop ordering */
break;
default:
/* Wait for the connection resource to be up */
if (is_recurring_action(action)) {
/* In case we ever get the recovery logic wrong, force
* recurring monitors to be restarted, even if just
* the connection was re-established
*/
order_start_then_action(remote_rsc, action,
pe_order_implies_then, data_set);
} else {
node_t *cluster_node = pe__current_node(remote_rsc);
if(task == monitor_rsc && state == remote_state_failed) {
/* We would only be here if we do not know the
* state of the resource on the remote node.
* Since we have no way to find out, it is
* necessary to fence the node.
*/
pe_fence_node(data_set, action->node, "resources are in an unknown state and the connection is unrecoverable");
}
if(cluster_node && state == remote_state_stopped) {
/* The connection is currently up, but is going
* down permanently.
*
* Make sure we check services are actually
* stopped _before_ we let the connection get
* closed
*/
order_action_then_stop(action, remote_rsc,
pe_order_runnable_left, data_set);
} else {
order_start_then_action(remote_rsc, action, pe_order_none,
data_set);
}
}
break;
}
}
static void
apply_remote_node_ordering(pe_working_set_t *data_set)
{
if (is_set(data_set->flags, pe_flag_have_remote_nodes) == FALSE) {
return;
}
for (GListPtr gIter = data_set->actions; gIter != NULL; gIter = gIter->next) {
action_t *action = (action_t *) gIter->data;
resource_t *remote = NULL;
// We are only interested in resource actions
if (action->rsc == NULL) {
continue;
}
/* Special case: If we are clearing the failcount of an actual
* remote connection resource, then make sure this happens before
* any start of the resource in this transition.
*/
if (action->rsc->is_remote_node &&
safe_str_eq(action->task, CRM_OP_CLEAR_FAILCOUNT)) {
custom_action_order(action->rsc,
NULL,
action,
action->rsc,
generate_op_key(action->rsc->id, RSC_START, 0),
NULL,
pe_order_optional,
data_set);
continue;
}
// We are only interested in actions allocated to a node
if (action->node == NULL) {
continue;
}
if (!pe__is_guest_or_remote_node(action->node)) {
continue;
}
/* We are only interested in real actions.
*
* @TODO This is probably wrong; pseudo-actions might be converted to
* real actions and vice versa later in update_actions() at the end of
* stage7().
*/
if (is_set(action->flags, pe_action_pseudo)) {
continue;
}
remote = action->node->details->remote_rsc;
if (remote == NULL) {
// Orphaned
continue;
}
/* Another special case: if a resource is moving to a Pacemaker Remote
* node, order the stop on the original node after any start of the
* remote connection. This ensures that if the connection fails to
* start, we leave the resource running on the original node.
*/
if (safe_str_eq(action->task, RSC_START)) {
for (GList *item = action->rsc->actions; item != NULL;
item = item->next) {
pe_action_t *rsc_action = item->data;
if ((rsc_action->node->details != action->node->details)
&& safe_str_eq(rsc_action->task, RSC_STOP)) {
custom_action_order(remote, start_key(remote), NULL,
action->rsc, NULL, rsc_action,
pe_order_optional, data_set);
}
}
}
/* The action occurs across a remote connection, so create
* ordering constraints that guarantee the action occurs while the node
* is active (after start, before stop ... things like that).
*
* This is somewhat brittle in that we need to make sure the results of
* this ordering are compatible with the result of get_router_node().
* It would probably be better to add XML_LRM_ATTR_ROUTER_NODE as part
* of this logic rather than action2xml().
*/
if (remote->container) {
crm_trace("Container ordering for %s", action->uuid);
apply_container_ordering(action, data_set);
} else {
crm_trace("Remote ordering for %s", action->uuid);
apply_remote_ordering(action, data_set);
}
}
}
static gboolean
order_first_probe_unneeded(pe_action_t * probe, pe_action_t * rh_action)
{
/* No need to probe the resource on the node that is being
* unfenced. Otherwise it might introduce transition loop
* since probe will be performed after the node is
* unfenced.
*/
if (safe_str_eq(rh_action->task, CRM_OP_FENCE)
&& probe->node && rh_action->node
&& probe->node->details == rh_action->node->details) {
const char *op = g_hash_table_lookup(rh_action->meta, "stonith_action");
if (safe_str_eq(op, "on")) {
return TRUE;
}
}
// Shutdown waits for probe to complete only if it's on the same node
if ((safe_str_eq(rh_action->task, CRM_OP_SHUTDOWN))
&& probe->node && rh_action->node
&& probe->node->details != rh_action->node->details) {
return TRUE;
}
return FALSE;
}
static void
order_first_probes(pe_working_set_t * data_set)
{
GListPtr gIter = NULL;
for (gIter = data_set->ordering_constraints; gIter != NULL; gIter = gIter->next) {
pe__ordering_t *order = gIter->data;
enum pe_ordering order_type = pe_order_optional;
pe_resource_t *lh_rsc = order->lh_rsc;
pe_resource_t *rh_rsc = order->rh_rsc;
pe_action_t *lh_action = order->lh_action;
pe_action_t *rh_action = order->rh_action;
const char *lh_action_task = order->lh_action_task;
const char *rh_action_task = order->rh_action_task;
GListPtr probes = NULL;
GListPtr rh_actions = NULL;
GListPtr pIter = NULL;
if (lh_rsc == NULL) {
continue;
} else if (rh_rsc && lh_rsc == rh_rsc) {
continue;
}
if (lh_action == NULL && lh_action_task == NULL) {
continue;
}
if (rh_action == NULL && rh_action_task == NULL) {
continue;
}
/* Technically probe is expected to return "not running", which could be
* the alternative of stop action if the status of the resource is
* unknown yet.
*/
if (lh_action && safe_str_neq(lh_action->task, RSC_STOP)) {
continue;
} else if (lh_action == NULL
&& lh_action_task
&& crm_ends_with(lh_action_task, "_" RSC_STOP "_0") == FALSE) {
continue;
}
/* Do not probe the resource inside of a stopping container. Otherwise
* it might introduce transition loop since probe will be performed
* after the container starts again.
*/
if (rh_rsc && lh_rsc->container == rh_rsc) {
if (rh_action && safe_str_eq(rh_action->task, RSC_STOP)) {
continue;
} else if (rh_action == NULL && rh_action_task
&& crm_ends_with(rh_action_task,"_" RSC_STOP "_0")) {
continue;
}
}
if (order->type == pe_order_none) {
continue;
}
// Preserve the order options for future filtering
if (is_set(order->type, pe_order_apply_first_non_migratable)) {
set_bit(order_type, pe_order_apply_first_non_migratable);
}
if (is_set(order->type, pe_order_same_node)) {
set_bit(order_type, pe_order_same_node);
}
// Keep the order types for future filtering
if (order->type == pe_order_anti_colocation
|| order->type == pe_order_load) {
order_type = order->type;
}
probes = pe__resource_actions(lh_rsc, NULL, RSC_STATUS, FALSE);
if (probes == NULL) {
continue;
}
if (rh_action) {
rh_actions = g_list_prepend(rh_actions, rh_action);
} else if (rh_rsc && rh_action_task) {
rh_actions = find_actions(rh_rsc->actions, rh_action_task, NULL);
}
if (rh_actions == NULL) {
g_list_free(probes);
continue;
}
crm_trace("Processing for LH probe based on ordering constraint %s -> %s"
" (id=%d, type=%.6x)",
lh_action ? lh_action->uuid : lh_action_task,
rh_action ? rh_action->uuid : rh_action_task,
order->id, order->type);
for (pIter = probes; pIter != NULL; pIter = pIter->next) {
pe_action_t *probe = (pe_action_t *) pIter->data;
GListPtr rIter = NULL;
for (rIter = rh_actions; rIter != NULL; rIter = rIter->next) {
pe_action_t *rh_action_iter = (pe_action_t *) rIter->data;
if (order_first_probe_unneeded(probe, rh_action_iter)) {
continue;
}
order_actions(probe, rh_action_iter, order_type);
}
}
g_list_free(rh_actions);
g_list_free(probes);
}
}
static void
order_then_probes(pe_working_set_t * data_set)
{
#if 0
GListPtr gIter = NULL;
for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) {
resource_t *rsc = (resource_t *) gIter->data;
/* Given "A then B", we would prefer to wait for A to be
* started before probing B.
*
* If A was a filesystem on which the binaries and data for B
* lived, it would have been useful if the author of B's agent
* could assume that A is running before B.monitor will be
* called.
*
* However we can't _only_ probe once A is running, otherwise
* we'd not detect the state of B if A could not be started
* for some reason.
*
* In practice however, we cannot even do an opportunistic
* version of this because B may be moving:
*
* B.probe -> B.start
* B.probe -> B.stop
* B.stop -> B.start
* A.stop -> A.start
* A.start -> B.probe
*
* So far so good, but if we add the result of this code:
*
* B.stop -> A.stop
*
* Then we get a loop:
*
* B.probe -> B.stop -> A.stop -> A.start -> B.probe
*
* We could kill the 'B.probe -> B.stop' dependency, but that
* could mean stopping B "too" soon, because B.start must wait
* for the probes to complete.
*
* Another option is to allow it only if A is a non-unique
* clone with clone-max == node-max (since we'll never be
* moving it). However, we could still be stopping one
* instance at the same time as starting another.
* The complexity of checking for allowed conditions combined
* with the ever narrowing usecase suggests that this code
* should remain disabled until someone gets smarter.
*/
action_t *start = NULL;
GListPtr actions = NULL;
GListPtr probes = NULL;
actions = pe__resource_actions(rsc, NULL, RSC_START, FALSE);
if (actions) {
start = actions->data;
g_list_free(actions);
}
if(start == NULL) {
crm_err("No start action for %s", rsc->id);
continue;
}
probes = pe__resource_actions(rsc, NULL, RSC_STATUS, FALSE);
for (actions = start->actions_before; actions != NULL; actions = actions->next) {
action_wrapper_t *before = (action_wrapper_t *) actions->data;
GListPtr pIter = NULL;
action_t *first = before->action;
resource_t *first_rsc = first->rsc;
if(first->required_runnable_before) {
GListPtr clone_actions = NULL;
for (clone_actions = first->actions_before; clone_actions != NULL; clone_actions = clone_actions->next) {
before = (action_wrapper_t *) clone_actions->data;
crm_trace("Testing %s -> %s (%p) for %s", first->uuid, before->action->uuid, before->action->rsc, start->uuid);
CRM_ASSERT(before->action->rsc);
first_rsc = before->action->rsc;
break;
}
} else if(safe_str_neq(first->task, RSC_START)) {
crm_trace("Not a start op %s for %s", first->uuid, start->uuid);
}
if(first_rsc == NULL) {
continue;
} else if(uber_parent(first_rsc) == uber_parent(start->rsc)) {
crm_trace("Same parent %s for %s", first_rsc->id, start->uuid);
continue;
} else if(FALSE && pe_rsc_is_clone(uber_parent(first_rsc)) == FALSE) {
crm_trace("Not a clone %s for %s", first_rsc->id, start->uuid);
continue;
}
crm_err("Applying %s before %s %d", first->uuid, start->uuid, uber_parent(first_rsc)->variant);
for (pIter = probes; pIter != NULL; pIter = pIter->next) {
action_t *probe = (action_t *) pIter->data;
crm_err("Ordering %s before %s", first->uuid, probe->uuid);
order_actions(first, probe, pe_order_optional);
}
}
}
#endif
}
static void
order_probes(pe_working_set_t * data_set)
{
order_first_probes(data_set);
order_then_probes(data_set);
}
gboolean
stage7(pe_working_set_t * data_set)
{
GListPtr gIter = NULL;
crm_trace("Applying ordering constraints");
/* Don't ask me why, but apparently they need to be processed in
* the order they were created in... go figure
*
* Also g_list_append() has horrendous performance characteristics
* So we need to use g_list_prepend() and then reverse the list here
*/
data_set->ordering_constraints = g_list_reverse(data_set->ordering_constraints);
for (gIter = data_set->ordering_constraints; gIter != NULL; gIter = gIter->next) {
pe__ordering_t *order = gIter->data;
resource_t *rsc = order->lh_rsc;
crm_trace("Applying ordering constraint: %d", order->id);
if (rsc != NULL) {
crm_trace("rsc_action-to-*");
rsc_order_first(rsc, order, data_set);
continue;
}
rsc = order->rh_rsc;
if (rsc != NULL) {
crm_trace("action-to-rsc_action");
rsc_order_then(order->lh_action, rsc, order);
} else {
crm_trace("action-to-action");
order_actions(order->lh_action, order->rh_action, order->type);
}
}
for (gIter = data_set->actions; gIter != NULL; gIter = gIter->next) {
action_t *action = (action_t *) gIter->data;
update_colo_start_chain(action, data_set);
}
crm_trace("Ordering probes");
order_probes(data_set);
crm_trace("Updating %d actions", g_list_length(data_set->actions));
for (gIter = data_set->actions; gIter != NULL; gIter = gIter->next) {
action_t *action = (action_t *) gIter->data;
update_action(action, data_set);
}
LogNodeActions(data_set, FALSE);
for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) {
resource_t *rsc = (resource_t *) gIter->data;
LogActions(rsc, data_set, FALSE);
}
return TRUE;
}
static int transition_id = -1;
/*!
* \internal
* \brief Log a message after calculating a transition
*
* \param[in] filename Where transition input is stored
*/
void
pcmk__log_transition_summary(const char *filename)
{
if (was_processing_error) {
crm_err("Calculated transition %d (with errors), saving inputs in %s",
transition_id, filename);
} else if (was_processing_warning) {
crm_warn("Calculated transition %d (with warnings), saving inputs in %s",
transition_id, filename);
} else {
crm_notice("Calculated transition %d, saving inputs in %s",
transition_id, filename);
}
if (crm_config_error) {
crm_notice("Configuration errors found during scheduler processing,"
" please run \"crm_verify -L\" to identify issues");
}
}
/*
* Create a dependency graph to send to the transitioner (via the controller)
*/
gboolean
stage8(pe_working_set_t * data_set)
{
GListPtr gIter = NULL;
const char *value = NULL;
transition_id++;
crm_trace("Creating transition graph %d.", transition_id);
data_set->graph = create_xml_node(NULL, XML_TAG_GRAPH);
value = pe_pref(data_set->config_hash, "cluster-delay");
crm_xml_add(data_set->graph, "cluster-delay", value);
value = pe_pref(data_set->config_hash, "stonith-timeout");
crm_xml_add(data_set->graph, "stonith-timeout", value);
crm_xml_add(data_set->graph, "failed-stop-offset", "INFINITY");
if (is_set(data_set->flags, pe_flag_start_failure_fatal)) {
crm_xml_add(data_set->graph, "failed-start-offset", "INFINITY");
} else {
crm_xml_add(data_set->graph, "failed-start-offset", "1");
}
value = pe_pref(data_set->config_hash, "batch-limit");
crm_xml_add(data_set->graph, "batch-limit", value);
crm_xml_add_int(data_set->graph, "transition_id", transition_id);
value = pe_pref(data_set->config_hash, "migration-limit");
if (crm_int_helper(value, NULL) > 0) {
crm_xml_add(data_set->graph, "migration-limit", value);
}
/* errors...
slist_iter(action, action_t, action_list, lpc,
if(action->optional == FALSE && action->runnable == FALSE) {
print_action("Ignoring", action, TRUE);
}
);
*/
gIter = data_set->resources;
for (; gIter != NULL; gIter = gIter->next) {
resource_t *rsc = (resource_t *) gIter->data;
pe_rsc_trace(rsc, "processing actions for rsc=%s", rsc->id);
rsc->cmds->expand(rsc, data_set);
}
crm_log_xml_trace(data_set->graph, "created resource-driven action list");
/* pseudo action to distribute list of nodes with maintenance state update */
add_maintenance_update(data_set);
/* catch any non-resource specific actions */
crm_trace("processing non-resource actions");
gIter = data_set->actions;
for (; gIter != NULL; gIter = gIter->next) {
action_t *action = (action_t *) gIter->data;
if (action->rsc
&& action->node
&& action->node->details->shutdown
&& is_not_set(action->rsc->flags, pe_rsc_maintenance)
&& is_not_set(action->flags, pe_action_optional)
&& is_not_set(action->flags, pe_action_runnable)
&& crm_str_eq(action->task, RSC_STOP, TRUE)
) {
/* Eventually we should just ignore the 'fence' case
* But for now it's the best way to detect (in CTS) when
* CIB resource updates are being lost
*/
if (is_set(data_set->flags, pe_flag_have_quorum)
|| data_set->no_quorum_policy == no_quorum_ignore) {
crm_crit("Cannot %s node '%s' because of %s:%s%s (%s)",
action->node->details->unclean ? "fence" : "shut down",
action->node->details->uname, action->rsc->id,
is_not_set(action->rsc->flags, pe_rsc_managed) ? " unmanaged" : " blocked",
is_set(action->rsc->flags, pe_rsc_failed) ? " failed" : "",
action->uuid);
}
}
graph_element_from_action(action, data_set);
}
crm_log_xml_trace(data_set->graph, "created generic action list");
crm_trace("Created transition graph %d.", transition_id);
return TRUE;
}
void
LogNodeActions(pe_working_set_t * data_set, gboolean terminal)
{
GListPtr gIter = NULL;
for (gIter = data_set->actions; gIter != NULL; gIter = gIter->next) {
char *node_name = NULL;
char *task = NULL;
action_t *action = (action_t *) gIter->data;
if (action->rsc != NULL) {
continue;
} else if (is_set(action->flags, pe_action_optional)) {
continue;
}
if (pe__is_guest_node(action->node)) {
node_name = crm_strdup_printf("%s (resource: %s)", action->node->details->uname, action->node->details->remote_rsc->container->id);
} else if(action->node) {
node_name = crm_strdup_printf("%s", action->node->details->uname);
}
if (safe_str_eq(action->task, CRM_OP_SHUTDOWN)) {
task = strdup("Shutdown");
} else if (safe_str_eq(action->task, CRM_OP_FENCE)) {
const char *op = g_hash_table_lookup(action->meta, "stonith_action");
task = crm_strdup_printf("Fence (%s)", op);
}
if(task == NULL) {
/* Nothing to report */
} else if(terminal && action->reason) {
printf(" * %s %s '%s'\n", task, node_name, action->reason);
} else if(terminal) {
printf(" * %s %s\n", task, node_name);
} else if(action->reason) {
crm_notice(" * %s %s '%s'\n", task, node_name, action->reason);
} else {
crm_notice(" * %s %s\n", task, node_name);
}
free(node_name);
free(task);
}
}