Page MenuHomeClusterLabs Projects

No OneTemporary

diff --git a/include/crm/common/action_relation_internal.h b/include/crm/common/action_relation_internal.h
index ee8c026091..7d7d7b19b6 100644
--- a/include/crm/common/action_relation_internal.h
+++ b/include/crm/common/action_relation_internal.h
@@ -1,115 +1,118 @@
/*
* Copyright 2023 the Pacemaker project contributors
*
* The version control history for this file may have further details.
*
* This source code is licensed under the GNU Lesser General Public License
* version 2.1 or later (LGPLv2.1+) WITHOUT ANY WARRANTY.
*/
#ifndef PCMK__CRM_COMMON_ACTION_RELATION_INTERNAL__H
# define PCMK__CRM_COMMON_ACTION_RELATION_INTERNAL__H
/*!
* Flags to indicate the relationship between two actions
*
* @COMPAT The values and semantics of these flags should not be changed until
* the deprecated enum pe_ordering is dropped from the public API.
*/
enum pcmk__action_relation_flags {
//! No relation (compare with equality rather than bit set)
pcmk__ar_none = 0U,
//! Actions are ordered (optionally, if no other flags are set)
pcmk__ar_ordered = (1U << 0),
//! Relation applies only if 'first' cannot be part of a live migration
pcmk__ar_if_first_unmigratable = (1U << 1),
/*!
* If 'then' is required, 'first' becomes required (and becomes unmigratable
* if 'then' is); also, if 'first' is a stop of a blocked resource, 'then'
* becomes unrunnable
*/
pcmk__ar_then_implies_first = (1U << 4),
/*!
* If 'first' is required, 'then' becomes required; if 'first' is a stop of
* a blocked resource, 'then' becomes unrunnable
*/
pcmk__ar_first_implies_then = (1U << 5),
/*!
* If 'then' is required and for a promoted instance, 'first' becomes
* required (and becomes unmigratable if 'then' is)
*/
pcmk__ar_promoted_then_implies_first = (1U << 6),
/*!
* 'first' is runnable only if 'then' is both runnable and migratable,
* and 'first' becomes required if 'then' is
*/
pcmk__ar_unmigratable_then_blocks = (1U << 7),
//! 'then' is runnable (and migratable) only if 'first' is runnable
pcmk__ar_unrunnable_first_blocks = (1U << 8),
//! If 'first' is unrunnable, 'then' becomes a real, unmigratable action
pcmk__ar_first_else_then = (1U << 9),
//! If 'first' is required, 'then' action for instance on same node is
pcmk__ar_first_implies_same_node_then = (1U << 10),
/*!
* Disable relation if 'first' is unrunnable and for an active resource,
* otherwise order actions and make 'then' unrunnable if 'first' is.
*
* This is used to order a bundle replica's start of its container before a
* probe of its remote connection resource, in case the connection uses the
* REMOTE_CONTAINER_HACK to replace the connection address with where the
* container is running.
*/
pcmk__ar_nested_remote_probe = (1U << 11),
/*!
* If 'first' is for a blocked resource, make 'then' unrunnable.
*
* If 'then' is required, make 'first' required, make 'first' unmigratable
* if 'then' is unmigratable, and make 'then' unrunnable if 'first' is
* unrunnable.
*
* If 'then' is unrunnable and for the same resource as 'first', make
* 'first' required if it is runnable, and make 'first' unmigratable if
* 'then' is unmigratable.
*
* This is used for "stop then start primitive" (restarts) and
* "stop group member then stop previous member".
*/
pcmk__ar_intermediate_stop = (1U << 12),
/*!
* The actions must be serialized if in the same transition but can be in
* either order. (In practice, we always arrange them as 'first' then
* 'then', so they end up being essentially the same as optional orderings.)
*
* @TODO Handle more intelligently -- for example, we could schedule the
* action with the fewest inputs first, so we're more likely to execute at
* least one if there is a failure during the transition. Or, we could
* prefer certain action types over others, or base it on resource priority.
*/
pcmk__ar_serialize = (1U << 14),
//! Relation applies only if actions are on same node
pcmk__ar_if_on_same_node = (1U << 15),
//! If 'then' is required, 'first' must be added to the transition graph
pcmk__ar_then_implies_first_graphed = (1U << 16),
//! If 'first' is required and runnable, 'then' must be in graph
pcmk__ar_first_implies_then_graphed = (1U << 17),
//! User-configured asymmetric ordering
pcmk__ar_asymmetric = (1U << 20),
+
+ //! Actions are ordered if on same node (or migration target for migrate_to)
+ pcmk__ar_if_on_same_node_or_target = (1U << 21),
};
#endif // PCMK__CRM_COMMON_ACTION_RELATION_INTERNAL__H
diff --git a/lib/pacemaker/pcmk_graph_producer.c b/lib/pacemaker/pcmk_graph_producer.c
index 2e43325afd..40b7a7f774 100644
--- a/lib/pacemaker/pcmk_graph_producer.c
+++ b/lib/pacemaker/pcmk_graph_producer.c
@@ -1,1096 +1,1095 @@
/*
* Copyright 2004-2023 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 <pacemaker-internal.h>
#include "libpacemaker_private.h"
// Convenience macros for logging action properties
#define action_type_str(flags) \
(pcmk_is_set((flags), pcmk_action_pseudo)? "pseudo-action" : "action")
#define action_optional_str(flags) \
(pcmk_is_set((flags), pcmk_action_optional)? "optional" : "required")
#define action_runnable_str(flags) \
(pcmk_is_set((flags), pcmk_action_runnable)? "runnable" : "unrunnable")
#define action_node_str(a) \
(((a)->node == NULL)? "no node" : (a)->node->details->uname)
/*!
* \internal
* \brief Add an XML node tag for a specified ID
*
* \param[in] id Node UUID to add
* \param[in,out] xml Parent XML tag to add to
*/
static xmlNode*
add_node_to_xml_by_id(const char *id, xmlNode *xml)
{
xmlNode *node_xml;
node_xml = create_xml_node(xml, XML_CIB_TAG_NODE);
crm_xml_add(node_xml, XML_ATTR_ID, id);
return node_xml;
}
/*!
* \internal
* \brief Add an XML node tag for a specified node
*
* \param[in] node Node to add
* \param[in,out] xml XML to add node to
*/
static void
add_node_to_xml(const pe_node_t *node, void *xml)
{
add_node_to_xml_by_id(node->details->id, (xmlNode *) xml);
}
/*!
* \internal
* \brief Count (optionally add to XML) nodes needing maintenance state update
*
* \param[in,out] xml Parent XML tag to add to, if any
* \param[in] data_set Working set for cluster
*
* \return Count of nodes added
* \note Only Pacemaker Remote nodes are considered currently
*/
static int
add_maintenance_nodes(xmlNode *xml, const pe_working_set_t *data_set)
{
xmlNode *maintenance = NULL;
int count = 0;
if (xml != NULL) {
maintenance = create_xml_node(xml, XML_GRAPH_TAG_MAINTENANCE);
}
for (const GList *iter = data_set->nodes; iter != NULL; iter = iter->next) {
const pe_node_t *node = iter->data;
if (pe__is_guest_or_remote_node(node) &&
(node->details->maintenance != node->details->remote_maintenance)) {
if (maintenance != NULL) {
crm_xml_add(add_node_to_xml_by_id(node->details->id,
maintenance),
XML_NODE_IS_MAINTENANCE,
(node->details->maintenance? "1" : "0"));
}
count++;
}
}
crm_trace("%s %d nodes in need of maintenance mode update in state",
((maintenance == NULL)? "Counted" : "Added"), count);
return count;
}
/*!
* \internal
* \brief Add pseudo action with nodes needing maintenance state update
*
* \param[in,out] data_set Working set for cluster
*/
static void
add_maintenance_update(pe_working_set_t *data_set)
{
pe_action_t *action = NULL;
if (add_maintenance_nodes(NULL, data_set) != 0) {
action = get_pseudo_op(PCMK_ACTION_MAINTENANCE_NODES, data_set);
pe__set_action_flags(action, pcmk_action_always_in_graph);
}
}
/*!
* \internal
* \brief Add XML with nodes that an action is expected to bring down
*
* If a specified action is expected to bring any nodes down, add an XML block
* with their UUIDs. When a node is lost, this allows the controller to
* determine whether it was expected.
*
* \param[in,out] xml Parent XML tag to add to
* \param[in] action Action to check for downed nodes
*/
static void
add_downed_nodes(xmlNode *xml, const pe_action_t *action)
{
CRM_CHECK((xml != NULL) && (action != NULL) && (action->node != NULL),
return);
if (pcmk__str_eq(action->task, PCMK_ACTION_DO_SHUTDOWN, pcmk__str_none)) {
/* Shutdown makes the action's node down */
xmlNode *downed = create_xml_node(xml, XML_GRAPH_TAG_DOWNED);
add_node_to_xml_by_id(action->node->details->id, downed);
} else if (pcmk__str_eq(action->task, PCMK_ACTION_STONITH,
pcmk__str_none)) {
/* Fencing makes the action's node and any hosted guest nodes down */
const char *fence = g_hash_table_lookup(action->meta, "stonith_action");
if (pcmk__is_fencing_action(fence)) {
xmlNode *downed = create_xml_node(xml, XML_GRAPH_TAG_DOWNED);
add_node_to_xml_by_id(action->node->details->id, downed);
pe_foreach_guest_node(action->node->details->data_set, action->node,
add_node_to_xml, downed);
}
} else if (action->rsc && action->rsc->is_remote_node
&& pcmk__str_eq(action->task, PCMK_ACTION_STOP,
pcmk__str_none)) {
/* Stopping a remote connection resource makes connected node down,
* unless it's part of a migration
*/
GList *iter;
pe_action_t *input;
bool migrating = false;
for (iter = action->actions_before; iter != NULL; iter = iter->next) {
input = ((pe_action_wrapper_t *) iter->data)->action;
if ((input->rsc != NULL)
&& pcmk__str_eq(action->rsc->id, input->rsc->id, pcmk__str_none)
&& pcmk__str_eq(input->task, PCMK_ACTION_MIGRATE_FROM,
pcmk__str_none)) {
migrating = true;
break;
}
}
if (!migrating) {
xmlNode *downed = create_xml_node(xml, XML_GRAPH_TAG_DOWNED);
add_node_to_xml_by_id(action->rsc->id, downed);
}
}
}
/*!
* \internal
* \brief Create a transition graph operation key for a clone action
*
* \param[in] action Clone action
* \param[in] interval_ms Action interval in milliseconds
*
* \return Newly allocated string with transition graph operation key
*/
static char *
clone_op_key(const pe_action_t *action, guint interval_ms)
{
if (pcmk__str_eq(action->task, PCMK_ACTION_NOTIFY, pcmk__str_none)) {
const char *n_type = g_hash_table_lookup(action->meta, "notify_type");
const char *n_task = g_hash_table_lookup(action->meta,
"notify_operation");
CRM_LOG_ASSERT((n_type != NULL) && (n_task != NULL));
return pcmk__notify_key(action->rsc->clone_name, n_type, n_task);
} else if (action->cancel_task != NULL) {
return pcmk__op_key(action->rsc->clone_name, action->cancel_task,
interval_ms);
} else {
return pcmk__op_key(action->rsc->clone_name, action->task, interval_ms);
}
}
/*!
* \internal
* \brief Add node details to transition graph action XML
*
* \param[in] action Scheduled action
* \param[in,out] xml Transition graph action XML for \p action
*/
static void
add_node_details(const pe_action_t *action, xmlNode *xml)
{
pe_node_t *router_node = pcmk__connection_host_for_action(action);
crm_xml_add(xml, XML_LRM_ATTR_TARGET, action->node->details->uname);
crm_xml_add(xml, XML_LRM_ATTR_TARGET_UUID, action->node->details->id);
if (router_node != NULL) {
crm_xml_add(xml, XML_LRM_ATTR_ROUTER_NODE, router_node->details->uname);
}
}
/*!
* \internal
* \brief Add resource details to transition graph action XML
*
* \param[in] action Scheduled action
* \param[in,out] action_xml Transition graph action XML for \p action
*/
static void
add_resource_details(const pe_action_t *action, xmlNode *action_xml)
{
xmlNode *rsc_xml = NULL;
const char *attr_list[] = {
XML_AGENT_ATTR_CLASS,
XML_AGENT_ATTR_PROVIDER,
XML_ATTR_TYPE
};
/* If a resource is locked to a node via shutdown-lock, mark its actions
* so the controller can preserve the lock when the action completes.
*/
if (pcmk__action_locks_rsc_to_node(action)) {
crm_xml_add_ll(action_xml, XML_CONFIG_ATTR_SHUTDOWN_LOCK,
(long long) action->rsc->lock_time);
}
// List affected resource
rsc_xml = create_xml_node(action_xml,
(const char *) action->rsc->xml->name);
if (pcmk_is_set(action->rsc->flags, pcmk_rsc_removed)
&& (action->rsc->clone_name != NULL)) {
/* Use the numbered instance name here, because if there is more
* than one instance on a node, we need to make sure the command
* goes to the right one.
*
* This is important even for anonymous clones, because the clone's
* unique meta-attribute might have just been toggled from on to
* off.
*/
crm_debug("Using orphan clone name %s instead of %s",
action->rsc->id, action->rsc->clone_name);
crm_xml_add(rsc_xml, XML_ATTR_ID, action->rsc->clone_name);
crm_xml_add(rsc_xml, XML_ATTR_ID_LONG, action->rsc->id);
} else if (!pcmk_is_set(action->rsc->flags, pcmk_rsc_unique)) {
const char *xml_id = ID(action->rsc->xml);
crm_debug("Using anonymous clone name %s for %s (aka %s)",
xml_id, action->rsc->id, action->rsc->clone_name);
/* ID is what we'd like client to use
* ID_LONG is what they might know it as instead
*
* ID_LONG is only strictly needed /here/ during the
* transition period until all nodes in the cluster
* are running the new software /and/ have rebooted
* once (meaning that they've only ever spoken to a DC
* supporting this feature).
*
* If anyone toggles the unique flag to 'on', the
* 'instance free' name will correspond to an orphan
* and fall into the clause above instead
*/
crm_xml_add(rsc_xml, XML_ATTR_ID, xml_id);
if ((action->rsc->clone_name != NULL)
&& !pcmk__str_eq(xml_id, action->rsc->clone_name,
pcmk__str_none)) {
crm_xml_add(rsc_xml, XML_ATTR_ID_LONG, action->rsc->clone_name);
} else {
crm_xml_add(rsc_xml, XML_ATTR_ID_LONG, action->rsc->id);
}
} else {
CRM_ASSERT(action->rsc->clone_name == NULL);
crm_xml_add(rsc_xml, XML_ATTR_ID, action->rsc->id);
}
for (int lpc = 0; lpc < PCMK__NELEM(attr_list); lpc++) {
crm_xml_add(rsc_xml, attr_list[lpc],
g_hash_table_lookup(action->rsc->meta, attr_list[lpc]));
}
}
/*!
* \internal
* \brief Add action attributes to transition graph action XML
*
* \param[in,out] action Scheduled action
* \param[in,out] action_xml Transition graph action XML for \p action
*/
static void
add_action_attributes(pe_action_t *action, xmlNode *action_xml)
{
xmlNode *args_xml = NULL;
/* We create free-standing XML to start, so we can sort the attributes
* before adding it to action_xml, which keeps the scheduler regression
* test graphs comparable.
*/
args_xml = create_xml_node(NULL, XML_TAG_ATTRS);
crm_xml_add(args_xml, XML_ATTR_CRM_VERSION, CRM_FEATURE_SET);
g_hash_table_foreach(action->extra, hash2field, args_xml);
if ((action->rsc != NULL) && (action->node != NULL)) {
// Get the resource instance attributes, evaluated properly for node
GHashTable *params = pe_rsc_params(action->rsc, action->node,
action->rsc->cluster);
pcmk__substitute_remote_addr(action->rsc, params);
g_hash_table_foreach(params, hash2smartfield, args_xml);
} else if ((action->rsc != NULL)
&& (action->rsc->variant <= pcmk_rsc_variant_primitive)) {
GHashTable *params = pe_rsc_params(action->rsc, NULL,
action->rsc->cluster);
g_hash_table_foreach(params, hash2smartfield, args_xml);
}
g_hash_table_foreach(action->meta, hash2metafield, args_xml);
if (action->rsc != NULL) {
pe_resource_t *parent = action->rsc;
while (parent != NULL) {
parent->cmds->add_graph_meta(parent, args_xml);
parent = parent->parent;
}
pcmk__add_bundle_meta_to_xml(args_xml, action);
} else if (pcmk__str_eq(action->task, PCMK_ACTION_STONITH, pcmk__str_none)
&& (action->node != NULL)) {
/* Pass the node's attributes as meta-attributes.
*
* @TODO: Determine whether it is still necessary to do this. It was
* added in 33d99707, probably for the libfence-based implementation in
* c9a90bd, which is no longer used.
*/
g_hash_table_foreach(action->node->details->attrs, hash2metafield,
args_xml);
}
sorted_xml(args_xml, action_xml, FALSE);
free_xml(args_xml);
}
/*!
* \internal
* \brief Create the transition graph XML for a scheduled action
*
* \param[in,out] parent Parent XML element to add action to
* \param[in,out] action Scheduled action
* \param[in] skip_details If false, add action details as sub-elements
* \param[in] data_set Cluster working set
*/
static void
create_graph_action(xmlNode *parent, pe_action_t *action, bool skip_details,
const pe_working_set_t *data_set)
{
bool needs_node_info = true;
bool needs_maintenance_info = false;
xmlNode *action_xml = NULL;
if ((action == NULL) || (data_set == NULL)) {
return;
}
// Create the top-level element based on task
if (pcmk__str_eq(action->task, PCMK_ACTION_STONITH, pcmk__str_none)) {
/* All fences need node info; guest node fences are pseudo-events */
if (pcmk_is_set(action->flags, pcmk_action_pseudo)) {
action_xml = create_xml_node(parent, XML_GRAPH_TAG_PSEUDO_EVENT);
} else {
action_xml = create_xml_node(parent, XML_GRAPH_TAG_CRM_EVENT);
}
} else if (pcmk__str_any_of(action->task,
PCMK_ACTION_DO_SHUTDOWN,
PCMK_ACTION_CLEAR_FAILCOUNT, NULL)) {
action_xml = create_xml_node(parent, XML_GRAPH_TAG_CRM_EVENT);
} else if (pcmk__str_eq(action->task, PCMK_ACTION_LRM_DELETE,
pcmk__str_none)) {
// CIB-only clean-up for shutdown locks
action_xml = create_xml_node(parent, XML_GRAPH_TAG_CRM_EVENT);
crm_xml_add(action_xml, PCMK__XA_MODE, XML_TAG_CIB);
} else if (pcmk_is_set(action->flags, pcmk_action_pseudo)) {
if (pcmk__str_eq(action->task, PCMK_ACTION_MAINTENANCE_NODES,
pcmk__str_none)) {
needs_maintenance_info = true;
}
action_xml = create_xml_node(parent, XML_GRAPH_TAG_PSEUDO_EVENT);
needs_node_info = false;
} else {
action_xml = create_xml_node(parent, XML_GRAPH_TAG_RSC_OP);
}
crm_xml_add_int(action_xml, XML_ATTR_ID, action->id);
crm_xml_add(action_xml, XML_LRM_ATTR_TASK, action->task);
if ((action->rsc != NULL) && (action->rsc->clone_name != NULL)) {
char *clone_key = NULL;
guint interval_ms;
if (pcmk__guint_from_hash(action->meta, XML_LRM_ATTR_INTERVAL_MS, 0,
&interval_ms) != pcmk_rc_ok) {
interval_ms = 0;
}
clone_key = clone_op_key(action, interval_ms);
crm_xml_add(action_xml, XML_LRM_ATTR_TASK_KEY, clone_key);
crm_xml_add(action_xml, "internal_" XML_LRM_ATTR_TASK_KEY,
action->uuid);
free(clone_key);
} else {
crm_xml_add(action_xml, XML_LRM_ATTR_TASK_KEY, action->uuid);
}
if (needs_node_info && (action->node != NULL)) {
add_node_details(action, action_xml);
g_hash_table_insert(action->meta, strdup(XML_LRM_ATTR_TARGET),
strdup(action->node->details->uname));
g_hash_table_insert(action->meta, strdup(XML_LRM_ATTR_TARGET_UUID),
strdup(action->node->details->id));
}
if (skip_details) {
return;
}
if ((action->rsc != NULL)
&& !pcmk_is_set(action->flags, pcmk_action_pseudo)) {
// This is a real resource action, so add resource details
add_resource_details(action, action_xml);
}
/* List any attributes in effect */
add_action_attributes(action, action_xml);
/* List any nodes this action is expected to make down */
if (needs_node_info && (action->node != NULL)) {
add_downed_nodes(action_xml, action);
}
if (needs_maintenance_info) {
add_maintenance_nodes(action_xml, data_set);
}
}
/*!
* \internal
* \brief Check whether an action should be added to the transition graph
*
* \param[in] action Action to check
*
* \return true if action should be added to graph, otherwise false
*/
static bool
should_add_action_to_graph(const pe_action_t *action)
{
if (!pcmk_is_set(action->flags, pcmk_action_runnable)) {
crm_trace("Ignoring action %s (%d): unrunnable",
action->uuid, action->id);
return false;
}
if (pcmk_is_set(action->flags, pcmk_action_optional)
&& !pcmk_is_set(action->flags, pcmk_action_always_in_graph)) {
crm_trace("Ignoring action %s (%d): optional",
action->uuid, action->id);
return false;
}
/* Actions for unmanaged resources should be excluded from the graph,
* with the exception of monitors and cancellation of recurring monitors.
*/
if ((action->rsc != NULL)
&& !pcmk_is_set(action->rsc->flags, pcmk_rsc_managed)
&& !pcmk__str_eq(action->task, PCMK_ACTION_MONITOR, pcmk__str_none)) {
const char *interval_ms_s;
/* A cancellation of a recurring monitor will get here because the task
* is cancel rather than monitor, but the interval can still be used to
* recognize it. The interval has been normalized to milliseconds by
* this point, so a string comparison is sufficient.
*/
interval_ms_s = g_hash_table_lookup(action->meta,
XML_LRM_ATTR_INTERVAL_MS);
if (pcmk__str_eq(interval_ms_s, "0", pcmk__str_null_matches)) {
crm_trace("Ignoring action %s (%d): for unmanaged resource (%s)",
action->uuid, action->id, action->rsc->id);
return false;
}
}
/* Always add pseudo-actions, fence actions, and shutdown actions (already
* determined to be required and runnable by this point)
*/
if (pcmk_is_set(action->flags, pcmk_action_pseudo)
|| pcmk__strcase_any_of(action->task, PCMK_ACTION_STONITH,
PCMK_ACTION_DO_SHUTDOWN, NULL)) {
return true;
}
if (action->node == NULL) {
pe_err("Skipping action %s (%d) "
"because it was not assigned to a node (bug?)",
action->uuid, action->id);
pcmk__log_action("Unassigned", action, false);
return false;
}
if (pcmk_is_set(action->flags, pcmk_action_on_dc)) {
crm_trace("Action %s (%d) should be dumped: "
"can run on DC instead of %s",
action->uuid, action->id, pe__node_name(action->node));
} else if (pe__is_guest_node(action->node)
&& !action->node->details->remote_requires_reset) {
crm_trace("Action %s (%d) should be dumped: "
"assuming will be runnable on guest %s",
action->uuid, action->id, pe__node_name(action->node));
} else if (!action->node->details->online) {
pe_err("Skipping action %s (%d) "
"because it was scheduled for offline node (bug?)",
action->uuid, action->id);
pcmk__log_action("Offline node", action, false);
return false;
} else if (action->node->details->unclean) {
pe_err("Skipping action %s (%d) "
"because it was scheduled for unclean node (bug?)",
action->uuid, action->id);
pcmk__log_action("Unclean node", action, false);
return false;
}
return true;
}
/*!
* \internal
* \brief Check whether an ordering's flags can change an action
*
* \param[in] ordering Ordering to check
*
* \return true if ordering has flags that can change an action, false otherwise
*/
static bool
ordering_can_change_actions(const pe_action_wrapper_t *ordering)
{
return pcmk_any_flags_set(ordering->type,
~(pcmk__ar_then_implies_first_graphed
|pcmk__ar_first_implies_then_graphed
|pcmk__ar_ordered));
}
/*!
* \internal
* \brief Check whether an action input should be in the transition graph
*
* \param[in] action Action to check
* \param[in,out] input Action input to check
*
* \return true if input should be in graph, false otherwise
* \note This function may not only check an input, but disable it under certian
* circumstances (load or anti-colocation orderings that are not needed).
*/
static bool
should_add_input_to_graph(const pe_action_t *action, pe_action_wrapper_t *input)
{
if (input->state == pe_link_dumped) {
return true;
}
if ((uint32_t) input->type == pcmk__ar_none) {
crm_trace("Ignoring %s (%d) input %s (%d): "
"ordering disabled",
action->uuid, action->id,
input->action->uuid, input->action->id);
return false;
} else if (!pcmk_is_set(input->action->flags, pcmk_action_runnable)
&& !ordering_can_change_actions(input)) {
crm_trace("Ignoring %s (%d) input %s (%d): "
"optional and input unrunnable",
action->uuid, action->id,
input->action->uuid, input->action->id);
return false;
} else if (!pcmk_is_set(input->action->flags, pcmk_action_runnable)
&& pcmk_is_set(input->type, pe_order_one_or_more)) {
crm_trace("Ignoring %s (%d) input %s (%d): "
"one-or-more and input unrunnable",
action->uuid, action->id,
input->action->uuid, input->action->id);
return false;
} else if (pcmk_is_set(input->type, pcmk__ar_unmigratable_then_blocks)
&& !pcmk_is_set(input->action->flags, pcmk_action_runnable)) {
crm_trace("Ignoring %s (%d) input %s (%d): "
"input blocked if 'then' unmigratable",
action->uuid, action->id,
input->action->uuid, input->action->id);
return false;
} else if (pcmk_is_set(input->type, pcmk__ar_if_first_unmigratable)
&& pcmk_is_set(input->action->flags, pcmk_action_migratable)) {
crm_trace("Ignoring %s (%d) input %s (%d): ordering applies "
"only if input is unmigratable, but it is migratable",
action->uuid, action->id,
input->action->uuid, input->action->id);
return false;
} else if (((uint32_t) input->type == pcmk__ar_ordered)
&& pcmk_is_set(input->action->flags, pcmk_action_migratable)
&& pcmk__ends_with(input->action->uuid, "_stop_0")) {
crm_trace("Ignoring %s (%d) input %s (%d): "
"optional but stop in migration",
action->uuid, action->id,
input->action->uuid, input->action->id);
return false;
- } else if (input->type == pe_order_load) {
+ } else if ((uint32_t) input->type == pcmk__ar_if_on_same_node_or_target) {
pe_node_t *input_node = input->action->node;
- // load orderings are relevant only if actions are for same node
-
if ((action->rsc != NULL)
&& pcmk__str_eq(action->task, PCMK_ACTION_MIGRATE_TO,
pcmk__str_none)) {
pe_node_t *assigned = action->rsc->allocated_to;
- /* For load_stopped -> migrate_to orderings, we care about where it
- * has been assigned to, not where it will be executed.
+ /* For load_stopped -> migrate_to orderings, we care about where
+ * the resource has been assigned, not where migrate_to will be
+ * executed.
*/
if (!pe__same_node(input_node, assigned)) {
crm_trace("Ignoring %s (%d) input %s (%d): "
- "load ordering node mismatch %s vs %s",
+ "migration target %s is not same as input node %s",
action->uuid, action->id,
input->action->uuid, input->action->id,
(assigned? assigned->details->uname : "<none>"),
(input_node? input_node->details->uname : "<none>"));
input->type = pcmk__ar_none;
return false;
}
} else if (!pe__same_node(input_node, action->node)) {
crm_trace("Ignoring %s (%d) input %s (%d): "
- "load ordering node mismatch %s vs %s",
+ "not on same node (%s vs %s)",
action->uuid, action->id,
input->action->uuid, input->action->id,
(action->node? action->node->details->uname : "<none>"),
(input_node? input_node->details->uname : "<none>"));
input->type = pcmk__ar_none;
return false;
} else if (pcmk_is_set(input->action->flags, pcmk_action_optional)) {
crm_trace("Ignoring %s (%d) input %s (%d): "
- "load ordering input optional",
+ "ordering optional",
action->uuid, action->id,
input->action->uuid, input->action->id);
input->type = pcmk__ar_none;
return false;
}
} else if (input->type == pe_order_anti_colocation) {
if (input->action->node && action->node
&& !pe__same_node(input->action->node, action->node)) {
crm_trace("Ignoring %s (%d) input %s (%d): "
"anti-colocation node mismatch %s vs %s",
action->uuid, action->id,
input->action->uuid, input->action->id,
pe__node_name(action->node),
pe__node_name(input->action->node));
input->type = pcmk__ar_none;
return false;
} else if (pcmk_is_set(input->action->flags, pcmk_action_optional)) {
crm_trace("Ignoring %s (%d) input %s (%d): "
"anti-colocation input optional",
action->uuid, action->id,
input->action->uuid, input->action->id);
input->type = pcmk__ar_none;
return false;
}
} else if (input->action->rsc
&& input->action->rsc != action->rsc
&& pcmk_is_set(input->action->rsc->flags, pcmk_rsc_failed)
&& !pcmk_is_set(input->action->rsc->flags, pcmk_rsc_managed)
&& pcmk__ends_with(input->action->uuid, "_stop_0")
&& action->rsc && pe_rsc_is_clone(action->rsc)) {
crm_warn("Ignoring requirement that %s complete before %s:"
" unmanaged failed resources cannot prevent clone shutdown",
input->action->uuid, action->uuid);
return false;
} else if (pcmk_is_set(input->action->flags, pcmk_action_optional)
&& !pcmk_any_flags_set(input->action->flags,
pcmk_action_always_in_graph
|pcmk_action_added_to_graph)
&& !should_add_action_to_graph(input->action)) {
crm_trace("Ignoring %s (%d) input %s (%d): "
"input optional",
action->uuid, action->id,
input->action->uuid, input->action->id);
return false;
}
crm_trace("%s (%d) input %s %s (%d) on %s should be dumped: %s %s %#.6x",
action->uuid, action->id, action_type_str(input->action->flags),
input->action->uuid, input->action->id,
action_node_str(input->action),
action_runnable_str(input->action->flags),
action_optional_str(input->action->flags), input->type);
return true;
}
/*!
* \internal
* \brief Check whether an ordering creates an ordering loop
*
* \param[in] init_action "First" action in ordering
* \param[in] action Callers should always set this the same as
* \p init_action (this function may use a different
* value for recursive calls)
* \param[in,out] input Action wrapper for "then" action in ordering
*
* \return true if the ordering creates a loop, otherwise false
*/
bool
pcmk__graph_has_loop(const pe_action_t *init_action, const pe_action_t *action,
pe_action_wrapper_t *input)
{
bool has_loop = false;
if (pcmk_is_set(input->action->flags, pcmk_action_detect_loop)) {
crm_trace("Breaking tracking loop: %s@%s -> %s@%s (%#.6x)",
input->action->uuid,
input->action->node? input->action->node->details->uname : "",
action->uuid,
action->node? action->node->details->uname : "",
input->type);
return false;
}
// Don't need to check inputs that won't be used
if (!should_add_input_to_graph(action, input)) {
return false;
}
if (input->action == init_action) {
crm_debug("Input loop found in %s@%s ->...-> %s@%s",
action->uuid,
action->node? action->node->details->uname : "",
init_action->uuid,
init_action->node? init_action->node->details->uname : "");
return true;
}
pe__set_action_flags(input->action, pcmk_action_detect_loop);
crm_trace("Checking inputs of action %s@%s input %s@%s (%#.6x)"
"for graph loop with %s@%s ",
action->uuid,
action->node? action->node->details->uname : "",
input->action->uuid,
input->action->node? input->action->node->details->uname : "",
input->type,
init_action->uuid,
init_action->node? init_action->node->details->uname : "");
// Recursively check input itself for loops
for (GList *iter = input->action->actions_before;
iter != NULL; iter = iter->next) {
if (pcmk__graph_has_loop(init_action, input->action,
(pe_action_wrapper_t *) iter->data)) {
// Recursive call already logged a debug message
has_loop = true;
break;
}
}
pe__clear_action_flags(input->action, pcmk_action_detect_loop);
if (!has_loop) {
crm_trace("No input loop found in %s@%s -> %s@%s (%#.6x)",
input->action->uuid,
input->action->node? input->action->node->details->uname : "",
action->uuid,
action->node? action->node->details->uname : "",
input->type);
}
return has_loop;
}
/*!
* \internal
* \brief Create a synapse XML element for a transition graph
*
* \param[in] action Action that synapse is for
* \param[in,out] data_set Cluster working set containing graph
*
* \return Newly added XML element for new graph synapse
*/
static xmlNode *
create_graph_synapse(const pe_action_t *action, pe_working_set_t *data_set)
{
int synapse_priority = 0;
xmlNode *syn = create_xml_node(data_set->graph, "synapse");
crm_xml_add_int(syn, XML_ATTR_ID, data_set->num_synapse);
data_set->num_synapse++;
if (action->rsc != NULL) {
synapse_priority = action->rsc->priority;
}
if (action->priority > synapse_priority) {
synapse_priority = action->priority;
}
if (synapse_priority > 0) {
crm_xml_add_int(syn, XML_CIB_ATTR_PRIORITY, synapse_priority);
}
return syn;
}
/*!
* \internal
* \brief Add an action to the transition graph XML if appropriate
*
* \param[in,out] data Action to possibly add
* \param[in,out] user_data Cluster working set
*
* \note This will de-duplicate the action inputs, meaning that the
* pe_action_wrapper_t:type flags can no longer be relied on to retain
* their original settings. That means this MUST be called after
* pcmk__apply_orderings() is complete, and nothing after this should rely
* on those type flags. (For example, some code looks for type equal to
* some flag rather than whether the flag is set, and some code looks for
* particular combinations of flags -- such code must be done before
* pcmk__create_graph().)
*/
static void
add_action_to_graph(gpointer data, gpointer user_data)
{
pe_action_t *action = (pe_action_t *) data;
pe_working_set_t *data_set = (pe_working_set_t *) user_data;
xmlNode *syn = NULL;
xmlNode *set = NULL;
xmlNode *in = NULL;
/* If we haven't already, de-duplicate inputs (even if we won't be adding
* the action to the graph, so that crm_simulate's dot graphs don't have
* duplicates).
*/
if (!pcmk_is_set(action->flags, pcmk_action_inputs_deduplicated)) {
pcmk__deduplicate_action_inputs(action);
pe__set_action_flags(action, pcmk_action_inputs_deduplicated);
}
if (pcmk_is_set(action->flags, pcmk_action_added_to_graph)
|| !should_add_action_to_graph(action)) {
return; // Already added, or shouldn't be
}
pe__set_action_flags(action, pcmk_action_added_to_graph);
crm_trace("Adding action %d (%s%s%s) to graph",
action->id, action->uuid,
((action->node == NULL)? "" : " on "),
((action->node == NULL)? "" : action->node->details->uname));
syn = create_graph_synapse(action, data_set);
set = create_xml_node(syn, "action_set");
in = create_xml_node(syn, "inputs");
create_graph_action(set, action, false, data_set);
for (GList *lpc = action->actions_before; lpc != NULL; lpc = lpc->next) {
pe_action_wrapper_t *input = (pe_action_wrapper_t *) lpc->data;
if (should_add_input_to_graph(action, input)) {
xmlNode *input_xml = create_xml_node(in, "trigger");
input->state = pe_link_dumped;
create_graph_action(input_xml, input->action, true, data_set);
}
}
}
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)%s%s",
transition_id,
(filename == NULL)? "" : ", saving inputs in ",
(filename == NULL)? "" : filename);
} else if (was_processing_warning) {
crm_warn("Calculated transition %d (with warnings)%s%s",
transition_id,
(filename == NULL)? "" : ", saving inputs in ",
(filename == NULL)? "" : filename);
} else {
crm_notice("Calculated transition %d%s%s",
transition_id,
(filename == NULL)? "" : ", saving inputs in ",
(filename == NULL)? "" : filename);
}
if (crm_config_error) {
crm_notice("Configuration errors found during scheduler processing,"
" please run \"crm_verify -L\" to identify issues");
}
}
/*!
* \internal
* \brief Add a resource's actions to the transition graph
*
* \param[in,out] rsc Resource whose actions should be added
*/
void
pcmk__add_rsc_actions_to_graph(pe_resource_t *rsc)
{
GList *iter = NULL;
CRM_ASSERT(rsc != NULL);
pe_rsc_trace(rsc, "Adding actions for %s to graph", rsc->id);
// First add the resource's own actions
g_list_foreach(rsc->actions, add_action_to_graph, rsc->cluster);
// Then recursively add its children's actions (appropriate to variant)
for (iter = rsc->children; iter != NULL; iter = iter->next) {
pe_resource_t *child_rsc = (pe_resource_t *) iter->data;
child_rsc->cmds->add_actions_to_graph(child_rsc);
}
}
/*!
* \internal
* \brief Create a transition graph with all cluster actions needed
*
* \param[in,out] data_set Cluster working set
*/
void
pcmk__create_graph(pe_working_set_t *data_set)
{
GList *iter = NULL;
const char *value = NULL;
long long limit = 0LL;
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 (pcmk_is_set(data_set->flags, pcmk_sched_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 ((pcmk__scan_ll(value, &limit, 0LL) == pcmk_rc_ok) && (limit > 0)) {
crm_xml_add(data_set->graph, "migration-limit", value);
}
if (data_set->recheck_by > 0) {
char *recheck_epoch = NULL;
recheck_epoch = crm_strdup_printf("%llu",
(long long) data_set->recheck_by);
crm_xml_add(data_set->graph, "recheck-by", recheck_epoch);
free(recheck_epoch);
}
/* The following code will de-duplicate action inputs, so nothing past this
* should rely on the action input type flags retaining their original
* values.
*/
// Add resource actions to graph
for (iter = data_set->resources; iter != NULL; iter = iter->next) {
pe_resource_t *rsc = (pe_resource_t *) iter->data;
pe_rsc_trace(rsc, "Processing actions for %s", rsc->id);
rsc->cmds->add_actions_to_graph(rsc);
}
// Add pseudo-action for list of nodes with maintenance state update
add_maintenance_update(data_set);
// Add non-resource (node) actions
for (iter = data_set->actions; iter != NULL; iter = iter->next) {
pe_action_t *action = (pe_action_t *) iter->data;
if ((action->rsc != NULL)
&& (action->node != NULL)
&& action->node->details->shutdown
&& !pcmk_is_set(action->rsc->flags, pcmk_rsc_maintenance)
&& !pcmk_any_flags_set(action->flags,
pcmk_action_optional|pcmk_action_runnable)
&& pcmk__str_eq(action->task, PCMK_ACTION_STOP, pcmk__str_none)) {
/* 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 (pcmk_is_set(data_set->flags, pcmk_sched_quorate)
|| (data_set->no_quorum_policy == pcmk_no_quorum_ignore)) {
const bool managed = pcmk_is_set(action->rsc->flags,
pcmk_rsc_managed);
const bool failed = pcmk_is_set(action->rsc->flags,
pcmk_rsc_failed);
crm_crit("Cannot %s %s because of %s:%s%s (%s)",
action->node->details->unclean? "fence" : "shut down",
pe__node_name(action->node), action->rsc->id,
(managed? " blocked" : " unmanaged"),
(failed? " failed" : ""), action->uuid);
}
}
add_action_to_graph((gpointer) action, (gpointer) data_set);
}
crm_log_xml_trace(data_set->graph, "graph");
}
diff --git a/lib/pacemaker/pcmk_sched_ordering.c b/lib/pacemaker/pcmk_sched_ordering.c
index 75e03b91a5..2b37eaa298 100644
--- a/lib/pacemaker/pcmk_sched_ordering.c
+++ b/lib/pacemaker/pcmk_sched_ordering.c
@@ -1,1498 +1,1499 @@
/*
* Copyright 2004-2023 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 <inttypes.h> // PRIx32
#include <stdbool.h>
#include <glib.h>
#include <crm/crm.h>
#include <pacemaker-internal.h>
#include "libpacemaker_private.h"
enum pe_order_kind {
pe_order_kind_optional,
pe_order_kind_mandatory,
pe_order_kind_serialize,
};
enum ordering_symmetry {
ordering_asymmetric, // the only relation in an asymmetric ordering
ordering_symmetric, // the normal relation in a symmetric ordering
ordering_symmetric_inverse, // the inverse relation in a symmetric ordering
};
#define EXPAND_CONSTRAINT_IDREF(__set, __rsc, __name) do { \
__rsc = pcmk__find_constraint_resource(data_set->resources, \
__name); \
if (__rsc == NULL) { \
pcmk__config_err("%s: No resource found for %s", __set, __name);\
return pcmk_rc_unpack_error; \
} \
} while (0)
static const char *
invert_action(const char *action)
{
if (pcmk__str_eq(action, PCMK_ACTION_START, pcmk__str_none)) {
return PCMK_ACTION_STOP;
} else if (pcmk__str_eq(action, PCMK_ACTION_STOP, pcmk__str_none)) {
return PCMK_ACTION_START;
} else if (pcmk__str_eq(action, PCMK_ACTION_PROMOTE, pcmk__str_none)) {
return PCMK_ACTION_DEMOTE;
} else if (pcmk__str_eq(action, PCMK_ACTION_DEMOTE, pcmk__str_none)) {
return PCMK_ACTION_PROMOTE;
} else if (pcmk__str_eq(action, PCMK_ACTION_PROMOTED, pcmk__str_none)) {
return PCMK_ACTION_DEMOTED;
} else if (pcmk__str_eq(action, PCMK_ACTION_DEMOTED, pcmk__str_none)) {
return PCMK_ACTION_PROMOTED;
} else if (pcmk__str_eq(action, PCMK_ACTION_RUNNING, pcmk__str_none)) {
return PCMK_ACTION_STOPPED;
} else if (pcmk__str_eq(action, PCMK_ACTION_STOPPED, pcmk__str_none)) {
return PCMK_ACTION_RUNNING;
}
crm_warn("Unknown action '%s' specified in order constraint", action);
return NULL;
}
static enum pe_order_kind
get_ordering_type(const xmlNode *xml_obj)
{
enum pe_order_kind kind_e = pe_order_kind_mandatory;
const char *kind = crm_element_value(xml_obj, XML_ORDER_ATTR_KIND);
if (kind == NULL) {
const char *score = crm_element_value(xml_obj, XML_RULE_ATTR_SCORE);
kind_e = pe_order_kind_mandatory;
if (score) {
// @COMPAT deprecated informally since 1.0.7, formally since 2.0.1
int score_i = char2score(score);
if (score_i == 0) {
kind_e = pe_order_kind_optional;
}
pe_warn_once(pe_wo_order_score,
"Support for 'score' in rsc_order is deprecated "
"and will be removed in a future release "
"(use 'kind' instead)");
}
} else if (pcmk__str_eq(kind, "Mandatory", pcmk__str_none)) {
kind_e = pe_order_kind_mandatory;
} else if (pcmk__str_eq(kind, "Optional", pcmk__str_none)) {
kind_e = pe_order_kind_optional;
} else if (pcmk__str_eq(kind, "Serialize", pcmk__str_none)) {
kind_e = pe_order_kind_serialize;
} else {
pcmk__config_err("Resetting '" XML_ORDER_ATTR_KIND "' for constraint "
"%s to 'Mandatory' because '%s' is not valid",
pcmk__s(ID(xml_obj), "missing ID"), kind);
}
return kind_e;
}
/*!
* \internal
* \brief Get ordering symmetry from XML
*
* \param[in] xml_obj Ordering XML
* \param[in] parent_kind Default ordering kind
* \param[in] parent_symmetrical_s Parent element's symmetrical setting, if any
*
* \retval ordering_symmetric Ordering is symmetric
* \retval ordering_asymmetric Ordering is asymmetric
*/
static enum ordering_symmetry
get_ordering_symmetry(const xmlNode *xml_obj, enum pe_order_kind parent_kind,
const char *parent_symmetrical_s)
{
int rc = pcmk_rc_ok;
bool symmetric = false;
enum pe_order_kind kind = parent_kind; // Default to parent's kind
// Check ordering XML for explicit kind
if ((crm_element_value(xml_obj, XML_ORDER_ATTR_KIND) != NULL)
|| (crm_element_value(xml_obj, XML_RULE_ATTR_SCORE) != NULL)) {
kind = get_ordering_type(xml_obj);
}
// Check ordering XML (and parent) for explicit symmetrical setting
rc = pcmk__xe_get_bool_attr(xml_obj, XML_CONS_ATTR_SYMMETRICAL, &symmetric);
if (rc != pcmk_rc_ok && parent_symmetrical_s != NULL) {
symmetric = crm_is_true(parent_symmetrical_s);
rc = pcmk_rc_ok;
}
if (rc == pcmk_rc_ok) {
if (symmetric) {
if (kind == pe_order_kind_serialize) {
pcmk__config_warn("Ignoring " XML_CONS_ATTR_SYMMETRICAL
" for '%s' because not valid with "
XML_ORDER_ATTR_KIND " of 'Serialize'",
ID(xml_obj));
} else {
return ordering_symmetric;
}
}
return ordering_asymmetric;
}
// Use default symmetry
if (kind == pe_order_kind_serialize) {
return ordering_asymmetric;
}
return ordering_symmetric;
}
/*!
* \internal
* \brief Get ordering flags appropriate to ordering kind
*
* \param[in] kind Ordering kind
* \param[in] first Action name for 'first' action
* \param[in] symmetry This ordering's symmetry role
*
* \return Minimal ordering flags appropriate to \p kind
*/
static uint32_t
ordering_flags_for_kind(enum pe_order_kind kind, const char *first,
enum ordering_symmetry symmetry)
{
uint32_t flags = pcmk__ar_none; // so we trace-log all flags set
switch (kind) {
case pe_order_kind_optional:
pe__set_order_flags(flags, pcmk__ar_ordered);
break;
case pe_order_kind_serialize:
/* This flag is not used anywhere directly but means the relation
* will not match an equality comparison against pcmk__ar_none or
* pcmk__ar_ordered.
*/
pe__set_order_flags(flags, pcmk__ar_serialize);
break;
case pe_order_kind_mandatory:
pe__set_order_flags(flags, pcmk__ar_ordered);
switch (symmetry) {
case ordering_asymmetric:
pe__set_order_flags(flags, pcmk__ar_asymmetric);
break;
case ordering_symmetric:
pe__set_order_flags(flags, pcmk__ar_first_implies_then);
if (pcmk__strcase_any_of(first, PCMK_ACTION_START,
PCMK_ACTION_PROMOTE, NULL)) {
pe__set_order_flags(flags,
pcmk__ar_unrunnable_first_blocks);
}
break;
case ordering_symmetric_inverse:
pe__set_order_flags(flags, pcmk__ar_then_implies_first);
break;
}
break;
}
return flags;
}
/*!
* \internal
* \brief Find resource corresponding to ID specified in ordering
*
* \param[in] xml Ordering XML
* \param[in] resource_attr XML attribute name for resource ID
* \param[in] instance_attr XML attribute name for instance number.
* This option is deprecated and will be removed in a
* future release.
* \param[in] data_set Cluster working set
*
* \return Resource corresponding to \p id, or NULL if none
*/
static pe_resource_t *
get_ordering_resource(const xmlNode *xml, const char *resource_attr,
const char *instance_attr,
const pe_working_set_t *data_set)
{
// @COMPAT: instance_attr and instance_id variables deprecated since 2.1.5
pe_resource_t *rsc = NULL;
const char *rsc_id = crm_element_value(xml, resource_attr);
const char *instance_id = crm_element_value(xml, instance_attr);
if (rsc_id == NULL) {
pcmk__config_err("Ignoring constraint '%s' without %s",
ID(xml), resource_attr);
return NULL;
}
rsc = pcmk__find_constraint_resource(data_set->resources, rsc_id);
if (rsc == NULL) {
pcmk__config_err("Ignoring constraint '%s' because resource '%s' "
"does not exist", ID(xml), rsc_id);
return NULL;
}
if (instance_id != NULL) {
pe_warn_once(pe_wo_order_inst,
"Support for " XML_ORDER_ATTR_FIRST_INSTANCE " and "
XML_ORDER_ATTR_THEN_INSTANCE " is deprecated and will be "
"removed in a future release.");
if (!pe_rsc_is_clone(rsc)) {
pcmk__config_err("Ignoring constraint '%s' because resource '%s' "
"is not a clone but instance '%s' was requested",
ID(xml), rsc_id, instance_id);
return NULL;
}
rsc = find_clone_instance(rsc, instance_id);
if (rsc == NULL) {
pcmk__config_err("Ignoring constraint '%s' because resource '%s' "
"does not have an instance '%s'",
"'%s'", ID(xml), rsc_id, instance_id);
return NULL;
}
}
return rsc;
}
/*!
* \internal
* \brief Determine minimum number of 'first' instances required in ordering
*
* \param[in] rsc 'First' resource in ordering
* \param[in] xml Ordering XML
*
* \return Minimum 'first' instances required (or 0 if not applicable)
*/
static int
get_minimum_first_instances(const pe_resource_t *rsc, const xmlNode *xml)
{
const char *clone_min = NULL;
bool require_all = false;
if (!pe_rsc_is_clone(rsc)) {
return 0;
}
clone_min = g_hash_table_lookup(rsc->meta, PCMK_META_CLONE_MIN);
if (clone_min != NULL) {
int clone_min_int = 0;
pcmk__scan_min_int(clone_min, &clone_min_int, 0);
return clone_min_int;
}
/* @COMPAT 1.1.13:
* require-all=false is deprecated equivalent of clone-min=1
*/
if (pcmk__xe_get_bool_attr(xml, "require-all", &require_all) != ENODATA) {
pe_warn_once(pe_wo_require_all,
"Support for require-all in ordering constraints "
"is deprecated and will be removed in a future release"
" (use clone-min clone meta-attribute instead)");
if (!require_all) {
return 1;
}
}
return 0;
}
/*!
* \internal
* \brief Create orderings for a constraint with clone-min > 0
*
* \param[in] id Ordering ID
* \param[in,out] rsc_first 'First' resource in ordering (a clone)
* \param[in] action_first 'First' action in ordering
* \param[in] rsc_then 'Then' resource in ordering
* \param[in] action_then 'Then' action in ordering
* \param[in] flags Ordering flags
* \param[in] clone_min Minimum required instances of 'first'
*/
static void
clone_min_ordering(const char *id,
pe_resource_t *rsc_first, const char *action_first,
pe_resource_t *rsc_then, const char *action_then,
uint32_t flags, int clone_min)
{
// Create a pseudo-action for when the minimum instances are active
char *task = crm_strdup_printf(PCMK_ACTION_CLONE_ONE_OR_MORE ":%s", id);
pe_action_t *clone_min_met = get_pseudo_op(task, rsc_first->cluster);
free(task);
/* Require the pseudo-action to have the required number of actions to be
* considered runnable before allowing the pseudo-action to be runnable.
*/
clone_min_met->required_runnable_before = clone_min;
pe__set_action_flags(clone_min_met, pcmk_action_min_runnable);
// Order the actions for each clone instance before the pseudo-action
for (GList *iter = rsc_first->children; iter != NULL; iter = iter->next) {
pe_resource_t *child = iter->data;
pcmk__new_ordering(child, pcmk__op_key(child->id, action_first, 0),
NULL, NULL, NULL, clone_min_met,
pe_order_one_or_more
|pcmk__ar_first_implies_then_graphed,
rsc_first->cluster);
}
// Order "then" action after the pseudo-action (if runnable)
pcmk__new_ordering(NULL, NULL, clone_min_met, rsc_then,
pcmk__op_key(rsc_then->id, action_then, 0),
NULL, flags|pcmk__ar_unrunnable_first_blocks,
rsc_first->cluster);
}
/*!
* \internal
* \brief Update ordering flags for restart-type=restart
*
* \param[in] rsc 'Then' resource in ordering
* \param[in] kind Ordering kind
* \param[in] flag Ordering flag to set (when applicable)
* \param[in,out] flags Ordering flag set to update
*
* \compat The restart-type resource meta-attribute is deprecated. Eventually,
* it will be removed, and pe_restart_ignore will be the only behavior,
* at which time this can just be removed entirely.
*/
#define handle_restart_type(rsc, kind, flag, flags) do { \
if (((kind) == pe_order_kind_optional) \
&& ((rsc)->restart_type == pe_restart_restart)) { \
pe__set_order_flags((flags), (flag)); \
} \
} while (0)
/*!
* \internal
* \brief Create new ordering for inverse of symmetric constraint
*
* \param[in] id Ordering ID (for logging only)
* \param[in] kind Ordering kind
* \param[in] rsc_first 'First' resource in ordering (a clone)
* \param[in] action_first 'First' action in ordering
* \param[in,out] rsc_then 'Then' resource in ordering
* \param[in] action_then 'Then' action in ordering
*/
static void
inverse_ordering(const char *id, enum pe_order_kind kind,
pe_resource_t *rsc_first, const char *action_first,
pe_resource_t *rsc_then, const char *action_then)
{
action_then = invert_action(action_then);
action_first = invert_action(action_first);
if ((action_then == NULL) || (action_first == NULL)) {
pcmk__config_warn("Cannot invert constraint '%s' "
"(please specify inverse manually)", id);
} else {
uint32_t flags = ordering_flags_for_kind(kind, action_first,
ordering_symmetric_inverse);
handle_restart_type(rsc_then, kind, pcmk__ar_then_implies_first, flags);
pcmk__order_resource_actions(rsc_then, action_then, rsc_first,
action_first, flags);
}
}
static void
unpack_simple_rsc_order(xmlNode *xml_obj, pe_working_set_t *data_set)
{
pe_resource_t *rsc_then = NULL;
pe_resource_t *rsc_first = NULL;
int min_required_before = 0;
enum pe_order_kind kind = pe_order_kind_mandatory;
uint32_t flags = pcmk__ar_none;
enum ordering_symmetry symmetry;
const char *action_then = NULL;
const char *action_first = NULL;
const char *id = NULL;
CRM_CHECK(xml_obj != NULL, return);
id = crm_element_value(xml_obj, XML_ATTR_ID);
if (id == NULL) {
pcmk__config_err("Ignoring <%s> constraint without " XML_ATTR_ID,
xml_obj->name);
return;
}
rsc_first = get_ordering_resource(xml_obj, XML_ORDER_ATTR_FIRST,
XML_ORDER_ATTR_FIRST_INSTANCE,
data_set);
if (rsc_first == NULL) {
return;
}
rsc_then = get_ordering_resource(xml_obj, XML_ORDER_ATTR_THEN,
XML_ORDER_ATTR_THEN_INSTANCE,
data_set);
if (rsc_then == NULL) {
return;
}
action_first = crm_element_value(xml_obj, XML_ORDER_ATTR_FIRST_ACTION);
if (action_first == NULL) {
action_first = PCMK_ACTION_START;
}
action_then = crm_element_value(xml_obj, XML_ORDER_ATTR_THEN_ACTION);
if (action_then == NULL) {
action_then = action_first;
}
kind = get_ordering_type(xml_obj);
symmetry = get_ordering_symmetry(xml_obj, kind, NULL);
flags = ordering_flags_for_kind(kind, action_first, symmetry);
handle_restart_type(rsc_then, kind, pcmk__ar_first_implies_then, flags);
/* If there is a minimum number of instances that must be runnable before
* the 'then' action is runnable, we use a pseudo-action for convenience:
* minimum number of clone instances have runnable actions ->
* pseudo-action is runnable -> dependency is runnable.
*/
min_required_before = get_minimum_first_instances(rsc_first, xml_obj);
if (min_required_before > 0) {
clone_min_ordering(id, rsc_first, action_first, rsc_then, action_then,
flags, min_required_before);
} else {
pcmk__order_resource_actions(rsc_first, action_first, rsc_then,
action_then, flags);
}
if (symmetry == ordering_symmetric) {
inverse_ordering(id, kind, rsc_first, action_first,
rsc_then, action_then);
}
}
/*!
* \internal
* \brief Create a new ordering between two actions
*
* \param[in,out] first_rsc Resource for 'first' action (if NULL and
* \p first_action is a resource action, that
* resource will be used)
* \param[in,out] first_action_task Action key for 'first' action (if NULL and
* \p first_action is not NULL, its UUID will
* be used)
* \param[in,out] first_action 'first' action (if NULL, \p first_rsc and
* \p first_action_task must be set)
*
* \param[in] then_rsc Resource for 'then' action (if NULL and
* \p then_action is a resource action, that
* resource will be used)
* \param[in,out] then_action_task Action key for 'then' action (if NULL and
* \p then_action is not NULL, its UUID will
* be used)
* \param[in] then_action 'then' action (if NULL, \p then_rsc and
* \p then_action_task must be set)
*
* \param[in] flags Group of enum pcmk__action_relation_flags
* \param[in,out] sched Cluster working set to add ordering to
*
* \note This function takes ownership of first_action_task and
* then_action_task, which do not need to be freed by the caller.
*/
void
pcmk__new_ordering(pe_resource_t *first_rsc, char *first_action_task,
pe_action_t *first_action, pe_resource_t *then_rsc,
char *then_action_task, pe_action_t *then_action,
uint32_t flags, pe_working_set_t *sched)
{
pe__ordering_t *order = NULL;
// One of action or resource must be specified for each side
CRM_CHECK(((first_action != NULL) || (first_rsc != NULL))
&& ((then_action != NULL) || (then_rsc != NULL)),
free(first_action_task); free(then_action_task); return);
if ((first_rsc == NULL) && (first_action != NULL)) {
first_rsc = first_action->rsc;
}
if ((then_rsc == NULL) && (then_action != NULL)) {
then_rsc = then_action->rsc;
}
order = calloc(1, sizeof(pe__ordering_t));
CRM_ASSERT(order != NULL);
order->id = sched->order_id++;
order->flags = flags;
order->lh_rsc = first_rsc;
order->rh_rsc = then_rsc;
order->lh_action = first_action;
order->rh_action = then_action;
order->lh_action_task = first_action_task;
order->rh_action_task = then_action_task;
if ((order->lh_action_task == NULL) && (first_action != NULL)) {
order->lh_action_task = strdup(first_action->uuid);
}
if ((order->rh_action_task == NULL) && (then_action != NULL)) {
order->rh_action_task = strdup(then_action->uuid);
}
if ((order->lh_rsc == NULL) && (first_action != NULL)) {
order->lh_rsc = first_action->rsc;
}
if ((order->rh_rsc == NULL) && (then_action != NULL)) {
order->rh_rsc = then_action->rsc;
}
pe_rsc_trace(first_rsc, "Created ordering %d for %s then %s",
(sched->order_id - 1),
pcmk__s(order->lh_action_task, "an underspecified action"),
pcmk__s(order->rh_action_task, "an underspecified action"));
sched->ordering_constraints = g_list_prepend(sched->ordering_constraints,
order);
pcmk__order_migration_equivalents(order);
}
/*!
* \brief Unpack a set in an ordering constraint
*
* \param[in] set Set XML to unpack
* \param[in] parent_kind rsc_order XML "kind" attribute
* \param[in] parent_symmetrical_s rsc_order XML "symmetrical" attribute
* \param[in,out] data_set Cluster working set
*
* \return Standard Pacemaker return code
*/
static int
unpack_order_set(const xmlNode *set, enum pe_order_kind parent_kind,
const char *parent_symmetrical_s, pe_working_set_t *data_set)
{
GList *set_iter = NULL;
GList *resources = NULL;
pe_resource_t *last = NULL;
pe_resource_t *resource = NULL;
int local_kind = parent_kind;
bool sequential = false;
uint32_t flags = pcmk__ar_ordered;
enum ordering_symmetry symmetry;
char *key = NULL;
const char *id = ID(set);
const char *action = crm_element_value(set, "action");
const char *sequential_s = crm_element_value(set, "sequential");
const char *kind_s = crm_element_value(set, XML_ORDER_ATTR_KIND);
if (action == NULL) {
action = PCMK_ACTION_START;
}
if (kind_s) {
local_kind = get_ordering_type(set);
}
if (sequential_s == NULL) {
sequential_s = "1";
}
sequential = crm_is_true(sequential_s);
symmetry = get_ordering_symmetry(set, parent_kind, parent_symmetrical_s);
flags = ordering_flags_for_kind(local_kind, action, symmetry);
for (const xmlNode *xml_rsc = first_named_child(set, XML_TAG_RESOURCE_REF);
xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) {
EXPAND_CONSTRAINT_IDREF(id, resource, ID(xml_rsc));
resources = g_list_append(resources, resource);
}
if (pcmk__list_of_1(resources)) {
crm_trace("Single set: %s", id);
goto done;
}
set_iter = resources;
while (set_iter != NULL) {
resource = (pe_resource_t *) set_iter->data;
set_iter = set_iter->next;
key = pcmk__op_key(resource->id, action, 0);
if (local_kind == pe_order_kind_serialize) {
/* Serialize before everything that comes after */
for (GList *iter = set_iter; iter != NULL; iter = iter->next) {
pe_resource_t *then_rsc = iter->data;
char *then_key = pcmk__op_key(then_rsc->id, action, 0);
pcmk__new_ordering(resource, strdup(key), NULL, then_rsc,
then_key, NULL, flags, data_set);
}
} else if (sequential) {
if (last != NULL) {
pcmk__order_resource_actions(last, action, resource, action,
flags);
}
last = resource;
}
free(key);
}
if (symmetry == ordering_asymmetric) {
goto done;
}
last = NULL;
action = invert_action(action);
flags = ordering_flags_for_kind(local_kind, action,
ordering_symmetric_inverse);
set_iter = resources;
while (set_iter != NULL) {
resource = (pe_resource_t *) set_iter->data;
set_iter = set_iter->next;
if (sequential) {
if (last != NULL) {
pcmk__order_resource_actions(resource, action, last, action,
flags);
}
last = resource;
}
}
done:
g_list_free(resources);
return pcmk_rc_ok;
}
/*!
* \brief Order two resource sets relative to each other
*
* \param[in] id Ordering ID (for logging)
* \param[in] set1 First listed set
* \param[in] set2 Second listed set
* \param[in] kind Ordering kind
* \param[in,out] data_set Cluster working set
* \param[in] symmetry Which ordering symmetry applies to this relation
*
* \return Standard Pacemaker return code
*/
static int
order_rsc_sets(const char *id, const xmlNode *set1, const xmlNode *set2,
enum pe_order_kind kind, pe_working_set_t *data_set,
enum ordering_symmetry symmetry)
{
const xmlNode *xml_rsc = NULL;
const xmlNode *xml_rsc_2 = NULL;
pe_resource_t *rsc_1 = NULL;
pe_resource_t *rsc_2 = NULL;
const char *action_1 = crm_element_value(set1, "action");
const char *action_2 = crm_element_value(set2, "action");
uint32_t flags = pcmk__ar_none;
bool require_all = true;
(void) pcmk__xe_get_bool_attr(set1, "require-all", &require_all);
if (action_1 == NULL) {
action_1 = PCMK_ACTION_START;
}
if (action_2 == NULL) {
action_2 = PCMK_ACTION_START;
}
if (symmetry == ordering_symmetric_inverse) {
action_1 = invert_action(action_1);
action_2 = invert_action(action_2);
}
if (pcmk__str_eq(PCMK_ACTION_STOP, action_1, pcmk__str_none)
|| pcmk__str_eq(PCMK_ACTION_DEMOTE, action_1, pcmk__str_none)) {
/* Assuming: A -> ( B || C) -> D
* The one-or-more logic only applies during the start/promote phase.
* During shutdown neither B nor can shutdown until D is down, so simply
* turn require_all back on.
*/
require_all = true;
}
flags = ordering_flags_for_kind(kind, action_1, symmetry);
/* If we have an unordered set1, whether it is sequential or not is
* irrelevant in regards to set2.
*/
if (!require_all) {
char *task = crm_strdup_printf(PCMK_ACTION_ONE_OR_MORE ":%s", ID(set1));
pe_action_t *unordered_action = get_pseudo_op(task, data_set);
free(task);
pe__set_action_flags(unordered_action, pcmk_action_min_runnable);
for (xml_rsc = first_named_child(set1, XML_TAG_RESOURCE_REF);
xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) {
EXPAND_CONSTRAINT_IDREF(id, rsc_1, ID(xml_rsc));
/* Add an ordering constraint between every element in set1 and the
* pseudo action. If any action in set1 is runnable the pseudo
* action will be runnable.
*/
pcmk__new_ordering(rsc_1, pcmk__op_key(rsc_1->id, action_1, 0),
NULL, NULL, NULL, unordered_action,
pe_order_one_or_more
|pcmk__ar_first_implies_then_graphed,
data_set);
}
for (xml_rsc_2 = first_named_child(set2, XML_TAG_RESOURCE_REF);
xml_rsc_2 != NULL; xml_rsc_2 = crm_next_same_xml(xml_rsc_2)) {
EXPAND_CONSTRAINT_IDREF(id, rsc_2, ID(xml_rsc_2));
/* Add an ordering constraint between the pseudo-action and every
* element in set2. If the pseudo-action is runnable, every action
* in set2 will be runnable.
*/
pcmk__new_ordering(NULL, NULL, unordered_action,
rsc_2, pcmk__op_key(rsc_2->id, action_2, 0),
NULL, flags|pcmk__ar_unrunnable_first_blocks,
data_set);
}
return pcmk_rc_ok;
}
if (pcmk__xe_attr_is_true(set1, "sequential")) {
if (symmetry == ordering_symmetric_inverse) {
// Get the first one
xml_rsc = first_named_child(set1, XML_TAG_RESOURCE_REF);
if (xml_rsc != NULL) {
EXPAND_CONSTRAINT_IDREF(id, rsc_1, ID(xml_rsc));
}
} else {
// Get the last one
const char *rid = NULL;
for (xml_rsc = first_named_child(set1, XML_TAG_RESOURCE_REF);
xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) {
rid = ID(xml_rsc);
}
EXPAND_CONSTRAINT_IDREF(id, rsc_1, rid);
}
}
if (pcmk__xe_attr_is_true(set2, "sequential")) {
if (symmetry == ordering_symmetric_inverse) {
// Get the last one
const char *rid = NULL;
for (xml_rsc = first_named_child(set2, XML_TAG_RESOURCE_REF);
xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) {
rid = ID(xml_rsc);
}
EXPAND_CONSTRAINT_IDREF(id, rsc_2, rid);
} else {
// Get the first one
xml_rsc = first_named_child(set2, XML_TAG_RESOURCE_REF);
if (xml_rsc != NULL) {
EXPAND_CONSTRAINT_IDREF(id, rsc_2, ID(xml_rsc));
}
}
}
if ((rsc_1 != NULL) && (rsc_2 != NULL)) {
pcmk__order_resource_actions(rsc_1, action_1, rsc_2, action_2, flags);
} else if (rsc_1 != NULL) {
for (xml_rsc = first_named_child(set2, XML_TAG_RESOURCE_REF);
xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) {
EXPAND_CONSTRAINT_IDREF(id, rsc_2, ID(xml_rsc));
pcmk__order_resource_actions(rsc_1, action_1, rsc_2, action_2,
flags);
}
} else if (rsc_2 != NULL) {
for (xml_rsc = first_named_child(set1, XML_TAG_RESOURCE_REF);
xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) {
EXPAND_CONSTRAINT_IDREF(id, rsc_1, ID(xml_rsc));
pcmk__order_resource_actions(rsc_1, action_1, rsc_2, action_2,
flags);
}
} else {
for (xml_rsc = first_named_child(set1, XML_TAG_RESOURCE_REF);
xml_rsc != NULL; xml_rsc = crm_next_same_xml(xml_rsc)) {
EXPAND_CONSTRAINT_IDREF(id, rsc_1, ID(xml_rsc));
for (xmlNode *xml_rsc_2 = first_named_child(set2,
XML_TAG_RESOURCE_REF);
xml_rsc_2 != NULL; xml_rsc_2 = crm_next_same_xml(xml_rsc_2)) {
EXPAND_CONSTRAINT_IDREF(id, rsc_2, ID(xml_rsc_2));
pcmk__order_resource_actions(rsc_1, action_1, rsc_2,
action_2, flags);
}
}
}
return pcmk_rc_ok;
}
/*!
* \internal
* \brief If an ordering constraint uses resource tags, expand them
*
* \param[in,out] xml_obj Ordering constraint XML
* \param[out] expanded_xml Equivalent XML with tags expanded
* \param[in] data_set Cluster working set
*
* \return Standard Pacemaker return code (specifically, pcmk_rc_ok on success,
* and pcmk_rc_unpack_error on invalid configuration)
*/
static int
unpack_order_tags(xmlNode *xml_obj, xmlNode **expanded_xml,
const pe_working_set_t *data_set)
{
const char *id_first = NULL;
const char *id_then = NULL;
const char *action_first = NULL;
const char *action_then = NULL;
pe_resource_t *rsc_first = NULL;
pe_resource_t *rsc_then = NULL;
pe_tag_t *tag_first = NULL;
pe_tag_t *tag_then = NULL;
xmlNode *rsc_set_first = NULL;
xmlNode *rsc_set_then = NULL;
bool any_sets = false;
// Check whether there are any resource sets with template or tag references
*expanded_xml = pcmk__expand_tags_in_sets(xml_obj, data_set);
if (*expanded_xml != NULL) {
crm_log_xml_trace(*expanded_xml, "Expanded rsc_order");
return pcmk_rc_ok;
}
id_first = crm_element_value(xml_obj, XML_ORDER_ATTR_FIRST);
id_then = crm_element_value(xml_obj, XML_ORDER_ATTR_THEN);
if ((id_first == NULL) || (id_then == NULL)) {
return pcmk_rc_ok;
}
if (!pcmk__valid_resource_or_tag(data_set, id_first, &rsc_first,
&tag_first)) {
pcmk__config_err("Ignoring constraint '%s' because '%s' is not a "
"valid resource or tag", ID(xml_obj), id_first);
return pcmk_rc_unpack_error;
}
if (!pcmk__valid_resource_or_tag(data_set, id_then, &rsc_then, &tag_then)) {
pcmk__config_err("Ignoring constraint '%s' because '%s' is not a "
"valid resource or tag", ID(xml_obj), id_then);
return pcmk_rc_unpack_error;
}
if ((rsc_first != NULL) && (rsc_then != NULL)) {
// Neither side references a template or tag
return pcmk_rc_ok;
}
action_first = crm_element_value(xml_obj, XML_ORDER_ATTR_FIRST_ACTION);
action_then = crm_element_value(xml_obj, XML_ORDER_ATTR_THEN_ACTION);
*expanded_xml = copy_xml(xml_obj);
// Convert template/tag reference in "first" into constraint resource_set
if (!pcmk__tag_to_set(*expanded_xml, &rsc_set_first, XML_ORDER_ATTR_FIRST,
true, data_set)) {
free_xml(*expanded_xml);
*expanded_xml = NULL;
return pcmk_rc_unpack_error;
}
if (rsc_set_first != NULL) {
if (action_first != NULL) {
// Move "first-action" into converted resource_set as "action"
crm_xml_add(rsc_set_first, "action", action_first);
xml_remove_prop(*expanded_xml, XML_ORDER_ATTR_FIRST_ACTION);
}
any_sets = true;
}
// Convert template/tag reference in "then" into constraint resource_set
if (!pcmk__tag_to_set(*expanded_xml, &rsc_set_then, XML_ORDER_ATTR_THEN,
true, data_set)) {
free_xml(*expanded_xml);
*expanded_xml = NULL;
return pcmk_rc_unpack_error;
}
if (rsc_set_then != NULL) {
if (action_then != NULL) {
// Move "then-action" into converted resource_set as "action"
crm_xml_add(rsc_set_then, "action", action_then);
xml_remove_prop(*expanded_xml, XML_ORDER_ATTR_THEN_ACTION);
}
any_sets = true;
}
if (any_sets) {
crm_log_xml_trace(*expanded_xml, "Expanded rsc_order");
} else {
free_xml(*expanded_xml);
*expanded_xml = NULL;
}
return pcmk_rc_ok;
}
/*!
* \internal
* \brief Unpack ordering constraint XML
*
* \param[in,out] xml_obj Ordering constraint XML to unpack
* \param[in,out] data_set Cluster working set
*/
void
pcmk__unpack_ordering(xmlNode *xml_obj, pe_working_set_t *data_set)
{
xmlNode *set = NULL;
xmlNode *last = NULL;
xmlNode *orig_xml = NULL;
xmlNode *expanded_xml = NULL;
const char *id = crm_element_value(xml_obj, XML_ATTR_ID);
const char *invert = crm_element_value(xml_obj, XML_CONS_ATTR_SYMMETRICAL);
enum pe_order_kind kind = get_ordering_type(xml_obj);
enum ordering_symmetry symmetry = get_ordering_symmetry(xml_obj, kind,
NULL);
// Expand any resource tags in the constraint XML
if (unpack_order_tags(xml_obj, &expanded_xml, data_set) != pcmk_rc_ok) {
return;
}
if (expanded_xml != NULL) {
orig_xml = xml_obj;
xml_obj = expanded_xml;
}
// If the constraint has resource sets, unpack them
for (set = first_named_child(xml_obj, XML_CONS_TAG_RSC_SET);
set != NULL; set = crm_next_same_xml(set)) {
set = expand_idref(set, data_set->input);
if ((set == NULL) // Configuration error, message already logged
|| (unpack_order_set(set, kind, invert, data_set) != pcmk_rc_ok)) {
if (expanded_xml != NULL) {
free_xml(expanded_xml);
}
return;
}
if (last != NULL) {
if (order_rsc_sets(id, last, set, kind, data_set,
symmetry) != pcmk_rc_ok) {
if (expanded_xml != NULL) {
free_xml(expanded_xml);
}
return;
}
if ((symmetry == ordering_symmetric)
&& (order_rsc_sets(id, set, last, kind, data_set,
ordering_symmetric_inverse) != pcmk_rc_ok)) {
if (expanded_xml != NULL) {
free_xml(expanded_xml);
}
return;
}
}
last = set;
}
if (expanded_xml) {
free_xml(expanded_xml);
xml_obj = orig_xml;
}
// If the constraint has no resource sets, unpack it as a simple ordering
if (last == NULL) {
return unpack_simple_rsc_order(xml_obj, data_set);
}
}
static bool
ordering_is_invalid(pe_action_t *action, pe_action_wrapper_t *input)
{
/* Prevent user-defined ordering constraints between resources
* running in a guest node and the resource that defines that node.
*/
if (!pcmk_is_set(input->type, pe_order_preserve)
&& (input->action->rsc != NULL)
&& pcmk__rsc_corresponds_to_guest(action->rsc, input->action->node)) {
crm_warn("Invalid ordering constraint between %s and %s",
input->action->rsc->id, action->rsc->id);
return true;
}
/* If there's an order like
* "rscB_stop node2"-> "load_stopped_node2" -> "rscA_migrate_to node1"
*
* then rscA is being migrated from node1 to node2, while rscB is being
* migrated from node2 to node1. If there would be a graph loop,
* break the order "load_stopped_node2" -> "rscA_migrate_to node1".
*/
- if ((input->type == pe_order_load) && action->rsc
+ if (((uint32_t) input->type == pcmk__ar_if_on_same_node_or_target)
+ && (action->rsc != NULL)
&& pcmk__str_eq(action->task, PCMK_ACTION_MIGRATE_TO, pcmk__str_none)
&& pcmk__graph_has_loop(action, action, input)) {
return true;
}
return false;
}
void
pcmk__disable_invalid_orderings(pe_working_set_t *data_set)
{
for (GList *iter = data_set->actions; iter != NULL; iter = iter->next) {
pe_action_t *action = (pe_action_t *) iter->data;
pe_action_wrapper_t *input = NULL;
for (GList *input_iter = action->actions_before;
input_iter != NULL; input_iter = input_iter->next) {
input = (pe_action_wrapper_t *) input_iter->data;
if (ordering_is_invalid(action, input)) {
input->type = pcmk__ar_none;
}
}
}
}
/*!
* \internal
* \brief Order stops on a node before the node's shutdown
*
* \param[in,out] node Node being shut down
* \param[in] shutdown_op Shutdown action for node
*/
void
pcmk__order_stops_before_shutdown(pe_node_t *node, pe_action_t *shutdown_op)
{
for (GList *iter = node->details->data_set->actions;
iter != NULL; iter = iter->next) {
pe_action_t *action = (pe_action_t *) iter->data;
// Only stops on the node shutting down are relevant
if (!pe__same_node(action->node, node)
|| !pcmk__str_eq(action->task, PCMK_ACTION_STOP, pcmk__str_none)) {
continue;
}
// Resources and nodes in maintenance mode won't be touched
if (pcmk_is_set(action->rsc->flags, pcmk_rsc_maintenance)) {
pe_rsc_trace(action->rsc,
"Not ordering %s before shutdown of %s because "
"resource in maintenance mode",
action->uuid, pe__node_name(node));
continue;
} else if (node->details->maintenance) {
pe_rsc_trace(action->rsc,
"Not ordering %s before shutdown of %s because "
"node in maintenance mode",
action->uuid, pe__node_name(node));
continue;
}
/* Don't touch a resource that is unmanaged or blocked, to avoid
* blocking the shutdown (though if another action depends on this one,
* we may still end up blocking)
*/
if (!pcmk_any_flags_set(action->rsc->flags,
pcmk_rsc_managed|pcmk_rsc_blocked)) {
pe_rsc_trace(action->rsc,
"Not ordering %s before shutdown of %s because "
"resource is unmanaged or blocked",
action->uuid, pe__node_name(node));
continue;
}
pe_rsc_trace(action->rsc, "Ordering %s before shutdown of %s",
action->uuid, pe__node_name(node));
pe__clear_action_flags(action, pcmk_action_optional);
pcmk__new_ordering(action->rsc, NULL, action, NULL,
strdup(PCMK_ACTION_DO_SHUTDOWN), shutdown_op,
pcmk__ar_ordered|pcmk__ar_unrunnable_first_blocks,
node->details->data_set);
}
}
/*!
* \brief Find resource actions matching directly or as child
*
* \param[in] rsc Resource to check
* \param[in] original_key Action key to search for (possibly referencing
* parent of \rsc)
*
* \return Newly allocated list of matching actions
* \note It is the caller's responsibility to free the result with g_list_free()
*/
static GList *
find_actions_by_task(const pe_resource_t *rsc, const char *original_key)
{
// Search under given task key directly
GList *list = find_actions(rsc->actions, original_key, NULL);
if (list == NULL) {
// Search again using this resource's ID
char *key = NULL;
char *task = NULL;
guint interval_ms = 0;
if (parse_op_key(original_key, NULL, &task, &interval_ms)) {
key = pcmk__op_key(rsc->id, task, interval_ms);
list = find_actions(rsc->actions, key, NULL);
free(key);
free(task);
} else {
crm_err("Invalid operation key (bug?): %s", original_key);
}
}
return list;
}
/*!
* \internal
* \brief Order relevant resource actions after a given action
*
* \param[in,out] first_action Action to order after (or NULL if none runnable)
* \param[in] rsc Resource whose actions should be ordered
* \param[in,out] order Ordering constraint being applied
*/
static void
order_resource_actions_after(pe_action_t *first_action,
const pe_resource_t *rsc, pe__ordering_t *order)
{
GList *then_actions = NULL;
uint32_t flags = pcmk__ar_none;
CRM_CHECK((rsc != NULL) && (order != NULL), return);
flags = order->flags;
pe_rsc_trace(rsc, "Applying ordering %d for 'then' resource %s",
order->id, rsc->id);
if (order->rh_action != NULL) {
then_actions = g_list_prepend(NULL, order->rh_action);
} else {
then_actions = find_actions_by_task(rsc, order->rh_action_task);
}
if (then_actions == NULL) {
pe_rsc_trace(rsc, "Ignoring ordering %d: no %s actions found for %s",
order->id, order->rh_action_task, rsc->id);
return;
}
if ((first_action != NULL) && (first_action->rsc == rsc)
&& pcmk_is_set(first_action->flags, pcmk_action_migration_abort)) {
pe_rsc_trace(rsc,
"Detected dangling migration ordering (%s then %s %s)",
first_action->uuid, order->rh_action_task, rsc->id);
pe__clear_order_flags(flags, pcmk__ar_first_implies_then);
}
if ((first_action == NULL)
&& !pcmk_is_set(flags, pcmk__ar_first_implies_then)) {
pe_rsc_debug(rsc,
"Ignoring ordering %d for %s: No first action found",
order->id, rsc->id);
g_list_free(then_actions);
return;
}
for (GList *iter = then_actions; iter != NULL; iter = iter->next) {
pe_action_t *then_action_iter = (pe_action_t *) iter->data;
if (first_action != NULL) {
order_actions(first_action, then_action_iter, flags);
} else {
pe__clear_action_flags(then_action_iter, pcmk_action_runnable);
crm_warn("%s of %s is unrunnable because there is no %s of %s "
"to order it after", then_action_iter->task, rsc->id,
order->lh_action_task, order->lh_rsc->id);
}
}
g_list_free(then_actions);
}
static void
rsc_order_first(pe_resource_t *first_rsc, pe__ordering_t *order)
{
GList *first_actions = NULL;
pe_action_t *first_action = order->lh_action;
pe_resource_t *then_rsc = order->rh_rsc;
CRM_ASSERT(first_rsc != NULL);
pe_rsc_trace(first_rsc, "Applying ordering constraint %d (first: %s)",
order->id, first_rsc->id);
if (first_action != NULL) {
first_actions = g_list_prepend(NULL, first_action);
} else {
first_actions = find_actions_by_task(first_rsc, order->lh_action_task);
}
if ((first_actions == NULL) && (first_rsc == then_rsc)) {
pe_rsc_trace(first_rsc,
"Ignoring constraint %d: first (%s for %s) not found",
order->id, order->lh_action_task, first_rsc->id);
} else if (first_actions == NULL) {
char *key = NULL;
char *op_type = NULL;
guint interval_ms = 0;
parse_op_key(order->lh_action_task, NULL, &op_type, &interval_ms);
key = pcmk__op_key(first_rsc->id, op_type, interval_ms);
if ((first_rsc->fns->state(first_rsc, TRUE) == pcmk_role_stopped)
&& pcmk__str_eq(op_type, PCMK_ACTION_STOP, pcmk__str_none)) {
free(key);
pe_rsc_trace(first_rsc,
"Ignoring constraint %d: first (%s for %s) not found",
order->id, order->lh_action_task, first_rsc->id);
} else if ((first_rsc->fns->state(first_rsc,
TRUE) == pcmk_role_unpromoted)
&& pcmk__str_eq(op_type, PCMK_ACTION_DEMOTE,
pcmk__str_none)) {
free(key);
pe_rsc_trace(first_rsc,
"Ignoring constraint %d: first (%s for %s) not found",
order->id, order->lh_action_task, first_rsc->id);
} else {
pe_rsc_trace(first_rsc,
"Creating first (%s for %s) for constraint %d ",
order->lh_action_task, first_rsc->id, order->id);
first_action = custom_action(first_rsc, key, op_type, NULL, TRUE,
TRUE, first_rsc->cluster);
first_actions = g_list_prepend(NULL, first_action);
}
free(op_type);
}
if (then_rsc == NULL) {
if (order->rh_action == NULL) {
pe_rsc_trace(first_rsc, "Ignoring constraint %d: then not found",
order->id);
return;
}
then_rsc = order->rh_action->rsc;
}
for (GList *iter = first_actions; iter != NULL; iter = iter->next) {
first_action = iter->data;
if (then_rsc == NULL) {
order_actions(first_action, order->rh_action, order->flags);
} else {
order_resource_actions_after(first_action, then_rsc, order);
}
}
g_list_free(first_actions);
}
// GFunc to call pcmk__block_colocation_dependents()
static void
block_colocation_dependents(gpointer data, gpointer user_data)
{
pcmk__block_colocation_dependents(data);
}
// GFunc to call pcmk__update_action_for_orderings()
static void
update_action_for_orderings(gpointer data, gpointer user_data)
{
pcmk__update_action_for_orderings((pe_action_t *) data,
(pe_working_set_t *) user_data);
}
/*!
* \internal
* \brief Apply all ordering constraints
*
* \param[in,out] sched Cluster working set
*/
void
pcmk__apply_orderings(pe_working_set_t *sched)
{
crm_trace("Applying ordering constraints");
/* Ordering constraints need to be processed in the order they were created.
* rsc_order_first() and order_resource_actions_after() require the relevant
* actions to already exist in some cases, but rsc_order_first() will create
* the 'first' action in certain cases. Thus calling rsc_order_first() can
* change the behavior of later-created orderings.
*
* Also, g_list_append() should be avoided for performance reasons, so we
* prepend orderings when creating them and reverse the list here.
*
* @TODO This is brittle and should be carefully redesigned so that the
* order of creation doesn't matter, and the reverse becomes unneeded.
*/
sched->ordering_constraints = g_list_reverse(sched->ordering_constraints);
for (GList *iter = sched->ordering_constraints;
iter != NULL; iter = iter->next) {
pe__ordering_t *order = iter->data;
pe_resource_t *rsc = order->lh_rsc;
if (rsc != NULL) {
rsc_order_first(rsc, order);
continue;
}
rsc = order->rh_rsc;
if (rsc != NULL) {
order_resource_actions_after(order->lh_action, rsc, order);
} else {
crm_trace("Applying ordering constraint %d (non-resource actions)",
order->id);
order_actions(order->lh_action, order->rh_action, order->flags);
}
}
g_list_foreach(sched->actions, block_colocation_dependents, NULL);
crm_trace("Ordering probes");
pcmk__order_probes(sched);
crm_trace("Updating %d actions", g_list_length(sched->actions));
g_list_foreach(sched->actions, update_action_for_orderings, sched);
pcmk__disable_invalid_orderings(sched);
}
/*!
* \internal
* \brief Order a given action after each action in a given list
*
* \param[in,out] after "After" action
* \param[in,out] list List of "before" actions
*/
void
pcmk__order_after_each(pe_action_t *after, GList *list)
{
const char *after_desc = (after->task == NULL)? after->uuid : after->task;
for (GList *iter = list; iter != NULL; iter = iter->next) {
pe_action_t *before = (pe_action_t *) iter->data;
const char *before_desc = before->task? before->task : before->uuid;
crm_debug("Ordering %s on %s before %s on %s",
before_desc, pe__node_name(before->node),
after_desc, pe__node_name(after->node));
order_actions(before, after, pcmk__ar_ordered);
}
}
/*!
* \internal
* \brief Order promotions and demotions for restarts of a clone or bundle
*
* \param[in,out] rsc Clone or bundle to order
*/
void
pcmk__promotable_restart_ordering(pe_resource_t *rsc)
{
// Order start and promote after all instances are stopped
pcmk__order_resource_actions(rsc, PCMK_ACTION_STOPPED,
rsc, PCMK_ACTION_START,
pcmk__ar_ordered);
pcmk__order_resource_actions(rsc, PCMK_ACTION_STOPPED,
rsc, PCMK_ACTION_PROMOTE,
pcmk__ar_ordered);
// Order stop, start, and promote after all instances are demoted
pcmk__order_resource_actions(rsc, PCMK_ACTION_DEMOTED,
rsc, PCMK_ACTION_STOP,
pcmk__ar_ordered);
pcmk__order_resource_actions(rsc, PCMK_ACTION_DEMOTED,
rsc, PCMK_ACTION_START,
pcmk__ar_ordered);
pcmk__order_resource_actions(rsc, PCMK_ACTION_DEMOTED,
rsc, PCMK_ACTION_PROMOTE,
pcmk__ar_ordered);
// Order promote after all instances are started
pcmk__order_resource_actions(rsc, PCMK_ACTION_RUNNING,
rsc, PCMK_ACTION_PROMOTE,
pcmk__ar_ordered);
// Order demote after all instances are demoted
pcmk__order_resource_actions(rsc, PCMK_ACTION_DEMOTE,
rsc, PCMK_ACTION_DEMOTED,
pcmk__ar_ordered);
}
diff --git a/lib/pacemaker/pcmk_sched_probes.c b/lib/pacemaker/pcmk_sched_probes.c
index daa783ea80..2db2b5036b 100644
--- a/lib/pacemaker/pcmk_sched_probes.c
+++ b/lib/pacemaker/pcmk_sched_probes.c
@@ -1,903 +1,903 @@
/*
* Copyright 2004-2023 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 <glib.h>
#include <crm/crm.h>
#include <crm/pengine/status.h>
#include <pacemaker-internal.h>
#include "libpacemaker_private.h"
/*!
* \internal
* \brief Add the expected result to a newly created probe
*
* \param[in,out] probe Probe action to add expected result to
* \param[in] rsc Resource that probe is for
* \param[in] node Node that probe will run on
*/
static void
add_expected_result(pe_action_t *probe, const pe_resource_t *rsc,
const pe_node_t *node)
{
// Check whether resource is currently active on node
pe_node_t *running = pe_find_node_id(rsc->running_on, node->details->id);
// The expected result is what we think the resource's current state is
if (running == NULL) {
pe__add_action_expected_result(probe, CRM_EX_NOT_RUNNING);
} else if (rsc->role == pcmk_role_promoted) {
pe__add_action_expected_result(probe, CRM_EX_PROMOTED);
}
}
/*!
* \internal
* \brief Create any needed robes on a node for a list of resources
*
* \param[in,out] rscs List of resources to create probes for
* \param[in,out] node Node to create probes on
*
* \return true if any probe was created, otherwise false
*/
bool
pcmk__probe_resource_list(GList *rscs, pe_node_t *node)
{
bool any_created = false;
for (GList *iter = rscs; iter != NULL; iter = iter->next) {
pe_resource_t *rsc = (pe_resource_t *) iter->data;
if (rsc->cmds->create_probe(rsc, node)) {
any_created = true;
}
}
return any_created;
}
/*!
* \internal
* \brief Order one resource's start after another's start-up probe
*
* \param[in,out] rsc1 Resource that might get start-up probe
* \param[in] rsc2 Resource that might be started
*/
static void
probe_then_start(pe_resource_t *rsc1, pe_resource_t *rsc2)
{
if ((rsc1->allocated_to != NULL)
&& (g_hash_table_lookup(rsc1->known_on,
rsc1->allocated_to->details->id) == NULL)) {
pcmk__new_ordering(rsc1,
pcmk__op_key(rsc1->id, PCMK_ACTION_MONITOR, 0),
NULL,
rsc2, pcmk__op_key(rsc2->id, PCMK_ACTION_START, 0),
NULL,
pcmk__ar_ordered, rsc1->cluster);
}
}
/*!
* \internal
* \brief Check whether a guest resource will stop
*
* \param[in] node Guest node to check
*
* \return true if guest resource will likely stop, otherwise false
*/
static bool
guest_resource_will_stop(const pe_node_t *node)
{
const pe_resource_t *guest_rsc = node->details->remote_rsc->container;
/* Ideally, we'd check whether the guest has a required stop, but that
* information doesn't exist yet, so approximate it ...
*/
return node->details->remote_requires_reset
|| node->details->unclean
|| pcmk_is_set(guest_rsc->flags, pcmk_rsc_failed)
|| (guest_rsc->next_role == pcmk_role_stopped)
// Guest is moving
|| ((guest_rsc->role > pcmk_role_stopped)
&& (guest_rsc->allocated_to != NULL)
&& (pe_find_node(guest_rsc->running_on,
guest_rsc->allocated_to->details->uname) == NULL));
}
/*!
* \internal
* \brief Create a probe action for a resource on a node
*
* \param[in,out] rsc Resource to create probe for
* \param[in,out] node Node to create probe on
*
* \return Newly created probe action
*/
static pe_action_t *
probe_action(pe_resource_t *rsc, pe_node_t *node)
{
pe_action_t *probe = NULL;
char *key = pcmk__op_key(rsc->id, PCMK_ACTION_MONITOR, 0);
crm_debug("Scheduling probe of %s %s on %s",
role2text(rsc->role), rsc->id, pe__node_name(node));
probe = custom_action(rsc, key, PCMK_ACTION_MONITOR, node, FALSE, TRUE,
rsc->cluster);
pe__clear_action_flags(probe, pcmk_action_optional);
pcmk__order_vs_unfence(rsc, node, probe, pcmk__ar_ordered);
add_expected_result(probe, rsc, node);
return probe;
}
/*!
* \internal
* \brief Create probes for a resource on a node, if needed
*
* \brief Schedule any probes needed for a resource on a node
*
* \param[in,out] rsc Resource to create probe for
* \param[in,out] node Node to create probe on
*
* \return true if any probe was created, otherwise false
*/
bool
pcmk__probe_rsc_on_node(pe_resource_t *rsc, pe_node_t *node)
{
uint32_t flags = pcmk__ar_ordered;
pe_action_t *probe = NULL;
pe_node_t *allowed = NULL;
pe_resource_t *top = uber_parent(rsc);
const char *reason = NULL;
CRM_ASSERT((rsc != NULL) && (node != NULL));
if (!pcmk_is_set(rsc->cluster->flags, pcmk_sched_probe_resources)) {
reason = "start-up probes are disabled";
goto no_probe;
}
if (pe__is_guest_or_remote_node(node)) {
const char *class = crm_element_value(rsc->xml, XML_AGENT_ATTR_CLASS);
if (pcmk__str_eq(class, PCMK_RESOURCE_CLASS_STONITH, pcmk__str_none)) {
reason = "Pacemaker Remote nodes cannot run stonith agents";
goto no_probe;
} else if (pe__is_guest_node(node)
&& pe__resource_contains_guest_node(rsc->cluster, rsc)) {
reason = "guest nodes cannot run resources containing guest nodes";
goto no_probe;
} else if (rsc->is_remote_node) {
reason = "Pacemaker Remote nodes cannot host remote connections";
goto no_probe;
}
}
// If this is a collective resource, probes are created for its children
if (rsc->children != NULL) {
return pcmk__probe_resource_list(rsc->children, node);
}
if ((rsc->container != NULL) && !rsc->is_remote_node) {
reason = "resource is inside a container";
goto no_probe;
} else if (pcmk_is_set(rsc->flags, pcmk_rsc_removed)) {
reason = "resource is orphaned";
goto no_probe;
} else if (g_hash_table_lookup(rsc->known_on, node->details->id) != NULL) {
reason = "resource state is already known";
goto no_probe;
}
allowed = g_hash_table_lookup(rsc->allowed_nodes, node->details->id);
if (rsc->exclusive_discover || top->exclusive_discover) {
// Exclusive discovery is enabled ...
if (allowed == NULL) {
// ... but this node is not allowed to run the resource
reason = "resource has exclusive discovery but is not allowed "
"on node";
goto no_probe;
} else if (allowed->rsc_discover_mode != pcmk_probe_exclusive) {
// ... but no constraint marks this node for discovery of resource
reason = "resource has exclusive discovery but is not enabled "
"on node";
goto no_probe;
}
}
if (allowed == NULL) {
allowed = node;
}
if (allowed->rsc_discover_mode == pcmk_probe_never) {
reason = "node has discovery disabled";
goto no_probe;
}
if (pe__is_guest_node(node)) {
pe_resource_t *guest = node->details->remote_rsc->container;
if (guest->role == pcmk_role_stopped) {
// The guest is stopped, so we know no resource is active there
reason = "node's guest is stopped";
probe_then_start(guest, top);
goto no_probe;
} else if (guest_resource_will_stop(node)) {
reason = "node's guest will stop";
// Order resource start after guest stop (in case it's restarting)
pcmk__new_ordering(guest,
pcmk__op_key(guest->id, PCMK_ACTION_STOP, 0),
NULL, top,
pcmk__op_key(top->id, PCMK_ACTION_START, 0),
NULL, pcmk__ar_ordered, rsc->cluster);
goto no_probe;
}
}
// We've eliminated all cases where a probe is not needed, so now it is
probe = probe_action(rsc, node);
/* Below, we will order the probe relative to start or reload. If this is a
* clone instance, the start or reload is for the entire clone rather than
* just the instance. Otherwise, the start or reload is for the resource
* itself.
*/
if (!pe_rsc_is_clone(top)) {
top = rsc;
}
/* Prevent a start if the resource can't be probed, but don't cause the
* resource or entire clone to stop if already active.
*/
if (!pcmk_is_set(probe->flags, pcmk_action_runnable)
&& (top->running_on == NULL)) {
pe__set_order_flags(flags, pcmk__ar_unrunnable_first_blocks);
}
// Start or reload after probing the resource
pcmk__new_ordering(rsc, NULL, probe,
top, pcmk__op_key(top->id, PCMK_ACTION_START, 0), NULL,
flags, rsc->cluster);
pcmk__new_ordering(rsc, NULL, probe, top, reload_key(rsc), NULL,
pcmk__ar_ordered, rsc->cluster);
return true;
no_probe:
pe_rsc_trace(rsc,
"Skipping probe for %s on %s because %s",
rsc->id, node->details->id, reason);
return false;
}
/*!
* \internal
* \brief Check whether a probe should be ordered before another action
*
* \param[in] probe Probe action to check
* \param[in] then Other action to check
*
* \return true if \p probe should be ordered before \p then, otherwise false
*/
static bool
probe_needed_before_action(const pe_action_t *probe, const pe_action_t *then)
{
// Probes on a node are performed after unfencing it, not before
if (pcmk__str_eq(then->task, PCMK_ACTION_STONITH, pcmk__str_none)
&& pe__same_node(probe->node, then->node)) {
const char *op = g_hash_table_lookup(then->meta, "stonith_action");
if (pcmk__str_eq(op, PCMK_ACTION_ON, pcmk__str_casei)) {
return false;
}
}
// Probes should be done on a node before shutting it down
if (pcmk__str_eq(then->task, PCMK_ACTION_DO_SHUTDOWN, pcmk__str_none)
&& (probe->node != NULL) && (then->node != NULL)
&& !pe__same_node(probe->node, then->node)) {
return false;
}
// Otherwise probes should always be done before any other action
return true;
}
/*!
* \internal
* \brief Add implicit "probe then X" orderings for "stop then X" orderings
*
* If the state of a resource is not known yet, a probe will be scheduled,
* expecting a "not running" result. If the probe fails, a stop will not be
* scheduled until the next transition. Thus, if there are ordering constraints
* like "stop this resource then do something else that's not for the same
* resource", add implicit "probe this resource then do something" equivalents
* so the relation is upheld until we know whether a stop is needed.
*
* \param[in,out] data_set Cluster working set
*/
static void
add_probe_orderings_for_stops(pe_working_set_t *data_set)
{
for (GList *iter = data_set->ordering_constraints; iter != NULL;
iter = iter->next) {
pe__ordering_t *order = iter->data;
uint32_t order_flags = pcmk__ar_ordered;
GList *probes = NULL;
GList *then_actions = NULL;
pe_action_t *first = NULL;
pe_action_t *then = NULL;
// Skip disabled orderings
if (order->flags == pcmk__ar_none) {
continue;
}
// Skip non-resource orderings, and orderings for the same resource
if ((order->lh_rsc == NULL) || (order->lh_rsc == order->rh_rsc)) {
continue;
}
// Skip invalid orderings (shouldn't be possible)
first = order->lh_action;
then = order->rh_action;
if (((first == NULL) && (order->lh_action_task == NULL))
|| ((then == NULL) && (order->rh_action_task == NULL))) {
continue;
}
// Skip orderings for first actions other than stop
if ((first != NULL) && !pcmk__str_eq(first->task, PCMK_ACTION_STOP,
pcmk__str_none)) {
continue;
} else if ((first == NULL)
&& !pcmk__ends_with(order->lh_action_task,
"_" PCMK_ACTION_STOP "_0")) {
continue;
}
/* Do not imply a probe ordering for a resource inside of a stopping
* container. Otherwise, it might introduce a transition loop, since a
* probe could be scheduled after the container starts again.
*/
if ((order->rh_rsc != NULL)
&& (order->lh_rsc->container == order->rh_rsc)) {
if ((then != NULL) && pcmk__str_eq(then->task, PCMK_ACTION_STOP,
pcmk__str_none)) {
continue;
} else if ((then == NULL)
&& pcmk__ends_with(order->rh_action_task,
"_" PCMK_ACTION_STOP "_0")) {
continue;
}
}
// Preserve certain order options for future filtering
if (pcmk_is_set(order->flags, pcmk__ar_if_first_unmigratable)) {
pe__set_order_flags(order_flags, pcmk__ar_if_first_unmigratable);
}
if (pcmk_is_set(order->flags, pcmk__ar_if_on_same_node)) {
pe__set_order_flags(order_flags, pcmk__ar_if_on_same_node);
}
// Preserve certain order types for future filtering
if ((order->flags == pe_order_anti_colocation)
- || (order->flags == pe_order_load)) {
+ || (order->flags == pcmk__ar_if_on_same_node_or_target)) {
order_flags = order->flags;
}
// List all scheduled probes for the first resource
probes = pe__resource_actions(order->lh_rsc, NULL, PCMK_ACTION_MONITOR,
FALSE);
if (probes == NULL) { // There aren't any
continue;
}
// List all relevant "then" actions
if (then != NULL) {
then_actions = g_list_prepend(NULL, then);
} else if (order->rh_rsc != NULL) {
then_actions = find_actions(order->rh_rsc->actions,
order->rh_action_task, NULL);
if (then_actions == NULL) { // There aren't any
g_list_free(probes);
continue;
}
}
crm_trace("Implying 'probe then' orderings for '%s then %s' "
"(id=%d, type=%.6x)",
((first == NULL)? order->lh_action_task : first->uuid),
((then == NULL)? order->rh_action_task : then->uuid),
order->id, order->flags);
for (GList *probe_iter = probes; probe_iter != NULL;
probe_iter = probe_iter->next) {
pe_action_t *probe = (pe_action_t *) probe_iter->data;
for (GList *then_iter = then_actions; then_iter != NULL;
then_iter = then_iter->next) {
pe_action_t *then = (pe_action_t *) then_iter->data;
if (probe_needed_before_action(probe, then)) {
order_actions(probe, then, order_flags);
}
}
}
g_list_free(then_actions);
g_list_free(probes);
}
}
/*!
* \internal
* \brief Add necessary orderings between probe and starts of clone instances
*
* , in additon to the ordering with the parent resource added upon creating
* the probe.
*
* \param[in,out] probe Probe as 'first' action in an ordering
* \param[in,out] after 'then' action wrapper in the ordering
*/
static void
add_start_orderings_for_probe(pe_action_t *probe, pe_action_wrapper_t *after)
{
uint32_t flags = pcmk__ar_ordered|pcmk__ar_unrunnable_first_blocks;
/* Although the ordering between the probe of the clone instance and the
* start of its parent has been added in pcmk__probe_rsc_on_node(), we
* avoided enforcing `pcmk__ar_unrunnable_first_blocks` order type for that
* as long as any of the clone instances are running to prevent them from
* being unexpectedly stopped.
*
* On the other hand, we still need to prevent any inactive instances from
* starting unless the probe is runnable so that we don't risk starting too
* many instances before we know the state on all nodes.
*/
if ((after->action->rsc->variant <= pcmk_rsc_variant_group)
|| pcmk_is_set(probe->flags, pcmk_action_runnable)
// The order type is already enforced for its parent.
|| pcmk_is_set(after->type, pcmk__ar_unrunnable_first_blocks)
|| (pe__const_top_resource(probe->rsc, false) != after->action->rsc)
|| !pcmk__str_eq(after->action->task, PCMK_ACTION_START,
pcmk__str_none)) {
return;
}
crm_trace("Adding probe start orderings for 'unrunnable %s@%s "
"then instances of %s@%s'",
probe->uuid, pe__node_name(probe->node),
after->action->uuid, pe__node_name(after->action->node));
for (GList *then_iter = after->action->actions_after; then_iter != NULL;
then_iter = then_iter->next) {
pe_action_wrapper_t *then = (pe_action_wrapper_t *) then_iter->data;
if (then->action->rsc->running_on
|| (pe__const_top_resource(then->action->rsc, false)
!= after->action->rsc)
|| !pcmk__str_eq(then->action->task, PCMK_ACTION_START,
pcmk__str_none)) {
continue;
}
crm_trace("Adding probe start ordering for 'unrunnable %s@%s "
"then %s@%s' (type=%#.6x)",
probe->uuid, pe__node_name(probe->node),
then->action->uuid, pe__node_name(then->action->node), flags);
/* Prevent the instance from starting if the instance can't, but don't
* cause any other intances to stop if already active.
*/
order_actions(probe, then->action, flags);
}
return;
}
/*!
* \internal
* \brief Order probes before restarts and re-promotes
*
* If a given ordering is a "probe then start" or "probe then promote" ordering,
* add an implicit "probe then stop/demote" ordering in case the action is part
* of a restart/re-promote, and do the same recursively for all actions ordered
* after the "then" action.
*
* \param[in,out] probe Probe as 'first' action in an ordering
* \param[in,out] after 'then' action in the ordering
*/
static void
add_restart_orderings_for_probe(pe_action_t *probe, pe_action_t *after)
{
GList *iter = NULL;
bool interleave = false;
pe_resource_t *compatible_rsc = NULL;
// Validate that this is a resource probe followed by some action
if ((after == NULL) || (probe == NULL) || (probe->rsc == NULL)
|| (probe->rsc->variant != pcmk_rsc_variant_primitive)
|| !pcmk__str_eq(probe->task, PCMK_ACTION_MONITOR, pcmk__str_none)) {
return;
}
// Avoid running into any possible loop
if (pcmk_is_set(after->flags, pcmk_action_detect_loop)) {
return;
}
pe__set_action_flags(after, pcmk_action_detect_loop);
crm_trace("Adding probe restart orderings for '%s@%s then %s@%s'",
probe->uuid, pe__node_name(probe->node),
after->uuid, pe__node_name(after->node));
/* Add restart orderings if "then" is for a different primitive.
* Orderings for collective resources will be added later.
*/
if ((after->rsc != NULL)
&& (after->rsc->variant == pcmk_rsc_variant_primitive)
&& (probe->rsc != after->rsc)) {
GList *then_actions = NULL;
if (pcmk__str_eq(after->task, PCMK_ACTION_START, pcmk__str_none)) {
then_actions = pe__resource_actions(after->rsc, NULL,
PCMK_ACTION_STOP, FALSE);
} else if (pcmk__str_eq(after->task, PCMK_ACTION_PROMOTE,
pcmk__str_none)) {
then_actions = pe__resource_actions(after->rsc, NULL,
PCMK_ACTION_DEMOTE, FALSE);
}
for (iter = then_actions; iter != NULL; iter = iter->next) {
pe_action_t *then = (pe_action_t *) iter->data;
// Skip pseudo-actions (for example, those implied by fencing)
if (!pcmk_is_set(then->flags, pcmk_action_pseudo)) {
order_actions(probe, then, pcmk__ar_ordered);
}
}
g_list_free(then_actions);
}
/* Detect whether "then" is an interleaved clone action. For these, we want
* to add orderings only for the relevant instance.
*/
if ((after->rsc != NULL)
&& (after->rsc->variant > pcmk_rsc_variant_group)) {
const char *interleave_s = g_hash_table_lookup(after->rsc->meta,
XML_RSC_ATTR_INTERLEAVE);
interleave = crm_is_true(interleave_s);
if (interleave) {
compatible_rsc = pcmk__find_compatible_instance(probe->rsc,
after->rsc,
pcmk_role_unknown,
false);
}
}
/* Now recursively do the same for all actions ordered after "then". This
* also handles collective resources since the collective action will be
* ordered before its individual instances' actions.
*/
for (iter = after->actions_after; iter != NULL; iter = iter->next) {
pe_action_wrapper_t *after_wrapper = (pe_action_wrapper_t *) iter->data;
/* pcmk__ar_first_implies_then is the reason why a required A.start
* implies/enforces B.start to be required too, which is the cause of
* B.restart/re-promote.
*
* Not sure about pcmk__ar_first_implies_same_node_then though. It's now
* only used for unfencing case, which tends to introduce transition
* loops...
*/
if (!pcmk_is_set(after_wrapper->type, pcmk__ar_first_implies_then)) {
/* The order type between a group/clone and its child such as
* B.start-> B_child.start is:
* pcmk__ar_then_implies_first_graphed
* |pcmk__ar_unrunnable_first_blocks
*
* Proceed through the ordering chain and build dependencies with
* its children.
*/
if ((after->rsc == NULL)
|| (after->rsc->variant < pcmk_rsc_variant_group)
|| (probe->rsc->parent == after->rsc)
|| (after_wrapper->action->rsc == NULL)
|| (after_wrapper->action->rsc->variant > pcmk_rsc_variant_group)
|| (after->rsc != after_wrapper->action->rsc->parent)) {
continue;
}
/* Proceed to the children of a group or a non-interleaved clone.
* For an interleaved clone, proceed only to the relevant child.
*/
if ((after->rsc->variant > pcmk_rsc_variant_group) && interleave
&& ((compatible_rsc == NULL)
|| (compatible_rsc != after_wrapper->action->rsc))) {
continue;
}
}
crm_trace("Recursively adding probe restart orderings for "
"'%s@%s then %s@%s' (type=%#.6x)",
after->uuid, pe__node_name(after->node),
after_wrapper->action->uuid,
pe__node_name(after_wrapper->action->node),
after_wrapper->type);
add_restart_orderings_for_probe(probe, after_wrapper->action);
}
}
/*!
* \internal
* \brief Clear the tracking flag on all scheduled actions
*
* \param[in,out] data_set Cluster working set
*/
static void
clear_actions_tracking_flag(pe_working_set_t *data_set)
{
for (GList *iter = data_set->actions; iter != NULL; iter = iter->next) {
pe_action_t *action = iter->data;
pe__clear_action_flags(action, pcmk_action_detect_loop);
}
}
/*!
* \internal
* \brief Add start and restart orderings for probes scheduled for a resource
*
* \param[in,out] data Resource whose probes should be ordered
* \param[in] user_data Unused
*/
static void
add_start_restart_orderings_for_rsc(gpointer data, gpointer user_data)
{
pe_resource_t *rsc = data;
GList *probes = NULL;
// For collective resources, order each instance recursively
if (rsc->variant != pcmk_rsc_variant_primitive) {
g_list_foreach(rsc->children, add_start_restart_orderings_for_rsc,
NULL);
return;
}
// Find all probes for given resource
probes = pe__resource_actions(rsc, NULL, PCMK_ACTION_MONITOR, FALSE);
// Add probe restart orderings for each probe found
for (GList *iter = probes; iter != NULL; iter = iter->next) {
pe_action_t *probe = (pe_action_t *) iter->data;
for (GList *then_iter = probe->actions_after; then_iter != NULL;
then_iter = then_iter->next) {
pe_action_wrapper_t *then = (pe_action_wrapper_t *) then_iter->data;
add_start_orderings_for_probe(probe, then);
add_restart_orderings_for_probe(probe, then->action);
clear_actions_tracking_flag(rsc->cluster);
}
}
g_list_free(probes);
}
/*!
* \internal
* \brief Add "A then probe B" orderings for "A then B" orderings
*
* \param[in,out] data_set Cluster working set
*
* \note This function is currently disabled (see next comment).
*/
static void
order_then_probes(pe_working_set_t *data_set)
{
#if 0
/* Given an ordering "A then B", we would prefer to wait for A to be started
* before probing B.
*
* For example, if A is a filesystem which B can't even run without, it
* would be helpful 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 after A is running, otherwise we wouldn't
* detect the state of B if A could not be started. We can't even do an
* opportunistic version of this, because B may be moving:
*
* A.stop -> A.start -> B.probe -> B.stop -> B.start
*
* and if we add B.stop -> A.stop here, we get a loop:
*
* A.stop -> A.start -> B.probe -> B.stop -> A.stop
*
* We could kill the "B.probe -> B.stop" dependency, but that could mean
* stopping B "too" soon, because B.start must wait for the probe, and
* we don't want to stop B if we can't start it.
*
* We could add the ordering only if A is an anonymous 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 use case suggests that this code should remain disabled until
* someone gets smarter.
*/
for (GList *iter = data_set->resources; iter != NULL; iter = iter->next) {
pe_resource_t *rsc = (pe_resource_t *) iter->data;
pe_action_t *start = NULL;
GList *actions = NULL;
GList *probes = NULL;
actions = pe__resource_actions(rsc, NULL, PCMK_ACTION_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, PCMK_ACTION_MONITOR, FALSE);
for (actions = start->actions_before; actions != NULL;
actions = actions->next) {
pe_action_wrapper_t *before = (pe_action_wrapper_t *) actions->data;
pe_action_t *first = before->action;
pe_resource_t *first_rsc = first->rsc;
if (first->required_runnable_before) {
for (GList *clone_actions = first->actions_before;
clone_actions != NULL;
clone_actions = clone_actions->next) {
before = (pe_action_wrapper_t *) clone_actions->data;
crm_trace("Testing '%s then %s' for %s",
first->uuid, before->action->uuid, start->uuid);
CRM_ASSERT(before->action->rsc != NULL);
first_rsc = before->action->rsc;
break;
}
} else if (!pcmk__str_eq(first->task, PCMK_ACTION_START,
pcmk__str_none)) {
crm_trace("Not a start op %s for %s", first->uuid, start->uuid);
}
if (first_rsc == NULL) {
continue;
} else if (pe__const_top_resource(first_rsc, false)
== pe__const_top_resource(start->rsc, false)) {
crm_trace("Same parent %s for %s", first_rsc->id, start->uuid);
continue;
} else if (!pe_rsc_is_clone(pe__const_top_resource(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,
pe__const_top_resource(first_rsc, false)->variant);
for (GList *probe_iter = probes; probe_iter != NULL;
probe_iter = probe_iter->next) {
pe_action_t *probe = (pe_action_t *) probe_iter->data;
crm_err("Ordering %s before %s", first->uuid, probe->uuid);
order_actions(first, probe, pcmk__ar_ordered);
}
}
}
#endif
}
void
pcmk__order_probes(pe_working_set_t *data_set)
{
// Add orderings for "probe then X"
g_list_foreach(data_set->resources, add_start_restart_orderings_for_rsc,
NULL);
add_probe_orderings_for_stops(data_set);
order_then_probes(data_set);
}
/*!
* \internal
* \brief Schedule any probes needed
*
* \param[in,out] data_set Cluster working set
*
* \note This may also schedule fencing of failed remote nodes.
*/
void
pcmk__schedule_probes(pe_working_set_t *data_set)
{
// Schedule probes on each node in the cluster as needed
for (GList *iter = data_set->nodes; iter != NULL; iter = iter->next) {
pe_node_t *node = (pe_node_t *) iter->data;
const char *probed = NULL;
if (!node->details->online) { // Don't probe offline nodes
if (pcmk__is_failed_remote_node(node)) {
pe_fence_node(data_set, node,
"the connection is unrecoverable", FALSE);
}
continue;
} else if (node->details->unclean) { // ... or nodes that need fencing
continue;
} else if (!node->details->rsc_discovery_enabled) {
// The user requested that probes not be done on this node
continue;
}
/* This is no longer needed for live clusters, since the probe_complete
* node attribute will never be in the CIB. However this is still useful
* for processing old saved CIBs (< 1.1.14), including the
* reprobe-target_rc regression test.
*/
probed = pe_node_attribute_raw(node, CRM_OP_PROBED);
if (probed != NULL && crm_is_true(probed) == FALSE) {
pe_action_t *probe_op = NULL;
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;
}
// Probe each resource in the cluster on this node, as needed
pcmk__probe_resource_list(data_set->resources, node);
}
}
diff --git a/lib/pacemaker/pcmk_sched_utilization.c b/lib/pacemaker/pcmk_sched_utilization.c
index 787488399b..c20b66a6dc 100644
--- a/lib/pacemaker/pcmk_sched_utilization.c
+++ b/lib/pacemaker/pcmk_sched_utilization.c
@@ -1,465 +1,465 @@
/*
* Copyright 2014-2023 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 <crm/msg_xml.h>
#include <pacemaker-internal.h>
#include "libpacemaker_private.h"
/*!
* \internal
* \brief Get integer utilization from a string
*
* \param[in] s String representation of a node utilization value
*
* \return Integer equivalent of \p s
* \todo It would make sense to restrict utilization values to nonnegative
* integers, but the documentation just says "integers" and we didn't
* restrict them initially, so for backward compatibility, allow any
* integer.
*/
static int
utilization_value(const char *s)
{
int value = 0;
if ((s != NULL) && (pcmk__scan_min_int(s, &value, INT_MIN) == EINVAL)) {
pe_warn("Using 0 for utilization instead of invalid value '%s'", value);
value = 0;
}
return value;
}
/*
* Functions for comparing node capacities
*/
struct compare_data {
const pe_node_t *node1;
const pe_node_t *node2;
bool node2_only;
int result;
};
/*!
* \internal
* \brief Compare a single utilization attribute for two nodes
*
* Compare one utilization attribute for two nodes, incrementing the result if
* the first node has greater capacity, and decrementing it if the second node
* has greater capacity.
*
* \param[in] key Utilization attribute name to compare
* \param[in] value Utilization attribute value to compare
* \param[in,out] user_data Comparison data (as struct compare_data*)
*/
static void
compare_utilization_value(gpointer key, gpointer value, gpointer user_data)
{
int node1_capacity = 0;
int node2_capacity = 0;
struct compare_data *data = user_data;
const char *node2_value = NULL;
if (data->node2_only) {
if (g_hash_table_lookup(data->node1->details->utilization, key)) {
return; // We've already compared this attribute
}
} else {
node1_capacity = utilization_value((const char *) value);
}
node2_value = g_hash_table_lookup(data->node2->details->utilization, key);
node2_capacity = utilization_value(node2_value);
if (node1_capacity > node2_capacity) {
data->result--;
} else if (node1_capacity < node2_capacity) {
data->result++;
}
}
/*!
* \internal
* \brief Compare utilization capacities of two nodes
*
* \param[in] node1 First node to compare
* \param[in] node2 Second node to compare
*
* \return Negative integer if node1 has more free capacity,
* 0 if the capacities are equal, or a positive integer
* if node2 has more free capacity
*/
int
pcmk__compare_node_capacities(const pe_node_t *node1, const pe_node_t *node2)
{
struct compare_data data = {
.node1 = node1,
.node2 = node2,
.node2_only = false,
.result = 0,
};
// Compare utilization values that node1 and maybe node2 have
g_hash_table_foreach(node1->details->utilization, compare_utilization_value,
&data);
// Compare utilization values that only node2 has
data.node2_only = true;
g_hash_table_foreach(node2->details->utilization, compare_utilization_value,
&data);
return data.result;
}
/*
* Functions for updating node capacities
*/
struct calculate_data {
GHashTable *current_utilization;
bool plus;
};
/*!
* \internal
* \brief Update a single utilization attribute with a new value
*
* \param[in] key Name of utilization attribute to update
* \param[in] value Value to add or substract
* \param[in,out] user_data Calculation data (as struct calculate_data *)
*/
static void
update_utilization_value(gpointer key, gpointer value, gpointer user_data)
{
int result = 0;
const char *current = NULL;
struct calculate_data *data = user_data;
current = g_hash_table_lookup(data->current_utilization, key);
if (data->plus) {
result = utilization_value(current) + utilization_value(value);
} else if (current) {
result = utilization_value(current) - utilization_value(value);
}
g_hash_table_replace(data->current_utilization,
strdup(key), pcmk__itoa(result));
}
/*!
* \internal
* \brief Subtract a resource's utilization from node capacity
*
* \param[in,out] current_utilization Current node utilization attributes
* \param[in] rsc Resource with utilization to subtract
*/
void
pcmk__consume_node_capacity(GHashTable *current_utilization,
const pe_resource_t *rsc)
{
struct calculate_data data = {
.current_utilization = current_utilization,
.plus = false,
};
g_hash_table_foreach(rsc->utilization, update_utilization_value, &data);
}
/*!
* \internal
* \brief Add a resource's utilization to node capacity
*
* \param[in,out] current_utilization Current node utilization attributes
* \param[in] rsc Resource with utilization to add
*/
void
pcmk__release_node_capacity(GHashTable *current_utilization,
const pe_resource_t *rsc)
{
struct calculate_data data = {
.current_utilization = current_utilization,
.plus = true,
};
g_hash_table_foreach(rsc->utilization, update_utilization_value, &data);
}
/*
* Functions for checking for sufficient node capacity
*/
struct capacity_data {
const pe_node_t *node;
const char *rsc_id;
bool is_enough;
};
/*!
* \internal
* \brief Check whether a single utilization attribute has sufficient capacity
*
* \param[in] key Name of utilization attribute to check
* \param[in] value Amount of utilization required
* \param[in,out] user_data Capacity data (as struct capacity_data *)
*/
static void
check_capacity(gpointer key, gpointer value, gpointer user_data)
{
int required = 0;
int remaining = 0;
const char *node_value_s = NULL;
struct capacity_data *data = user_data;
node_value_s = g_hash_table_lookup(data->node->details->utilization, key);
required = utilization_value(value);
remaining = utilization_value(node_value_s);
if (required > remaining) {
crm_debug("Remaining capacity for %s on %s (%d) is insufficient "
"for resource %s usage (%d)",
(const char *) key, pe__node_name(data->node), remaining,
data->rsc_id, required);
data->is_enough = false;
}
}
/*!
* \internal
* \brief Check whether a node has sufficient capacity for a resource
*
* \param[in] node Node to check
* \param[in] rsc_id ID of resource to check (for debug logs only)
* \param[in] utilization Required utilization amounts
*
* \return true if node has sufficient capacity for resource, otherwise false
*/
static bool
have_enough_capacity(const pe_node_t *node, const char *rsc_id,
GHashTable *utilization)
{
struct capacity_data data = {
.node = node,
.rsc_id = rsc_id,
.is_enough = true,
};
g_hash_table_foreach(utilization, check_capacity, &data);
return data.is_enough;
}
/*!
* \internal
* \brief Sum the utilization requirements of a list of resources
*
* \param[in] orig_rsc Resource being assigned (for logging purposes)
* \param[in] rscs Resources whose utilization should be summed
*
* \return Newly allocated hash table with sum of all utilization values
* \note It is the caller's responsibility to free the return value using
* g_hash_table_destroy().
*/
static GHashTable *
sum_resource_utilization(const pe_resource_t *orig_rsc, GList *rscs)
{
GHashTable *utilization = pcmk__strkey_table(free, free);
for (GList *iter = rscs; iter != NULL; iter = iter->next) {
pe_resource_t *rsc = (pe_resource_t *) iter->data;
rsc->cmds->add_utilization(rsc, orig_rsc, rscs, utilization);
}
return utilization;
}
/*!
* \internal
* \brief Ban resource from nodes with insufficient utilization capacity
*
* \param[in,out] rsc Resource to check
*
* \return Allowed node for \p rsc with most spare capacity, if there are no
* nodes with enough capacity for \p rsc and all its colocated resources
*/
const pe_node_t *
pcmk__ban_insufficient_capacity(pe_resource_t *rsc)
{
bool any_capable = false;
char *rscs_id = NULL;
pe_node_t *node = NULL;
const pe_node_t *most_capable_node = NULL;
GList *colocated_rscs = NULL;
GHashTable *unassigned_utilization = NULL;
GHashTableIter iter;
CRM_CHECK(rsc != NULL, return NULL);
// The default placement strategy ignores utilization
if (pcmk__str_eq(rsc->cluster->placement_strategy, "default",
pcmk__str_casei)) {
return NULL;
}
// Check whether any resources are colocated with this one
colocated_rscs = rsc->cmds->colocated_resources(rsc, NULL, NULL);
if (colocated_rscs == NULL) {
return NULL;
}
rscs_id = crm_strdup_printf("%s and its colocated resources", rsc->id);
// If rsc isn't in the list, add it so we include its utilization
if (g_list_find(colocated_rscs, rsc) == NULL) {
colocated_rscs = g_list_append(colocated_rscs, rsc);
}
// Sum utilization of colocated resources that haven't been assigned yet
unassigned_utilization = sum_resource_utilization(rsc, colocated_rscs);
// Check whether any node has enough capacity for all the resources
g_hash_table_iter_init(&iter, rsc->allowed_nodes);
while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) {
if (!pcmk__node_available(node, true, false)) {
continue;
}
if (have_enough_capacity(node, rscs_id, unassigned_utilization)) {
any_capable = true;
}
// Keep track of node with most free capacity
if ((most_capable_node == NULL)
|| (pcmk__compare_node_capacities(node, most_capable_node) < 0)) {
most_capable_node = node;
}
}
if (any_capable) {
// If so, ban resource from any node with insufficient capacity
g_hash_table_iter_init(&iter, rsc->allowed_nodes);
while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) {
if (pcmk__node_available(node, true, false)
&& !have_enough_capacity(node, rscs_id,
unassigned_utilization)) {
pe_rsc_debug(rsc, "%s does not have enough capacity for %s",
pe__node_name(node), rscs_id);
resource_location(rsc, node, -INFINITY, "__limit_utilization__",
rsc->cluster);
}
}
most_capable_node = NULL;
} else {
// Otherwise, ban from nodes with insufficient capacity for rsc alone
g_hash_table_iter_init(&iter, rsc->allowed_nodes);
while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) {
if (pcmk__node_available(node, true, false)
&& !have_enough_capacity(node, rsc->id, rsc->utilization)) {
pe_rsc_debug(rsc, "%s does not have enough capacity for %s",
pe__node_name(node), rsc->id);
resource_location(rsc, node, -INFINITY, "__limit_utilization__",
rsc->cluster);
}
}
}
g_hash_table_destroy(unassigned_utilization);
g_list_free(colocated_rscs);
free(rscs_id);
pe__show_node_scores(true, rsc, "Post-utilization", rsc->allowed_nodes,
rsc->cluster);
return most_capable_node;
}
/*!
* \internal
* \brief Create a new load_stopped pseudo-op for a node
*
* \param[in,out] node Node to create op for
*
* \return Newly created load_stopped op
*/
static pe_action_t *
new_load_stopped_op(pe_node_t *node)
{
char *load_stopped_task = crm_strdup_printf(PCMK_ACTION_LOAD_STOPPED "_%s",
node->details->uname);
pe_action_t *load_stopped = get_pseudo_op(load_stopped_task,
node->details->data_set);
if (load_stopped->node == NULL) {
load_stopped->node = pe__copy_node(node);
pe__clear_action_flags(load_stopped, pcmk_action_optional);
}
free(load_stopped_task);
return load_stopped;
}
/*!
* \internal
* \brief Create utilization-related internal constraints for a resource
*
* \param[in,out] rsc Resource to create constraints for
* \param[in] allowed_nodes List of allowed next nodes for \p rsc
*/
void
pcmk__create_utilization_constraints(pe_resource_t *rsc,
const GList *allowed_nodes)
{
const GList *iter = NULL;
pe_action_t *load_stopped = NULL;
pe_rsc_trace(rsc, "Creating utilization constraints for %s - strategy: %s",
rsc->id, rsc->cluster->placement_strategy);
// "stop rsc then load_stopped" constraints for current nodes
for (iter = rsc->running_on; iter != NULL; iter = iter->next) {
load_stopped = new_load_stopped_op(iter->data);
pcmk__new_ordering(rsc, stop_key(rsc), NULL, NULL, NULL, load_stopped,
- pe_order_load, rsc->cluster);
+ pcmk__ar_if_on_same_node_or_target, rsc->cluster);
}
// "load_stopped then start/migrate_to rsc" constraints for allowed nodes
for (iter = allowed_nodes; iter; iter = iter->next) {
load_stopped = new_load_stopped_op(iter->data);
pcmk__new_ordering(NULL, NULL, load_stopped, rsc, start_key(rsc), NULL,
- pe_order_load, rsc->cluster);
+ pcmk__ar_if_on_same_node_or_target, rsc->cluster);
pcmk__new_ordering(NULL, NULL, load_stopped,
rsc,
pcmk__op_key(rsc->id, PCMK_ACTION_MIGRATE_TO, 0),
NULL,
- pe_order_load, rsc->cluster);
+ pcmk__ar_if_on_same_node_or_target, rsc->cluster);
}
}
/*!
* \internal
* \brief Output node capacities if enabled
*
* \param[in] desc Prefix for output
* \param[in,out] data_set Cluster working set
*/
void
pcmk__show_node_capacities(const char *desc, pe_working_set_t *data_set)
{
if (!pcmk_is_set(data_set->flags, pcmk_sched_show_utilization)) {
return;
}
for (const GList *iter = data_set->nodes; iter != NULL; iter = iter->next) {
const pe_node_t *node = (const pe_node_t *) iter->data;
pcmk__output_t *out = data_set->priv;
out->message(out, "node-capacity", node, desc);
}
}
diff --git a/lib/pacemaker/pcmk_simulate.c b/lib/pacemaker/pcmk_simulate.c
index b403ad977a..3709d530d8 100644
--- a/lib/pacemaker/pcmk_simulate.c
+++ b/lib/pacemaker/pcmk_simulate.c
@@ -1,1006 +1,1006 @@
/*
* Copyright 2021-2023 the Pacemaker project contributors
*
* The version control history for this file may have further details.
*
* This source code is licensed under the GNU Lesser General Public License
* version 2.1 or later (LGPLv2.1+) WITHOUT ANY WARRANTY.
*/
#include <crm_internal.h>
#include <crm/cib/internal.h>
#include <crm/common/output.h>
#include <crm/common/results.h>
#include <crm/pengine/pe_types.h>
#include <pacemaker-internal.h>
#include <pacemaker.h>
#include <stdint.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <unistd.h>
#include "libpacemaker_private.h"
static pcmk__output_t *out = NULL;
static cib_t *fake_cib = NULL;
static GList *fake_resource_list = NULL;
static const GList *fake_op_fail_list = NULL;
static void set_effective_date(pe_working_set_t *data_set, bool print_original,
const char *use_date);
/*!
* \internal
* \brief Create an action name for use in a dot graph
*
* \param[in] action Action to create name for
* \param[in] verbose If true, add action ID to name
*
* \return Newly allocated string with action name
* \note It is the caller's responsibility to free the result.
*/
static char *
create_action_name(const pe_action_t *action, bool verbose)
{
char *action_name = NULL;
const char *prefix = "";
const char *action_host = NULL;
const char *clone_name = NULL;
const char *task = action->task;
if (action->node != NULL) {
action_host = action->node->details->uname;
} else if (!pcmk_is_set(action->flags, pcmk_action_pseudo)) {
action_host = "<none>";
}
if (pcmk__str_eq(action->task, PCMK_ACTION_CANCEL, pcmk__str_none)) {
prefix = "Cancel ";
task = action->cancel_task;
}
if (action->rsc != NULL) {
clone_name = action->rsc->clone_name;
}
if (clone_name != NULL) {
char *key = NULL;
guint interval_ms = 0;
if (pcmk__guint_from_hash(action->meta,
XML_LRM_ATTR_INTERVAL_MS, 0,
&interval_ms) != pcmk_rc_ok) {
interval_ms = 0;
}
if (pcmk__strcase_any_of(action->task, PCMK_ACTION_NOTIFY,
PCMK_ACTION_NOTIFIED, NULL)) {
const char *n_type = g_hash_table_lookup(action->meta,
"notify_key_type");
const char *n_task = g_hash_table_lookup(action->meta,
"notify_key_operation");
CRM_ASSERT(n_type != NULL);
CRM_ASSERT(n_task != NULL);
key = pcmk__notify_key(clone_name, n_type, n_task);
} else {
key = pcmk__op_key(clone_name, task, interval_ms);
}
if (action_host != NULL) {
action_name = crm_strdup_printf("%s%s %s",
prefix, key, action_host);
} else {
action_name = crm_strdup_printf("%s%s", prefix, key);
}
free(key);
} else if (pcmk__str_eq(action->task, PCMK_ACTION_STONITH,
pcmk__str_none)) {
const char *op = g_hash_table_lookup(action->meta, "stonith_action");
action_name = crm_strdup_printf("%s%s '%s' %s",
prefix, action->task, op, action_host);
} else if (action->rsc && action_host) {
action_name = crm_strdup_printf("%s%s %s",
prefix, action->uuid, action_host);
} else if (action_host) {
action_name = crm_strdup_printf("%s%s %s",
prefix, action->task, action_host);
} else {
action_name = crm_strdup_printf("%s", action->uuid);
}
if (verbose) {
char *with_id = crm_strdup_printf("%s (%d)", action_name, action->id);
free(action_name);
action_name = with_id;
}
return action_name;
}
/*!
* \internal
* \brief Display the status of a cluster
*
* \param[in,out] data_set Cluster working set
* \param[in] show_opts How to modify display (as pcmk_show_opt_e flags)
* \param[in] section_opts Sections to display (as pcmk_section_e flags)
* \param[in] title What to use as list title
* \param[in] print_spacer Whether to display a spacer first
*/
static void
print_cluster_status(pe_working_set_t *data_set, uint32_t show_opts,
uint32_t section_opts, const char *title,
bool print_spacer)
{
pcmk__output_t *out = data_set->priv;
GList *all = NULL;
crm_exit_t stonith_rc = 0;
enum pcmk_pacemakerd_state state = pcmk_pacemakerd_state_invalid;
section_opts |= pcmk_section_nodes | pcmk_section_resources;
show_opts |= pcmk_show_inactive_rscs | pcmk_show_failed_detail;
all = g_list_prepend(all, (gpointer) "*");
PCMK__OUTPUT_SPACER_IF(out, print_spacer);
out->begin_list(out, NULL, NULL, "%s", title);
out->message(out, "cluster-status",
data_set, state, stonith_rc, NULL,
false, section_opts, show_opts, NULL, all, all);
out->end_list(out);
g_list_free(all);
}
/*!
* \internal
* \brief Display a summary of all actions scheduled in a transition
*
* \param[in,out] data_set Cluster working set (fully scheduled)
* \param[in] print_spacer Whether to display a spacer first
*/
static void
print_transition_summary(pe_working_set_t *data_set, bool print_spacer)
{
pcmk__output_t *out = data_set->priv;
PCMK__OUTPUT_SPACER_IF(out, print_spacer);
out->begin_list(out, NULL, NULL, "Transition Summary");
pcmk__output_actions(data_set);
out->end_list(out);
}
/*!
* \internal
* \brief Reset a cluster working set's input, output, date, and flags
*
* \param[in,out] data_set Cluster working set
* \param[in] input What to set as cluster input
* \param[in] out What to set as cluster output object
* \param[in] use_date What to set as cluster's current timestamp
* \param[in] flags Group of enum pcmk_scheduler_flags to set
*/
static void
reset(pe_working_set_t *data_set, xmlNodePtr input, pcmk__output_t *out,
const char *use_date, unsigned int flags)
{
data_set->input = input;
data_set->priv = out;
set_effective_date(data_set, true, use_date);
if (pcmk_is_set(flags, pcmk_sim_sanitized)) {
pe__set_working_set_flags(data_set, pcmk_sched_sanitized);
}
if (pcmk_is_set(flags, pcmk_sim_show_scores)) {
pe__set_working_set_flags(data_set, pcmk_sched_output_scores);
}
if (pcmk_is_set(flags, pcmk_sim_show_utilization)) {
pe__set_working_set_flags(data_set, pcmk_sched_show_utilization);
}
}
/*!
* \brief Write out a file in dot(1) format describing the actions that will
* be taken by the scheduler in response to an input CIB file.
*
* \param[in,out] data_set Working set for the cluster
* \param[in] dot_file The filename to write
* \param[in] all_actions Write all actions, even those that are optional
* or are on unmanaged resources
* \param[in] verbose Add extra information, such as action IDs, to the
* output
*
* \return Standard Pacemaker return code
*/
static int
write_sim_dotfile(pe_working_set_t *data_set, const char *dot_file,
bool all_actions, bool verbose)
{
GList *iter = NULL;
FILE *dot_strm = fopen(dot_file, "w");
if (dot_strm == NULL) {
return errno;
}
fprintf(dot_strm, " digraph \"g\" {\n");
for (iter = data_set->actions; iter != NULL; iter = iter->next) {
pe_action_t *action = (pe_action_t *) iter->data;
const char *style = "dashed";
const char *font = "black";
const char *color = "black";
char *action_name = create_action_name(action, verbose);
if (pcmk_is_set(action->flags, pcmk_action_pseudo)) {
font = "orange";
}
if (pcmk_is_set(action->flags, pcmk_action_added_to_graph)) {
style = "bold";
color = "green";
} else if ((action->rsc != NULL)
&& !pcmk_is_set(action->rsc->flags, pcmk_rsc_managed)) {
color = "red";
font = "purple";
if (!all_actions) {
goto do_not_write;
}
} else if (pcmk_is_set(action->flags, pcmk_action_optional)) {
color = "blue";
if (!all_actions) {
goto do_not_write;
}
} else {
color = "red";
CRM_LOG_ASSERT(!pcmk_is_set(action->flags, pcmk_action_runnable));
}
pe__set_action_flags(action, pcmk_action_added_to_graph);
fprintf(dot_strm, "\"%s\" [ style=%s color=\"%s\" fontcolor=\"%s\"]\n",
action_name, style, color, font);
do_not_write:
free(action_name);
}
for (iter = data_set->actions; iter != NULL; iter = iter->next) {
pe_action_t *action = (pe_action_t *) iter->data;
for (GList *before_iter = action->actions_before;
before_iter != NULL; before_iter = before_iter->next) {
pe_action_wrapper_t *before = before_iter->data;
char *before_name = NULL;
char *after_name = NULL;
const char *style = "dashed";
bool optional = true;
if (before->state == pe_link_dumped) {
optional = false;
style = "bold";
} else if ((uint32_t) before->type == pcmk__ar_none) {
continue;
} else if (pcmk_is_set(before->action->flags,
pcmk_action_added_to_graph)
&& pcmk_is_set(action->flags, pcmk_action_added_to_graph)
- && before->type != pe_order_load) {
+ && (uint32_t) before->type != pcmk__ar_if_on_same_node_or_target) {
optional = false;
}
if (all_actions || !optional) {
before_name = create_action_name(before->action, verbose);
after_name = create_action_name(action, verbose);
fprintf(dot_strm, "\"%s\" -> \"%s\" [ style = %s]\n",
before_name, after_name, style);
free(before_name);
free(after_name);
}
}
}
fprintf(dot_strm, "}\n");
fflush(dot_strm);
fclose(dot_strm);
return pcmk_rc_ok;
}
/*!
* \brief Profile the configuration updates and scheduler actions in a single
* CIB file, printing the profiling timings.
*
* \note \p data_set->priv must have been set to a valid \p pcmk__output_t
* object before this function is called.
*
* \param[in] xml_file The CIB file to profile
* \param[in] repeat Number of times to run
* \param[in,out] data_set Working set for the cluster
* \param[in] use_date The date to set the cluster's time to (may be NULL)
*/
static void
profile_file(const char *xml_file, long long repeat, pe_working_set_t *data_set,
const char *use_date)
{
pcmk__output_t *out = data_set->priv;
xmlNode *cib_object = NULL;
clock_t start = 0;
clock_t end;
unsigned long long data_set_flags = pcmk_sched_no_compat;
CRM_ASSERT(out != NULL);
cib_object = filename2xml(xml_file);
start = clock();
if (pcmk_find_cib_element(cib_object, XML_CIB_TAG_STATUS) == NULL) {
create_xml_node(cib_object, XML_CIB_TAG_STATUS);
}
if (cli_config_update(&cib_object, NULL, FALSE) == FALSE) {
free_xml(cib_object);
return;
}
if (validate_xml(cib_object, NULL, FALSE) != TRUE) {
free_xml(cib_object);
return;
}
if (pcmk_is_set(data_set->flags, pcmk_sched_output_scores)) {
data_set_flags |= pcmk_sched_output_scores;
}
if (pcmk_is_set(data_set->flags, pcmk_sched_show_utilization)) {
data_set_flags |= pcmk_sched_show_utilization;
}
for (int i = 0; i < repeat; ++i) {
xmlNode *input = (repeat == 1)? cib_object : copy_xml(cib_object);
data_set->input = input;
set_effective_date(data_set, false, use_date);
pcmk__schedule_actions(input, data_set_flags, data_set);
pe_reset_working_set(data_set);
}
end = clock();
out->message(out, "profile", xml_file, start, end);
}
void
pcmk__profile_dir(const char *dir, long long repeat, pe_working_set_t *data_set,
const char *use_date)
{
pcmk__output_t *out = data_set->priv;
struct dirent **namelist;
int file_num = scandir(dir, &namelist, 0, alphasort);
CRM_ASSERT(out != NULL);
if (file_num > 0) {
struct stat prop;
char buffer[FILENAME_MAX];
out->begin_list(out, NULL, NULL, "Timings");
while (file_num--) {
if ('.' == namelist[file_num]->d_name[0]) {
free(namelist[file_num]);
continue;
} else if (!pcmk__ends_with_ext(namelist[file_num]->d_name,
".xml")) {
free(namelist[file_num]);
continue;
}
snprintf(buffer, sizeof(buffer), "%s/%s",
dir, namelist[file_num]->d_name);
if (stat(buffer, &prop) == 0 && S_ISREG(prop.st_mode)) {
profile_file(buffer, repeat, data_set, use_date);
}
free(namelist[file_num]);
}
free(namelist);
out->end_list(out);
}
}
/*!
* \brief Set the date of the cluster, either to the value given by
* \p use_date, or to the "execution-date" value in the CIB.
*
* \note \p data_set->priv must have been set to a valid \p pcmk__output_t
* object before this function is called.
*
* \param[in,out] data_set Working set for the cluster
* \param[in] print_original If \p true, the "execution-date" should
* also be printed
* \param[in] use_date The date to set the cluster's time to
* (may be NULL)
*/
static void
set_effective_date(pe_working_set_t *data_set, bool print_original,
const char *use_date)
{
pcmk__output_t *out = data_set->priv;
time_t original_date = 0;
CRM_ASSERT(out != NULL);
crm_element_value_epoch(data_set->input, "execution-date", &original_date);
if (use_date) {
data_set->now = crm_time_new(use_date);
out->info(out, "Setting effective cluster time: %s", use_date);
crm_time_log(LOG_NOTICE, "Pretending 'now' is", data_set->now,
crm_time_log_date | crm_time_log_timeofday);
} else if (original_date != 0) {
data_set->now = pcmk__copy_timet(original_date);
if (print_original) {
char *when = crm_time_as_string(data_set->now,
crm_time_log_date|crm_time_log_timeofday);
out->info(out, "Using the original execution date of: %s", when);
free(when);
}
}
}
/*!
* \internal
* \brief Simulate successfully executing a pseudo-action in a graph
*
* \param[in,out] graph Graph to update with pseudo-action result
* \param[in,out] action Pseudo-action to simulate executing
*
* \return Standard Pacemaker return code
*/
static int
simulate_pseudo_action(pcmk__graph_t *graph, pcmk__graph_action_t *action)
{
const char *node = crm_element_value(action->xml, XML_LRM_ATTR_TARGET);
const char *task = crm_element_value(action->xml, XML_LRM_ATTR_TASK_KEY);
pcmk__set_graph_action_flags(action, pcmk__graph_action_confirmed);
out->message(out, "inject-pseudo-action", node, task);
pcmk__update_graph(graph, action);
return pcmk_rc_ok;
}
/*!
* \internal
* \brief Simulate executing a resource action in a graph
*
* \param[in,out] graph Graph to update with resource action result
* \param[in,out] action Resource action to simulate executing
*
* \return Standard Pacemaker return code
*/
static int
simulate_resource_action(pcmk__graph_t *graph, pcmk__graph_action_t *action)
{
int rc;
lrmd_event_data_t *op = NULL;
int target_outcome = PCMK_OCF_OK;
const char *rtype = NULL;
const char *rclass = NULL;
const char *resource = NULL;
const char *rprovider = NULL;
const char *resource_config_name = NULL;
const char *operation = crm_element_value(action->xml, "operation");
const char *target_rc_s = crm_meta_value(action->params,
XML_ATTR_TE_TARGET_RC);
xmlNode *cib_node = NULL;
xmlNode *cib_resource = NULL;
xmlNode *action_rsc = first_named_child(action->xml, XML_CIB_TAG_RESOURCE);
char *node = crm_element_value_copy(action->xml, XML_LRM_ATTR_TARGET);
char *uuid = NULL;
const char *router_node = crm_element_value(action->xml,
XML_LRM_ATTR_ROUTER_NODE);
// Certain actions don't need to be displayed or history entries
if (pcmk__str_eq(operation, CRM_OP_REPROBE, pcmk__str_none)) {
crm_debug("No history injection for %s op on %s", operation, node);
goto done; // Confirm action and update graph
}
if (action_rsc == NULL) { // Shouldn't be possible
crm_log_xml_err(action->xml, "Bad");
free(node);
return EPROTO;
}
/* A resource might be known by different names in the configuration and in
* the action (for example, a clone instance). Grab the configuration name
* (which is preferred when writing history), and if necessary, the instance
* name.
*/
resource_config_name = crm_element_value(action_rsc, XML_ATTR_ID);
if (resource_config_name == NULL) { // Shouldn't be possible
crm_log_xml_err(action->xml, "No ID");
free(node);
return EPROTO;
}
resource = resource_config_name;
if (pe_find_resource(fake_resource_list, resource) == NULL) {
const char *longname = crm_element_value(action_rsc, XML_ATTR_ID_LONG);
if ((longname != NULL)
&& (pe_find_resource(fake_resource_list, longname) != NULL)) {
resource = longname;
}
}
// Certain actions need to be displayed but don't need history entries
if (pcmk__strcase_any_of(operation, PCMK_ACTION_DELETE,
PCMK_ACTION_META_DATA, NULL)) {
out->message(out, "inject-rsc-action", resource, operation, node,
(guint) 0);
goto done; // Confirm action and update graph
}
rclass = crm_element_value(action_rsc, XML_AGENT_ATTR_CLASS);
rtype = crm_element_value(action_rsc, XML_ATTR_TYPE);
rprovider = crm_element_value(action_rsc, XML_AGENT_ATTR_PROVIDER);
pcmk__scan_min_int(target_rc_s, &target_outcome, 0);
CRM_ASSERT(fake_cib->cmds->query(fake_cib, NULL, NULL,
cib_sync_call|cib_scope_local) == pcmk_ok);
// Ensure the action node is in the CIB
uuid = crm_element_value_copy(action->xml, XML_LRM_ATTR_TARGET_UUID);
cib_node = pcmk__inject_node(fake_cib, node,
((router_node == NULL)? uuid: node));
free(uuid);
CRM_ASSERT(cib_node != NULL);
// Add a history entry for the action
cib_resource = pcmk__inject_resource_history(out, cib_node, resource,
resource_config_name,
rclass, rtype, rprovider);
if (cib_resource == NULL) {
crm_err("Could not simulate action %d history for resource %s",
action->id, resource);
free(node);
free_xml(cib_node);
return EINVAL;
}
// Simulate and display an executor event for the action result
op = pcmk__event_from_graph_action(cib_resource, action, PCMK_EXEC_DONE,
target_outcome, "User-injected result");
out->message(out, "inject-rsc-action", resource, op->op_type, node,
op->interval_ms);
// Check whether action is in a list of desired simulated failures
for (const GList *iter = fake_op_fail_list;
iter != NULL; iter = iter->next) {
const char *spec = (const char *) iter->data;
char *key = NULL;
const char *match_name = NULL;
// Allow user to specify anonymous clone with or without instance number
key = crm_strdup_printf(PCMK__OP_FMT "@%s=", resource, op->op_type,
op->interval_ms, node);
if (strncasecmp(key, spec, strlen(key)) == 0) {
match_name = resource;
}
free(key);
// If not found, try the resource's name in the configuration
if ((match_name == NULL)
&& (strcmp(resource, resource_config_name) != 0)) {
key = crm_strdup_printf(PCMK__OP_FMT "@%s=", resource_config_name,
op->op_type, op->interval_ms, node);
if (strncasecmp(key, spec, strlen(key)) == 0) {
match_name = resource_config_name;
}
free(key);
}
if (match_name == NULL) {
continue; // This failed action entry doesn't match
}
// ${match_name}_${task}_${interval_in_ms}@${node}=${rc}
rc = sscanf(spec, "%*[^=]=%d", (int *) &op->rc);
if (rc != 1) {
out->err(out, "Invalid failed operation '%s' "
"(result code must be integer)", spec);
continue; // Keep checking other list entries
}
out->info(out, "Pretending action %d failed with rc=%d",
action->id, op->rc);
pcmk__set_graph_action_flags(action, pcmk__graph_action_failed);
graph->abort_priority = INFINITY;
pcmk__inject_failcount(out, cib_node, match_name, op->op_type,
op->interval_ms, op->rc);
break;
}
pcmk__inject_action_result(cib_resource, op, target_outcome);
lrmd_free_event(op);
rc = fake_cib->cmds->modify(fake_cib, XML_CIB_TAG_STATUS, cib_node,
cib_sync_call|cib_scope_local);
CRM_ASSERT(rc == pcmk_ok);
done:
free(node);
free_xml(cib_node);
pcmk__set_graph_action_flags(action, pcmk__graph_action_confirmed);
pcmk__update_graph(graph, action);
return pcmk_rc_ok;
}
/*!
* \internal
* \brief Simulate successfully executing a cluster action
*
* \param[in,out] graph Graph to update with action result
* \param[in,out] action Cluster action to simulate
*
* \return Standard Pacemaker return code
*/
static int
simulate_cluster_action(pcmk__graph_t *graph, pcmk__graph_action_t *action)
{
const char *node = crm_element_value(action->xml, XML_LRM_ATTR_TARGET);
const char *task = crm_element_value(action->xml, XML_LRM_ATTR_TASK);
xmlNode *rsc = first_named_child(action->xml, XML_CIB_TAG_RESOURCE);
pcmk__set_graph_action_flags(action, pcmk__graph_action_confirmed);
out->message(out, "inject-cluster-action", node, task, rsc);
pcmk__update_graph(graph, action);
return pcmk_rc_ok;
}
/*!
* \internal
* \brief Simulate successfully executing a fencing action
*
* \param[in,out] graph Graph to update with action result
* \param[in,out] action Fencing action to simulate
*
* \return Standard Pacemaker return code
*/
static int
simulate_fencing_action(pcmk__graph_t *graph, pcmk__graph_action_t *action)
{
const char *op = crm_meta_value(action->params, "stonith_action");
char *target = crm_element_value_copy(action->xml, XML_LRM_ATTR_TARGET);
out->message(out, "inject-fencing-action", target, op);
if (!pcmk__str_eq(op, PCMK_ACTION_ON, pcmk__str_casei)) {
int rc = pcmk_ok;
GString *xpath = g_string_sized_new(512);
// Set node state to offline
xmlNode *cib_node = pcmk__inject_node_state_change(fake_cib, target,
false);
CRM_ASSERT(cib_node != NULL);
crm_xml_add(cib_node, XML_ATTR_ORIGIN, __func__);
rc = fake_cib->cmds->replace(fake_cib, XML_CIB_TAG_STATUS, cib_node,
cib_sync_call|cib_scope_local);
CRM_ASSERT(rc == pcmk_ok);
// Simulate controller clearing node's resource history and attributes
pcmk__g_strcat(xpath,
"//" XML_CIB_TAG_STATE
"[@" XML_ATTR_UNAME "='", target, "']/" XML_CIB_TAG_LRM,
NULL);
fake_cib->cmds->remove(fake_cib, (const char *) xpath->str, NULL,
cib_xpath|cib_sync_call|cib_scope_local);
g_string_truncate(xpath, 0);
pcmk__g_strcat(xpath,
"//" XML_CIB_TAG_STATE
"[@" XML_ATTR_UNAME "='", target, "']"
"/" XML_TAG_TRANSIENT_NODEATTRS, NULL);
fake_cib->cmds->remove(fake_cib, (const char *) xpath->str, NULL,
cib_xpath|cib_sync_call|cib_scope_local);
free_xml(cib_node);
g_string_free(xpath, TRUE);
}
pcmk__set_graph_action_flags(action, pcmk__graph_action_confirmed);
pcmk__update_graph(graph, action);
free(target);
return pcmk_rc_ok;
}
enum pcmk__graph_status
pcmk__simulate_transition(pe_working_set_t *data_set, cib_t *cib,
const GList *op_fail_list)
{
pcmk__graph_t *transition = NULL;
enum pcmk__graph_status graph_rc;
pcmk__graph_functions_t simulation_fns = {
simulate_pseudo_action,
simulate_resource_action,
simulate_cluster_action,
simulate_fencing_action,
};
out = data_set->priv;
fake_cib = cib;
fake_op_fail_list = op_fail_list;
if (!out->is_quiet(out)) {
out->begin_list(out, NULL, NULL, "Executing Cluster Transition");
}
pcmk__set_graph_functions(&simulation_fns);
transition = pcmk__unpack_graph(data_set->graph, crm_system_name);
pcmk__log_graph(LOG_DEBUG, transition);
fake_resource_list = data_set->resources;
do {
graph_rc = pcmk__execute_graph(transition);
} while (graph_rc == pcmk__graph_active);
fake_resource_list = NULL;
if (graph_rc != pcmk__graph_complete) {
out->err(out, "Transition failed: %s",
pcmk__graph_status2text(graph_rc));
pcmk__log_graph(LOG_ERR, transition);
out->err(out, "An invalid transition was produced");
}
pcmk__free_graph(transition);
if (!out->is_quiet(out)) {
// If not quiet, we'll need the resulting CIB for later display
xmlNode *cib_object = NULL;
int rc = fake_cib->cmds->query(fake_cib, NULL, &cib_object,
cib_sync_call|cib_scope_local);
CRM_ASSERT(rc == pcmk_ok);
pe_reset_working_set(data_set);
data_set->input = cib_object;
out->end_list(out);
}
return graph_rc;
}
int
pcmk__simulate(pe_working_set_t *data_set, pcmk__output_t *out,
const pcmk_injections_t *injections, unsigned int flags,
uint32_t section_opts, const char *use_date,
const char *input_file, const char *graph_file,
const char *dot_file)
{
int printed = pcmk_rc_no_output;
int rc = pcmk_rc_ok;
xmlNodePtr input = NULL;
cib_t *cib = NULL;
rc = cib__signon_query(out, &cib, &input);
if (rc != pcmk_rc_ok) {
goto simulate_done;
}
reset(data_set, input, out, use_date, flags);
cluster_status(data_set);
if ((cib->variant == cib_native)
&& pcmk_is_set(section_opts, pcmk_section_times)) {
if (pcmk__our_nodename == NULL) {
// Currently used only in the times section
pcmk__query_node_name(out, 0, &pcmk__our_nodename, 0);
}
data_set->localhost = pcmk__our_nodename;
}
if (!out->is_quiet(out)) {
const bool show_pending = pcmk_is_set(flags, pcmk_sim_show_pending);
if (pcmk_is_set(data_set->flags, pcmk_sched_in_maintenance)) {
printed = out->message(out, "maint-mode", data_set->flags);
}
if (data_set->disabled_resources || data_set->blocked_resources) {
PCMK__OUTPUT_SPACER_IF(out, printed == pcmk_rc_ok);
printed = out->info(out,
"%d of %d resource instances DISABLED and "
"%d BLOCKED from further action due to failure",
data_set->disabled_resources,
data_set->ninstances,
data_set->blocked_resources);
}
/* Most formatted output headers use caps for each word, but this one
* only has the first word capitalized for compatibility with pcs.
*/
print_cluster_status(data_set, (show_pending? pcmk_show_pending : 0),
section_opts, "Current cluster status",
(printed == pcmk_rc_ok));
printed = pcmk_rc_ok;
}
// If the user requested any injections, handle them
if ((injections->node_down != NULL)
|| (injections->node_fail != NULL)
|| (injections->node_up != NULL)
|| (injections->op_inject != NULL)
|| (injections->ticket_activate != NULL)
|| (injections->ticket_grant != NULL)
|| (injections->ticket_revoke != NULL)
|| (injections->ticket_standby != NULL)
|| (injections->watchdog != NULL)) {
PCMK__OUTPUT_SPACER_IF(out, printed == pcmk_rc_ok);
pcmk__inject_scheduler_input(data_set, cib, injections);
printed = pcmk_rc_ok;
rc = cib->cmds->query(cib, NULL, &input, cib_sync_call);
if (rc != pcmk_rc_ok) {
rc = pcmk_legacy2rc(rc);
goto simulate_done;
}
cleanup_calculations(data_set);
reset(data_set, input, out, use_date, flags);
cluster_status(data_set);
}
if (input_file != NULL) {
rc = write_xml_file(input, input_file, FALSE);
if (rc < 0) {
rc = pcmk_legacy2rc(rc);
goto simulate_done;
}
}
if (pcmk_any_flags_set(flags, pcmk_sim_process | pcmk_sim_simulate)) {
pcmk__output_t *logger_out = NULL;
unsigned long long data_set_flags = pcmk_sched_no_compat;
if (pcmk_is_set(data_set->flags, pcmk_sched_output_scores)) {
data_set_flags |= pcmk_sched_output_scores;
}
if (pcmk_is_set(data_set->flags, pcmk_sched_show_utilization)) {
data_set_flags |= pcmk_sched_show_utilization;
}
if (pcmk_all_flags_set(data_set->flags,
pcmk_sched_output_scores
|pcmk_sched_show_utilization)) {
PCMK__OUTPUT_SPACER_IF(out, printed == pcmk_rc_ok);
out->begin_list(out, NULL, NULL,
"Assignment Scores and Utilization Information");
printed = pcmk_rc_ok;
} else if (pcmk_is_set(data_set->flags, pcmk_sched_output_scores)) {
PCMK__OUTPUT_SPACER_IF(out, printed == pcmk_rc_ok);
out->begin_list(out, NULL, NULL, "Assignment Scores");
printed = pcmk_rc_ok;
} else if (pcmk_is_set(data_set->flags, pcmk_sched_show_utilization)) {
PCMK__OUTPUT_SPACER_IF(out, printed == pcmk_rc_ok);
out->begin_list(out, NULL, NULL, "Utilization Information");
printed = pcmk_rc_ok;
} else {
rc = pcmk__log_output_new(&logger_out);
if (rc != pcmk_rc_ok) {
goto simulate_done;
}
pe__register_messages(logger_out);
pcmk__register_lib_messages(logger_out);
data_set->priv = logger_out;
}
pcmk__schedule_actions(input, data_set_flags, data_set);
if (logger_out == NULL) {
out->end_list(out);
} else {
logger_out->finish(logger_out, CRM_EX_OK, true, NULL);
pcmk__output_free(logger_out);
data_set->priv = out;
}
input = NULL; /* Don't try and free it twice */
if (graph_file != NULL) {
rc = write_xml_file(data_set->graph, graph_file, FALSE);
if (rc < 0) {
rc = pcmk_rc_graph_error;
goto simulate_done;
}
}
if (dot_file != NULL) {
rc = write_sim_dotfile(data_set, dot_file,
pcmk_is_set(flags, pcmk_sim_all_actions),
pcmk_is_set(flags, pcmk_sim_verbose));
if (rc != pcmk_rc_ok) {
rc = pcmk_rc_dot_error;
goto simulate_done;
}
}
if (!out->is_quiet(out)) {
print_transition_summary(data_set, printed == pcmk_rc_ok);
}
}
rc = pcmk_rc_ok;
if (!pcmk_is_set(flags, pcmk_sim_simulate)) {
goto simulate_done;
}
PCMK__OUTPUT_SPACER_IF(out, printed == pcmk_rc_ok);
if (pcmk__simulate_transition(data_set, cib, injections->op_fail)
!= pcmk__graph_complete) {
rc = pcmk_rc_invalid_transition;
}
if (out->is_quiet(out)) {
goto simulate_done;
}
set_effective_date(data_set, true, use_date);
if (pcmk_is_set(flags, pcmk_sim_show_scores)) {
pe__set_working_set_flags(data_set, pcmk_sched_output_scores);
}
if (pcmk_is_set(flags, pcmk_sim_show_utilization)) {
pe__set_working_set_flags(data_set, pcmk_sched_show_utilization);
}
cluster_status(data_set);
print_cluster_status(data_set, 0, section_opts, "Revised Cluster Status",
true);
simulate_done:
cib__clean_up_connection(&cib);
return rc;
}
int
pcmk_simulate(xmlNodePtr *xml, pe_working_set_t *data_set,
const pcmk_injections_t *injections, unsigned int flags,
unsigned int section_opts, const char *use_date,
const char *input_file, const char *graph_file,
const char *dot_file)
{
pcmk__output_t *out = NULL;
int rc = pcmk_rc_ok;
rc = pcmk__xml_output_new(&out, xml);
if (rc != pcmk_rc_ok) {
return rc;
}
pe__register_messages(out);
pcmk__register_lib_messages(out);
rc = pcmk__simulate(data_set, out, injections, flags, section_opts,
use_date, input_file, graph_file, dot_file);
pcmk__xml_output_finish(out, xml);
return rc;
}

File Metadata

Mime Type
text/x-diff
Expires
Fri, Sep 5, 9:30 AM (12 h, 35 m)
Storage Engine
blob
Storage Format
Raw Data
Storage Handle
2308701
Default Alt Text
(183 KB)

Event Timeline