diff --git a/include/pcmki/pcmki_sched_allocate.h b/include/pcmki/pcmki_sched_allocate.h index 3033a5829f..e0791c7480 100644 --- a/include/pcmki/pcmki_sched_allocate.h +++ b/include/pcmki/pcmki_sched_allocate.h @@ -1,174 +1,171 @@ /* * Copyright 2004-2021 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 SCHED_ALLOCATE__H # define SCHED_ALLOCATE__H # include # include # include # include # include # include struct resource_alloc_functions_s { GHashTable *(*merge_weights) (pe_resource_t *, const char *, GHashTable *, const char *, float, enum pe_weights); pe_node_t *(*allocate) (pe_resource_t *, pe_node_t *, pe_working_set_t *); void (*create_actions) (pe_resource_t *, pe_working_set_t *); gboolean(*create_probe) (pe_resource_t *, pe_node_t *, pe_action_t *, gboolean, pe_working_set_t *); void (*internal_constraints) (pe_resource_t *, pe_working_set_t *); void (*rsc_colocation_lh) (pe_resource_t *, pe_resource_t *, pcmk__colocation_t *, pe_working_set_t *); void (*rsc_colocation_rh) (pe_resource_t *, pe_resource_t *, pcmk__colocation_t *, pe_working_set_t *); void (*rsc_location) (pe_resource_t *, pe__location_t *); enum pe_action_flags (*action_flags) (pe_action_t *, pe_node_t *); enum pe_graph_flags (*update_actions) (pe_action_t *, pe_action_t *, pe_node_t *, enum pe_action_flags, enum pe_action_flags, enum pe_ordering, pe_working_set_t *data_set); void (*expand) (pe_resource_t *, pe_working_set_t *); void (*append_meta) (pe_resource_t * rsc, xmlNode * xml); }; GHashTable *pcmk__native_merge_weights(pe_resource_t *rsc, const char *rhs, GHashTable *nodes, const char *attr, float factor, uint32_t flags); GHashTable *pcmk__group_merge_weights(pe_resource_t *rsc, const char *rhs, GHashTable *nodes, const char *attr, float factor, uint32_t flags); pe_node_t *pcmk__native_allocate(pe_resource_t *rsc, pe_node_t *preferred, pe_working_set_t *data_set); extern void native_create_actions(pe_resource_t * rsc, pe_working_set_t * data_set); extern void native_internal_constraints(pe_resource_t * rsc, pe_working_set_t * data_set); void native_rsc_colocation_lh(pe_resource_t *lh_rsc, pe_resource_t *rh_rsc, pcmk__colocation_t *constraint, pe_working_set_t *data_set); void native_rsc_colocation_rh(pe_resource_t *lh_rsc, pe_resource_t *rh_rsc, pcmk__colocation_t *constraint, pe_working_set_t *data_set); extern void rsc_ticket_constraint(pe_resource_t * lh_rsc, rsc_ticket_t * rsc_ticket, pe_working_set_t * data_set); extern enum pe_action_flags native_action_flags(pe_action_t * action, pe_node_t * node); void native_rsc_location(pe_resource_t *rsc, pe__location_t *constraint); extern void native_expand(pe_resource_t * rsc, pe_working_set_t * data_set); extern gboolean native_create_probe(pe_resource_t * rsc, pe_node_t * node, pe_action_t * complete, gboolean force, pe_working_set_t * data_set); extern void native_append_meta(pe_resource_t * rsc, xmlNode * xml); pe_node_t *pcmk__group_allocate(pe_resource_t *rsc, pe_node_t *preferred, pe_working_set_t *data_set); extern void group_create_actions(pe_resource_t * rsc, pe_working_set_t * data_set); extern void group_internal_constraints(pe_resource_t * rsc, pe_working_set_t * data_set); void group_rsc_colocation_lh(pe_resource_t *lh_rsc, pe_resource_t *rh_rsc, pcmk__colocation_t *constraint, pe_working_set_t *data_set); void group_rsc_colocation_rh(pe_resource_t *lh_rsc, pe_resource_t *rh_rsc, pcmk__colocation_t *constraint, pe_working_set_t *data_set); extern enum pe_action_flags group_action_flags(pe_action_t * action, pe_node_t * node); void group_rsc_location(pe_resource_t *rsc, pe__location_t *constraint); extern void group_expand(pe_resource_t * rsc, pe_working_set_t * data_set); extern void group_append_meta(pe_resource_t * rsc, xmlNode * xml); pe_node_t *pcmk__bundle_allocate(pe_resource_t *rsc, pe_node_t *preferred, pe_working_set_t *data_set); void pcmk__bundle_create_actions(pe_resource_t *rsc, pe_working_set_t *data_set); gboolean pcmk__bundle_create_probe(pe_resource_t *rsc, pe_node_t *node, pe_action_t *complete, gboolean force, pe_working_set_t *data_set); void pcmk__bundle_internal_constraints(pe_resource_t *rsc, pe_working_set_t *data_set); void pcmk__bundle_rsc_colocation_lh(pe_resource_t *lh_rsc, pe_resource_t *rh_rsc, pcmk__colocation_t *constraint, pe_working_set_t *data_set); void pcmk__bundle_rsc_colocation_rh(pe_resource_t *lh_rsc, pe_resource_t *rh_rsc, pcmk__colocation_t *constraint, pe_working_set_t *data_set); void pcmk__bundle_rsc_location(pe_resource_t *rsc, pe__location_t *constraint); enum pe_action_flags pcmk__bundle_action_flags(pe_action_t *action, pe_node_t *node); void pcmk__bundle_expand(pe_resource_t *rsc, pe_working_set_t *data_set); void pcmk__bundle_append_meta(pe_resource_t *rsc, xmlNode *xml); pe_node_t *pcmk__clone_allocate(pe_resource_t *rsc, pe_node_t *preferred, pe_working_set_t *data_set); extern void clone_create_actions(pe_resource_t * rsc, pe_working_set_t * data_set); extern void clone_internal_constraints(pe_resource_t * rsc, pe_working_set_t * data_set); void clone_rsc_colocation_lh(pe_resource_t *lh_rsc, pe_resource_t *rh_rsc, pcmk__colocation_t *constraint, pe_working_set_t *data_set); void clone_rsc_colocation_rh(pe_resource_t *lh_rsc, pe_resource_t *rh_rsc, pcmk__colocation_t *constraint, pe_working_set_t *data_set); void clone_rsc_location(pe_resource_t *rsc, pe__location_t *constraint); extern enum pe_action_flags clone_action_flags(pe_action_t * action, pe_node_t * node); extern void clone_expand(pe_resource_t * rsc, pe_working_set_t * data_set); extern gboolean clone_create_probe(pe_resource_t * rsc, pe_node_t * node, pe_action_t * complete, gboolean force, pe_working_set_t * data_set); extern void clone_append_meta(pe_resource_t * rsc, xmlNode * xml); void pcmk__add_promotion_scores(pe_resource_t *rsc); pe_node_t *pcmk__set_instance_roles(pe_resource_t *rsc, pe_working_set_t *data_set); void create_promotable_actions(pe_resource_t *rsc, pe_working_set_t *data_set); void promote_demote_constraints(pe_resource_t *rsc, pe_working_set_t *data_set); void promotable_constraints(pe_resource_t *rsc, pe_working_set_t *data_set); void promotable_colocation_rh(pe_resource_t *lh_rsc, pe_resource_t *rh_rsc, pcmk__colocation_t *constraint, pe_working_set_t *data_set); /* extern resource_object_functions_t resource_variants[]; */ extern resource_alloc_functions_t resource_class_alloc_functions[]; void LogNodeActions(pe_working_set_t * data_set); void LogActions(pe_resource_t * rsc, pe_working_set_t * data_set); void pcmk__bundle_log_actions(pe_resource_t *rsc, pe_working_set_t *data_set); -extern void rsc_stonith_ordering(pe_resource_t * rsc, pe_action_t * stonith_op, - pe_working_set_t * data_set); - enum pe_graph_flags native_update_actions(pe_action_t *first, pe_action_t *then, pe_node_t *node, enum pe_action_flags flags, enum pe_action_flags filter, enum pe_ordering type, pe_working_set_t *data_set); enum pe_graph_flags group_update_actions(pe_action_t *first, pe_action_t *then, pe_node_t *node, enum pe_action_flags flags, enum pe_action_flags filter, enum pe_ordering type, pe_working_set_t *data_set); enum pe_graph_flags pcmk__multi_update_actions(pe_action_t *first, pe_action_t *then, pe_node_t *node, enum pe_action_flags flags, enum pe_action_flags filter, enum pe_ordering type, pe_working_set_t *data_set); gboolean update_action(pe_action_t *action, pe_working_set_t *data_set); void complex_set_cmds(pe_resource_t * rsc); void pcmk__log_transition_summary(const char *filename); void clone_create_pseudo_actions( pe_resource_t * rsc, GList *children, notify_data_t **start_notify, notify_data_t **stop_notify, pe_working_set_t * data_set); #endif diff --git a/include/pcmki/pcmki_scheduler.h b/include/pcmki/pcmki_scheduler.h index 8150252150..74e602f73f 100644 --- a/include/pcmki/pcmki_scheduler.h +++ b/include/pcmki/pcmki_scheduler.h @@ -1,132 +1,130 @@ /* * Copyright 2014-2021 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 PENGINE__H # define PENGINE__H typedef struct rsc_ticket_s rsc_ticket_t; typedef struct lrm_agent_s lrm_agent_t; # include # include # include # include # include # include # include enum pe_stop_fail { pesf_block, pesf_stonith, pesf_ignore }; enum pe_weights { pe_weights_none = 0x0, pe_weights_init = 0x1, pe_weights_forward = 0x4, pe_weights_positive = 0x8, pe_weights_rollback = 0x10, }; typedef struct { const char *id; const char *node_attribute; pe_resource_t *rsc_lh; pe_resource_t *rsc_rh; int role_lh; int role_rh; int score; bool influence; // Whether rsc_lh should influence active rsc_rh placement } pcmk__colocation_t; enum loss_ticket_policy_e { loss_ticket_stop, loss_ticket_demote, loss_ticket_fence, loss_ticket_freeze }; struct rsc_ticket_s { const char *id; pe_resource_t *rsc_lh; pe_ticket_t *ticket; enum loss_ticket_policy_e loss_policy; int role_lh; }; extern gboolean stage0(pe_working_set_t * data_set); extern gboolean probe_resources(pe_working_set_t * data_set); extern gboolean stage2(pe_working_set_t * data_set); extern gboolean stage3(pe_working_set_t * data_set); extern gboolean stage4(pe_working_set_t * data_set); extern gboolean stage5(pe_working_set_t * data_set); extern gboolean stage6(pe_working_set_t * data_set); extern gboolean stage7(pe_working_set_t * data_set); extern gboolean stage8(pe_working_set_t * data_set); extern gboolean unpack_constraints(xmlNode * xml_constraints, pe_working_set_t * data_set); extern gboolean shutdown_constraints(pe_node_t * node, pe_action_t * shutdown_op, pe_working_set_t * data_set); -void pcmk__order_vs_fence(pe_action_t *stonith_op, pe_working_set_t *data_set); - extern int custom_action_order(pe_resource_t * lh_rsc, char *lh_task, pe_action_t * lh_action, pe_resource_t * rh_rsc, char *rh_task, pe_action_t * rh_action, enum pe_ordering type, pe_working_set_t * data_set); extern int new_rsc_order(pe_resource_t * lh_rsc, const char *lh_task, pe_resource_t * rh_rsc, const char *rh_task, enum pe_ordering type, pe_working_set_t * data_set); # define order_start_start(rsc1,rsc2, type) \ new_rsc_order(rsc1, CRMD_ACTION_START, rsc2, CRMD_ACTION_START, type, data_set) # define order_stop_stop(rsc1, rsc2, type) \ new_rsc_order(rsc1, CRMD_ACTION_STOP, rsc2, CRMD_ACTION_STOP, type, data_set) extern void graph_element_from_action(pe_action_t * action, pe_working_set_t * data_set); extern void add_maintenance_update(pe_working_set_t *data_set); xmlNode *pcmk__schedule_actions(pe_working_set_t *data_set, xmlNode *xml_input, crm_time_t *now); bool pcmk__ordering_is_invalid(pe_action_t *action, pe_action_wrapper_t *input); extern const char *transition_idle_timeout; /*! * \internal * \brief Check whether colocation's left-hand preferences should be considered * * \param[in] colocation Colocation constraint * \param[in] rsc Right-hand instance (normally this will be * colocation->rsc_rh, which NULL will be treated as, * but for clones or bundles with multiple instances * this can be a particular instance) * * \return true if colocation influence should be effective, otherwise false */ static inline bool pcmk__colocation_has_influence(const pcmk__colocation_t *colocation, const pe_resource_t *rsc) { if (rsc == NULL) { rsc = colocation->rsc_rh; } /* The left hand of a colocation influences the right hand's location * if the influence option is true, or the right hand is not yet active. */ return colocation->influence || (rsc->running_on == NULL); } #endif diff --git a/lib/pacemaker/Makefile.am b/lib/pacemaker/Makefile.am index af6b43ea8c..314497ff6e 100644 --- a/lib/pacemaker/Makefile.am +++ b/lib/pacemaker/Makefile.am @@ -1,51 +1,54 @@ # # Copyright 2004-2021 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 $(top_srcdir)/mk/common.mk AM_CPPFLAGS += -I$(top_builddir) -I$(top_srcdir) +noinst_HEADERS = libpacemaker_private.h + ## libraries lib_LTLIBRARIES = libpacemaker.la ## SOURCES libpacemaker_la_LDFLAGS = -version-info 3:1:2 libpacemaker_la_CFLAGS = $(CFLAGS_HARDENED_LIB) libpacemaker_la_LDFLAGS += $(LDFLAGS_HARDENED_LIB) libpacemaker_la_LIBADD = $(top_builddir)/lib/pengine/libpe_status.la \ $(top_builddir)/lib/cib/libcib.la \ $(top_builddir)/lib/lrmd/liblrmd.la \ $(top_builddir)/lib/common/libcrmcommon.la # -L$(top_builddir)/lib/pils -lpils -export-dynamic -module -avoid-version # Use += rather than backlashed continuation lines for parsing by bumplibs libpacemaker_la_SOURCES = libpacemaker_la_SOURCES += pcmk_cluster_queries.c libpacemaker_la_SOURCES += pcmk_fence.c libpacemaker_la_SOURCES += pcmk_graph_consumer.c libpacemaker_la_SOURCES += pcmk_graph_logging.c libpacemaker_la_SOURCES += pcmk_graph_producer.c libpacemaker_la_SOURCES += pcmk_output.c libpacemaker_la_SOURCES += pcmk_output_utils.c libpacemaker_la_SOURCES += pcmk_resource.c libpacemaker_la_SOURCES += pcmk_sched_allocate.c libpacemaker_la_SOURCES += pcmk_sched_bundle.c libpacemaker_la_SOURCES += pcmk_sched_clone.c libpacemaker_la_SOURCES += pcmk_sched_constraints.c +libpacemaker_la_SOURCES += pcmk_sched_fencing.c libpacemaker_la_SOURCES += pcmk_sched_group.c libpacemaker_la_SOURCES += pcmk_sched_messages.c libpacemaker_la_SOURCES += pcmk_sched_native.c libpacemaker_la_SOURCES += pcmk_sched_notif.c libpacemaker_la_SOURCES += pcmk_sched_promotable.c libpacemaker_la_SOURCES += pcmk_sched_transition.c libpacemaker_la_SOURCES += pcmk_sched_utilization.c libpacemaker_la_SOURCES += pcmk_sched_utils.c libpacemaker_la_SOURCES += pcmk_simulate.c diff --git a/lib/pacemaker/libpacemaker_private.h b/lib/pacemaker/libpacemaker_private.h new file mode 100644 index 0000000000..ee64ec0b9c --- /dev/null +++ b/lib/pacemaker/libpacemaker_private.h @@ -0,0 +1,37 @@ +/* + * Copyright 2021 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__LIBPACEMAKER_PRIVATE__H +# define PCMK__LIBPACEMAKER_PRIVATE__H + +/* This header is for the sole use of libpacemaker, so that functions can be + * declared with G_GNUC_INTERNAL for efficiency. + */ + +#include // pe_action_t, pe_node_t, pe_working_set_t + +G_GNUC_INTERNAL +void pcmk__order_vs_fence(pe_action_t *stonith_op, pe_working_set_t *data_set); + +G_GNUC_INTERNAL +void pcmk__order_vs_unfence(pe_resource_t *rsc, pe_node_t *node, + pe_action_t *action, enum pe_ordering order, + pe_working_set_t *data_set); + +G_GNUC_INTERNAL +void pcmk__fence_guest(pe_node_t *node, pe_working_set_t *data_set); + +G_GNUC_INTERNAL +bool pcmk__node_unfenced(pe_node_t *node); + +G_GNUC_INTERNAL +bool pcmk__is_unfence_device(const pe_resource_t *rsc, + const pe_working_set_t *data_set); + +#endif // PCMK__LIBPACEMAKER_PRIVATE__H diff --git a/lib/pacemaker/pcmk_graph_producer.c b/lib/pacemaker/pcmk_graph_producer.c index 67f591b66f..a52a2bb872 100644 --- a/lib/pacemaker/pcmk_graph_producer.c +++ b/lib/pacemaker/pcmk_graph_producer.c @@ -1,1866 +1,1846 @@ /* * Copyright 2004-2021 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 #include #include #include #include #include #include #include void update_colo_start_chain(pe_action_t *action, pe_working_set_t *data_set); gboolean rsc_update_action(pe_action_t * first, pe_action_t * then, enum pe_ordering type); static enum pe_action_flags get_action_flags(pe_action_t * action, pe_node_t * node) { enum pe_action_flags flags = action->flags; if (action->rsc) { flags = action->rsc->cmds->action_flags(action, NULL); if (pe_rsc_is_clone(action->rsc) && node) { /* We only care about activity on $node */ enum pe_action_flags clone_flags = action->rsc->cmds->action_flags(action, node); /* Go to great lengths to ensure the correct value for pe_action_runnable... * * If we are a clone, then for _ordering_ constraints, it's only relevant * if we are runnable _anywhere_. * * This only applies to _runnable_ though, and only for ordering constraints. * If this function is ever used during colocation, then we'll need additional logic * * Not very satisfying, but it's logical and appears to work well. */ if (!pcmk_is_set(clone_flags, pe_action_runnable) && pcmk_is_set(flags, pe_action_runnable)) { pe__set_raw_action_flags(clone_flags, action->rsc->id, pe_action_runnable); } flags = clone_flags; } } return flags; } static char * convert_non_atomic_uuid(char *old_uuid, pe_resource_t * rsc, gboolean allow_notify, gboolean free_original) { guint interval_ms = 0; char *uuid = NULL; char *rid = NULL; char *raw_task = NULL; int task = no_action; CRM_ASSERT(rsc); pe_rsc_trace(rsc, "Processing %s", old_uuid); if (old_uuid == NULL) { return NULL; } else if (strstr(old_uuid, "notify") != NULL) { goto done; /* no conversion */ } else if (rsc->variant < pe_group) { goto done; /* no conversion */ } CRM_ASSERT(parse_op_key(old_uuid, &rid, &raw_task, &interval_ms)); if (interval_ms > 0) { goto done; /* no conversion */ } task = text2task(raw_task); switch (task) { case stop_rsc: case start_rsc: case action_notify: case action_promote: case action_demote: break; case stopped_rsc: case started_rsc: case action_notified: case action_promoted: case action_demoted: task--; break; case monitor_rsc: case shutdown_crm: case stonith_node: task = no_action; break; default: crm_err("Unknown action: %s", raw_task); task = no_action; break; } if (task != no_action) { if (pcmk_is_set(rsc->flags, pe_rsc_notify) && allow_notify) { uuid = pcmk__notify_key(rid, "confirmed-post", task2text(task + 1)); } else { uuid = pcmk__op_key(rid, task2text(task + 1), 0); } pe_rsc_trace(rsc, "Converted %s -> %s", old_uuid, uuid); } done: if (uuid == NULL) { uuid = strdup(old_uuid); } if (free_original) { free(old_uuid); } free(raw_task); free(rid); return uuid; } static pe_action_t * rsc_expand_action(pe_action_t * action) { gboolean notify = FALSE; pe_action_t *result = action; pe_resource_t *rsc = action->rsc; if (rsc == NULL) { return action; } if ((rsc->parent == NULL) || (pe_rsc_is_clone(rsc) && (rsc->parent->variant == pe_container))) { /* Only outermost resources have notification actions. * The exception is those in bundles. */ notify = pcmk_is_set(rsc->flags, pe_rsc_notify); } if (rsc->variant >= pe_group) { /* Expand 'start' -> 'started' */ char *uuid = NULL; uuid = convert_non_atomic_uuid(action->uuid, rsc, notify, FALSE); if (uuid) { pe_rsc_trace(rsc, "Converting %s to %s %d", action->uuid, uuid, pcmk_is_set(rsc->flags, pe_rsc_notify)); result = find_first_action(rsc->actions, uuid, NULL, NULL); if (result == NULL) { crm_err("Couldn't expand %s to %s in %s", action->uuid, uuid, rsc->id); result = action; } free(uuid); } } return result; } static enum pe_graph_flags graph_update_action(pe_action_t * first, pe_action_t * then, pe_node_t * node, enum pe_action_flags first_flags, enum pe_action_flags then_flags, pe_action_wrapper_t *order, pe_working_set_t *data_set) { enum pe_graph_flags changed = pe_graph_none; enum pe_ordering type = order->type; /* TODO: Do as many of these in parallel as possible */ if (pcmk_is_set(type, pe_order_implies_then_on_node)) { /* Normally we want the _whole_ 'then' clone to * restart if 'first' is restarted, so then->node is * needed. * * However for unfencing, we want to limit this to * instances on the same node as 'first' (the * unfencing operation), so first->node is supplied. * * Swap the node, from then on we can can treat it * like any other 'pe_order_implies_then' */ pe__clear_order_flags(type, pe_order_implies_then_on_node); pe__set_order_flags(type, pe_order_implies_then); node = first->node; pe_rsc_trace(then->rsc, "%s then %s: mapped pe_order_implies_then_on_node to " "pe_order_implies_then on %s", first->uuid, then->uuid, node->details->uname); } if (type & pe_order_implies_then) { if (then->rsc) { changed |= then->rsc->cmds->update_actions(first, then, node, first_flags & pe_action_optional, pe_action_optional, pe_order_implies_then, data_set); } else if (!pcmk_is_set(first_flags, pe_action_optional) && pcmk_is_set(then->flags, pe_action_optional)) { pe__clear_action_flags(then, pe_action_optional); pe__set_graph_flags(changed, first, pe_graph_updated_then); } pe_rsc_trace(then->rsc, "%s then %s: %s after pe_order_implies_then", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if ((type & pe_order_restart) && then->rsc) { enum pe_action_flags restart = (pe_action_optional | pe_action_runnable); changed |= then->rsc->cmds->update_actions(first, then, node, first_flags, restart, pe_order_restart, data_set); pe_rsc_trace(then->rsc, "%s then %s: %s after pe_order_restart", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (type & pe_order_implies_first) { if (first->rsc) { changed |= first->rsc->cmds->update_actions(first, then, node, first_flags, pe_action_optional, pe_order_implies_first, data_set); } else if (!pcmk_is_set(first_flags, pe_action_optional) && pcmk_is_set(first->flags, pe_action_runnable)) { pe__clear_action_flags(first, pe_action_runnable); pe__set_graph_flags(changed, first, pe_graph_updated_first); } pe_rsc_trace(then->rsc, "%s then %s: %s after pe_order_implies_first", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (type & pe_order_promoted_implies_first) { if (then->rsc) { changed |= then->rsc->cmds->update_actions(first, then, node, first_flags & pe_action_optional, pe_action_optional, pe_order_promoted_implies_first, data_set); } pe_rsc_trace(then->rsc, "%s then %s: %s after pe_order_promoted_implies_first", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (type & pe_order_one_or_more) { if (then->rsc) { changed |= then->rsc->cmds->update_actions(first, then, node, first_flags, pe_action_runnable, pe_order_one_or_more, data_set); } else if (pcmk_is_set(first_flags, pe_action_runnable)) { // We have another runnable instance of "first" then->runnable_before++; /* Mark "then" as runnable if it requires a certain number of * "before" instances to be runnable, and they now are. */ if ((then->runnable_before >= then->required_runnable_before) && !pcmk_is_set(then->flags, pe_action_runnable)) { pe__set_action_flags(then, pe_action_runnable); pe__set_graph_flags(changed, first, pe_graph_updated_then); } } pe_rsc_trace(then->rsc, "%s then %s: %s after pe_order_one_or_more", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (then->rsc && pcmk_is_set(type, pe_order_probe)) { if (!pcmk_is_set(first_flags, pe_action_runnable) && (first->rsc->running_on != NULL)) { pe_rsc_trace(then->rsc, "%s then %s: ignoring because first is stopping", first->uuid, then->uuid); type = pe_order_none; order->type = pe_order_none; } else { changed |= then->rsc->cmds->update_actions(first, then, node, first_flags, pe_action_runnable, pe_order_runnable_left, data_set); } pe_rsc_trace(then->rsc, "%s then %s: %s after pe_order_probe", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (type & pe_order_runnable_left) { if (then->rsc) { changed |= then->rsc->cmds->update_actions(first, then, node, first_flags, pe_action_runnable, pe_order_runnable_left, data_set); } else if (!pcmk_is_set(first_flags, pe_action_runnable) && pcmk_is_set(then->flags, pe_action_runnable)) { pe__clear_action_flags(then, pe_action_runnable); pe__set_graph_flags(changed, first, pe_graph_updated_then); } pe_rsc_trace(then->rsc, "%s then %s: %s after pe_order_runnable_left", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (type & pe_order_implies_first_migratable) { if (then->rsc) { changed |= then->rsc->cmds->update_actions(first, then, node, first_flags, pe_action_optional, pe_order_implies_first_migratable, data_set); } pe_rsc_trace(then->rsc, "%s then %s: %s after " "pe_order_implies_first_migratable", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (type & pe_order_pseudo_left) { if (then->rsc) { changed |= then->rsc->cmds->update_actions(first, then, node, first_flags, pe_action_optional, pe_order_pseudo_left, data_set); } pe_rsc_trace(then->rsc, "%s then %s: %s after pe_order_pseudo_left", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (type & pe_order_optional) { if (then->rsc) { changed |= then->rsc->cmds->update_actions(first, then, node, first_flags, pe_action_runnable, pe_order_optional, data_set); } pe_rsc_trace(then->rsc, "%s then %s: %s after pe_order_optional", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if (type & pe_order_asymmetrical) { if (then->rsc) { changed |= then->rsc->cmds->update_actions(first, then, node, first_flags, pe_action_runnable, pe_order_asymmetrical, data_set); } pe_rsc_trace(then->rsc, "%s then %s: %s after pe_order_asymmetrical", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } if ((first->flags & pe_action_runnable) && (type & pe_order_implies_then_printed) && (first_flags & pe_action_optional) == 0) { pe_rsc_trace(then->rsc, "%s will be in graph because %s is required", then->uuid, first->uuid); pe__set_action_flags(then, pe_action_print_always); // Don't bother marking 'then' as changed just for this } if (pcmk_is_set(type, pe_order_implies_first_printed) && !pcmk_is_set(then_flags, pe_action_optional)) { pe_rsc_trace(then->rsc, "%s will be in graph because %s is required", first->uuid, then->uuid); pe__set_action_flags(first, pe_action_print_always); // Don't bother marking 'first' as changed just for this } if ((type & pe_order_implies_then || type & pe_order_implies_first || type & pe_order_restart) && first->rsc && pcmk__str_eq(first->task, RSC_STOP, pcmk__str_casei) && !pcmk_is_set(first->rsc->flags, pe_rsc_managed) && pcmk_is_set(first->rsc->flags, pe_rsc_block) && !pcmk_is_set(first->flags, pe_action_runnable)) { if (pcmk_is_set(then->flags, pe_action_runnable)) { pe__clear_action_flags(then, pe_action_runnable); pe__set_graph_flags(changed, first, pe_graph_updated_then); } pe_rsc_trace(then->rsc, "%s then %s: %s after checking whether first " "is blocked, unmanaged, unrunnable stop", first->uuid, then->uuid, (changed? "changed" : "unchanged")); } return changed; } static void mark_start_blocked(pe_resource_t *rsc, pe_resource_t *reason, pe_working_set_t *data_set) { GList *gIter = rsc->actions; char *reason_text = crm_strdup_printf("colocation with %s", reason->id); for (; gIter != NULL; gIter = gIter->next) { pe_action_t *action = (pe_action_t *) gIter->data; if (!pcmk__str_eq(action->task, RSC_START, pcmk__str_casei)) { continue; } if (pcmk_is_set(action->flags, pe_action_runnable)) { pe__clear_action_flags(action, pe_action_runnable); pe_action_set_reason(action, reason_text, false); update_colo_start_chain(action, data_set); update_action(action, data_set); } } free(reason_text); } void update_colo_start_chain(pe_action_t *action, pe_working_set_t *data_set) { GList *gIter = NULL; pe_resource_t *rsc = NULL; if (!pcmk_is_set(action->flags, pe_action_runnable) && pcmk__str_eq(action->task, RSC_START, pcmk__str_casei)) { rsc = uber_parent(action->rsc); if (rsc->parent) { /* For bundles, uber_parent() returns the clone, not the bundle, so * the existence of rsc->parent implies this is a bundle. * In this case, we need the bundle resource, so that we can check * if all containers are stopped/stopping. */ rsc = rsc->parent; } } if (rsc == NULL || rsc->rsc_cons_lhs == NULL) { return; } /* if rsc has children, all the children need to have start set to * unrunnable before we follow the colo chain for the parent. */ for (gIter = rsc->children; gIter != NULL; gIter = gIter->next) { pe_resource_t *child = (pe_resource_t *)gIter->data; pe_action_t *start = find_first_action(child->actions, NULL, RSC_START, NULL); if ((start == NULL) || pcmk_is_set(start->flags, pe_action_runnable)) { return; } } for (gIter = rsc->rsc_cons_lhs; gIter != NULL; gIter = gIter->next) { pcmk__colocation_t *colocate_with = (pcmk__colocation_t *) gIter->data; if (colocate_with->score == INFINITY) { mark_start_blocked(colocate_with->rsc_lh, action->rsc, data_set); } } } // Convenience macros for logging action properties #define action_type_str(flags) \ (pcmk_is_set((flags), pe_action_pseudo)? "pseudo-action" : "action") #define action_optional_str(flags) \ (pcmk_is_set((flags), pe_action_optional)? "optional" : "required") #define action_runnable_str(flags) \ (pcmk_is_set((flags), pe_action_runnable)? "runnable" : "unrunnable") #define action_node_str(a) \ (((a)->node == NULL)? "no node" : (a)->node->details->uname) gboolean update_action(pe_action_t *then, pe_working_set_t *data_set) { GList *lpc = NULL; enum pe_graph_flags changed = pe_graph_none; int last_flags = then->flags; pe_rsc_trace(then->rsc, "Updating %s %s (%s %s) on %s", action_type_str(then->flags), then->uuid, action_optional_str(then->flags), action_runnable_str(then->flags), action_node_str(then)); if (pcmk_is_set(then->flags, pe_action_requires_any)) { /* initialize current known runnable before actions to 0 * from here as graph_update_action is called for each of * then's before actions, this number will increment as * runnable 'first' actions are encountered */ then->runnable_before = 0; /* for backwards compatibility with previous options that use * the 'requires_any' flag, initialize required to 1 if it is * not set. */ if (then->required_runnable_before == 0) { then->required_runnable_before = 1; } pe__clear_action_flags(then, pe_action_runnable); /* We are relying on the pe_order_one_or_more clause of * graph_update_action(), called as part of the: * * 'if (first == other->action)' * * block below, to set this back if appropriate */ } for (lpc = then->actions_before; lpc != NULL; lpc = lpc->next) { pe_action_wrapper_t *other = (pe_action_wrapper_t *) lpc->data; pe_action_t *first = other->action; pe_node_t *then_node = then->node; pe_node_t *first_node = first->node; enum pe_action_flags then_flags = 0; enum pe_action_flags first_flags = 0; if (first->rsc && first->rsc->variant == pe_group && pcmk__str_eq(first->task, RSC_START, pcmk__str_casei)) { first_node = first->rsc->fns->location(first->rsc, NULL, FALSE); if (first_node) { pe_rsc_trace(first->rsc, "Found node %s for 'first' %s", first_node->details->uname, first->uuid); } } if (then->rsc && then->rsc->variant == pe_group && pcmk__str_eq(then->task, RSC_START, pcmk__str_casei)) { then_node = then->rsc->fns->location(then->rsc, NULL, FALSE); if (then_node) { pe_rsc_trace(then->rsc, "Found node %s for 'then' %s", then_node->details->uname, then->uuid); } } /* Disable constraint if it only applies when on same node, but isn't */ if (pcmk_is_set(other->type, pe_order_same_node) && (first_node != NULL) && (then_node != NULL) && (first_node->details != then_node->details)) { pe_rsc_trace(then->rsc, "Disabled ordering %s on %s then %s on %s: not same node", other->action->uuid, first_node->details->uname, then->uuid, then_node->details->uname); other->type = pe_order_none; continue; } pe__clear_graph_flags(changed, then, pe_graph_updated_first); if (first->rsc && pcmk_is_set(other->type, pe_order_then_cancels_first) && !pcmk_is_set(then->flags, pe_action_optional)) { /* 'then' is required, so we must abandon 'first' * (e.g. a required stop cancels any agent reload). */ pe__set_action_flags(other->action, pe_action_optional); if (!strcmp(first->task, CRMD_ACTION_RELOAD_AGENT)) { pe__clear_resource_flags(first->rsc, pe_rsc_reload); } } if (first->rsc && then->rsc && (first->rsc != then->rsc) && (is_parent(then->rsc, first->rsc) == FALSE)) { first = rsc_expand_action(first); } if (first != other->action) { pe_rsc_trace(then->rsc, "Ordering %s after %s instead of %s", then->uuid, first->uuid, other->action->uuid); } first_flags = get_action_flags(first, then_node); then_flags = get_action_flags(then, first_node); pe_rsc_trace(then->rsc, "%s then %s: type=0x%.6x filter=0x%.6x " "(%s %s %s on %s 0x%.6x then 0x%.6x)", first->uuid, then->uuid, other->type, first_flags, action_optional_str(first_flags), action_runnable_str(first_flags), action_type_str(first_flags), action_node_str(first), first->flags, then->flags); if (first == other->action) { /* * 'first' was not expanded (e.g. from 'start' to 'running'), which could mean it: * - has no associated resource, * - was a primitive, * - was pre-expanded (e.g. 'running' instead of 'start') * * The third argument here to graph_update_action() is a node which is used under two conditions: * - Interleaving, in which case first->node and * then->node are equal (and NULL) * - If 'then' is a clone, to limit the scope of the * constraint to instances on the supplied node * */ pe_node_t *node = then->node; changed |= graph_update_action(first, then, node, first_flags, then_flags, other, data_set); /* 'first' was for a complex resource (clone, group, etc), * create a new dependency if necessary */ } else if (order_actions(first, then, other->type)) { /* This was the first time 'first' and 'then' were associated, * start again to get the new actions_before list */ pe__set_graph_flags(changed, then, pe_graph_updated_then|pe_graph_disable); } if (changed & pe_graph_disable) { pe_rsc_trace(then->rsc, "Disabled ordering %s then %s in favor of %s then %s", other->action->uuid, then->uuid, first->uuid, then->uuid); pe__clear_graph_flags(changed, then, pe_graph_disable); other->type = pe_order_none; } if (changed & pe_graph_updated_first) { GList *lpc2 = NULL; crm_trace("Re-processing %s and its 'after' actions since it changed", first->uuid); for (lpc2 = first->actions_after; lpc2 != NULL; lpc2 = lpc2->next) { pe_action_wrapper_t *other = (pe_action_wrapper_t *) lpc2->data; update_action(other->action, data_set); } update_action(first, data_set); } } if (pcmk_is_set(then->flags, pe_action_requires_any)) { if (last_flags != then->flags) { pe__set_graph_flags(changed, then, pe_graph_updated_then); } else { pe__clear_graph_flags(changed, then, pe_graph_updated_then); } } if (changed & pe_graph_updated_then) { crm_trace("Re-processing %s and its 'after' actions since it changed", then->uuid); if (pcmk_is_set(last_flags, pe_action_runnable) && !pcmk_is_set(then->flags, pe_action_runnable)) { update_colo_start_chain(then, data_set); } update_action(then, data_set); for (lpc = then->actions_after; lpc != NULL; lpc = lpc->next) { pe_action_wrapper_t *other = (pe_action_wrapper_t *) lpc->data; update_action(other->action, data_set); } } return FALSE; } gboolean shutdown_constraints(pe_node_t * node, pe_action_t * shutdown_op, pe_working_set_t * data_set) { /* add the stop to the before lists so it counts as a pre-req * for the shutdown */ GList *lpc = NULL; for (lpc = data_set->actions; lpc != NULL; lpc = lpc->next) { pe_action_t *action = (pe_action_t *) lpc->data; if (action->rsc == NULL || action->node == NULL) { continue; } else if (action->node->details != node->details) { continue; } else if (pcmk_is_set(action->rsc->flags, pe_rsc_maintenance)) { pe_rsc_trace(action->rsc, "Skipping %s: maintenance mode", action->uuid); continue; } else if (node->details->maintenance) { pe_rsc_trace(action->rsc, "Skipping %s: node %s is in maintenance mode", action->uuid, node->details->uname); continue; } else if (!pcmk__str_eq(action->task, RSC_STOP, pcmk__str_casei)) { continue; } else if (!pcmk_any_flags_set(action->rsc->flags, pe_rsc_managed|pe_rsc_block)) { /* * If another action depends on this one, we may still end up blocking */ pe_rsc_trace(action->rsc, "Skipping %s: unmanaged", action->uuid); continue; } pe_rsc_trace(action->rsc, "Ordering %s before shutdown on %s", action->uuid, node->details->uname); pe__clear_action_flags(action, pe_action_optional); custom_action_order(action->rsc, NULL, action, NULL, strdup(CRM_OP_SHUTDOWN), shutdown_op, pe_order_optional | pe_order_runnable_left, data_set); } return TRUE; } -/*! - * \internal - * \brief Order all actions appropriately relative to a fencing operation - * - * Ensure start operations of affected resources are ordered after fencing, - * imply stop and demote operations of affected resources by marking them as - * pseudo-actions, etc. - * - * \param[in] stonith_op Fencing operation - * \param[in,out] data_set Working set of cluster - */ -void -pcmk__order_vs_fence(pe_action_t *stonith_op, pe_working_set_t *data_set) -{ - CRM_CHECK(stonith_op && data_set, return); - for (GList *r = data_set->resources; r != NULL; r = r->next) { - rsc_stonith_ordering((pe_resource_t *) r->data, stonith_op, data_set); - } -} - static pe_node_t * get_router_node(pe_action_t *action) { pe_node_t *began_on = NULL; pe_node_t *ended_on = NULL; bool partial_migration = FALSE; const char *task = action->task; if (pcmk__str_eq(task, CRM_OP_FENCE, pcmk__str_casei) || !pe__is_guest_or_remote_node(action->node)) { return NULL; } CRM_ASSERT(action->node->details->remote_rsc != NULL); began_on = pe__current_node(action->node->details->remote_rsc); ended_on = action->node->details->remote_rsc->allocated_to; if (action->node->details->remote_rsc && (action->node->details->remote_rsc->container == NULL) && action->node->details->remote_rsc->partial_migration_target) { partial_migration = TRUE; } if (began_on == NULL) { crm_trace("Routing %s for %s through remote connection's " "next node %s (starting)%s", action->task, (action->rsc? action->rsc->id : "no resource"), (ended_on? ended_on->details->uname : "none"), partial_migration? " (partial migration)" : ""); return ended_on; } if (ended_on == NULL) { crm_trace("Routing %s for %s through remote connection's " "current node %s (stopping)%s", action->task, (action->rsc? action->rsc->id : "no resource"), (began_on? began_on->details->uname : "none"), partial_migration? " (partial migration)" : ""); return began_on; } if (began_on->details == ended_on->details) { crm_trace("Routing %s for %s through remote connection's " "current node %s (not moving)%s", action->task, (action->rsc? action->rsc->id : "no resource"), (began_on? began_on->details->uname : "none"), partial_migration? " (partial migration)" : ""); return began_on; } /* If we get here, the remote connection is moving during this transition. * This means some actions for resources behind the connection will get * routed through the cluster node the connection reource is currently on, * and others are routed through the cluster node the connection will end up * on. */ if (pcmk__str_eq(task, "notify", pcmk__str_casei)) { task = g_hash_table_lookup(action->meta, "notify_operation"); } /* * Stop, demote, and migration actions must occur before the connection can * move (these actions are required before the remote resource can stop). In * this case, we know these actions have to be routed through the initial * cluster node the connection resource lived on before the move takes * place. * * The exception is a partial migration of a (non-guest) remote connection * resource; in that case, all actions (even these) will be ordered after * the connection's pseudo-start on the migration target, so the target is * the router node. */ if (pcmk__strcase_any_of(task, "cancel", "stop", "demote", "migrate_from", "migrate_to", NULL) && !partial_migration) { crm_trace("Routing %s for %s through remote connection's " "current node %s (moving)%s", action->task, (action->rsc? action->rsc->id : "no resource"), (began_on? began_on->details->uname : "none"), partial_migration? " (partial migration)" : ""); return began_on; } /* Everything else (start, promote, monitor, probe, refresh, * clear failcount, delete, ...) must occur after the connection starts on * the node it is moving to. */ crm_trace("Routing %s for %s through remote connection's " "next node %s (moving)%s", action->task, (action->rsc? action->rsc->id : "no resource"), (ended_on? ended_on->details->uname : "none"), partial_migration? " (partial migration)" : ""); return ended_on; } /*! * \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_UUID, 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 Add XML with nodes that need an update of their maintenance state * * \param[in,out] xml Parent XML tag to add to * \param[in] data_set Working set for cluster */ static int add_maintenance_nodes(xmlNode *xml, const pe_working_set_t *data_set) { GList *gIter = NULL; xmlNode *maintenance = xml?create_xml_node(xml, XML_GRAPH_TAG_MAINTENANCE):NULL; int count = 0; for (gIter = data_set->nodes; gIter != NULL; gIter = gIter->next) { pe_node_t *node = (pe_node_t *) gIter->data; struct pe_node_shared_s *details = node->details; if (!pe__is_guest_or_remote_node(node)) { continue; /* just remote nodes need to know atm */ } if (details->maintenance != details->remote_maintenance) { if (maintenance) { crm_xml_add( add_node_to_xml_by_id(node->details->id, maintenance), XML_NODE_IS_MAINTENANCE, details->maintenance?"1":"0"); } count++; } } crm_trace("%s %d nodes to adjust maintenance-mode " "to transition", maintenance?"Added":"Counted", count); return count; } /*! * \internal * \brief Add pseudo action with nodes needing maintenance state update * * \param[in,out] data_set Working set for cluster */ void add_maintenance_update(pe_working_set_t *data_set) { pe_action_t *action = NULL; if (add_maintenance_nodes(NULL, data_set)) { crm_trace("adding maintenance state update pseudo action"); action = get_pseudo_op(CRM_OP_MAINTENANCE_NODES, data_set); pe__set_action_flags(action, pe_action_print_always); } } /*! * \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 * \param[in] data_set Working set for cluster */ static void add_downed_nodes(xmlNode *xml, const pe_action_t *action, const pe_working_set_t *data_set) { CRM_CHECK(xml && action && action->node && data_set, return); if (pcmk__str_eq(action->task, CRM_OP_SHUTDOWN, pcmk__str_casei)) { /* 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, CRM_OP_FENCE, pcmk__str_casei)) { /* 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__strcase_any_of(fence, "off", "reboot", NULL)) { 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(data_set, action->node, add_node_to_xml, downed); } } else if (action->rsc && action->rsc->is_remote_node && pcmk__str_eq(action->task, CRMD_ACTION_STOP, pcmk__str_casei)) { /* Stopping a remote connection resource makes connected node down, * unless it's part of a migration */ GList *iter; pe_action_t *input; gboolean migrating = FALSE; for (iter = action->actions_before; iter != NULL; iter = iter->next) { input = ((pe_action_wrapper_t *) iter->data)->action; if (input->rsc && pcmk__str_eq(action->rsc->id, input->rsc->id, pcmk__str_casei) && pcmk__str_eq(input->task, CRMD_ACTION_MIGRATED, pcmk__str_casei)) { 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); } } } static bool should_lock_action(pe_action_t *action) { // Only actions taking place on resource's lock node are locked if ((action->rsc->lock_node == NULL) || (action->node == NULL) || (action->node->details != action->rsc->lock_node->details)) { return false; } /* During shutdown, only stops are locked (otherwise, another action such as * a demote would cause the controller to clear the lock) */ if (action->node->details->shutdown && action->task && strcmp(action->task, RSC_STOP)) { return false; } return true; } static xmlNode * action2xml(pe_action_t * action, gboolean as_input, pe_working_set_t *data_set) { gboolean needs_node_info = TRUE; gboolean needs_maintenance_info = FALSE; xmlNode *action_xml = NULL; xmlNode *args_xml = NULL; #if ENABLE_VERSIONED_ATTRS pe_rsc_action_details_t *rsc_details = NULL; #endif if (action == NULL) { return NULL; } if (pcmk__str_eq(action->task, CRM_OP_FENCE, pcmk__str_casei)) { /* All fences need node info; guest node fences are pseudo-events */ action_xml = create_xml_node(NULL, pcmk_is_set(action->flags, pe_action_pseudo)? XML_GRAPH_TAG_PSEUDO_EVENT : XML_GRAPH_TAG_CRM_EVENT); } else if (pcmk__str_eq(action->task, CRM_OP_SHUTDOWN, pcmk__str_casei)) { action_xml = create_xml_node(NULL, XML_GRAPH_TAG_CRM_EVENT); } else if (pcmk__str_eq(action->task, CRM_OP_CLEAR_FAILCOUNT, pcmk__str_casei)) { action_xml = create_xml_node(NULL, XML_GRAPH_TAG_CRM_EVENT); } else if (pcmk__str_eq(action->task, CRM_OP_LRM_REFRESH, pcmk__str_casei)) { action_xml = create_xml_node(NULL, XML_GRAPH_TAG_CRM_EVENT); } else if (pcmk__str_eq(action->task, CRM_OP_LRM_DELETE, pcmk__str_casei)) { // CIB-only clean-up for shutdown locks action_xml = create_xml_node(NULL, XML_GRAPH_TAG_CRM_EVENT); crm_xml_add(action_xml, PCMK__XA_MODE, XML_TAG_CIB); /* } else if(pcmk__str_eq(action->task, RSC_PROBED, pcmk__str_casei)) { */ /* action_xml = create_xml_node(NULL, XML_GRAPH_TAG_CRM_EVENT); */ } else if (pcmk_is_set(action->flags, pe_action_pseudo)) { if (pcmk__str_eq(action->task, CRM_OP_MAINTENANCE_NODES, pcmk__str_casei)) { needs_maintenance_info = TRUE; } action_xml = create_xml_node(NULL, XML_GRAPH_TAG_PSEUDO_EVENT); needs_node_info = FALSE; } else { action_xml = create_xml_node(NULL, XML_GRAPH_TAG_RSC_OP); #if ENABLE_VERSIONED_ATTRS rsc_details = pe_rsc_action_details(action); #endif } 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; } if (pcmk__str_eq(action->task, RSC_NOTIFY, pcmk__str_casei)) { 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_CHECK(n_type != NULL, crm_err("No notify type value found for %s", action->uuid)); CRM_CHECK(n_task != NULL, crm_err("No notify operation value found for %s", action->uuid)); clone_key = pcmk__notify_key(action->rsc->clone_name, n_type, n_task); } else if(action->cancel_task) { clone_key = pcmk__op_key(action->rsc->clone_name, action->cancel_task, interval_ms); } else { clone_key = pcmk__op_key(action->rsc->clone_name, action->task, interval_ms); } CRM_CHECK(clone_key != NULL, crm_err("Could not generate a key for %s", action->uuid)); 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) { pe_node_t *router_node = get_router_node(action); crm_xml_add(action_xml, XML_LRM_ATTR_TARGET, action->node->details->uname); crm_xml_add(action_xml, XML_LRM_ATTR_TARGET_UUID, action->node->details->id); if (router_node) { crm_xml_add(action_xml, XML_LRM_ATTR_ROUTER_NODE, router_node->details->uname); } 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)); } /* No details if this action is only being listed in the inputs section */ if (as_input) { return action_xml; } if (action->rsc && !pcmk_is_set(action->flags, pe_action_pseudo)) { int lpc = 0; 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 (should_lock_action(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, crm_element_name(action->rsc->xml)); if (pcmk_is_set(action->rsc->flags, pe_rsc_orphan) && action->rsc->clone_name) { /* Do not use the 'instance free' name here as that * might interfere with the instance we plan to keep. * Ie. if there are more than two named /anonymous/ * instances on a given node, we need to make sure the * command goes to the right one. * * Keep this block, even when everyone is using * 'instance free' anonymous clone names - it means * we'll do the right thing if anyone toggles the * unique flag 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, pe_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 && !pcmk__str_eq(xml_id, action->rsc->clone_name, pcmk__str_casei)) { 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 (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])); } } /* List any attributes in effect */ 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) { // Get the resource instance attributes, evaluated properly for node GHashTable *params = pe_rsc_params(action->rsc, action->node, data_set); /* REMOTE_CONTAINER_HACK: If this is a remote connection resource with * addr="#uname", pull the actual value from the parameters evaluated * without a node (which was put there earlier in stage8() when the * bundle's expand() method was called). */ const char *remote_addr = g_hash_table_lookup(params, XML_RSC_ATTR_REMOTE_RA_ADDR); if (pcmk__str_eq(remote_addr, "#uname", pcmk__str_none)) { GHashTable *base = pe_rsc_params(action->rsc, NULL, data_set); remote_addr = g_hash_table_lookup(base, XML_RSC_ATTR_REMOTE_RA_ADDR); if (remote_addr != NULL) { g_hash_table_insert(params, strdup(XML_RSC_ATTR_REMOTE_RA_ADDR), strdup(remote_addr)); } } g_hash_table_foreach(params, hash2smartfield, args_xml); #if ENABLE_VERSIONED_ATTRS { xmlNode *versioned_parameters = create_xml_node(NULL, XML_TAG_RSC_VER_ATTRS); pe_get_versioned_attributes(versioned_parameters, action->rsc, action->node, data_set); if (xml_has_children(versioned_parameters)) { add_node_copy(action_xml, versioned_parameters); } free_xml(versioned_parameters); } #endif } else if(action->rsc && action->rsc->variant <= pe_native) { GHashTable *params = pe_rsc_params(action->rsc, NULL, data_set); g_hash_table_foreach(params, hash2smartfield, args_xml); #if ENABLE_VERSIONED_ATTRS if (xml_has_children(action->rsc->versioned_parameters)) { add_node_copy(action_xml, action->rsc->versioned_parameters); } #endif } #if ENABLE_VERSIONED_ATTRS if (rsc_details) { if (xml_has_children(rsc_details->versioned_parameters)) { add_node_copy(action_xml, rsc_details->versioned_parameters); } if (xml_has_children(rsc_details->versioned_meta)) { add_node_copy(action_xml, rsc_details->versioned_meta); } } #endif g_hash_table_foreach(action->meta, hash2metafield, args_xml); if (action->rsc != NULL) { const char *value = g_hash_table_lookup(action->rsc->meta, "external-ip"); pe_resource_t *parent = action->rsc; while (parent != NULL) { parent->cmds->append_meta(parent, args_xml); parent = parent->parent; } if(value) { hash2smartfield((gpointer)"pcmk_external_ip", (gpointer)value, (gpointer)args_xml); } if (action->node && /* make clang analyzer happy */ pe__is_guest_node(action->node)) { pe_node_t *host = NULL; enum action_tasks task = text2task(action->task); if(task == action_notify || task == action_notified) { const char *n_task = g_hash_table_lookup(action->meta, "notify_operation"); task = text2task(n_task); } // Differentiate between up and down actions switch (task) { case stop_rsc: case stopped_rsc: case action_demote: case action_demoted: host = pe__current_node(action->node->details->remote_rsc->container); break; case start_rsc: case started_rsc: case monitor_rsc: case action_promote: case action_promoted: host = action->node->details->remote_rsc->container->allocated_to; break; default: break; } if(host) { hash2metafield((gpointer)XML_RSC_ATTR_TARGET, (gpointer)g_hash_table_lookup(action->rsc->meta, XML_RSC_ATTR_TARGET), (gpointer)args_xml); hash2metafield((gpointer) PCMK__ENV_PHYSICAL_HOST, (gpointer)host->details->uname, (gpointer)args_xml); } } } else if (pcmk__str_eq(action->task, CRM_OP_FENCE, pcmk__str_casei) && action->node) { /* 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); /* List any nodes this action is expected to make down */ if (needs_node_info && (action->node != NULL)) { add_downed_nodes(action_xml, action, data_set); } if (needs_maintenance_info) { add_maintenance_nodes(action_xml, data_set); } crm_log_xml_trace(action_xml, "dumped action"); return action_xml; } static bool should_dump_action(pe_action_t *action) { CRM_CHECK(action != NULL, return false); if (pcmk_is_set(action->flags, pe_action_dumped)) { crm_trace("Action %s (%d) already dumped", action->uuid, action->id); return false; } else if (pcmk_is_set(action->flags, pe_action_pseudo) && pcmk__str_eq(action->task, CRM_OP_PROBED, pcmk__str_casei)) { GList *lpc = NULL; /* This is a horrible but convenient hack * * It mimimizes the number of actions with unsatisfied inputs * (i.e. not included in the graph) * * This in turn, means we can be more concise when printing * aborted/incomplete graphs. * * It also makes it obvious which node is preventing * probe_complete from running (presumably because it is only * partially up) * * For these reasons we tolerate such perversions */ for (lpc = action->actions_after; lpc != NULL; lpc = lpc->next) { pe_action_wrapper_t *wrapper = (pe_action_wrapper_t *) lpc->data; if (!pcmk_is_set(wrapper->action->flags, pe_action_runnable)) { /* Only interested in runnable operations */ } else if (!pcmk__str_eq(wrapper->action->task, RSC_START, pcmk__str_casei)) { /* Only interested in start operations */ } else if (pcmk_is_set(wrapper->action->flags, pe_action_dumped) || should_dump_action(wrapper->action)) { crm_trace("Action %s (%d) should be dumped: " "dependency of %s (%d)", action->uuid, action->id, wrapper->action->uuid, wrapper->action->id); return true; } } } if (!pcmk_is_set(action->flags, pe_action_runnable)) { crm_trace("Ignoring action %s (%d): unrunnable", action->uuid, action->id); return false; } else if (pcmk_is_set(action->flags, pe_action_optional) && !pcmk_is_set(action->flags, pe_action_print_always)) { crm_trace("Ignoring action %s (%d): optional", action->uuid, action->id); return false; // Monitors should be dumped even for unmanaged resources } else if (action->rsc && !pcmk_is_set(action->rsc->flags, pe_rsc_managed) && !pcmk__str_eq(action->task, RSC_STATUS, pcmk__str_casei)) { const char *interval_ms_s = g_hash_table_lookup(action->meta, XML_LRM_ATTR_INTERVAL_MS); // Cancellation of recurring monitors should still be dumped 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; } } if (pcmk_is_set(action->flags, pe_action_pseudo) || pcmk__strcase_any_of(action->task, CRM_OP_FENCE, CRM_OP_SHUTDOWN, NULL)) { /* skip the next checks */ return true; } if (action->node == NULL) { pe_err("Skipping action %s (%d) " "because it was not allocated to a node (bug?)", action->uuid, action->id); log_action(LOG_DEBUG, "Unallocated action", action, false); return false; } else if (pcmk_is_set(action->flags, pe_action_dc)) { crm_trace("Action %s (%d) should be dumped: " "can run on DC instead of %s", action->uuid, action->id, action->node->details->uname); } 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 node %s", action->uuid, action->id, action->node->details->uname); } else if (action->node->details->online == false) { pe_err("Skipping action %s (%d) " "because it was scheduled for offline node (bug?)", action->uuid, action->id); log_action(LOG_DEBUG, "Action for offline node", action, FALSE); return false; #if 0 /* but this would also affect resources that can be safely * migrated before a fencing op */ } else if (action->node->details->unclean == false) { pe_err("Skipping action %s (%d) " "because it was scheduled for unclean node (bug?)", action->uuid, action->id); log_action(LOG_DEBUG, "Action for unclean node", action, false); return false; #endif } return true; } /* lowest to highest */ static gint sort_action_id(gconstpointer a, gconstpointer b) { const pe_action_wrapper_t *action_wrapper2 = (const pe_action_wrapper_t *)a; const pe_action_wrapper_t *action_wrapper1 = (const pe_action_wrapper_t *)b; if (a == NULL) { return 1; } if (b == NULL) { return -1; } if (action_wrapper1->action->id > action_wrapper2->action->id) { return -1; } if (action_wrapper1->action->id < action_wrapper2->action->id) { return 1; } return 0; } /*! * \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(pe_action_wrapper_t *ordering) { return pcmk_any_flags_set(ordering->type, ~(pe_order_implies_first_printed |pe_order_implies_then_printed |pe_order_optional)); } /*! * \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 check_dump_input(pe_action_t *action, pe_action_wrapper_t *input) { if (input->state == pe_link_dumped) { return true; } if (input->type == pe_order_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, pe_action_runnable) && !ordering_can_change_actions(input) && !pcmk__str_eq(input->action->uuid, CRM_OP_PROBED, pcmk__str_casei)) { 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, pe_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(action->flags, pe_action_pseudo) && pcmk_is_set(input->type, pe_order_stonith_stop)) { crm_trace("Ignoring %s (%d) input %s (%d): " "stonith stop but action is pseudo", action->uuid, action->id, input->action->uuid, input->action->id); return false; } else if (pcmk_is_set(input->type, pe_order_implies_first_migratable) && !pcmk_is_set(input->action->flags, pe_action_runnable)) { crm_trace("Ignoring %s (%d) input %s (%d): " "implies input migratable but input unrunnable", action->uuid, action->id, input->action->uuid, input->action->id); return false; } else if (pcmk_is_set(input->type, pe_order_apply_first_non_migratable) && pcmk_is_set(input->action->flags, pe_action_migrate_runnable)) { crm_trace("Ignoring %s (%d) input %s (%d): " "only if input unmigratable but input unrunnable", action->uuid, action->id, input->action->uuid, input->action->id); return false; } else if ((input->type == pe_order_optional) && pcmk_is_set(input->action->flags, pe_action_migrate_runnable) && 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) { pe_node_t *input_node = input->action->node; // load orderings are relevant only if actions are for same node if (action->rsc && pcmk__str_eq(action->task, RSC_MIGRATE, pcmk__str_casei)) { pe_node_t *allocated = action->rsc->allocated_to; /* For load_stopped -> migrate_to orderings, we care about where it * has been allocated to, not where it will be executed. */ if ((input_node == NULL) || (allocated == NULL) || (input_node->details != allocated->details)) { crm_trace("Ignoring %s (%d) input %s (%d): " "load ordering node mismatch %s vs %s", action->uuid, action->id, input->action->uuid, input->action->id, (allocated? allocated->details->uname : ""), (input_node? input_node->details->uname : "")); input->type = pe_order_none; return false; } } else if ((input_node == NULL) || (action->node == NULL) || (input_node->details != action->node->details)) { crm_trace("Ignoring %s (%d) input %s (%d): " "load ordering node mismatch %s vs %s", action->uuid, action->id, input->action->uuid, input->action->id, (action->node? action->node->details->uname : ""), (input_node? input_node->details->uname : "")); input->type = pe_order_none; return false; } else if (pcmk_is_set(input->action->flags, pe_action_optional)) { crm_trace("Ignoring %s (%d) input %s (%d): " "load ordering input optional", action->uuid, action->id, input->action->uuid, input->action->id); input->type = pe_order_none; return false; } } else if (input->type == pe_order_anti_colocation) { if (input->action->node && action->node && (input->action->node->details != action->node->details)) { 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, action->node->details->uname, input->action->node->details->uname); input->type = pe_order_none; return false; } else if (pcmk_is_set(input->action->flags, pe_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 = pe_order_none; return false; } } else if (input->action->rsc && input->action->rsc != action->rsc && pcmk_is_set(input->action->rsc->flags, pe_rsc_failed) && !pcmk_is_set(input->action->rsc->flags, pe_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, pe_action_optional) && !pcmk_any_flags_set(input->action->flags, pe_action_print_always|pe_action_dumped) && !should_dump_action(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 0x%.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; } static bool graph_has_loop(pe_action_t *init_action, pe_action_t *action, pe_action_wrapper_t *input) { bool has_loop = false; if (pcmk_is_set(input->action->flags, pe_action_tracking)) { crm_trace("Breaking tracking loop: %s@%s -> %s@%s (0x%.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 (!check_dump_input(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, pe_action_tracking); crm_trace("Checking inputs of action %s@%s input %s@%s (0x%.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 (graph_has_loop(init_action, input->action, (pe_action_wrapper_t *) iter->data)) { // Recursive call already logged a debug message has_loop = true; goto done; } } done: pe__clear_action_flags(input->action, pe_action_tracking); if (!has_loop) { crm_trace("No input loop found in %s@%s -> %s@%s (0x%.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; } bool pcmk__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) && action->rsc && action->rsc->fillers && input->action->rsc && input->action->node && input->action->node->details->remote_rsc && (input->action->node->details->remote_rsc->container == action->rsc)) { 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 && pcmk__str_eq(action->task, RSC_MIGRATE, pcmk__str_casei) && graph_has_loop(action, action, input)) { return true; } return false; } // Remove duplicate inputs (regardless of flags) static void deduplicate_inputs(pe_action_t *action) { GList *item = NULL; GList *next = NULL; pe_action_wrapper_t *last_input = NULL; action->actions_before = g_list_sort(action->actions_before, sort_action_id); for (item = action->actions_before; item != NULL; item = next) { pe_action_wrapper_t *input = (pe_action_wrapper_t *) item->data; next = item->next; if (last_input && (input->action->id == last_input->action->id)) { crm_trace("Input %s (%d) duplicate skipped for action %s (%d)", input->action->uuid, input->action->id, action->uuid, action->id); /* For the purposes of scheduling, the ordering flags no longer * matter, but crm_simulate looks at certain ones when creating a * dot graph. Combining the flags is sufficient for that purpose. */ last_input->type |= input->type; if (input->state == pe_link_dumped) { last_input->state = pe_link_dumped; } free(item->data); action->actions_before = g_list_delete_link(action->actions_before, item); } else { last_input = input; input->state = pe_link_not_dumped; } } } /*! * \internal * \brief Add an action to the transition graph XML if appropriate * * \param[in] action Action to possibly add * \param[in] data_set 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 stage7() * 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 stage8().) */ void graph_element_from_action(pe_action_t *action, pe_working_set_t *data_set) { GList *lpc = NULL; int synapse_priority = 0; xmlNode *syn = NULL; xmlNode *set = NULL; xmlNode *in = NULL; xmlNode *xml_action = NULL; pe_action_wrapper_t *input = NULL; /* If we haven't already, de-duplicate inputs -- even if we won't be dumping * the action, so that crm_simulate dot graphs don't have duplicates. */ if (!pcmk_is_set(action->flags, pe_action_dedup)) { deduplicate_inputs(action); pe__set_action_flags(action, pe_action_dedup); } if (should_dump_action(action) == FALSE) { return; } pe__set_action_flags(action, pe_action_dumped); syn = create_xml_node(data_set->graph, "synapse"); set = create_xml_node(syn, "action_set"); in = create_xml_node(syn, "inputs"); 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); } xml_action = action2xml(action, FALSE, data_set); add_node_nocopy(set, crm_element_name(xml_action), xml_action); for (lpc = action->actions_before; lpc != NULL; lpc = lpc->next) { input = (pe_action_wrapper_t *) lpc->data; if (check_dump_input(action, input)) { xmlNode *input_xml = create_xml_node(in, "trigger"); input->state = pe_link_dumped; xml_action = action2xml(input->action, TRUE, data_set); add_node_nocopy(input_xml, crm_element_name(xml_action), xml_action); } } } diff --git a/lib/pacemaker/pcmk_sched_allocate.c b/lib/pacemaker/pcmk_sched_allocate.c index f46474b5ca..bc111d68df 100644 --- a/lib/pacemaker/pcmk_sched_allocate.c +++ b/lib/pacemaker/pcmk_sched_allocate.c @@ -1,2987 +1,2902 @@ /* * Copyright 2004-2021 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 #include #include #include #include #include #include #include #include #include +#include "libpacemaker_private.h" CRM_TRACE_INIT_DATA(pacemaker); extern bool pcmk__is_daemon; void set_alloc_actions(pe_working_set_t * data_set); extern void ReloadRsc(pe_resource_t * rsc, pe_node_t *node, pe_working_set_t * data_set); extern gboolean DeleteRsc(pe_resource_t * rsc, pe_node_t * node, gboolean optional, pe_working_set_t * data_set); static void apply_remote_node_ordering(pe_working_set_t *data_set); static enum remote_connection_state get_remote_node_state(pe_node_t *node); enum remote_connection_state { remote_state_unknown = 0, remote_state_alive = 1, remote_state_resting = 2, remote_state_failed = 3, remote_state_stopped = 4 }; static const char * state2text(enum remote_connection_state state) { switch (state) { case remote_state_unknown: return "unknown"; case remote_state_alive: return "alive"; case remote_state_resting: return "resting"; case remote_state_failed: return "failed"; case remote_state_stopped: return "stopped"; } return "impossible"; } resource_alloc_functions_t resource_class_alloc_functions[] = { { pcmk__native_merge_weights, pcmk__native_allocate, native_create_actions, native_create_probe, native_internal_constraints, native_rsc_colocation_lh, native_rsc_colocation_rh, native_rsc_location, native_action_flags, native_update_actions, native_expand, native_append_meta, }, { pcmk__group_merge_weights, pcmk__group_allocate, group_create_actions, native_create_probe, group_internal_constraints, group_rsc_colocation_lh, group_rsc_colocation_rh, group_rsc_location, group_action_flags, group_update_actions, group_expand, group_append_meta, }, { pcmk__native_merge_weights, pcmk__clone_allocate, clone_create_actions, clone_create_probe, clone_internal_constraints, clone_rsc_colocation_lh, clone_rsc_colocation_rh, clone_rsc_location, clone_action_flags, pcmk__multi_update_actions, clone_expand, clone_append_meta, }, { pcmk__native_merge_weights, pcmk__bundle_allocate, pcmk__bundle_create_actions, pcmk__bundle_create_probe, pcmk__bundle_internal_constraints, pcmk__bundle_rsc_colocation_lh, pcmk__bundle_rsc_colocation_rh, pcmk__bundle_rsc_location, pcmk__bundle_action_flags, pcmk__multi_update_actions, pcmk__bundle_expand, pcmk__bundle_append_meta, } }; static gboolean check_rsc_parameters(pe_resource_t * rsc, pe_node_t * node, xmlNode * rsc_entry, gboolean active_here, pe_working_set_t * data_set) { int attr_lpc = 0; gboolean force_restart = FALSE; gboolean delete_resource = FALSE; gboolean changed = FALSE; const char *value = NULL; const char *old_value = NULL; const char *attr_list[] = { XML_ATTR_TYPE, XML_AGENT_ATTR_CLASS, XML_AGENT_ATTR_PROVIDER }; for (; attr_lpc < PCMK__NELEM(attr_list); attr_lpc++) { value = crm_element_value(rsc->xml, attr_list[attr_lpc]); old_value = crm_element_value(rsc_entry, attr_list[attr_lpc]); if (value == old_value /* i.e. NULL */ || pcmk__str_eq(value, old_value, pcmk__str_none)) { continue; } changed = TRUE; trigger_unfencing(rsc, node, "Device definition changed", NULL, data_set); if (active_here) { force_restart = TRUE; crm_notice("Forcing restart of %s on %s, %s changed: %s -> %s", rsc->id, node->details->uname, attr_list[attr_lpc], crm_str(old_value), crm_str(value)); } } if (force_restart) { /* make sure the restart happens */ stop_action(rsc, node, FALSE); pe__set_resource_flags(rsc, pe_rsc_start_pending); delete_resource = TRUE; } else if (changed) { delete_resource = TRUE; } return delete_resource; } static void CancelXmlOp(pe_resource_t * rsc, xmlNode * xml_op, pe_node_t * active_node, const char *reason, pe_working_set_t * data_set) { guint interval_ms = 0; pe_action_t *cancel = NULL; const char *task = NULL; const char *call_id = NULL; CRM_CHECK(xml_op != NULL, return); CRM_CHECK(active_node != NULL, return); task = crm_element_value(xml_op, XML_LRM_ATTR_TASK); call_id = crm_element_value(xml_op, XML_LRM_ATTR_CALLID); crm_element_value_ms(xml_op, XML_LRM_ATTR_INTERVAL_MS, &interval_ms); crm_info("Action " PCMK__OP_FMT " on %s will be stopped: %s", rsc->id, task, interval_ms, active_node->details->uname, (reason? reason : "unknown")); cancel = pe_cancel_op(rsc, task, interval_ms, active_node, data_set); add_hash_param(cancel->meta, XML_LRM_ATTR_CALLID, call_id); custom_action_order(rsc, stop_key(rsc), NULL, rsc, NULL, cancel, pe_order_optional, data_set); } static gboolean check_action_definition(pe_resource_t * rsc, pe_node_t * active_node, xmlNode * xml_op, pe_working_set_t * data_set) { char *key = NULL; guint interval_ms = 0; const op_digest_cache_t *digest_data = NULL; gboolean did_change = FALSE; const char *task = crm_element_value(xml_op, XML_LRM_ATTR_TASK); const char *digest_secure = NULL; CRM_CHECK(active_node != NULL, return FALSE); crm_element_value_ms(xml_op, XML_LRM_ATTR_INTERVAL_MS, &interval_ms); if (interval_ms > 0) { xmlNode *op_match = NULL; /* we need to reconstruct the key because of the way we used to construct resource IDs */ key = pcmk__op_key(rsc->id, task, interval_ms); pe_rsc_trace(rsc, "Checking parameters for %s", key); op_match = find_rsc_op_entry(rsc, key); if ((op_match == NULL) && pcmk_is_set(data_set->flags, pe_flag_stop_action_orphans)) { CancelXmlOp(rsc, xml_op, active_node, "orphan", data_set); free(key); return TRUE; } else if (op_match == NULL) { pe_rsc_debug(rsc, "Orphan action detected: %s on %s", key, active_node->details->uname); free(key); return TRUE; } free(key); key = NULL; } crm_trace("Testing " PCMK__OP_FMT " on %s", rsc->id, task, interval_ms, active_node->details->uname); if ((interval_ms == 0) && pcmk__str_eq(task, RSC_STATUS, pcmk__str_casei)) { /* Reload based on the start action not a probe */ task = RSC_START; } else if ((interval_ms == 0) && pcmk__str_eq(task, RSC_MIGRATED, pcmk__str_casei)) { /* Reload based on the start action not a migrate */ task = RSC_START; } else if ((interval_ms == 0) && pcmk__str_eq(task, RSC_PROMOTE, pcmk__str_casei)) { /* Reload based on the start action not a promote */ task = RSC_START; } digest_data = rsc_action_digest_cmp(rsc, xml_op, active_node, data_set); if (pcmk_is_set(data_set->flags, pe_flag_sanitized)) { digest_secure = crm_element_value(xml_op, XML_LRM_ATTR_SECURE_DIGEST); } if(digest_data->rc != RSC_DIGEST_MATCH && digest_secure && digest_data->digest_secure_calc && strcmp(digest_data->digest_secure_calc, digest_secure) == 0) { if (!pcmk__is_daemon && data_set->priv != NULL) { pcmk__output_t *out = data_set->priv; out->info(out, "Only 'private' parameters to " PCMK__OP_FMT " on %s changed: %s", rsc->id, task, interval_ms, active_node->details->uname, crm_element_value(xml_op, XML_ATTR_TRANSITION_MAGIC)); } } else if (digest_data->rc == RSC_DIGEST_RESTART) { /* Changes that force a restart */ pe_action_t *required = NULL; did_change = TRUE; key = pcmk__op_key(rsc->id, task, interval_ms); crm_log_xml_info(digest_data->params_restart, "params:restart"); required = custom_action(rsc, key, task, NULL, FALSE, TRUE, data_set); pe_action_set_reason(required, "resource definition change", true); trigger_unfencing(rsc, active_node, "Device parameters changed", NULL, data_set); } else if ((digest_data->rc == RSC_DIGEST_ALL) || (digest_data->rc == RSC_DIGEST_UNKNOWN)) { // Changes that can potentially be handled by an agent reload const char *digest_restart = crm_element_value(xml_op, XML_LRM_ATTR_RESTART_DIGEST); did_change = TRUE; trigger_unfencing(rsc, active_node, "Device parameters changed (reload)", NULL, data_set); crm_log_xml_info(digest_data->params_all, "params:reload"); key = pcmk__op_key(rsc->id, task, interval_ms); if (interval_ms > 0) { pe_action_t *op = NULL; #if 0 /* Always reload/restart the entire resource */ ReloadRsc(rsc, active_node, data_set); #else /* Re-sending the recurring op is sufficient - the old one will be cancelled automatically */ op = custom_action(rsc, key, task, active_node, TRUE, TRUE, data_set); pe__set_action_flags(op, pe_action_reschedule); #endif } else if (digest_restart) { pe_rsc_trace(rsc, "Reloading '%s' action for resource %s", task, rsc->id); /* Reload this resource */ ReloadRsc(rsc, active_node, data_set); free(key); } else { pe_action_t *required = NULL; pe_rsc_trace(rsc, "Resource %s doesn't support agent reloads", rsc->id); /* Re-send the start/demote/promote op * Recurring ops will be detected independently */ required = custom_action(rsc, key, task, NULL, FALSE, TRUE, data_set); pe_action_set_reason(required, "resource definition change", true); } } return did_change; } /*! * \internal * \brief Do deferred action checks after allocation * * \param[in] data_set Working set for cluster */ static void check_params(pe_resource_t *rsc, pe_node_t *node, xmlNode *rsc_op, enum pe_check_parameters check, pe_working_set_t *data_set) { const char *reason = NULL; op_digest_cache_t *digest_data = NULL; switch (check) { case pe_check_active: if (check_action_definition(rsc, node, rsc_op, data_set) && pe_get_failcount(node, rsc, NULL, pe_fc_effective, NULL, data_set)) { reason = "action definition changed"; } break; case pe_check_last_failure: digest_data = rsc_action_digest_cmp(rsc, rsc_op, node, data_set); switch (digest_data->rc) { case RSC_DIGEST_UNKNOWN: crm_trace("Resource %s history entry %s on %s has no digest to compare", rsc->id, ID(rsc_op), node->details->id); break; case RSC_DIGEST_MATCH: break; default: reason = "resource parameters have changed"; break; } break; } if (reason) { pe__clear_failcount(rsc, node, reason, data_set); } } static void check_actions_for(xmlNode * rsc_entry, pe_resource_t * rsc, pe_node_t * node, pe_working_set_t * data_set) { GList *gIter = NULL; int offset = -1; int stop_index = 0; int start_index = 0; const char *task = NULL; xmlNode *rsc_op = NULL; GList *op_list = NULL; GList *sorted_op_list = NULL; CRM_CHECK(node != NULL, return); if (pcmk_is_set(rsc->flags, pe_rsc_orphan)) { pe_resource_t *parent = uber_parent(rsc); if(parent == NULL || pe_rsc_is_clone(parent) == FALSE || pcmk_is_set(parent->flags, pe_rsc_unique)) { pe_rsc_trace(rsc, "Skipping param check for %s and deleting: orphan", rsc->id); DeleteRsc(rsc, node, FALSE, data_set); } else { pe_rsc_trace(rsc, "Skipping param check for %s (orphan clone)", rsc->id); } return; } else if (pe_find_node_id(rsc->running_on, node->details->id) == NULL) { if (check_rsc_parameters(rsc, node, rsc_entry, FALSE, data_set)) { DeleteRsc(rsc, node, FALSE, data_set); } pe_rsc_trace(rsc, "Skipping param check for %s: no longer active on %s", rsc->id, node->details->uname); return; } pe_rsc_trace(rsc, "Processing %s on %s", rsc->id, node->details->uname); if (check_rsc_parameters(rsc, node, rsc_entry, TRUE, data_set)) { DeleteRsc(rsc, node, FALSE, data_set); } for (rsc_op = pcmk__xe_first_child(rsc_entry); rsc_op != NULL; rsc_op = pcmk__xe_next(rsc_op)) { if (pcmk__str_eq((const char *)rsc_op->name, XML_LRM_TAG_RSC_OP, pcmk__str_none)) { op_list = g_list_prepend(op_list, rsc_op); } } sorted_op_list = g_list_sort(op_list, sort_op_by_callid); calculate_active_ops(sorted_op_list, &start_index, &stop_index); for (gIter = sorted_op_list; gIter != NULL; gIter = gIter->next) { xmlNode *rsc_op = (xmlNode *) gIter->data; guint interval_ms = 0; offset++; if (start_index < stop_index) { /* stopped */ continue; } else if (offset < start_index) { /* action occurred prior to a start */ continue; } task = crm_element_value(rsc_op, XML_LRM_ATTR_TASK); crm_element_value_ms(rsc_op, XML_LRM_ATTR_INTERVAL_MS, &interval_ms); if ((interval_ms > 0) && (pcmk_is_set(rsc->flags, pe_rsc_maintenance) || node->details->maintenance)) { // Maintenance mode cancels recurring operations CancelXmlOp(rsc, rsc_op, node, "maintenance mode", data_set); } else if ((interval_ms > 0) || pcmk__strcase_any_of(task, RSC_STATUS, RSC_START, RSC_PROMOTE, RSC_MIGRATED, NULL)) { /* If a resource operation failed, and the operation's definition * has changed, clear any fail count so they can be retried fresh. */ if (pe__bundle_needs_remote_name(rsc, data_set)) { /* We haven't allocated resources to nodes yet, so if the * REMOTE_CONTAINER_HACK is used, we may calculate the digest * based on the literal "#uname" value rather than the properly * substituted value. That would mistakenly make the action * definition appear to have been changed. Defer the check until * later in this case. */ pe__add_param_check(rsc_op, rsc, node, pe_check_active, data_set); } else if (check_action_definition(rsc, node, rsc_op, data_set) && pe_get_failcount(node, rsc, NULL, pe_fc_effective, NULL, data_set)) { pe__clear_failcount(rsc, node, "action definition changed", data_set); } } } g_list_free(sorted_op_list); } static GList * find_rsc_list(GList *result, pe_resource_t * rsc, const char *id, gboolean renamed_clones, gboolean partial, pe_working_set_t * data_set) { GList *gIter = NULL; gboolean match = FALSE; if (id == NULL) { return NULL; } if (rsc == NULL) { if (data_set == NULL) { return NULL; } for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) { pe_resource_t *child = (pe_resource_t *) gIter->data; result = find_rsc_list(result, child, id, renamed_clones, partial, NULL); } return result; } if (partial) { if (strstr(rsc->id, id)) { match = TRUE; } else if (renamed_clones && rsc->clone_name && strstr(rsc->clone_name, id)) { match = TRUE; } } else { if (strcmp(rsc->id, id) == 0) { match = TRUE; } else if (renamed_clones && rsc->clone_name && strcmp(rsc->clone_name, id) == 0) { match = TRUE; } } if (match) { result = g_list_prepend(result, rsc); } if (rsc->children) { gIter = rsc->children; for (; gIter != NULL; gIter = gIter->next) { pe_resource_t *child = (pe_resource_t *) gIter->data; result = find_rsc_list(result, child, id, renamed_clones, partial, NULL); } } return result; } static void check_actions(pe_working_set_t * data_set) { const char *id = NULL; pe_node_t *node = NULL; xmlNode *lrm_rscs = NULL; xmlNode *status = get_object_root(XML_CIB_TAG_STATUS, data_set->input); xmlNode *node_state = NULL; for (node_state = pcmk__xe_first_child(status); node_state != NULL; node_state = pcmk__xe_next(node_state)) { if (pcmk__str_eq((const char *)node_state->name, XML_CIB_TAG_STATE, pcmk__str_none)) { id = crm_element_value(node_state, XML_ATTR_ID); lrm_rscs = find_xml_node(node_state, XML_CIB_TAG_LRM, FALSE); lrm_rscs = find_xml_node(lrm_rscs, XML_LRM_TAG_RESOURCES, FALSE); node = pe_find_node_id(data_set->nodes, id); if (node == NULL) { continue; /* Still need to check actions for a maintenance node to cancel existing monitor operations */ } else if (can_run_resources(node) == FALSE && node->details->maintenance == FALSE) { crm_trace("Skipping param check for %s: can't run resources", node->details->uname); continue; } crm_trace("Processing node %s", node->details->uname); if (node->details->online || pcmk_is_set(data_set->flags, pe_flag_stonith_enabled)) { xmlNode *rsc_entry = NULL; for (rsc_entry = pcmk__xe_first_child(lrm_rscs); rsc_entry != NULL; rsc_entry = pcmk__xe_next(rsc_entry)) { if (pcmk__str_eq((const char *)rsc_entry->name, XML_LRM_TAG_RESOURCE, pcmk__str_none)) { if (xml_has_children(rsc_entry)) { GList *gIter = NULL; GList *result = NULL; const char *rsc_id = ID(rsc_entry); CRM_CHECK(rsc_id != NULL, return); result = find_rsc_list(NULL, NULL, rsc_id, TRUE, FALSE, data_set); for (gIter = result; gIter != NULL; gIter = gIter->next) { pe_resource_t *rsc = (pe_resource_t *) gIter->data; if (rsc->variant != pe_native) { continue; } check_actions_for(rsc_entry, rsc, node, data_set); } g_list_free(result); } } } } } } } static void apply_placement_constraints(pe_working_set_t * data_set) { for (GList *gIter = data_set->placement_constraints; gIter != NULL; gIter = gIter->next) { pe__location_t *cons = gIter->data; cons->rsc_lh->cmds->rsc_location(cons->rsc_lh, cons); } } static gboolean failcount_clear_action_exists(pe_node_t * node, pe_resource_t * rsc) { gboolean rc = FALSE; GList *list = pe__resource_actions(rsc, node, CRM_OP_CLEAR_FAILCOUNT, TRUE); if (list) { rc = TRUE; } g_list_free(list); return rc; } static void common_apply_stickiness(pe_resource_t * rsc, pe_node_t * node, pe_working_set_t * data_set) { if (rsc->children) { GList *gIter = rsc->children; for (; gIter != NULL; gIter = gIter->next) { pe_resource_t *child_rsc = (pe_resource_t *) gIter->data; common_apply_stickiness(child_rsc, node, data_set); } return; } if (pcmk_is_set(rsc->flags, pe_rsc_managed) && rsc->stickiness != 0 && pcmk__list_of_1(rsc->running_on)) { pe_node_t *current = pe_find_node_id(rsc->running_on, node->details->id); pe_node_t *match = pe_hash_table_lookup(rsc->allowed_nodes, node->details->id); if (current == NULL) { } else if ((match != NULL) || pcmk_is_set(data_set->flags, pe_flag_symmetric_cluster)) { pe_resource_t *sticky_rsc = rsc; resource_location(sticky_rsc, node, rsc->stickiness, "stickiness", data_set); pe_rsc_debug(sticky_rsc, "Resource %s: preferring current location" " (node=%s, weight=%d)", sticky_rsc->id, node->details->uname, rsc->stickiness); } else { GHashTableIter iter; pe_node_t *nIter = NULL; pe_rsc_debug(rsc, "Ignoring stickiness for %s: the cluster is asymmetric" " and node %s is not explicitly allowed", rsc->id, node->details->uname); g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **)&nIter)) { crm_err("%s[%s] = %d", rsc->id, nIter->details->uname, nIter->weight); } } } /* Check the migration threshold only if a failcount clear action * has not already been placed for this resource on the node. * There is no sense in potentially forcing the resource from this * node if the failcount is being reset anyway. * * @TODO A clear_failcount operation can be scheduled in stage4() via * check_actions_for(), or in stage5() via check_params(). This runs in * stage2(), so it cannot detect those, meaning we might check the migration * threshold when we shouldn't -- worst case, we stop or move the resource, * then move it back next transition. */ if (failcount_clear_action_exists(node, rsc) == FALSE) { pe_resource_t *failed = NULL; if (pcmk__threshold_reached(rsc, node, data_set, &failed)) { resource_location(failed, node, -INFINITY, "__fail_limit__", data_set); } } } void complex_set_cmds(pe_resource_t * rsc) { GList *gIter = rsc->children; rsc->cmds = &resource_class_alloc_functions[rsc->variant]; for (; gIter != NULL; gIter = gIter->next) { pe_resource_t *child_rsc = (pe_resource_t *) gIter->data; complex_set_cmds(child_rsc); } } void set_alloc_actions(pe_working_set_t * data_set) { GList *gIter = data_set->resources; for (; gIter != NULL; gIter = gIter->next) { pe_resource_t *rsc = (pe_resource_t *) gIter->data; complex_set_cmds(rsc); } } static void calculate_system_health(gpointer gKey, gpointer gValue, gpointer user_data) { const char *key = (const char *)gKey; const char *value = (const char *)gValue; int *system_health = (int *)user_data; if (!gKey || !gValue || !user_data) { return; } if (pcmk__starts_with(key, "#health")) { int score; /* Convert the value into an integer */ score = char2score(value); /* Add it to the running total */ *system_health = pe__add_scores(score, *system_health); } } static gboolean apply_system_health(pe_working_set_t * data_set) { GList *gIter = NULL; const char *health_strategy = pe_pref(data_set->config_hash, "node-health-strategy"); int base_health = 0; if (pcmk__str_eq(health_strategy, "none", pcmk__str_null_matches | pcmk__str_casei)) { /* Prevent any accidental health -> score translation */ pcmk__score_red = 0; pcmk__score_yellow = 0; pcmk__score_green = 0; return TRUE; } else if (pcmk__str_eq(health_strategy, "migrate-on-red", pcmk__str_casei)) { /* Resources on nodes which have health values of red are * weighted away from that node. */ pcmk__score_red = -INFINITY; pcmk__score_yellow = 0; pcmk__score_green = 0; } else if (pcmk__str_eq(health_strategy, "only-green", pcmk__str_casei)) { /* Resources on nodes which have health values of red or yellow * are forced away from that node. */ pcmk__score_red = -INFINITY; pcmk__score_yellow = -INFINITY; pcmk__score_green = 0; } else if (pcmk__str_eq(health_strategy, "progressive", pcmk__str_casei)) { /* Same as the above, but use the r/y/g scores provided by the user * Defaults are provided by the pe_prefs table * Also, custom health "base score" can be used */ base_health = char2score(pe_pref(data_set->config_hash, "node-health-base")); } else if (pcmk__str_eq(health_strategy, "custom", pcmk__str_casei)) { /* Requires the admin to configure the rsc_location constaints for * processing the stored health scores */ /* TODO: Check for the existence of appropriate node health constraints */ return TRUE; } else { crm_err("Unknown node health strategy: %s", health_strategy); return FALSE; } crm_info("Applying automated node health strategy: %s", health_strategy); for (gIter = data_set->nodes; gIter != NULL; gIter = gIter->next) { int system_health = base_health; pe_node_t *node = (pe_node_t *) gIter->data; /* Search through the node hash table for system health entries. */ g_hash_table_foreach(node->details->attrs, calculate_system_health, &system_health); crm_info(" Node %s has an combined system health of %d", node->details->uname, system_health); /* If the health is non-zero, then create a new rsc2node so that the * weight will be added later on. */ if (system_health != 0) { GList *gIter2 = data_set->resources; for (; gIter2 != NULL; gIter2 = gIter2->next) { pe_resource_t *rsc = (pe_resource_t *) gIter2->data; rsc2node_new(health_strategy, rsc, system_health, NULL, node, data_set); } } } return TRUE; } gboolean stage0(pe_working_set_t * data_set) { xmlNode *cib_constraints = get_object_root(XML_CIB_TAG_CONSTRAINTS, data_set->input); if (data_set->input == NULL) { return FALSE; } if (!pcmk_is_set(data_set->flags, pe_flag_have_status)) { crm_trace("Calculating status"); cluster_status(data_set); } set_alloc_actions(data_set); apply_system_health(data_set); unpack_constraints(cib_constraints, data_set); return TRUE; } /* * Check nodes for resources started outside of the LRM */ gboolean probe_resources(pe_working_set_t * data_set) { pe_action_t *probe_node_complete = NULL; for (GList *gIter = data_set->nodes; gIter != NULL; gIter = gIter->next) { pe_node_t *node = (pe_node_t *) gIter->data; const char *probed = pe_node_attribute_raw(node, CRM_OP_PROBED); if (node->details->online == FALSE) { if (pe__is_remote_node(node) && node->details->remote_rsc && (get_remote_node_state(node) == remote_state_failed)) { pe_fence_node(data_set, node, "the connection is unrecoverable", FALSE); } continue; } else if (node->details->unclean) { continue; } else if (node->details->rsc_discovery_enabled == FALSE) { /* resource discovery is disabled for this node */ continue; } if (probed != NULL && crm_is_true(probed) == FALSE) { pe_action_t *probe_op = custom_action(NULL, crm_strdup_printf("%s-%s", CRM_OP_REPROBE, node->details->uname), CRM_OP_REPROBE, node, FALSE, TRUE, data_set); add_hash_param(probe_op->meta, XML_ATTR_TE_NOWAIT, XML_BOOLEAN_TRUE); continue; } for (GList *gIter2 = data_set->resources; gIter2 != NULL; gIter2 = gIter2->next) { pe_resource_t *rsc = (pe_resource_t *) gIter2->data; rsc->cmds->create_probe(rsc, node, probe_node_complete, FALSE, data_set); } } return TRUE; } static void rsc_discover_filter(pe_resource_t *rsc, pe_node_t *node) { pe_resource_t *top = uber_parent(rsc); pe_node_t *match; if (rsc->exclusive_discover == FALSE && top->exclusive_discover == FALSE) { return; } g_list_foreach(rsc->children, (GFunc) rsc_discover_filter, node); match = g_hash_table_lookup(rsc->allowed_nodes, node->details->id); if (match && match->rsc_discover_mode != pe_discover_exclusive) { match->weight = -INFINITY; } } static time_t shutdown_time(pe_node_t *node, pe_working_set_t *data_set) { const char *shutdown = pe_node_attribute_raw(node, XML_CIB_ATTR_SHUTDOWN); time_t result = 0; if (shutdown) { long long result_ll; if (pcmk__scan_ll(shutdown, &result_ll, 0LL) == pcmk_rc_ok) { result = (time_t) result_ll; } } return result? result : get_effective_time(data_set); } static void apply_shutdown_lock(pe_resource_t *rsc, pe_working_set_t *data_set) { const char *class; // Only primitives and (uncloned) groups may be locked if (rsc->variant == pe_group) { g_list_foreach(rsc->children, (GFunc) apply_shutdown_lock, data_set); } else if (rsc->variant != pe_native) { return; } // Fence devices and remote connections can't be locked class = crm_element_value(rsc->xml, XML_AGENT_ATTR_CLASS); if (pcmk__str_eq(class, PCMK_RESOURCE_CLASS_STONITH, pcmk__str_null_matches) || pe__resource_is_remote_conn(rsc, data_set)) { return; } if (rsc->lock_node != NULL) { // The lock was obtained from resource history if (rsc->running_on != NULL) { /* The resource was started elsewhere even though it is now * considered locked. This shouldn't be possible, but as a * failsafe, we don't want to disturb the resource now. */ pe_rsc_info(rsc, "Cancelling shutdown lock because %s is already active", rsc->id); pe__clear_resource_history(rsc, rsc->lock_node, data_set); rsc->lock_node = NULL; rsc->lock_time = 0; } // Only a resource active on exactly one node can be locked } else if (pcmk__list_of_1(rsc->running_on)) { pe_node_t *node = rsc->running_on->data; if (node->details->shutdown) { if (node->details->unclean) { pe_rsc_debug(rsc, "Not locking %s to unclean %s for shutdown", rsc->id, node->details->uname); } else { rsc->lock_node = node; rsc->lock_time = shutdown_time(node, data_set); } } } if (rsc->lock_node == NULL) { // No lock needed return; } if (data_set->shutdown_lock > 0) { time_t lock_expiration = rsc->lock_time + data_set->shutdown_lock; pe_rsc_info(rsc, "Locking %s to %s due to shutdown (expires @%lld)", rsc->id, rsc->lock_node->details->uname, (long long) lock_expiration); pe__update_recheck_time(++lock_expiration, data_set); } else { pe_rsc_info(rsc, "Locking %s to %s due to shutdown", rsc->id, rsc->lock_node->details->uname); } // If resource is locked to one node, ban it from all other nodes for (GList *item = data_set->nodes; item != NULL; item = item->next) { pe_node_t *node = item->data; if (strcmp(node->details->uname, rsc->lock_node->details->uname)) { resource_location(rsc, node, -CRM_SCORE_INFINITY, XML_CONFIG_ATTR_SHUTDOWN_LOCK, data_set); } } } /* * \internal * \brief Stage 2 of cluster status: apply node-specific criteria * * Count known nodes, and apply location constraints, stickiness, and exclusive * resource discovery. */ gboolean stage2(pe_working_set_t * data_set) { GList *gIter = NULL; if (pcmk_is_set(data_set->flags, pe_flag_shutdown_lock)) { g_list_foreach(data_set->resources, (GFunc) apply_shutdown_lock, data_set); } if (!pcmk_is_set(data_set->flags, pe_flag_no_compat)) { // @COMPAT API backward compatibility for (gIter = data_set->nodes; gIter != NULL; gIter = gIter->next) { pe_node_t *node = (pe_node_t *) gIter->data; if (node && (node->weight >= 0) && node->details->online && (node->details->type != node_ping)) { data_set->max_valid_nodes++; } } } apply_placement_constraints(data_set); gIter = data_set->nodes; for (; gIter != NULL; gIter = gIter->next) { GList *gIter2 = NULL; pe_node_t *node = (pe_node_t *) gIter->data; gIter2 = data_set->resources; for (; gIter2 != NULL; gIter2 = gIter2->next) { pe_resource_t *rsc = (pe_resource_t *) gIter2->data; common_apply_stickiness(rsc, node, data_set); rsc_discover_filter(rsc, node); } } return TRUE; } /* * Create internal resource constraints before allocation */ gboolean stage3(pe_working_set_t * data_set) { GList *gIter = data_set->resources; for (; gIter != NULL; gIter = gIter->next) { pe_resource_t *rsc = (pe_resource_t *) gIter->data; rsc->cmds->internal_constraints(rsc, data_set); } return TRUE; } /* * Check for orphaned or redefined actions */ gboolean stage4(pe_working_set_t * data_set) { check_actions(data_set); return TRUE; } static void * convert_const_pointer(const void *ptr) { /* Worst function ever */ return (void *)ptr; } static gint sort_rsc_process_order(gconstpointer a, gconstpointer b, gpointer data) { int rc = 0; int r1_weight = -INFINITY; int r2_weight = -INFINITY; const char *reason = "existence"; GList *nodes = (GList *) data; const pe_resource_t *resource1 = a; const pe_resource_t *resource2 = b; pe_node_t *r1_node = NULL; pe_node_t *r2_node = NULL; GList *gIter = NULL; GHashTable *r1_nodes = NULL; GHashTable *r2_nodes = NULL; reason = "priority"; r1_weight = resource1->priority; r2_weight = resource2->priority; if (r1_weight > r2_weight) { rc = -1; goto done; } if (r1_weight < r2_weight) { rc = 1; goto done; } reason = "no node list"; if (nodes == NULL) { goto done; } r1_nodes = pcmk__native_merge_weights(convert_const_pointer(resource1), resource1->id, NULL, NULL, 1, pe_weights_forward | pe_weights_init); pe__show_node_weights(true, NULL, resource1->id, r1_nodes, resource1->cluster); r2_nodes = pcmk__native_merge_weights(convert_const_pointer(resource2), resource2->id, NULL, NULL, 1, pe_weights_forward | pe_weights_init); pe__show_node_weights(true, NULL, resource2->id, r2_nodes, resource2->cluster); /* Current location score */ reason = "current location"; r1_weight = -INFINITY; r2_weight = -INFINITY; if (resource1->running_on) { r1_node = pe__current_node(resource1); r1_node = g_hash_table_lookup(r1_nodes, r1_node->details->id); if (r1_node != NULL) { r1_weight = r1_node->weight; } } if (resource2->running_on) { r2_node = pe__current_node(resource2); r2_node = g_hash_table_lookup(r2_nodes, r2_node->details->id); if (r2_node != NULL) { r2_weight = r2_node->weight; } } if (r1_weight > r2_weight) { rc = -1; goto done; } if (r1_weight < r2_weight) { rc = 1; goto done; } reason = "score"; for (gIter = nodes; gIter != NULL; gIter = gIter->next) { pe_node_t *node = (pe_node_t *) gIter->data; r1_node = NULL; r2_node = NULL; r1_weight = -INFINITY; if (r1_nodes) { r1_node = g_hash_table_lookup(r1_nodes, node->details->id); } if (r1_node) { r1_weight = r1_node->weight; } r2_weight = -INFINITY; if (r2_nodes) { r2_node = g_hash_table_lookup(r2_nodes, node->details->id); } if (r2_node) { r2_weight = r2_node->weight; } if (r1_weight > r2_weight) { rc = -1; goto done; } if (r1_weight < r2_weight) { rc = 1; goto done; } } done: crm_trace("%s (%d) on %s %c %s (%d) on %s: %s", resource1->id, r1_weight, r1_node ? r1_node->details->id : "n/a", rc < 0 ? '>' : rc > 0 ? '<' : '=', resource2->id, r2_weight, r2_node ? r2_node->details->id : "n/a", reason); if (r1_nodes) { g_hash_table_destroy(r1_nodes); } if (r2_nodes) { g_hash_table_destroy(r2_nodes); } return rc; } static void allocate_resources(pe_working_set_t * data_set) { GList *gIter = NULL; if (pcmk_is_set(data_set->flags, pe_flag_have_remote_nodes)) { /* Allocate remote connection resources first (which will also allocate * any colocation dependencies). If the connection is migrating, always * prefer the partial migration target. */ for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) { pe_resource_t *rsc = (pe_resource_t *) gIter->data; if (rsc->is_remote_node == FALSE) { continue; } pe_rsc_trace(rsc, "Allocating remote connection resource '%s'", rsc->id); rsc->cmds->allocate(rsc, rsc->partial_migration_target, data_set); } } /* now do the rest of the resources */ for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) { pe_resource_t *rsc = (pe_resource_t *) gIter->data; if (rsc->is_remote_node == TRUE) { continue; } pe_rsc_trace(rsc, "Allocating %s resource '%s'", crm_element_name(rsc->xml), rsc->id); rsc->cmds->allocate(rsc, NULL, data_set); } } /* We always use pe_order_preserve with these convenience functions to exempt * internally generated constraints from the prohibition of user constraints * involving remote connection resources. * * The start ordering additionally uses pe_order_runnable_left so that the * specified action is not runnable if the start is not runnable. */ static inline void order_start_then_action(pe_resource_t *lh_rsc, pe_action_t *rh_action, enum pe_ordering extra, pe_working_set_t *data_set) { if (lh_rsc && rh_action && data_set) { custom_action_order(lh_rsc, start_key(lh_rsc), NULL, rh_action->rsc, NULL, rh_action, pe_order_preserve | pe_order_runnable_left | extra, data_set); } } static inline void order_action_then_stop(pe_action_t *lh_action, pe_resource_t *rh_rsc, enum pe_ordering extra, pe_working_set_t *data_set) { if (lh_action && rh_rsc && data_set) { custom_action_order(lh_action->rsc, NULL, lh_action, rh_rsc, stop_key(rh_rsc), NULL, pe_order_preserve | extra, data_set); } } // Clear fail counts for orphaned rsc on all online nodes static void cleanup_orphans(pe_resource_t * rsc, pe_working_set_t * data_set) { GList *gIter = NULL; for (gIter = data_set->nodes; gIter != NULL; gIter = gIter->next) { pe_node_t *node = (pe_node_t *) gIter->data; if (node->details->online && pe_get_failcount(node, rsc, NULL, pe_fc_effective, NULL, data_set)) { pe_action_t *clear_op = NULL; clear_op = pe__clear_failcount(rsc, node, "it is orphaned", data_set); /* We can't use order_action_then_stop() here because its * pe_order_preserve breaks things */ custom_action_order(clear_op->rsc, NULL, clear_op, rsc, stop_key(rsc), NULL, pe_order_optional, data_set); } } } gboolean stage5(pe_working_set_t * data_set) { pcmk__output_t *out = data_set->priv; GList *gIter = NULL; if (!pcmk__str_eq(data_set->placement_strategy, "default", pcmk__str_casei)) { GList *nodes = g_list_copy(data_set->nodes); nodes = sort_nodes_by_weight(nodes, NULL, data_set); data_set->resources = g_list_sort_with_data(data_set->resources, sort_rsc_process_order, nodes); g_list_free(nodes); } gIter = data_set->nodes; for (; gIter != NULL; gIter = gIter->next) { pe_node_t *node = (pe_node_t *) gIter->data; if (pcmk_is_set(data_set->flags, pe_flag_show_utilization)) { out->message(out, "node-capacity", node, "Original"); } } crm_trace("Allocating services"); /* Take (next) highest resource, assign it and create its actions */ allocate_resources(data_set); gIter = data_set->nodes; for (; gIter != NULL; gIter = gIter->next) { pe_node_t *node = (pe_node_t *) gIter->data; if (pcmk_is_set(data_set->flags, pe_flag_show_utilization)) { out->message(out, "node-capacity", node, "Remaining"); } } // Process deferred action checks pe__foreach_param_check(data_set, check_params); pe__free_param_checks(data_set); if (pcmk_is_set(data_set->flags, pe_flag_startup_probes)) { crm_trace("Calculating needed probes"); /* This code probably needs optimization * ptest -x with 100 nodes, 100 clones and clone-max=100: With probes: ptest[14781]: 2010/09/27_17:56:46 notice: TRACE: do_calculations: pengine.c:258 Calculate cluster status ptest[14781]: 2010/09/27_17:56:46 notice: TRACE: do_calculations: pengine.c:278 Applying placement constraints ptest[14781]: 2010/09/27_17:56:47 notice: TRACE: do_calculations: pengine.c:285 Create internal constraints ptest[14781]: 2010/09/27_17:56:47 notice: TRACE: do_calculations: pengine.c:292 Check actions ptest[14781]: 2010/09/27_17:56:48 notice: TRACE: do_calculations: pengine.c:299 Allocate resources ptest[14781]: 2010/09/27_17:56:48 notice: TRACE: stage5: allocate.c:881 Allocating services ptest[14781]: 2010/09/27_17:56:49 notice: TRACE: stage5: allocate.c:894 Calculating needed probes ptest[14781]: 2010/09/27_17:56:51 notice: TRACE: stage5: allocate.c:899 Creating actions ptest[14781]: 2010/09/27_17:56:52 notice: TRACE: stage5: allocate.c:905 Creating done ptest[14781]: 2010/09/27_17:56:52 notice: TRACE: do_calculations: pengine.c:306 Processing fencing and shutdown cases ptest[14781]: 2010/09/27_17:56:52 notice: TRACE: do_calculations: pengine.c:313 Applying ordering constraints 36s ptest[14781]: 2010/09/27_17:57:28 notice: TRACE: do_calculations: pengine.c:320 Create transition graph Without probes: ptest[14637]: 2010/09/27_17:56:21 notice: TRACE: do_calculations: pengine.c:258 Calculate cluster status ptest[14637]: 2010/09/27_17:56:22 notice: TRACE: do_calculations: pengine.c:278 Applying placement constraints ptest[14637]: 2010/09/27_17:56:22 notice: TRACE: do_calculations: pengine.c:285 Create internal constraints ptest[14637]: 2010/09/27_17:56:22 notice: TRACE: do_calculations: pengine.c:292 Check actions ptest[14637]: 2010/09/27_17:56:23 notice: TRACE: do_calculations: pengine.c:299 Allocate resources ptest[14637]: 2010/09/27_17:56:23 notice: TRACE: stage5: allocate.c:881 Allocating services ptest[14637]: 2010/09/27_17:56:24 notice: TRACE: stage5: allocate.c:899 Creating actions ptest[14637]: 2010/09/27_17:56:25 notice: TRACE: stage5: allocate.c:905 Creating done ptest[14637]: 2010/09/27_17:56:25 notice: TRACE: do_calculations: pengine.c:306 Processing fencing and shutdown cases ptest[14637]: 2010/09/27_17:56:25 notice: TRACE: do_calculations: pengine.c:313 Applying ordering constraints ptest[14637]: 2010/09/27_17:56:25 notice: TRACE: do_calculations: pengine.c:320 Create transition graph */ probe_resources(data_set); } crm_trace("Handle orphans"); if (pcmk_is_set(data_set->flags, pe_flag_stop_rsc_orphans)) { for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) { pe_resource_t *rsc = (pe_resource_t *) gIter->data; /* There's no need to recurse into rsc->children because those * should just be unallocated clone instances. */ if (pcmk_is_set(rsc->flags, pe_rsc_orphan)) { cleanup_orphans(rsc, data_set); } } } crm_trace("Creating actions"); for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) { pe_resource_t *rsc = (pe_resource_t *) gIter->data; rsc->cmds->create_actions(rsc, data_set); } crm_trace("Creating done"); return TRUE; } static gboolean is_managed(const pe_resource_t * rsc) { GList *gIter = rsc->children; if (pcmk_is_set(rsc->flags, pe_rsc_managed)) { return TRUE; } for (; gIter != NULL; gIter = gIter->next) { pe_resource_t *child_rsc = (pe_resource_t *) gIter->data; if (is_managed(child_rsc)) { return TRUE; } } return FALSE; } static gboolean any_managed_resources(pe_working_set_t * data_set) { GList *gIter = data_set->resources; for (; gIter != NULL; gIter = gIter->next) { pe_resource_t *rsc = (pe_resource_t *) gIter->data; if (is_managed(rsc)) { return TRUE; } } return FALSE; } -/*! - * \internal - * \brief Create pseudo-op for guest node fence, and order relative to it - * - * \param[in] node Guest node to fence - * \param[in] data_set Working set of CIB state - */ -static void -fence_guest(pe_node_t *node, pe_working_set_t *data_set) -{ - pe_resource_t *container = node->details->remote_rsc->container; - pe_action_t *stop = NULL; - pe_action_t *stonith_op = NULL; - - /* The fence action is just a label; we don't do anything differently for - * off vs. reboot. We specify it explicitly, rather than let it default to - * cluster's default action, because we are not _initiating_ fencing -- we - * are creating a pseudo-event to describe fencing that is already occurring - * by other means (container recovery). - */ - const char *fence_action = "off"; - - /* Check whether guest's container resource has any explicit stop or - * start (the stop may be implied by fencing of the guest's host). - */ - if (container) { - stop = find_first_action(container->actions, NULL, CRMD_ACTION_STOP, NULL); - - if (find_first_action(container->actions, NULL, CRMD_ACTION_START, NULL)) { - fence_action = "reboot"; - } - } - - /* Create a fence pseudo-event, so we have an event to order actions - * against, and the controller can always detect it. - */ - stonith_op = pe_fence_op(node, fence_action, FALSE, "guest is unclean", FALSE, data_set); - pe__set_action_flags(stonith_op, pe_action_pseudo|pe_action_runnable); - - /* We want to imply stops/demotes after the guest is stopped, not wait until - * it is restarted, so we always order pseudo-fencing after stop, not start - * (even though start might be closer to what is done for a real reboot). - */ - if ((stop != NULL) && pcmk_is_set(stop->flags, pe_action_pseudo)) { - pe_action_t *parent_stonith_op = pe_fence_op(stop->node, NULL, FALSE, NULL, FALSE, data_set); - crm_info("Implying guest node %s is down (action %d) after %s fencing", - node->details->uname, stonith_op->id, stop->node->details->uname); - order_actions(parent_stonith_op, stonith_op, - pe_order_runnable_left|pe_order_implies_then); - - } else if (stop) { - order_actions(stop, stonith_op, - pe_order_runnable_left|pe_order_implies_then); - crm_info("Implying guest node %s is down (action %d) " - "after container %s is stopped (action %d)", - node->details->uname, stonith_op->id, - container->id, stop->id); - } else { - /* If we're fencing the guest node but there's no stop for the guest - * resource, we must think the guest is already stopped. However, we may - * think so because its resource history was just cleaned. To avoid - * unnecessarily considering the guest node down if it's really up, - * order the pseudo-fencing after any stop of the connection resource, - * which will be ordered after any container (re-)probe. - */ - stop = find_first_action(node->details->remote_rsc->actions, NULL, - RSC_STOP, NULL); - - if (stop) { - order_actions(stop, stonith_op, pe_order_optional); - crm_info("Implying guest node %s is down (action %d) " - "after connection is stopped (action %d)", - node->details->uname, stonith_op->id, stop->id); - } else { - /* Not sure why we're fencing, but everything must already be - * cleanly stopped. - */ - crm_info("Implying guest node %s is down (action %d) ", - node->details->uname, stonith_op->id); - } - } - - /* Order/imply other actions relative to pseudo-fence as with real fence */ - pcmk__order_vs_fence(stonith_op, data_set); -} - /* * Create dependencies for stonith and shutdown operations */ gboolean stage6(pe_working_set_t * data_set) { pe_action_t *dc_down = NULL; pe_action_t *stonith_op = NULL; gboolean integrity_lost = FALSE; gboolean need_stonith = TRUE; GList *gIter; GList *stonith_ops = NULL; GList *shutdown_ops = NULL; /* Remote ordering constraints need to happen prior to calculating fencing * because it is one more place we will mark the node as dirty. * * A nice side effect of doing them early is that apply_*_ordering() can be * simpler because pe_fence_node() has already done some of the work. */ crm_trace("Creating remote ordering constraints"); apply_remote_node_ordering(data_set); crm_trace("Processing fencing and shutdown cases"); if (any_managed_resources(data_set) == FALSE) { crm_notice("Delaying fencing operations until there are resources to manage"); need_stonith = FALSE; } /* Check each node for stonith/shutdown */ for (gIter = data_set->nodes; gIter != NULL; gIter = gIter->next) { pe_node_t *node = (pe_node_t *) gIter->data; /* Guest nodes are "fenced" by recovering their container resource, * so handle them separately. */ if (pe__is_guest_node(node)) { if (node->details->remote_requires_reset && need_stonith && pe_can_fence(data_set, node)) { - fence_guest(node, data_set); + pcmk__fence_guest(node, data_set); } continue; } stonith_op = NULL; if (node->details->unclean && need_stonith && pe_can_fence(data_set, node)) { stonith_op = pe_fence_op(node, NULL, FALSE, "node is unclean", FALSE, data_set); pe_warn("Scheduling Node %s for STONITH", node->details->uname); pcmk__order_vs_fence(stonith_op, data_set); if (node->details->is_dc) { // Remember if the DC is being fenced dc_down = stonith_op; } else { if (!pcmk_is_set(data_set->flags, pe_flag_concurrent_fencing) && (stonith_ops != NULL)) { /* Concurrent fencing is disabled, so order each non-DC * fencing in a chain. If there is any DC fencing or * shutdown, it will be ordered after the last action in the * chain later. */ order_actions((pe_action_t *) stonith_ops->data, stonith_op, pe_order_optional); } // Remember all non-DC fencing actions in a separate list stonith_ops = g_list_prepend(stonith_ops, stonith_op); } } else if (node->details->online && node->details->shutdown && /* TODO define what a shutdown op means for a remote node. * For now we do not send shutdown operations for remote nodes, but * if we can come up with a good use for this in the future, we will. */ pe__is_guest_or_remote_node(node) == FALSE) { pe_action_t *down_op = sched_shutdown_op(node, data_set); if (node->details->is_dc) { // Remember if the DC is being shut down dc_down = down_op; } else { // Remember non-DC shutdowns for later ordering shutdown_ops = g_list_prepend(shutdown_ops, down_op); } } if (node->details->unclean && stonith_op == NULL) { integrity_lost = TRUE; pe_warn("Node %s is unclean!", node->details->uname); } } if (integrity_lost) { if (!pcmk_is_set(data_set->flags, pe_flag_stonith_enabled)) { pe_warn("YOUR RESOURCES ARE NOW LIKELY COMPROMISED"); pe_err("ENABLE STONITH TO KEEP YOUR RESOURCES SAFE"); } else if (!pcmk_is_set(data_set->flags, pe_flag_have_quorum)) { crm_notice("Cannot fence unclean nodes until quorum is" " attained (or no-quorum-policy is set to ignore)"); } } if (dc_down != NULL) { /* Order any non-DC shutdowns before any DC shutdown, to avoid repeated * DC elections. However, we don't want to order non-DC shutdowns before * a DC *fencing*, because even though we don't want a node that's * shutting down to become DC, the DC fencing could be ordered before a * clone stop that's also ordered before the shutdowns, thus leading to * a graph loop. */ if (pcmk__str_eq(dc_down->task, CRM_OP_SHUTDOWN, pcmk__str_casei)) { for (gIter = shutdown_ops; gIter != NULL; gIter = gIter->next) { pe_action_t *node_stop = (pe_action_t *) gIter->data; crm_debug("Ordering shutdown on %s before %s on DC %s", node_stop->node->details->uname, dc_down->task, dc_down->node->details->uname); order_actions(node_stop, dc_down, pe_order_optional); } } // Order any non-DC fencing before any DC fencing or shutdown if (pcmk_is_set(data_set->flags, pe_flag_concurrent_fencing)) { /* With concurrent fencing, order each non-DC fencing action * separately before any DC fencing or shutdown. */ for (gIter = stonith_ops; gIter != NULL; gIter = gIter->next) { order_actions((pe_action_t *) gIter->data, dc_down, pe_order_optional); } } else if (stonith_ops) { /* Without concurrent fencing, the non-DC fencing actions are * already ordered relative to each other, so we just need to order * the DC fencing after the last action in the chain (which is the * first item in the list). */ order_actions((pe_action_t *) stonith_ops->data, dc_down, pe_order_optional); } } g_list_free(stonith_ops); g_list_free(shutdown_ops); return TRUE; } /* * Determine the sets of independent actions and the correct order for the * actions in each set. * * Mark dependencies of un-runnable actions un-runnable * */ static GList * find_actions_by_task(GList *actions, pe_resource_t * rsc, const char *original_key) { GList *list = NULL; list = find_actions(actions, original_key, NULL); if (list == NULL) { /* we're potentially searching a child of the original resource */ char *key = NULL; char *task = NULL; guint interval_ms = 0; if (parse_op_key(original_key, NULL, &task, &interval_ms)) { key = pcmk__op_key(rsc->id, task, interval_ms); list = find_actions(actions, key, NULL); } else { crm_err("search key: %s", original_key); } free(key); free(task); } return list; } static void rsc_order_then(pe_action_t *lh_action, pe_resource_t *rsc, pe__ordering_t *order) { GList *gIter = NULL; GList *rh_actions = NULL; pe_action_t *rh_action = NULL; enum pe_ordering type; CRM_CHECK(rsc != NULL, return); CRM_CHECK(order != NULL, return); type = order->type; rh_action = order->rh_action; crm_trace("Applying ordering constraint %d (then: %s)", order->id, rsc->id); if (rh_action != NULL) { rh_actions = g_list_prepend(NULL, rh_action); } else if (rsc != NULL) { rh_actions = find_actions_by_task(rsc->actions, rsc, order->rh_action_task); } if (rh_actions == NULL) { pe_rsc_trace(rsc, "Ignoring constraint %d: then (%s for %s) not found", order->id, order->rh_action_task, rsc->id); return; } if ((lh_action != NULL) && (lh_action->rsc == rsc) && pcmk_is_set(lh_action->flags, pe_action_dangle)) { pe_rsc_trace(rsc, "Detected dangling operation %s -> %s", lh_action->uuid, order->rh_action_task); pe__clear_order_flags(type, pe_order_implies_then); } gIter = rh_actions; for (; gIter != NULL; gIter = gIter->next) { pe_action_t *rh_action_iter = (pe_action_t *) gIter->data; if (lh_action) { order_actions(lh_action, rh_action_iter, type); } else if (type & pe_order_implies_then) { pe__clear_action_flags(rh_action_iter, pe_action_runnable); crm_warn("Unrunnable %s 0x%.6x", rh_action_iter->uuid, type); } else { crm_warn("neither %s 0x%.6x", rh_action_iter->uuid, type); } } g_list_free(rh_actions); } static void rsc_order_first(pe_resource_t *lh_rsc, pe__ordering_t *order, pe_working_set_t *data_set) { GList *lh_actions = NULL; pe_action_t *lh_action = order->lh_action; pe_resource_t *rh_rsc = order->rh_rsc; CRM_ASSERT(lh_rsc != NULL); pe_rsc_trace(lh_rsc, "Applying ordering constraint %d (first: %s)", order->id, lh_rsc->id); if (lh_action != NULL) { lh_actions = g_list_prepend(NULL, lh_action); } else { lh_actions = find_actions_by_task(lh_rsc->actions, lh_rsc, order->lh_action_task); } if ((lh_actions == NULL) && (lh_rsc == rh_rsc)) { pe_rsc_trace(lh_rsc, "Ignoring constraint %d: first (%s for %s) not found", order->id, order->lh_action_task, lh_rsc->id); } else if (lh_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(lh_rsc->id, op_type, interval_ms); if (lh_rsc->fns->state(lh_rsc, TRUE) == RSC_ROLE_STOPPED && pcmk__str_eq(op_type, RSC_STOP, pcmk__str_casei)) { free(key); pe_rsc_trace(lh_rsc, "Ignoring constraint %d: first (%s for %s) not found", order->id, order->lh_action_task, lh_rsc->id); } else if ((lh_rsc->fns->state(lh_rsc, TRUE) == RSC_ROLE_UNPROMOTED) && pcmk__str_eq(op_type, RSC_DEMOTE, pcmk__str_casei)) { free(key); pe_rsc_trace(lh_rsc, "Ignoring constraint %d: first (%s for %s) not found", order->id, order->lh_action_task, lh_rsc->id); } else { pe_rsc_trace(lh_rsc, "Creating first (%s for %s) for constraint %d ", order->lh_action_task, lh_rsc->id, order->id); lh_action = custom_action(lh_rsc, key, op_type, NULL, TRUE, TRUE, data_set); lh_actions = g_list_prepend(NULL, lh_action); } free(op_type); } if (rh_rsc == NULL) { if (order->rh_action == NULL) { pe_rsc_trace(lh_rsc, "Ignoring constraint %d: then not found", order->id); return; } rh_rsc = order->rh_action->rsc; } for (GList *gIter = lh_actions; gIter != NULL; gIter = gIter->next) { lh_action = (pe_action_t *) gIter->data; if (rh_rsc == NULL) { order_actions(lh_action, order->rh_action, order->type); } else { rsc_order_then(lh_action, rh_rsc, order); } } g_list_free(lh_actions); } extern void update_colo_start_chain(pe_action_t *action, pe_working_set_t *data_set); static int is_recurring_action(pe_action_t *action) { guint interval_ms; if (pcmk__guint_from_hash(action->meta, XML_LRM_ATTR_INTERVAL_MS, 0, &interval_ms) != pcmk_rc_ok) { return 0; } return (interval_ms > 0); } static void apply_container_ordering(pe_action_t *action, pe_working_set_t *data_set) { /* VMs are also classified as containers for these purposes... in * that they both involve a 'thing' running on a real or remote * cluster node. * * This allows us to be smarter about the type and extent of * recovery actions required in various scenarios */ pe_resource_t *remote_rsc = NULL; pe_resource_t *container = NULL; enum action_tasks task = text2task(action->task); CRM_ASSERT(action->rsc); CRM_ASSERT(action->node); CRM_ASSERT(pe__is_guest_or_remote_node(action->node)); remote_rsc = action->node->details->remote_rsc; CRM_ASSERT(remote_rsc); container = remote_rsc->container; CRM_ASSERT(container); if (pcmk_is_set(container->flags, pe_rsc_failed)) { pe_fence_node(data_set, action->node, "container failed", FALSE); } crm_trace("Order %s action %s relative to %s%s for %s%s", action->task, action->uuid, pcmk_is_set(remote_rsc->flags, pe_rsc_failed)? "failed " : "", remote_rsc->id, pcmk_is_set(container->flags, pe_rsc_failed)? "failed " : "", container->id); if (pcmk__strcase_any_of(action->task, CRMD_ACTION_MIGRATE, CRMD_ACTION_MIGRATED, NULL)) { /* Migration ops map to "no_action", but we need to apply the same * ordering as for stop or demote (see get_router_node()). */ task = stop_rsc; } switch (task) { case start_rsc: case action_promote: /* Force resource recovery if the container is recovered */ order_start_then_action(container, action, pe_order_implies_then, data_set); /* Wait for the connection resource to be up too */ order_start_then_action(remote_rsc, action, pe_order_none, data_set); break; case stop_rsc: case action_demote: if (pcmk_is_set(container->flags, pe_rsc_failed)) { /* When the container representing a guest node fails, any stop * or demote actions for resources running on the guest node * are implied by the container stopping. This is similar to * how fencing operations work for cluster nodes and remote * nodes. */ } else { /* Ensure the operation happens before the connection is brought * down. * * If we really wanted to, we could order these after the * connection start, IFF the container's current role was * stopped (otherwise we re-introduce an ordering loop when the * connection is restarting). */ order_action_then_stop(action, remote_rsc, pe_order_none, data_set); } break; default: /* Wait for the connection resource to be up */ if (is_recurring_action(action)) { /* In case we ever get the recovery logic wrong, force * recurring monitors to be restarted, even if just * the connection was re-established */ if(task != no_action) { order_start_then_action(remote_rsc, action, pe_order_implies_then, data_set); } } else { order_start_then_action(remote_rsc, action, pe_order_none, data_set); } break; } } static enum remote_connection_state get_remote_node_state(pe_node_t *node) { pe_resource_t *remote_rsc = NULL; pe_node_t *cluster_node = NULL; CRM_ASSERT(node); remote_rsc = node->details->remote_rsc; CRM_ASSERT(remote_rsc); cluster_node = pe__current_node(remote_rsc); /* If the cluster node the remote connection resource resides on * is unclean or went offline, we can't process any operations * on that remote node until after it starts elsewhere. */ if(remote_rsc->next_role == RSC_ROLE_STOPPED || remote_rsc->allocated_to == NULL) { /* The connection resource is not going to run anywhere */ if (cluster_node && cluster_node->details->unclean) { /* The remote connection is failed because its resource is on a * failed node and can't be recovered elsewhere, so we must fence. */ return remote_state_failed; } if (!pcmk_is_set(remote_rsc->flags, pe_rsc_failed)) { /* Connection resource is cleanly stopped */ return remote_state_stopped; } /* Connection resource is failed */ if ((remote_rsc->next_role == RSC_ROLE_STOPPED) && remote_rsc->remote_reconnect_ms && node->details->remote_was_fenced && !pe__shutdown_requested(node)) { /* We won't know whether the connection is recoverable until the * reconnect interval expires and we reattempt connection. */ return remote_state_unknown; } /* The remote connection is in a failed state. If there are any * resources known to be active on it (stop) or in an unknown state * (probe), we must assume the worst and fence it. */ return remote_state_failed; } else if (cluster_node == NULL) { /* Connection is recoverable but not currently running anywhere, see if we can recover it first */ return remote_state_unknown; } else if(cluster_node->details->unclean == TRUE || cluster_node->details->online == FALSE) { /* Connection is running on a dead node, see if we can recover it first */ return remote_state_resting; } else if (pcmk__list_of_multiple(remote_rsc->running_on) && remote_rsc->partial_migration_source && remote_rsc->partial_migration_target) { /* We're in the middle of migrating a connection resource, * wait until after the resource migrates before performing * any actions. */ return remote_state_resting; } return remote_state_alive; } /*! * \internal * \brief Order actions on remote node relative to actions for the connection */ static void apply_remote_ordering(pe_action_t *action, pe_working_set_t *data_set) { pe_resource_t *remote_rsc = NULL; enum action_tasks task = text2task(action->task); enum remote_connection_state state = get_remote_node_state(action->node); enum pe_ordering order_opts = pe_order_none; if (action->rsc == NULL) { return; } CRM_ASSERT(action->node); CRM_ASSERT(pe__is_guest_or_remote_node(action->node)); remote_rsc = action->node->details->remote_rsc; CRM_ASSERT(remote_rsc); crm_trace("Order %s action %s relative to %s%s (state: %s)", action->task, action->uuid, pcmk_is_set(remote_rsc->flags, pe_rsc_failed)? "failed " : "", remote_rsc->id, state2text(state)); if (pcmk__strcase_any_of(action->task, CRMD_ACTION_MIGRATE, CRMD_ACTION_MIGRATED, NULL)) { /* Migration ops map to "no_action", but we need to apply the same * ordering as for stop or demote (see get_router_node()). */ task = stop_rsc; } switch (task) { case start_rsc: case action_promote: order_opts = pe_order_none; if (state == remote_state_failed) { /* Force recovery, by making this action required */ pe__set_order_flags(order_opts, pe_order_implies_then); } /* Ensure connection is up before running this action */ order_start_then_action(remote_rsc, action, order_opts, data_set); break; case stop_rsc: if(state == remote_state_alive) { order_action_then_stop(action, remote_rsc, pe_order_implies_first, data_set); } else if(state == remote_state_failed) { /* The resource is active on the node, but since we don't have a * valid connection, the only way to stop the resource is by * fencing the node. There is no need to order the stop relative * to the remote connection, since the stop will become implied * by the fencing. */ pe_fence_node(data_set, action->node, "resources are active and the connection is unrecoverable", FALSE); } else if(remote_rsc->next_role == RSC_ROLE_STOPPED) { /* State must be remote_state_unknown or remote_state_stopped. * Since the connection is not coming back up in this * transition, stop this resource first. */ order_action_then_stop(action, remote_rsc, pe_order_implies_first, data_set); } else { /* The connection is going to be started somewhere else, so * stop this resource after that completes. */ order_start_then_action(remote_rsc, action, pe_order_none, data_set); } break; case action_demote: /* Only order this demote relative to the connection start if the * connection isn't being torn down. Otherwise, the demote would be * blocked because the connection start would not be allowed. */ if(state == remote_state_resting || state == remote_state_unknown) { order_start_then_action(remote_rsc, action, pe_order_none, data_set); } /* Otherwise we can rely on the stop ordering */ break; default: /* Wait for the connection resource to be up */ if (is_recurring_action(action)) { /* In case we ever get the recovery logic wrong, force * recurring monitors to be restarted, even if just * the connection was re-established */ order_start_then_action(remote_rsc, action, pe_order_implies_then, data_set); } else { pe_node_t *cluster_node = pe__current_node(remote_rsc); if(task == monitor_rsc && state == remote_state_failed) { /* We would only be here if we do not know the * state of the resource on the remote node. * Since we have no way to find out, it is * necessary to fence the node. */ pe_fence_node(data_set, action->node, "resources are in an unknown state and the connection is unrecoverable", FALSE); } if(cluster_node && state == remote_state_stopped) { /* The connection is currently up, but is going * down permanently. * * Make sure we check services are actually * stopped _before_ we let the connection get * closed */ order_action_then_stop(action, remote_rsc, pe_order_runnable_left, data_set); } else { order_start_then_action(remote_rsc, action, pe_order_none, data_set); } } break; } } static void apply_remote_node_ordering(pe_working_set_t *data_set) { if (!pcmk_is_set(data_set->flags, pe_flag_have_remote_nodes)) { return; } for (GList *gIter = data_set->actions; gIter != NULL; gIter = gIter->next) { pe_action_t *action = (pe_action_t *) gIter->data; pe_resource_t *remote = NULL; // We are only interested in resource actions if (action->rsc == NULL) { continue; } /* Special case: If we are clearing the failcount of an actual * remote connection resource, then make sure this happens before * any start of the resource in this transition. */ if (action->rsc->is_remote_node && pcmk__str_eq(action->task, CRM_OP_CLEAR_FAILCOUNT, pcmk__str_casei)) { custom_action_order(action->rsc, NULL, action, action->rsc, pcmk__op_key(action->rsc->id, RSC_START, 0), NULL, pe_order_optional, data_set); continue; } // We are only interested in actions allocated to a node if (action->node == NULL) { continue; } if (!pe__is_guest_or_remote_node(action->node)) { continue; } /* We are only interested in real actions. * * @TODO This is probably wrong; pseudo-actions might be converted to * real actions and vice versa later in update_actions() at the end of * stage7(). */ if (pcmk_is_set(action->flags, pe_action_pseudo)) { continue; } remote = action->node->details->remote_rsc; if (remote == NULL) { // Orphaned continue; } /* Another special case: if a resource is moving to a Pacemaker Remote * node, order the stop on the original node after any start of the * remote connection. This ensures that if the connection fails to * start, we leave the resource running on the original node. */ if (pcmk__str_eq(action->task, RSC_START, pcmk__str_casei)) { for (GList *item = action->rsc->actions; item != NULL; item = item->next) { pe_action_t *rsc_action = item->data; if ((rsc_action->node->details != action->node->details) && pcmk__str_eq(rsc_action->task, RSC_STOP, pcmk__str_casei)) { custom_action_order(remote, start_key(remote), NULL, action->rsc, NULL, rsc_action, pe_order_optional, data_set); } } } /* The action occurs across a remote connection, so create * ordering constraints that guarantee the action occurs while the node * is active (after start, before stop ... things like that). * * This is somewhat brittle in that we need to make sure the results of * this ordering are compatible with the result of get_router_node(). * It would probably be better to add XML_LRM_ATTR_ROUTER_NODE as part * of this logic rather than action2xml(). */ if (remote->container) { crm_trace("Container ordering for %s", action->uuid); apply_container_ordering(action, data_set); } else { crm_trace("Remote ordering for %s", action->uuid); apply_remote_ordering(action, data_set); } } } static gboolean order_first_probe_unneeded(pe_action_t * probe, pe_action_t * rh_action) { /* No need to probe the resource on the node that is being * unfenced. Otherwise it might introduce transition loop * since probe will be performed after the node is * unfenced. */ if (pcmk__str_eq(rh_action->task, CRM_OP_FENCE, pcmk__str_casei) && probe->node && rh_action->node && probe->node->details == rh_action->node->details) { const char *op = g_hash_table_lookup(rh_action->meta, "stonith_action"); if (pcmk__str_eq(op, "on", pcmk__str_casei)) { return TRUE; } } // Shutdown waits for probe to complete only if it's on the same node if ((pcmk__str_eq(rh_action->task, CRM_OP_SHUTDOWN, pcmk__str_casei)) && probe->node && rh_action->node && probe->node->details != rh_action->node->details) { return TRUE; } return FALSE; } static void order_first_probes_imply_stops(pe_working_set_t * data_set) { GList *gIter = NULL; for (gIter = data_set->ordering_constraints; gIter != NULL; gIter = gIter->next) { pe__ordering_t *order = gIter->data; enum pe_ordering order_type = pe_order_optional; pe_resource_t *lh_rsc = order->lh_rsc; pe_resource_t *rh_rsc = order->rh_rsc; pe_action_t *lh_action = order->lh_action; pe_action_t *rh_action = order->rh_action; const char *lh_action_task = order->lh_action_task; const char *rh_action_task = order->rh_action_task; GList *probes = NULL; GList *rh_actions = NULL; GList *pIter = NULL; if (lh_rsc == NULL) { continue; } else if (rh_rsc && lh_rsc == rh_rsc) { continue; } if (lh_action == NULL && lh_action_task == NULL) { continue; } if (rh_action == NULL && rh_action_task == NULL) { continue; } /* Technically probe is expected to return "not running", which could be * the alternative of stop action if the status of the resource is * unknown yet. */ if (lh_action && !pcmk__str_eq(lh_action->task, RSC_STOP, pcmk__str_casei)) { continue; } else if (lh_action == NULL && lh_action_task && !pcmk__ends_with(lh_action_task, "_" RSC_STOP "_0")) { continue; } /* Do not probe the resource inside of a stopping container. Otherwise * it might introduce transition loop since probe will be performed * after the container starts again. */ if (rh_rsc && lh_rsc->container == rh_rsc) { if (rh_action && pcmk__str_eq(rh_action->task, RSC_STOP, pcmk__str_casei)) { continue; } else if (rh_action == NULL && rh_action_task && pcmk__ends_with(rh_action_task,"_" RSC_STOP "_0")) { continue; } } if (order->type == pe_order_none) { continue; } // Preserve the order options for future filtering if (pcmk_is_set(order->type, pe_order_apply_first_non_migratable)) { pe__set_order_flags(order_type, pe_order_apply_first_non_migratable); } if (pcmk_is_set(order->type, pe_order_same_node)) { pe__set_order_flags(order_type, pe_order_same_node); } // Keep the order types for future filtering if (order->type == pe_order_anti_colocation || order->type == pe_order_load) { order_type = order->type; } probes = pe__resource_actions(lh_rsc, NULL, RSC_STATUS, FALSE); if (probes == NULL) { continue; } if (rh_action) { rh_actions = g_list_prepend(rh_actions, rh_action); } else if (rh_rsc && rh_action_task) { rh_actions = find_actions(rh_rsc->actions, rh_action_task, NULL); } if (rh_actions == NULL) { g_list_free(probes); continue; } crm_trace("Processing for LH probe based on ordering constraint %s -> %s" " (id=%d, type=%.6x)", lh_action ? lh_action->uuid : lh_action_task, rh_action ? rh_action->uuid : rh_action_task, order->id, order->type); for (pIter = probes; pIter != NULL; pIter = pIter->next) { pe_action_t *probe = (pe_action_t *) pIter->data; GList *rIter = NULL; for (rIter = rh_actions; rIter != NULL; rIter = rIter->next) { pe_action_t *rh_action_iter = (pe_action_t *) rIter->data; if (order_first_probe_unneeded(probe, rh_action_iter)) { continue; } order_actions(probe, rh_action_iter, order_type); } } g_list_free(rh_actions); g_list_free(probes); } } static void order_first_probe_then_restart_repromote(pe_action_t * probe, pe_action_t * after, pe_working_set_t * data_set) { GList *gIter = NULL; bool interleave = FALSE; pe_resource_t *compatible_rsc = NULL; if (probe == NULL || probe->rsc == NULL || probe->rsc->variant != pe_native) { return; } if (after == NULL // Avoid running into any possible loop || pcmk_is_set(after->flags, pe_action_tracking)) { return; } if (!pcmk__str_eq(probe->task, RSC_STATUS, pcmk__str_casei)) { return; } pe__set_action_flags(after, pe_action_tracking); crm_trace("Processing based on %s %s -> %s %s", probe->uuid, probe->node ? probe->node->details->uname: "", after->uuid, after->node ? after->node->details->uname : ""); if (after->rsc /* Better not build a dependency directly with a clone/group. * We are going to proceed through the ordering chain and build * dependencies with its children. */ && after->rsc->variant == pe_native && probe->rsc != after->rsc) { GList *then_actions = NULL; enum pe_ordering probe_order_type = pe_order_optional; if (pcmk__str_eq(after->task, RSC_START, pcmk__str_casei)) { then_actions = pe__resource_actions(after->rsc, NULL, RSC_STOP, FALSE); } else if (pcmk__str_eq(after->task, RSC_PROMOTE, pcmk__str_casei)) { then_actions = pe__resource_actions(after->rsc, NULL, RSC_DEMOTE, FALSE); } for (gIter = then_actions; gIter != NULL; gIter = gIter->next) { pe_action_t *then = (pe_action_t *) gIter->data; // Skip any pseudo action which for example is implied by fencing if (pcmk_is_set(then->flags, pe_action_pseudo)) { continue; } order_actions(probe, then, probe_order_type); } g_list_free(then_actions); } if (after->rsc && after->rsc->variant > pe_group) { const char *interleave_s = g_hash_table_lookup(after->rsc->meta, XML_RSC_ATTR_INTERLEAVE); interleave = crm_is_true(interleave_s); if (interleave) { /* For an interleaved clone, we should build a dependency only * with the relevant clone child. */ compatible_rsc = find_compatible_child(probe->rsc, after->rsc, RSC_ROLE_UNKNOWN, FALSE, data_set); } } for (gIter = after->actions_after; gIter != NULL; gIter = gIter->next) { pe_action_wrapper_t *after_wrapper = (pe_action_wrapper_t *) gIter->data; /* pe_order_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 pe_order_implies_then_on_node though. It's now only * used for unfencing case, which tends to introduce transition * loops... */ if (!pcmk_is_set(after_wrapper->type, pe_order_implies_then)) { /* The order type between a group/clone and its child such as * B.start-> B_child.start is: * pe_order_implies_first_printed | pe_order_runnable_left * * Proceed through the ordering chain and build dependencies with * its children. */ if (after->rsc == NULL || after->rsc->variant < pe_group || probe->rsc->parent == after->rsc || after_wrapper->action->rsc == NULL || after_wrapper->action->rsc->variant > pe_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 > pe_group && interleave == TRUE && (compatible_rsc == NULL || compatible_rsc != after_wrapper->action->rsc)) { continue; } } crm_trace("Proceeding through %s %s -> %s %s (type=0x%.6x)", after->uuid, after->node ? after->node->details->uname: "", after_wrapper->action->uuid, after_wrapper->action->node ? after_wrapper->action->node->details->uname : "", after_wrapper->type); order_first_probe_then_restart_repromote(probe, after_wrapper->action, data_set); } } static void clear_actions_tracking_flag(pe_working_set_t * data_set) { GList *gIter = NULL; for (gIter = data_set->actions; gIter != NULL; gIter = gIter->next) { pe_action_t *action = (pe_action_t *) gIter->data; if (pcmk_is_set(action->flags, pe_action_tracking)) { pe__clear_action_flags(action, pe_action_tracking); } } } static void order_first_rsc_probes(pe_resource_t * rsc, pe_working_set_t * data_set) { GList *gIter = NULL; GList *probes = NULL; g_list_foreach(rsc->children, (GFunc) order_first_rsc_probes, data_set); if (rsc->variant != pe_native) { return; } probes = pe__resource_actions(rsc, NULL, RSC_STATUS, FALSE); for (gIter = probes; gIter != NULL; gIter= gIter->next) { pe_action_t *probe = (pe_action_t *) gIter->data; GList *aIter = NULL; for (aIter = probe->actions_after; aIter != NULL; aIter = aIter->next) { pe_action_wrapper_t *after_wrapper = (pe_action_wrapper_t *) aIter->data; order_first_probe_then_restart_repromote(probe, after_wrapper->action, data_set); clear_actions_tracking_flag(data_set); } } g_list_free(probes); } static void order_first_probes(pe_working_set_t * data_set) { GList *gIter = NULL; for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) { pe_resource_t *rsc = (pe_resource_t *) gIter->data; order_first_rsc_probes(rsc, data_set); } order_first_probes_imply_stops(data_set); } static void order_then_probes(pe_working_set_t * data_set) { #if 0 GList *gIter = NULL; for (gIter = data_set->resources; gIter != NULL; gIter = gIter->next) { pe_resource_t *rsc = (pe_resource_t *) gIter->data; /* Given "A then B", we would prefer to wait for A to be * started before probing B. * * If A was a filesystem on which the binaries and data for B * lived, it would have been useful if the author of B's agent * could assume that A is running before B.monitor will be * called. * * However we can't _only_ probe once A is running, otherwise * we'd not detect the state of B if A could not be started * for some reason. * * In practice however, we cannot even do an opportunistic * version of this because B may be moving: * * B.probe -> B.start * B.probe -> B.stop * B.stop -> B.start * A.stop -> A.start * A.start -> B.probe * * So far so good, but if we add the result of this code: * * B.stop -> A.stop * * Then we get a loop: * * B.probe -> B.stop -> A.stop -> A.start -> B.probe * * We could kill the 'B.probe -> B.stop' dependency, but that * could mean stopping B "too" soon, because B.start must wait * for the probes to complete. * * Another option is to allow it only if A is a non-unique * clone with clone-max == node-max (since we'll never be * moving it). However, we could still be stopping one * instance at the same time as starting another. * The complexity of checking for allowed conditions combined * with the ever narrowing usecase suggests that this code * should remain disabled until someone gets smarter. */ pe_action_t *start = NULL; GList *actions = NULL; GList *probes = NULL; actions = pe__resource_actions(rsc, NULL, RSC_START, FALSE); if (actions) { start = actions->data; g_list_free(actions); } if(start == NULL) { crm_err("No start action for %s", rsc->id); continue; } probes = pe__resource_actions(rsc, NULL, RSC_STATUS, FALSE); for (actions = start->actions_before; actions != NULL; actions = actions->next) { pe_action_wrapper_t *before = (pe_action_wrapper_t *) actions->data; GList *pIter = NULL; pe_action_t *first = before->action; pe_resource_t *first_rsc = first->rsc; if(first->required_runnable_before) { GList *clone_actions = NULL; for (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 -> %s (%p) for %s", first->uuid, before->action->uuid, before->action->rsc, start->uuid); CRM_ASSERT(before->action->rsc); first_rsc = before->action->rsc; break; } } else if(!pcmk__str_eq(first->task, RSC_START, pcmk__str_casei)) { crm_trace("Not a start op %s for %s", first->uuid, start->uuid); } if(first_rsc == NULL) { continue; } else if(uber_parent(first_rsc) == uber_parent(start->rsc)) { crm_trace("Same parent %s for %s", first_rsc->id, start->uuid); continue; } else if(FALSE && pe_rsc_is_clone(uber_parent(first_rsc)) == FALSE) { crm_trace("Not a clone %s for %s", first_rsc->id, start->uuid); continue; } crm_err("Applying %s before %s %d", first->uuid, start->uuid, uber_parent(first_rsc)->variant); for (pIter = probes; pIter != NULL; pIter = pIter->next) { pe_action_t *probe = (pe_action_t *) pIter->data; crm_err("Ordering %s before %s", first->uuid, probe->uuid); order_actions(first, probe, pe_order_optional); } } } #endif } static void order_probes(pe_working_set_t * data_set) { order_first_probes(data_set); order_then_probes(data_set); } gboolean stage7(pe_working_set_t * data_set) { pcmk__output_t *prev_out = data_set->priv; pcmk__output_t *out = NULL; GList *gIter = NULL; crm_trace("Applying ordering constraints"); /* Don't ask me why, but apparently they need to be processed in * the order they were created in... go figure * * Also g_list_append() has horrendous performance characteristics * So we need to use g_list_prepend() and then reverse the list here */ data_set->ordering_constraints = g_list_reverse(data_set->ordering_constraints); for (gIter = data_set->ordering_constraints; gIter != NULL; gIter = gIter->next) { pe__ordering_t *order = gIter->data; pe_resource_t *rsc = order->lh_rsc; if (rsc != NULL) { rsc_order_first(rsc, order, data_set); continue; } rsc = order->rh_rsc; if (rsc != NULL) { rsc_order_then(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->type); } } g_list_foreach(data_set->actions, (GFunc) update_colo_start_chain, data_set); crm_trace("Ordering probes"); order_probes(data_set); crm_trace("Updating %d actions", g_list_length(data_set->actions)); g_list_foreach(data_set->actions, (GFunc) update_action, data_set); // Check for invalid orderings for (gIter = data_set->actions; gIter != NULL; gIter = gIter->next) { pe_action_t *action = (pe_action_t *) gIter->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 (pcmk__ordering_is_invalid(action, input)) { input->type = pe_order_none; } } } /* stage7 only ever outputs to the log, so ignore whatever output object was * previously set and just log instead. */ out = pcmk__new_logger(); if (out == NULL) { return FALSE; } pcmk__output_set_log_level(out, LOG_NOTICE); data_set->priv = out; out->begin_list(out, NULL, NULL, "Actions"); LogNodeActions(data_set); g_list_foreach(data_set->resources, (GFunc) LogActions, data_set); out->end_list(out); out->finish(out, CRM_EX_OK, true, NULL); pcmk__output_free(out); data_set->priv = prev_out; return TRUE; } static int transition_id = -1; /*! * \internal * \brief Log a message after calculating a transition * * \param[in] filename Where transition input is stored */ void pcmk__log_transition_summary(const char *filename) { if (was_processing_error) { crm_err("Calculated transition %d (with errors)%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"); } } /* * Create a dependency graph to send to the transitioner (via the controller) */ gboolean stage8(pe_working_set_t * data_set) { GList *gIter = 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, pe_flag_start_failure_fatal)) { crm_xml_add(data_set->graph, "failed-start-offset", "INFINITY"); } else { crm_xml_add(data_set->graph, "failed-start-offset", "1"); } value = pe_pref(data_set->config_hash, "batch-limit"); crm_xml_add(data_set->graph, "batch-limit", value); crm_xml_add_int(data_set->graph, "transition_id", transition_id); value = pe_pref(data_set->config_hash, "migration-limit"); if ((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. */ gIter = data_set->resources; for (; gIter != NULL; gIter = gIter->next) { pe_resource_t *rsc = (pe_resource_t *) gIter->data; pe_rsc_trace(rsc, "processing actions for rsc=%s", rsc->id); rsc->cmds->expand(rsc, data_set); } crm_log_xml_trace(data_set->graph, "created resource-driven action list"); /* pseudo action to distribute list of nodes with maintenance state update */ add_maintenance_update(data_set); /* catch any non-resource specific actions */ crm_trace("processing non-resource actions"); gIter = data_set->actions; for (; gIter != NULL; gIter = gIter->next) { pe_action_t *action = (pe_action_t *) gIter->data; if (action->rsc && action->node && action->node->details->shutdown && !pcmk_is_set(action->rsc->flags, pe_rsc_maintenance) && !pcmk_any_flags_set(action->flags, pe_action_optional|pe_action_runnable) && pcmk__str_eq(action->task, RSC_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, pe_flag_have_quorum) || data_set->no_quorum_policy == no_quorum_ignore) { crm_crit("Cannot %s node '%s' because of %s:%s%s (%s)", action->node->details->unclean ? "fence" : "shut down", action->node->details->uname, action->rsc->id, pcmk_is_set(action->rsc->flags, pe_rsc_managed)? " blocked" : " unmanaged", pcmk_is_set(action->rsc->flags, pe_rsc_failed)? " failed" : "", action->uuid); } } graph_element_from_action(action, data_set); } crm_log_xml_trace(data_set->graph, "created generic action list"); crm_trace("Created transition graph %d.", transition_id); return TRUE; } void LogNodeActions(pe_working_set_t * data_set) { pcmk__output_t *out = data_set->priv; GList *gIter = NULL; for (gIter = data_set->actions; gIter != NULL; gIter = gIter->next) { char *node_name = NULL; char *task = NULL; pe_action_t *action = (pe_action_t *) gIter->data; if (action->rsc != NULL) { continue; } else if (pcmk_is_set(action->flags, pe_action_optional)) { continue; } if (pe__is_guest_node(action->node)) { node_name = crm_strdup_printf("%s (resource: %s)", action->node->details->uname, action->node->details->remote_rsc->container->id); } else if(action->node) { node_name = crm_strdup_printf("%s", action->node->details->uname); } if (pcmk__str_eq(action->task, CRM_OP_SHUTDOWN, pcmk__str_casei)) { task = strdup("Shutdown"); } else if (pcmk__str_eq(action->task, CRM_OP_FENCE, pcmk__str_casei)) { const char *op = g_hash_table_lookup(action->meta, "stonith_action"); task = crm_strdup_printf("Fence (%s)", op); } out->message(out, "node-action", task, node_name, action->reason); free(node_name); free(task); } } diff --git a/lib/pacemaker/pcmk_sched_fencing.c b/lib/pacemaker/pcmk_sched_fencing.c new file mode 100644 index 0000000000..307b871d94 --- /dev/null +++ b/lib/pacemaker/pcmk_sched_fencing.c @@ -0,0 +1,486 @@ +/* + * Copyright 2004-2021 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 + +#include + +#include +#include +#include +#include "libpacemaker_private.h" + +/*! + * \internal + * \brief Check whether a resource is known on a particular node + * + * \param[in] rsc Resource to check + * \param[in] node Node to check + * + * \return TRUE if resource (or parent if an anonymous clone) is known + */ +static bool +rsc_is_known_on(pe_resource_t *rsc, const pe_node_t *node) +{ + if (pe_hash_table_lookup(rsc->known_on, node->details->id)) { + return TRUE; + + } else if ((rsc->variant == pe_native) + && pe_rsc_is_anon_clone(rsc->parent) + && pe_hash_table_lookup(rsc->parent->known_on, node->details->id)) { + /* We check only the parent, not the uber-parent, because we cannot + * assume that the resource is known if it is in an anonymously cloned + * group (which may be only partially known). + */ + return TRUE; + } + return FALSE; +} + +/*! + * \internal + * \brief Order a resource's start and promote actions relative to fencing + * + * \param[in] rsc Resource to be ordered + * \param[in] stonith_op Fence action + * \param[in] data_set Cluster working set + */ +static void +order_start_vs_fencing(pe_resource_t *rsc, pe_action_t *stonith_op, + pe_working_set_t *data_set) +{ + pe_node_t *target; + GList *gIter = NULL; + + CRM_CHECK(stonith_op && stonith_op->node, return); + target = stonith_op->node; + + for (gIter = rsc->actions; gIter != NULL; gIter = gIter->next) { + pe_action_t *action = (pe_action_t *) gIter->data; + + switch (action->needs) { + case rsc_req_nothing: + // Anything other than start or promote requires nothing + break; + + case rsc_req_stonith: + order_actions(stonith_op, action, pe_order_optional); + break; + + case rsc_req_quorum: + if (pcmk__str_eq(action->task, RSC_START, pcmk__str_casei) + && pe_hash_table_lookup(rsc->allowed_nodes, target->details->id) + && !rsc_is_known_on(rsc, target)) { + + /* If we don't know the status of the resource on the node + * we're about to shoot, we have to assume it may be active + * there. Order the resource start after the fencing. This + * is analogous to waiting for all the probes for a resource + * to complete before starting it. + * + * The most likely explanation is that the DC died and took + * its status with it. + */ + pe_rsc_debug(rsc, "Ordering %s after %s recovery", action->uuid, + target->details->uname); + order_actions(stonith_op, action, + pe_order_optional | pe_order_runnable_left); + } + break; + } + } +} + +/*! + * \internal + * \brief Order a resource's stop and demote actions relative to fencing + * + * \param[in] rsc Resource to be ordered + * \param[in] stonith_op Fence action + * \param[in] data_set Cluster working set + */ +static void +order_stop_vs_fencing(pe_resource_t *rsc, pe_action_t *stonith_op, + pe_working_set_t *data_set) +{ + GList *gIter = NULL; + GList *action_list = NULL; + bool order_implicit = false; + + pe_resource_t *top = uber_parent(rsc); + pe_action_t *parent_stop = NULL; + pe_node_t *target; + + CRM_CHECK(stonith_op && stonith_op->node, return); + target = stonith_op->node; + + /* Get a list of stop actions potentially implied by the fencing */ + action_list = pe__resource_actions(rsc, target, RSC_STOP, FALSE); + + /* If resource requires fencing, implicit actions must occur after fencing. + * + * Implied stops and demotes of resources running on guest nodes are always + * ordered after fencing, even if the resource does not require fencing, + * because guest node "fencing" is actually just a resource stop. + */ + if (pcmk_is_set(rsc->flags, pe_rsc_needs_fencing) + || pe__is_guest_node(target)) { + + order_implicit = true; + } + + if (action_list && order_implicit) { + parent_stop = find_first_action(top->actions, NULL, RSC_STOP, NULL); + } + + for (gIter = action_list; gIter != NULL; gIter = gIter->next) { + pe_action_t *action = (pe_action_t *) gIter->data; + + // The stop would never complete, so convert it into a pseudo-action. + pe__set_action_flags(action, pe_action_pseudo|pe_action_runnable); + + if (order_implicit) { + pe__set_action_flags(action, pe_action_implied_by_stonith); + + /* Order the stonith before the parent stop (if any). + * + * Also order the stonith before the resource stop, unless the + * resource is inside a bundle -- that would cause a graph loop. + * We can rely on the parent stop's ordering instead. + * + * User constraints must not order a resource in a guest node + * relative to the guest node container resource. The + * pe_order_preserve flag marks constraints as generated by the + * cluster and thus immune to that check (and is irrelevant if + * target is not a guest). + */ + if (!pe_rsc_is_bundled(rsc)) { + order_actions(stonith_op, action, pe_order_preserve); + } + order_actions(stonith_op, parent_stop, pe_order_preserve); + } + + if (pcmk_is_set(rsc->flags, pe_rsc_failed)) { + crm_notice("Stop of failed resource %s is implicit %s %s is fenced", + rsc->id, (order_implicit? "after" : "because"), + target->details->uname); + } else { + crm_info("%s is implicit %s %s is fenced", + action->uuid, (order_implicit? "after" : "because"), + target->details->uname); + } + + if (pcmk_is_set(rsc->flags, pe_rsc_notify)) { + /* Create a second notification that will be delivered + * immediately after the node is fenced + * + * Basic problem: + * - C is a clone active on the node to be shot and stopping on another + * - R is a resource that depends on C + * + * + C.stop depends on R.stop + * + C.stopped depends on STONITH + * + C.notify depends on C.stopped + * + C.healthy depends on C.notify + * + R.stop depends on C.healthy + * + * The extra notification here changes + * + C.healthy depends on C.notify + * into: + * + C.healthy depends on C.notify' + * + C.notify' depends on STONITH' + * thus breaking the loop + */ + create_secondary_notification(action, rsc, stonith_op, data_set); + } + +#if 0 + /* It might be a good idea to stop healthy resources on a node about to + * be fenced, when possible. + * + * However, fencing must be done before a failed resource's + * (pseudo-)stop action, so that could create a loop. For example, given + * a group of A and B running on node N with a failed stop of B: + * + * fence N -> stop B (pseudo-op) -> stop A -> fence N + * + * The block below creates the stop A -> fence N ordering and therefore + * must (at least for now) be disabled. Instead, run the block above and + * treat all resources on N as B would be (i.e., as a pseudo-op after + * the fencing). + * + * @TODO Maybe break the "A requires B" dependency in update_action() + * and use this block for healthy resources instead of the above. + */ + crm_info("Moving healthy resource %s off %s before fencing", + rsc->id, node->details->uname); + custom_action_order(rsc, stop_key(rsc), NULL, NULL, + strdup(CRM_OP_FENCE), stonith_op, + pe_order_optional, data_set); +#endif + } + + g_list_free(action_list); + + /* Get a list of demote actions potentially implied by the fencing */ + action_list = pe__resource_actions(rsc, target, RSC_DEMOTE, FALSE); + + for (gIter = action_list; gIter != NULL; gIter = gIter->next) { + pe_action_t *action = (pe_action_t *) gIter->data; + + if (!(action->node->details->online) || action->node->details->unclean + || pcmk_is_set(rsc->flags, pe_rsc_failed)) { + + if (pcmk_is_set(rsc->flags, pe_rsc_failed)) { + pe_rsc_info(rsc, + "Demote of failed resource %s is implicit after %s is fenced", + rsc->id, target->details->uname); + } else { + pe_rsc_info(rsc, "%s is implicit after %s is fenced", + action->uuid, target->details->uname); + } + + /* The demote would never complete and is now implied by the + * fencing, so convert it into a pseudo-action. + */ + pe__set_action_flags(action, pe_action_pseudo|pe_action_runnable); + + if (pe_rsc_is_bundled(rsc)) { + // Do nothing, let recovery be ordered after parent's implied stop + + } else if (order_implicit) { + order_actions(stonith_op, action, pe_order_preserve|pe_order_optional); + } + } + } + + g_list_free(action_list); +} + +/*! + * \internal + * \brief Order resource actions properly relative to fencing + * + * \param[in] rsc Resource whose actions should be ordered + * \param[in] stonith_op Fencing operation to be ordered against + * \param[in] data_set Cluster working set + */ +static void +rsc_stonith_ordering(pe_resource_t *rsc, pe_action_t *stonith_op, + pe_working_set_t *data_set) +{ + if (rsc->children) { + GList *gIter = NULL; + + for (gIter = rsc->children; gIter != NULL; gIter = gIter->next) { + pe_resource_t *child_rsc = (pe_resource_t *) gIter->data; + + rsc_stonith_ordering(child_rsc, stonith_op, data_set); + } + + } else if (!pcmk_is_set(rsc->flags, pe_rsc_managed)) { + pe_rsc_trace(rsc, + "Skipping fencing constraints for unmanaged resource: %s", + rsc->id); + + } else { + order_start_vs_fencing(rsc, stonith_op, data_set); + order_stop_vs_fencing(rsc, stonith_op, data_set); + } +} + +/*! + * \internal + * \brief Order all actions appropriately relative to a fencing operation + * + * Ensure start operations of affected resources are ordered after fencing, + * imply stop and demote operations of affected resources by marking them as + * pseudo-actions, etc. + * + * \param[in] stonith_op Fencing operation + * \param[in,out] data_set Working set of cluster + */ +void +pcmk__order_vs_fence(pe_action_t *stonith_op, pe_working_set_t *data_set) +{ + CRM_CHECK(stonith_op && data_set, return); + for (GList *r = data_set->resources; r != NULL; r = r->next) { + rsc_stonith_ordering((pe_resource_t *) r->data, stonith_op, data_set); + } +} + +/*! + * \internal + * \brief Order an action after unfencing + * + * \param[in] rsc Resource that action is for + * \param[in] node Node that action is on + * \param[in] action Action to be ordered after unfencing + * \param[in] order Ordering flags + * \param[in] data_set Cluster working set + */ +void +pcmk__order_vs_unfence(pe_resource_t *rsc, pe_node_t *node, pe_action_t *action, + enum pe_ordering order, pe_working_set_t *data_set) +{ + /* When unfencing is in use, we order unfence actions before any probe or + * start of resources that require unfencing, and also of fence devices. + * + * This might seem to violate the principle that fence devices require + * only quorum. However, fence agents that unfence often don't have enough + * information to even probe or start unless the node is first unfenced. + */ + if (pcmk__is_unfence_device(rsc, data_set) + || pcmk_is_set(rsc->flags, pe_rsc_needs_unfencing)) { + + /* Start with an optional ordering. Requiring unfencing would result in + * the node being unfenced, and all its resources being stopped, + * whenever a new resource is added -- which would be highly suboptimal. + */ + pe_action_t *unfence = pe_fence_op(node, "on", TRUE, NULL, FALSE, data_set); + + order_actions(unfence, action, order); + + if (!pcmk__node_unfenced(node)) { + // But unfencing is required if it has never been done + char *reason = crm_strdup_printf("required by %s %s", + rsc->id, action->task); + + trigger_unfencing(NULL, node, reason, NULL, data_set); + free(reason); + } + } +} + +/*! + * \internal + * \brief Create pseudo-op for guest node fence, and order relative to it + * + * \param[in] node Guest node to fence + * \param[in] data_set Working set of CIB state + */ +void +pcmk__fence_guest(pe_node_t *node, pe_working_set_t *data_set) +{ + pe_resource_t *container = node->details->remote_rsc->container; + pe_action_t *stop = NULL; + pe_action_t *stonith_op = NULL; + + /* The fence action is just a label; we don't do anything differently for + * off vs. reboot. We specify it explicitly, rather than let it default to + * cluster's default action, because we are not _initiating_ fencing -- we + * are creating a pseudo-event to describe fencing that is already occurring + * by other means (container recovery). + */ + const char *fence_action = "off"; + + /* Check whether guest's container resource has any explicit stop or + * start (the stop may be implied by fencing of the guest's host). + */ + if (container) { + stop = find_first_action(container->actions, NULL, CRMD_ACTION_STOP, + NULL); + + if (find_first_action(container->actions, NULL, CRMD_ACTION_START, + NULL)) { + fence_action = "reboot"; + } + } + + /* Create a fence pseudo-event, so we have an event to order actions + * against, and the controller can always detect it. + */ + stonith_op = pe_fence_op(node, fence_action, FALSE, "guest is unclean", + FALSE, data_set); + pe__set_action_flags(stonith_op, pe_action_pseudo|pe_action_runnable); + + /* We want to imply stops/demotes after the guest is stopped, not wait until + * it is restarted, so we always order pseudo-fencing after stop, not start + * (even though start might be closer to what is done for a real reboot). + */ + if ((stop != NULL) && pcmk_is_set(stop->flags, pe_action_pseudo)) { + pe_action_t *parent_stonith_op = pe_fence_op(stop->node, NULL, FALSE, + NULL, FALSE, data_set); + + crm_info("Implying guest node %s is down (action %d) after %s fencing", + node->details->uname, stonith_op->id, + stop->node->details->uname); + order_actions(parent_stonith_op, stonith_op, + pe_order_runnable_left|pe_order_implies_then); + + } else if (stop) { + order_actions(stop, stonith_op, + pe_order_runnable_left|pe_order_implies_then); + crm_info("Implying guest node %s is down (action %d) " + "after container %s is stopped (action %d)", + node->details->uname, stonith_op->id, + container->id, stop->id); + } else { + /* If we're fencing the guest node but there's no stop for the guest + * resource, we must think the guest is already stopped. However, we may + * think so because its resource history was just cleaned. To avoid + * unnecessarily considering the guest node down if it's really up, + * order the pseudo-fencing after any stop of the connection resource, + * which will be ordered after any container (re-)probe. + */ + stop = find_first_action(node->details->remote_rsc->actions, NULL, + RSC_STOP, NULL); + + if (stop) { + order_actions(stop, stonith_op, pe_order_optional); + crm_info("Implying guest node %s is down (action %d) " + "after connection is stopped (action %d)", + node->details->uname, stonith_op->id, stop->id); + } else { + /* Not sure why we're fencing, but everything must already be + * cleanly stopped. + */ + crm_info("Implying guest node %s is down (action %d) ", + node->details->uname, stonith_op->id); + } + } + + // Order/imply other actions relative to pseudo-fence as with real fence + pcmk__order_vs_fence(stonith_op, data_set); +} + +/*! + * \internal + * \brief Check whether node has already been unfenced + * + * \param[in] node Node to check + * + * \return true if node has a nonzero #node-unfenced attribute (or none), + * otherwise false + */ +bool +pcmk__node_unfenced(pe_node_t *node) +{ + const char *unfenced = pe_node_attribute_raw(node, CRM_ATTR_UNFENCED); + + return !pcmk__str_eq(unfenced, "0", pcmk__str_null_matches); +} + +/*! + * \internal + * \brief Check whether a resource is a fencing device that supports unfencing + * + * \param[in] rsc Resource to check + * \param[in] data_set Cluster working set + * + * \return true if \p rsc is a fencing device that supports unfencing, + * otherwise false + */ +bool +pcmk__is_unfence_device(const pe_resource_t *rsc, + const pe_working_set_t *data_set) +{ + return pcmk_is_set(rsc->flags, pe_rsc_fence_device) + && pcmk_is_set(data_set->flags, pe_flag_enable_unfencing); +} diff --git a/lib/pacemaker/pcmk_sched_native.c b/lib/pacemaker/pcmk_sched_native.c index 5121f2d0ff..d7194bc5da 100644 --- a/lib/pacemaker/pcmk_sched_native.c +++ b/lib/pacemaker/pcmk_sched_native.c @@ -1,3210 +1,2900 @@ /* * Copyright 2004-2021 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 #include #include #include #include #include #include +#include "libpacemaker_private.h" + // The controller removes the resource from the CIB, making this redundant // #define DELETE_THEN_REFRESH 1 #define INFINITY_HACK (INFINITY * -100) #define VARIANT_NATIVE 1 #include extern bool pcmk__is_daemon; static void Recurring(pe_resource_t *rsc, pe_action_t *start, pe_node_t *node, pe_working_set_t *data_set); static void RecurringOp(pe_resource_t *rsc, pe_action_t *start, pe_node_t *node, xmlNode *operation, pe_working_set_t *data_set); static void Recurring_Stopped(pe_resource_t *rsc, pe_action_t *start, pe_node_t *node, pe_working_set_t *data_set); static void RecurringOp_Stopped(pe_resource_t *rsc, pe_action_t *start, pe_node_t *node, xmlNode *operation, pe_working_set_t *data_set); void ReloadRsc(pe_resource_t * rsc, pe_node_t *node, pe_working_set_t * data_set); gboolean DeleteRsc(pe_resource_t * rsc, pe_node_t * node, gboolean optional, pe_working_set_t * data_set); gboolean StopRsc(pe_resource_t * rsc, pe_node_t * next, gboolean optional, pe_working_set_t * data_set); gboolean StartRsc(pe_resource_t * rsc, pe_node_t * next, gboolean optional, pe_working_set_t * data_set); gboolean DemoteRsc(pe_resource_t * rsc, pe_node_t * next, gboolean optional, pe_working_set_t * data_set); gboolean PromoteRsc(pe_resource_t * rsc, pe_node_t * next, gboolean optional, pe_working_set_t * data_set); gboolean RoleError(pe_resource_t * rsc, pe_node_t * next, gboolean optional, pe_working_set_t * data_set); gboolean NullOp(pe_resource_t * rsc, pe_node_t * next, gboolean optional, pe_working_set_t * data_set); /* This array says what the *next* role should be when transitioning from one * role to another. For example going from Stopped to Promoted, the next role is * RSC_ROLE_UNPROMOTED, because the resource must be started before being promoted. * The current state then becomes Started, which is fed into this array again, * giving a next role of RSC_ROLE_PROMOTED. */ static enum rsc_role_e rsc_state_matrix[RSC_ROLE_MAX][RSC_ROLE_MAX] = { /* Current state Next state*/ /* Unknown Stopped Started Unpromoted Promoted */ /* Unknown */ { RSC_ROLE_UNKNOWN, RSC_ROLE_STOPPED, RSC_ROLE_STOPPED, RSC_ROLE_STOPPED, RSC_ROLE_STOPPED }, /* Stopped */ { RSC_ROLE_STOPPED, RSC_ROLE_STOPPED, RSC_ROLE_STARTED, RSC_ROLE_UNPROMOTED, RSC_ROLE_UNPROMOTED }, /* Started */ { RSC_ROLE_STOPPED, RSC_ROLE_STOPPED, RSC_ROLE_STARTED, RSC_ROLE_UNPROMOTED, RSC_ROLE_PROMOTED }, /* Unpromoted */ { RSC_ROLE_STOPPED, RSC_ROLE_STOPPED, RSC_ROLE_STOPPED, RSC_ROLE_UNPROMOTED, RSC_ROLE_PROMOTED }, /* Promoted */ { RSC_ROLE_STOPPED, RSC_ROLE_UNPROMOTED, RSC_ROLE_UNPROMOTED, RSC_ROLE_UNPROMOTED, RSC_ROLE_PROMOTED }, }; typedef gboolean (*rsc_transition_fn)(pe_resource_t *rsc, pe_node_t *next, gboolean optional, pe_working_set_t *data_set); // This array picks the function needed to transition from one role to another static rsc_transition_fn rsc_action_matrix[RSC_ROLE_MAX][RSC_ROLE_MAX] = { /* Current state Next state */ /* Unknown Stopped Started Unpromoted Promoted */ /* Unknown */ { RoleError, StopRsc, RoleError, RoleError, RoleError, }, /* Stopped */ { RoleError, NullOp, StartRsc, StartRsc, RoleError, }, /* Started */ { RoleError, StopRsc, NullOp, NullOp, PromoteRsc, }, /* Unpromoted */ { RoleError, StopRsc, StopRsc, NullOp, PromoteRsc, }, /* Promoted */ { RoleError, DemoteRsc, DemoteRsc, DemoteRsc, NullOp, }, }; #define clear_node_weights_flags(nw_flags, nw_rsc, flags_to_clear) do { \ flags = pcmk__clear_flags_as(__func__, __LINE__, LOG_TRACE, \ "Node weight", (nw_rsc)->id, (flags), \ (flags_to_clear), #flags_to_clear); \ } while (0) static gboolean native_choose_node(pe_resource_t * rsc, pe_node_t * prefer, pe_working_set_t * data_set) { GList *nodes = NULL; pe_node_t *chosen = NULL; pe_node_t *best = NULL; int multiple = 1; int length = 0; gboolean result = FALSE; process_utilization(rsc, &prefer, data_set); if (!pcmk_is_set(rsc->flags, pe_rsc_provisional)) { return rsc->allocated_to ? TRUE : FALSE; } // Sort allowed nodes by weight if (rsc->allowed_nodes) { length = g_hash_table_size(rsc->allowed_nodes); } if (length > 0) { nodes = g_hash_table_get_values(rsc->allowed_nodes); nodes = sort_nodes_by_weight(nodes, pe__current_node(rsc), data_set); // First node in sorted list has the best score best = g_list_nth_data(nodes, 0); } if (prefer && nodes) { chosen = g_hash_table_lookup(rsc->allowed_nodes, prefer->details->id); if (chosen == NULL) { pe_rsc_trace(rsc, "Preferred node %s for %s was unknown", prefer->details->uname, rsc->id); /* Favor the preferred node as long as its weight is at least as good as * the best allowed node's. * * An alternative would be to favor the preferred node even if the best * node is better, when the best node's weight is less than INFINITY. */ } else if ((chosen->weight < 0) || (chosen->weight < best->weight)) { pe_rsc_trace(rsc, "Preferred node %s for %s was unsuitable", chosen->details->uname, rsc->id); chosen = NULL; } else if (!can_run_resources(chosen)) { pe_rsc_trace(rsc, "Preferred node %s for %s was unavailable", chosen->details->uname, rsc->id); chosen = NULL; } else { pe_rsc_trace(rsc, "Chose preferred node %s for %s (ignoring %d candidates)", chosen->details->uname, rsc->id, length); } } if ((chosen == NULL) && nodes) { /* Either there is no preferred node, or the preferred node is not * available, but there are other nodes allowed to run the resource. */ chosen = best; pe_rsc_trace(rsc, "Chose node %s for %s from %d candidates", chosen ? chosen->details->uname : "", rsc->id, length); if (!pe_rsc_is_unique_clone(rsc->parent) && chosen && (chosen->weight > 0) && can_run_resources(chosen)) { /* If the resource is already running on a node, prefer that node if * it is just as good as the chosen node. * * We don't do this for unique clone instances, because * distribute_children() has already assigned instances to their * running nodes when appropriate, and if we get here, we don't want * remaining unallocated instances to prefer a node that's already * running another instance. */ pe_node_t *running = pe__current_node(rsc); if (running && (can_run_resources(running) == FALSE)) { pe_rsc_trace(rsc, "Current node for %s (%s) can't run resources", rsc->id, running->details->uname); } else if (running) { for (GList *iter = nodes->next; iter; iter = iter->next) { pe_node_t *tmp = (pe_node_t *) iter->data; if (tmp->weight != chosen->weight) { // The nodes are sorted by weight, so no more are equal break; } if (tmp->details == running->details) { // Scores are equal, so prefer the current node chosen = tmp; } multiple++; } } } } if (multiple > 1) { static char score[33]; int log_level = (chosen->weight >= INFINITY)? LOG_WARNING : LOG_INFO; score2char_stack(chosen->weight, score, sizeof(score)); do_crm_log(log_level, "Chose node %s for %s from %d nodes with score %s", chosen->details->uname, rsc->id, multiple, score); } result = native_assign_node(rsc, chosen, FALSE); g_list_free(nodes); return result; } /*! * \internal * \brief Find score of highest-scored node that matches colocation attribute * * \param[in] rsc Resource whose allowed nodes should be searched * \param[in] attr Colocation attribute name (must not be NULL) * \param[in] value Colocation attribute value to require */ static int best_node_score_matching_attr(const pe_resource_t *rsc, const char *attr, const char *value) { GHashTableIter iter; pe_node_t *node = NULL; int best_score = -INFINITY; const char *best_node = NULL; // Find best allowed node with matching attribute g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **) &node)) { if ((node->weight > best_score) && can_run_resources(node) && pcmk__str_eq(value, pe_node_attribute_raw(node, attr), pcmk__str_casei)) { best_score = node->weight; best_node = node->details->uname; } } if (!pcmk__str_eq(attr, CRM_ATTR_UNAME, pcmk__str_casei)) { if (best_node == NULL) { crm_info("No allowed node for %s matches node attribute %s=%s", rsc->id, attr, value); } else { crm_info("Allowed node %s for %s had best score (%d) " "of those matching node attribute %s=%s", best_node, rsc->id, best_score, attr, value); } } return best_score; } /*! * \internal * \brief Add resource's colocation matches to current node allocation scores * * For each node in a given table, if any of a given resource's allowed nodes * have a matching value for the colocation attribute, add the highest of those * nodes' scores to the node's score. * * \param[in,out] nodes Hash table of nodes with allocation scores so far * \param[in] rsc Resource whose allowed nodes should be compared * \param[in] attr Colocation attribute that must match (NULL for default) * \param[in] factor Factor by which to multiply scores being added * \param[in] only_positive Whether to add only positive scores */ static void add_node_scores_matching_attr(GHashTable *nodes, const pe_resource_t *rsc, const char *attr, float factor, bool only_positive) { GHashTableIter iter; pe_node_t *node = NULL; if (attr == NULL) { attr = CRM_ATTR_UNAME; } // Iterate through each node g_hash_table_iter_init(&iter, nodes); while (g_hash_table_iter_next(&iter, NULL, (void **)&node)) { float weight_f = 0; int weight = 0; int score = 0; int new_score = 0; score = best_node_score_matching_attr(rsc, attr, pe_node_attribute_raw(node, attr)); if ((factor < 0) && (score < 0)) { /* Negative preference for a node with a negative score * should not become a positive preference. * * @TODO Consider filtering only if weight is -INFINITY */ crm_trace("%s: Filtering %d + %f * %d (double negative disallowed)", node->details->uname, node->weight, factor, score); continue; } if (node->weight == INFINITY_HACK) { crm_trace("%s: Filtering %d + %f * %d (node was marked unusable)", node->details->uname, node->weight, factor, score); continue; } weight_f = factor * score; // Round the number; see http://c-faq.com/fp/round.html weight = (int) ((weight_f < 0)? (weight_f - 0.5) : (weight_f + 0.5)); /* Small factors can obliterate the small scores that are often actually * used in configurations. If the score and factor are nonzero, ensure * that the result is nonzero as well. */ if ((weight == 0) && (score != 0)) { if (factor > 0.0) { weight = 1; } else if (factor < 0.0) { weight = -1; } } new_score = pe__add_scores(weight, node->weight); if (only_positive && (new_score < 0) && (node->weight > 0)) { crm_trace("%s: Filtering %d + %f * %d = %d " "(negative disallowed, marking node unusable)", node->details->uname, node->weight, factor, score, new_score); node->weight = INFINITY_HACK; continue; } if (only_positive && (new_score < 0) && (node->weight == 0)) { crm_trace("%s: Filtering %d + %f * %d = %d (negative disallowed)", node->details->uname, node->weight, factor, score, new_score); continue; } crm_trace("%s: %d + %f * %d = %d", node->details->uname, node->weight, factor, score, new_score); node->weight = new_score; } } static inline bool is_nonempty_group(pe_resource_t *rsc) { return rsc && (rsc->variant == pe_group) && (rsc->children != NULL); } /*! * \internal * \brief Incorporate colocation constraint scores into node weights * * \param[in,out] rsc Resource being placed * \param[in] rhs ID of 'with' resource * \param[in,out] nodes Nodes, with scores as of this point * \param[in] attr Colocation attribute (ID by default) * \param[in] factor Incorporate scores multiplied by this factor * \param[in] flags Bitmask of enum pe_weights values * * \return Nodes, with scores modified by this constraint * \note This function assumes ownership of the nodes argument. The caller * should free the returned copy rather than the original. */ GHashTable * pcmk__native_merge_weights(pe_resource_t *rsc, const char *rhs, GHashTable *nodes, const char *attr, float factor, uint32_t flags) { GHashTable *work = NULL; // Avoid infinite recursion if (pcmk_is_set(rsc->flags, pe_rsc_merging)) { pe_rsc_info(rsc, "%s: Breaking dependency loop at %s", rhs, rsc->id); return nodes; } pe__set_resource_flags(rsc, pe_rsc_merging); if (pcmk_is_set(flags, pe_weights_init)) { if (is_nonempty_group(rsc)) { GList *last = g_list_last(rsc->children); pe_resource_t *last_rsc = last->data; pe_rsc_trace(rsc, "%s: Merging scores from group %s " "using last member %s (at %.6f)", rhs, rsc->id, last_rsc->id, factor); work = pcmk__native_merge_weights(last_rsc, rhs, NULL, attr, factor, flags); } else { work = pcmk__copy_node_table(rsc->allowed_nodes); } clear_node_weights_flags(flags, rsc, pe_weights_init); } else if (is_nonempty_group(rsc)) { /* The first member of the group will recursively incorporate any * constraints involving other members (including the group internal * colocation). * * @TODO The indirect colocations from the dependent group's other * members will be incorporated at full strength rather than by * factor, so the group's combined stickiness will be treated as * (factor + (#members - 1)) * stickiness. It is questionable what * the right approach should be. */ pe_rsc_trace(rsc, "%s: Merging scores from first member of group %s " "(at %.6f)", rhs, rsc->id, factor); work = pcmk__copy_node_table(nodes); work = pcmk__native_merge_weights(rsc->children->data, rhs, work, attr, factor, flags); } else { pe_rsc_trace(rsc, "%s: Merging scores from %s (at %.6f)", rhs, rsc->id, factor); work = pcmk__copy_node_table(nodes); add_node_scores_matching_attr(work, rsc, attr, factor, pcmk_is_set(flags, pe_weights_positive)); } if (can_run_any(work)) { GList *gIter = NULL; int multiplier = (factor < 0)? -1 : 1; if (pcmk_is_set(flags, pe_weights_forward)) { gIter = rsc->rsc_cons; pe_rsc_trace(rsc, "Checking additional %d optional '%s with' constraints", g_list_length(gIter), rsc->id); } else if (is_nonempty_group(rsc)) { pe_resource_t *last_rsc = g_list_last(rsc->children)->data; gIter = last_rsc->rsc_cons_lhs; pe_rsc_trace(rsc, "Checking additional %d optional 'with group %s' " "constraints using last member %s", g_list_length(gIter), rsc->id, last_rsc->id); } else { gIter = rsc->rsc_cons_lhs; pe_rsc_trace(rsc, "Checking additional %d optional 'with %s' constraints", g_list_length(gIter), rsc->id); } for (; gIter != NULL; gIter = gIter->next) { pe_resource_t *other = NULL; pcmk__colocation_t *constraint = (pcmk__colocation_t *) gIter->data; if (pcmk_is_set(flags, pe_weights_forward)) { other = constraint->rsc_rh; } else if (!pcmk__colocation_has_influence(constraint, NULL)) { continue; } else { other = constraint->rsc_lh; } pe_rsc_trace(rsc, "Optionally merging score of '%s' constraint (%s with %s)", constraint->id, constraint->rsc_lh->id, constraint->rsc_rh->id); work = pcmk__native_merge_weights(other, rhs, work, constraint->node_attribute, multiplier * constraint->score / (float) INFINITY, flags|pe_weights_rollback); pe__show_node_weights(true, NULL, rhs, work, rsc->cluster); } } else if (pcmk_is_set(flags, pe_weights_rollback)) { pe_rsc_info(rsc, "%s: Rolling back optional scores from %s", rhs, rsc->id); g_hash_table_destroy(work); pe__clear_resource_flags(rsc, pe_rsc_merging); return nodes; } if (pcmk_is_set(flags, pe_weights_positive)) { pe_node_t *node = NULL; GHashTableIter iter; g_hash_table_iter_init(&iter, work); while (g_hash_table_iter_next(&iter, NULL, (void **)&node)) { if (node->weight == INFINITY_HACK) { node->weight = 1; } } } if (nodes) { g_hash_table_destroy(nodes); } pe__clear_resource_flags(rsc, pe_rsc_merging); return work; } -static inline bool -node_has_been_unfenced(pe_node_t *node) -{ - const char *unfenced = pe_node_attribute_raw(node, CRM_ATTR_UNFENCED); - - return !pcmk__str_eq(unfenced, "0", pcmk__str_null_matches); -} - -static inline bool -is_unfence_device(pe_resource_t *rsc, pe_working_set_t *data_set) -{ - return pcmk_is_set(rsc->flags, pe_rsc_fence_device) - && pcmk_is_set(data_set->flags, pe_flag_enable_unfencing); -} - pe_node_t * pcmk__native_allocate(pe_resource_t *rsc, pe_node_t *prefer, pe_working_set_t *data_set) { GList *gIter = NULL; if (rsc->parent && !pcmk_is_set(rsc->parent->flags, pe_rsc_allocating)) { /* never allocate children on their own */ pe_rsc_debug(rsc, "Escalating allocation of %s to its parent: %s", rsc->id, rsc->parent->id); rsc->parent->cmds->allocate(rsc->parent, prefer, data_set); } if (!pcmk_is_set(rsc->flags, pe_rsc_provisional)) { return rsc->allocated_to; } if (pcmk_is_set(rsc->flags, pe_rsc_allocating)) { pe_rsc_debug(rsc, "Dependency loop detected involving %s", rsc->id); return NULL; } pe__set_resource_flags(rsc, pe_rsc_allocating); pe__show_node_weights(true, rsc, "Pre-alloc", rsc->allowed_nodes, data_set); for (gIter = rsc->rsc_cons; gIter != NULL; gIter = gIter->next) { pcmk__colocation_t *constraint = (pcmk__colocation_t *) gIter->data; GHashTable *archive = NULL; pe_resource_t *rsc_rh = constraint->rsc_rh; if ((constraint->role_lh >= RSC_ROLE_PROMOTED) || (constraint->score < 0 && constraint->score > -INFINITY)) { archive = pcmk__copy_node_table(rsc->allowed_nodes); } pe_rsc_trace(rsc, "%s: Allocating %s first (constraint=%s score=%d role=%s)", rsc->id, rsc_rh->id, constraint->id, constraint->score, role2text(constraint->role_lh)); rsc_rh->cmds->allocate(rsc_rh, NULL, data_set); rsc->cmds->rsc_colocation_lh(rsc, rsc_rh, constraint, data_set); if (archive && can_run_any(rsc->allowed_nodes) == FALSE) { pe_rsc_info(rsc, "%s: Rolling back scores from %s", rsc->id, rsc_rh->id); g_hash_table_destroy(rsc->allowed_nodes); rsc->allowed_nodes = archive; archive = NULL; } if (archive) { g_hash_table_destroy(archive); } } pe__show_node_weights(true, rsc, "Post-coloc", rsc->allowed_nodes, data_set); for (gIter = rsc->rsc_cons_lhs; gIter != NULL; gIter = gIter->next) { pcmk__colocation_t *constraint = (pcmk__colocation_t *) gIter->data; if (!pcmk__colocation_has_influence(constraint, NULL)) { continue; } pe_rsc_trace(rsc, "Merging score of '%s' constraint (%s with %s)", constraint->id, constraint->rsc_lh->id, constraint->rsc_rh->id); rsc->allowed_nodes = constraint->rsc_lh->cmds->merge_weights(constraint->rsc_lh, rsc->id, rsc->allowed_nodes, constraint->node_attribute, (float)constraint->score / INFINITY, pe_weights_rollback); } if (rsc->next_role == RSC_ROLE_STOPPED) { pe_rsc_trace(rsc, "Making sure %s doesn't get allocated", rsc->id); /* make sure it doesn't come up again */ resource_location(rsc, NULL, -INFINITY, XML_RSC_ATTR_TARGET_ROLE, data_set); } else if(rsc->next_role > rsc->role && !pcmk_is_set(data_set->flags, pe_flag_have_quorum) && data_set->no_quorum_policy == no_quorum_freeze) { crm_notice("Resource %s cannot be elevated from %s to %s: no-quorum-policy=freeze", rsc->id, role2text(rsc->role), role2text(rsc->next_role)); pe__set_next_role(rsc, rsc->role, "no-quorum-policy=freeze"); } pe__show_node_weights(!pcmk_is_set(data_set->flags, pe_flag_show_scores), rsc, __func__, rsc->allowed_nodes, data_set); if (pcmk_is_set(data_set->flags, pe_flag_stonith_enabled) && !pcmk_is_set(data_set->flags, pe_flag_have_stonith_resource)) { pe__clear_resource_flags(rsc, pe_rsc_managed); } if (!pcmk_is_set(rsc->flags, pe_rsc_managed)) { const char *reason = NULL; pe_node_t *assign_to = NULL; pe__set_next_role(rsc, rsc->role, "unmanaged"); assign_to = pe__current_node(rsc); if (assign_to == NULL) { reason = "inactive"; } else if (rsc->role == RSC_ROLE_PROMOTED) { reason = "promoted"; } else if (pcmk_is_set(rsc->flags, pe_rsc_failed)) { reason = "failed"; } else { reason = "active"; } pe_rsc_info(rsc, "Unmanaged resource %s allocated to %s: %s", rsc->id, (assign_to? assign_to->details->uname : "no node"), reason); native_assign_node(rsc, assign_to, TRUE); } else if (pcmk_is_set(data_set->flags, pe_flag_stop_everything)) { pe_rsc_debug(rsc, "Forcing %s to stop", rsc->id); native_assign_node(rsc, NULL, TRUE); } else if (pcmk_is_set(rsc->flags, pe_rsc_provisional) && native_choose_node(rsc, prefer, data_set)) { pe_rsc_trace(rsc, "Allocated resource %s to %s", rsc->id, rsc->allocated_to->details->uname); } else if (rsc->allocated_to == NULL) { if (!pcmk_is_set(rsc->flags, pe_rsc_orphan)) { pe_rsc_info(rsc, "Resource %s cannot run anywhere", rsc->id); } else if (rsc->running_on != NULL) { pe_rsc_info(rsc, "Stopping orphan resource %s", rsc->id); } } else { pe_rsc_debug(rsc, "Pre-Allocated resource %s to %s", rsc->id, rsc->allocated_to->details->uname); } pe__clear_resource_flags(rsc, pe_rsc_allocating); if (rsc->is_remote_node) { pe_node_t *remote_node = pe_find_node(data_set->nodes, rsc->id); CRM_ASSERT(remote_node != NULL); if (rsc->allocated_to && rsc->next_role != RSC_ROLE_STOPPED) { crm_trace("Setting Pacemaker Remote node %s to ONLINE", remote_node->details->id); remote_node->details->online = TRUE; /* We shouldn't consider an unseen remote-node unclean if we are going * to try and connect to it. Otherwise we get an unnecessary fence */ if (remote_node->details->unseen == TRUE) { remote_node->details->unclean = FALSE; } } else { crm_trace("Setting Pacemaker Remote node %s to SHUTDOWN (next role %s, %sallocated)", remote_node->details->id, role2text(rsc->next_role), (rsc->allocated_to? "" : "un")); remote_node->details->shutdown = TRUE; } } return rsc->allocated_to; } static gboolean is_op_dup(pe_resource_t *rsc, const char *name, guint interval_ms) { gboolean dup = FALSE; const char *id = NULL; const char *value = NULL; xmlNode *operation = NULL; guint interval2_ms = 0; CRM_ASSERT(rsc); for (operation = pcmk__xe_first_child(rsc->ops_xml); operation != NULL; operation = pcmk__xe_next(operation)) { if (pcmk__str_eq((const char *)operation->name, "op", pcmk__str_none)) { value = crm_element_value(operation, "name"); if (!pcmk__str_eq(value, name, pcmk__str_casei)) { continue; } value = crm_element_value(operation, XML_LRM_ATTR_INTERVAL); interval2_ms = crm_parse_interval_spec(value); if (interval_ms != interval2_ms) { continue; } if (id == NULL) { id = ID(operation); } else { pcmk__config_err("Operation %s is duplicate of %s (do not use " "same name and interval combination more " "than once per resource)", ID(operation), id); dup = TRUE; } } } return dup; } static bool op_cannot_recur(const char *name) { return pcmk__strcase_any_of(name, RSC_STOP, RSC_START, RSC_DEMOTE, RSC_PROMOTE, NULL); } static void RecurringOp(pe_resource_t * rsc, pe_action_t * start, pe_node_t * node, xmlNode * operation, pe_working_set_t * data_set) { char *key = NULL; const char *name = NULL; const char *role = NULL; const char *interval_spec = NULL; const char *node_uname = node? node->details->uname : "n/a"; guint interval_ms = 0; pe_action_t *mon = NULL; gboolean is_optional = TRUE; GList *possible_matches = NULL; CRM_ASSERT(rsc); /* Only process for the operations without role="Stopped" */ role = crm_element_value(operation, "role"); if (role && text2role(role) == RSC_ROLE_STOPPED) { return; } interval_spec = crm_element_value(operation, XML_LRM_ATTR_INTERVAL); interval_ms = crm_parse_interval_spec(interval_spec); if (interval_ms == 0) { return; } name = crm_element_value(operation, "name"); if (is_op_dup(rsc, name, interval_ms)) { crm_trace("Not creating duplicate recurring action %s for %dms %s", ID(operation), interval_ms, name); return; } if (op_cannot_recur(name)) { pcmk__config_err("Ignoring %s because action '%s' cannot be recurring", ID(operation), name); return; } key = pcmk__op_key(rsc->id, name, interval_ms); if (find_rsc_op_entry(rsc, key) == NULL) { crm_trace("Not creating recurring action %s for disabled resource %s", ID(operation), rsc->id); free(key); return; } pe_rsc_trace(rsc, "Creating recurring action %s for %s in role %s on %s", ID(operation), rsc->id, role2text(rsc->next_role), node_uname); if (start != NULL) { pe_rsc_trace(rsc, "Marking %s %s due to %s", key, pcmk_is_set(start->flags, pe_action_optional)? "optional" : "mandatory", start->uuid); is_optional = (rsc->cmds->action_flags(start, NULL) & pe_action_optional); } else { pe_rsc_trace(rsc, "Marking %s optional", key); is_optional = TRUE; } /* start a monitor for an already active resource */ possible_matches = find_actions_exact(rsc->actions, key, node); if (possible_matches == NULL) { is_optional = FALSE; pe_rsc_trace(rsc, "Marking %s mandatory: not active", key); } else { GList *gIter = NULL; for (gIter = possible_matches; gIter != NULL; gIter = gIter->next) { pe_action_t *op = (pe_action_t *) gIter->data; if (pcmk_is_set(op->flags, pe_action_reschedule)) { is_optional = FALSE; break; } } g_list_free(possible_matches); } if (((rsc->next_role == RSC_ROLE_PROMOTED) && (role == NULL)) || (role != NULL && text2role(role) != rsc->next_role)) { int log_level = LOG_TRACE; const char *result = "Ignoring"; if (is_optional) { char *after_key = NULL; pe_action_t *cancel_op = NULL; // It's running, so cancel it log_level = LOG_INFO; result = "Cancelling"; cancel_op = pe_cancel_op(rsc, name, interval_ms, node, data_set); switch (rsc->role) { case RSC_ROLE_UNPROMOTED: case RSC_ROLE_STARTED: if (rsc->next_role == RSC_ROLE_PROMOTED) { after_key = promote_key(rsc); } else if (rsc->next_role == RSC_ROLE_STOPPED) { after_key = stop_key(rsc); } break; case RSC_ROLE_PROMOTED: after_key = demote_key(rsc); break; default: break; } if (after_key) { custom_action_order(rsc, NULL, cancel_op, rsc, after_key, NULL, pe_order_runnable_left, data_set); } } do_crm_log(log_level, "%s action %s (%s vs. %s)", result, key, role ? role : role2text(RSC_ROLE_UNPROMOTED), role2text(rsc->next_role)); free(key); return; } mon = custom_action(rsc, key, name, node, is_optional, TRUE, data_set); key = mon->uuid; if (is_optional) { pe_rsc_trace(rsc, "%s\t %s (optional)", node_uname, mon->uuid); } if ((start == NULL) || !pcmk_is_set(start->flags, pe_action_runnable)) { pe_rsc_debug(rsc, "%s\t %s (cancelled : start un-runnable)", node_uname, mon->uuid); pe__clear_action_flags(mon, pe_action_runnable); } else if (node == NULL || node->details->online == FALSE || node->details->unclean) { pe_rsc_debug(rsc, "%s\t %s (cancelled : no node available)", node_uname, mon->uuid); pe__clear_action_flags(mon, pe_action_runnable); } else if (!pcmk_is_set(mon->flags, pe_action_optional)) { pe_rsc_info(rsc, " Start recurring %s (%us) for %s on %s", mon->task, interval_ms / 1000, rsc->id, node_uname); } if (rsc->next_role == RSC_ROLE_PROMOTED) { char *running_promoted = pcmk__itoa(PCMK_OCF_RUNNING_PROMOTED); add_hash_param(mon->meta, XML_ATTR_TE_TARGET_RC, running_promoted); free(running_promoted); } if ((node == NULL) || pcmk_is_set(rsc->flags, pe_rsc_managed)) { custom_action_order(rsc, start_key(rsc), NULL, NULL, strdup(key), mon, pe_order_implies_then | pe_order_runnable_left, data_set); custom_action_order(rsc, reload_key(rsc), NULL, NULL, strdup(key), mon, pe_order_implies_then | pe_order_runnable_left, data_set); if (rsc->next_role == RSC_ROLE_PROMOTED) { custom_action_order(rsc, promote_key(rsc), NULL, rsc, NULL, mon, pe_order_optional | pe_order_runnable_left, data_set); } else if (rsc->role == RSC_ROLE_PROMOTED) { custom_action_order(rsc, demote_key(rsc), NULL, rsc, NULL, mon, pe_order_optional | pe_order_runnable_left, data_set); } } } static void Recurring(pe_resource_t * rsc, pe_action_t * start, pe_node_t * node, pe_working_set_t * data_set) { if (!pcmk_is_set(rsc->flags, pe_rsc_maintenance) && (node == NULL || node->details->maintenance == FALSE)) { xmlNode *operation = NULL; for (operation = pcmk__xe_first_child(rsc->ops_xml); operation != NULL; operation = pcmk__xe_next(operation)) { if (pcmk__str_eq((const char *)operation->name, "op", pcmk__str_none)) { RecurringOp(rsc, start, node, operation, data_set); } } } } static void RecurringOp_Stopped(pe_resource_t * rsc, pe_action_t * start, pe_node_t * node, xmlNode * operation, pe_working_set_t * data_set) { char *key = NULL; const char *name = NULL; const char *role = NULL; const char *interval_spec = NULL; const char *node_uname = node? node->details->uname : "n/a"; guint interval_ms = 0; GList *possible_matches = NULL; GList *gIter = NULL; /* Only process for the operations with role="Stopped" */ role = crm_element_value(operation, "role"); if (role == NULL || text2role(role) != RSC_ROLE_STOPPED) { return; } interval_spec = crm_element_value(operation, XML_LRM_ATTR_INTERVAL); interval_ms = crm_parse_interval_spec(interval_spec); if (interval_ms == 0) { return; } name = crm_element_value(operation, "name"); if (is_op_dup(rsc, name, interval_ms)) { crm_trace("Not creating duplicate recurring action %s for %dms %s", ID(operation), interval_ms, name); return; } if (op_cannot_recur(name)) { pcmk__config_err("Ignoring %s because action '%s' cannot be recurring", ID(operation), name); return; } key = pcmk__op_key(rsc->id, name, interval_ms); if (find_rsc_op_entry(rsc, key) == NULL) { crm_trace("Not creating recurring action %s for disabled resource %s", ID(operation), rsc->id); free(key); return; } // @TODO add support if (!pcmk_is_set(rsc->flags, pe_rsc_unique)) { crm_notice("Ignoring %s (recurring monitors for Stopped role are " "not supported for anonymous clones)", ID(operation)); return; } pe_rsc_trace(rsc, "Creating recurring action %s for %s in role %s on nodes where it should not be running", ID(operation), rsc->id, role2text(rsc->next_role)); /* if the monitor exists on the node where the resource will be running, cancel it */ if (node != NULL) { possible_matches = find_actions_exact(rsc->actions, key, node); if (possible_matches) { pe_action_t *cancel_op = NULL; g_list_free(possible_matches); cancel_op = pe_cancel_op(rsc, name, interval_ms, node, data_set); if ((rsc->next_role == RSC_ROLE_STARTED) || (rsc->next_role == RSC_ROLE_UNPROMOTED)) { /* rsc->role == RSC_ROLE_STOPPED: cancel the monitor before start */ /* rsc->role == RSC_ROLE_STARTED: for a migration, cancel the monitor on the target node before start */ custom_action_order(rsc, NULL, cancel_op, rsc, start_key(rsc), NULL, pe_order_runnable_left, data_set); } pe_rsc_info(rsc, "Cancel action %s (%s vs. %s) on %s", key, role, role2text(rsc->next_role), node_uname); } } for (gIter = data_set->nodes; gIter != NULL; gIter = gIter->next) { pe_node_t *stop_node = (pe_node_t *) gIter->data; const char *stop_node_uname = stop_node->details->uname; gboolean is_optional = TRUE; gboolean probe_is_optional = TRUE; gboolean stop_is_optional = TRUE; pe_action_t *stopped_mon = NULL; char *rc_inactive = NULL; GList *probe_complete_ops = NULL; GList *stop_ops = NULL; GList *local_gIter = NULL; if (node && pcmk__str_eq(stop_node_uname, node_uname, pcmk__str_casei)) { continue; } pe_rsc_trace(rsc, "Creating recurring action %s for %s on %s", ID(operation), rsc->id, crm_str(stop_node_uname)); /* start a monitor for an already stopped resource */ possible_matches = find_actions_exact(rsc->actions, key, stop_node); if (possible_matches == NULL) { pe_rsc_trace(rsc, "Marking %s mandatory on %s: not active", key, crm_str(stop_node_uname)); is_optional = FALSE; } else { pe_rsc_trace(rsc, "Marking %s optional on %s: already active", key, crm_str(stop_node_uname)); is_optional = TRUE; g_list_free(possible_matches); } stopped_mon = custom_action(rsc, strdup(key), name, stop_node, is_optional, TRUE, data_set); rc_inactive = pcmk__itoa(PCMK_OCF_NOT_RUNNING); add_hash_param(stopped_mon->meta, XML_ATTR_TE_TARGET_RC, rc_inactive); free(rc_inactive); if (pcmk_is_set(rsc->flags, pe_rsc_managed)) { GList *probes = pe__resource_actions(rsc, stop_node, RSC_STATUS, FALSE); GList *pIter = NULL; for (pIter = probes; pIter != NULL; pIter = pIter->next) { pe_action_t *probe = (pe_action_t *) pIter->data; order_actions(probe, stopped_mon, pe_order_runnable_left); crm_trace("%s then %s on %s", probe->uuid, stopped_mon->uuid, stop_node->details->uname); } g_list_free(probes); } if (probe_complete_ops) { g_list_free(probe_complete_ops); } stop_ops = pe__resource_actions(rsc, stop_node, RSC_STOP, TRUE); for (local_gIter = stop_ops; local_gIter != NULL; local_gIter = local_gIter->next) { pe_action_t *stop = (pe_action_t *) local_gIter->data; if (!pcmk_is_set(stop->flags, pe_action_optional)) { stop_is_optional = FALSE; } if (!pcmk_is_set(stop->flags, pe_action_runnable)) { crm_debug("%s\t %s (cancelled : stop un-runnable)", crm_str(stop_node_uname), stopped_mon->uuid); pe__clear_action_flags(stopped_mon, pe_action_runnable); } if (pcmk_is_set(rsc->flags, pe_rsc_managed)) { custom_action_order(rsc, stop_key(rsc), stop, NULL, strdup(key), stopped_mon, pe_order_implies_then | pe_order_runnable_left, data_set); } } if (stop_ops) { g_list_free(stop_ops); } if (is_optional == FALSE && probe_is_optional && stop_is_optional && !pcmk_is_set(rsc->flags, pe_rsc_managed)) { pe_rsc_trace(rsc, "Marking %s optional on %s due to unmanaged", key, crm_str(stop_node_uname)); pe__set_action_flags(stopped_mon, pe_action_optional); } if (pcmk_is_set(stopped_mon->flags, pe_action_optional)) { pe_rsc_trace(rsc, "%s\t %s (optional)", crm_str(stop_node_uname), stopped_mon->uuid); } if (stop_node->details->online == FALSE || stop_node->details->unclean) { pe_rsc_debug(rsc, "%s\t %s (cancelled : no node available)", crm_str(stop_node_uname), stopped_mon->uuid); pe__clear_action_flags(stopped_mon, pe_action_runnable); } if (pcmk_is_set(stopped_mon->flags, pe_action_runnable) && !pcmk_is_set(stopped_mon->flags, pe_action_optional)) { crm_notice(" Start recurring %s (%us) for %s on %s", stopped_mon->task, interval_ms / 1000, rsc->id, crm_str(stop_node_uname)); } } free(key); } static void Recurring_Stopped(pe_resource_t * rsc, pe_action_t * start, pe_node_t * node, pe_working_set_t * data_set) { if (!pcmk_is_set(rsc->flags, pe_rsc_maintenance) && (node == NULL || node->details->maintenance == FALSE)) { xmlNode *operation = NULL; for (operation = pcmk__xe_first_child(rsc->ops_xml); operation != NULL; operation = pcmk__xe_next(operation)) { if (pcmk__str_eq((const char *)operation->name, "op", pcmk__str_none)) { RecurringOp_Stopped(rsc, start, node, operation, data_set); } } } } static void handle_migration_actions(pe_resource_t * rsc, pe_node_t *current, pe_node_t *chosen, pe_working_set_t * data_set) { pe_action_t *migrate_to = NULL; pe_action_t *migrate_from = NULL; pe_action_t *start = NULL; pe_action_t *stop = NULL; gboolean partial = rsc->partial_migration_target ? TRUE : FALSE; pe_rsc_trace(rsc, "Processing migration actions %s moving from %s to %s . partial migration = %s", rsc->id, current->details->id, chosen->details->id, partial ? "TRUE" : "FALSE"); start = start_action(rsc, chosen, TRUE); stop = stop_action(rsc, current, TRUE); if (partial == FALSE) { migrate_to = custom_action(rsc, pcmk__op_key(rsc->id, RSC_MIGRATE, 0), RSC_MIGRATE, current, TRUE, TRUE, data_set); } migrate_from = custom_action(rsc, pcmk__op_key(rsc->id, RSC_MIGRATED, 0), RSC_MIGRATED, chosen, TRUE, TRUE, data_set); if ((migrate_to && migrate_from) || (migrate_from && partial)) { pe__set_action_flags(start, pe_action_migrate_runnable); pe__set_action_flags(stop, pe_action_migrate_runnable); // This is easier than trying to delete it from the graph pe__set_action_flags(start, pe_action_pseudo); /* order probes before migrations */ if (partial) { pe__set_action_flags(migrate_from, pe_action_migrate_runnable); migrate_from->needs = start->needs; custom_action_order(rsc, pcmk__op_key(rsc->id, RSC_STATUS, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_MIGRATED, 0), NULL, pe_order_optional, data_set); } else { pe__set_action_flags(migrate_from, pe_action_migrate_runnable); pe__set_action_flags(migrate_to, pe_action_migrate_runnable); migrate_to->needs = start->needs; custom_action_order(rsc, pcmk__op_key(rsc->id, RSC_STATUS, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_MIGRATE, 0), NULL, pe_order_optional, data_set); custom_action_order(rsc, pcmk__op_key(rsc->id, RSC_MIGRATE, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_MIGRATED, 0), NULL, pe_order_optional|pe_order_implies_first_migratable, data_set); } custom_action_order(rsc, pcmk__op_key(rsc->id, RSC_MIGRATED, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_STOP, 0), NULL, pe_order_optional|pe_order_implies_first_migratable, data_set); custom_action_order(rsc, pcmk__op_key(rsc->id, RSC_MIGRATED, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_START, 0), NULL, pe_order_optional|pe_order_implies_first_migratable|pe_order_pseudo_left, data_set); } if (migrate_to) { add_hash_param(migrate_to->meta, XML_LRM_ATTR_MIGRATE_SOURCE, current->details->uname); add_hash_param(migrate_to->meta, XML_LRM_ATTR_MIGRATE_TARGET, chosen->details->uname); /* Pacemaker Remote connections don't require pending to be recorded in * the CIB. We can reduce CIB writes by not setting PENDING for them. */ if (rsc->is_remote_node == FALSE) { /* migrate_to takes place on the source node, but can * have an effect on the target node depending on how * the agent is written. Because of this, we have to maintain * a record that the migrate_to occurred, in case the source node * loses membership while the migrate_to action is still in-flight. */ add_hash_param(migrate_to->meta, XML_OP_ATTR_PENDING, "true"); } } if (migrate_from) { add_hash_param(migrate_from->meta, XML_LRM_ATTR_MIGRATE_SOURCE, current->details->uname); add_hash_param(migrate_from->meta, XML_LRM_ATTR_MIGRATE_TARGET, chosen->details->uname); } } void native_create_actions(pe_resource_t * rsc, pe_working_set_t * data_set) { pe_action_t *start = NULL; pe_node_t *chosen = NULL; pe_node_t *current = NULL; gboolean need_stop = FALSE; bool need_promote = FALSE; gboolean is_moving = FALSE; gboolean allow_migrate = pcmk_is_set(rsc->flags, pe_rsc_allow_migrate)? TRUE : FALSE; GList *gIter = NULL; unsigned int num_all_active = 0; unsigned int num_clean_active = 0; bool multiply_active = FALSE; enum rsc_role_e role = RSC_ROLE_UNKNOWN; enum rsc_role_e next_role = RSC_ROLE_UNKNOWN; CRM_ASSERT(rsc); chosen = rsc->allocated_to; next_role = rsc->next_role; if (next_role == RSC_ROLE_UNKNOWN) { pe__set_next_role(rsc, (chosen == NULL)? RSC_ROLE_STOPPED : RSC_ROLE_STARTED, "allocation"); } pe_rsc_trace(rsc, "Creating all actions for %s transition from %s to %s (%s) on %s", rsc->id, role2text(rsc->role), role2text(rsc->next_role), ((next_role == RSC_ROLE_UNKNOWN)? "implicit" : "explicit"), ((chosen == NULL)? "no node" : chosen->details->uname)); current = pe__find_active_on(rsc, &num_all_active, &num_clean_active); for (gIter = rsc->dangling_migrations; gIter != NULL; gIter = gIter->next) { pe_node_t *dangling_source = (pe_node_t *) gIter->data; pe_action_t *stop = NULL; pe_rsc_trace(rsc, "Creating stop action %sfor %s on %s due to dangling migration", pcmk_is_set(data_set->flags, pe_flag_remove_after_stop)? "and cleanup " : "", rsc->id, dangling_source->details->uname); stop = stop_action(rsc, dangling_source, FALSE); pe__set_action_flags(stop, pe_action_dangle); if (pcmk_is_set(data_set->flags, pe_flag_remove_after_stop)) { DeleteRsc(rsc, dangling_source, FALSE, data_set); } } if ((num_all_active == 2) && (num_clean_active == 2) && chosen && rsc->partial_migration_source && rsc->partial_migration_target && (current->details == rsc->partial_migration_source->details) && (chosen->details == rsc->partial_migration_target->details)) { /* The chosen node is still the migration target from a partial * migration. Attempt to continue the migration instead of recovering * by stopping the resource everywhere and starting it on a single node. */ pe_rsc_trace(rsc, "Will attempt to continue with partial migration " "to target %s from %s", rsc->partial_migration_target->details->id, rsc->partial_migration_source->details->id); } else if (!pcmk_is_set(rsc->flags, pe_rsc_needs_fencing)) { /* If a resource has "requires" set to nothing or quorum, don't consider * it active on unclean nodes (similar to how all resources behave when * stonith-enabled is false). We can start such resources elsewhere * before fencing completes, and if we considered the resource active on * the failed node, we would attempt recovery for being active on * multiple nodes. */ multiply_active = (num_clean_active > 1); } else { multiply_active = (num_all_active > 1); } if (multiply_active) { if (rsc->partial_migration_target && rsc->partial_migration_source) { // Migration was in progress, but we've chosen a different target crm_notice("Resource %s can no longer migrate from %s to %s " "(will stop on both nodes)", rsc->id, rsc->partial_migration_source->details->uname, rsc->partial_migration_target->details->uname); } else { const char *class = crm_element_value(rsc->xml, XML_AGENT_ATTR_CLASS); // Resource was (possibly) incorrectly multiply active pe_proc_err("%s resource %s might be active on %u nodes (%s)", crm_str(class), rsc->id, num_all_active, recovery2text(rsc->recovery_type)); crm_notice("See https://wiki.clusterlabs.org/wiki/FAQ#Resource_is_Too_Active for more information"); } if (rsc->recovery_type == recovery_stop_start) { need_stop = TRUE; } /* If by chance a partial migration is in process, but the migration * target is not chosen still, clear all partial migration data. */ rsc->partial_migration_source = rsc->partial_migration_target = NULL; allow_migrate = FALSE; } if (pcmk_is_set(rsc->flags, pe_rsc_start_pending)) { pe_rsc_trace(rsc, "Creating start action for %s to represent already pending start", rsc->id); start = start_action(rsc, chosen, TRUE); pe__set_action_flags(start, pe_action_print_always); } if (current && chosen && current->details != chosen->details) { pe_rsc_trace(rsc, "Moving %s from %s to %s", rsc->id, crm_str(current->details->uname), crm_str(chosen->details->uname)); is_moving = TRUE; need_stop = TRUE; } else if (pcmk_is_set(rsc->flags, pe_rsc_failed)) { if (pcmk_is_set(rsc->flags, pe_rsc_stop)) { need_stop = TRUE; pe_rsc_trace(rsc, "Recovering %s", rsc->id); } else { pe_rsc_trace(rsc, "Recovering %s by demotion", rsc->id); if (rsc->next_role == RSC_ROLE_PROMOTED) { need_promote = TRUE; } } } else if (pcmk_is_set(rsc->flags, pe_rsc_block)) { pe_rsc_trace(rsc, "Blocking further actions on %s", rsc->id); need_stop = TRUE; } else if (rsc->role > RSC_ROLE_STARTED && current != NULL && chosen != NULL) { pe_rsc_trace(rsc, "Creating start action for promoted resource %s", rsc->id); start = start_action(rsc, chosen, TRUE); if (!pcmk_is_set(start->flags, pe_action_optional)) { // Recovery of a promoted resource pe_rsc_trace(rsc, "%s restart is required for recovery", rsc->id); need_stop = TRUE; } } /* Create any additional actions required when bringing resource down and * back up to same level. */ role = rsc->role; while (role != RSC_ROLE_STOPPED) { next_role = rsc_state_matrix[role][RSC_ROLE_STOPPED]; pe_rsc_trace(rsc, "Creating %s action to take %s down from %s to %s", (need_stop? "required" : "optional"), rsc->id, role2text(role), role2text(next_role)); if (rsc_action_matrix[role][next_role] (rsc, current, !need_stop, data_set) == FALSE) { break; } role = next_role; } while ((rsc->role <= rsc->next_role) && (role != rsc->role) && !pcmk_is_set(rsc->flags, pe_rsc_block)) { bool required = need_stop; next_role = rsc_state_matrix[role][rsc->role]; if ((next_role == RSC_ROLE_PROMOTED) && need_promote) { required = true; } pe_rsc_trace(rsc, "Creating %s action to take %s up from %s to %s", (required? "required" : "optional"), rsc->id, role2text(role), role2text(next_role)); if (rsc_action_matrix[role][next_role](rsc, chosen, !required, data_set) == FALSE) { break; } role = next_role; } role = rsc->role; /* Required steps from this role to the next */ while (role != rsc->next_role) { next_role = rsc_state_matrix[role][rsc->next_role]; pe_rsc_trace(rsc, "Creating action to take %s from %s to %s (ending at %s)", rsc->id, role2text(role), role2text(next_role), role2text(rsc->next_role)); if (rsc_action_matrix[role][next_role] (rsc, chosen, FALSE, data_set) == FALSE) { break; } role = next_role; } if (pcmk_is_set(rsc->flags, pe_rsc_block)) { pe_rsc_trace(rsc, "Not creating recurring monitors for blocked resource %s", rsc->id); } else if ((rsc->next_role != RSC_ROLE_STOPPED) || !pcmk_is_set(rsc->flags, pe_rsc_managed)) { pe_rsc_trace(rsc, "Creating recurring monitors for %s resource %s", ((rsc->next_role == RSC_ROLE_STOPPED)? "unmanaged" : "active"), rsc->id); start = start_action(rsc, chosen, TRUE); Recurring(rsc, start, chosen, data_set); Recurring_Stopped(rsc, start, chosen, data_set); } else { pe_rsc_trace(rsc, "Creating recurring monitors for inactive resource %s", rsc->id); Recurring_Stopped(rsc, NULL, NULL, data_set); } /* if we are stuck in a partial migration, where the target * of the partial migration no longer matches the chosen target. * A full stop/start is required */ if (rsc->partial_migration_target && (chosen == NULL || rsc->partial_migration_target->details != chosen->details)) { pe_rsc_trace(rsc, "Not allowing partial migration of %s to continue", rsc->id); allow_migrate = FALSE; } else if (!is_moving || !pcmk_is_set(rsc->flags, pe_rsc_managed) || pcmk_any_flags_set(rsc->flags, pe_rsc_failed|pe_rsc_start_pending) || (current && current->details->unclean) || rsc->next_role < RSC_ROLE_STARTED) { allow_migrate = FALSE; } if (allow_migrate) { handle_migration_actions(rsc, current, chosen, data_set); } } static void rsc_avoids_remote_nodes(pe_resource_t *rsc) { GHashTableIter iter; pe_node_t *node = NULL; g_hash_table_iter_init(&iter, rsc->allowed_nodes); while (g_hash_table_iter_next(&iter, NULL, (void **)&node)) { if (node->details->remote_rsc) { node->weight = -INFINITY; } } } /*! * \internal * \brief Return allowed nodes as (possibly sorted) list * * Convert a resource's hash table of allowed nodes to a list. If printing to * stdout, sort the list, to keep action ID numbers consistent for regression * test output (while avoiding the performance hit on a live cluster). * * \param[in] rsc Resource to check for allowed nodes * \param[in] data_set Cluster working set * * \return List of resource's allowed nodes * \note Callers should take care not to rely on the list being sorted. */ static GList * allowed_nodes_as_list(pe_resource_t *rsc, pe_working_set_t *data_set) { GList *allowed_nodes = NULL; if (rsc->allowed_nodes) { allowed_nodes = g_hash_table_get_values(rsc->allowed_nodes); } if (!pcmk__is_daemon) { allowed_nodes = g_list_sort(allowed_nodes, sort_node_uname); } return allowed_nodes; } void native_internal_constraints(pe_resource_t * rsc, pe_working_set_t * data_set) { /* This function is on the critical path and worth optimizing as much as possible */ pe_resource_t *top = NULL; GList *allowed_nodes = NULL; bool check_unfencing = FALSE; bool check_utilization = FALSE; if (!pcmk_is_set(rsc->flags, pe_rsc_managed)) { pe_rsc_trace(rsc, "Skipping native constraints for unmanaged resource: %s", rsc->id); return; } top = uber_parent(rsc); // Whether resource requires unfencing check_unfencing = !pcmk_is_set(rsc->flags, pe_rsc_fence_device) && pcmk_is_set(data_set->flags, pe_flag_enable_unfencing) && pcmk_is_set(rsc->flags, pe_rsc_needs_unfencing); // Whether a non-default placement strategy is used check_utilization = (g_hash_table_size(rsc->utilization) > 0) && !pcmk__str_eq(data_set->placement_strategy, "default", pcmk__str_casei); // Order stops before starts (i.e. restart) custom_action_order(rsc, pcmk__op_key(rsc->id, RSC_STOP, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_START, 0), NULL, pe_order_optional|pe_order_implies_then|pe_order_restart, data_set); // Promotable ordering: demote before stop, start before promote if (pcmk_is_set(top->flags, pe_rsc_promotable) || (rsc->role > RSC_ROLE_UNPROMOTED)) { custom_action_order(rsc, pcmk__op_key(rsc->id, RSC_DEMOTE, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_STOP, 0), NULL, pe_order_promoted_implies_first, data_set); custom_action_order(rsc, pcmk__op_key(rsc->id, RSC_START, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_PROMOTE, 0), NULL, pe_order_runnable_left, data_set); } // Don't clear resource history if probing on same node custom_action_order(rsc, pcmk__op_key(rsc->id, CRM_OP_LRM_DELETE, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_STATUS, 0), NULL, pe_order_same_node|pe_order_then_cancels_first, data_set); // Certain checks need allowed nodes if (check_unfencing || check_utilization || rsc->container) { allowed_nodes = allowed_nodes_as_list(rsc, data_set); } if (check_unfencing) { /* Check if the node needs to be unfenced first */ for (GList *item = allowed_nodes; item; item = item->next) { pe_node_t *node = item->data; pe_action_t *unfence = pe_fence_op(node, "on", TRUE, NULL, FALSE, data_set); crm_debug("Ordering any stops of %s before %s, and any starts after", rsc->id, unfence->uuid); /* * It would be more efficient to order clone resources once, * rather than order each instance, but ordering the instance * allows us to avoid unnecessary dependencies that might conflict * with user constraints. * * @TODO: This constraint can still produce a transition loop if the * resource has a stop scheduled on the node being unfenced, and * there is a user ordering constraint to start some other resource * (which will be ordered after the unfence) before stopping this * resource. An example is "start some slow-starting cloned service * before stopping an associated virtual IP that may be moving to * it": * stop this -> unfencing -> start that -> stop this */ custom_action_order(rsc, stop_key(rsc), NULL, NULL, strdup(unfence->uuid), unfence, pe_order_optional|pe_order_same_node, data_set); custom_action_order(NULL, strdup(unfence->uuid), unfence, rsc, start_key(rsc), NULL, pe_order_implies_then_on_node|pe_order_same_node, data_set); } } if (check_utilization) { GList *gIter = NULL; pe_rsc_trace(rsc, "Creating utilization constraints for %s - strategy: %s", rsc->id, data_set->placement_strategy); for (gIter = rsc->running_on; gIter != NULL; gIter = gIter->next) { pe_node_t *current = (pe_node_t *) gIter->data; char *load_stopped_task = crm_strdup_printf(LOAD_STOPPED "_%s", current->details->uname); pe_action_t *load_stopped = get_pseudo_op(load_stopped_task, data_set); if (load_stopped->node == NULL) { load_stopped->node = pe__copy_node(current); pe__clear_action_flags(load_stopped, pe_action_optional); } custom_action_order(rsc, stop_key(rsc), NULL, NULL, load_stopped_task, load_stopped, pe_order_load, data_set); } for (GList *item = allowed_nodes; item; item = item->next) { pe_node_t *next = item->data; char *load_stopped_task = crm_strdup_printf(LOAD_STOPPED "_%s", next->details->uname); pe_action_t *load_stopped = get_pseudo_op(load_stopped_task, data_set); if (load_stopped->node == NULL) { load_stopped->node = pe__copy_node(next); pe__clear_action_flags(load_stopped, pe_action_optional); } custom_action_order(NULL, strdup(load_stopped_task), load_stopped, rsc, start_key(rsc), NULL, pe_order_load, data_set); custom_action_order(NULL, strdup(load_stopped_task), load_stopped, rsc, pcmk__op_key(rsc->id, RSC_MIGRATE, 0), NULL, pe_order_load, data_set); free(load_stopped_task); } } if (rsc->container) { pe_resource_t *remote_rsc = NULL; if (rsc->is_remote_node) { // rsc is the implicit remote connection for a guest or bundle node /* Do not allow a guest resource to live on a Pacemaker Remote node, * to avoid nesting remotes. However, allow bundles to run on remote * nodes. */ if (!pcmk_is_set(rsc->flags, pe_rsc_allow_remote_remotes)) { rsc_avoids_remote_nodes(rsc->container); } /* If someone cleans up a guest or bundle node's container, we will * likely schedule a (re-)probe of the container and recovery of the * connection. Order the connection stop after the container probe, * so that if we detect the container running, we will trigger a new * transition and avoid the unnecessary recovery. */ new_rsc_order(rsc->container, RSC_STATUS, rsc, RSC_STOP, pe_order_optional, data_set); /* A user can specify that a resource must start on a Pacemaker Remote * node by explicitly configuring it with the container=NODENAME * meta-attribute. This is of questionable merit, since location * constraints can accomplish the same thing. But we support it, so here * we check whether a resource (that is not itself a remote connection) * has container set to a remote node or guest node resource. */ } else if (rsc->container->is_remote_node) { remote_rsc = rsc->container; } else { remote_rsc = pe__resource_contains_guest_node(data_set, rsc->container); } if (remote_rsc) { /* Force the resource on the Pacemaker Remote node instead of * colocating the resource with the container resource. */ for (GList *item = allowed_nodes; item; item = item->next) { pe_node_t *node = item->data; if (node->details->remote_rsc != remote_rsc) { node->weight = -INFINITY; } } } else { /* This resource is either a filler for a container that does NOT * represent a Pacemaker Remote node, or a Pacemaker Remote * connection resource for a guest node or bundle. */ int score; crm_trace("Order and colocate %s relative to its container %s", rsc->id, rsc->container->id); custom_action_order(rsc->container, pcmk__op_key(rsc->container->id, RSC_START, 0), NULL, rsc, pcmk__op_key(rsc->id, RSC_START, 0), NULL, pe_order_implies_then|pe_order_runnable_left, data_set); custom_action_order(rsc, pcmk__op_key(rsc->id, RSC_STOP, 0), NULL, rsc->container, pcmk__op_key(rsc->container->id, RSC_STOP, 0), NULL, pe_order_implies_first, data_set); if (pcmk_is_set(rsc->flags, pe_rsc_allow_remote_remotes)) { score = 10000; /* Highly preferred but not essential */ } else { score = INFINITY; /* Force them to run on the same host */ } pcmk__new_colocation("resource-with-container", NULL, score, rsc, rsc->container, NULL, NULL, true, data_set); } } if (rsc->is_remote_node || pcmk_is_set(rsc->flags, pe_rsc_fence_device)) { /* don't allow remote nodes to run stonith devices * or remote connection resources.*/ rsc_avoids_remote_nodes(rsc); } g_list_free(allowed_nodes); } void native_rsc_colocation_lh(pe_resource_t *rsc_lh, pe_resource_t *rsc_rh, pcmk__colocation_t *constraint, pe_working_set_t *data_set) { if (rsc_lh == NULL) { pe_err("rsc_lh was NULL for %s", constraint->id); return; } else if (constraint->rsc_rh == NULL) { pe_err("rsc_rh was NULL for %s", constraint->id); return; } pe_rsc_trace(rsc_lh, "Processing colocation constraint between %s and %s", rsc_lh->id, rsc_rh->id); rsc_rh->cmds->rsc_colocation_rh(rsc_lh, rsc_rh, constraint, data_set); } enum filter_colocation_res filter_colocation_constraint(pe_resource_t * rsc_lh, pe_resource_t * rsc_rh, pcmk__colocation_t *constraint, gboolean preview) { /* rh side must be allocated before we can process constraint */ if (!preview && pcmk_is_set(rsc_rh->flags, pe_rsc_provisional)) { return influence_nothing; } if ((constraint->role_lh >= RSC_ROLE_UNPROMOTED) && rsc_lh->parent && pcmk_is_set(rsc_lh->parent->flags, pe_rsc_promotable) && !pcmk_is_set(rsc_lh->flags, pe_rsc_provisional)) { /* LH and RH resources have already been allocated, place the correct * priority on LH rsc for the given promotable clone resource role */ return influence_rsc_priority; } if (!preview && !pcmk_is_set(rsc_lh->flags, pe_rsc_provisional)) { // Log an error if we violated a mandatory colocation constraint const pe_node_t *rh_node = rsc_rh->allocated_to; if (rsc_lh->allocated_to == NULL) { // Dependent resource isn't allocated, so constraint doesn't matter return influence_nothing; } if (constraint->score >= INFINITY) { // Dependent resource must colocate with rh_node if ((rh_node == NULL) || (rh_node->details != rsc_lh->allocated_to->details)) { crm_err("%s must be colocated with %s but is not (%s vs. %s)", rsc_lh->id, rsc_rh->id, rsc_lh->allocated_to->details->uname, (rh_node? rh_node->details->uname : "unallocated")); } } else if (constraint->score <= -INFINITY) { // Dependent resource must anti-colocate with rh_node if ((rh_node != NULL) && (rsc_lh->allocated_to->details == rh_node->details)) { crm_err("%s and %s must be anti-colocated but are allocated " "to the same node (%s)", rsc_lh->id, rsc_rh->id, rh_node->details->uname); } } return influence_nothing; } if (constraint->score > 0 && constraint->role_lh != RSC_ROLE_UNKNOWN && constraint->role_lh != rsc_lh->next_role) { crm_trace("LH: Skipping constraint: \"%s\" state filter nextrole is %s", role2text(constraint->role_lh), role2text(rsc_lh->next_role)); return influence_nothing; } if (constraint->score > 0 && constraint->role_rh != RSC_ROLE_UNKNOWN && constraint->role_rh != rsc_rh->next_role) { crm_trace("RH: Skipping constraint: \"%s\" state filter", role2text(constraint->role_rh)); return influence_nothing; } if (constraint->score < 0 && constraint->role_lh != RSC_ROLE_UNKNOWN && constraint->role_lh == rsc_lh->next_role) { crm_trace("LH: Skipping negative constraint: \"%s\" state filter", role2text(constraint->role_lh)); return influence_nothing; } if (constraint->score < 0 && constraint->role_rh != RSC_ROLE_UNKNOWN && constraint->role_rh == rsc_rh->next_role) { crm_trace("RH: Skipping negative constraint: \"%s\" state filter", role2text(constraint->role_rh)); return influence_nothing; } return influence_rsc_location; } static void influence_priority(pe_resource_t *rsc_lh, pe_resource_t *rsc_rh, pcmk__colocation_t *constraint) { const char *rh_value = NULL; const char *lh_value = NULL; const char *attribute = CRM_ATTR_ID; int score_multiplier = 1; if (!rsc_rh->allocated_to || !rsc_lh->allocated_to) { return; } if (constraint->node_attribute != NULL) { attribute = constraint->node_attribute; } lh_value = pe_node_attribute_raw(rsc_lh->allocated_to, attribute); rh_value = pe_node_attribute_raw(rsc_rh->allocated_to, attribute); if (!pcmk__str_eq(lh_value, rh_value, pcmk__str_casei)) { if ((constraint->score == INFINITY) && (constraint->role_lh == RSC_ROLE_PROMOTED)) { rsc_lh->priority = -INFINITY; } return; } if (constraint->role_rh && (constraint->role_rh != rsc_rh->next_role)) { return; } if (constraint->role_lh == RSC_ROLE_UNPROMOTED) { score_multiplier = -1; } rsc_lh->priority = pe__add_scores(score_multiplier * constraint->score, rsc_lh->priority); } static void colocation_match(pe_resource_t *rsc_lh, pe_resource_t *rsc_rh, pcmk__colocation_t *constraint) { const char *attribute = CRM_ATTR_ID; const char *value = NULL; GHashTable *work = NULL; GHashTableIter iter; pe_node_t *node = NULL; if (constraint->node_attribute != NULL) { attribute = constraint->node_attribute; } if (rsc_rh->allocated_to) { value = pe_node_attribute_raw(rsc_rh->allocated_to, attribute); } else if (constraint->score < 0) { // Nothing to do (anti-colocation with something that is not running) return; } work = pcmk__copy_node_table(rsc_lh->allowed_nodes); g_hash_table_iter_init(&iter, work); while (g_hash_table_iter_next(&iter, NULL, (void **)&node)) { if (rsc_rh->allocated_to == NULL) { pe_rsc_trace(rsc_lh, "%s: %s@%s -= %d (%s inactive)", constraint->id, rsc_lh->id, node->details->uname, constraint->score, rsc_rh->id); node->weight = pe__add_scores(-constraint->score, node->weight); } else if (pcmk__str_eq(pe_node_attribute_raw(node, attribute), value, pcmk__str_casei)) { if (constraint->score < CRM_SCORE_INFINITY) { pe_rsc_trace(rsc_lh, "%s: %s@%s += %d", constraint->id, rsc_lh->id, node->details->uname, constraint->score); node->weight = pe__add_scores(constraint->score, node->weight); } } else if (constraint->score >= CRM_SCORE_INFINITY) { pe_rsc_trace(rsc_lh, "%s: %s@%s -= %d (%s mismatch)", constraint->id, rsc_lh->id, node->details->uname, constraint->score, attribute); node->weight = pe__add_scores(-constraint->score, node->weight); } } if (can_run_any(work) || constraint->score <= -INFINITY || constraint->score >= INFINITY) { g_hash_table_destroy(rsc_lh->allowed_nodes); rsc_lh->allowed_nodes = work; work = NULL; } else { pe_rsc_info(rsc_lh, "%s: Rolling back scores from %s (no available nodes)", rsc_lh->id, rsc_rh->id); } if (work) { g_hash_table_destroy(work); } } void native_rsc_colocation_rh(pe_resource_t *rsc_lh, pe_resource_t *rsc_rh, pcmk__colocation_t *constraint, pe_working_set_t *data_set) { enum filter_colocation_res filter_results; CRM_ASSERT(rsc_lh); CRM_ASSERT(rsc_rh); filter_results = filter_colocation_constraint(rsc_lh, rsc_rh, constraint, FALSE); pe_rsc_trace(rsc_lh, "%s %s with %s (%s, score=%d, filter=%d)", ((constraint->score > 0)? "Colocating" : "Anti-colocating"), rsc_lh->id, rsc_rh->id, constraint->id, constraint->score, filter_results); switch (filter_results) { case influence_rsc_priority: influence_priority(rsc_lh, rsc_rh, constraint); break; case influence_rsc_location: colocation_match(rsc_lh, rsc_rh, constraint); break; case influence_nothing: default: return; } } static gboolean filter_rsc_ticket(pe_resource_t * rsc_lh, rsc_ticket_t * rsc_ticket) { if (rsc_ticket->role_lh != RSC_ROLE_UNKNOWN && rsc_ticket->role_lh != rsc_lh->role) { pe_rsc_trace(rsc_lh, "LH: Skipping constraint: \"%s\" state filter", role2text(rsc_ticket->role_lh)); return FALSE; } return TRUE; } void rsc_ticket_constraint(pe_resource_t * rsc_lh, rsc_ticket_t * rsc_ticket, pe_working_set_t * data_set) { if (rsc_ticket == NULL) { pe_err("rsc_ticket was NULL"); return; } if (rsc_lh == NULL) { pe_err("rsc_lh was NULL for %s", rsc_ticket->id); return; } if (rsc_ticket->ticket->granted && rsc_ticket->ticket->standby == FALSE) { return; } if (rsc_lh->children) { GList *gIter = rsc_lh->children; pe_rsc_trace(rsc_lh, "Processing ticket dependencies from %s", rsc_lh->id); for (; gIter != NULL; gIter = gIter->next) { pe_resource_t *child_rsc = (pe_resource_t *) gIter->data; rsc_ticket_constraint(child_rsc, rsc_ticket, data_set); } return; } pe_rsc_trace(rsc_lh, "%s: Processing ticket dependency on %s (%s, %s)", rsc_lh->id, rsc_ticket->ticket->id, rsc_ticket->id, role2text(rsc_ticket->role_lh)); if ((rsc_ticket->ticket->granted == FALSE) && (rsc_lh->running_on != NULL)) { GList *gIter = NULL; switch (rsc_ticket->loss_policy) { case loss_ticket_stop: resource_location(rsc_lh, NULL, -INFINITY, "__loss_of_ticket__", data_set); break; case loss_ticket_demote: // Promotion score will be set to -INFINITY in promotion_order() if (rsc_ticket->role_lh != RSC_ROLE_PROMOTED) { resource_location(rsc_lh, NULL, -INFINITY, "__loss_of_ticket__", data_set); } break; case loss_ticket_fence: if (filter_rsc_ticket(rsc_lh, rsc_ticket) == FALSE) { return; } resource_location(rsc_lh, NULL, -INFINITY, "__loss_of_ticket__", data_set); for (gIter = rsc_lh->running_on; gIter != NULL; gIter = gIter->next) { pe_node_t *node = (pe_node_t *) gIter->data; pe_fence_node(data_set, node, "deadman ticket was lost", FALSE); } break; case loss_ticket_freeze: if (filter_rsc_ticket(rsc_lh, rsc_ticket) == FALSE) { return; } if (rsc_lh->running_on != NULL) { pe__clear_resource_flags(rsc_lh, pe_rsc_managed); pe__set_resource_flags(rsc_lh, pe_rsc_block); } break; } } else if (rsc_ticket->ticket->granted == FALSE) { if ((rsc_ticket->role_lh != RSC_ROLE_PROMOTED) || (rsc_ticket->loss_policy == loss_ticket_stop)) { resource_location(rsc_lh, NULL, -INFINITY, "__no_ticket__", data_set); } } else if (rsc_ticket->ticket->standby) { if ((rsc_ticket->role_lh != RSC_ROLE_PROMOTED) || (rsc_ticket->loss_policy == loss_ticket_stop)) { resource_location(rsc_lh, NULL, -INFINITY, "__ticket_standby__", data_set); } } } enum pe_action_flags native_action_flags(pe_action_t * action, pe_node_t * node) { return action->flags; } static inline bool is_primitive_action(pe_action_t *action) { return action && action->rsc && (action->rsc->variant == pe_native); } /*! * \internal * \brief Clear a single action flag and set reason text * * \param[in] action Action whose flag should be cleared * \param[in] flag Action flag that should be cleared * \param[in] reason Action that is the reason why flag is being cleared */ #define clear_action_flag_because(action, flag, reason) do { \ if (pcmk_is_set((action)->flags, (flag))) { \ pe__clear_action_flags(action, flag); \ if ((action)->rsc != (reason)->rsc) { \ char *reason_text = pe__action2reason((reason), (flag)); \ pe_action_set_reason((action), reason_text, \ ((flag) == pe_action_migrate_runnable)); \ free(reason_text); \ } \ } \ } while (0) /*! * \internal * \brief Set action bits appropriately when pe_restart_order is used * * \param[in] first 'First' action in an ordering with pe_restart_order * \param[in] then 'Then' action in an ordering with pe_restart_order * \param[in] filter What ordering flags to care about * * \note pe_restart_order is set for "stop resource before starting it" and * "stop later group member before stopping earlier group member" */ static void handle_restart_ordering(pe_action_t *first, pe_action_t *then, enum pe_action_flags filter) { const char *reason = NULL; CRM_ASSERT(is_primitive_action(first)); CRM_ASSERT(is_primitive_action(then)); // We need to update the action in two cases: // ... if 'then' is required if (pcmk_is_set(filter, pe_action_optional) && !pcmk_is_set(then->flags, pe_action_optional)) { reason = "restart"; } /* ... if 'then' is unrunnable action on same resource (if a resource * should restart but can't start, we still want to stop) */ if (pcmk_is_set(filter, pe_action_runnable) && !pcmk_is_set(then->flags, pe_action_runnable) && pcmk_is_set(then->rsc->flags, pe_rsc_managed) && (first->rsc == then->rsc)) { reason = "stop"; } if (reason == NULL) { return; } pe_rsc_trace(first->rsc, "Handling %s -> %s for %s", first->uuid, then->uuid, reason); // Make 'first' required if it is runnable if (pcmk_is_set(first->flags, pe_action_runnable)) { clear_action_flag_because(first, pe_action_optional, then); } // Make 'first' required if 'then' is required if (!pcmk_is_set(then->flags, pe_action_optional)) { clear_action_flag_because(first, pe_action_optional, then); } // Make 'first' unmigratable if 'then' is unmigratable if (!pcmk_is_set(then->flags, pe_action_migrate_runnable)) { clear_action_flag_because(first, pe_action_migrate_runnable, then); } // Make 'then' unrunnable if 'first' is required but unrunnable if (!pcmk_is_set(first->flags, pe_action_optional) && !pcmk_is_set(first->flags, pe_action_runnable)) { clear_action_flag_because(then, pe_action_runnable, first); } } enum pe_graph_flags native_update_actions(pe_action_t *first, pe_action_t *then, pe_node_t *node, enum pe_action_flags flags, enum pe_action_flags filter, enum pe_ordering type, pe_working_set_t *data_set) { /* flags == get_action_flags(first, then_node) called from update_action() */ enum pe_graph_flags changed = pe_graph_none; enum pe_action_flags then_flags = then->flags; enum pe_action_flags first_flags = first->flags; if (type & pe_order_asymmetrical) { pe_resource_t *then_rsc = then->rsc; enum rsc_role_e then_rsc_role = then_rsc ? then_rsc->fns->state(then_rsc, TRUE) : 0; if (!then_rsc) { /* ignore */ } else if ((then_rsc_role == RSC_ROLE_STOPPED) && pcmk__str_eq(then->task, RSC_STOP, pcmk__str_casei)) { /* ignore... if 'then' is supposed to be stopped after 'first', but * then is already stopped, there is nothing to be done when non-symmetrical. */ } else if ((then_rsc_role >= RSC_ROLE_STARTED) && pcmk__str_eq(then->task, RSC_START, pcmk__str_casei) && pcmk_is_set(then->flags, pe_action_optional) && then->node && pcmk__list_of_1(then_rsc->running_on) && then->node->details == ((pe_node_t *) then_rsc->running_on->data)->details) { /* Ignore. If 'then' is supposed to be started after 'first', but * 'then' is already started, there is nothing to be done when * asymmetrical -- unless the start is mandatory, which indicates * the resource is restarting, and the ordering is still needed. */ } else if (!(first->flags & pe_action_runnable)) { /* prevent 'then' action from happening if 'first' is not runnable and * 'then' has not yet occurred. */ clear_action_flag_because(then, pe_action_optional, first); clear_action_flag_because(then, pe_action_runnable, first); } else { /* ignore... then is allowed to start/stop if it wants to. */ } } if (pcmk_is_set(type, pe_order_implies_first) && !pcmk_is_set(then_flags, pe_action_optional)) { // Then is required, and implies first should be, too if (pcmk_is_set(filter, pe_action_optional) && !pcmk_is_set(flags, pe_action_optional) && pcmk_is_set(first_flags, pe_action_optional)) { clear_action_flag_because(first, pe_action_optional, then); } if (pcmk_is_set(flags, pe_action_migrate_runnable) && !pcmk_is_set(then->flags, pe_action_migrate_runnable)) { clear_action_flag_because(first, pe_action_migrate_runnable, then); } } if (type & pe_order_promoted_implies_first) { if ((filter & pe_action_optional) && ((then->flags & pe_action_optional) == FALSE) && (then->rsc != NULL) && (then->rsc->role == RSC_ROLE_PROMOTED)) { clear_action_flag_because(first, pe_action_optional, then); if (pcmk_is_set(first->flags, pe_action_migrate_runnable) && !pcmk_is_set(then->flags, pe_action_migrate_runnable)) { clear_action_flag_because(first, pe_action_migrate_runnable, then); } } } if ((type & pe_order_implies_first_migratable) && pcmk_is_set(filter, pe_action_optional)) { if (((then->flags & pe_action_migrate_runnable) == FALSE) || ((then->flags & pe_action_runnable) == FALSE)) { clear_action_flag_because(first, pe_action_runnable, then); } if ((then->flags & pe_action_optional) == 0) { clear_action_flag_because(first, pe_action_optional, then); } } if ((type & pe_order_pseudo_left) && pcmk_is_set(filter, pe_action_optional)) { if ((first->flags & pe_action_runnable) == FALSE) { clear_action_flag_because(then, pe_action_migrate_runnable, first); pe__clear_action_flags(then, pe_action_pseudo); } } if (pcmk_is_set(type, pe_order_runnable_left) && pcmk_is_set(filter, pe_action_runnable) && pcmk_is_set(then->flags, pe_action_runnable) && !pcmk_is_set(flags, pe_action_runnable)) { clear_action_flag_because(then, pe_action_runnable, first); clear_action_flag_because(then, pe_action_migrate_runnable, first); } if (pcmk_is_set(type, pe_order_implies_then) && pcmk_is_set(filter, pe_action_optional) && pcmk_is_set(then->flags, pe_action_optional) && !pcmk_is_set(flags, pe_action_optional) && !pcmk_is_set(first->flags, pe_action_migrate_runnable)) { clear_action_flag_because(then, pe_action_optional, first); } if (pcmk_is_set(type, pe_order_restart)) { handle_restart_ordering(first, then, filter); } if (then_flags != then->flags) { pe__set_graph_flags(changed, first, pe_graph_updated_then); pe_rsc_trace(then->rsc, "%s on %s: flags are now 0x%.6x (was 0x%.6x) " "because of 'first' %s (0x%.6x)", then->uuid, then->node? then->node->details->uname : "no node", then->flags, then_flags, first->uuid, first->flags); if(then->rsc && then->rsc->parent) { /* "X_stop then X_start" doesn't get handled for cloned groups unless we do this */ update_action(then, data_set); } } if (first_flags != first->flags) { pe__set_graph_flags(changed, first, pe_graph_updated_first); pe_rsc_trace(first->rsc, "%s on %s: flags are now 0x%.6x (was 0x%.6x) " "because of 'then' %s (0x%.6x)", first->uuid, first->node? first->node->details->uname : "no node", first->flags, first_flags, then->uuid, then->flags); } return changed; } void native_rsc_location(pe_resource_t *rsc, pe__location_t *constraint) { GList *gIter = NULL; bool need_role = false; CRM_CHECK((constraint != NULL) && (rsc != NULL), return); // If a role was specified, ensure constraint is applicable need_role = (constraint->role_filter > RSC_ROLE_UNKNOWN); if (need_role && (constraint->role_filter != rsc->next_role)) { pe_rsc_trace(rsc, "Not applying %s to %s because role will be %s not %s", constraint->id, rsc->id, role2text(rsc->next_role), role2text(constraint->role_filter)); return; } if (constraint->node_list_rh == NULL) { pe_rsc_trace(rsc, "Not applying %s to %s because no nodes match", constraint->id, rsc->id); return; } pe_rsc_trace(rsc, "Applying %s%s%s to %s", constraint->id, (need_role? " for role " : ""), (need_role? role2text(constraint->role_filter) : ""), rsc->id); for (gIter = constraint->node_list_rh; gIter != NULL; gIter = gIter->next) { pe_node_t *node = (pe_node_t *) gIter->data; pe_node_t *other_node = NULL; other_node = (pe_node_t *) pe_hash_table_lookup(rsc->allowed_nodes, node->details->id); if (other_node != NULL) { pe_rsc_trace(rsc, "* + %d on %s", node->weight, node->details->uname); other_node->weight = pe__add_scores(other_node->weight, node->weight); } else { pe_rsc_trace(rsc, "* = %d on %s", node->weight, node->details->uname); other_node = pe__copy_node(node); g_hash_table_insert(rsc->allowed_nodes, (gpointer) other_node->details->id, other_node); } if (other_node->rsc_discover_mode < constraint->discover_mode) { if (constraint->discover_mode == pe_discover_exclusive) { rsc->exclusive_discover = TRUE; } /* exclusive > never > always... always is default */ other_node->rsc_discover_mode = constraint->discover_mode; } } } void native_expand(pe_resource_t * rsc, pe_working_set_t * data_set) { GList *gIter = NULL; CRM_ASSERT(rsc); pe_rsc_trace(rsc, "Processing actions from %s", rsc->id); for (gIter = rsc->actions; gIter != NULL; gIter = gIter->next) { pe_action_t *action = (pe_action_t *) gIter->data; crm_trace("processing action %d for rsc=%s", action->id, rsc->id); graph_element_from_action(action, data_set); } for (gIter = rsc->children; gIter != NULL; gIter = gIter->next) { pe_resource_t *child_rsc = (pe_resource_t *) gIter->data; child_rsc->cmds->expand(child_rsc, data_set); } } #define STOP_SANITY_ASSERT(lineno) do { \ if(current && current->details->unclean) { \ /* It will be a pseudo op */ \ } else if(stop == NULL) { \ crm_err("%s:%d: No stop action exists for %s", \ __func__, lineno, rsc->id); \ CRM_ASSERT(stop != NULL); \ } else if (pcmk_is_set(stop->flags, pe_action_optional)) { \ crm_err("%s:%d: Action %s is still optional", \ __func__, lineno, stop->uuid); \ CRM_ASSERT(!pcmk_is_set(stop->flags, pe_action_optional)); \ } \ } while(0) void LogActions(pe_resource_t * rsc, pe_working_set_t * data_set) { pcmk__output_t *out = data_set->priv; pe_node_t *next = NULL; pe_node_t *current = NULL; gboolean moving = FALSE; if(rsc->variant == pe_container) { pcmk__bundle_log_actions(rsc, data_set); return; } if (rsc->children) { g_list_foreach(rsc->children, (GFunc) LogActions, data_set); return; } next = rsc->allocated_to; if (rsc->running_on) { current = pe__current_node(rsc); if (rsc->role == RSC_ROLE_STOPPED) { /* * This can occur when resources are being recovered * We fiddle with the current role in native_create_actions() */ rsc->role = RSC_ROLE_STARTED; } } if ((current == NULL) && pcmk_is_set(rsc->flags, pe_rsc_orphan)) { /* Don't log stopped orphans */ return; } out->message(out, "rsc-action", rsc, current, next, moving); } gboolean StopRsc(pe_resource_t * rsc, pe_node_t * next, gboolean optional, pe_working_set_t * data_set) { GList *gIter = NULL; CRM_ASSERT(rsc); pe_rsc_trace(rsc, "%s", rsc->id); for (gIter = rsc->running_on; gIter != NULL; gIter = gIter->next) { pe_node_t *current = (pe_node_t *) gIter->data; pe_action_t *stop; if (rsc->partial_migration_target) { if (rsc->partial_migration_target->details == current->details) { pe_rsc_trace(rsc, "Filtered %s -> %s %s", current->details->uname, next->details->uname, rsc->id); continue; } else { pe_rsc_trace(rsc, "Forced on %s %s", current->details->uname, rsc->id); optional = FALSE; } } pe_rsc_trace(rsc, "%s on %s", rsc->id, current->details->uname); stop = stop_action(rsc, current, optional); if(rsc->allocated_to == NULL) { pe_action_set_reason(stop, "node availability", TRUE); } if (!pcmk_is_set(rsc->flags, pe_rsc_managed)) { pe__clear_action_flags(stop, pe_action_runnable); } if (pcmk_is_set(data_set->flags, pe_flag_remove_after_stop)) { DeleteRsc(rsc, current, optional, data_set); } if (pcmk_is_set(rsc->flags, pe_rsc_needs_unfencing)) { pe_action_t *unfence = pe_fence_op(current, "on", TRUE, NULL, FALSE, data_set); order_actions(stop, unfence, pe_order_implies_first); - if (!node_has_been_unfenced(current)) { + if (!pcmk__node_unfenced(current)) { pe_proc_err("Stopping %s until %s can be unfenced", rsc->id, current->details->uname); } } } return TRUE; } -static void -order_after_unfencing(pe_resource_t *rsc, pe_node_t *node, pe_action_t *action, - enum pe_ordering order, pe_working_set_t *data_set) -{ - /* When unfencing is in use, we order unfence actions before any probe or - * start of resources that require unfencing, and also of fence devices. - * - * This might seem to violate the principle that fence devices require - * only quorum. However, fence agents that unfence often don't have enough - * information to even probe or start unless the node is first unfenced. - */ - if (is_unfence_device(rsc, data_set) - || pcmk_is_set(rsc->flags, pe_rsc_needs_unfencing)) { - - /* Start with an optional ordering. Requiring unfencing would result in - * the node being unfenced, and all its resources being stopped, - * whenever a new resource is added -- which would be highly suboptimal. - */ - pe_action_t *unfence = pe_fence_op(node, "on", TRUE, NULL, FALSE, data_set); - - order_actions(unfence, action, order); - - if (!node_has_been_unfenced(node)) { - // But unfencing is required if it has never been done - char *reason = crm_strdup_printf("required by %s %s", - rsc->id, action->task); - - trigger_unfencing(NULL, node, reason, NULL, data_set); - free(reason); - } - } -} - gboolean StartRsc(pe_resource_t * rsc, pe_node_t * next, gboolean optional, pe_working_set_t * data_set) { pe_action_t *start = NULL; CRM_ASSERT(rsc); pe_rsc_trace(rsc, "%s on %s %d %d", rsc->id, next ? next->details->uname : "N/A", optional, next ? next->weight : 0); start = start_action(rsc, next, TRUE); - order_after_unfencing(rsc, next, start, pe_order_implies_then, data_set); + pcmk__order_vs_unfence(rsc, next, start, pe_order_implies_then, data_set); if (pcmk_is_set(start->flags, pe_action_runnable) && !optional) { pe__clear_action_flags(start, pe_action_optional); } return TRUE; } gboolean PromoteRsc(pe_resource_t * rsc, pe_node_t * next, gboolean optional, pe_working_set_t * data_set) { GList *gIter = NULL; gboolean runnable = TRUE; GList *action_list = NULL; CRM_ASSERT(rsc); CRM_CHECK(next != NULL, return FALSE); pe_rsc_trace(rsc, "%s on %s", rsc->id, next->details->uname); action_list = pe__resource_actions(rsc, next, RSC_START, TRUE); for (gIter = action_list; gIter != NULL; gIter = gIter->next) { pe_action_t *start = (pe_action_t *) gIter->data; if (!pcmk_is_set(start->flags, pe_action_runnable)) { runnable = FALSE; } } g_list_free(action_list); if (runnable) { promote_action(rsc, next, optional); return TRUE; } pe_rsc_debug(rsc, "%s\tPromote %s (canceled)", next->details->uname, rsc->id); action_list = pe__resource_actions(rsc, next, RSC_PROMOTE, TRUE); for (gIter = action_list; gIter != NULL; gIter = gIter->next) { pe_action_t *promote = (pe_action_t *) gIter->data; pe__clear_action_flags(promote, pe_action_runnable); } g_list_free(action_list); return TRUE; } gboolean DemoteRsc(pe_resource_t * rsc, pe_node_t * next, gboolean optional, pe_working_set_t * data_set) { GList *gIter = NULL; CRM_ASSERT(rsc); pe_rsc_trace(rsc, "%s", rsc->id); /* CRM_CHECK(rsc->next_role == RSC_ROLE_UNPROMOTED, return FALSE); */ for (gIter = rsc->running_on; gIter != NULL; gIter = gIter->next) { pe_node_t *current = (pe_node_t *) gIter->data; pe_rsc_trace(rsc, "%s on %s", rsc->id, next ? next->details->uname : "N/A"); demote_action(rsc, current, optional); } return TRUE; } gboolean RoleError(pe_resource_t * rsc, pe_node_t * next, gboolean optional, pe_working_set_t * data_set) { CRM_ASSERT(rsc); crm_err("%s on %s", rsc->id, next ? next->details->uname : "N/A"); CRM_CHECK(FALSE, return FALSE); return FALSE; } gboolean NullOp(pe_resource_t * rsc, pe_node_t * next, gboolean optional, pe_working_set_t * data_set) { CRM_ASSERT(rsc); pe_rsc_trace(rsc, "%s", rsc->id); return FALSE; } gboolean DeleteRsc(pe_resource_t * rsc, pe_node_t * node, gboolean optional, pe_working_set_t * data_set) { if (pcmk_is_set(rsc->flags, pe_rsc_failed)) { pe_rsc_trace(rsc, "Resource %s not deleted from %s: failed", rsc->id, node->details->uname); return FALSE; } else if (node == NULL) { pe_rsc_trace(rsc, "Resource %s not deleted: NULL node", rsc->id); return FALSE; } else if (node->details->unclean || node->details->online == FALSE) { pe_rsc_trace(rsc, "Resource %s not deleted from %s: unrunnable", rsc->id, node->details->uname); return FALSE; } crm_notice("Removing %s from %s", rsc->id, node->details->uname); delete_action(rsc, node, optional); new_rsc_order(rsc, RSC_STOP, rsc, RSC_DELETE, optional ? pe_order_implies_then : pe_order_optional, data_set); new_rsc_order(rsc, RSC_DELETE, rsc, RSC_START, optional ? pe_order_implies_then : pe_order_optional, data_set); return TRUE; } gboolean native_create_probe(pe_resource_t * rsc, pe_node_t * node, pe_action_t * complete, gboolean force, pe_working_set_t * data_set) { enum pe_ordering flags = pe_order_optional; char *key = NULL; pe_action_t *probe = NULL; pe_node_t *running = NULL; pe_node_t *allowed = NULL; pe_resource_t *top = uber_parent(rsc); static const char *rc_promoted = NULL; static const char *rc_inactive = NULL; if (rc_inactive == NULL) { rc_inactive = pcmk__itoa(PCMK_OCF_NOT_RUNNING); rc_promoted = pcmk__itoa(PCMK_OCF_RUNNING_PROMOTED); } CRM_CHECK(node != NULL, return FALSE); if (!force && !pcmk_is_set(data_set->flags, pe_flag_startup_probes)) { pe_rsc_trace(rsc, "Skipping active resource detection for %s", rsc->id); return FALSE; } 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_casei)) { pe_rsc_trace(rsc, "Skipping probe for %s on %s because Pacemaker Remote nodes cannot run stonith agents", rsc->id, node->details->id); return FALSE; } else if (pe__is_guest_node(node) && pe__resource_contains_guest_node(data_set, rsc)) { pe_rsc_trace(rsc, "Skipping probe for %s on %s because guest nodes cannot run resources containing guest nodes", rsc->id, node->details->id); return FALSE; } else if (rsc->is_remote_node) { pe_rsc_trace(rsc, "Skipping probe for %s on %s because Pacemaker Remote nodes cannot host remote connections", rsc->id, node->details->id); return FALSE; } } if (rsc->children) { GList *gIter = NULL; gboolean any_created = FALSE; for (gIter = rsc->children; gIter != NULL; gIter = gIter->next) { pe_resource_t *child_rsc = (pe_resource_t *) gIter->data; any_created = child_rsc->cmds->create_probe(child_rsc, node, complete, force, data_set) || any_created; } return any_created; } else if ((rsc->container) && (!rsc->is_remote_node)) { pe_rsc_trace(rsc, "Skipping %s: it is within container %s", rsc->id, rsc->container->id); return FALSE; } if (pcmk_is_set(rsc->flags, pe_rsc_orphan)) { pe_rsc_trace(rsc, "Skipping orphan: %s", rsc->id); return FALSE; } // Check whether resource is already known on node if (!force && g_hash_table_lookup(rsc->known_on, node->details->id)) { pe_rsc_trace(rsc, "Skipping known: %s on %s", rsc->id, node->details->uname); return FALSE; } allowed = g_hash_table_lookup(rsc->allowed_nodes, node->details->id); if (rsc->exclusive_discover || top->exclusive_discover) { if (allowed == NULL) { /* exclusive discover is enabled and this node is not in the allowed list. */ pe_rsc_trace(rsc, "Skipping probe for %s on node %s, A", rsc->id, node->details->id); return FALSE; } else if (allowed->rsc_discover_mode != pe_discover_exclusive) { /* exclusive discover is enabled and this node is not marked * as a node this resource should be discovered on */ pe_rsc_trace(rsc, "Skipping probe for %s on node %s, B", rsc->id, node->details->id); return FALSE; } } if(allowed == NULL && node->rsc_discover_mode == pe_discover_never) { /* If this node was allowed to host this resource it would * have been explicitly added to the 'allowed_nodes' list. * However it wasn't and the node has discovery disabled, so * no need to probe for this resource. */ pe_rsc_trace(rsc, "Skipping probe for %s on node %s, C", rsc->id, node->details->id); return FALSE; } if (allowed && allowed->rsc_discover_mode == pe_discover_never) { /* this resource is marked as not needing to be discovered on this node */ pe_rsc_trace(rsc, "Skipping probe for %s on node %s, discovery mode", rsc->id, node->details->id); return FALSE; } if (pe__is_guest_node(node)) { pe_resource_t *remote = node->details->remote_rsc->container; if(remote->role == RSC_ROLE_STOPPED) { /* If the container is stopped, then we know anything that * might have been inside it is also stopped and there is * no need to probe. * * If we don't know the container's state on the target * either: * * - the container is running, the transition will abort * and we'll end up in a different case next time, or * * - the container is stopped * * Either way there is no need to probe. * */ if(remote->allocated_to && g_hash_table_lookup(remote->known_on, remote->allocated_to->details->id) == NULL) { /* For safety, we order the 'rsc' start after 'remote' * has been probed. * * Using 'top' helps for groups, but we may need to * follow the start's ordering chain backwards. */ custom_action_order(remote, pcmk__op_key(remote->id, RSC_STATUS, 0), NULL, top, pcmk__op_key(top->id, RSC_START, 0), NULL, pe_order_optional, data_set); } pe_rsc_trace(rsc, "Skipping probe for %s on node %s, %s is stopped", rsc->id, node->details->id, remote->id); return FALSE; /* Here we really we want to check if remote->stop is required, * but that information doesn't exist yet */ } else if(node->details->remote_requires_reset || node->details->unclean || pcmk_is_set(remote->flags, pe_rsc_failed) || remote->next_role == RSC_ROLE_STOPPED || (remote->allocated_to && pe_find_node(remote->running_on, remote->allocated_to->details->uname) == NULL) ) { /* The container is stopping or restarting, don't start * 'rsc' until 'remote' stops as this also implies that * 'rsc' is stopped - avoiding the need to probe */ custom_action_order(remote, pcmk__op_key(remote->id, RSC_STOP, 0), NULL, top, pcmk__op_key(top->id, RSC_START, 0), NULL, pe_order_optional, data_set); pe_rsc_trace(rsc, "Skipping probe for %s on node %s, %s is stopping, restarting or moving", rsc->id, node->details->id, remote->id); return FALSE; /* } else { * The container is running so there is no problem probing it */ } } key = pcmk__op_key(rsc->id, RSC_STATUS, 0); probe = custom_action(rsc, key, RSC_STATUS, node, FALSE, TRUE, data_set); pe__clear_action_flags(probe, pe_action_optional); - order_after_unfencing(rsc, node, probe, pe_order_optional, data_set); + pcmk__order_vs_unfence(rsc, node, probe, pe_order_optional, data_set); /* * We need to know if it's running_on (not just known_on) this node * to correctly determine the target rc. */ running = pe_find_node_id(rsc->running_on, node->details->id); if (running == NULL) { add_hash_param(probe->meta, XML_ATTR_TE_TARGET_RC, rc_inactive); } else if (rsc->role == RSC_ROLE_PROMOTED) { add_hash_param(probe->meta, XML_ATTR_TE_TARGET_RC, rc_promoted); } crm_debug("Probing %s on %s (%s) %d %p", rsc->id, node->details->uname, role2text(rsc->role), pcmk_is_set(probe->flags, pe_action_runnable), rsc->running_on); - if (is_unfence_device(rsc, data_set) || !pe_rsc_is_clone(top)) { + if (pcmk__is_unfence_device(rsc, data_set) || !pe_rsc_is_clone(top)) { top = rsc; } else { crm_trace("Probing %s on %s (%s) as %s", rsc->id, node->details->uname, role2text(rsc->role), top->id); } if (!pcmk_is_set(probe->flags, pe_action_runnable) && (rsc->running_on == NULL)) { /* Prevent the start from occurring if rsc isn't active, but * don't cause it to stop if it was active already */ pe__set_order_flags(flags, pe_order_runnable_left); } custom_action_order(rsc, NULL, probe, top, pcmk__op_key(top->id, RSC_START, 0), NULL, flags, data_set); // Order the probe before any agent reload custom_action_order(rsc, NULL, probe, top, reload_key(rsc), NULL, pe_order_optional, data_set); #if 0 // complete is always null currently - if (!is_unfence_device(rsc, data_set)) { + if (!pcmk__is_unfence_device(rsc, data_set)) { /* Normally rsc.start depends on probe complete which depends * on rsc.probe. But this can't be the case for fence devices * with unfencing, as it would create graph loops. * * So instead we explicitly order 'rsc.probe then rsc.start' */ order_actions(probe, complete, pe_order_implies_then); } #endif return TRUE; } -/*! - * \internal - * \brief Check whether a resource is known on a particular node - * - * \param[in] rsc Resource to check - * \param[in] node Node to check - * - * \return TRUE if resource (or parent if an anonymous clone) is known - */ -static bool -rsc_is_known_on(pe_resource_t *rsc, const pe_node_t *node) -{ - if (pe_hash_table_lookup(rsc->known_on, node->details->id)) { - return TRUE; - - } else if ((rsc->variant == pe_native) - && pe_rsc_is_anon_clone(rsc->parent) - && pe_hash_table_lookup(rsc->parent->known_on, node->details->id)) { - /* We check only the parent, not the uber-parent, because we cannot - * assume that the resource is known if it is in an anonymously cloned - * group (which may be only partially known). - */ - return TRUE; - } - return FALSE; -} - -/*! - * \internal - * \brief Order a resource's start and promote actions relative to fencing - * - * \param[in] rsc Resource to be ordered - * \param[in] stonith_op Fence action - * \param[in] data_set Cluster information - */ -static void -native_start_constraints(pe_resource_t * rsc, pe_action_t * stonith_op, pe_working_set_t * data_set) -{ - pe_node_t *target; - GList *gIter = NULL; - - CRM_CHECK(stonith_op && stonith_op->node, return); - target = stonith_op->node; - - for (gIter = rsc->actions; gIter != NULL; gIter = gIter->next) { - pe_action_t *action = (pe_action_t *) gIter->data; - - switch (action->needs) { - case rsc_req_nothing: - // Anything other than start or promote requires nothing - break; - - case rsc_req_stonith: - order_actions(stonith_op, action, pe_order_optional); - break; - - case rsc_req_quorum: - if (pcmk__str_eq(action->task, RSC_START, pcmk__str_casei) - && pe_hash_table_lookup(rsc->allowed_nodes, target->details->id) - && !rsc_is_known_on(rsc, target)) { - - /* If we don't know the status of the resource on the node - * we're about to shoot, we have to assume it may be active - * there. Order the resource start after the fencing. This - * is analogous to waiting for all the probes for a resource - * to complete before starting it. - * - * The most likely explanation is that the DC died and took - * its status with it. - */ - pe_rsc_debug(rsc, "Ordering %s after %s recovery", action->uuid, - target->details->uname); - order_actions(stonith_op, action, - pe_order_optional | pe_order_runnable_left); - } - break; - } - } -} - -static void -native_stop_constraints(pe_resource_t * rsc, pe_action_t * stonith_op, pe_working_set_t * data_set) -{ - GList *gIter = NULL; - GList *action_list = NULL; - bool order_implicit = false; - - pe_resource_t *top = uber_parent(rsc); - pe_action_t *parent_stop = NULL; - pe_node_t *target; - - CRM_CHECK(stonith_op && stonith_op->node, return); - target = stonith_op->node; - - /* Get a list of stop actions potentially implied by the fencing */ - action_list = pe__resource_actions(rsc, target, RSC_STOP, FALSE); - - /* If resource requires fencing, implicit actions must occur after fencing. - * - * Implied stops and demotes of resources running on guest nodes are always - * ordered after fencing, even if the resource does not require fencing, - * because guest node "fencing" is actually just a resource stop. - */ - if (pcmk_is_set(rsc->flags, pe_rsc_needs_fencing) - || pe__is_guest_node(target)) { - - order_implicit = true; - } - - if (action_list && order_implicit) { - parent_stop = find_first_action(top->actions, NULL, RSC_STOP, NULL); - } - - for (gIter = action_list; gIter != NULL; gIter = gIter->next) { - pe_action_t *action = (pe_action_t *) gIter->data; - - // The stop would never complete, so convert it into a pseudo-action. - pe__set_action_flags(action, pe_action_pseudo|pe_action_runnable); - - if (order_implicit) { - pe__set_action_flags(action, pe_action_implied_by_stonith); - - /* Order the stonith before the parent stop (if any). - * - * Also order the stonith before the resource stop, unless the - * resource is inside a bundle -- that would cause a graph loop. - * We can rely on the parent stop's ordering instead. - * - * User constraints must not order a resource in a guest node - * relative to the guest node container resource. The - * pe_order_preserve flag marks constraints as generated by the - * cluster and thus immune to that check (and is irrelevant if - * target is not a guest). - */ - if (!pe_rsc_is_bundled(rsc)) { - order_actions(stonith_op, action, pe_order_preserve); - } - order_actions(stonith_op, parent_stop, pe_order_preserve); - } - - if (pcmk_is_set(rsc->flags, pe_rsc_failed)) { - crm_notice("Stop of failed resource %s is implicit %s %s is fenced", - rsc->id, (order_implicit? "after" : "because"), - target->details->uname); - } else { - crm_info("%s is implicit %s %s is fenced", - action->uuid, (order_implicit? "after" : "because"), - target->details->uname); - } - - if (pcmk_is_set(rsc->flags, pe_rsc_notify)) { - /* Create a second notification that will be delivered - * immediately after the node is fenced - * - * Basic problem: - * - C is a clone active on the node to be shot and stopping on another - * - R is a resource that depends on C - * - * + C.stop depends on R.stop - * + C.stopped depends on STONITH - * + C.notify depends on C.stopped - * + C.healthy depends on C.notify - * + R.stop depends on C.healthy - * - * The extra notification here changes - * + C.healthy depends on C.notify - * into: - * + C.healthy depends on C.notify' - * + C.notify' depends on STONITH' - * thus breaking the loop - */ - create_secondary_notification(action, rsc, stonith_op, data_set); - } - -/* From Bug #1601, successful fencing must be an input to a failed resources stop action. - - However given group(rA, rB) running on nodeX and B.stop has failed, - A := stop healthy resource (rA.stop) - B := stop failed resource (pseudo operation B.stop) - C := stonith nodeX - A requires B, B requires C, C requires A - This loop would prevent the cluster from making progress. - - This block creates the "C requires A" dependency and therefore must (at least - for now) be disabled. - - Instead, run the block above and treat all resources on nodeX as B would be - (marked as a pseudo op depending on the STONITH). - - TODO: Break the "A requires B" dependency in update_action() and re-enable this block - - } else if(is_stonith == FALSE) { - crm_info("Moving healthy resource %s" - " off %s before fencing", - rsc->id, node->details->uname); - - * stop healthy resources before the - * stonith op - * - custom_action_order( - rsc, stop_key(rsc), NULL, - NULL,strdup(CRM_OP_FENCE),stonith_op, - pe_order_optional, data_set); -*/ - } - - g_list_free(action_list); - - /* Get a list of demote actions potentially implied by the fencing */ - action_list = pe__resource_actions(rsc, target, RSC_DEMOTE, FALSE); - - for (gIter = action_list; gIter != NULL; gIter = gIter->next) { - pe_action_t *action = (pe_action_t *) gIter->data; - - if (action->node->details->online == FALSE || action->node->details->unclean == TRUE - || pcmk_is_set(rsc->flags, pe_rsc_failed)) { - - if (pcmk_is_set(rsc->flags, pe_rsc_failed)) { - pe_rsc_info(rsc, - "Demote of failed resource %s is implicit after %s is fenced", - rsc->id, target->details->uname); - } else { - pe_rsc_info(rsc, "%s is implicit after %s is fenced", - action->uuid, target->details->uname); - } - - /* The demote would never complete and is now implied by the - * fencing, so convert it into a pseudo-action. - */ - pe__set_action_flags(action, pe_action_pseudo|pe_action_runnable); - - if (pe_rsc_is_bundled(rsc)) { - /* Do nothing, let the recovery be ordered after the parent's implied stop */ - - } else if (order_implicit) { - order_actions(stonith_op, action, pe_order_preserve|pe_order_optional); - } - } - } - - g_list_free(action_list); -} - -void -rsc_stonith_ordering(pe_resource_t * rsc, pe_action_t * stonith_op, pe_working_set_t * data_set) -{ - if (rsc->children) { - GList *gIter = NULL; - - for (gIter = rsc->children; gIter != NULL; gIter = gIter->next) { - pe_resource_t *child_rsc = (pe_resource_t *) gIter->data; - - rsc_stonith_ordering(child_rsc, stonith_op, data_set); - } - - } else if (!pcmk_is_set(rsc->flags, pe_rsc_managed)) { - pe_rsc_trace(rsc, "Skipping fencing constraints for unmanaged resource: %s", rsc->id); - - } else { - native_start_constraints(rsc, stonith_op, data_set); - native_stop_constraints(rsc, stonith_op, data_set); - } -} - void ReloadRsc(pe_resource_t * rsc, pe_node_t *node, pe_working_set_t * data_set) { GList *gIter = NULL; pe_action_t *reload = NULL; if (rsc->children) { for (gIter = rsc->children; gIter != NULL; gIter = gIter->next) { pe_resource_t *child_rsc = (pe_resource_t *) gIter->data; ReloadRsc(child_rsc, node, data_set); } return; } else if (rsc->variant > pe_native) { /* Complex resource with no children */ return; } else if (!pcmk_is_set(rsc->flags, pe_rsc_managed)) { pe_rsc_trace(rsc, "%s: unmanaged", rsc->id); return; } else if (pcmk_is_set(rsc->flags, pe_rsc_failed)) { /* We don't need to specify any particular actions here, normal failure * recovery will apply. */ pe_rsc_trace(rsc, "%s: preventing agent reload because failed", rsc->id); return; } else if (pcmk_is_set(rsc->flags, pe_rsc_start_pending)) { /* If a resource's configuration changed while a start was pending, * force a full restart. */ pe_rsc_trace(rsc, "%s: preventing agent reload because start pending", rsc->id); stop_action(rsc, node, FALSE); return; } else if (node == NULL) { pe_rsc_trace(rsc, "%s: not active", rsc->id); return; } pe_rsc_trace(rsc, "Processing %s", rsc->id); pe__set_resource_flags(rsc, pe_rsc_reload); reload = custom_action(rsc, reload_key(rsc), CRMD_ACTION_RELOAD_AGENT, node, FALSE, TRUE, data_set); pe_action_set_reason(reload, "resource definition change", FALSE); custom_action_order(NULL, NULL, reload, rsc, stop_key(rsc), NULL, pe_order_optional|pe_order_then_cancels_first, data_set); custom_action_order(NULL, NULL, reload, rsc, demote_key(rsc), NULL, pe_order_optional|pe_order_then_cancels_first, data_set); } void native_append_meta(pe_resource_t * rsc, xmlNode * xml) { char *value = g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_INCARNATION); pe_resource_t *parent; if (value) { char *name = NULL; name = crm_meta_name(XML_RSC_ATTR_INCARNATION); crm_xml_add(xml, name, value); free(name); } value = g_hash_table_lookup(rsc->meta, XML_RSC_ATTR_REMOTE_NODE); if (value) { char *name = NULL; name = crm_meta_name(XML_RSC_ATTR_REMOTE_NODE); crm_xml_add(xml, name, value); free(name); } for (parent = rsc; parent != NULL; parent = parent->parent) { if (parent->container) { crm_xml_add(xml, CRM_META"_"XML_RSC_ATTR_CONTAINER, parent->container->id); } } }