diff --git a/include/crm/cluster/election_internal.h b/include/crm/cluster/election_internal.h index 3eb90825f2..665c3a0357 100644 --- a/include/crm/cluster/election_internal.h +++ b/include/crm/cluster/election_internal.h @@ -1,85 +1,86 @@ /* - * Copyright 2009-2019 the Pacemaker project contributors + * Copyright 2009-2022 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 CRM_COMMON_ELECTION__H # define CRM_COMMON_ELECTION__H #ifdef __cplusplus extern "C" { #endif /** * \file * \brief Functions for conducting elections * * An election is useful for a daemon that runs on all nodes but needs any one * instance to perform a special role. * * Elections are closely tied to the cluster peer cache. Peers in the cache that * are active members are eligible to vote. Elections are named for logging * purposes, but only one election may exist at any time, so typically an * election would be created at daemon start-up and freed at shutdown. * * Pacemaker's election procedure has been heavily adapted from the * Invitation Algorithm variant of the Garcia-Molina Bully Algorithm: * * https://en.wikipedia.org/wiki/Bully_algorithm * * Elections are conducted via cluster messages. There are two types of * messages: a "vote" is a declaration of the voting node's candidacy, and is * always broadcast; a "no-vote" is a concession by the responding node, and is * always a reply to the preferred node's vote. (These correspond to "invite" * and "accept" in the traditional algorithm.) * * A vote together with any no-vote replies to it is considered an election * round. Rounds are numbered with a simple counter unique to each node * (this would be the group number in the traditional algorithm). Concurrent * election rounds are possible. * * An election round is started when any node broadcasts a vote. When a node * receives another node's vote, it compares itself against the sending node * according to certain metrics, and either starts a new round (if it prefers * itself) or replies to the other node with a no-vote (if it prefers that * node). * * If a node receives no-votes from all other active nodes, it declares itself * the winner. The library API does not notify other nodes of this; callers * must implement that if desired. */ typedef struct election_s election_t; /*! Possible election states */ enum election_result { election_start = 0, /*! new election needed */ election_in_progress, /*! election started but not all peers have voted */ election_lost, /*! local node lost most recent election */ election_won, /*! local node won most recent election */ election_error, /*! election message or election object invalid */ }; void election_fini(election_t *e); void election_reset(election_t *e); election_t *election_init(const char *name, const char *uname, guint period_ms, GSourceFunc cb); void election_timeout_set_period(election_t *e, guint period_ms); void election_timeout_stop(election_t *e); void election_vote(election_t *e); bool election_check(election_t *e); void election_remove(election_t *e, const char *uname); -enum election_result election_state(election_t *e); -enum election_result election_count_vote(election_t *e, xmlNode *vote, bool can_win); +enum election_result election_state(const election_t *e); +enum election_result election_count_vote(election_t *e, const xmlNode *message, + bool can_win); void election_clear_dampening(election_t *e); #ifdef __cplusplus } #endif #endif diff --git a/lib/cluster/cluster.c b/lib/cluster/cluster.c index 066232c81f..20974b5e05 100644 --- a/lib/cluster/cluster.c +++ b/lib/cluster/cluster.c @@ -1,410 +1,410 @@ /* * Copyright 2004-2022 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU Lesser General Public License * version 2.1 or later (LGPLv2.1+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include "crmcluster_private.h" CRM_TRACE_INIT_DATA(cluster); /*! * \brief Get (and set if needed) a node's UUID * - * \param[in] peer Node to check + * \param[in,out] peer Node to check * * \return Node UUID of \p peer, or NULL if unknown */ const char * crm_peer_uuid(crm_node_t *peer) { char *uuid = NULL; // Check simple cases first, to avoid any calls that might block if (peer == NULL) { return NULL; } if (peer->uuid != NULL) { return peer->uuid; } switch (get_cluster_type()) { case pcmk_cluster_corosync: #if SUPPORT_COROSYNC uuid = pcmk__corosync_uuid(peer); #endif break; case pcmk_cluster_unknown: case pcmk_cluster_invalid: crm_err("Unsupported cluster type"); break; } peer->uuid = uuid; return peer->uuid; } /*! * \brief Connect to the cluster layer * - * \param[in] Initialized cluster object to connect + * \param[in,out] Initialized cluster object to connect * * \return TRUE on success, otherwise FALSE */ gboolean crm_cluster_connect(crm_cluster_t *cluster) { enum cluster_type_e type = get_cluster_type(); crm_notice("Connecting to %s cluster infrastructure", name_for_cluster_type(type)); switch (type) { case pcmk_cluster_corosync: #if SUPPORT_COROSYNC if (is_corosync_cluster()) { crm_peer_init(); return pcmk__corosync_connect(cluster); } #endif break; default: break; } return FALSE; } /*! * \brief Disconnect from the cluster layer * - * \param[in] cluster Cluster object to disconnect + * \param[in,out] cluster Cluster object to disconnect */ void crm_cluster_disconnect(crm_cluster_t *cluster) { enum cluster_type_e type = get_cluster_type(); crm_info("Disconnecting from %s cluster infrastructure", name_for_cluster_type(type)); switch (type) { case pcmk_cluster_corosync: #if SUPPORT_COROSYNC if (is_corosync_cluster()) { crm_peer_destroy(); pcmk__corosync_disconnect(cluster); } #endif break; default: break; } } /*! * \brief Allocate a new \p crm_cluster_t object * * \return A newly allocated \p crm_cluster_t object (guaranteed not \p NULL) * \note The caller is responsible for freeing the return value using * \p pcmk_cluster_free(). */ crm_cluster_t * pcmk_cluster_new(void) { crm_cluster_t *cluster = calloc(1, sizeof(crm_cluster_t)); CRM_ASSERT(cluster != NULL); return cluster; } /*! * \brief Free a \p crm_cluster_t object and its dynamically allocated members * * \param[in,out] cluster Cluster object to free */ void pcmk_cluster_free(crm_cluster_t *cluster) { if (cluster == NULL) { return; } free(cluster->uuid); free(cluster->uname); free(cluster); } /*! * \brief Send an XML message via the cluster messaging layer * * \param[in] node Cluster node to send message to * \param[in] service Message type to use in message host info * \param[in] data XML message to send * \param[in] ordered Ignored for currently supported messaging layers * * \return TRUE on success, otherwise FALSE */ gboolean send_cluster_message(const crm_node_t *node, enum crm_ais_msg_types service, xmlNode *data, gboolean ordered) { switch (get_cluster_type()) { case pcmk_cluster_corosync: #if SUPPORT_COROSYNC return pcmk__cpg_send_xml(data, node, service); #endif break; default: break; } return FALSE; } /*! * \brief Get the local node's name * * \return Local node's name * \note This will fatally exit if local node name cannot be known. */ const char * get_local_node_name(void) { static char *name = NULL; if (name == NULL) { name = get_node_name(0); } return name; } /*! * \brief Get the node name corresponding to a cluster node ID * * \param[in] nodeid Node ID to check (or 0 for local node) * * \return Node name corresponding to \p nodeid * \note This will fatally exit if \p nodeid is 0 and local node name cannot be * known. */ char * get_node_name(uint32_t nodeid) { char *name = NULL; enum cluster_type_e stack = get_cluster_type(); switch (stack) { # if SUPPORT_COROSYNC case pcmk_cluster_corosync: name = pcmk__corosync_name(0, nodeid); break; # endif default: crm_err("Unknown cluster type: %s (%d)", name_for_cluster_type(stack), stack); } if ((name == NULL) && (nodeid == 0)) { name = pcmk_hostname(); if (name == NULL) { // @TODO Maybe let the caller decide what to do crm_err("Could not obtain the local %s node name", name_for_cluster_type(stack)); crm_exit(CRM_EX_FATAL); } crm_notice("Defaulting to uname -n for the local %s node name", name_for_cluster_type(stack)); } if (name == NULL) { crm_notice("Could not obtain a node name for %s node with id %u", name_for_cluster_type(stack), nodeid); } return name; } /*! * \brief Get the node name corresponding to a node UUID * * \param[in] uuid UUID of desired node * * \return name of desired node * * \note This relies on the remote peer cache being populated with all * remote nodes in the cluster, so callers should maintain that cache. */ const char * crm_peer_uname(const char *uuid) { GHashTableIter iter; crm_node_t *node = NULL; CRM_CHECK(uuid != NULL, return NULL); /* remote nodes have the same uname and uuid */ if (g_hash_table_lookup(crm_remote_peer_cache, uuid)) { return uuid; } /* avoid blocking calls where possible */ g_hash_table_iter_init(&iter, crm_peer_cache); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &node)) { if (pcmk__str_eq(node->uuid, uuid, pcmk__str_casei)) { if (node->uname != NULL) { return node->uname; } break; } } node = NULL; #if SUPPORT_COROSYNC if (is_corosync_cluster()) { long long id; if ((pcmk__scan_ll(uuid, &id, 0LL) != pcmk_rc_ok) || (id < 1LL) || (id > UINT32_MAX)) { crm_err("Invalid Corosync node ID '%s'", uuid); return NULL; } node = pcmk__search_cluster_node_cache((uint32_t) id, NULL); if (node != NULL) { crm_info("Setting uuid for node %s[%u] to %s", node->uname, node->id, uuid); node->uuid = strdup(uuid); return node->uname; } return NULL; } #endif return NULL; } /*! * \brief Add a node's UUID as an XML attribute * * \param[in,out] xml XML element to add UUID to * \param[in] attr XML attribute name to set - * \param[in] node Node whose UUID should be used as attribute value + * \param[in,out] node Node whose UUID should be used as attribute value */ void set_uuid(xmlNode *xml, const char *attr, crm_node_t *node) { crm_xml_add(xml, attr, crm_peer_uuid(node)); } /*! * \brief Get a log-friendly string equivalent of a cluster type * * \param[in] type Cluster type * * \return Log-friendly string corresponding to \p type */ const char * name_for_cluster_type(enum cluster_type_e type) { switch (type) { case pcmk_cluster_corosync: return "corosync"; case pcmk_cluster_unknown: return "unknown"; case pcmk_cluster_invalid: return "invalid"; } crm_err("Invalid cluster type: %d", type); return "invalid"; } /*! * \brief Get (and validate) the local cluster type * * \return Local cluster type * \note This will fatally exit if the local cluster type is invalid. */ enum cluster_type_e get_cluster_type(void) { bool detected = false; const char *cluster = NULL; static enum cluster_type_e cluster_type = pcmk_cluster_unknown; /* Return the previous calculation, if any */ if (cluster_type != pcmk_cluster_unknown) { return cluster_type; } cluster = pcmk__env_option(PCMK__ENV_CLUSTER_TYPE); #if SUPPORT_COROSYNC /* If nothing is defined in the environment, try corosync (if supported) */ if (cluster == NULL) { crm_debug("Testing with Corosync"); cluster_type = pcmk__corosync_detect(); if (cluster_type != pcmk_cluster_unknown) { detected = true; goto done; } } #endif /* Something was defined in the environment, test it against what we support */ crm_info("Verifying cluster type: '%s'", ((cluster == NULL)? "-unspecified-" : cluster)); if (cluster == NULL) { #if SUPPORT_COROSYNC } else if (pcmk__str_eq(cluster, "corosync", pcmk__str_casei)) { cluster_type = pcmk_cluster_corosync; #endif } else { cluster_type = pcmk_cluster_invalid; goto done; /* Keep the compiler happy when no stacks are supported */ } done: if (cluster_type == pcmk_cluster_unknown) { crm_notice("Could not determine the current cluster type"); } else if (cluster_type == pcmk_cluster_invalid) { crm_notice("This installation does not support the '%s' cluster infrastructure: terminating.", cluster); crm_exit(CRM_EX_FATAL); } else { crm_info("%s an active '%s' cluster", (detected? "Detected" : "Assuming"), name_for_cluster_type(cluster_type)); } return cluster_type; } /*! * \brief Check whether the local cluster is a Corosync cluster * * \return TRUE if the local cluster is a Corosync cluster, otherwise FALSE */ gboolean is_corosync_cluster(void) { return get_cluster_type() == pcmk_cluster_corosync; } diff --git a/lib/cluster/corosync.c b/lib/cluster/corosync.c index 9fa7bfb8f9..08280cef7a 100644 --- a/lib/cluster/corosync.c +++ b/lib/cluster/corosync.c @@ -1,814 +1,814 @@ /* * Copyright 2004-2022 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU Lesser General Public License * version 2.1 or later (LGPLv2.1+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include #include // PRIu64 #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* PCMK__SPECIAL_PID* */ #include "crmcluster_private.h" static quorum_handle_t pcmk_quorum_handle = 0; static gboolean (*quorum_app_callback)(unsigned long long seq, gboolean quorate) = NULL; /*! * \internal * \brief Get the Corosync UUID associated with a Pacemaker node * * \param[in] node Pacemaker node * * \return Newly allocated string with node's Corosync UUID, or NULL if unknown * \note It is the caller's responsibility to free the result with free(). */ char * -pcmk__corosync_uuid(crm_node_t *node) +pcmk__corosync_uuid(const crm_node_t *node) { if ((node != NULL) && is_corosync_cluster()) { if (node->id > 0) { return crm_strdup_printf("%u", node->id); } else { crm_info("Node %s is not yet known by Corosync", node->uname); } } return NULL; } static bool node_name_is_valid(const char *key, const char *name) { int octet; if (name == NULL) { crm_trace("%s is empty", key); return false; } else if (sscanf(name, "%d.%d.%d.%d", &octet, &octet, &octet, &octet) == 4) { crm_trace("%s contains an IPv4 address (%s), ignoring", key, name); return false; } else if (strstr(name, ":") != NULL) { crm_trace("%s contains an IPv6 address (%s), ignoring", key, name); return false; } crm_trace("'%s: %s' is valid", key, name); return true; } /* * \internal * \brief Get Corosync node name corresponding to a node ID * * \param[in] cmap_handle Connection to Corosync CMAP * \param[in] nodeid Node ID to check * * \return Newly allocated string with name or (if no name) IP address * associated with first address assigned to a Corosync node ID (or NULL * if unknown) * \note It is the caller's responsibility to free the result with free(). */ char * pcmk__corosync_name(uint64_t /*cmap_handle_t */ cmap_handle, uint32_t nodeid) { // Originally based on corosync-quorumtool.c:node_name() int lpc = 0; cs_error_t rc = CS_OK; int retries = 0; char *name = NULL; cmap_handle_t local_handle = 0; int fd = -1; uid_t found_uid = 0; gid_t found_gid = 0; pid_t found_pid = 0; int rv; if (nodeid == 0) { nodeid = get_local_nodeid(0); } if (cmap_handle == 0 && local_handle == 0) { retries = 0; crm_trace("Initializing CMAP connection"); do { rc = pcmk__init_cmap(&local_handle); if (rc != CS_OK) { retries++; crm_debug("API connection setup failed: %s. Retrying in %ds", cs_strerror(rc), retries); sleep(retries); } } while (retries < 5 && rc != CS_OK); if (rc != CS_OK) { crm_warn("Could not connect to Cluster Configuration Database API, error %s", cs_strerror(rc)); local_handle = 0; } } if (cmap_handle == 0) { cmap_handle = local_handle; rc = cmap_fd_get(cmap_handle, &fd); if (rc != CS_OK) { crm_err("Could not obtain the CMAP API connection: %s (%d)", cs_strerror(rc), rc); goto bail; } /* CMAP provider run as root (in given user namespace, anyway)? */ if (!(rv = crm_ipc_is_authentic_process(fd, (uid_t) 0,(gid_t) 0, &found_pid, &found_uid, &found_gid))) { crm_err("CMAP provider is not authentic:" " process %lld (uid: %lld, gid: %lld)", (long long) PCMK__SPECIAL_PID_AS_0(found_pid), (long long) found_uid, (long long) found_gid); goto bail; } else if (rv < 0) { crm_err("Could not verify authenticity of CMAP provider: %s (%d)", strerror(-rv), -rv); goto bail; } } while (name == NULL && cmap_handle != 0) { uint32_t id = 0; char *key = NULL; key = crm_strdup_printf("nodelist.node.%d.nodeid", lpc); rc = cmap_get_uint32(cmap_handle, key, &id); crm_trace("Checking %u vs %u from %s", nodeid, id, key); free(key); if (rc != CS_OK) { break; } if (nodeid == id) { crm_trace("Searching for node name for %u in nodelist.node.%d %s", nodeid, lpc, pcmk__s(name, "")); if (name == NULL) { key = crm_strdup_printf("nodelist.node.%d.name", lpc); cmap_get_string(cmap_handle, key, &name); crm_trace("%s = %s", key, pcmk__s(name, "")); free(key); } if (name == NULL) { key = crm_strdup_printf("nodelist.node.%d.ring0_addr", lpc); cmap_get_string(cmap_handle, key, &name); crm_trace("%s = %s", key, pcmk__s(name, "")); if (!node_name_is_valid(key, name)) { free(name); name = NULL; } free(key); } break; } lpc++; } bail: if(local_handle) { cmap_finalize(local_handle); } if (name == NULL) { crm_info("Unable to get node name for nodeid %u", nodeid); } return name; } /*! * \internal * \brief Disconnect from Corosync cluster * - * \param[in] cluster Cluster connection to disconnect + * \param[in,out] cluster Cluster connection to disconnect */ void pcmk__corosync_disconnect(crm_cluster_t *cluster) { cluster_disconnect_cpg(cluster); if (pcmk_quorum_handle) { quorum_finalize(pcmk_quorum_handle); pcmk_quorum_handle = 0; } crm_notice("Disconnected from Corosync"); } /*! * \internal * \brief Dispatch function for quorum connection file descriptor * * \param[in] user_data Ignored * * \return 0 on success, -1 on error (per mainloop_io_t interface) */ static int quorum_dispatch_cb(gpointer user_data) { int rc = quorum_dispatch(pcmk_quorum_handle, CS_DISPATCH_ALL); if (rc < 0) { crm_err("Connection to the Quorum API failed: %d", rc); quorum_finalize(pcmk_quorum_handle); pcmk_quorum_handle = 0; return -1; } return 0; } /*! * \internal * \brief Notification callback for Corosync quorum connection * * \param[in] handle Corosync quorum connection * \param[in] quorate Whether cluster is quorate * \param[in] ring_id Corosync ring ID * \param[in] view_list_entries Number of entries in \p view_list * \param[in] view_list Corosync node IDs in membership */ static void quorum_notification_cb(quorum_handle_t handle, uint32_t quorate, uint64_t ring_id, uint32_t view_list_entries, uint32_t *view_list) { int i; GHashTableIter iter; crm_node_t *node = NULL; static gboolean init_phase = TRUE; if (quorate != crm_have_quorum) { if (quorate) { crm_notice("Quorum acquired " CRM_XS " membership=%" PRIu64 " members=%lu", ring_id, (long unsigned int)view_list_entries); } else { crm_warn("Quorum lost " CRM_XS " membership=%" PRIu64 " members=%lu", ring_id, (long unsigned int)view_list_entries); } crm_have_quorum = quorate; } else { crm_info("Quorum %s " CRM_XS " membership=%" PRIu64 " members=%lu", (quorate? "retained" : "still lost"), ring_id, (long unsigned int)view_list_entries); } if (view_list_entries == 0 && init_phase) { crm_info("Corosync membership is still forming, ignoring"); return; } init_phase = FALSE; /* Reset last_seen for all cached nodes so we can tell which ones aren't * in the view list */ g_hash_table_iter_init(&iter, crm_peer_cache); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &node)) { node->last_seen = 0; } /* Update the peer cache for each node in view list */ for (i = 0; i < view_list_entries; i++) { uint32_t id = view_list[i]; crm_debug("Member[%d] %u ", i, id); /* Get this node's peer cache entry (adding one if not already there) */ node = crm_get_peer(id, NULL); if (node->uname == NULL) { char *name = pcmk__corosync_name(0, id); crm_info("Obtaining name for new node %u", id); node = crm_get_peer(id, name); free(name); } /* Update the node state (including updating last_seen to ring_id) */ pcmk__update_peer_state(__func__, node, CRM_NODE_MEMBER, ring_id); } /* Remove any peer cache entries we didn't update */ pcmk__reap_unseen_nodes(ring_id); if (quorum_app_callback) { quorum_app_callback(ring_id, quorate); } } /*! * \internal * \brief Connect to Corosync quorum service * * \param[in] dispatch Connection dispatch callback * \param[in] destroy Connection destroy callback */ void pcmk__corosync_quorum_connect(gboolean (*dispatch)(unsigned long long, gboolean), void (*destroy)(gpointer)) { cs_error_t rc; int fd = 0; int quorate = 0; uint32_t quorum_type = 0; struct mainloop_fd_callbacks quorum_fd_callbacks; uid_t found_uid = 0; gid_t found_gid = 0; pid_t found_pid = 0; int rv; quorum_fd_callbacks.dispatch = quorum_dispatch_cb; quorum_fd_callbacks.destroy = destroy; crm_debug("Configuring Pacemaker to obtain quorum from Corosync"); { #if 0 // New way but not supported by all Corosync 2 versions quorum_model_v0_data_t quorum_model_data = { .model = QUORUM_MODEL_V0, .quorum_notify_fn = quorum_notification_cb, }; rc = quorum_model_initialize(&pcmk_quorum_handle, QUORUM_MODEL_V0, (quorum_model_data_t *) &quorum_model_data, &quorum_type, NULL); #else quorum_callbacks_t quorum_callbacks = { .quorum_notify_fn = quorum_notification_cb, }; rc = quorum_initialize(&pcmk_quorum_handle, &quorum_callbacks, &quorum_type); #endif } if (rc != CS_OK) { crm_err("Could not connect to the Quorum API: %s (%d)", cs_strerror(rc), rc); goto bail; } else if (quorum_type != QUORUM_SET) { crm_err("Corosync quorum is not configured"); goto bail; } rc = quorum_fd_get(pcmk_quorum_handle, &fd); if (rc != CS_OK) { crm_err("Could not obtain the Quorum API connection: %s (%d)", strerror(rc), rc); goto bail; } /* Quorum provider run as root (in given user namespace, anyway)? */ if (!(rv = crm_ipc_is_authentic_process(fd, (uid_t) 0,(gid_t) 0, &found_pid, &found_uid, &found_gid))) { crm_err("Quorum provider is not authentic:" " process %lld (uid: %lld, gid: %lld)", (long long) PCMK__SPECIAL_PID_AS_0(found_pid), (long long) found_uid, (long long) found_gid); rc = CS_ERR_ACCESS; goto bail; } else if (rv < 0) { crm_err("Could not verify authenticity of Quorum provider: %s (%d)", strerror(-rv), -rv); rc = CS_ERR_ACCESS; goto bail; } rc = quorum_getquorate(pcmk_quorum_handle, &quorate); if (rc != CS_OK) { crm_err("Could not obtain the current Quorum API state: %d", rc); goto bail; } if (quorate) { crm_notice("Quorum acquired"); } else { crm_warn("No quorum"); } quorum_app_callback = dispatch; crm_have_quorum = quorate; rc = quorum_trackstart(pcmk_quorum_handle, CS_TRACK_CHANGES | CS_TRACK_CURRENT); if (rc != CS_OK) { crm_err("Could not setup Quorum API notifications: %d", rc); goto bail; } mainloop_add_fd("quorum", G_PRIORITY_HIGH, fd, dispatch, &quorum_fd_callbacks); pcmk__corosync_add_nodes(NULL); bail: if (rc != CS_OK) { quorum_finalize(pcmk_quorum_handle); } } /*! * \internal * \brief Connect to Corosync cluster layer * - * \param[in] cluster Initialized cluster object to connect + * \param[in,out] cluster Initialized cluster object to connect */ gboolean pcmk__corosync_connect(crm_cluster_t *cluster) { crm_node_t *peer = NULL; enum cluster_type_e stack = get_cluster_type(); crm_peer_init(); if (stack != pcmk_cluster_corosync) { crm_err("Invalid cluster type: %s " CRM_XS " stack=%d", name_for_cluster_type(stack), stack); return FALSE; } if (!cluster_connect_cpg(cluster)) { // Error message was logged by cluster_connect_cpg() return FALSE; } crm_info("Connection to %s established", name_for_cluster_type(stack)); cluster->nodeid = get_local_nodeid(0); if (cluster->nodeid == 0) { crm_err("Could not determine local node ID"); return FALSE; } cluster->uname = get_node_name(0); if (cluster->uname == NULL) { crm_err("Could not determine local node name"); return FALSE; } // Ensure local node always exists in peer cache peer = crm_get_peer(cluster->nodeid, cluster->uname); cluster->uuid = pcmk__corosync_uuid(peer); return TRUE; } /*! * \internal * \brief Check whether a Corosync cluster is active * * \return pcmk_cluster_corosync if Corosync is found, else pcmk_cluster_unknown */ enum cluster_type_e pcmk__corosync_detect(void) { int rc = CS_OK; cmap_handle_t handle; rc = pcmk__init_cmap(&handle); switch(rc) { case CS_OK: break; case CS_ERR_SECURITY: crm_debug("Failed to initialize the cmap API: Permission denied (%d)", rc); /* It's there, we just can't talk to it. * Good enough for us to identify as 'corosync' */ return pcmk_cluster_corosync; default: crm_info("Failed to initialize the cmap API: %s (%d)", pcmk__cs_err_str(rc), rc); return pcmk_cluster_unknown; } cmap_finalize(handle); return pcmk_cluster_corosync; } /*! * \brief Check whether a Corosync cluster peer is active * * \param[in] node Node to check * * \return TRUE if \p node is an active Corosync peer, otherwise FALSE */ gboolean crm_is_corosync_peer_active(const crm_node_t *node) { if (node == NULL) { crm_trace("Corosync peer inactive: NULL"); return FALSE; } else if (!pcmk__str_eq(node->state, CRM_NODE_MEMBER, pcmk__str_casei)) { crm_trace("Corosync peer %s inactive: state=%s", node->uname, node->state); return FALSE; } else if (!pcmk_is_set(node->processes, crm_proc_cpg)) { crm_trace("Corosync peer %s inactive: processes=%.16x", node->uname, node->processes); return FALSE; } return TRUE; } /*! * \internal * \brief Load Corosync node list (via CMAP) into peer cache and optionally XML * - * \param[in] xml_parent If not NULL, add a entry to this for each node + * \param[in,out] xml_parent If not NULL, add entry here for each node * * \return true if any nodes were found, false otherwise */ bool pcmk__corosync_add_nodes(xmlNode *xml_parent) { int lpc = 0; cs_error_t rc = CS_OK; int retries = 0; bool any = false; cmap_handle_t cmap_handle; int fd = -1; uid_t found_uid = 0; gid_t found_gid = 0; pid_t found_pid = 0; int rv; do { rc = pcmk__init_cmap(&cmap_handle); if (rc != CS_OK) { retries++; crm_debug("API connection setup failed: %s. Retrying in %ds", cs_strerror(rc), retries); sleep(retries); } } while (retries < 5 && rc != CS_OK); if (rc != CS_OK) { crm_warn("Could not connect to Cluster Configuration Database API, error %d", rc); return false; } rc = cmap_fd_get(cmap_handle, &fd); if (rc != CS_OK) { crm_err("Could not obtain the CMAP API connection: %s (%d)", cs_strerror(rc), rc); goto bail; } /* CMAP provider run as root (in given user namespace, anyway)? */ if (!(rv = crm_ipc_is_authentic_process(fd, (uid_t) 0,(gid_t) 0, &found_pid, &found_uid, &found_gid))) { crm_err("CMAP provider is not authentic:" " process %lld (uid: %lld, gid: %lld)", (long long) PCMK__SPECIAL_PID_AS_0(found_pid), (long long) found_uid, (long long) found_gid); goto bail; } else if (rv < 0) { crm_err("Could not verify authenticity of CMAP provider: %s (%d)", strerror(-rv), -rv); goto bail; } crm_peer_init(); crm_trace("Initializing Corosync node list"); for (lpc = 0; TRUE; lpc++) { uint32_t nodeid = 0; char *name = NULL; char *key = NULL; key = crm_strdup_printf("nodelist.node.%d.nodeid", lpc); rc = cmap_get_uint32(cmap_handle, key, &nodeid); free(key); if (rc != CS_OK) { break; } name = pcmk__corosync_name(cmap_handle, nodeid); if (name != NULL) { GHashTableIter iter; crm_node_t *node = NULL; g_hash_table_iter_init(&iter, crm_peer_cache); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &node)) { if(node && node->uname && strcasecmp(node->uname, name) == 0) { if (node->id && node->id != nodeid) { crm_crit("Nodes %u and %u share the same name '%s': shutting down", node->id, nodeid, name); crm_exit(CRM_EX_FATAL); } } } } if (nodeid > 0 || name != NULL) { crm_trace("Initializing node[%d] %u = %s", lpc, nodeid, name); crm_get_peer(nodeid, name); } if (nodeid > 0 && name != NULL) { any = true; if (xml_parent) { xmlNode *node = create_xml_node(xml_parent, XML_CIB_TAG_NODE); crm_xml_set_id(node, "%u", nodeid); crm_xml_add(node, XML_ATTR_UNAME, name); } } free(name); } bail: cmap_finalize(cmap_handle); return any; } /*! * \internal * \brief Get cluster name from Corosync configuration (via CMAP) * * \return Newly allocated string with cluster name if configured, or NULL */ char * pcmk__corosync_cluster_name(void) { cmap_handle_t handle; char *cluster_name = NULL; cs_error_t rc = CS_OK; int fd = -1; uid_t found_uid = 0; gid_t found_gid = 0; pid_t found_pid = 0; int rv; rc = pcmk__init_cmap(&handle); if (rc != CS_OK) { crm_info("Failed to initialize the cmap API: %s (%d)", cs_strerror(rc), rc); return NULL; } rc = cmap_fd_get(handle, &fd); if (rc != CS_OK) { crm_err("Could not obtain the CMAP API connection: %s (%d)", cs_strerror(rc), rc); goto bail; } /* CMAP provider run as root (in given user namespace, anyway)? */ if (!(rv = crm_ipc_is_authentic_process(fd, (uid_t) 0,(gid_t) 0, &found_pid, &found_uid, &found_gid))) { crm_err("CMAP provider is not authentic:" " process %lld (uid: %lld, gid: %lld)", (long long) PCMK__SPECIAL_PID_AS_0(found_pid), (long long) found_uid, (long long) found_gid); goto bail; } else if (rv < 0) { crm_err("Could not verify authenticity of CMAP provider: %s (%d)", strerror(-rv), -rv); goto bail; } rc = cmap_get_string(handle, "totem.cluster_name", &cluster_name); if (rc != CS_OK) { crm_info("Cannot get totem.cluster_name: %s (%d)", cs_strerror(rc), rc); } else { crm_debug("cmap totem.cluster_name = '%s'", cluster_name); } bail: cmap_finalize(handle); return cluster_name; } /*! * \internal * \brief Check (via CMAP) whether Corosync configuration has a node list * * \return true if Corosync has node list, otherwise false */ bool pcmk__corosync_has_nodelist(void) { cs_error_t cs_rc = CS_OK; int retries = 0; cmap_handle_t cmap_handle; cmap_iter_handle_t iter_handle; char key_name[CMAP_KEYNAME_MAXLEN + 1]; int fd = -1; uid_t found_uid = 0; gid_t found_gid = 0; pid_t found_pid = 0; int rc = pcmk_ok; static bool got_result = false; static bool result = false; if (got_result) { return result; } // Connect to CMAP do { cs_rc = pcmk__init_cmap(&cmap_handle); if (cs_rc != CS_OK) { retries++; crm_debug("CMAP connection failed: %s (rc=%d, retrying in %ds)", cs_strerror(cs_rc), cs_rc, retries); sleep(retries); } } while ((retries < 5) && (cs_rc != CS_OK)); if (cs_rc != CS_OK) { crm_warn("Assuming Corosync does not have node list: " "CMAP connection failed (%s) " CRM_XS " rc=%d", cs_strerror(cs_rc), cs_rc); return false; } // Get CMAP connection file descriptor cs_rc = cmap_fd_get(cmap_handle, &fd); if (cs_rc != CS_OK) { crm_warn("Assuming Corosync does not have node list: " "CMAP unusable (%s) " CRM_XS " rc=%d", cs_strerror(cs_rc), cs_rc); goto bail; } // Check whether CMAP connection is authentic (i.e. provided by root) rc = crm_ipc_is_authentic_process(fd, (uid_t) 0, (gid_t) 0, &found_pid, &found_uid, &found_gid); if (rc == 0) { crm_warn("Assuming Corosync does not have node list: " "CMAP provider is inauthentic " CRM_XS " pid=%lld uid=%lld gid=%lld", (long long) PCMK__SPECIAL_PID_AS_0(found_pid), (long long) found_uid, (long long) found_gid); goto bail; } else if (rc < 0) { crm_warn("Assuming Corosync does not have node list: " "Could not verify CMAP authenticity (%s) " CRM_XS " rc=%d", pcmk_strerror(rc), rc); goto bail; } // Check whether nodelist section is presetn cs_rc = cmap_iter_init(cmap_handle, "nodelist", &iter_handle); if (cs_rc != CS_OK) { crm_warn("Assuming Corosync does not have node list: " "CMAP not readable (%s) " CRM_XS " rc=%d", cs_strerror(cs_rc), cs_rc); goto bail; } cs_rc = cmap_iter_next(cmap_handle, iter_handle, key_name, NULL, NULL); if (cs_rc == CS_OK) { result = true; } cmap_iter_finalize(cmap_handle, iter_handle); got_result = true; crm_debug("Corosync %s node list", (result? "has" : "does not have")); bail: cmap_finalize(cmap_handle); return result; } diff --git a/lib/cluster/cpg.c b/lib/cluster/cpg.c index 9de7a3af28..2af4a5059d 100644 --- a/lib/cluster/cpg.c +++ b/lib/cluster/cpg.c @@ -1,1091 +1,1092 @@ /* * Copyright 2004-2022 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU Lesser General Public License * version 2.1 or later (LGPLv2.1+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* PCMK__SPECIAL_PID* */ #include "crmcluster_private.h" /* @TODO Once we can update the public API to require crm_cluster_t* in more * functions, we can ditch this in favor of cluster->cpg_handle. */ static cpg_handle_t pcmk_cpg_handle = 0; // @TODO These could be moved to crm_cluster_t* at that time as well static bool cpg_evicted = false; static GList *cs_message_queue = NULL; static int cs_message_timer = 0; struct pcmk__cpg_host_s { uint32_t id; uint32_t pid; gboolean local; enum crm_ais_msg_types type; uint32_t size; char uname[MAX_NAME]; } __attribute__ ((packed)); typedef struct pcmk__cpg_host_s pcmk__cpg_host_t; struct pcmk__cpg_msg_s { struct qb_ipc_response_header header __attribute__ ((aligned(8))); uint32_t id; gboolean is_compressed; pcmk__cpg_host_t host; pcmk__cpg_host_t sender; uint32_t size; uint32_t compressed_size; /* 584 bytes */ char data[0]; } __attribute__ ((packed)); typedef struct pcmk__cpg_msg_s pcmk__cpg_msg_t; static void crm_cs_flush(gpointer data); #define msg_data_len(msg) (msg->is_compressed?msg->compressed_size:msg->size) #define cs_repeat(rc, counter, max, code) do { \ rc = code; \ if ((rc == CS_ERR_TRY_AGAIN) || (rc == CS_ERR_QUEUE_FULL)) { \ counter++; \ crm_debug("Retrying operation after %ds", counter); \ sleep(counter); \ } else { \ break; \ } \ } while (counter < max) /*! * \brief Disconnect from Corosync CPG * - * \param[in] Cluster to disconnect + * \param[in,out] cluster Cluster to disconnect */ void cluster_disconnect_cpg(crm_cluster_t *cluster) { pcmk_cpg_handle = 0; if (cluster->cpg_handle) { crm_trace("Disconnecting CPG"); cpg_leave(cluster->cpg_handle, &cluster->group); cpg_finalize(cluster->cpg_handle); cluster->cpg_handle = 0; } else { crm_info("No CPG connection"); } } /*! * \brief Get the local Corosync node ID (via CPG) * * \param[in] handle CPG connection to use (or 0 to use new connection) * * \return Corosync ID of local node (or 0 if not known) */ uint32_t get_local_nodeid(cpg_handle_t handle) { cs_error_t rc = CS_OK; int retries = 0; static uint32_t local_nodeid = 0; cpg_handle_t local_handle = handle; cpg_model_v1_data_t cpg_model_info = {CPG_MODEL_V1, NULL, NULL, NULL, 0}; int fd = -1; uid_t found_uid = 0; gid_t found_gid = 0; pid_t found_pid = 0; int rv; if(local_nodeid != 0) { return local_nodeid; } if(handle == 0) { crm_trace("Creating connection"); cs_repeat(rc, retries, 5, cpg_model_initialize(&local_handle, CPG_MODEL_V1, (cpg_model_data_t *)&cpg_model_info, NULL)); if (rc != CS_OK) { crm_err("Could not connect to the CPG API: %s (%d)", cs_strerror(rc), rc); return 0; } rc = cpg_fd_get(local_handle, &fd); if (rc != CS_OK) { crm_err("Could not obtain the CPG API connection: %s (%d)", cs_strerror(rc), rc); goto bail; } /* CPG provider run as root (in given user namespace, anyway)? */ if (!(rv = crm_ipc_is_authentic_process(fd, (uid_t) 0,(gid_t) 0, &found_pid, &found_uid, &found_gid))) { crm_err("CPG provider is not authentic:" " process %lld (uid: %lld, gid: %lld)", (long long) PCMK__SPECIAL_PID_AS_0(found_pid), (long long) found_uid, (long long) found_gid); goto bail; } else if (rv < 0) { crm_err("Could not verify authenticity of CPG provider: %s (%d)", strerror(-rv), -rv); goto bail; } } if (rc == CS_OK) { retries = 0; crm_trace("Performing lookup"); cs_repeat(rc, retries, 5, cpg_local_get(local_handle, &local_nodeid)); } if (rc != CS_OK) { crm_err("Could not get local node id from the CPG API: %s (%d)", pcmk__cs_err_str(rc), rc); } bail: if(handle == 0) { crm_trace("Closing connection"); cpg_finalize(local_handle); } crm_debug("Local nodeid is %u", local_nodeid); return local_nodeid; } /*! * \internal * \brief Callback function for Corosync message queue timer * * \param[in] data CPG handle * * \return FALSE (to indicate to glib that timer should not be removed) */ static gboolean crm_cs_flush_cb(gpointer data) { cs_message_timer = 0; crm_cs_flush(data); return FALSE; } // Send no more than this many CPG messages in one flush #define CS_SEND_MAX 200 /*! * \internal * \brief Send messages in Corosync CPG message queue * * \param[in] data CPG handle */ static void crm_cs_flush(gpointer data) { unsigned int sent = 0; guint queue_len = 0; cs_error_t rc = 0; cpg_handle_t *handle = (cpg_handle_t *) data; if (*handle == 0) { crm_trace("Connection is dead"); return; } queue_len = g_list_length(cs_message_queue); if (((queue_len % 1000) == 0) && (queue_len > 1)) { crm_err("CPG queue has grown to %d", queue_len); } else if (queue_len == CS_SEND_MAX) { crm_warn("CPG queue has grown to %d", queue_len); } if (cs_message_timer != 0) { /* There is already a timer, wait until it goes off */ crm_trace("Timer active %d", cs_message_timer); return; } while ((cs_message_queue != NULL) && (sent < CS_SEND_MAX)) { struct iovec *iov = cs_message_queue->data; rc = cpg_mcast_joined(*handle, CPG_TYPE_AGREED, iov, 1); if (rc != CS_OK) { break; } sent++; crm_trace("CPG message sent, size=%llu", (unsigned long long) iov->iov_len); cs_message_queue = g_list_remove(cs_message_queue, iov); free(iov->iov_base); free(iov); } queue_len -= sent; do_crm_log((queue_len > 5)? LOG_INFO : LOG_TRACE, "Sent %u CPG message%s (%d still queued): %s (rc=%d)", sent, pcmk__plural_s(sent), queue_len, pcmk__cs_err_str(rc), (int) rc); if (cs_message_queue) { uint32_t delay_ms = 100; if (rc != CS_OK) { /* Proportionally more if sending failed but cap at 1s */ delay_ms = QB_MIN(1000, CS_SEND_MAX + (10 * queue_len)); } cs_message_timer = g_timeout_add(delay_ms, crm_cs_flush_cb, data); } } /*! * \internal * \brief Dispatch function for CPG handle * - * \param[in] user_data Cluster object + * \param[in,out] user_data Cluster object * * \return 0 on success, -1 on error (per mainloop_io_t interface) */ static int pcmk_cpg_dispatch(gpointer user_data) { cs_error_t rc = CS_OK; crm_cluster_t *cluster = (crm_cluster_t *) user_data; rc = cpg_dispatch(cluster->cpg_handle, CS_DISPATCH_ONE); if (rc != CS_OK) { crm_err("Connection to the CPG API failed: %s (%d)", pcmk__cs_err_str(rc), rc); cpg_finalize(cluster->cpg_handle); cluster->cpg_handle = 0; return -1; } else if (cpg_evicted) { crm_err("Evicted from CPG membership"); return -1; } return 0; } static inline const char * ais_dest(const pcmk__cpg_host_t *host) { if (host->local) { return "local"; } else if (host->size > 0) { return host->uname; } else { return ""; } } static inline const char * msg_type2text(enum crm_ais_msg_types type) { const char *text = "unknown"; switch (type) { case crm_msg_none: text = "unknown"; break; case crm_msg_ais: text = "ais"; break; case crm_msg_cib: text = "cib"; break; case crm_msg_crmd: text = "crmd"; break; case crm_msg_pe: text = "pengine"; break; case crm_msg_te: text = "tengine"; break; case crm_msg_lrmd: text = "lrmd"; break; case crm_msg_attrd: text = "attrd"; break; case crm_msg_stonithd: text = "stonithd"; break; case crm_msg_stonith_ng: text = "stonith-ng"; break; } return text; } /*! * \internal * \brief Check whether a Corosync CPG message is valid * * \param[in] msg Corosync CPG message to check * * \return true if \p msg is valid, otherwise false */ static bool check_message_sanity(const pcmk__cpg_msg_t *msg) { int32_t payload_size = msg->header.size - sizeof(pcmk__cpg_msg_t); if (payload_size < 1) { crm_err("%sCPG message %d from %s invalid: " "Claimed size of %d bytes is too small " CRM_XS " from %s[%u] to %s@%s", (msg->is_compressed? "Compressed " : ""), msg->id, ais_dest(&(msg->sender)), (int) msg->header.size, msg_type2text(msg->sender.type), msg->sender.pid, msg_type2text(msg->host.type), ais_dest(&(msg->host))); return false; } if (msg->header.error != CS_OK) { crm_err("%sCPG message %d from %s invalid: " "Sender indicated error %d " CRM_XS " from %s[%u] to %s@%s", (msg->is_compressed? "Compressed " : ""), msg->id, ais_dest(&(msg->sender)), msg->header.error, msg_type2text(msg->sender.type), msg->sender.pid, msg_type2text(msg->host.type), ais_dest(&(msg->host))); return false; } if (msg_data_len(msg) != payload_size) { crm_err("%sCPG message %d from %s invalid: " "Total size %d inconsistent with payload size %d " CRM_XS " from %s[%u] to %s@%s", (msg->is_compressed? "Compressed " : ""), msg->id, ais_dest(&(msg->sender)), (int) msg->header.size, (int) msg_data_len(msg), msg_type2text(msg->sender.type), msg->sender.pid, msg_type2text(msg->host.type), ais_dest(&(msg->host))); return false; } if (!msg->is_compressed && /* msg->size != (strlen(msg->data) + 1) would be a stronger check, * but checking the last byte or two should be quick */ (((msg->size > 1) && (msg->data[msg->size - 2] == '\0')) || (msg->data[msg->size - 1] != '\0'))) { crm_err("CPG message %d from %s invalid: " "Payload does not end at byte %llu " CRM_XS " from %s[%u] to %s@%s", msg->id, ais_dest(&(msg->sender)), (unsigned long long) msg->size, msg_type2text(msg->sender.type), msg->sender.pid, msg_type2text(msg->host.type), ais_dest(&(msg->host))); return false; } crm_trace("Verified %d-byte %sCPG message %d from %s[%u]@%s to %s@%s", (int) msg->header.size, (msg->is_compressed? "compressed " : ""), msg->id, msg_type2text(msg->sender.type), msg->sender.pid, ais_dest(&(msg->sender)), msg_type2text(msg->host.type), ais_dest(&(msg->host))); return true; } /*! * \brief Extract text data from a Corosync CPG message * - * \param[in] handle CPG connection (to get local node ID if not yet known) - * \param[in] nodeid Corosync ID of node that sent message - * \param[in] pid Process ID of message sender (for logging only) - * \param[in] content CPG message - * \param[out] kind If not NULL, will be set to CPG header ID - * (which should be an enum crm_ais_msg_class value, - * currently always crm_class_cluster) - * \param[out] from If not NULL, will be set to sender uname - * (valid for the lifetime of \p content) + * \param[in] handle CPG connection (to get local node ID if not known) + * \param[in] nodeid Corosync ID of node that sent message + * \param[in] pid Process ID of message sender (for logging only) + * \param[in,out] content CPG message + * \param[out] kind If not NULL, will be set to CPG header ID + * (which should be an enum crm_ais_msg_class value, + * currently always crm_class_cluster) + * \param[out] from If not NULL, will be set to sender uname + * (valid for the lifetime of \p content) * * \return Newly allocated string with message data * \note It is the caller's responsibility to free the return value with free(). */ char * pcmk_message_common_cs(cpg_handle_t handle, uint32_t nodeid, uint32_t pid, void *content, uint32_t *kind, const char **from) { char *data = NULL; pcmk__cpg_msg_t *msg = (pcmk__cpg_msg_t *) content; if(handle) { // Do filtering and field massaging uint32_t local_nodeid = get_local_nodeid(handle); const char *local_name = get_local_node_name(); if (msg->sender.id > 0 && msg->sender.id != nodeid) { crm_err("Nodeid mismatch from %d.%d: claimed nodeid=%u", nodeid, pid, msg->sender.id); return NULL; } else if (msg->host.id != 0 && (local_nodeid != msg->host.id)) { /* Not for us */ crm_trace("Not for us: %u != %u", msg->host.id, local_nodeid); return NULL; } else if (msg->host.size != 0 && !pcmk__str_eq(msg->host.uname, local_name, pcmk__str_casei)) { /* Not for us */ crm_trace("Not for us: %s != %s", msg->host.uname, local_name); return NULL; } msg->sender.id = nodeid; if (msg->sender.size == 0) { crm_node_t *peer = crm_get_peer(nodeid, NULL); if (peer == NULL) { crm_err("Peer with nodeid=%u is unknown", nodeid); } else if (peer->uname == NULL) { crm_err("No uname for peer with nodeid=%u", nodeid); } else { crm_notice("Fixing uname for peer with nodeid=%u", nodeid); msg->sender.size = strlen(peer->uname); memset(msg->sender.uname, 0, MAX_NAME); memcpy(msg->sender.uname, peer->uname, msg->sender.size); } } } crm_trace("Got new%s message (size=%d, %d, %d)", msg->is_compressed ? " compressed" : "", msg_data_len(msg), msg->size, msg->compressed_size); if (kind != NULL) { *kind = msg->header.id; } if (from != NULL) { *from = msg->sender.uname; } if (msg->is_compressed && msg->size > 0) { int rc = BZ_OK; char *uncompressed = NULL; unsigned int new_size = msg->size + 1; if (!check_message_sanity(msg)) { goto badmsg; } crm_trace("Decompressing message data"); uncompressed = calloc(1, new_size); rc = BZ2_bzBuffToBuffDecompress(uncompressed, &new_size, msg->data, msg->compressed_size, 1, 0); if (rc != BZ_OK) { crm_err("Decompression failed: %s " CRM_XS " bzerror=%d", bz2_strerror(rc), rc); free(uncompressed); goto badmsg; } CRM_ASSERT(rc == BZ_OK); CRM_ASSERT(new_size == msg->size); data = uncompressed; } else if (!check_message_sanity(msg)) { goto badmsg; } else { data = strdup(msg->data); } // Is this necessary? crm_get_peer(msg->sender.id, msg->sender.uname); crm_trace("Payload: %.200s", data); return data; badmsg: crm_err("Invalid message (id=%d, dest=%s:%s, from=%s:%s.%d):" " min=%d, total=%d, size=%d, bz2_size=%d", msg->id, ais_dest(&(msg->host)), msg_type2text(msg->host.type), ais_dest(&(msg->sender)), msg_type2text(msg->sender.type), msg->sender.pid, (int)sizeof(pcmk__cpg_msg_t), msg->header.size, msg->size, msg->compressed_size); free(data); return NULL; } /*! * \internal * \brief Compare cpg_address objects by node ID * * \param[in] first First cpg_address structure to compare * \param[in] second Second cpg_address structure to compare * * \return Negative number if first's node ID is lower, * positive number if first's node ID is greater, * or 0 if both node IDs are equal */ static int cmp_member_list_nodeid(const void *first, const void *second) { const struct cpg_address *const a = *((const struct cpg_address **) first), *const b = *((const struct cpg_address **) second); if (a->nodeid < b->nodeid) { return -1; } else if (a->nodeid > b->nodeid) { return 1; } /* don't bother with "reason" nor "pid" */ return 0; } /*! * \internal * \brief Get a readable string equivalent of a cpg_reason_t value * * \param[in] reason CPG reason value * * \return Readable string suitable for logging */ static const char * cpgreason2str(cpg_reason_t reason) { switch (reason) { case CPG_REASON_JOIN: return " via cpg_join"; case CPG_REASON_LEAVE: return " via cpg_leave"; case CPG_REASON_NODEDOWN: return " via cluster exit"; case CPG_REASON_NODEUP: return " via cluster join"; case CPG_REASON_PROCDOWN: return " for unknown reason"; default: break; } return ""; } /*! * \internal * \brief Get a log-friendly node name * * \param[in] peer Node to check * * \return Node's uname, or readable string if not known */ static inline const char * -peer_name(crm_node_t *peer) +peer_name(const crm_node_t *peer) { if (peer == NULL) { return "unknown node"; } else if (peer->uname == NULL) { return "peer node"; } else { return peer->uname; } } /*! * \internal * \brief Process a CPG peer's leaving the cluster * * \param[in] cpg_group_name CPG group name (for logging) * \param[in] event_counter Event number (for logging) * \param[in] local_nodeid Node ID of local node * \param[in] cpg_peer CPG peer that left * \param[in] sorted_member_list List of remaining members, qsort()-ed by ID * \param[in] member_list_entries Number of entries in \p sorted_member_list */ static void node_left(const char *cpg_group_name, int event_counter, uint32_t local_nodeid, const struct cpg_address *cpg_peer, const struct cpg_address **sorted_member_list, size_t member_list_entries) { crm_node_t *peer = pcmk__search_cluster_node_cache(cpg_peer->nodeid, NULL); const struct cpg_address **rival = NULL; /* Most CPG-related Pacemaker code assumes that only one process on a node * can be in the process group, but Corosync does not impose this * limitation, and more than one can be a member in practice due to a * daemon attempting to start while another instance is already running. * * Check for any such duplicate instances, because we don't want to process * their leaving as if our actual peer left. If the peer that left still has * an entry in sorted_member_list (with a different PID), we will ignore the * leaving. * * @TODO Track CPG members' PIDs so we can tell exactly who left. */ if (peer != NULL) { rival = bsearch(&cpg_peer, sorted_member_list, member_list_entries, sizeof(const struct cpg_address *), cmp_member_list_nodeid); } if (rival == NULL) { crm_info("Group %s event %d: %s (node %u pid %u) left%s", cpg_group_name, event_counter, peer_name(peer), cpg_peer->nodeid, cpg_peer->pid, cpgreason2str(cpg_peer->reason)); if (peer != NULL) { crm_update_peer_proc(__func__, peer, crm_proc_cpg, OFFLINESTATUS); } } else if (cpg_peer->nodeid == local_nodeid) { crm_warn("Group %s event %d: duplicate local pid %u left%s", cpg_group_name, event_counter, cpg_peer->pid, cpgreason2str(cpg_peer->reason)); } else { crm_warn("Group %s event %d: " "%s (node %u) duplicate pid %u left%s (%u remains)", cpg_group_name, event_counter, peer_name(peer), cpg_peer->nodeid, cpg_peer->pid, cpgreason2str(cpg_peer->reason), (*rival)->pid); } } /*! * \brief Handle a CPG configuration change event * * \param[in] handle CPG connection * \param[in] cpg_name CPG group name * \param[in] member_list List of current CPG members * \param[in] member_list_entries Number of entries in \p member_list * \param[in] left_list List of CPG members that left * \param[in] left_list_entries Number of entries in \p left_list * \param[in] joined_list List of CPG members that joined * \param[in] joined_list_entries Number of entries in \p joined_list */ void pcmk_cpg_membership(cpg_handle_t handle, const struct cpg_name *groupName, const struct cpg_address *member_list, size_t member_list_entries, const struct cpg_address *left_list, size_t left_list_entries, const struct cpg_address *joined_list, size_t joined_list_entries) { int i; gboolean found = FALSE; static int counter = 0; uint32_t local_nodeid = get_local_nodeid(handle); const struct cpg_address **sorted; sorted = malloc(member_list_entries * sizeof(const struct cpg_address *)); CRM_ASSERT(sorted != NULL); for (size_t iter = 0; iter < member_list_entries; iter++) { sorted[iter] = member_list + iter; } /* so that the cross-matching multiply-subscribed nodes is then cheap */ qsort(sorted, member_list_entries, sizeof(const struct cpg_address *), cmp_member_list_nodeid); for (i = 0; i < left_list_entries; i++) { node_left(groupName->value, counter, local_nodeid, &left_list[i], sorted, member_list_entries); } free(sorted); sorted = NULL; for (i = 0; i < joined_list_entries; i++) { crm_info("Group %s event %d: node %u pid %u joined%s", groupName->value, counter, joined_list[i].nodeid, joined_list[i].pid, cpgreason2str(joined_list[i].reason)); } for (i = 0; i < member_list_entries; i++) { crm_node_t *peer = crm_get_peer(member_list[i].nodeid, NULL); if (member_list[i].nodeid == local_nodeid && member_list[i].pid != getpid()) { // See the note in node_left() crm_warn("Group %s event %d: detected duplicate local pid %u", groupName->value, counter, member_list[i].pid); continue; } crm_info("Group %s event %d: %s (node %u pid %u) is member", groupName->value, counter, peer_name(peer), member_list[i].nodeid, member_list[i].pid); /* If the caller left auto-reaping enabled, this will also update the * state to member. */ peer = crm_update_peer_proc(__func__, peer, crm_proc_cpg, ONLINESTATUS); if (peer && peer->state && strcmp(peer->state, CRM_NODE_MEMBER)) { /* The node is a CPG member, but we currently think it's not a * cluster member. This is possible only if auto-reaping was * disabled. The node may be joining, and we happened to get the CPG * notification before the quorum notification; or the node may have * just died, and we are processing its final messages; or a bug * has affected the peer cache. */ time_t now = time(NULL); if (peer->when_lost == 0) { // Track when we first got into this contradictory state peer->when_lost = now; } else if (now > (peer->when_lost + 60)) { // If it persists for more than a minute, update the state crm_warn("Node %u is member of group %s but was believed offline", member_list[i].nodeid, groupName->value); pcmk__update_peer_state(__func__, peer, CRM_NODE_MEMBER, 0); } } if (local_nodeid == member_list[i].nodeid) { found = TRUE; } } if (!found) { crm_err("Local node was evicted from group %s", groupName->value); cpg_evicted = true; } counter++; } /*! * \brief Connect to Corosync CPG * - * \param[in] cluster Cluster object + * \param[in,out] cluster Cluster object * * \return TRUE on success, otherwise FALSE */ gboolean cluster_connect_cpg(crm_cluster_t *cluster) { cs_error_t rc; int fd = -1; int retries = 0; uint32_t id = 0; crm_node_t *peer = NULL; cpg_handle_t handle = 0; const char *message_name = pcmk__message_name(crm_system_name); uid_t found_uid = 0; gid_t found_gid = 0; pid_t found_pid = 0; int rv; struct mainloop_fd_callbacks cpg_fd_callbacks = { .dispatch = pcmk_cpg_dispatch, .destroy = cluster->destroy, }; cpg_model_v1_data_t cpg_model_info = { .model = CPG_MODEL_V1, .cpg_deliver_fn = cluster->cpg.cpg_deliver_fn, .cpg_confchg_fn = cluster->cpg.cpg_confchg_fn, .cpg_totem_confchg_fn = NULL, .flags = 0, }; cpg_evicted = false; cluster->group.length = 0; cluster->group.value[0] = 0; /* group.value is char[128] */ strncpy(cluster->group.value, message_name, 127); cluster->group.value[127] = 0; cluster->group.length = 1 + QB_MIN(127, strlen(cluster->group.value)); cs_repeat(rc, retries, 30, cpg_model_initialize(&handle, CPG_MODEL_V1, (cpg_model_data_t *)&cpg_model_info, NULL)); if (rc != CS_OK) { crm_err("Could not connect to the CPG API: %s (%d)", cs_strerror(rc), rc); goto bail; } rc = cpg_fd_get(handle, &fd); if (rc != CS_OK) { crm_err("Could not obtain the CPG API connection: %s (%d)", cs_strerror(rc), rc); goto bail; } /* CPG provider run as root (in given user namespace, anyway)? */ if (!(rv = crm_ipc_is_authentic_process(fd, (uid_t) 0,(gid_t) 0, &found_pid, &found_uid, &found_gid))) { crm_err("CPG provider is not authentic:" " process %lld (uid: %lld, gid: %lld)", (long long) PCMK__SPECIAL_PID_AS_0(found_pid), (long long) found_uid, (long long) found_gid); rc = CS_ERR_ACCESS; goto bail; } else if (rv < 0) { crm_err("Could not verify authenticity of CPG provider: %s (%d)", strerror(-rv), -rv); rc = CS_ERR_ACCESS; goto bail; } id = get_local_nodeid(handle); if (id == 0) { crm_err("Could not get local node id from the CPG API"); goto bail; } cluster->nodeid = id; retries = 0; cs_repeat(rc, retries, 30, cpg_join(handle, &cluster->group)); if (rc != CS_OK) { crm_err("Could not join the CPG group '%s': %d", message_name, rc); goto bail; } pcmk_cpg_handle = handle; cluster->cpg_handle = handle; mainloop_add_fd("corosync-cpg", G_PRIORITY_MEDIUM, fd, cluster, &cpg_fd_callbacks); bail: if (rc != CS_OK) { cpg_finalize(handle); return FALSE; } peer = crm_get_peer(id, NULL); crm_update_peer_proc(__func__, peer, crm_proc_cpg, ONLINESTATUS); return TRUE; } /*! * \internal * \brief Send an XML message via Corosync CPG * * \param[in] msg XML message to send * \param[in] node Cluster node to send message to * \param[in] dest Type of message to send * * \return TRUE on success, otherwise FALSE */ gboolean -pcmk__cpg_send_xml(xmlNode *msg, crm_node_t *node, enum crm_ais_msg_types dest) +pcmk__cpg_send_xml(xmlNode *msg, const crm_node_t *node, + enum crm_ais_msg_types dest) { gboolean rc = TRUE; char *data = NULL; data = dump_xml_unformatted(msg); rc = send_cluster_text(crm_class_cluster, data, FALSE, node, dest); free(data); return rc; } /*! * \internal * \brief Send string data via Corosync CPG * * \param[in] msg_class Message class (to set as CPG header ID) * \param[in] data Data to send * \param[in] local What to set as host "local" value (which is never used) * \param[in] node Cluster node to send message to * \param[in] dest Type of message to send * * \return TRUE on success, otherwise FALSE */ gboolean send_cluster_text(enum crm_ais_msg_class msg_class, const char *data, gboolean local, const crm_node_t *node, enum crm_ais_msg_types dest) { static int msg_id = 0; static int local_pid = 0; static int local_name_len = 0; static const char *local_name = NULL; char *target = NULL; struct iovec *iov; pcmk__cpg_msg_t *msg = NULL; enum crm_ais_msg_types sender = text2msg_type(crm_system_name); switch (msg_class) { case crm_class_cluster: break; default: crm_err("Invalid message class: %d", msg_class); return FALSE; } CRM_CHECK(dest != crm_msg_ais, return FALSE); if (local_name == NULL) { local_name = get_local_node_name(); } if ((local_name_len == 0) && (local_name != NULL)) { local_name_len = strlen(local_name); } if (data == NULL) { data = ""; } if (local_pid == 0) { local_pid = getpid(); } if (sender == crm_msg_none) { sender = local_pid; } msg = calloc(1, sizeof(pcmk__cpg_msg_t)); msg_id++; msg->id = msg_id; msg->header.id = msg_class; msg->header.error = CS_OK; msg->host.type = dest; msg->host.local = local; if (node) { if (node->uname) { target = strdup(node->uname); msg->host.size = strlen(node->uname); memset(msg->host.uname, 0, MAX_NAME); memcpy(msg->host.uname, node->uname, msg->host.size); } else { target = crm_strdup_printf("%u", node->id); } msg->host.id = node->id; } else { target = strdup("all"); } msg->sender.id = 0; msg->sender.type = sender; msg->sender.pid = local_pid; msg->sender.size = local_name_len; memset(msg->sender.uname, 0, MAX_NAME); if ((local_name != NULL) && (msg->sender.size != 0)) { memcpy(msg->sender.uname, local_name, msg->sender.size); } msg->size = 1 + strlen(data); msg->header.size = sizeof(pcmk__cpg_msg_t) + msg->size; if (msg->size < CRM_BZ2_THRESHOLD) { msg = pcmk__realloc(msg, msg->header.size); memcpy(msg->data, data, msg->size); } else { char *compressed = NULL; unsigned int new_size = 0; char *uncompressed = strdup(data); if (pcmk__compress(uncompressed, (unsigned int) msg->size, 0, &compressed, &new_size) == pcmk_rc_ok) { msg->header.size = sizeof(pcmk__cpg_msg_t) + new_size; msg = pcmk__realloc(msg, msg->header.size); memcpy(msg->data, compressed, new_size); msg->is_compressed = TRUE; msg->compressed_size = new_size; } else { // cppcheck seems not to understand the abort logic in pcmk__realloc // cppcheck-suppress memleak msg = pcmk__realloc(msg, msg->header.size); memcpy(msg->data, data, msg->size); } free(uncompressed); free(compressed); } iov = calloc(1, sizeof(struct iovec)); iov->iov_base = msg; iov->iov_len = msg->header.size; if (msg->compressed_size) { crm_trace("Queueing CPG message %u to %s (%llu bytes, %d bytes compressed payload): %.200s", msg->id, target, (unsigned long long) iov->iov_len, msg->compressed_size, data); } else { crm_trace("Queueing CPG message %u to %s (%llu bytes, %d bytes payload): %.200s", msg->id, target, (unsigned long long) iov->iov_len, msg->size, data); } free(target); cs_message_queue = g_list_append(cs_message_queue, iov); crm_cs_flush(&pcmk_cpg_handle); return TRUE; } /*! * \brief Get the message type equivalent of a string * * \param[in] text String of message type * * \return Message type equivalent of \p text */ enum crm_ais_msg_types text2msg_type(const char *text) { int type = crm_msg_none; CRM_CHECK(text != NULL, return type); text = pcmk__message_name(text); if (pcmk__str_eq(text, "ais", pcmk__str_casei)) { type = crm_msg_ais; } else if (pcmk__str_eq(text, CRM_SYSTEM_CIB, pcmk__str_casei)) { type = crm_msg_cib; } else if (pcmk__strcase_any_of(text, CRM_SYSTEM_CRMD, CRM_SYSTEM_DC, NULL)) { type = crm_msg_crmd; } else if (pcmk__str_eq(text, CRM_SYSTEM_TENGINE, pcmk__str_casei)) { type = crm_msg_te; } else if (pcmk__str_eq(text, CRM_SYSTEM_PENGINE, pcmk__str_casei)) { type = crm_msg_pe; } else if (pcmk__str_eq(text, CRM_SYSTEM_LRMD, pcmk__str_casei)) { type = crm_msg_lrmd; } else if (pcmk__str_eq(text, CRM_SYSTEM_STONITHD, pcmk__str_casei)) { type = crm_msg_stonithd; } else if (pcmk__str_eq(text, "stonith-ng", pcmk__str_casei)) { type = crm_msg_stonith_ng; } else if (pcmk__str_eq(text, "attrd", pcmk__str_casei)) { type = crm_msg_attrd; } else { /* This will normally be a transient client rather than * a cluster daemon. Set the type to the pid of the client */ int scan_rc = sscanf(text, "%d", &type); if (scan_rc != 1 || type <= crm_msg_stonith_ng) { /* Ensure it's sane */ type = crm_msg_none; } } return type; } diff --git a/lib/cluster/crmcluster_private.h b/lib/cluster/crmcluster_private.h index bb5ea7a586..6933b7304c 100644 --- a/lib/cluster/crmcluster_private.h +++ b/lib/cluster/crmcluster_private.h @@ -1,47 +1,47 @@ /* - * Copyright 2020 the Pacemaker project contributors + * Copyright 2020-2022 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__CRMCLUSTER_PRIVATE__H # define PCMK__CRMCLUSTER_PRIVATE__H /* This header is for the sole use of libcrmcluster, so that functions can be * declared with G_GNUC_INTERNAL for efficiency. */ #include // uint32_t, uint64_t #include // G_GNUC_INTERNAL, gboolean #include // xmlNode #include // cluster_type_e, crm_node_t G_GNUC_INTERNAL enum cluster_type_e pcmk__corosync_detect(void); G_GNUC_INTERNAL bool pcmk__corosync_has_nodelist(void); G_GNUC_INTERNAL -char *pcmk__corosync_uuid(crm_node_t *peer); +char *pcmk__corosync_uuid(const crm_node_t *peer); G_GNUC_INTERNAL char *pcmk__corosync_name(uint64_t /*cmap_handle_t */ cmap_handle, uint32_t nodeid); G_GNUC_INTERNAL gboolean pcmk__corosync_connect(crm_cluster_t *cluster); G_GNUC_INTERNAL void pcmk__corosync_disconnect(crm_cluster_t *cluster); G_GNUC_INTERNAL -gboolean pcmk__cpg_send_xml(xmlNode *msg, crm_node_t *node, +gboolean pcmk__cpg_send_xml(xmlNode *msg, const crm_node_t *node, enum crm_ais_msg_types dest); #endif // PCMK__CRMCLUSTER_PRIVATE__H diff --git a/lib/cluster/election.c b/lib/cluster/election.c index 42eefb007b..ebbae720ae 100644 --- a/lib/cluster/election.c +++ b/lib/cluster/election.c @@ -1,725 +1,727 @@ /* - * Copyright 2004-2021 the Pacemaker project contributors + * Copyright 2004-2022 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU Lesser General Public License * version 2.1 or later (LGPLv2.1+) WITHOUT ANY WARRANTY. */ #include #include #include #include #include #include #include #include #include #define STORM_INTERVAL 2 /* in seconds */ struct election_s { enum election_result state; guint count; // How many times local node has voted char *name; // Descriptive name for this election char *uname; // Local node's name GSourceFunc cb; // Function to call if election is won GHashTable *voted; // Key = node name, value = how node voted mainloop_timer_t *timeout; // When to abort if all votes not received int election_wins; // Track wins, for storm detection bool wrote_blackbox; // Write a storm blackbox at most once time_t expires; // When storm detection period ends time_t last_election_loss; // When dampening period ends }; static void election_complete(election_t *e) { e->state = election_won; if (e->cb != NULL) { e->cb(e); } election_reset(e); } static gboolean election_timer_cb(gpointer user_data) { election_t *e = user_data; crm_info("%s timed out, declaring local node as winner", e->name); election_complete(e); return FALSE; } /*! * \brief Get current state of an election * * \param[in] e Election object * * \return Current state of \e */ enum election_result -election_state(election_t *e) +election_state(const election_t *e) { return (e == NULL)? election_error : e->state; } /*! * \brief Create a new election object * * Every node that wishes to participate in an election must create an election * object. Typically, this should be done once, at start-up. A caller should * only create a single election object. * * \param[in] name Label for election (for logging) * \param[in] uname Local node's name * \param[in] period_ms How long to wait for all peers to vote * \param[in] cb Function to call if local node wins election * * \return Newly allocated election object on success, NULL on error * \note The caller is responsible for freeing the returned value using * election_fini(). */ election_t * election_init(const char *name, const char *uname, guint period_ms, GSourceFunc cb) { election_t *e = NULL; static guint count = 0; CRM_CHECK(uname != NULL, return NULL); e = calloc(1, sizeof(election_t)); if (e == NULL) { crm_perror(LOG_CRIT, "Cannot create election"); return NULL; } e->uname = strdup(uname); if (e->uname == NULL) { crm_perror(LOG_CRIT, "Cannot create election"); free(e); return NULL; } e->name = name? crm_strdup_printf("election-%s", name) : crm_strdup_printf("election-%u", count++); e->cb = cb; e->timeout = mainloop_timer_add(e->name, period_ms, FALSE, election_timer_cb, e); crm_trace("Created %s", e->name); return e; } /*! * \brief Disregard any previous vote by specified peer * * This discards any recorded vote from a specified peer. Election users should * call this whenever a voting peer becomes inactive. * - * \param[in] e Election object - * \param[in] uname Name of peer to disregard + * \param[in,out] e Election object + * \param[in] uname Name of peer to disregard */ void election_remove(election_t *e, const char *uname) { if ((e != NULL) && (uname != NULL) && (e->voted != NULL)) { crm_trace("Discarding %s (no-)vote from lost peer %s", e->name, uname); g_hash_table_remove(e->voted, uname); } } /*! * \brief Stop election timer and disregard all votes * - * \param[in] e Election object + * \param[in,out] e Election object */ void election_reset(election_t *e) { if (e != NULL) { crm_trace("Resetting election %s", e->name); mainloop_timer_stop(e->timeout); if (e->voted) { crm_trace("Destroying voted cache with %d members", g_hash_table_size(e->voted)); g_hash_table_destroy(e->voted); e->voted = NULL; } } } /*! * \brief Free an election object * * Free all memory associated with an election object, stopping its * election timer (if running). * - * \param[in] e Election object + * \param[in,out] e Election object */ void election_fini(election_t *e) { if (e != NULL) { election_reset(e); crm_trace("Destroying %s", e->name); mainloop_timer_del(e->timeout); free(e->uname); free(e->name); free(e); } } static void election_timeout_start(election_t *e) { if (e != NULL) { mainloop_timer_start(e->timeout); } } /*! * \brief Stop an election's timer, if running * - * \param[in] e Election object + * \param[in,out] e Election object */ void election_timeout_stop(election_t *e) { if (e != NULL) { mainloop_timer_stop(e->timeout); } } /*! * \brief Change an election's timeout (restarting timer if running) * - * \param[in] e Election object - * \param[in] period New timeout + * \param[in,out] e Election object + * \param[in] period New timeout */ void election_timeout_set_period(election_t *e, guint period) { if (e != NULL) { mainloop_timer_set_period(e->timeout, period); } else { crm_err("No election defined"); } } static int get_uptime(struct timeval *output) { static time_t expires = 0; static struct rusage info; time_t tm_now = time(NULL); if (expires < tm_now) { int rc = 0; info.ru_utime.tv_sec = 0; info.ru_utime.tv_usec = 0; rc = getrusage(RUSAGE_SELF, &info); output->tv_sec = 0; output->tv_usec = 0; if (rc < 0) { crm_perror(LOG_ERR, "Could not calculate the current uptime"); expires = 0; return -1; } crm_debug("Current CPU usage is: %lds, %ldus", (long)info.ru_utime.tv_sec, (long)info.ru_utime.tv_usec); } expires = tm_now + STORM_INTERVAL; /* N seconds after the last _access_ */ output->tv_sec = info.ru_utime.tv_sec; output->tv_usec = info.ru_utime.tv_usec; return 1; } static int compare_age(struct timeval your_age) { struct timeval our_age; get_uptime(&our_age); /* If an error occurred, our_age will be compared as {0,0} */ if (our_age.tv_sec > your_age.tv_sec) { crm_debug("Win: %ld vs %ld (seconds)", (long)our_age.tv_sec, (long)your_age.tv_sec); return 1; } else if (our_age.tv_sec < your_age.tv_sec) { crm_debug("Lose: %ld vs %ld (seconds)", (long)our_age.tv_sec, (long)your_age.tv_sec); return -1; } else if (our_age.tv_usec > your_age.tv_usec) { crm_debug("Win: %ld.%06ld vs %ld.%06ld (usec)", (long)our_age.tv_sec, (long)our_age.tv_usec, (long)your_age.tv_sec, (long)your_age.tv_usec); return 1; } else if (our_age.tv_usec < your_age.tv_usec) { crm_debug("Lose: %ld.%06ld vs %ld.%06ld (usec)", (long)our_age.tv_sec, (long)our_age.tv_usec, (long)your_age.tv_sec, (long)your_age.tv_usec); return -1; } return 0; } /*! * \brief Start a new election by offering local node's candidacy * * Broadcast a "vote" election message containing the local node's ID, * (incremented) election counter, and uptime, and start the election timer. * - * \param[in] e Election object + * \param[in,out] e Election object + * * \note Any nodes agreeing to the candidacy will send a "no-vote" reply, and if * all active peers do so, or if the election times out, the local node * wins the election. (If we lose to any peer vote, we will stop the * timer, so a timeout means we did not lose -- either some peer did not * vote, or we did not call election_check() in time.) */ void election_vote(election_t *e) { struct timeval age; xmlNode *vote = NULL; crm_node_t *our_node; if (e == NULL) { crm_trace("Election vote requested, but no election available"); return; } our_node = crm_get_peer(0, e->uname); if ((our_node == NULL) || (crm_is_peer_active(our_node) == FALSE)) { crm_trace("Cannot vote in %s yet: local node not connected to cluster", e->name); return; } election_reset(e); e->state = election_in_progress; vote = create_request(CRM_OP_VOTE, NULL, NULL, CRM_SYSTEM_CRMD, CRM_SYSTEM_CRMD, NULL); e->count++; crm_xml_add(vote, F_CRM_ELECTION_OWNER, our_node->uuid); crm_xml_add_int(vote, F_CRM_ELECTION_ID, e->count); get_uptime(&age); crm_xml_add_timeval(vote, F_CRM_ELECTION_AGE_S, F_CRM_ELECTION_AGE_US, &age); send_cluster_message(NULL, crm_msg_crmd, vote, TRUE); free_xml(vote); crm_debug("Started %s round %d", e->name, e->count); election_timeout_start(e); return; } /*! * \brief Check whether local node has won an election * * If all known peers have sent no-vote messages, stop the election timer, set * the election state to won, and call any registered win callback. * - * \param[in] e Election object + * \param[in,out] e Election object * * \return TRUE if local node has won, FALSE otherwise * \note If all known peers have sent no-vote messages, but the election owner * does not call this function, the election will not be won (and the * callback will not be called) until the election times out. * \note This should be called when election_count_vote() returns * \c election_in_progress. */ bool election_check(election_t *e) { int voted_size = 0; int num_members = 0; if (e == NULL) { crm_trace("Election check requested, but no election available"); return FALSE; } if (e->voted == NULL) { crm_trace("%s check requested, but no votes received yet", e->name); return FALSE; } voted_size = g_hash_table_size(e->voted); num_members = crm_active_peers(); /* in the case of #voted > #members, it is better to * wait for the timeout and give the cluster time to * stabilize */ if (voted_size >= num_members) { /* we won and everyone has voted */ election_timeout_stop(e); if (voted_size > num_members) { GHashTableIter gIter; const crm_node_t *node; char *key = NULL; crm_warn("Received too many votes in %s", e->name); g_hash_table_iter_init(&gIter, crm_peer_cache); while (g_hash_table_iter_next(&gIter, NULL, (gpointer *) & node)) { if (crm_is_peer_active(node)) { crm_warn("* expected vote: %s", node->uname); } } g_hash_table_iter_init(&gIter, e->voted); while (g_hash_table_iter_next(&gIter, (gpointer *) & key, NULL)) { crm_warn("* actual vote: %s", key); } } crm_info("%s won by local node", e->name); election_complete(e); return TRUE; } else { crm_debug("%s still waiting on %d of %d votes", e->name, num_members - voted_size, num_members); } return FALSE; } #define LOSS_DAMPEN 2 /* in seconds */ struct vote { const char *op; const char *from; const char *version; const char *election_owner; int election_id; struct timeval age; }; /*! * \brief Unpack an election message * - * \param[in] e Election object + * \param[in] e Election object (for logging only) * \param[in] message Election message XML * \param[out] vote Parsed fields from message * * \return TRUE if election message and election are valid, FALSE otherwise * \note The parsed struct's pointer members are valid only for the lifetime of * the message argument. */ static bool -parse_election_message(election_t *e, xmlNode *message, struct vote *vote) +parse_election_message(const election_t *e, const xmlNode *message, + struct vote *vote) { CRM_CHECK(message && vote, return FALSE); vote->election_id = -1; vote->age.tv_sec = -1; vote->age.tv_usec = -1; vote->op = crm_element_value(message, F_CRM_TASK); vote->from = crm_element_value(message, F_CRM_HOST_FROM); vote->version = crm_element_value(message, F_CRM_VERSION); vote->election_owner = crm_element_value(message, F_CRM_ELECTION_OWNER); crm_element_value_int(message, F_CRM_ELECTION_ID, &(vote->election_id)); if ((vote->op == NULL) || (vote->from == NULL) || (vote->version == NULL) || (vote->election_owner == NULL) || (vote->election_id < 0)) { crm_warn("Invalid %s message from %s in %s ", (vote->op? vote->op : "election"), (vote->from? vote->from : "unspecified node"), (e? e->name : "election")); return FALSE; } // Op-specific validation if (pcmk__str_eq(vote->op, CRM_OP_VOTE, pcmk__str_none)) { // Only vote ops have uptime crm_element_value_timeval(message, F_CRM_ELECTION_AGE_S, F_CRM_ELECTION_AGE_US, &(vote->age)); if ((vote->age.tv_sec < 0) || (vote->age.tv_usec < 0)) { crm_warn("Cannot count %s %s from %s because it is missing uptime", (e? e->name : "election"), vote->op, vote->from); return FALSE; } } else if (!pcmk__str_eq(vote->op, CRM_OP_NOVOTE, pcmk__str_none)) { crm_info("Cannot process %s message from %s because %s is not a known election op", (e? e->name : "election"), vote->from, vote->op); return FALSE; } // Election validation if (e == NULL) { crm_info("Cannot count %s from %s because no election available", vote->op, vote->from); return FALSE; } /* If the membership cache is NULL, we REALLY shouldn't be voting -- * the question is how we managed to get here. */ if (crm_peer_cache == NULL) { crm_info("Cannot count %s %s from %s because no peer information available", e->name, vote->op, vote->from); return FALSE; } return TRUE; } static void record_vote(election_t *e, struct vote *vote) { char *voter_copy = NULL; char *vote_copy = NULL; CRM_ASSERT(e && vote && vote->from && vote->op); if (e->voted == NULL) { e->voted = pcmk__strkey_table(free, free); } voter_copy = strdup(vote->from); vote_copy = strdup(vote->op); CRM_ASSERT(voter_copy && vote_copy); g_hash_table_replace(e->voted, voter_copy, vote_copy); } static void send_no_vote(crm_node_t *peer, struct vote *vote) { // @TODO probably shouldn't hardcode CRM_SYSTEM_CRMD and crm_msg_crmd xmlNode *novote = create_request(CRM_OP_NOVOTE, NULL, vote->from, CRM_SYSTEM_CRMD, CRM_SYSTEM_CRMD, NULL); crm_xml_add(novote, F_CRM_ELECTION_OWNER, vote->election_owner); crm_xml_add_int(novote, F_CRM_ELECTION_ID, vote->election_id); send_cluster_message(peer, crm_msg_crmd, novote, TRUE); free_xml(novote); } /*! * \brief Process an election message (vote or no-vote) from a peer * - * \param[in] e Election object - * \param[in] vote Election message XML from peer - * \param[in] can_win Whether to consider the local node eligible for winning + * \param[in,out] e Election object + * \param[in] message Election message XML from peer + * \param[in] can_win Whether local node is eligible to win * * \return Election state after new vote is considered * \note If the peer message is a vote, and we prefer the peer to win, this will * send a no-vote reply to the peer. * \note The situations "we lost to this vote" from "this is a late no-vote * after we've already lost" both return election_lost. If a caller needs * to distinguish them, it should save the current state before calling * this function, and then compare the result. */ enum election_result -election_count_vote(election_t *e, xmlNode *message, bool can_win) +election_count_vote(election_t *e, const xmlNode *message, bool can_win) { int log_level = LOG_INFO; gboolean done = FALSE; gboolean we_lose = FALSE; const char *reason = "unknown"; bool we_are_owner = FALSE; crm_node_t *our_node = NULL, *your_node = NULL; time_t tm_now = time(NULL); struct vote vote; CRM_CHECK(message != NULL, return election_error); if (parse_election_message(e, message, &vote) == FALSE) { return election_error; } your_node = crm_get_peer(0, vote.from); our_node = crm_get_peer(0, e->uname); we_are_owner = (our_node != NULL) && pcmk__str_eq(our_node->uuid, vote.election_owner, pcmk__str_none); if (!can_win) { reason = "Not eligible"; we_lose = TRUE; } else if (our_node == NULL || crm_is_peer_active(our_node) == FALSE) { reason = "We are not part of the cluster"; log_level = LOG_ERR; we_lose = TRUE; } else if (we_are_owner && (vote.election_id != e->count)) { log_level = LOG_TRACE; reason = "Superseded"; done = TRUE; } else if (your_node == NULL || crm_is_peer_active(your_node) == FALSE) { /* Possibly we cached the message in the FSA queue at a point that it wasn't */ reason = "Peer is not part of our cluster"; log_level = LOG_WARNING; done = TRUE; } else if (pcmk__str_eq(vote.op, CRM_OP_NOVOTE, pcmk__str_none) || pcmk__str_eq(vote.from, e->uname, pcmk__str_none)) { /* Receiving our own broadcast vote, or a no-vote from peer, is a vote * for us to win */ if (!we_are_owner) { crm_warn("Cannot count %s round %d %s from %s because we are not election owner (%s)", e->name, vote.election_id, vote.op, vote.from, vote.election_owner); return election_error; } if (e->state != election_in_progress) { // Should only happen if we already lost crm_debug("Not counting %s round %d %s from %s because no election in progress", e->name, vote.election_id, vote.op, vote.from); return e->state; } record_vote(e, &vote); reason = "Recorded"; done = TRUE; } else { // A peer vote requires a comparison to determine which node is better int age_result = compare_age(vote.age); int version_result = compare_version(vote.version, CRM_FEATURE_SET); if (version_result < 0) { reason = "Version"; we_lose = TRUE; } else if (version_result > 0) { reason = "Version"; } else if (age_result < 0) { reason = "Uptime"; we_lose = TRUE; } else if (age_result > 0) { reason = "Uptime"; } else if (strcasecmp(e->uname, vote.from) > 0) { reason = "Host name"; we_lose = TRUE; } else { reason = "Host name"; } } if (e->expires < tm_now) { e->election_wins = 0; e->expires = tm_now + STORM_INTERVAL; } else if (done == FALSE && we_lose == FALSE) { int peers = 1 + g_hash_table_size(crm_peer_cache); /* If every node has to vote down every other node, thats N*(N-1) total elections * Allow some leeway before _really_ complaining */ e->election_wins++; if (e->election_wins > (peers * peers)) { crm_warn("%s election storm detected: %d wins in %d seconds", e->name, e->election_wins, STORM_INTERVAL); e->election_wins = 0; e->expires = tm_now + STORM_INTERVAL; if (e->wrote_blackbox == FALSE) { /* It's questionable whether a black box (from every node in the * cluster) would be truly helpful in diagnosing an election * storm. It's also highly doubtful a production environment * would get multiple election storms from distinct causes, so * saving one blackbox per process lifetime should be * sufficient. Alternatives would be to save a timestamp of the * last blackbox write instead of a boolean, and write a new one * if some amount of time has passed; or to save a storm count, * write a blackbox on every Nth occurrence. */ crm_write_blackbox(0, NULL); e->wrote_blackbox = TRUE; } } } if (done) { do_crm_log(log_level + 1, "Processed %s round %d %s (current round %d) from %s (%s)", e->name, vote.election_id, vote.op, e->count, vote.from, reason); return e->state; } else if (we_lose == FALSE) { /* We track the time of the last election loss to implement an election * dampening period, reducing the likelihood of an election storm. If * this node has lost within the dampening period, don't start a new * election, even if we win against a peer's vote -- the peer we lost to * should win again. * * @TODO This has a problem case: if an election winner immediately * leaves the cluster, and a new election is immediately called, all * nodes could lose, with no new winner elected. The ideal solution * would be to tie the election structure with the peer caches, which * would allow us to clear the dampening when the previous winner * leaves (and would allow other improvements as well). */ if ((e->last_election_loss == 0) || ((tm_now - e->last_election_loss) > (time_t) LOSS_DAMPEN)) { do_crm_log(log_level, "%s round %d (owner node ID %s) pass: %s from %s (%s)", e->name, vote.election_id, vote.election_owner, vote.op, vote.from, reason); e->last_election_loss = 0; election_timeout_stop(e); /* Start a new election by voting down this, and other, peers */ e->state = election_start; return e->state; } else { char *loss_time = ctime(&e->last_election_loss); if (loss_time) { // Show only HH:MM:SS loss_time += 11; loss_time[8] = '\0'; } crm_info("Ignoring %s round %d (owner node ID %s) pass vs %s because we lost less than %ds ago at %s", e->name, vote.election_id, vote.election_owner, vote.from, LOSS_DAMPEN, (loss_time? loss_time : "unknown")); } } e->last_election_loss = tm_now; do_crm_log(log_level, "%s round %d (owner node ID %s) lost: %s from %s (%s)", e->name, vote.election_id, vote.election_owner, vote.op, vote.from, reason); election_reset(e); send_no_vote(your_node, &vote); e->state = election_lost; return e->state; } /*! * \brief Reset any election dampening currently in effect * - * \param[in] e Election object to clear + * \param[in,out] e Election object to clear */ void election_clear_dampening(election_t *e) { e->last_election_loss = 0; } diff --git a/lib/cluster/membership.c b/lib/cluster/membership.c index 1c7565532b..c93e5a3b78 100644 --- a/lib/cluster/membership.c +++ b/lib/cluster/membership.c @@ -1,1300 +1,1300 @@ /* * Copyright 2004-2022 the Pacemaker project contributors * * The version control history for this file may have further details. * * This source code is licensed under the GNU Lesser General Public License * version 2.1 or later (LGPLv2.1+) WITHOUT ANY WARRANTY. */ #include #ifndef _GNU_SOURCE # define _GNU_SOURCE #endif #include #include #include #include #include #include #include #include #include #include #include #include "crmcluster_private.h" /* The peer cache remembers cluster nodes that have been seen. * This is managed mostly automatically by libcluster, based on * cluster membership events. * * Because cluster nodes can have conflicting names or UUIDs, * the hash table key is a uniquely generated ID. */ GHashTable *crm_peer_cache = NULL; /* * The remote peer cache tracks pacemaker_remote nodes. While the * value has the same type as the peer cache's, it is tracked separately for * three reasons: pacemaker_remote nodes can't have conflicting names or UUIDs, * so the name (which is also the UUID) is used as the hash table key; there * is no equivalent of membership events, so management is not automatic; and * most users of the peer cache need to exclude pacemaker_remote nodes. * * That said, using a single cache would be more logical and less error-prone, * so it would be a good idea to merge them one day. * * libcluster provides two avenues for populating the cache: * crm_remote_peer_get() and crm_remote_peer_cache_remove() directly manage it, * while crm_remote_peer_cache_refresh() populates it via the CIB. */ GHashTable *crm_remote_peer_cache = NULL; /* * The known node cache tracks cluster and remote nodes that have been seen in * the CIB. It is useful mainly when a caller needs to know about a node that * may no longer be in the membership, but doesn't want to add the node to the * main peer cache tables. */ static GHashTable *known_node_cache = NULL; unsigned long long crm_peer_seq = 0; gboolean crm_have_quorum = FALSE; static gboolean crm_autoreap = TRUE; // Flag setting and clearing for crm_node_t:flags #define set_peer_flags(peer, flags_to_set) do { \ (peer)->flags = pcmk__set_flags_as(__func__, __LINE__, LOG_TRACE, \ "Peer", (peer)->uname, \ (peer)->flags, (flags_to_set), \ #flags_to_set); \ } while (0) #define clear_peer_flags(peer, flags_to_clear) do { \ (peer)->flags = pcmk__clear_flags_as(__func__, __LINE__, \ LOG_TRACE, \ "Peer", (peer)->uname, \ (peer)->flags, (flags_to_clear), \ #flags_to_clear); \ } while (0) static void update_peer_uname(crm_node_t *node, const char *uname); int crm_remote_peer_cache_size(void) { if (crm_remote_peer_cache == NULL) { return 0; } return g_hash_table_size(crm_remote_peer_cache); } /*! * \brief Get a remote node peer cache entry, creating it if necessary * * \param[in] node_name Name of remote node * * \return Cache entry for node on success, NULL (and set errno) otherwise * * \note When creating a new entry, this will leave the node state undetermined, * so the caller should also call pcmk__update_peer_state() if the state * is known. */ crm_node_t * crm_remote_peer_get(const char *node_name) { crm_node_t *node; if (node_name == NULL) { errno = -EINVAL; return NULL; } /* Return existing cache entry if one exists */ node = g_hash_table_lookup(crm_remote_peer_cache, node_name); if (node) { return node; } /* Allocate a new entry */ node = calloc(1, sizeof(crm_node_t)); if (node == NULL) { return NULL; } /* Populate the essential information */ set_peer_flags(node, crm_remote_node); node->uuid = strdup(node_name); if (node->uuid == NULL) { free(node); errno = -ENOMEM; return NULL; } /* Add the new entry to the cache */ g_hash_table_replace(crm_remote_peer_cache, node->uuid, node); crm_trace("added %s to remote cache", node_name); /* Update the entry's uname, ensuring peer status callbacks are called */ update_peer_uname(node, node_name); return node; } void crm_remote_peer_cache_remove(const char *node_name) { if (g_hash_table_remove(crm_remote_peer_cache, node_name)) { crm_trace("removed %s from remote peer cache", node_name); } } /*! * \internal * \brief Return node status based on a CIB status entry * * \param[in] node_state XML of node state * * \return CRM_NODE_LOST if XML_NODE_IN_CLUSTER is false in node_state, * CRM_NODE_MEMBER otherwise * \note Unlike most boolean XML attributes, this one defaults to true, for * backward compatibility with older controllers that don't set it. */ static const char * -remote_state_from_cib(xmlNode *node_state) +remote_state_from_cib(const xmlNode *node_state) { bool status = false; if (pcmk__xe_get_bool_attr(node_state, XML_NODE_IN_CLUSTER, &status) == pcmk_rc_ok && !status) { return CRM_NODE_LOST; } else { return CRM_NODE_MEMBER; } } /* user data for looping through remote node xpath searches */ struct refresh_data { const char *field; /* XML attribute to check for node name */ gboolean has_state; /* whether to update node state based on XML */ }; /*! * \internal * \brief Process one pacemaker_remote node xpath search result * * \param[in] result XML search result * \param[in] user_data what to look for in the XML */ static void remote_cache_refresh_helper(xmlNode *result, void *user_data) { - struct refresh_data *data = user_data; + const struct refresh_data *data = user_data; const char *remote = crm_element_value(result, data->field); const char *state = NULL; crm_node_t *node; CRM_CHECK(remote != NULL, return); /* Determine node's state, if the result has it */ if (data->has_state) { state = remote_state_from_cib(result); } /* Check whether cache already has entry for node */ node = g_hash_table_lookup(crm_remote_peer_cache, remote); if (node == NULL) { /* Node is not in cache, so add a new entry for it */ node = crm_remote_peer_get(remote); CRM_ASSERT(node); if (state) { pcmk__update_peer_state(__func__, node, state, 0); } } else if (pcmk_is_set(node->flags, crm_node_dirty)) { /* Node is in cache and hasn't been updated already, so mark it clean */ clear_peer_flags(node, crm_node_dirty); if (state) { pcmk__update_peer_state(__func__, node, state, 0); } } } static void mark_dirty(gpointer key, gpointer value, gpointer user_data) { set_peer_flags((crm_node_t *) value, crm_node_dirty); } static gboolean is_dirty(gpointer key, gpointer value, gpointer user_data) { return pcmk_is_set(((crm_node_t*)value)->flags, crm_node_dirty); } /*! * \brief Repopulate the remote peer cache based on CIB XML * * \param[in] xmlNode CIB XML to parse */ void crm_remote_peer_cache_refresh(xmlNode *cib) { struct refresh_data data; crm_peer_init(); /* First, we mark all existing cache entries as dirty, * so that later we can remove any that weren't in the CIB. * We don't empty the cache, because we need to detect changes in state. */ g_hash_table_foreach(crm_remote_peer_cache, mark_dirty, NULL); /* Look for guest nodes and remote nodes in the status section */ data.field = "id"; data.has_state = TRUE; crm_foreach_xpath_result(cib, PCMK__XP_REMOTE_NODE_STATUS, remote_cache_refresh_helper, &data); /* Look for guest nodes and remote nodes in the configuration section, * because they may have just been added and not have a status entry yet. * In that case, the cached node state will be left NULL, so that the * peer status callback isn't called until we're sure the node started * successfully. */ data.field = "value"; data.has_state = FALSE; crm_foreach_xpath_result(cib, PCMK__XP_GUEST_NODE_CONFIG, remote_cache_refresh_helper, &data); data.field = "id"; data.has_state = FALSE; crm_foreach_xpath_result(cib, PCMK__XP_REMOTE_NODE_CONFIG, remote_cache_refresh_helper, &data); /* Remove all old cache entries that weren't seen in the CIB */ g_hash_table_foreach_remove(crm_remote_peer_cache, is_dirty, NULL); } gboolean crm_is_peer_active(const crm_node_t * node) { if(node == NULL) { return FALSE; } if (pcmk_is_set(node->flags, crm_remote_node)) { /* remote nodes are never considered active members. This * guarantees they will never be considered for DC membership.*/ return FALSE; } #if SUPPORT_COROSYNC if (is_corosync_cluster()) { return crm_is_corosync_peer_active(node); } #endif crm_err("Unhandled cluster type: %s", name_for_cluster_type(get_cluster_type())); return FALSE; } static gboolean crm_reap_dead_member(gpointer key, gpointer value, gpointer user_data) { crm_node_t *node = value; crm_node_t *search = user_data; if (search == NULL) { return FALSE; } else if (search->id && node->id != search->id) { return FALSE; } else if (search->id == 0 && !pcmk__str_eq(node->uname, search->uname, pcmk__str_casei)) { return FALSE; } else if (crm_is_peer_active(value) == FALSE) { crm_info("Removing node with name %s and id %u from membership cache", (node->uname? node->uname : "unknown"), node->id); return TRUE; } return FALSE; } /*! * \brief Remove all peer cache entries matching a node ID and/or uname * * \param[in] id ID of node to remove (or 0 to ignore) * \param[in] name Uname of node to remove (or NULL to ignore) * * \return Number of cache entries removed */ guint reap_crm_member(uint32_t id, const char *name) { int matches = 0; crm_node_t search = { 0, }; if (crm_peer_cache == NULL) { crm_trace("Membership cache not initialized, ignoring purge request"); return 0; } search.id = id; pcmk__str_update(&search.uname, name); matches = g_hash_table_foreach_remove(crm_peer_cache, crm_reap_dead_member, &search); if(matches) { crm_notice("Purged %d peer%s with id=%u%s%s from the membership cache", matches, pcmk__plural_s(matches), search.id, (search.uname? " and/or uname=" : ""), (search.uname? search.uname : "")); } else { crm_info("No peers with id=%u%s%s to purge from the membership cache", search.id, (search.uname? " and/or uname=" : ""), (search.uname? search.uname : "")); } free(search.uname); return matches; } static void count_peer(gpointer key, gpointer value, gpointer user_data) { guint *count = user_data; crm_node_t *node = value; if (crm_is_peer_active(node)) { *count = *count + 1; } } guint crm_active_peers(void) { guint count = 0; if (crm_peer_cache) { g_hash_table_foreach(crm_peer_cache, count_peer, &count); } return count; } static void destroy_crm_node(gpointer data) { crm_node_t *node = data; crm_trace("Destroying entry for node %u: %s", node->id, node->uname); free(node->uname); free(node->state); free(node->uuid); free(node->expected); free(node); } void crm_peer_init(void) { if (crm_peer_cache == NULL) { crm_peer_cache = pcmk__strikey_table(free, destroy_crm_node); } if (crm_remote_peer_cache == NULL) { crm_remote_peer_cache = pcmk__strikey_table(NULL, destroy_crm_node); } if (known_node_cache == NULL) { known_node_cache = pcmk__strikey_table(free, destroy_crm_node); } } void crm_peer_destroy(void) { if (crm_peer_cache != NULL) { crm_trace("Destroying peer cache with %d members", g_hash_table_size(crm_peer_cache)); g_hash_table_destroy(crm_peer_cache); crm_peer_cache = NULL; } if (crm_remote_peer_cache != NULL) { crm_trace("Destroying remote peer cache with %d members", g_hash_table_size(crm_remote_peer_cache)); g_hash_table_destroy(crm_remote_peer_cache); crm_remote_peer_cache = NULL; } if (known_node_cache != NULL) { crm_trace("Destroying known node cache with %d members", g_hash_table_size(known_node_cache)); g_hash_table_destroy(known_node_cache); known_node_cache = NULL; } } static void (*peer_status_callback)(enum crm_status_type, crm_node_t *, const void *) = NULL; /*! * \brief Set a client function that will be called after peer status changes * * \param[in] dispatch Pointer to function to use as callback * * \note Previously, client callbacks were responsible for peer cache * management. This is no longer the case, and client callbacks should do * only client-specific handling. Callbacks MUST NOT add or remove entries * in the peer caches. */ void crm_set_status_callback(void (*dispatch) (enum crm_status_type, crm_node_t *, const void *)) { peer_status_callback = dispatch; } /*! * \brief Tell the library whether to automatically reap lost nodes * * If TRUE (the default), calling crm_update_peer_proc() will also update the * peer state to CRM_NODE_MEMBER or CRM_NODE_LOST, and pcmk__update_peer_state() * will reap peers whose state changes to anything other than CRM_NODE_MEMBER. * Callers should leave this enabled unless they plan to manage the cache * separately on their own. * * \param[in] autoreap TRUE to enable automatic reaping, FALSE to disable */ void crm_set_autoreap(gboolean autoreap) { crm_autoreap = autoreap; } static void dump_peer_hash(int level, const char *caller) { GHashTableIter iter; const char *id = NULL; crm_node_t *node = NULL; g_hash_table_iter_init(&iter, crm_peer_cache); while (g_hash_table_iter_next(&iter, (gpointer *) &id, (gpointer *) &node)) { do_crm_log(level, "%s: Node %u/%s = %p - %s", caller, node->id, node->uname, node, id); } } static gboolean hash_find_by_data(gpointer key, gpointer value, gpointer user_data) { return value == user_data; } /*! * \internal * \brief Search caches for a node (cluster or Pacemaker Remote) * * \param[in] id If not 0, cluster node ID to search for * \param[in] uname If not NULL, node name to search for * \param[in] flags Bitmask of enum crm_get_peer_flags * * \return Node cache entry if found, otherwise NULL */ crm_node_t * pcmk__search_node_caches(unsigned int id, const char *uname, uint32_t flags) { crm_node_t *node = NULL; CRM_ASSERT(id > 0 || uname != NULL); crm_peer_init(); if ((uname != NULL) && pcmk_is_set(flags, CRM_GET_PEER_REMOTE)) { node = g_hash_table_lookup(crm_remote_peer_cache, uname); } if ((node == NULL) && pcmk_is_set(flags, CRM_GET_PEER_CLUSTER)) { node = pcmk__search_cluster_node_cache(id, uname); } return node; } /*! * \brief Get a node cache entry (cluster or Pacemaker Remote) * * \param[in] id If not 0, cluster node ID to search for * \param[in] uname If not NULL, node name to search for * \param[in] flags Bitmask of enum crm_get_peer_flags * * \return (Possibly newly created) node cache entry */ crm_node_t * crm_get_peer_full(unsigned int id, const char *uname, int flags) { crm_node_t *node = NULL; CRM_ASSERT(id > 0 || uname != NULL); crm_peer_init(); if (pcmk_is_set(flags, CRM_GET_PEER_REMOTE)) { node = g_hash_table_lookup(crm_remote_peer_cache, uname); } if ((node == NULL) && pcmk_is_set(flags, CRM_GET_PEER_CLUSTER)) { node = crm_get_peer(id, uname); } return node; } /*! * \internal * \brief Search cluster node cache * * \param[in] id If not 0, cluster node ID to search for * \param[in] uname If not NULL, node name to search for * * \return Cluster node cache entry if found, otherwise NULL */ crm_node_t * pcmk__search_cluster_node_cache(unsigned int id, const char *uname) { GHashTableIter iter; crm_node_t *node = NULL; crm_node_t *by_id = NULL; crm_node_t *by_name = NULL; CRM_ASSERT(id > 0 || uname != NULL); crm_peer_init(); if (uname != NULL) { g_hash_table_iter_init(&iter, crm_peer_cache); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &node)) { if(node->uname && strcasecmp(node->uname, uname) == 0) { crm_trace("Name match: %s = %p", node->uname, node); by_name = node; break; } } } if (id > 0) { g_hash_table_iter_init(&iter, crm_peer_cache); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &node)) { if(node->id == id) { crm_trace("ID match: %u = %p", node->id, node); by_id = node; break; } } } node = by_id; /* Good default */ if(by_id == by_name) { /* Nothing to do if they match (both NULL counts) */ crm_trace("Consistent: %p for %u/%s", by_id, id, uname); } else if(by_id == NULL && by_name) { crm_trace("Only one: %p for %u/%s", by_name, id, uname); if(id && by_name->id) { dump_peer_hash(LOG_WARNING, __func__); crm_crit("Node %u and %u share the same name '%s'", id, by_name->id, uname); node = NULL; /* Create a new one */ } else { node = by_name; } } else if(by_name == NULL && by_id) { crm_trace("Only one: %p for %u/%s", by_id, id, uname); if(uname && by_id->uname) { dump_peer_hash(LOG_WARNING, __func__); crm_crit("Node '%s' and '%s' share the same cluster nodeid %u: assuming '%s' is correct", uname, by_id->uname, id, uname); } } else if(uname && by_id->uname) { if(pcmk__str_eq(uname, by_id->uname, pcmk__str_casei)) { crm_notice("Node '%s' has changed its ID from %u to %u", by_id->uname, by_name->id, by_id->id); g_hash_table_foreach_remove(crm_peer_cache, hash_find_by_data, by_name); } else { crm_warn("Node '%s' and '%s' share the same cluster nodeid: %u %s", by_id->uname, by_name->uname, id, uname); dump_peer_hash(LOG_INFO, __func__); crm_abort(__FILE__, __func__, __LINE__, "member weirdness", TRUE, TRUE); } } else if(id && by_name->id) { crm_warn("Node %u and %u share the same name: '%s'", by_id->id, by_name->id, uname); } else { /* Simple merge */ /* Only corosync-based clusters use node IDs. The functions that call * pcmk__update_peer_state() and crm_update_peer_proc() only know * nodeid, so 'by_id' is authoritative when merging. */ dump_peer_hash(LOG_DEBUG, __func__); crm_info("Merging %p into %p", by_name, by_id); g_hash_table_foreach_remove(crm_peer_cache, hash_find_by_data, by_name); } return node; } #if SUPPORT_COROSYNC static guint remove_conflicting_peer(crm_node_t *node) { int matches = 0; GHashTableIter iter; crm_node_t *existing_node = NULL; if (node->id == 0 || node->uname == NULL) { return 0; } if (!pcmk__corosync_has_nodelist()) { return 0; } g_hash_table_iter_init(&iter, crm_peer_cache); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &existing_node)) { if (existing_node->id > 0 && existing_node->id != node->id && existing_node->uname != NULL && strcasecmp(existing_node->uname, node->uname) == 0) { if (crm_is_peer_active(existing_node)) { continue; } crm_warn("Removing cached offline node %u/%s which has conflicting uname with %u", existing_node->id, existing_node->uname, node->id); g_hash_table_iter_remove(&iter); matches++; } } return matches; } #endif /*! * \brief Get a cluster node cache entry * * \param[in] id If not 0, cluster node ID to search for * \param[in] uname If not NULL, node name to search for * * \return (Possibly newly created) cluster node cache entry */ /* coverity[-alloc] Memory is referenced in one or both hashtables */ crm_node_t * crm_get_peer(unsigned int id, const char *uname) { crm_node_t *node = NULL; char *uname_lookup = NULL; CRM_ASSERT(id > 0 || uname != NULL); crm_peer_init(); node = pcmk__search_cluster_node_cache(id, uname); /* if uname wasn't provided, and find_peer did not turn up a uname based on id. * we need to do a lookup of the node name using the id in the cluster membership. */ if ((node == NULL || node->uname == NULL) && (uname == NULL)) { uname_lookup = get_node_name(id); } if (uname_lookup) { uname = uname_lookup; crm_trace("Inferred a name of '%s' for node %u", uname, id); /* try to turn up the node one more time now that we know the uname. */ if (node == NULL) { node = pcmk__search_cluster_node_cache(id, uname); } } if (node == NULL) { char *uniqueid = crm_generate_uuid(); node = calloc(1, sizeof(crm_node_t)); CRM_ASSERT(node); crm_info("Created entry %s/%p for node %s/%u (%d total)", uniqueid, node, uname, id, 1 + g_hash_table_size(crm_peer_cache)); g_hash_table_replace(crm_peer_cache, uniqueid, node); } if(id > 0 && uname && (node->id == 0 || node->uname == NULL)) { crm_info("Node %u is now known as %s", id, uname); } if(id > 0 && node->id == 0) { node->id = id; } if (uname && (node->uname == NULL)) { update_peer_uname(node, uname); } if(node->uuid == NULL) { const char *uuid = crm_peer_uuid(node); if (uuid) { crm_info("Node %u has uuid %s", id, uuid); } else { crm_info("Cannot obtain a UUID for node %u/%s", id, node->uname); } } free(uname_lookup); return node; } /*! * \internal * \brief Update a node's uname * - * \param[in] node Node object to update - * \param[in] uname New name to set + * \param[in,out] node Node object to update + * \param[in] uname New name to set * * \note This function should not be called within a peer cache iteration, * because in some cases it can remove conflicting cache entries, * which would invalidate the iterator. */ static void update_peer_uname(crm_node_t *node, const char *uname) { CRM_CHECK(uname != NULL, crm_err("Bug: can't update node name without name"); return); CRM_CHECK(node != NULL, crm_err("Bug: can't update node name to %s without node", uname); return); if (pcmk__str_eq(uname, node->uname, pcmk__str_casei)) { crm_debug("Node uname '%s' did not change", uname); return; } for (const char *c = uname; *c; ++c) { if ((*c >= 'A') && (*c <= 'Z')) { crm_warn("Node names with capitals are discouraged, consider changing '%s'", uname); break; } } pcmk__str_update(&node->uname, uname); if (peer_status_callback != NULL) { peer_status_callback(crm_status_uname, node, NULL); } #if SUPPORT_COROSYNC if (is_corosync_cluster() && !pcmk_is_set(node->flags, crm_remote_node)) { remove_conflicting_peer(node); } #endif } /*! * \internal * \brief Get log-friendly string equivalent of a process flag * * \param[in] proc Process flag * * \return Log-friendly string equivalent of \p proc */ static inline const char * proc2text(enum crm_proc_flag proc) { const char *text = "unknown"; switch (proc) { case crm_proc_none: text = "none"; break; case crm_proc_based: text = "pacemaker-based"; break; case crm_proc_controld: text = "pacemaker-controld"; break; case crm_proc_schedulerd: text = "pacemaker-schedulerd"; break; case crm_proc_execd: text = "pacemaker-execd"; break; case crm_proc_attrd: text = "pacemaker-attrd"; break; case crm_proc_fenced: text = "pacemaker-fenced"; break; case crm_proc_cpg: text = "corosync-cpg"; break; } return text; } /*! * \internal * \brief Update a node's process information (and potentially state) * - * \param[in] source Caller's function name (for log messages) - * \param[in] node Node object to update - * \param[in] flag Bitmask of new process information - * \param[in] status node status (online, offline, etc.) + * \param[in] source Caller's function name (for log messages) + * \param[in,out] node Node object to update + * \param[in] flag Bitmask of new process information + * \param[in] status node status (online, offline, etc.) * * \return NULL if any node was reaped from peer caches, value of node otherwise * * \note If this function returns NULL, the supplied node object was likely * freed and should not be used again. This function should not be * called within a cache iteration if reaping is possible, otherwise * reaping could invalidate the iterator. */ crm_node_t * crm_update_peer_proc(const char *source, crm_node_t * node, uint32_t flag, const char *status) { uint32_t last = 0; gboolean changed = FALSE; CRM_CHECK(node != NULL, crm_err("%s: Could not set %s to %s for NULL", source, proc2text(flag), status); return NULL); /* Pacemaker doesn't spawn processes on remote nodes */ if (pcmk_is_set(node->flags, crm_remote_node)) { return node; } last = node->processes; if (status == NULL) { node->processes = flag; if (node->processes != last) { changed = TRUE; } } else if (pcmk__str_eq(status, ONLINESTATUS, pcmk__str_casei)) { if ((node->processes & flag) != flag) { node->processes = pcmk__set_flags_as(__func__, __LINE__, LOG_TRACE, "Peer process", node->uname, node->processes, flag, "processes"); changed = TRUE; } } else if (node->processes & flag) { node->processes = pcmk__clear_flags_as(__func__, __LINE__, LOG_TRACE, "Peer process", node->uname, node->processes, flag, "processes"); changed = TRUE; } if (changed) { if (status == NULL && flag <= crm_proc_none) { crm_info("%s: Node %s[%u] - all processes are now offline", source, node->uname, node->id); } else { crm_info("%s: Node %s[%u] - %s is now %s", source, node->uname, node->id, proc2text(flag), status); } /* Call the client callback first, then update the peer state, * in case the node will be reaped */ if (peer_status_callback != NULL) { peer_status_callback(crm_status_processes, node, &last); } /* The client callback shouldn't touch the peer caches, * but as a safety net, bail if the peer cache was destroyed. */ if (crm_peer_cache == NULL) { return NULL; } if (crm_autoreap) { const char *peer_state = NULL; if (pcmk_is_set(node->processes, crm_get_cluster_proc())) { peer_state = CRM_NODE_MEMBER; } else { peer_state = CRM_NODE_LOST; } node = pcmk__update_peer_state(__func__, node, peer_state, 0); } } else { crm_trace("%s: Node %s[%u] - %s is unchanged (%s)", source, node->uname, node->id, proc2text(flag), status); } return node; } /*! * \internal * \brief Update a cluster node cache entry's expected join state * * \param[in] source Caller's function name (for logging) * \param[in,out] node Node to update * \param[in] expected Node's new join state */ void pcmk__update_peer_expected(const char *source, crm_node_t *node, const char *expected) { char *last = NULL; gboolean changed = FALSE; CRM_CHECK(node != NULL, crm_err("%s: Could not set 'expected' to %s", source, expected); return); /* Remote nodes don't participate in joins */ if (pcmk_is_set(node->flags, crm_remote_node)) { return; } last = node->expected; if (expected != NULL && !pcmk__str_eq(node->expected, expected, pcmk__str_casei)) { node->expected = strdup(expected); changed = TRUE; } if (changed) { crm_info("%s: Node %s[%u] - expected state is now %s (was %s)", source, node->uname, node->id, expected, last); free(last); } else { crm_trace("%s: Node %s[%u] - expected state is unchanged (%s)", source, node->uname, node->id, expected); } } /*! * \internal * \brief Update a node's state and membership information * - * \param[in] source Caller's function name (for log messages) - * \param[in] node Node object to update - * \param[in] state Node's new state - * \param[in] membership Node's new membership ID - * \param[in] iter If not NULL, pointer to node's peer cache iterator + * \param[in] source Caller's function name (for log messages) + * \param[in,out] node Node object to update + * \param[in] state Node's new state + * \param[in] membership Node's new membership ID + * \param[in,out] iter If not NULL, pointer to node's peer cache iterator * * \return NULL if any node was reaped, value of node otherwise * * \note If this function returns NULL, the supplied node object was likely * freed and should not be used again. This function may be called from * within a peer cache iteration if the iterator is supplied. */ static crm_node_t * update_peer_state_iter(const char *source, crm_node_t *node, const char *state, uint64_t membership, GHashTableIter *iter) { gboolean is_member; CRM_CHECK(node != NULL, crm_err("Could not set state for unknown host to %s" CRM_XS " source=%s", state, source); return NULL); is_member = pcmk__str_eq(state, CRM_NODE_MEMBER, pcmk__str_casei); if (is_member) { node->when_lost = 0; if (membership) { node->last_seen = membership; } } if (state && !pcmk__str_eq(node->state, state, pcmk__str_casei)) { char *last = node->state; node->state = strdup(state); crm_notice("Node %s state is now %s " CRM_XS " nodeid=%u previous=%s source=%s", node->uname, state, node->id, (last? last : "unknown"), source); if (peer_status_callback != NULL) { peer_status_callback(crm_status_nstate, node, last); } free(last); if (crm_autoreap && !is_member && !pcmk_is_set(node->flags, crm_remote_node)) { /* We only autoreap from the peer cache, not the remote peer cache, * because the latter should be managed only by * crm_remote_peer_cache_refresh(). */ if(iter) { crm_notice("Purged 1 peer with id=%u and/or uname=%s from the membership cache", node->id, node->uname); g_hash_table_iter_remove(iter); } else { reap_crm_member(node->id, node->uname); } node = NULL; } } else { crm_trace("Node %s state is unchanged (%s) " CRM_XS " nodeid=%u source=%s", node->uname, state, node->id, source); } return node; } /*! * \brief Update a node's state and membership information * - * \param[in] source Caller's function name (for log messages) - * \param[in] node Node object to update - * \param[in] state Node's new state - * \param[in] membership Node's new membership ID + * \param[in] source Caller's function name (for log messages) + * \param[in,out] node Node object to update + * \param[in] state Node's new state + * \param[in] membership Node's new membership ID * * \return NULL if any node was reaped, value of node otherwise * * \note If this function returns NULL, the supplied node object was likely * freed and should not be used again. This function should not be * called within a cache iteration if reaping is possible, * otherwise reaping could invalidate the iterator. */ crm_node_t * pcmk__update_peer_state(const char *source, crm_node_t *node, const char *state, uint64_t membership) { return update_peer_state_iter(source, node, state, membership, NULL); } /*! * \internal * \brief Reap all nodes from cache whose membership information does not match * * \param[in] membership Membership ID of nodes to keep */ void pcmk__reap_unseen_nodes(uint64_t membership) { GHashTableIter iter; crm_node_t *node = NULL; crm_trace("Reaping unseen nodes..."); g_hash_table_iter_init(&iter, crm_peer_cache); while (g_hash_table_iter_next(&iter, NULL, (gpointer *)&node)) { if (node->last_seen != membership) { if (node->state) { /* * Calling update_peer_state_iter() allows us to * remove the node from crm_peer_cache without * invalidating our iterator */ update_peer_state_iter(__func__, node, CRM_NODE_LOST, membership, &iter); } else { crm_info("State of node %s[%u] is still unknown", node->uname, node->id); } } } } static crm_node_t * find_known_node(const char *id, const char *uname) { GHashTableIter iter; crm_node_t *node = NULL; crm_node_t *by_id = NULL; crm_node_t *by_name = NULL; if (uname) { g_hash_table_iter_init(&iter, known_node_cache); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &node)) { if (node->uname && strcasecmp(node->uname, uname) == 0) { crm_trace("Name match: %s = %p", node->uname, node); by_name = node; break; } } } if (id) { g_hash_table_iter_init(&iter, known_node_cache); while (g_hash_table_iter_next(&iter, NULL, (gpointer *) &node)) { if(strcasecmp(node->uuid, id) == 0) { crm_trace("ID match: %s= %p", id, node); by_id = node; break; } } } node = by_id; /* Good default */ if (by_id == by_name) { /* Nothing to do if they match (both NULL counts) */ crm_trace("Consistent: %p for %s/%s", by_id, id, uname); } else if (by_id == NULL && by_name) { crm_trace("Only one: %p for %s/%s", by_name, id, uname); if (id) { node = NULL; } else { node = by_name; } } else if (by_name == NULL && by_id) { crm_trace("Only one: %p for %s/%s", by_id, id, uname); if (uname) { node = NULL; } } else if (uname && by_id->uname && pcmk__str_eq(uname, by_id->uname, pcmk__str_casei)) { /* Multiple nodes have the same uname in the CIB. * Return by_id. */ } else if (id && by_name->uuid && pcmk__str_eq(id, by_name->uuid, pcmk__str_casei)) { /* Multiple nodes have the same id in the CIB. * Return by_name. */ node = by_name; } else { node = NULL; } if (node == NULL) { crm_debug("Couldn't find node%s%s%s%s", id? " " : "", id? id : "", uname? " with name " : "", uname? uname : ""); } return node; } static void known_node_cache_refresh_helper(xmlNode *xml_node, void *user_data) { const char *id = crm_element_value(xml_node, XML_ATTR_ID); const char *uname = crm_element_value(xml_node, XML_ATTR_UNAME); crm_node_t * node = NULL; CRM_CHECK(id != NULL && uname !=NULL, return); node = find_known_node(id, uname); if (node == NULL) { char *uniqueid = crm_generate_uuid(); node = calloc(1, sizeof(crm_node_t)); CRM_ASSERT(node != NULL); node->uname = strdup(uname); CRM_ASSERT(node->uname != NULL); node->uuid = strdup(id); CRM_ASSERT(node->uuid != NULL); g_hash_table_replace(known_node_cache, uniqueid, node); } else if (pcmk_is_set(node->flags, crm_node_dirty)) { pcmk__str_update(&node->uname, uname); /* Node is in cache and hasn't been updated already, so mark it clean */ clear_peer_flags(node, crm_node_dirty); } } static void refresh_known_node_cache(xmlNode *cib) { crm_peer_init(); g_hash_table_foreach(known_node_cache, mark_dirty, NULL); crm_foreach_xpath_result(cib, PCMK__XP_MEMBER_NODE_CONFIG, known_node_cache_refresh_helper, NULL); /* Remove all old cache entries that weren't seen in the CIB */ g_hash_table_foreach_remove(known_node_cache, is_dirty, NULL); } void pcmk__refresh_node_caches_from_cib(xmlNode *cib) { crm_remote_peer_cache_refresh(cib); refresh_known_node_cache(cib); } /*! * \internal * \brief Search known node cache * * \param[in] id If not 0, cluster node ID to search for * \param[in] uname If not NULL, node name to search for * \param[in] flags Bitmask of enum crm_get_peer_flags * * \return Known node cache entry if found, otherwise NULL */ crm_node_t * pcmk__search_known_node_cache(unsigned int id, const char *uname, uint32_t flags) { crm_node_t *node = NULL; char *id_str = NULL; CRM_ASSERT(id > 0 || uname != NULL); node = pcmk__search_node_caches(id, uname, flags); if (node || !(flags & CRM_GET_PEER_CLUSTER)) { return node; } if (id > 0) { id_str = crm_strdup_printf("%u", id); } node = find_known_node(id_str, uname); free(id_str); return node; } // Deprecated functions kept only for backward API compatibility // LCOV_EXCL_START #include int crm_terminate_member(int nodeid, const char *uname, void *unused) { return stonith_api_kick(nodeid, uname, 120, TRUE); } int crm_terminate_member_no_mainloop(int nodeid, const char *uname, int *connection) { return stonith_api_kick(nodeid, uname, 120, TRUE); } // LCOV_EXCL_STOP // End deprecated API