/** * Copyright (c) 2013, Willem-Hendrik Thiart * Use of this source code is governed by a BSD-style license that can be * found in the LICENSE file. * * @file * @brief Implementation of a Raft server * @author Willem Thiart himself@willemthiart.com */ #include #include #include #include #include #include #include "raft.h" #include "raft_private.h" #ifndef min #define min(a, b) ((a) < (b) ? (a) : (b)) #endif #ifndef max #define max(a, b) ((a) < (b) ? (b) : (a)) #endif #ifndef __GNUC__ #define __attribute__(a) #endif void *(*raft_malloc)(size_t) = malloc; void *(*raft_calloc)(size_t, size_t) = calloc; void *(*raft_realloc)(void *, size_t) = realloc; void (*raft_free)(void *) = free; void raft_set_heap_functions(void *(*_malloc)(size_t), void *(*_calloc)(size_t, size_t), void *(*_realloc)(void *, size_t), void (*_free)(void *)) { raft_malloc = _malloc; raft_calloc = _calloc; raft_realloc = _realloc; raft_free = _free; } static void raft_log_node(raft_server_t *me_, raft_node_id_t id, const char *fmt, ...) __attribute__ ((format (printf, 3, 4))); static void raft_log_node(raft_server_t *me_, raft_node_id_t id, const char *fmt, ...) { raft_server_private_t* me = (raft_server_private_t*)me_; if (me->cb.log == NULL) return; char buf[1024]; va_list args; va_start(args, fmt); int n = vsnprintf(buf, sizeof(buf), fmt, args); if (n < 0) { buf[0] = '\0'; } va_end(args); me->cb.log(me_, id, me->udata, buf); } #define raft_log(me, ...) (raft_log_node(me, RAFT_NODE_ID_NONE, __VA_ARGS__)) void raft_randomize_election_timeout(raft_server_t* me_) { raft_server_private_t* me = (raft_server_private_t*)me_; /* [election_timeout, 2 * election_timeout) */ me->election_timeout_rand = me->election_timeout + rand() % me->election_timeout; raft_log(me_, "randomize election timeout to %d", me->election_timeout_rand); } void raft_update_quorum_meta(raft_server_t* me_, raft_msg_id_t id) { raft_server_private_t* me = (raft_server_private_t*) me_; // Make sure that timeout is greater than 'randomized election timeout' me->quorum_timeout = me->election_timeout * 2; me->last_acked_msg_id = id; } raft_server_t* raft_new_with_log(const raft_log_impl_t *log_impl, void *log_arg) { raft_server_private_t* me = raft_calloc(1, sizeof(raft_server_private_t)); if (!me) return NULL; me->current_term = 0; me->voted_for = -1; me->timeout_elapsed = 0; me->request_timeout = 200; me->election_timeout = 1000; me->node_transferring_leader_to = RAFT_NODE_ID_NONE; raft_update_quorum_meta((raft_server_t*)me, me->msg_id); raft_randomize_election_timeout((raft_server_t*)me); me->log_impl = log_impl; me->log = me->log_impl->init(me, log_arg); if (!me->log) { raft_free(me); return NULL; } me->voting_cfg_change_log_idx = -1; raft_set_state((raft_server_t*)me, RAFT_STATE_FOLLOWER); me->leader_id = RAFT_NODE_ID_NONE; me->snapshot_in_progress = 0; raft_set_snapshot_metadata((raft_server_t*)me, 0, 0); return (raft_server_t*)me; } raft_server_t* raft_new(void) { return raft_new_with_log(&raft_log_internal_impl, NULL); } void raft_set_callbacks(raft_server_t* me_, raft_cbs_t* funcs, void* udata) { raft_server_private_t* me = (raft_server_private_t*)me_; memcpy(&me->cb, funcs, sizeof(raft_cbs_t)); me->udata = udata; } void raft_destroy(raft_server_t* me_) { raft_server_private_t* me = (raft_server_private_t*)me_; me->log_impl->free(me->log); raft_free(me_); } void raft_clear(raft_server_t* me_) { raft_server_private_t* me = (raft_server_private_t*)me_; me->current_term = 0; me->voted_for = -1; me->timeout_elapsed = 0; raft_randomize_election_timeout(me_); me->voting_cfg_change_log_idx = -1; raft_set_state((raft_server_t*)me, RAFT_STATE_FOLLOWER); me->leader_id = RAFT_NODE_ID_NONE; me->commit_idx = 0; me->last_applied_idx = 0; me->num_nodes = 0; me->node = NULL; me->log_impl->reset(me->log, 1, 1); } raft_node_t* raft_add_node_internal(raft_server_t* me_, raft_entry_t *ety, void* udata, raft_node_id_t id, int is_self) { raft_server_private_t* me = (raft_server_private_t*)me_; if (id == RAFT_NODE_ID_NONE) return NULL; /* set to voting if node already exists */ raft_node_t* node = raft_get_node(me_, id); if (node) { if (!raft_node_is_voting(node)) { raft_node_set_voting(node, 1); return node; } else /* we shouldn't add a node twice */ return NULL; } node = raft_node_new(udata, id); if (!node) return NULL; void* p = raft_realloc(me->nodes, sizeof(void*) * (me->num_nodes + 1)); if (!p) { raft_node_free(node); return NULL; } me->num_nodes++; me->nodes = p; me->nodes[me->num_nodes - 1] = node; if (is_self) me->node = me->nodes[me->num_nodes - 1]; node = me->nodes[me->num_nodes - 1]; if (me->cb.notify_membership_event) me->cb.notify_membership_event(me_, raft_get_udata(me_), node, ety, RAFT_MEMBERSHIP_ADD); return node; } static raft_node_t* raft_add_non_voting_node_internal(raft_server_t* me_, raft_entry_t *ety, void* udata, raft_node_id_t id, int is_self) { if (raft_get_node(me_, id)) return NULL; raft_node_t* node = raft_add_node_internal(me_, ety, udata, id, is_self); if (!node) return NULL; raft_node_set_voting(node, 0); return node; } raft_node_t* raft_add_node(raft_server_t* me_, void* udata, raft_node_id_t id, int is_self) { return raft_add_node_internal(me_, NULL, udata, id, is_self); } raft_node_t* raft_add_non_voting_node(raft_server_t* me_, void* udata, raft_node_id_t id, int is_self) { return raft_add_non_voting_node_internal(me_, NULL, udata, id, is_self); } void raft_remove_node(raft_server_t* me_, raft_node_t* node) { raft_server_private_t* me = (raft_server_private_t*)me_; if (me->cb.notify_membership_event) me->cb.notify_membership_event(me_, raft_get_udata(me_), node, NULL, RAFT_MEMBERSHIP_REMOVE); assert(node); int i, found = 0; for (i = 0; i < me->num_nodes; i++) { if (me->nodes[i] == node) { found = 1; break; } } assert(found); memmove(&me->nodes[i], &me->nodes[i + 1], sizeof(*me->nodes) * (me->num_nodes - i - 1)); me->num_nodes--; if (me->leader_id == raft_node_get_id(node)) { me->leader_id = RAFT_NODE_ID_NONE; } raft_node_free(node); } void raft_handle_append_cfg_change(raft_server_t* me_, raft_entry_t* ety, raft_index_t idx) { raft_server_private_t* me = (raft_server_private_t*)me_; if (!raft_entry_is_cfg_change(ety)) return; if (!me->cb.get_node_id) return; void* udata = raft_get_udata(me_); raft_node_id_t node_id = me->cb.get_node_id(me_, udata, ety, idx); raft_node_t* node = raft_get_node(me_, node_id); int is_self = node_id == raft_get_nodeid(me_); switch (ety->type) { case RAFT_LOGTYPE_ADD_NONVOTING_NODE: if (!is_self) { if (node && !raft_node_is_active(node)) { raft_node_set_active(node, 1); } else if (!node) { node = raft_add_non_voting_node_internal(me_, ety, NULL, node_id, is_self); assert(node); } } break; case RAFT_LOGTYPE_ADD_NODE: node = raft_add_node_internal(me_, ety, NULL, node_id, is_self); assert(node); assert(raft_node_is_voting(node)); break; case RAFT_LOGTYPE_REMOVE_NODE: if (node) { raft_node_set_active(node, 0); } break; default: assert(0); } } void raft_handle_remove_cfg_change(raft_server_t* me_, raft_entry_t* ety, const raft_index_t idx) { raft_server_private_t* me = (raft_server_private_t*) me_; if (!raft_entry_is_cfg_change(ety)) return; if (!me->cb.get_node_id) return; void* udata = raft_get_udata(me_); raft_node_id_t node_id = me->cb.get_node_id(me_, udata, ety, idx); raft_node_t* node = raft_get_node(me_, node_id); switch (ety->type) { case RAFT_LOGTYPE_REMOVE_NODE: raft_node_set_active(node, 1); break; case RAFT_LOGTYPE_ADD_NONVOTING_NODE: assert(node_id != raft_get_nodeid(me_)); raft_remove_node(me_, node); break; case RAFT_LOGTYPE_ADD_NODE: raft_node_set_voting(node, 0); break; default: assert(0); break; } } int raft_delete_entry_from_idx(raft_server_t* me_, raft_index_t idx) { raft_server_private_t* me = (raft_server_private_t*)me_; assert(raft_get_commit_idx(me_) < idx); if (idx <= me->voting_cfg_change_log_idx) me->voting_cfg_change_log_idx = -1; return me->log_impl->pop(me->log, idx, (func_entry_notify_f) raft_handle_remove_cfg_change, me_); } int raft_election_start(raft_server_t* me_) { raft_server_private_t* me = (raft_server_private_t*)me_; raft_log(me_, "election starting: %d %d, term: %ld ci: %ld", me->election_timeout_rand, me->timeout_elapsed, me->current_term, raft_get_current_idx(me_)); me->leader_id = RAFT_NODE_ID_NONE; me->timeout_elapsed = 0; raft_randomize_election_timeout(me_); return raft_become_precandidate(me_); } int raft_become_leader(raft_server_t* me_) { raft_server_private_t* me = (raft_server_private_t*)me_; int i; raft_log(me_, "becoming leader term:%ld", raft_get_current_term(me_)); if (me->cb.notify_state_event) me->cb.notify_state_event(me_, raft_get_udata(me_), RAFT_STATE_LEADER); raft_reset_transfer_leader(me_); raft_index_t next_idx = raft_get_current_idx(me_) + 1; if (raft_get_current_term(me_) > 1) { raft_entry_t *noop = raft_entry_new(0); noop->term = raft_get_current_term(me_); noop->type = RAFT_LOGTYPE_NO_OP; int e = raft_append_entry(me_, noop); raft_entry_release(noop); if (0 != e) return e; // Commit noop immediately if this is a single node cluster if (raft_is_single_node_voting_cluster(me_)) { raft_set_commit_idx(me_, raft_get_current_idx(me_)); } } raft_set_state(me_, RAFT_STATE_LEADER); raft_update_quorum_meta(me_, me->msg_id); me->timeout_elapsed = 0; for (i = 0; i < me->num_nodes; i++) { raft_node_t* node = me->nodes[i]; if (me->node == node) continue; raft_node_set_next_idx(node, next_idx); raft_node_set_match_idx(node, 0); raft_send_appendentries(me_, node); } return 0; } int raft_become_precandidate(raft_server_t* me_) { raft_server_private_t* me = (raft_server_private_t*)me_; raft_log(me_, "becoming pre-candidate, next term : %ld", me->current_term + 1); if (me->cb.notify_state_event) me->cb.notify_state_event(me_, me->udata, RAFT_STATE_PRECANDIDATE); for (int i = 0; i < me->num_nodes; i++) raft_node_vote_for_me(me->nodes[i], 0); raft_set_state(me_, RAFT_STATE_PRECANDIDATE); for (int i = 0; i < me->num_nodes; i++) { raft_node_t* node = me->nodes[i]; if (me->node != node && raft_node_is_voting(node)) raft_send_requestvote(me_, node); } return 0; } int raft_become_candidate(raft_server_t* me_) { raft_server_private_t* me = (raft_server_private_t*)me_; int i; raft_log(me_, "becoming candidate"); if (me->cb.notify_state_event) me->cb.notify_state_event(me_, raft_get_udata(me_), RAFT_STATE_CANDIDATE); raft_reset_transfer_leader(me_); int e = raft_set_current_term(me_, raft_get_current_term(me_) + 1); if (0 != e) return e; for (i = 0; i < me->num_nodes; i++) raft_node_vote_for_me(me->nodes[i], 0); if (raft_node_is_voting(me->node)) raft_vote(me_, me->node); me->leader_id = RAFT_NODE_ID_NONE; raft_set_state(me_, RAFT_STATE_CANDIDATE); for (i = 0; i < me->num_nodes; i++) { raft_node_t* node = me->nodes[i]; if (me->node != node && raft_node_is_voting(node)) { raft_send_requestvote(me_, node); } } return 0; } void raft_become_follower(raft_server_t* me_) { raft_server_private_t* me = (raft_server_private_t*)me_; raft_log(me_, "becoming follower"); if (me->cb.notify_state_event) me->cb.notify_state_event(me_, raft_get_udata(me_), RAFT_STATE_FOLLOWER); raft_reset_transfer_leader(me_); raft_set_state(me_, RAFT_STATE_FOLLOWER); raft_randomize_election_timeout(me_); me->timeout_elapsed = 0; me->leader_id = RAFT_NODE_ID_NONE; } static int msgid_cmp(const void *a, const void *b) { raft_msg_id_t va = *((raft_msg_id_t*) a); raft_msg_id_t vb = *((raft_msg_id_t*) b); return va > vb ? -1 : 1; } static raft_msg_id_t quorum_msg_id(raft_server_t* me_) { raft_server_private_t* me = (raft_server_private_t*) me_; raft_msg_id_t msg_ids[me->num_nodes]; int num_voters = 0; for (int i = 0; i < me->num_nodes; i++) { raft_node_t* node = me->nodes[i]; if (!raft_node_is_voting(node)) continue; if (me->node == node) { msg_ids[num_voters++] = me->msg_id; } else { msg_ids[num_voters++] = raft_node_get_last_acked_msgid(node); } } assert(num_voters == raft_get_num_voting_nodes(me_)); /** * Sort the acknowledged msg_ids in the descending order and return * the median value. Median value means it's the highest msg_id * acknowledged by the majority. */ qsort(msg_ids, num_voters, sizeof(raft_msg_id_t), msgid_cmp); return msg_ids[num_voters / 2]; } int raft_periodic(raft_server_t* me_, int msec_since_last_period) { raft_server_private_t* me = (raft_server_private_t*)me_; me->timeout_elapsed += msec_since_last_period; /* Only one voting node means it's safe for us to become the leader */ if (raft_is_single_node_voting_cluster(me_) && !raft_is_leader(me_)) { // need to update term on new leadership int e = raft_set_current_term(me_, raft_get_current_term(me_) + 1); if (e != 0) { return e; } e = raft_become_leader(me_); if (e != 0) { return e; } } if (me->state == RAFT_STATE_LEADER) { if (me->request_timeout <= me->timeout_elapsed) { me->msg_id++; raft_send_appendentries_all(me_); } me->quorum_timeout -= msec_since_last_period; if (me->quorum_timeout < 0) { /** * Check-quorum implementation * * Periodically (every quorum_timeout), we enter this check to * verify quorum exists. The current 'quorum msg id' should be * greater than the last time we were here. It means we've got * responses for append entry requests from the cluster, so we * conclude that cluster is operational and quorum exists. In that * case, we save the current quorum msg id and update the * quorum_timeout timer. Otherwise, it means quorum does not exist. * We should step down and become a follower. */ raft_msg_id_t quorum_id = quorum_msg_id(me_); if (me->last_acked_msg_id == quorum_id) { raft_log(me_, "quorum does not exist, stepping down"); raft_become_follower(me_); } raft_update_quorum_meta(me_, quorum_id); } if (me->node_transferring_leader_to != RAFT_NODE_ID_NONE) { me->transfer_leader_time -= msec_since_last_period; if (me->transfer_leader_time < 0) { if (me->cb.notify_state_event) me->cb.notify_state_event(me_, raft_get_udata(me_), RAFT_STATE_LEADERSHIP_TRANSFER_FAILED); raft_reset_transfer_leader(me_); } } } else if ((me->election_timeout_rand <= me->timeout_elapsed || me->timeout_now) && /* Don't become the leader when building snapshots or bad things will * happen when we get a client request */ !raft_snapshot_is_in_progress(me_)) { int e = raft_election_start(me_); if (0 != e) return e; me->timeout_now = 0; } if (me->last_applied_idx < raft_get_commit_idx(me_) && raft_is_apply_allowed(me_)) { int e = raft_apply_all(me_); if (0 != e) return e; } raft_process_read_queue(me_); return 0; } raft_entry_t* raft_get_entry_from_idx(raft_server_t* me_, raft_index_t etyidx) { raft_server_private_t* me = (raft_server_private_t*)me_; return me->log_impl->get(me->log, etyidx); } int raft_voting_change_is_in_progress(raft_server_t* me_) { return ((raft_server_private_t*)me_)->voting_cfg_change_log_idx != -1; } int raft_recv_appendentries_response(raft_server_t* me_, raft_node_t* node, msg_appendentries_response_t* r) { raft_server_private_t* me = (raft_server_private_t*)me_; raft_log_node(me_, raft_node_get_id(node), "received appendentries response %s ci:%ld rci:%ld msgid:%lu", r->success == 1 ? "SUCCESS" : "fail", raft_get_current_idx(me_), r->current_idx, r->msg_id); if (!node) return -1; if (!raft_is_leader(me_)) return RAFT_ERR_NOT_LEADER; if (raft_node_get_last_acked_msgid(node) > r->msg_id) { // this was received out of order and is now irrelevant. return 0; } /* If response contains term T > currentTerm: set currentTerm = T and convert to follower (§5.3) */ if (me->current_term < r->term) { int e = raft_set_current_term(me_, r->term); if (0 != e) return e; raft_become_follower(me_); return 0; } else if (me->current_term != r->term) return 0; // if we got here, it means that the follower has acked us as a leader, even if it cant accept the append_entry raft_node_set_last_ack(node, r->msg_id, r->term); raft_index_t match_idx = raft_node_get_match_idx(node); if (0 == r->success) { /* Stale response -- ignore */ if (r->current_idx < match_idx) return 0; raft_index_t next = min(r->current_idx + 1, raft_get_current_idx(me_)); assert(0 < next); raft_node_set_next_idx(node, next); /* retry */ raft_send_appendentries(me_, node); return 0; } if (me->cb.send_timeoutnow && raft_get_transfer_leader(me_) == raft_node_get_id(node) && raft_get_current_idx(me_) == r->current_idx) { me->cb.send_timeoutnow(me_, node); } if (!raft_node_is_voting(node) && !raft_voting_change_is_in_progress(me_) && raft_get_current_idx(me_) <= r->current_idx + 1 && !raft_node_is_voting_committed(node) && raft_node_is_addition_committed(node) && me->cb.node_has_sufficient_logs && 0 == raft_node_has_sufficient_logs(node) ) { int e = me->cb.node_has_sufficient_logs(me_, me->udata, node); if (0 == e) raft_node_set_has_sufficient_logs(node); } if (r->current_idx <= match_idx) return 0; assert(r->current_idx <= raft_get_current_idx(me_)); raft_node_set_next_idx(node, r->current_idx + 1); raft_node_set_match_idx(node, r->current_idx); /* Update commit idx */ raft_index_t point = r->current_idx; if (point) { raft_entry_t* ety = raft_get_entry_from_idx(me_, point); if (raft_get_commit_idx(me_) < point && ety->term == me->current_term) { int votes = raft_node_is_voting(me->node) ? 1 : 0; for (int i = 0; i < me->num_nodes; i++) { raft_node_t* follower = me->nodes[i]; if (me->node != follower && raft_node_is_voting(follower) && point <= raft_node_get_match_idx(follower)) { votes++; } } if (raft_get_num_voting_nodes(me_) / 2 < votes) raft_set_commit_idx(me_, point); } if (ety) raft_entry_release(ety); } /* Aggressively send remaining entries */ if (raft_node_get_next_idx(node) <= raft_get_current_idx(me_)) raft_send_appendentries(me_, node); /* periodic applies committed entries lazily */ return 0; } int raft_recv_appendentries( raft_server_t* me_, raft_node_t* node, msg_appendentries_t* ae, msg_appendentries_response_t *r ) { raft_server_private_t* me = (raft_server_private_t*)me_; int e = 0; if (0 < ae->n_entries) { raft_log_node( me_, ae->leader_id, "recvd appendentries li:%d t:%ld ci:%ld lc:%ld pli:%ld plt:%ld #%d", ae->leader_id, ae->term, raft_get_current_idx(me_), ae->leader_commit, ae->prev_log_idx, ae->prev_log_term, ae->n_entries); } r->msg_id = ae->msg_id; r->success = 0; if (raft_is_candidate(me_) && me->current_term == ae->term) { raft_become_follower(me_); } else if (me->current_term < ae->term) { e = raft_set_current_term(me_, ae->term); if (0 != e) goto out; raft_become_follower(me_); } else if (ae->term < me->current_term) { /* 1. Reply false if term < currentTerm (§5.1) */ raft_log_node(me_, ae->leader_id, "AE term %ld is less than current term %ld", ae->term, me->current_term); goto out; } if (node != NULL) { raft_node_update_max_seen_msg_id(node, ae->msg_id); } /* update current leader because ae->term is up to date */ me->leader_id = ae->leader_id; me->timeout_elapsed = 0; /* Not the first appendentries we've received */ /* NOTE: the log starts at 1 */ if (0 < ae->prev_log_idx) { raft_entry_t* ety = raft_get_entry_from_idx(me_, ae->prev_log_idx); /* Is a snapshot */ if (ae->prev_log_idx == me->snapshot_last_idx) { if (me->snapshot_last_term != ae->prev_log_term) { /* Should never happen; something is seriously wrong! */ raft_log_node(me_, ae->leader_id, "Snapshot AE prev conflicts with committed entry"); e = RAFT_ERR_SHUTDOWN; if (ety) raft_entry_release(ety); goto out; } } /* 2. Reply false if log doesn't contain an entry at prevLogIndex whose term matches prevLogTerm (§5.3) */ else if (!ety) { raft_log_node(me_, ae->leader_id, "AE no log at prev_idx %ld", ae->prev_log_idx); goto out; } else if (ety->term != ae->prev_log_term) { raft_log_node(me_, ae->leader_id, "AE term doesn't match prev_term (ie. %ld vs %ld) ci:%ld comi:%ld lcomi:%ld pli:%ld", ety->term, ae->prev_log_term, raft_get_current_idx(me_), raft_get_commit_idx(me_), ae->leader_commit, ae->prev_log_idx); if (ae->prev_log_idx <= raft_get_commit_idx(me_)) { /* Should never happen; something is seriously wrong! */ raft_log_node(me_, ae->leader_id, "AE prev conflicts with committed entry"); e = RAFT_ERR_SHUTDOWN; raft_entry_release(ety); goto out; } /* Delete all the following log entries because they don't match */ e = raft_delete_entry_from_idx(me_, ae->prev_log_idx); raft_entry_release(ety); goto out; } if (ety) raft_entry_release(ety); } r->success = 1; r->current_idx = ae->prev_log_idx; /* synchronize msg_id to leader. not really needed for raft, but needed for virtraft for msg_id to be increasing * cluster wide so that can verify read_queue correctness easily. Otherwise, it be fine for msg_id to be unique to * each raft_server_t. */ if (ae->msg_id > me->msg_id) { me->msg_id = ae->msg_id; } /* 3. If an existing entry conflicts with a new one (same index but different terms), delete the existing entry and all that follow it (§5.3) */ int i; for (i = 0; i < ae->n_entries; i++) { raft_entry_t* ety = ae->entries[i]; raft_index_t ety_index = ae->prev_log_idx + 1 + i; raft_entry_t* existing_ety = raft_get_entry_from_idx(me_, ety_index); raft_term_t existing_term = existing_ety ? existing_ety->term : 0; if (existing_ety) raft_entry_release(existing_ety); if (existing_ety && existing_term != ety->term) { if (ety_index <= raft_get_commit_idx(me_)) { /* Should never happen; something is seriously wrong! */ raft_log_node(me_, ae->leader_id, "AE entry conflicts with committed entry ci:%ld comi:%ld lcomi:%ld pli:%ld", raft_get_current_idx(me_), raft_get_commit_idx(me_), ae->leader_commit, ae->prev_log_idx); e = RAFT_ERR_SHUTDOWN; goto out; } e = raft_delete_entry_from_idx(me_, ety_index); if (0 != e) goto out; break; } else if (!existing_ety) break; r->current_idx = ety_index; } /* Pick up remainder in case of mismatch or missing entry */ for (; i < ae->n_entries; i++) { e = raft_append_entry(me_, ae->entries[i]); if (0 != e) goto out; r->current_idx = ae->prev_log_idx + 1 + i; } /* 4. If leaderCommit > commitIndex, set commitIndex = min(leaderCommit, index of most recent entry) */ if (raft_get_commit_idx(me_) < ae->leader_commit) { raft_index_t last_log_idx = max(raft_get_current_idx(me_), 1); raft_set_commit_idx(me_, min(last_log_idx, ae->leader_commit)); } out: r->term = me->current_term; if (0 == r->success) r->current_idx = raft_get_current_idx(me_); return e; } int raft_recv_requestvote(raft_server_t* me_, raft_node_t* node, msg_requestvote_t* vr, msg_requestvote_response_t *r) { raft_server_private_t* me = (raft_server_private_t*)me_; int e = 0; r->prevote = vr->prevote; r->request_term = vr->term; r->vote_granted = 0; /* Reject request if we have a leader */ if (!vr->transfer_leader && me->leader_id != RAFT_NODE_ID_NONE && me->leader_id != vr->candidate_id && me->timeout_elapsed < me->election_timeout) { goto done; } /* Update the term only if this is not a prevote request */ if (!vr->prevote && raft_get_current_term(me_) < vr->term) { e = raft_set_current_term(me_, vr->term); if (0 != e) { goto done; } raft_become_follower(me_); } if (me->current_term > vr->term) { goto done; } if (me->current_term == vr->term && (me->voted_for != -1 && me->voted_for != vr->candidate_id)) { goto done; } /* Below we check if log is more up-to-date... */ raft_index_t current_idx = raft_get_current_idx(me_); raft_term_t ety_term = raft_get_last_log_term(me_); if (vr->last_log_term < ety_term || (vr->last_log_term == ety_term && vr->last_log_idx < current_idx)) { goto done; } r->vote_granted = 1; if (!vr->prevote) { /* It shouldn't be possible for a leader or candidate to grant a vote * Both states would have voted for themselves */ assert(!(raft_is_leader(me_) || raft_is_candidate(me_))); e = raft_vote_for_nodeid(me_, vr->candidate_id); if (0 != e) r->vote_granted = 0; /* must be in an election. */ me->leader_id = RAFT_NODE_ID_NONE; me->timeout_elapsed = 0; } done: raft_log_node(me_, vr->candidate_id, "node requested vote: %d, t:%ld, pv:%d replying: %s", vr->candidate_id, vr->term, vr->prevote, r->vote_granted == 1 ? "granted" : r->vote_granted == 0 ? "not granted" : "unknown"); r->term = raft_get_current_term(me_); return e; } int raft_votes_is_majority(const int num_nodes, const int nvotes) { if (num_nodes < nvotes) return 0; int half = num_nodes / 2; return half + 1 <= nvotes; } int raft_recv_requestvote_response(raft_server_t* me_, raft_node_t* node, msg_requestvote_response_t* r) { raft_server_private_t* me = (raft_server_private_t*)me_; raft_log_node(me_, raft_node_get_id(node), "node responded to requestvote status:%s pv:%d rt:%ld ct:%ld rt:%ld", r->vote_granted == 1 ? "granted" : r->vote_granted == 0 ? "not granted" : "unknown", r->prevote, r->request_term, me->current_term, r->term); if (r->term > me->current_term) { int e = raft_set_current_term(me_, r->term); if (0 != e) return e; raft_become_follower(me_); return 0; } if (r->prevote) { /* Validate prevote is not stale */ if (!raft_is_precandidate(me_) || r->request_term != me->current_term + 1) return 0; } else { /* Validate reqvote is not stale */ if (!raft_is_candidate(me_) || r->request_term != me->current_term) return 0; } if (r->vote_granted) { if (node) raft_node_vote_for_me(node, 1); int votes = raft_get_nvotes_for_me(me_); int nodes = raft_get_num_voting_nodes(me_); if (raft_votes_is_majority(nodes, votes)) { int e = raft_is_precandidate(me_) ? raft_become_candidate(me_) : raft_become_leader(me_); if (0 != e) return e; } } return 0; } int raft_recv_entry(raft_server_t* me_, msg_entry_t* ety, msg_entry_response_t *r) { raft_server_private_t* me = (raft_server_private_t*)me_; int i; if (raft_entry_is_voting_cfg_change(ety)) { /* Only one voting cfg change at a time */ if (raft_voting_change_is_in_progress(me_)) return RAFT_ERR_ONE_VOTING_CHANGE_ONLY; /* Multi-threading: need to fail here because user might be * snapshotting membership settings. */ if (!raft_is_apply_allowed(me_)) return RAFT_ERR_SNAPSHOT_IN_PROGRESS; } if (!raft_is_leader(me_)) return RAFT_ERR_NOT_LEADER; if (raft_get_transfer_leader(me_) != RAFT_NODE_ID_NONE) return RAFT_ERR_LEADER_TRANSFER_IN_PROGRESS; raft_log(me_, "received entry t:%ld id: %d idx: %ld", me->current_term, ety->id, raft_get_current_idx(me_) + 1); ety->term = me->current_term; int e = raft_append_entry(me_, ety); if (0 != e) return e; for (i = 0; i < me->num_nodes; i++) { raft_node_t* node = me->nodes[i]; if (me->node == node || !node) continue; /* Only send new entries. * Don't send the entry to peers who are behind, to prevent them from * becoming congested. */ raft_index_t next_idx = raft_node_get_next_idx(node); if (next_idx == raft_get_current_idx(me_)) raft_send_appendentries(me_, node); } /* if we are the only voter, commit now, as no appendentries_response will occur */ if (raft_is_single_node_voting_cluster(me_)) { raft_set_commit_idx(me_, raft_get_current_idx(me_)); } r->id = ety->id; r->idx = raft_get_current_idx(me_); r->term = me->current_term; /* FIXME: is this required if raft_append_entry does this too? */ if (raft_entry_is_voting_cfg_change(ety)) me->voting_cfg_change_log_idx = raft_get_current_idx(me_); return 0; } int raft_send_requestvote(raft_server_t* me_, raft_node_t* node) { raft_server_private_t* me = (raft_server_private_t*)me_; msg_requestvote_t rv; int e = 0; assert(node); assert(node != me->node); raft_node_id_t id = raft_node_get_id(node); raft_log_node(me_, id, "sending requestvote to: %d", id); if (raft_is_precandidate(me_)) { rv.prevote = 1; rv.term = me->current_term + 1; } else { rv.prevote = 0; rv.term = me->current_term; } rv.last_log_idx = raft_get_current_idx(me_); rv.last_log_term = raft_get_last_log_term(me_); rv.candidate_id = raft_get_nodeid(me_); rv.transfer_leader = (me->node_transferring_leader_to != RAFT_NODE_ID_NONE); if (me->cb.send_requestvote) e = me->cb.send_requestvote(me_, me->udata, node, &rv); return e; } int raft_append_entry(raft_server_t* me_, raft_entry_t* ety) { raft_server_private_t* me = (raft_server_private_t*)me_; int e = me->log_impl->append(me->log, ety); if (e < 0) return e; if (raft_entry_is_voting_cfg_change(ety)) me->voting_cfg_change_log_idx = raft_get_current_idx(me_) - 1; if (raft_entry_is_cfg_change(ety)) { raft_handle_append_cfg_change(me_, ety, raft_get_current_idx(me_)); } return 0; } int raft_apply_entry(raft_server_t* me_) { raft_server_private_t* me = (raft_server_private_t*)me_; if (!raft_is_apply_allowed(me_)) return -1; /* Don't apply after the commit_idx */ if (me->last_applied_idx == raft_get_commit_idx(me_)) return -1; raft_index_t log_idx = me->last_applied_idx + 1; raft_entry_t* ety = raft_get_entry_from_idx(me_, log_idx); if (!ety) return -1; raft_log(me_, "applying log: %ld, id: %d size: %d", log_idx, ety->id, ety->data_len); me->last_applied_idx++; if (me->cb.applylog) { int e = me->cb.applylog(me_, me->udata, ety, me->last_applied_idx); if (RAFT_ERR_SHUTDOWN == e) { raft_entry_release(ety); return RAFT_ERR_SHUTDOWN; } } /* voting cfg change is now complete. * TODO: there seem to be a possible off-by-one bug hidden here, requiring * checking log_idx >= voting_cfg_change_log_idx rather than plain ==. */ if (log_idx >= me->voting_cfg_change_log_idx) me->voting_cfg_change_log_idx = -1; if (!raft_entry_is_cfg_change(ety)) goto exit; raft_node_id_t node_id = me->cb.get_node_id(me_, raft_get_udata(me_), ety, log_idx); raft_node_t* node = raft_get_node(me_, node_id); switch (ety->type) { case RAFT_LOGTYPE_ADD_NODE: raft_node_set_addition_committed(node, 1); raft_node_set_voting_committed(node, 1); /* Membership Change: confirm connection with cluster */ raft_node_set_has_sufficient_logs(node); if (node_id == raft_get_nodeid(me_)) me->connected = RAFT_NODE_STATUS_CONNECTED; break; case RAFT_LOGTYPE_ADD_NONVOTING_NODE: raft_node_set_addition_committed(node, 1); break; case RAFT_LOGTYPE_REMOVE_NODE: if (node) { raft_remove_node(me_, node); } break; default: break; } exit: raft_entry_release(ety); return 0; } raft_entry_t** raft_get_entries_from_idx(raft_server_t* me_, raft_index_t idx, int* n_etys) { if (raft_get_current_idx(me_) < idx) { *n_etys = 0; return NULL; } raft_server_private_t* me = (raft_server_private_t*)me_; raft_index_t size = raft_get_current_idx(me_) - idx + 1; raft_entry_t **e = raft_malloc(size * sizeof(raft_entry_t*)); int n = me->log_impl->get_batch(me->log, idx, (int) size, e); if (n < 1) { raft_free(e); *n_etys = 0; return NULL; } *n_etys = n; return e; } int raft_send_appendentries(raft_server_t* me_, raft_node_t* node) { raft_server_private_t* me = (raft_server_private_t*)me_; assert(node); assert(node != me->node); if (!raft_node_is_active(node)) { return 0; } raft_index_t next_idx = raft_node_get_next_idx(node); /* figure out if the client needs a snapshot sent */ if (me->snapshot_last_idx > 0 && next_idx <= me->snapshot_last_idx) { if (me->cb.send_snapshot) me->cb.send_snapshot(me_, me->udata, node); return RAFT_ERR_NEEDS_SNAPSHOT; } if (!me->cb.send_appendentries) return -1; msg_appendentries_t ae = { .term = me->current_term, .leader_id = raft_get_nodeid(me_), .leader_commit = raft_get_commit_idx(me_), .msg_id = ++me->msg_id, }; ae.entries = raft_get_entries_from_idx(me_, next_idx, &ae.n_entries); assert((!ae.entries && 0 == ae.n_entries) || (ae.entries && 0 < ae.n_entries)); /* previous log is the log just before the new logs */ if (next_idx > 1) { raft_entry_t* prev_ety = raft_get_entry_from_idx(me_, next_idx - 1); if (!prev_ety) { ae.prev_log_idx = me->snapshot_last_idx; ae.prev_log_term = me->snapshot_last_term; } else { ae.prev_log_idx = next_idx - 1; ae.prev_log_term = prev_ety->term; raft_entry_release(prev_ety); } } raft_log_node(me_, raft_node_get_id(node), "sending appendentries: ci:%lu comi:%lu t:%lu lc:%lu pli:%lu plt:%lu msgid:%lu #%d", raft_get_current_idx(me_), raft_get_commit_idx(me_), ae.term, ae.leader_commit, ae.prev_log_idx, ae.prev_log_term, ae.msg_id, ae.n_entries); int res = me->cb.send_appendentries(me_, me->udata, node, &ae); raft_entry_release_list(ae.entries, ae.n_entries); raft_free(ae.entries); return res; } int raft_send_appendentries_all(raft_server_t* me_) { raft_server_private_t* me = (raft_server_private_t*)me_; int i, e; int ret = 0; me->timeout_elapsed = 0; for (i = 0; i < me->num_nodes; i++) { if (me->node == me->nodes[i]) continue; e = raft_send_appendentries(me_, me->nodes[i]); if (0 != e) ret = e; } return ret; } int raft_get_nvotes_for_me(raft_server_t* me_) { raft_server_private_t* me = (raft_server_private_t*)me_; int i, votes; for (i = 0, votes = 0; i < me->num_nodes; i++) { if (me->node != me->nodes[i] && raft_node_is_voting(me->nodes[i]) && raft_node_has_vote_for_me(me->nodes[i])) { votes += 1; } } if (raft_node_is_voting(me->node)) votes += 1; return votes; } int raft_vote(raft_server_t* me_, raft_node_t* node) { return raft_vote_for_nodeid(me_, node ? raft_node_get_id(node) : -1); } int raft_vote_for_nodeid(raft_server_t* me_, const raft_node_id_t nodeid) { raft_server_private_t* me = (raft_server_private_t*)me_; if (me->cb.persist_vote) { int e = me->cb.persist_vote(me_, me->udata, nodeid); if (0 != e) return e; } me->voted_for = nodeid; return 0; } int raft_msg_entry_response_committed(raft_server_t* me_, const msg_entry_response_t* r) { raft_entry_t* ety = raft_get_entry_from_idx(me_, r->idx); if (!ety) return 0; raft_term_t ety_term = ety->term; raft_entry_release(ety); /* entry from another leader has invalidated this entry message */ if (r->term != ety_term) return -1; return r->idx <= raft_get_commit_idx(me_); } int raft_apply_all(raft_server_t* me_) { if (!raft_is_apply_allowed(me_)) return 0; while (raft_get_last_applied_idx(me_) < raft_get_commit_idx(me_)) { int e = raft_apply_entry(me_); if (0 != e) return e; } return 0; } int raft_entry_is_voting_cfg_change(raft_entry_t* ety) { return RAFT_LOGTYPE_ADD_NODE == ety->type || RAFT_LOGTYPE_REMOVE_NODE == ety->type; } int raft_entry_is_cfg_change(raft_entry_t* ety) { return ( RAFT_LOGTYPE_ADD_NODE == ety->type || RAFT_LOGTYPE_ADD_NONVOTING_NODE == ety->type || RAFT_LOGTYPE_REMOVE_NODE == ety->type); } int raft_poll_entry(raft_server_t* me_) { raft_server_private_t* me = (raft_server_private_t*)me_; /* We should never drop uncommitted entries */ assert(me->log_impl->first_idx(me->log) <= raft_get_commit_idx(me_)); int e = me->log_impl->poll(me->log, me->log_impl->first_idx(me->log)); if (e != 0) return e; return 0; } int raft_pop_entry(raft_server_t* me_) { raft_server_private_t* me = (raft_server_private_t*)me_; raft_index_t cur_idx = me->log_impl->current_idx(me->log); return me->log_impl->pop(me->log, cur_idx, (func_entry_notify_f) raft_handle_remove_cfg_change, me_); } raft_index_t raft_get_first_entry_idx(raft_server_t* me_) { raft_server_private_t* me = (raft_server_private_t*)me_; assert(0 < raft_get_current_idx(me_)); if (me->snapshot_last_idx == 0) return 1; return me->snapshot_last_idx; } raft_index_t raft_get_num_snapshottable_logs(raft_server_t *me_) { raft_server_private_t* me = (raft_server_private_t*)me_; if (raft_get_log_count(me_) <= 1) return 0; return raft_get_commit_idx(me_) - me->log_impl->first_idx(me->log) + 1; } int raft_begin_snapshot(raft_server_t *me_, int flags) { raft_server_private_t* me = (raft_server_private_t*)me_; if (raft_get_num_snapshottable_logs(me_) == 0) return -1; raft_index_t snapshot_target = raft_get_commit_idx(me_); if (!snapshot_target) return -1; raft_entry_t* ety = raft_get_entry_from_idx(me_, snapshot_target); if (!ety) return -1; raft_term_t ety_term = ety->term; raft_entry_release(ety); /* we need to get all the way to the commit idx */ int e = raft_apply_all(me_); if (e != 0) return e; assert(raft_get_commit_idx(me_) == raft_get_last_applied_idx(me_)); raft_set_snapshot_metadata(me_, ety_term, snapshot_target); me->snapshot_in_progress = 1; me->snapshot_flags = flags; raft_log(me_, "begin snapshot sli:%ld slt:%ld slogs:%ld", me->snapshot_last_idx, me->snapshot_last_term, raft_get_num_snapshottable_logs(me_)); return 0; } int raft_cancel_snapshot(raft_server_t *me_) { raft_server_private_t* me = (raft_server_private_t*)me_; if (!me->snapshot_in_progress) return -1; me->snapshot_last_idx = me->saved_snapshot_last_idx; me->snapshot_last_term = me->saved_snapshot_last_term; me->snapshot_in_progress = 0; return 0; } int raft_end_snapshot(raft_server_t *me_) { raft_server_private_t* me = (raft_server_private_t*)me_; if (!me->snapshot_in_progress || me->snapshot_last_idx == 0) return -1; // TODO: What is the purpose of this assert? Looks wrong. // assert(raft_get_num_snapshottable_logs(me_) != 0); /* If needed, remove compacted logs */ int e = me->log_impl->poll(me->log, me->snapshot_last_idx + 1); if (e != 0) return e; me->snapshot_in_progress = 0; raft_log(me_, "end snapshot base:%ld commit-index:%ld current-index:%ld", me->log_impl->first_idx(me->log) - 1, raft_get_commit_idx(me_), raft_get_current_idx(me_)); if (!raft_is_leader(me_)) return 0; for (int i = 0; i < me->num_nodes; i++) { raft_node_t* node = me->nodes[i]; if (me->node == node || !raft_node_is_active(node)) continue; raft_index_t next_idx = raft_node_get_next_idx(node); /* figure out if the client needs a snapshot sent */ if (me->snapshot_last_idx > 0 && next_idx <= me->snapshot_last_idx) { if (me->cb.send_snapshot) me->cb.send_snapshot(me_, me->udata, node); } } return 0; } int raft_begin_load_snapshot( raft_server_t *me_, raft_term_t last_included_term, raft_index_t last_included_index) { raft_server_private_t* me = (raft_server_private_t*)me_; if (last_included_index == -1) return -1; if (last_included_index == 0 || last_included_term == 0) return -1; /* loading the snapshot will break cluster safety */ if (last_included_index < me->last_applied_idx) return -1; /* snapshot was unnecessary */ if (last_included_index < raft_get_current_idx(me_)) return -1; if (last_included_term == me->snapshot_last_term && last_included_index == me->snapshot_last_idx) return RAFT_ERR_SNAPSHOT_ALREADY_LOADED; if (me->current_term < last_included_term) { raft_set_current_term(me_, last_included_term); me->current_term = last_included_term; } raft_set_state((raft_server_t*)me, RAFT_STATE_FOLLOWER); me->leader_id = RAFT_NODE_ID_NONE; me->log_impl->reset(me->log, last_included_index + 1, last_included_term); if (raft_get_commit_idx(me_) < last_included_index) raft_set_commit_idx(me_, last_included_index); me->last_applied_idx = last_included_index; raft_set_snapshot_metadata(me_, last_included_term, me->last_applied_idx); /* remove all nodes but self */ int i, my_node_by_idx = 0; for (i = 0; i < me->num_nodes; i++) { if (raft_get_nodeid(me_) == raft_node_get_id(me->nodes[i])) my_node_by_idx = i; else { raft_node_free(me->nodes[i]); me->nodes[i] = NULL; } } /* this will be realloc'd by a raft_add_node */ me->nodes[0] = me->nodes[my_node_by_idx]; me->num_nodes = 1; raft_log(me_, "loaded snapshot sli:%ld slt:%ld slogs:%ld", me->snapshot_last_idx, me->snapshot_last_term, raft_get_num_snapshottable_logs(me_)); return 0; } int raft_end_load_snapshot(raft_server_t *me_) { raft_server_private_t* me = (raft_server_private_t*)me_; int i; /* Set nodes' voting status as committed */ for (i = 0; i < me->num_nodes; i++) { raft_node_t* node = me->nodes[i]; raft_node_set_voting_committed(node, raft_node_is_voting(node)); raft_node_set_addition_committed(node, 1); if (raft_node_is_voting(node)) raft_node_set_has_sufficient_logs(node); } return 0; } void *raft_get_log(raft_server_t *me_) { raft_server_private_t* me = (raft_server_private_t*)me_; return me->log; } raft_entry_t *raft_entry_new(unsigned int data_len) { raft_entry_t *ety = raft_calloc(1, sizeof(raft_entry_t) + data_len); ety->data_len = data_len; ety->refs = 1; return ety; } void raft_entry_hold(raft_entry_t *ety) { assert(ety->refs < UINT16_MAX); ety->refs++; } void raft_entry_release(raft_entry_t *ety) { assert(ety->refs > 0); ety->refs--; if (!ety->refs) { if (ety->free_func) { ety->free_func(ety); } else { raft_free(ety); } } } void raft_entry_release_list(raft_entry_t **ety_list, size_t len) { size_t i; for (i = 0; i < len; i++) { raft_entry_release(ety_list[i]); } } void raft_queue_read_request(raft_server_t* me_, func_read_request_callback_f cb, void *cb_arg) { raft_server_private_t* me = (raft_server_private_t*) me_; raft_read_request_t *req = raft_malloc(sizeof(raft_read_request_t)); req->read_idx = raft_get_current_idx(me_); req->read_term = raft_get_current_term(me_); req->msg_id = ++me->msg_id; req->cb = cb; req->cb_arg = cb_arg; req->next = NULL; if (!me->read_queue_head) me->read_queue_head = req; if (me->read_queue_tail) me->read_queue_tail->next = req; me->read_queue_tail = req; raft_send_appendentries_all(me_); } static void pop_read_queue(raft_server_private_t *me, int can_read) { raft_read_request_t *p = me->read_queue_head; p->cb(p->cb_arg, can_read); /* remove entry and update head/tail */ if (p->next) { me->read_queue_head = p->next; if (!me->read_queue_head->next) me->read_queue_tail = me->read_queue_head; } else { me->read_queue_head = NULL; me->read_queue_tail = NULL; } raft_free(p); } void raft_process_read_queue(raft_server_t* me_) { raft_server_private_t* me = (raft_server_private_t*) me_; if (!me->read_queue_head) return; /* As a follower we drop all queued read requests */ if (raft_is_follower(me_)) { while (me->read_queue_head) { pop_read_queue(me, 0); } return; } /* As a leader we can process requests that fulfill these conditions: * 1) Heartbeat acknowledged by majority * 2) We're on the same term (note: is this needed or over cautious?) */ if (!raft_is_leader(me_)) return; /* Quickly bail if nothing to do */ if (!me->read_queue_head) return; if (raft_get_num_voting_nodes(me_) > 1) { raft_entry_t *ety = raft_get_entry_from_idx(me_, raft_get_commit_idx(me_)); if (!ety) return; raft_term_t ety_term = ety->term; raft_entry_release(ety); /* Don't read if we did not commit an entry this term yet! * A new term has a NO_OP committed so if everything is well * we can except that to happen. */ if (ety_term < me->current_term) return; } raft_msg_id_t last_acked_msgid = quorum_msg_id(me_); raft_index_t last_applied_idx = me->last_applied_idx; /* Special case: the log's first index is 1, so we need to account * for that in case we read before anything was ever committed. * * Note that this also implies a single node because adding nodes would * bump the log and commit index. */ if (!me->commit_idx && !me->last_applied_idx && raft_get_current_idx(me_) == 1) last_applied_idx = 1; while (me->read_queue_head && me->read_queue_head->msg_id <= last_acked_msgid && me->read_queue_head->read_idx <= last_applied_idx) { pop_read_queue(me, me->read_queue_head->read_term == me->current_term); } } /* invoke a leadership transfer * node_id = targeted node we are transfering to * timeout = how long this should be allowed to take in milliseconds, as calculated by calls to raft_periodic) * return an error if leadership transfer is already in progress of the targeted node_id is unknown */ int raft_transfer_leader(raft_server_t* me_, raft_node_id_t node_id, long timeout) { raft_server_private_t* me = (raft_server_private_t*) me_; if (me->node_transferring_leader_to != RAFT_NODE_ID_NONE) { return RAFT_ERR_LEADER_TRANSFER_IN_PROGRESS; } raft_node_t * target = raft_get_node(me_, node_id); if (target == NULL) { return RAFT_ERR_INVALID_NODEID; } if (me->cb.send_timeoutnow && raft_get_current_idx(me_) == raft_node_get_match_idx(target)) { me->cb.send_timeoutnow(me_, target); } me->node_transferring_leader_to = node_id; if (timeout == 0) { me->transfer_leader_time = me->election_timeout; } else { me->transfer_leader_time = timeout; } return 0; }