wdgzyjx/myxce
1
0
Fork 0
forked from lei.chen/myxce
Code Pull requests Activity

feat(ha-single-leader-role): delete dispatch-gate COMMITTING branch + collapse role predicates to PRIMARY (task 4.2)

Browse source
This commit is contained in:
Rongsong Shen 2026-06-13 12:59:58 +00:00
parent a5bf86112a
commit 71ce41f486
3 changed files with 35 additions and 54 deletions
Download patch file Download diff file Expand all files Collapse all files

59
src/controller/ha.c
View file

@ -198,7 +198,7 @@ static int raft_cb_send_vote_request(void *ctx, int peer_idx, uint64_t term)
*
* from_idx is the absolute candidate index of the requester.
* Applies the same checks as handle_request_vote:
* - role gate: PRIMARY or COMMITTING → deny
* - role gate: PRIMARY → deny
* - fresh heartbeat on STANDBY → deny
* - in_catch_up → deny
* - freshness: candidate config_tag/lsn must be ≥ ours
@ -212,7 +212,7 @@ static int raft_cb_vote_eligible(void *ctx, int from_idx, uint64_t term)
/* Role gate */
ctrl_role_t role = (ctrl_role_t)atomic_load_explicit(&ha->role,
memory_order_acquire);
if (role == CTRL_PRIMARY || role == CTRL_PRIMARY_COMMITTING)
if (role == CTRL_PRIMARY)
return 0;
/* Fresh heartbeat on STANDBY → primary is alive, deny */
@ -1534,26 +1534,6 @@ int ha_dispatch_gate(ha_state_t *ha, int opcode)
return 0;
return -XCE_ERR_NOT_READY;
case CTRL_PRIMARY_COMMITTING:
/* Req 5.1, 5.2, 8.1: Primary won election but has not yet received
* majority replication ack. Allow only replication-related and
* administrative opcodes; deny job dispatch and queue management
* with XCE_ERR_NOT_READY. */
switch (opcode) {
case OPCode_heartbeat:
case OPCode_replConfig:
case OPCode_replConfigPush:
case OPCode_upgradeConfigAck:
case OPCode_hello:
case OPCode_walAck:
case OPCode_replEntry:
case OPCode_replAck:
case OPCode_replCatchUpDone:
return 0;
default:
return -XCE_ERR_NOT_READY;
}
case CTRL_STANDBY: {
/* Allow ops that standbys legitimately handle */
switch (opcode) {
@ -1957,7 +1937,7 @@ static int ha_heartbeat_is_fresh(ha_state_t *ha)
*
* Decision order (fail-fast):
* 0. role gate (Req 6.1, 6.2):
* - CTRL_PRIMARY or CTRL_PRIMARY_COMMITTING → deny
* - CTRL_PRIMARY → deny
* - CTRL_STANDBY with fresh heartbeat → deny
* 1. in_catch_up → deny (Req 2.5)
* 2. req term < my term → deny (Req 2.2)
@ -1988,20 +1968,19 @@ int handle_request_vote(ha_state_t *ha,
* Raft §5.1: "If RPC request or response contains term T > currentTerm:
* set currentTerm = T, convert to follower."
*
* Without this ordering, a node in CTRL_PRIMARY_COMMITTING at term=N
* would deny a vote request at term=N+1 without ever seeing the higher
* term. Both nodes then independently promote to PRIMARY via their
* committing timeout → split-brain. */
* Without this ordering, a CTRL_PRIMARY node at term=N would deny a vote
* request at term=N+1 without ever seeing the higher term, leaving two
* nodes both believing they lead → split-brain. */
uint64_t my_term = atomic_load_explicit(&ha->term, memory_order_acquire);
if (term > my_term) {
atomic_store_explicit(&ha->term, term, memory_order_release);
ha->voted_for_term = 0;
memset(ha->voted_for_name, 0, sizeof(ha->voted_for_name));
/* Step down from leader/committing: a higher-term election is in
* progress and we must not block it with the role gate below. */
/* Step down from leader: a higher-term election is in progress and we
* must not block it with the role gate below. */
ctrl_role_t cur_role = (ctrl_role_t)atomic_load_explicit(&ha->role,
memory_order_acquire);
if (cur_role == CTRL_PRIMARY || cur_role == CTRL_PRIMARY_COMMITTING) {
if (cur_role == CTRL_PRIMARY) {
ha_do_demote(ha);
INFO("HA: stepping down (term %" PRIu64 " → %" PRIu64
") — higher-term election in progress", my_term, term);
@ -2017,12 +1996,12 @@ int handle_request_vote(ha_state_t *ha,
/* Req 6.1, 6.2: role-based deny gate.
* Evaluated AFTER the term catch-up above: if we just stepped down due to
* a higher term, role is now CTRL_STANDBY and we fall through to voting.
* PRIMARY and PRIMARY_COMMITTING at the current term still deny votes
* (protecting the live primary from same-term challengers). */
* PRIMARY at the current term still denies votes (protecting the live
* primary from same-term challengers). */
ctrl_role_t my_role = (ctrl_role_t)atomic_load_explicit(&ha->role,
memory_order_acquire);
if (my_role == CTRL_PRIMARY || my_role == CTRL_PRIMARY_COMMITTING)
return 0; /* deny — I am primary (or committing); protect against challengers */
if (my_role == CTRL_PRIMARY)
return 0; /* deny — I am primary; protect against same-term challengers */
if (my_role == CTRL_STANDBY && ha->primary_ctrl_fd >= 0 &&
ha_heartbeat_is_fresh(ha))
return 0; /* deny — my primary is alive (active primary connection) */
@ -2507,7 +2486,7 @@ static void on_vote_connected(endpoint_t *ep)
#endif
/* Set state-aware dispatcher — routes by ha->role at message-arrival time.
* Never reassigned: ELECTING → vote reply; COMMITTING/PRIMARY → repl ack. */
* Never reassigned: ELECTING → vote reply; PRIMARY → repl ack. */
ep->msg_handler = peer_ctrl_msg_handler;
}
@ -2642,7 +2621,7 @@ done:
/*
* repl_ack_decode_and_dispatch: decode an inbound packet on a promoted vote
* endpoint while role == CTRL_PRIMARY_COMMITTING or CTRL_PRIMARY.
* endpoint while role == CTRL_PRIMARY.
*
* - OPCode_replCatchUpDone → INFO no-op (ack-counting removed in task 4.1)
* - OPCode_reply with stateReportMsg → ha_process_standby_state (Task 3.1)
@ -2820,7 +2799,6 @@ static int repl_ack_decode_and_dispatch(endpoint_t *ep, vote_ep_ctx_t *vctx)
* peer_ctrl_msg_handler: state-aware iomgr handler set once in on_vote_connected.
* Routes by ha->role at dispatch time — no explicit handler reassignment needed.
* CTRL_ELECTING → vote_reply_decode_and_dispatch
* CTRL_PRIMARY_COMMITTING /
* CTRL_PRIMARY → repl_ack_decode_and_dispatch
* any other role → drop packet (WARN)
*/
@ -2836,7 +2814,6 @@ static int peer_ctrl_msg_handler(void *ep_raw)
case CTRL_ELECTING:
return vote_reply_decode_and_dispatch(ep, vctx);
case CTRL_PRIMARY_COMMITTING:
case CTRL_PRIMARY:
return repl_ack_decode_and_dispatch(ep, vctx);
@ -3068,7 +3045,7 @@ void vote_ep_ctx_set_cluster_for_test(endpoint_t *ep, cluster_t *c)
/*
* raft_cb_become_leader_for_test: drive the leader promotion path without
* a live cluster. Passes ha as ctx; sets ha->raft_c = NULL to suppress
* production-only calls (phase2, push_config, ctrlhello, committing timer).
* production-only calls (phase2, push_config, ctrlhello).
*/
void raft_cb_become_leader_for_test(ha_state_t *ha, uint64_t term)
{
@ -3105,7 +3082,7 @@ static int vote_ep_err_handler(void *ep_raw)
ctrl_role_t role = (ctrl_role_t)atomic_load_explicit(&ha->role,
memory_order_acquire);
if (role == CTRL_PRIMARY_COMMITTING || role == CTRL_PRIMARY) {
if (role == CTRL_PRIMARY) {
/* Promoted connection dropped — remove from standby_endpoints */
if (ha->standby_endpoints) {
pthread_mutex_lock(&ha->lock);
@ -3902,7 +3879,7 @@ static int primary_heartbeat_fn(void *arg)
ctrl_role_t role = (ctrl_role_t)atomic_load_explicit(&ha->role,
memory_order_acquire);
if (role != CTRL_PRIMARY && role != CTRL_PRIMARY_COMMITTING)
if (role != CTRL_PRIMARY)
return 0; /* demoted — stop sending NOPs */
/* 1. Heartbeat all connected standbys — idle-triggered NOP (Req 3.1-3.4).