// Produce a reply to a RAFT-style RequestVote RPC; this is MongoDB ReplSetFresh command
// The caller should validate that the message is for the correct set, and has the required data
void TopologyCoordinatorImpl::prepareRequestVoteResponse(const Date_t now,
const BSONObj& cmdObj,
const OpTime& lastOpApplied,
std::string& errmsg,
BSONObjBuilder& result) {
string who = cmdObj["who"].String();
int cfgver = cmdObj["cfgver"].Int();
OpTime opTime(cmdObj["opTime"].Date());
bool weAreFresher = false;
if( _currentConfig.getConfigVersion() > cfgver ) {
log() << "replSet member " << who << " is not yet aware its cfg version "
<< cfgver << " is stale";
result.append("info", "config version stale");
weAreFresher = true;
}
// check not only our own optime, but any other member we can reach
else if( opTime < _commitOkayThrough ||
opTime < _latestKnownOpTime()) {
weAreFresher = true;
}
result.appendDate("opTime", lastOpApplied.asDate());
result.append("fresher", weAreFresher);
bool doVeto = _shouldVeto(cmdObj, errmsg);
result.append("veto",doVeto);
if (doVeto) {
result.append("errmsg", errmsg);
}
}
// update internal state with heartbeat response, and run topology checks
void TopologyCoordinatorImpl::updateHeartbeatInfo(Date_t now, const HeartbeatInfo& newInfo) {
// Fill in the new heartbeat data for the appropriate member
for (Member *m = _otherMembers.head(); m; m=m->next()) {
if (m->id() == newInfo.id()) {
m->get_hbinfo().updateFromLastPoll(newInfo);
break;
}
}
// Don't bother to make any changes if we are an election candidate
if (_busyWithElectSelf) return;
// ex-checkelectableset begins here
unsigned int latestOp = _latestKnownOpTime().getSecs();
// make sure the electable set is up-to-date
if (_aMajoritySeemsToBeUp()
&& !_currentConfig.self->arbiterOnly // not an arbiter
&& (_currentConfig.self->priority > 0) // not priority 0
&& (_stepDownUntil <= now) // stepDown timer has expired
&& (_memberState == MemberState::RS_SECONDARY)
// we are within 10 seconds of primary
&& (latestOp == 0 || _lastApplied.getSecs() >= latestOp - 10)) {
_electableSet.insert(_currentConfig.self->_id);
}
else {
_electableSet.erase(_currentConfig.self->_id);
}
// check if we should ask the primary (possibly ourselves) to step down
const Member* highestPriority = _getHighestPriorityElectable();
const Member* primary = _currentPrimary;
if (primary && highestPriority &&
highestPriority->config().priority > primary->config().priority &&
// if we're stepping down to allow another member to become primary, we
// better have another member (latestOp), and it should be up-to-date
latestOp != 0 && highestPriority->hbinfo().opTime.getSecs() >= latestOp - 10) {
log() << "stepping down " << primary->fullName() << " (priority " <<
primary->config().priority << "), " << highestPriority->fullName() <<
" is priority " << highestPriority->config().priority << " and " <<
(latestOp - highestPriority->hbinfo().opTime.getSecs()) << " seconds behind";
// Are we primary?
if (primary->h().isSelf()) {
// replSetStepDown tries to acquire the same lock
// msgCheckNewState takes, so we can't call replSetStepDown on
// ourselves.
// XXX Eric: schedule relinquish
//rs->relinquish();
}
else {
// We are not primary. Step down the remote node.
BSONObj cmd = BSON( "replSetStepDown" << 1 );
/* ScopedConn conn(primary->fullName());
BSONObj result;
// XXX Eric: schedule stepdown command
try {
if (!conn.runCommand("admin", cmd, result, 0)) {
log() << "stepping down " << primary->fullName()
<< " failed: " << result << endl;
}
}
catch (DBException &e) {
log() << "stepping down " << primary->fullName() << " threw exception: "
<< e.toString() << endl;
}
*/
}
}
// ex-checkauth begins here
{
int down = 0, authIssue = 0, total = 0;
for( Member *m = _otherMembers.head(); m; m=m->next() ) {
total++;
// all authIssue servers will also be not up
if (!m->hbinfo().up()) {
down++;
if (m->hbinfo().authIssue) {
authIssue++;
}
}
}
// if all nodes are down or failed auth AND at least one failed
// auth, go into recovering. If all nodes are down, stay a
// secondary.
if (authIssue > 0 && down == total) {
log() << "replset error could not reach/authenticate against any members";
if (_currentPrimary == _self) {
log() << "auth problems, relinquishing primary" << rsLog;
// XXX Eric: schedule relinquish
//rs->relinquish();
}
_blockSync = true;
// syncing is how we get into SECONDARY state, so we'll be stuck in
// RECOVERING until we unblock
_changeMemberState(MemberState::RS_RECOVERING);
}
else {
_blockSync = false;
}
}
// If a remote is primary, check that it is still up.
if (_currentPrimary && _currentPrimary->id() != _self->id()) {
if (!_currentPrimary->hbinfo().up() ||
!_currentPrimary->hbinfo().hbstate.primary()) {
_currentPrimary = NULL;
}
}
// Scan the member list's heartbeat data for who is primary, and update ourselves if it's
// not what _currentPrimary is.
{
const Member* remotePrimary(NULL);
Member* m = _otherMembers.head();
while (m) {
DEV verify( m != _self );
if( m->state().primary() && m->hbinfo().up() ) {
if( remotePrimary ) {
/* two other nodes think they are primary (asynchronously polled) -- wait for things to settle down. */
log() << "replSet info two primaries (transiently)" << rsLog;
return;
}
remotePrimary = m;
}
m = m->next();
}
if (remotePrimary) {
// If it's the same as last time, don't do anything further.
if (_currentPrimary == remotePrimary) {
return;
}
// Clear last heartbeat message on ourselves (why?)
_self->lhb() = "";
// insanity: this is what actually puts arbiters into ARBITER state
if (_currentConfig.self->arbiterOnly) {
_changeMemberState(MemberState::RS_ARBITER);
return;
}
// If we are also primary, this is a problem. Determine who should step down.
if (_memberState == MemberState::RS_PRIMARY) {
OpTime remoteElectionTime = remotePrimary->hbinfo().electionTime;
log() << "replset: another primary seen with election time "
<< remoteElectionTime;
// Step down whoever has the older election time.
if (remoteElectionTime > _electionTime) {
log() << "stepping down; another primary was elected more recently";
// XXX Eric: schedule a relinquish
//rs->relinquish();
// after completion, set currentprimary to remotePrimary.
}
else {
// else, stick around
log() << "another PRIMARY detected but it should step down"
" since it was elected earlier than me";
return;
}
}
_currentPrimary = remotePrimary;
return;
}
/* didn't find anyone who is currently primary */
}
// If we are primary, check if we can still see majority of the set;
// stepdown if we can't.
if (_currentPrimary) {
/* we must be primary */
fassert(18505, _currentPrimary == _self);
if (_shouldRelinquish()) {
log() << "can't see a majority of the set, relinquishing primary" << rsLog;
// XXX Eric: schedule a relinquish
//rs->relinquish();
}
return;
}
// At this point, there is no primary anywhere. Check to see if we should become an
// election candidate.
// If we can't elect ourselves due to config, can't become a candidate.
if (!_currentConfig.self->arbiterOnly // not an arbiter
&& (_currentConfig.self->priority > 0) // not priority 0
&& (_stepDownUntil <= now) // stepDown timer has expired
&& (_memberState == MemberState::RS_SECONDARY)) {
OCCASIONALLY log() << "replSet I don't see a primary and I can't elect myself";
return;
}
// If we can't see a majority, can't become a candidate.
if (!_aMajoritySeemsToBeUp()) {
static Date_t last;
static int n = 0;
int ll = 0;
if( ++n > 5 ) ll++;
if( last + 60 > now ) ll++;
LOG(ll) << "replSet can't see a majority, will not try to elect self" << rsLog;
last = now;
return;
}
// If we can't elect ourselves due to the current electable set;
// we are in the set if we are within 10 seconds of the latest known op (via heartbeats)
if (!(_electableSet.find(_self->id()) != _electableSet.end())) {
// we are too far behind to become primary
return;
}
// All checks passed, become a candidate and start election proceedings.
// don't try to do further elections & such while we are already working on one.
_busyWithElectSelf = true;
// XXX: schedule an election
/*
try {
rs->elect.electSelf();
}
catch(RetryAfterSleepException&) {
// we want to process new inbounds before trying this again. so we just put a checkNewstate in the queue for eval later.
requeue();
}
catch(...) {
log() << "replSet error unexpected assertion in rs manager" << rsLog;
}
}
*/
_busyWithElectSelf = false;
}
