static int test_reason(
struct ticket_config *tk,
struct booth_site *sender,
struct booth_site *leader,
struct boothc_ticket_msg *msg
)
{
int reason;
reason = ntohl(msg->header.reason);
if (reason == OR_TKT_LOST) {
if (tk->state == ST_INIT &&
tk->leader == no_leader) {
tk_log_warn("%s claims that the ticket is lost, "
"but it's in %s state (reject sent)",
site_string(sender),
state_to_string(tk->state)
);
return RLT_YOU_OUTDATED;
}
if (ticket_seems_ok(tk)) {
tk_log_warn("%s claims that the ticket is lost, "
"but it is ok here (reject sent)",
site_string(sender));
return RLT_TERM_STILL_VALID;
}
}
return 0;
}
static void log_reacquire_reason(struct ticket_config *tk)
{
int valid;
const char *where_granted = "\0";
char buff[64];
valid = is_time_set(&tk->term_expires) && !is_past(&tk->term_expires);
if (tk->leader == local) {
where_granted = "granted here";
} else {
snprintf(buff, sizeof(buff), "granted to %s",
site_string(tk->leader));
where_granted = buff;
}
if (!valid) {
tk_log_warn("%s, but not valid "
"anymore (will try to reacquire)", where_granted);
}
if (tk->is_granted && tk->leader != local) {
if (tk->leader && tk->leader != no_leader) {
tk_log_error("granted here, but also %s, "
"that's really too bad (will try to reacquire)",
where_granted);
} else {
tk_log_warn("granted here, but we're "
"not recorded as the grantee (will try to reacquire)");
}
}
}
/* For leader. */
static int process_ACK(
struct ticket_config *tk,
struct booth_site *sender,
struct booth_site *leader,
struct boothc_ticket_msg *msg
)
{
uint32_t term;
int req;
term = ntohl(msg->ticket.term);
if (newer_term(tk, sender, leader, msg, 0)) {
/* unexpected higher term */
tk_log_warn("got higher term from %s (%d vs. %d)",
site_string(sender),
term, tk->current_term);
return 0;
}
/* Don't send a reject. */
if (term < tk->current_term) {
/* Doesn't know what he's talking about - perhaps
* doesn't receive our packets? */
tk_log_warn("unexpected term "
"from %s (%d vs. %d) (ignoring)",
site_string(sender),
term, tk->current_term);
return 0;
}
/* if the ticket is to be revoked, further processing is not
* interesting (and dangerous) */
if (tk->next_state == ST_INIT || tk->state == ST_INIT)
return 0;
req = ntohl(msg->header.request);
if ((req == OP_UPDATE || req == OP_HEARTBEAT) &&
term == tk->current_term &&
leader == tk->leader) {
if (majority_of_bits(tk, tk->acks_received)) {
/* OK, at least half of the nodes are reachable;
* Update the ticket and send update messages out
*/
return leader_update_ticket(tk);
}
}
return 0;
}
/* §5.2 */
static int answer_REQ_VOTE(
struct ticket_config *tk,
struct booth_site *sender,
struct booth_site *leader,
struct boothc_ticket_msg *msg
)
{
int valid;
struct boothc_ticket_msg omsg;
cmd_result_t inappr_reason;
int reason;
inappr_reason = test_reason(tk, sender, leader, msg);
if (inappr_reason)
return send_reject(sender, tk, inappr_reason, msg);
valid = term_time_left(tk);
reason = ntohl(msg->header.reason);
/* valid tickets are not allowed only if the sender thinks
* the ticket got lost */
if (sender != tk->leader && valid && reason != OR_STEPDOWN) {
tk_log_warn("election from %s with reason %s rejected "
"(we have %s as ticket owner), ticket still valid for %ds",
site_string(sender), state_to_string(reason),
site_string(tk->leader), valid);
return send_reject(sender, tk, RLT_TERM_STILL_VALID, msg);
}
if (term_too_low(tk, sender, leader, msg))
return 0;
/* set this, so that we know not to send status for the
* ticket */
tk->in_election = 1;
/* reset ticket's leader on not valid tickets */
if (!valid)
set_leader(tk, NULL);
/* if it's a newer term or ... */
if (newer_term(tk, sender, leader, msg, 1)) {
clear_election(tk);
goto vote_for_sender;
}
/* ... we didn't vote yet, then vote for the sender */
/* §5.2, §5.4 */
if (!tk->voted_for) {
vote_for_sender:
tk->voted_for = sender;
record_vote(tk, sender, leader);
}
init_ticket_msg(&omsg, OP_VOTE_FOR, OP_REQ_VOTE, RLT_SUCCESS, 0, tk);
omsg.ticket.leader = htonl(get_node_id(tk->voted_for));
return booth_udp_send_auth(sender, &omsg, sendmsglen(&omsg));
}
static int process_UPDATE (
struct ticket_config *tk,
struct booth_site *sender,
struct booth_site *leader,
struct boothc_ticket_msg *msg
)
{
if (is_owned(tk) && sender != tk->leader) {
tk_log_warn("different leader %s wants to update "
"our ticket, sending reject",
site_string(leader));
return send_reject(sender, tk, RLT_TERM_OUTDATED, msg);
}
tk_log_debug("leader %s wants to update our ticket",
site_string(leader));
become_follower(tk, msg);
set_leader(tk, leader);
ticket_write(tk);
/* run ticket_cron if the ticket expires */
set_ticket_wakeup(tk);
return send_msg(OP_ACK, tk, sender, msg);
}
/** Try to get the ticket for the local site.
* */
int do_grant_ticket(struct ticket_config *tk, int options)
{
int rv;
tk_log_info("granting ticket");
if (tk->leader == local)
return RLT_SUCCESS;
if (is_owned(tk))
return RLT_OVERGRANT;
set_future_time(&tk->delay_commit, tk->term_duration + tk->acquire_after);
if (options & OPT_IMMEDIATE) {
tk_log_warn("granting ticket immediately! If there are "
"unreachable sites, _hope_ you are sure that they don't "
"have the ticket!");
time_reset(&tk->delay_commit);
}
rv = acquire_ticket(tk, OR_ADMIN);
if (rv) {
time_reset(&tk->delay_commit);
return rv;
} else {
return RLT_MORE;
}
}
static int do_ext_prog(struct ticket_config *tk,
int start_election)
{
int rv = 0;
if (!tk_test.path)
return 0;
switch(tk_test.progstate) {
case EXTPROG_IDLE:
rv = run_handler(tk);
if (rv == RUNCMD_ERR) {
tk_log_warn("couldn't run external test, not allowed to acquire ticket");
ext_prog_failed(tk, start_election);
}
break;
case EXTPROG_RUNNING:
/* should never get here, but just in case */
rv = RUNCMD_MORE;
break;
case EXTPROG_EXITED:
rv = tk_test_exit_status(tk);
if (rv) {
ext_prog_failed(tk, start_election);
}
break;
case EXTPROG_IGNORE:
/* nothing to do here */
break;
}
return rv;
}
int check_attr_prereq(struct ticket_config *tk, grant_type_e grant_type)
{
GList *el;
struct attr_prereq *ap;
struct geo_attr *geo_ap;
for (el = g_list_first(tk->attr_prereqs); el; el = g_list_next(el))
{
ap = (struct attr_prereq *)el->data;
if (ap->grant_type != grant_type)
continue;
geo_ap = (struct geo_attr *)g_hash_table_lookup(tk->attr, ap->attr_name);
switch(ap->op) {
case ATTR_OP_EQ:
if (!attr_found(geo_ap, ap))
goto fail;
break;
case ATTR_OP_NE:
if (attr_found(geo_ap, ap))
goto fail;
break;
default:
break;
}
}
return 0;
fail:
tk_log_warn("'%s' attr-prereq failed", ap->attr_name);
return 1;
}
void update_ticket_state(struct ticket_config *tk, struct booth_site *sender)
{
if (tk->state == ST_CANDIDATE) {
tk_log_info("learned from %s about "
"newer ticket, stopping elections",
site_string(sender));
/* there could be rejects coming from others; don't log
* warnings unnecessarily */
tk->expect_more_rejects = 1;
}
if (tk->leader == local || tk->is_granted) {
/* message from a live leader with valid ticket? */
if (sender == tk->leader && term_time_left(tk)) {
if (tk->is_granted) {
tk_log_warn("ticket was granted here, "
"but it's live at %s (revoking here)",
site_string(sender));
} else {
tk_log_info("ticket live at %s",
site_string(sender));
}
disown_ticket(tk);
ticket_write(tk);
set_state(tk, ST_FOLLOWER);
set_next_state(tk, ST_FOLLOWER);
} else {
if (tk->state == ST_CANDIDATE) {
set_state(tk, ST_FOLLOWER);
}
set_next_state(tk, ST_LEADER);
}
} else {
if (!tk->leader || tk->leader == no_leader) {
if (sender)
tk_log_info("ticket is not granted");
else
tk_log_info("ticket is not granted (from CIB)");
set_state(tk, ST_INIT);
} else {
if (sender)
tk_log_info("ticket granted to %s (says %s)",
site_string(tk->leader),
tk->leader == sender ? "they" : site_string(sender));
else
tk_log_info("ticket granted to %s (from CIB)",
site_string(tk->leader));
set_state(tk, ST_FOLLOWER);
/* just make sure that we check the ticket soon */
set_next_state(tk, ST_FOLLOWER);
}
}
}
/** Try to get the ticket for the local site.
* */
int do_grant_ticket(struct ticket_config *tk, int options)
{
int rv;
tk_log_info("granting ticket");
if (tk->leader == local)
return RLT_SUCCESS;
if (is_owned(tk)) {
if (is_manual(tk) && (options & OPT_IMMEDIATE)) {
/* -F flag has been used while granting a manual ticket.
* The ticket will be granted and may end up being granted
* on multiple sites */
tk_log_warn("manual ticket forced to be granted! be aware that "
"you may end up having two sites holding the same manual "
"ticket! revoke the ticket from the unnecessary site!");
} else {
return RLT_OVERGRANT;
}
}
set_future_time(&tk->delay_commit, tk->term_duration + tk->acquire_after);
if (options & OPT_IMMEDIATE) {
tk_log_warn("granting ticket immediately! If there are "
"unreachable sites, _hope_ you are sure that they don't "
"have the ticket!");
time_reset(&tk->delay_commit);
}
rv = acquire_ticket(tk, OR_ADMIN);
if (rv) {
time_reset(&tk->delay_commit);
return rv;
} else {
return RLT_MORE;
}
}
inline static void record_vote(struct ticket_config *tk,
struct booth_site *who,
struct booth_site *vote)
{
tk_log_debug("site %s votes for %s",
site_string(who),
site_string(vote));
if (!tk->votes_for[who->index]) {
tk->votes_for[who->index] = vote;
tk->votes_received |= who->bitmask;
} else {
if (tk->votes_for[who->index] != vote)
tk_log_warn("%s voted previously "
"for %s and now wants to vote for %s (ignored)",
site_string(who),
site_string(tk->votes_for[who->index]),
site_string(vote));
}
}
/* For follower. */
static int answer_HEARTBEAT (
struct ticket_config *tk,
struct booth_site *sender,
struct booth_site *leader,
struct boothc_ticket_msg *msg
)
{
uint32_t term;
term = ntohl(msg->ticket.term);
tk_log_debug("heartbeat from leader: %s, have %s; term %d vs %d",
site_string(leader), ticket_leader_string(tk),
term, tk->current_term);
if (term < tk->current_term) {
if (sender == tk->leader) {
tk_log_info("trusting leader %s with a lower term (%d vs %d)",
site_string(leader), term, tk->current_term);
} else if (is_owned(tk)) {
tk_log_warn("different leader %s with a lower term "
"(%d vs %d), sending reject",
site_string(leader), term, tk->current_term);
return send_reject(sender, tk, RLT_TERM_OUTDATED, msg);
}
}
/* got heartbeat, no rejects expected anymore */
tk->expect_more_rejects = 0;
/* Needed? */
newer_term(tk, sender, leader, msg, 0);
become_follower(tk, msg);
/* Racy??? */
assert(sender == leader || !leader);
set_leader(tk, leader);
/* Ack the heartbeat (we comply). */
return send_msg(OP_ACK, tk, sender, msg);
}
int raft_answer(
struct ticket_config *tk,
struct booth_site *sender,
struct booth_site *leader,
struct boothc_ticket_msg *msg
)
{
int cmd, req;
int rv;
rv = 0;
cmd = ntohl(msg->header.cmd);
req = ntohl(msg->header.request);
if (req)
tk_log_debug("got %s (req %s) from %s",
state_to_string(cmd),
state_to_string(req),
site_string(sender));
else
tk_log_debug("got %s from %s",
state_to_string(cmd),
site_string(sender));
/* don't process tickets with invalid term
*/
if (cmd != OP_STATUS &&
msg_term_invalid(tk, sender, leader, msg))
return 0;
switch (cmd) {
case OP_REQ_VOTE:
rv = answer_REQ_VOTE(tk, sender, leader, msg);
break;
case OP_VOTE_FOR:
rv = process_VOTE_FOR(tk, sender, leader, msg);
break;
case OP_ACK:
if (tk->leader == local &&
tk->state == ST_LEADER)
rv = process_ACK(tk, sender, leader, msg);
break;
case OP_HEARTBEAT:
if ((tk->leader != local || !term_time_left(tk)) &&
(tk->state == ST_INIT || tk->state == ST_FOLLOWER ||
tk->state == ST_CANDIDATE))
rv = answer_HEARTBEAT(tk, sender, leader, msg);
else {
tk_log_warn("unexpected message %s, from %s",
state_to_string(cmd),
site_string(sender));
if (ticket_seems_ok(tk))
send_reject(sender, tk, RLT_TERM_STILL_VALID, msg);
rv = -EINVAL;
}
break;
case OP_UPDATE:
if (((tk->leader != local && tk->leader == leader) || !is_owned(tk)) &&
(tk->state == ST_INIT || tk->state == ST_FOLLOWER ||
tk->state == ST_CANDIDATE)) {
rv = process_UPDATE(tk, sender, leader, msg);
} else {
tk_log_warn("unexpected message %s, from %s",
state_to_string(cmd),
site_string(sender));
if (ticket_seems_ok(tk))
send_reject(sender, tk, RLT_TERM_STILL_VALID, msg);
rv = -EINVAL;
}
break;
case OP_REJECTED:
rv = process_REJECTED(tk, sender, leader, msg);
break;
case OP_REVOKE:
rv = process_REVOKE(tk, sender, leader, msg);
break;
case OP_MY_INDEX:
rv = process_MY_INDEX(tk, sender, leader, msg);
break;
case OP_STATUS:
if (!tk->in_election)
rv = send_msg(OP_MY_INDEX, tk, sender, msg);
break;
default:
tk_log_error("unknown message %s, from %s",
state_to_string(cmd), site_string(sender));
rv = -EINVAL;
}
return rv;
}
/* reply to STATUS */
static int process_MY_INDEX (
struct ticket_config *tk,
struct booth_site *sender,
struct booth_site *leader,
struct boothc_ticket_msg *msg
)
{
int i;
int expired;
expired = !msg_term_time(msg);
/* test against the last valid(!) ticket we have */
i = my_last_term(tk) - ntohl(msg->ticket.term);
if (i > 0) {
/* let them know about our newer ticket */
send_msg(OP_MY_INDEX, tk, sender, msg);
if (tk->state == ST_LEADER) {
tk_log_info("sending ticket update to %s",
site_string(sender));
return send_msg(OP_UPDATE, tk, sender, msg);
}
}
/* we have a newer or equal ticket and theirs is expired,
* nothing more to do here */
if (i >= 0 && expired) {
return 0;
}
if (tk->state == ST_LEADER) {
/* we're the leader, thread carefully */
if (expired) {
/* if their ticket is expired,
* nothing more to do */
return 0;
}
if (i < 0) {
/* they have a newer ticket, trouble if we're already leader
* for it */
tk_log_warn("from %s: more up to date ticket at %s",
site_string(sender),
site_string(leader)
);
return leader_handle_newer_ticket(tk, sender, leader, msg);
} else {
/* we have the ticket and we don't care */
return 0;
}
} else if (tk->state == ST_CANDIDATE) {
if (leader == local) {
/* a belated MY_INDEX, we're already trying to get the
* ticket */
return 0;
}
}
/* their ticket is either newer or not expired, don't
* ignore it */
update_ticket_from_msg(tk, sender, msg);
set_leader(tk, leader);
update_ticket_state(tk, sender);
save_committed_tkt(tk);
set_ticket_wakeup(tk);
return 0;
}
static int process_REJECTED(
struct ticket_config *tk,
struct booth_site *sender,
struct booth_site *leader,
struct boothc_ticket_msg *msg
)
{
uint32_t rv;
rv = ntohl(msg->header.result);
if (tk->state == ST_CANDIDATE &&
leader == local) {
/* the sender has us as the leader (!)
* the elections will time out, then we can try again
*/
tk_log_warn("ticket was granted to us "
"(and we didn't know)");
tk->expect_more_rejects = 1;
return 0;
}
if (tk->state == ST_CANDIDATE &&
rv == RLT_TERM_OUTDATED) {
tk_log_warn("ticket outdated (term %d), granted to %s",
ntohl(msg->ticket.term),
site_string(leader)
);
set_leader(tk, leader);
tk->expect_more_rejects = 1;
become_follower(tk, msg);
return 0;
}
if (tk->state == ST_CANDIDATE &&
rv == RLT_TERM_STILL_VALID) {
if (tk->lost_leader == leader) {
if (tk->election_reason == OR_TKT_LOST) {
tk_log_warn("%s still has the ticket valid, "
"we'll backup a bit",
site_string(sender));
} else {
tk_log_warn("%s unexpectedly rejects elections",
site_string(sender));
}
} else {
tk_log_warn("ticket was granted to %s "
"(and we didn't know)",
site_string(leader));
}
set_leader(tk, leader);
become_follower(tk, msg);
tk->expect_more_rejects = 1;
return 0;
}
if (tk->state == ST_CANDIDATE &&
rv == RLT_YOU_OUTDATED) {
set_leader(tk, leader);
tk->expect_more_rejects = 1;
if (leader && leader != no_leader) {
tk_log_warn("our ticket is outdated, granted to %s",
site_string(leader));
become_follower(tk, msg);
} else {
tk_log_warn("our ticket is outdated and revoked");
update_ticket_from_msg(tk, sender, msg);
set_state(tk, ST_INIT);
}
return 0;
}
if (!tk->expect_more_rejects) {
tk_log_warn("from %s: in state %s, got %s (unexpected reject)",
site_string(sender),
state_to_string(tk->state),
state_to_string(rv));
}
return 0;
}