ceph源码分析--Monitor对osd report进行报down处理

上一篇讲到Monitor的tick处理，顺带讲解Monitor对osd report超时的处理

该部分在PGMonitor的tick中调用

void Monitor::tick()
{
// ok go.
dout(11) << "tick" << dendl;
···
//调用paxosservice的tick()
for (vector<PaxosService*>::iterator p = paxos_service.begin(); p != paxos_service.end(); ++p) {
(*p)->tick();
(*p)->maybe_trim();
}

···
new_tick();
}

在这里调用了PGMonitor的tick()

void PGMonitor::tick()
{
if (!is_active()) return;
if (mon->osdmon()->osdmap.require_osd_release >= CEPH_RELEASE_LUMINOUS) {
return;
}
//处理osd的report超时
handle_osd_timeouts();

if (!pg_map.pg_sum_deltas.empty()) {
utime_t age = ceph_clock_now() - pg_map.stamp;
if (age > 2 * g_conf->mon_delta_reset_interval) {
dout(10) << " clearing pg_map delta (" << age << " > " << g_conf->mon_delta_reset_interval << " seconds old)" << dendl;
pg_map.clear_delta();
}
}

/* If we have deltas for pools, run through pgmap's 'per_pool_sum_delta' and
* clear any deltas that are old enough.
*
* Note that 'per_pool_sum_delta' keeps a pool id as key, and a pair containing
* the calc'ed stats delta and an absolute timestamp from when those stats were
* obtained -- the timestamp IS NOT a delta itself.
*/
if (!pg_map.per_pool_sum_deltas.empty()) {
ceph::unordered_map<uint64_t,pair<pool_stat_t,utime_t> >::iterator it;
for (it = pg_map.per_pool_sum_delta.begin();
it != pg_map.per_pool_sum_delta.end(); ) {
utime_t age = ceph_clock_now() - it->second.second;
if (age > 2*g_conf->mon_delta_reset_interval) {
dout(10) << " clearing pg_map delta for pool " << it->first
<< " (" << age << " > " << g_conf->mon_delta_reset_interval
<< " seconds old)" << dendl;
pg_map.per_pool_sum_deltas.erase(it->first);
pg_map.per_pool_sum_deltas_stamps.erase(it->first);
pg_map.per_pool_sum_delta.erase((it++)->first);
} else {
++it;
}
}
}

dout(10) << pg_map << dendl;
}

调用了void PGMonitor::handle_osd_timeouts()

void PGMonitor::handle_osd_timeouts()
{
if (!mon->is_leader())
return;
if (did_delete)
return;

utime_t now(ceph_clock_now());
utime_t timeo(g_conf->mon_osd_report_timeout, 0);
if (now - mon->get_leader_since() < timeo) {
// We haven't been the leader for long enough to consider OSD timeouts
//
return;
}

if (mon->osdmon()->is_writeable())
//调用了osdmon的handle_osd_timeouts
mon->osdmon()->handle_osd_timeouts(now, last_osd_report);
}

bool OSDMonitor::handle_osd_timeouts(const utime_t &now,std::map

bool OSDMonitor::handle_osd_timeouts(const utime_t &now,
std::map<int,utime_t> &last_osd_report)
{
utime_t timeo(g_conf->mon_osd_report_timeout, 0);
if (now - mon->get_leader_since() < timeo) {
// We haven't been the leader for long enough to consider OSD timeouts
return false;
}

int max_osd = osdmap.get_max_osd();
bool new_down = false;

for (int i=0; i < max_osd; ++i) {
dout(30) << __func__ << ": checking up on osd " << i << dendl;
if (!osdmap.exists(i)) {
last_osd_report.erase(i); // if any
continue;
}
if (!osdmap.is_up(i))
continue;
const std::map<int,utime_t>::const_iterator t = last_osd_report.find(i);
if (t == last_osd_report.end()) {
// it wasn't in the map; start the timer.
last_osd_report[i] = now;
} else if (can_mark_down(i)) {
utime_t diff = now - t->second;
if (diff > timeo) {
mon->clog->info() << "osd." << i << " marked down after no beacon for "
<< diff << " seconds";
derr << "no beacon from osd." << i << " since " << t->second
<< ", " << diff << " seconds ago.  marking down" << dendl;
pending_inc.new_state[i] = CEPH_OSD_UP;
new_down = true;
}
}
}
return new_down;
}