Merge remote-tracking branch 'origin/wip-mds-names'

src/common/config_opts.h

@@ -219,6 +219,7 @@ OPTION(mds_dir_max_commit_size, OPT_INT, 90) // MB
 OPTION(mds_decay_halflife, OPT_FLOAT, 5)
 OPTION(mds_beacon_interval, OPT_FLOAT, 4)
 OPTION(mds_beacon_grace, OPT_FLOAT, 15)
+OPTION(mds_enforce_unique_name, OPT_BOOL, true)
 OPTION(mds_blacklist_interval, OPT_FLOAT, 24.0*60.0)  // how long to blacklist failed nodes
 OPTION(mds_session_timeout, OPT_FLOAT, 60)    // cap bits and leases time out if client idle
 OPTION(mds_session_autoclose, OPT_FLOAT, 300) // autoclose idle session

src/mds/MDS.cc

@@ -906,6 +906,20 @@ void MDS::handle_mds_map(MMDSMap *m)
       if (want_state == MDSMap::STATE_BOOT) {
         dout(10) << "not in map yet" << dendl;
       } else {
+	// did i get kicked by someone else?
+	if (g_conf->mds_enforce_unique_name) {
+	  if (uint64_t existing = mdsmap->find_mds_gid_by_name(name)) {
+	    MDSMap::mds_info_t& i = mdsmap->get_info_gid(existing);
+	    if (i.global_id > monc->get_global_id()) {
+	      dout(1) << "handle_mds_map i (" << addr
+		      << ") dne in the mdsmap, new instance has larger gid " << i.global_id
+		      << ", suicide" << dendl;
+	      suicide();
+	      goto out;
+	    }
+	  }
+	}
+
         dout(1) << "handle_mds_map i (" << addr
             << ") dne in the mdsmap, respawning myself" << dendl;
         respawn();

src/mds/MDSMap.h

@@ -236,6 +236,16 @@ public:
     assert(up.count(m) && mds_info.count(up[m]));
     return mds_info[up[m]];
   }
+  uint64_t find_mds_gid_by_name(const string& s) {
+    for (map<uint64_t,mds_info_t>::const_iterator p = mds_info.begin();
+	 p != mds_info.end();
+	 ++p) {
+      if (p->second.name == s) {
+	return p->first;
+      }
+    }
+    return 0;
+  }
 
   // counts
   unsigned get_num_in_mds() {

src/mon/MDSMonitor.cc

@@ -354,6 +354,13 @@ bool MDSMonitor::prepare_beacon(MMDSBeacon *m)
 
   // boot?
   if (state == MDSMap::STATE_BOOT) {
+    // zap previous instance of this name?
+    if (g_conf->mds_enforce_unique_name) {
+      while (uint64_t existing = pending_mdsmap.find_mds_gid_by_name(m->get_name())) {
+	fail_mds_gid(existing);
+      }
+    }
+
     // add
     MDSMap::mds_info_t& info = pending_mdsmap.mds_info[gid];
     info.global_id = gid;
@@ -376,7 +383,6 @@ bool MDSMonitor::prepare_beacon(MMDSBeacon *m)
       }
     }
 
-
     // initialize the beacon timer
     last_beacon[gid].stamp = ceph_clock_now(g_ceph_context);
     last_beacon[gid].seq = seq;
@@ -672,10 +678,30 @@ bool MDSMonitor::preprocess_command(MMonCommand *m)
     return false;
 }
 
+void MDSMonitor::fail_mds_gid(uint64_t gid)
+{
+  assert(pending_mdsmap.mds_info.count(gid));
+  MDSMap::mds_info_t& info = pending_mdsmap.mds_info[gid];
+  dout(10) << "fail_mds_gid " << gid << " mds." << info.name << " rank " << info.rank << dendl;
+
+  utime_t until = ceph_clock_now(g_ceph_context);
+  until += g_conf->mds_blacklist_interval;
+
+  pending_mdsmap.last_failure_osd_epoch = mon->osdmon()->blacklist(info.addr, until);
+  mon->osdmon()->propose_pending();
+
+  if (info.rank >= 0) {
+    pending_mdsmap.up.erase(info.rank);
+    pending_mdsmap.failed.insert(info.rank);
+  }
+
+  pending_mdsmap.mds_info.erase(gid);
+}
+
 int MDSMonitor::fail_mds(std::ostream &ss, const std::string &arg)
 {
   std::string err;
-  int w = strict_strtol(arg.c_str(), 10, &err);
+  int w = strict_strtoll(arg.c_str(), 10, &err);
   if (!err.empty()) {
     // Try to interpret the arg as an MDS name
     const MDSMap::mds_info_t *mds_info = mdsmap.find_by_name(arg);
@@ -688,18 +714,12 @@ int MDSMonitor::fail_mds(std::ostream &ss, const std::string &arg)
 
   if (pending_mdsmap.up.count(w)) {
     uint64_t gid = pending_mdsmap.up[w];
-    if (pending_mdsmap.mds_info.count(gid)) {
-      utime_t until = ceph_clock_now(g_ceph_context);
-      until += g_conf->mds_blacklist_interval;
-      MDSMap::mds_info_t& info = pending_mdsmap.mds_info[pending_mdsmap.up[w]];
-      pending_mdsmap.last_failure_osd_epoch = mon->osdmon()->blacklist(info.addr, until);
-      mon->osdmon()->propose_pending();
-
-      pending_mdsmap.mds_info.erase(gid);
-    }
-    pending_mdsmap.up.erase(w);
-    pending_mdsmap.failed.insert(w);
+    if (pending_mdsmap.mds_info.count(gid))
+      fail_mds_gid(gid);
     ss << "failed mds." << w;
+  } else if (pending_mdsmap.mds_info.count(w)) {
+    fail_mds_gid(w);
+    ss << "failed mds gid " << w;
   }
   return 0;
 }

src/mon/MDSMonitor.h

@@ -93,6 +93,8 @@ class MDSMonitor : public PaxosService {
   void get_health(list<pair<health_status_t,string> >& summary,
 		  list<pair<health_status_t,string> > *detail) const;
   int fail_mds(std::ostream &ss, const std::string &arg);
+  void fail_mds_gid(uint64_t gid);
+
   int cluster_fail(std::ostream &ss);
 
   bool preprocess_command(MMonCommand *m);