Imported Upstream version 1.6.5

17 files changed, 160 insertions, 73 deletions

diff --git a/db/dbcommands.cpp b/db/dbcommands.cpp
index 96374d9..7bd7203 100644
--- a/db/dbcommands.cpp
+++ b/db/dbcommands.cpp
@@ -377,7 +377,7 @@ namespace mongo {
             }
             
             {
-                BSONObjBuilder bb( result.subobjStart( "cursros" ) );
+                BSONObjBuilder bb( result.subobjStart( "cursors" ) );
                 ClientCursor::appendStats( bb );
                 bb.done();
             }
diff --git a/db/dbhelpers.cpp b/db/dbhelpers.cpp
index 124deb8..205787e 100644
--- a/db/dbhelpers.cpp
+++ b/db/dbhelpers.cpp
@@ -301,6 +301,10 @@ namespace mongo {
 
             c->checkLocation();
 
+            if ( yield && ! cc->yieldSometimes() ){
+                // cursor got finished by someone else, so we're done
+                break;
+            }
         }
 
         return num;
diff --git a/db/geo/2d.cpp b/db/geo/2d.cpp
index 19efafd..60818fc 100644
--- a/db/geo/2d.cpp
+++ b/db/geo/2d.cpp
@@ -1257,7 +1257,7 @@ namespace mongo {
 
             _want._min = Point( _g , _bl );
             _want._max = Point( _g , _tr );
-            
+
             uassert( 13064 , "need an area > 0 " , _want.area() > 0 );
 
             _state = START;
@@ -1268,12 +1268,14 @@ namespace mongo {
             
             GEODEBUG( "center : " << center.toString() << "\t" << _prefix );
 
-	    {
-	      GeoHash a(0LL,32);
-	      GeoHash b(0LL,32);
-	      b.move(1,1);
-	      _fudge = _g->distance(a,b);
-	    }
+            {
+                GeoHash a(0LL,32);
+                GeoHash b(0LL,32);
+                b.move(1,1);
+                _fudge = _g->distance(a,b);
+            }
+
+            _wantLen = _fudge + std::max((_want._max._x - _want._min._x), (_want._max._y - _want._min._y));
 
             ok();
         }
@@ -1308,32 +1310,47 @@ namespace mongo {
                         _state = DONE;
                         return;
                     }
-                    
-                    Box cur( _g , _prefix );
-                    if ( cur._min._x + _fudge < _want._min._x &&
-                         cur._min._y + _fudge < _want._min._y &&
-                         cur._max._x - _fudge > _want._max._x &&
-                         cur._max._y - _fudge > _want._max._y ){
-                        
-                        _state = DONE;
-                        GeoHash temp = _prefix.commonPrefix( cur._max.hash( _g ) );
 
-                        GEODEBUG( "box done : " << cur.toString() << " prefix:" << _prefix << " common:" << temp );
-                        
-                        if ( temp == _prefix )
-                            return;
-                        _prefix = temp;
-                        GEODEBUG( "\t one more loop" );
-                        continue;
-                    }
-                    else {
+                    if (_g->sizeEdge(_prefix) < _wantLen){
                         _prefix = _prefix.up();
+                    } else {
+                        for (int i=-1; i<=1; i++){
+                            for (int j=-1; j<=1; j++){
+
+                                if (i == 0 && j == 0)
+                                    continue; // main box
+
+                                GeoHash newBox = _prefix;
+                                newBox.move(i, j);
+
+                                PREFIXDEBUG(newBox, _g);
+
+                                Box cur( _g , newBox );
+                                if (_want.intersects(cur)){
+                                    // TODO consider splitting into quadrants
+                                    getPointsForPrefix(newBox);
+                                } else  {
+                                    GEODEBUG("skipping box");
+                                }
+                            }
+                        }
+                        _state = DONE;
                     }
+                    
                 }
                 return;
             }
 
         }
+
+        void getPointsForPrefix(const GeoHash& prefix){
+            if ( ! BtreeLocation::initial( *_id , _spec , _min , _max , prefix , _found , this ) ){
+                return;
+            }
+
+            while ( _min.hasPrefix( prefix ) && _min.advance( -1 , _found , this ) );
+            while ( _max.hasPrefix( prefix ) && _max.advance( 1 , _found , this ) );
+        }
         
         virtual bool checkDistance( const GeoHash& h , double& d ){
             bool res = _want.inside( Point( _g , h ) , _fudge );
@@ -1346,6 +1363,7 @@ namespace mongo {
         GeoHash _bl;
         GeoHash _tr;
         Box _want;
+        double _wantLen;
 
         int _found;
         
diff --git a/db/instance.cpp b/db/instance.cpp
index 9e81464..a6873f2 100644
--- a/db/instance.cpp
+++ b/db/instance.cpp
@@ -499,6 +499,7 @@ namespace mongo {
         if( ntoreturn ) 
             ss << " ntoreturn:" << ntoreturn;
 
+		time_t start = 0;
         int pass = 0;        
         bool exhaust = false;
         QueryResult* msgdata;
@@ -511,6 +512,17 @@ namespace mongo {
             catch ( GetMoreWaitException& ) { 
                 exhaust = false;
                 massert(13073, "shutting down", !inShutdown() );
+				if( pass == 0 ) { 
+  				    start = time(0);
+				}
+				else { 
+				  if( time(0) - start >= 4 ) {
+					// after about 4 seconds, return.  this is a sanity check.  pass stops at 1000 normally 
+					// for DEV this helps and also if sleep is highly inaccurate on a platform.  we want to 
+					// return occasionally so slave can checkpoint.
+					pass = 10000;
+				  }
+				}
                 pass++;
                 DEV 
                     sleepmillis(20);
diff --git a/db/namespace.cpp b/db/namespace.cpp
index de3f4df..8a1ab6f 100644
--- a/db/namespace.cpp
+++ b/db/namespace.cpp
@@ -564,8 +564,10 @@ namespace mongo {
     }
 
     void renameNamespace( const char *from, const char *to ) {
-		NamespaceIndex *ni = nsindex( from );
-		assert( ni && ni->details( from ) && !ni->details( to ) );
+        NamespaceIndex *ni = nsindex( from );
+		assert( ni );
+        assert( ni->details( from ) );
+        assert( ! ni->details( to ) );
 		
 		// Our namespace and index details will move to a different 
 		// memory location.  The only references to namespace and 
diff --git a/db/query.cpp b/db/query.cpp
index 5bd7b00..154fd15 100644
--- a/db/query.cpp
+++ b/db/query.cpp
@@ -320,7 +320,6 @@ namespace mongo {
 
             while ( 1 ) {
                 if ( !c->ok() ) {
-//                    log() << "TEMP Tailable : " << c->tailable() << ' ' << (queryOptions & QueryOption_AwaitData) << endl;
                     if ( c->tailable() ) {
                         /* when a tailable cursor hits "EOF", ok() goes false, and current() is null.  however 
                            advance() can still be retries as a reactivation attempt.  when there is new data, it will 
@@ -787,8 +786,10 @@ namespace mongo {
             if ( _pq.isExplain()) {
                 _eb.noteScan( _c.get(), _nscanned, _nscannedObjects, _n, scanAndOrderRequired(), _curop.elapsedMillis(), useHints && !_pq.getHint().eoo() );
             } else {
-                _response.appendData( _buf.buf(), _buf.len() );
-                _buf.decouple();
+                if (_buf.len()) {
+                    _response.appendData( _buf.buf(), _buf.len() );
+                    _buf.decouple();
+                }
             }
             if ( stop ) {
                 setStop();
diff --git a/db/queryutil.cpp b/db/queryutil.cpp
index 007a1ce..2153046 100644
--- a/db/queryutil.cpp
+++ b/db/queryutil.cpp
@@ -982,23 +982,16 @@ namespace mongo {
             BSONElement kk = k.next();
             int number = (int) kk.number();
             bool forward = ( number >= 0 ? 1 : -1 ) * ( _direction >= 0 ? 1 : -1 ) > 0;
-            BSONElement e = obj.getField( kk.fieldName() );
-            if ( e.eoo() ) {
-                e = staticNull.firstElement();
-            }
-            if ( e.type() == Array ) {
-                BSONObjIterator j( e.embeddedObject() );
-                bool match = false;
-                while( j.more() ) {
-                    if ( matchesElement( j.next(), i, forward ) ) {
-                        match = true;
-                        break;
-                    }
-                }
-                if ( !match ) {
-                    return false;
+            BSONElementSet keys;
+            obj.getFieldsDotted( kk.fieldName(), keys );
+            bool match = false;
+            for( BSONElementSet::const_iterator j = keys.begin(); j != keys.end(); ++j ) {
+                if ( matchesElement( *j, i, forward ) ) {
+                    match = true;
+                    break;
                 }
-            } else if ( !matchesElement( e, i, forward ) ) {
+            }
+            if ( !match ) {
                 return false;
             }
         }
diff --git a/db/repl.cpp b/db/repl.cpp
index 085ae64..ea0eab9 100644
--- a/db/repl.cpp
+++ b/db/repl.cpp
@@ -658,6 +658,8 @@ namespace mongo {
                 ReplSource tmp(c->current());
                 if ( tmp.hostName != cmdLine.source ) {
                     log() << "repl: --source " << cmdLine.source << " != " << tmp.hostName << " from local.sources collection" << endl;
+                    log() << "repl: for instructions on changing this slave's source, see:" << endl;
+                    log() << "http://dochub.mongodb.org/core/masterslave" << endl;
                     log() << "repl: terminating mongod after 30 seconds" << endl;
                     sleepsecs(30);
                     dbexit( EXIT_REPLICATION_ERROR );
diff --git a/db/repl/connections.h b/db/repl/connections.h
index 95defe4..cdf2fad 100644
--- a/db/repl/connections.h
+++ b/db/repl/connections.h
@@ -44,19 +44,36 @@ namespace mongo {
         /** throws assertions if connect failure etc. */
         ScopedConn(string hostport);
         ~ScopedConn();
-        DBClientConnection* operator->();
+
+        /* If we were to run a query and not exhaust the cursor, future use of the connection would be problematic.
+           So here what we do is wrapper known safe methods and not allow cursor-style queries at all.  This makes 
+           ScopedConn limited in functionality but very safe.  More non-cursor wrappers can be added here if needed.
+           */
+
+        bool runCommand(const string &dbname, const BSONObj& cmd, BSONObj &info, int options=0) {
+            return conn()->runCommand(dbname, cmd, info, options);
+        }
+        unsigned long long count(const string &ns) { 
+            return conn()->count(ns); 
+        }
+        BSONObj findOne(const string &ns, const Query& q, const BSONObj *fieldsToReturn = 0, int queryOptions = 0) { 
+            return conn()->findOne(ns, q, fieldsToReturn, queryOptions);
+        }
+
     private:
         auto_ptr<scoped_lock> connLock;
         static mutex mapMutex;
         struct X { 
             mutex z;
             DBClientConnection cc;
-            X() : z("X"), cc(/*reconnect*/true, 0, /*timeout*/10) { 
+            X() : z("X"), cc(/*reconnect*/ true, 0, 
+                             /*timeout*/ theReplSet ? theReplSet->config().ho.heartbeatTimeoutMillis/1000.0 : 10.0) { 
                 cc._logLevel = 2;
             }
         } *x;
         typedef map<string,ScopedConn::X*> M;
         static M& _map;
+        DBClientConnection* conn() { return &x->cc; }
     };
 
     inline ScopedConn::ScopedConn(string hostport) {
@@ -84,8 +101,8 @@ namespace mongo {
         // conLock releases...
     }
 
-    inline DBClientConnection* ScopedConn::operator->() { 
+    /*inline DBClientConnection* ScopedConn::operator->() { 
         return &x->cc; 
-    }
+    }*/
 
 }
diff --git a/db/repl/consensus.cpp b/db/repl/consensus.cpp
index 4044538..1519c26 100644
--- a/db/repl/consensus.cpp
+++ b/db/repl/consensus.cpp
@@ -134,6 +134,9 @@ namespace mongo {
         OID round = cmd["round"].OID();
         int myver = rs.config().version;
 
+        const Member* primary = rs.box.getPrimary();
+        const Member* hopeful = rs.findById(whoid);
+
         int vote = 0;
         if( set != rs.name() ) { 
             log() << "replSet error received an elect request for '" << set << "' but our set name is '" << rs.name() << "'" << rsLog;
@@ -147,6 +150,16 @@ namespace mongo {
             log() << "replSet info got stale version # during election" << rsLog;
             vote = -10000;
         }
+        else if( !hopeful ) {
+            log() << "couldn't find member with id " << whoid << rsLog;
+            vote = -10000;
+        }
+        else if( primary && primary->hbinfo().opTime > hopeful->hbinfo().opTime ) {
+            // other members might be aware of more up-to-date nodes
+            log() << hopeful->fullName() << " is trying to elect itself but " <<
+                primary->fullName() << " is already primary and more up-to-date" << rsLog;
+            vote = -10000;
+        }
         else {
             try {
                 vote = yea(whoid);
@@ -165,7 +178,7 @@ namespace mongo {
     void ReplSetImpl::_getTargets(list<Target>& L, int& configVersion) {
         configVersion = config().version;
         for( Member *m = head(); m; m=m->next() )
-            if( m->hbinfo().up() )
+            if( m->hbinfo().maybeUp() )
                 L.push_back( Target(m->fullName()) );
     }
 
diff --git a/db/repl/health.cpp b/db/repl/health.cpp
index 72396fe..c75221c 100644
--- a/db/repl/health.cpp
+++ b/db/repl/health.cpp
@@ -19,6 +19,7 @@
 #include "health.h"
 #include "../../util/background.h"
 #include "../../client/dbclient.h"
+#include "../../client/connpool.h"
 #include "../commands.h"
 #include "../../util/concurrency/value.h"
 #include "../../util/concurrency/task.h"
@@ -186,7 +187,7 @@ namespace mongo {
         //const bo fields = BSON( "o" << false << "o2" << false );
         const bo fields;
 
-        ScopedConn conn(m->fullName());        
+        ScopedDbConnection conn(m->fullName());
 
         auto_ptr<DBClientCursor> c = conn->query(rsoplog, Query().sort("$natural",1), 20, 0, &fields);
         if( c.get() == 0 ) { 
@@ -245,8 +246,6 @@ namespace mongo {
         ss << _table();
         ss << p(time_t_to_String_short(time(0)) + " current time");
 
-        //ss << "</pre>\n";
-
         if( !otEnd.isNull() ) {
             ss << "<p>Log length in time: ";
             unsigned d = otEnd.getSecs() - otFirst.getSecs();
@@ -259,6 +258,7 @@ namespace mongo {
             ss << "</p>\n";
         }
 
+        conn.done();
     }
 
     void ReplSetImpl::_summarizeAsHtml(stringstream& s) const { 
diff --git a/db/repl/heartbeat.cpp b/db/repl/heartbeat.cpp
index 4f28897..b39fad7 100644
--- a/db/repl/heartbeat.cpp
+++ b/db/repl/heartbeat.cpp
@@ -134,7 +134,7 @@ namespace mongo {
         assert( theReplSet == 0 || !theReplSet->lockedByMe() );
 
         ScopedConn conn(memberFullName);
-        return conn->runCommand("admin", cmd, result);
+        return conn.runCommand("admin", cmd, result, 0);
     }
 
     /* poll every other set member to check its status */
diff --git a/db/repl/multicmd.h b/db/repl/multicmd.h
index 61c9b5f..9eb9a17 100644
--- a/db/repl/multicmd.h
+++ b/db/repl/multicmd.h
@@ -43,7 +43,7 @@ namespace mongo {
         void run() {
             try { 
                 ScopedConn c(d.toHost);
-                d.ok = c->runCommand("admin", cmd, d.result);
+                d.ok = c.runCommand("admin", cmd, d.result);
             }
             catch(DBException&) { 
                 DEV log() << "dev caught dbexception on multiCommand " << d.toHost << rsLog;
diff --git a/db/repl/rs_config.cpp b/db/repl/rs_config.cpp
index 85c9a46..371507d 100644
--- a/db/repl/rs_config.cpp
+++ b/db/repl/rs_config.cpp
@@ -302,9 +302,8 @@ namespace mongo {
         clear();
         int level = 2;
         DEV level = 0;
-        //log(0) << "replSet load config from: " << h.toString() << rsLog;
 
-        auto_ptr<DBClientCursor> c;
+        BSONObj cfg;
         int v = -5;
         try {
             if( h.isSelf() ) {
@@ -337,13 +336,28 @@ namespace mongo {
             }
 
             v = -4;
-            ScopedConn conn(h.toString());
-            v = -3;
-            c = conn->query(rsConfigNs);
-            if( c.get() == 0 ) {
-                version = v; return;
+            unsigned long long count = 0;
+            try {
+                ScopedConn conn(h.toString());
+                v = -3;
+                cfg = conn.findOne(rsConfigNs, Query()).getOwned();
+                count = conn.count(rsConfigNs);
+            }
+            catch ( DBException& e) {
+                if ( !h.isSelf() ) {
+                    throw;
+                }
+
+                // on startup, socket is not listening yet
+                DBDirectClient cli;
+                cfg = cli.findOne( rsConfigNs, Query() ).getOwned();
+                count = cli.count(rsConfigNs);
             }
-            if( !c->more() ) {
+
+            if( count > 1 )
+                uasserted(13109, str::stream() << "multiple rows in " << rsConfigNs << " not supported host: " << h.toString());
+            
+            if( cfg.isEmpty() ) {
                 version = EMPTYCONFIG;
                 return;
             }
@@ -355,9 +369,7 @@ namespace mongo {
             return;
         }
 
-        BSONObj o = c->nextSafe();
-        uassert(13109, "multiple rows in " + rsConfigNs + " not supported", !c->more());
-        from(o);
+        from(cfg);
         checkRsConfig();
         _ok = true;
         log(level) << "replSet load config ok from " << (h.isSelf() ? "self" : h.toString()) << rsLog;
diff --git a/db/repl/rs_member.h b/db/repl/rs_member.h
index 6a797b5..099cb22 100644
--- a/db/repl/rs_member.h
+++ b/db/repl/rs_member.h
@@ -67,7 +67,6 @@ namespace mongo {
     public:
         HeartbeatInfo() : _id(0xffffffff),hbstate(MemberState::RS_UNKNOWN),health(-1.0),downSince(0),skew(INT_MIN) { }
         HeartbeatInfo(unsigned id);
-        bool up() const { return health > 0; }
         unsigned id() const { return _id; }
         MemberState hbstate;
         double health;
@@ -78,6 +77,11 @@ namespace mongo {
         OpTime opTime;
         int skew;
 
+        bool up() const { return health > 0; }
+
+        /** health is set to -1 on startup.  that means we haven't even checked yet.  0 means we checked and it failed. */
+        bool maybeUp() const { return health != 0; }
+
         long long timeDown() const; // ms
 
         /* true if changed in a way of interest to the repl set manager. */
diff --git a/db/repl/rs_sync.cpp b/db/repl/rs_sync.cpp
index 9ea65cf..9de3f60 100644
--- a/db/repl/rs_sync.cpp
+++ b/db/repl/rs_sync.cpp
@@ -70,7 +70,14 @@ namespace mongo {
                 return false;
             }
 
-            r.query(rsoplog, bo());
+            {
+                BSONObjBuilder q;
+                q.appendDate("$gte", applyGTE.asDate());
+                BSONObjBuilder query;
+                query.append("ts", q.done());
+                BSONObj queryObj = query.done();
+                r.query(rsoplog, queryObj);
+            }
             assert( r.haveCursor() );
 
             /* we lock outside the loop to avoid the overhead of locking on every operation.  server isn't usable yet anyway! */
diff --git a/db/update.cpp b/db/update.cpp
index d4a038b..e178e0f 100644
--- a/db/update.cpp
+++ b/db/update.cpp
@@ -517,10 +517,12 @@ namespace mongo {
         DEBUGUPDATE( "\t\t createNewFromMods root: " << root );
         BSONObjIteratorSorted es( obj );
         BSONElement e = es.next();
-
+        
         ModStateHolder::iterator m = _mods.lower_bound( root );
-        ModStateHolder::iterator mend = _mods.lower_bound( root + '{' );
-
+        StringBuilder buf(root.size() + 2 );
+        buf << root << (char)255;
+        ModStateHolder::iterator mend = _mods.lower_bound( buf.str() );
+        
         set<string> onedownseen;
         
         while ( e.type() && m != mend ){