lots of options for initial sync SERVER-1829

3 files changed, 100 insertions, 20 deletions

diff --git a/db/repl/rs_config.cpp b/db/repl/rs_config.cpp
index f454579c451..172d38ca4dc 100644
--- a/db/repl/rs_config.cpp
+++ b/db/repl/rs_config.cpp
@@ -138,6 +138,32 @@ namespace mongo {
         uassert(13438, "bad slaveDelay value", slaveDelay >= 0 && slaveDelay <= 3600 * 24 * 366);
         uassert(13439, "priority must be 0 when hidden=true", priority == 0 || !hidden);
         uassert(13477, "priority must be 0 when buildIndexes=false", buildIndexes || priority == 0);
+
+        if (!initialSync.isEmpty()) {
+            static const string legal[] = {"state", "name", "_id","optime"};
+            static const set<string> legals(legal, legal + 4);
+            assertOnlyHas(initialSync, legals);
+            
+            if (initialSync.hasElement("state")) {
+                uassert(13525, "initialSync source state must be 1 or 2",
+                        initialSync["state"].isNumber() &&
+                        (initialSync["state"].Number() == 1 ||
+                         initialSync["state"].Number() == 2));
+            }
+            if (initialSync.hasElement("name")) {
+                uassert(13526, "initialSync source name must be a string",
+                        initialSync["name"].type() == mongo::String);
+            }
+            if (initialSync.hasElement("_id")) {
+                uassert(13527, "initialSync source _id must be a number",
+                        initialSync["_id"].isNumber());
+            }
+            if (initialSync.hasElement("optime")) {
+                uassert(13528, "initialSync source optime must be a timestamp",
+                        initialSync["optime"].type() == mongo::Timestamp ||
+                        initialSync["optime"].type() == mongo::Date);
+            }
+        }
     }
 
     /** @param o old config
@@ -251,8 +277,10 @@ namespace mongo {
             BSONObj mobj = members[i].Obj();
             MemberCfg m;
             try {
-                static const string legal[] = {"_id","votes","priority","host","hidden","slaveDelay","arbiterOnly","buildIndexes","tags"};
-                static const set<string> legals(legal, legal + 8);
+                static const string legal[] =
+                    {"_id","votes","priority","host", "hidden","slaveDelay",
+                     "arbiterOnly","buildIndexes","tags","initialSync"};
+                static const set<string> legals(legal, legal + 10);
                 assertOnlyHas(mobj, legals);
 
                 try { 
@@ -286,6 +314,9 @@ namespace mongo {
                     for( unsigned i = 0; i < v.size(); i++ )
                         m.tags.insert( v[i].String() );
                 }
+                if( mobj.hasElement("initialSync")) {
+                    m.initialSync = mobj["initialSync"].Obj();
+                }
                 m.check();
             }
             catch( const char * p ) { 
diff --git a/db/repl/rs_config.h b/db/repl/rs_config.h
index 13b8c399c37..aef32389832 100644
--- a/db/repl/rs_config.h
+++ b/db/repl/rs_config.h
@@ -51,6 +51,7 @@ namespace mongo {
             bool hidden;          /* if set, don't advertise to drives in isMaster. for non-primaries (priority 0) */
             bool buildIndexes;    /* if false, do not create any non-_id indexes */
             set<string> tags;     /* tagging for data center, rack, etc. */
+            BSONObj initialSync;  /* directions for initial sync source */
 
             void check() const;   /* check validity, assert if not. */
             BSONObj asBson() const;
diff --git a/db/repl/rs_initialsync.cpp b/db/repl/rs_initialsync.cpp
index 92b6c6c9b97..46430bae004 100644
--- a/db/repl/rs_initialsync.cpp
+++ b/db/repl/rs_initialsync.cpp
@@ -104,32 +104,80 @@ namespace mongo {
     }
 
     /**
-     * Choose a member to sync from. Prefers a secondary, if available.
+     * Choose a member to sync from. 
      *
-     * TODO: make a config setting like "cloneFromPrimary" or something to force
-     * machines to clone from the primary.  We might want to integrate this with
-     * tags functionality, if we only want it to be able to clone from certain
-     * machines.
+     * The initalSync option is an object with 1 k/v pair:
+     * 
+     * "state" : 1|2
+     * "name" : "host"
+     * "_id" : N
+     * "optime" : t
+     *
+     * All except optime are exact matches.  "optime" will find a secondary with
+     * an optime >= to the optime given.
      */
     const Member* ReplSetImpl::getMemberToSyncTo() {
-        for( Member *m = head(); m; m = m->next() ) {
-            if (m->hbinfo().up() && m->state() == MemberState::RS_SECONDARY) {
-                sethbmsg( str::stream() << "syncing to secondary: " << m->fullName(), 0);
-                return const_cast<Member*>(m);
-            }
-        }
+        BSONObj sync = myConfig().initialSync;
+        bool secondaryOnly = false, isOpTime = false;
+        char *name = 0;
+        int id = -1;
+        OpTime optime;
 
-        // can't find secondary, try primary
         StateBox::SP sp = box.get();
         assert( !sp.state.primary() ); // wouldn't make sense if we were.
 
-        // we just checked in _syncDoInitialSync, but we could have lost a
-        // primary since then
-        if (!sp.primary) {
-            return 0;
+        // if it exists, we've already checked that these fields are valid in
+        // rs_config.cpp
+        if ( !sync.isEmpty() ) {
+            if (sync.hasElement("state")) {
+                if (sync["state"].Number() == 1) {
+                    if (sp.primary) {
+                        sethbmsg( str::stream() << "syncing to primary: " << sp.primary->fullName(), 0);
+                        return const_cast<Member*>(sp.primary);
+                    }
+                    else {
+                        sethbmsg("couldn't clone from primary");
+                        return NULL;
+                    }
+                }
+                else {
+                    secondaryOnly = true;
+                }
+            }
+            if (sync.hasElement("name")) {
+                name = (char*)sync["name"].valuestr();
+            }
+            if (sync.hasElement("_id")) {
+                id = sync["_id"].Number();
+            }
+            if (sync.hasElement("optime")) {
+                isOpTime = true;
+                optime = sync["optime"]._opTime();
+            }
         }
-        sethbmsg( str::stream() << "syncing to primary: " << sp.primary->fullName(), 0);
-        return const_cast<Member*>(sp.primary);
+        
+        for( Member *m = head(); m; m = m->next() ) {
+            if (!m->hbinfo().up() ||
+                (m->state() != MemberState::RS_SECONDARY &&
+                 m->state() != MemberState::RS_PRIMARY) ||
+                (secondaryOnly && m->state() != MemberState::RS_SECONDARY) ||
+                (id != -1 && (int)m->id() != id) ||
+                (name != 0 && strcmp(name, m->fullName().c_str()) != 0) ||
+                (isOpTime && optime >= m->hbinfo().opTime)) {
+                continue;
+            }
+
+            sethbmsg( str::stream() << "syncing to: " << m->fullName(), 0);
+            return const_cast<Member*>(m);
+        }
+        
+        sethbmsg( str::stream() << "couldn't find a member matching the sync criteria: " <<
+                  "\nstate? " << (secondaryOnly ? "2" : "none") <<
+                  "\nname? " << (name ? name : "none") <<
+                  "\n_id? " << id <<
+                  "\noptime? " << optime.toStringPretty() );
+        
+        return NULL;
     }
     
     /**