1 files changed, 97 insertions, 0 deletions
diff --git a/ndb/src/kernel/blocks/Start.txt b/ndb/src/kernel/blocks/Start.txt
new file mode 100644
index 00000000000..545296d44f1
--- /dev/null
+++ b/ndb/src/kernel/blocks/Start.txt
@@ -0,0 +1,97 @@
+
+--- Start phase 1 - Qmgr -------------------------------------------
+
+1) Set timer 1 - TimeToWaitAlive
+
+2) Send CM_REGREQ to all connected(and connecting) nodes
+
+3) Wait until -
+a) The precident answers CM_REGCONF
+b) All nodes has answered and I'm the candidate -> election won
+c) 30s has passed and I'm the candidate -> election won
+d) TimeToWaitAlive has passed -> Failure to start
+
+When receiving CM_REGCONF
+4) Send CM_NODEINFOREQ to all connected(and connecting) nodes 
+   reported in CM_REGCONF
+
+5) Wait until -
+a) All CM_NODEINFO_CONF has arrived
+b) TimeToWaitAlive has passed -> Failure to start
+
+6) Send CM_ACKADD to president
+
+7) Wait until - 
+a) Receive CM_ADD(CommitNew) from president -> I'm in the qmgr cluster
+b) TimeToWaitAlive has passed -> Failure to start
+
+NOTE:
+30s is hardcoded in 3c.
+TimeToWaitAlive should be atleast X sec greater than 30s. i.e. 30+X sec
+to support "partial starts"
+
+NOTE:
+In 3b, a more correct number (instead of all) would be
+N-NG+1 where N is #nodes and NG is #node groups = (N/R where R is # replicas)
+But Qmgr has no notion about node groups or replicas
+
+--- Start phase X - Qmgr -------------------------------------------
+
+President - When accepting a CM_REGREQ
+1) Send CM_REGCONF to starting node
+2) Send CM_ADD(Prepare) to all started nodes + starting node
+3) Send CM_ADD(AddCommit) to all started nodes
+4) Send CM_ADD(CommitNew) to starting node
+
+Cluster participant -
+1) Wait for both CM_NODEINFOREQ from starting and CM_ADD(Prepare) from pres.
+2) Send CM_ACKADD(Prepare)
+3) Wait for CM_ADD(AddCommit) from president
+4) Send CM_ACKADD(AddCommit) 
+ 
+--- Start phase 2 - NdbCntr ----------------------------------------
+
+- Use same TimeToWaitAliveTimer
+
+1) Check sysfile (DIH_RESTART_REQ)
+2) Read nodes (from Qmgr) P = qmgr president
+
+3) Send CNTR_MASTER_REQ to cntr(P)
+   including info in DIH_RESTART_REF/CONF
+
+4) Wait until -
+a) Receiving CNTR_MASTER_CONF -> continue
+b) Receiving CNTR_MASTER_REF -> P = node specified in REF, goto 3
+c) TimeToWaitAlive has passed -> Failure to start
+
+4) Run ndb-startphase 1
+
+--
+Initial start/System restart NdbCntr (on qmgr president node)
+
+1) Wait until -
+a) Receiving all CNTR_MASTER_REQ (all = those in READ_NODES_CONF)
+b) TimeToWaitAlive has passed -> Failure to start
+
+2) Wait until -
+a) Enough nodes (at least 1 in each node group and 1 full node group)
+   has sent me CNTR_MASTER_REQ
+b) TimeToWaitAlive has passed -> Failure to start
+
+3) Decide what kind of start to perform (initial / system restart)
+   Decide who should be the master (the one with greatest GCI)
+   Send CNTR_MASTER_CONF(initial/system restart) to all nodes included in start
+
+-- 
+Running NdbCntr
+
+When receiving CNTR_MASTER_REQ
+1) If I'm not master send CNTR_MASTER_REF (including master node id)
+2) If I'm master
+   Coordinate parallell node restarts
+   send CNTR_MASTER_CONF (node restart)
+
+NOTE:
+2a Specified with a command line/config parameter the system could 
+   start using only one node in each node group (if possible w.r.t LCP/GCP)
+