From: John David Duncan
Date: September 26 2011 9:16pm
Subject: bzr push into mysql-5.5-cluster branch (john.duncan:3554 to 3555)
List-Archive: http://lists.mysql.com/commits/141147
Message-Id: <201109262116.p8QLGJOh017769@acsmt356.oracle.com>
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-Transfer-Encoding: 7bit

 3555 John David Duncan	2011-09-26
      Fix for race condition in ndb/memcache S scheduler.
      It was possible for an ndb_send_thread to start up and try to use a WorkerConnection that the main thread had not yet initialized.  
      This should fix http://pb2.no.oracle.com/?template=mysql_show_test_failure&search=yes&push_id=2586144&test_id=2586293&test_run=n_mix&test_suite=ndb_memcache&test_case=basic

    modified:
      storage/ndb/memcache/src/schedulers/S_sched.cc
      storage/ndb/memcache/src/schedulers/S_sched.h
 3554 John David Duncan	2011-09-25
      work on const-conversion issues found by stricter compiler

    modified:
      storage/ndb/memcache/include/DataTypeHandler.h
      storage/ndb/memcache/src/DataTypeHandler.cc
      storage/ndb/memcache/src/Record.cc
=== modified file 'storage/ndb/memcache/src/schedulers/S_sched.cc'
--- a/storage/ndb/memcache/src/schedulers/S_sched.cc	2011-09-22 18:27:10 +0000
+++ b/storage/ndb/memcache/src/schedulers/S_sched.cc	2011-09-26 21:12:27 +0000
@@ -95,6 +95,10 @@ void S::SchedulerGlobal::init(int _nthre
       * wc_cell = new WorkerConnection(this, t, c);
     }
   }
+    
+  /* Start the send & poll threads for each connection */
+  for(int i = 0 ; i < nclusters ; i++) 
+    clusters[i]->startThreads();
   
   /* Log message for startup */
   logger->log(LOG_WARNING, 0, "Scheduler: starting for %d cluster%s; "
@@ -430,7 +434,8 @@ void S::SchedulerWorker::add_stats(const
 /* Cluster methods */
 S::Cluster::Cluster(SchedulerGlobal *global, int _id) : 
   cluster_id(_id), 
-  nreferences(0)
+  nreferences(0),
+  threads_started(false)
 {
   DEBUG_PRINT("%d", cluster_id);
   
@@ -478,6 +483,18 @@ S::Cluster::Cluster(SchedulerGlobal *glo
 }
 
 
+void S::Cluster::startThreads() {
+  /* Threads are started only once and persist across reconfiguration.
+     But, this method will be called again for each reconf. */
+  if(threads_started == false) {
+    for(int i = 0 ; i < nconnections; i++) {
+      connections[i]->startThreads();
+    }
+    threads_started = true;
+  }
+}
+
+
 S::Cluster::~Cluster() {
   DEBUG_PRINT("Shutting down cluster %d", cluster_id);
   for(int i = 0; i < nconnections ; i++) {
@@ -626,14 +643,17 @@ S::Connection::Connection(S::Cluster & _
   /* Initialize the queues for sent and resceduled items */
   sentqueue = new Queue<NdbInstance>(nInst);
   reschedulequeue = new Queue<NdbInstance>(nInst);
-    
+}
+
+
+void S::Connection::startThreads() {
   /* Start the poll thread */
   pthread_create( & poll_thread_id, NULL, run_poll_thread, (void *) this);
-
+  
   /* Start the send thread */
   pthread_create( & send_thread_id, NULL, run_send_thread, (void *) this);
 }
-
+  
 
 S::Connection::~Connection() {
   /* Shut down a connection. 

=== modified file 'storage/ndb/memcache/src/schedulers/S_sched.h'
--- a/storage/ndb/memcache/src/schedulers/S_sched.h	2011-09-12 10:05:07 +0000
+++ b/storage/ndb/memcache/src/schedulers/S_sched.h	2011-09-26 21:12:27 +0000
@@ -121,7 +121,9 @@ public:
   ~Cluster();
   void add_stats(const char *, ADD_STAT, const void *);
   WorkerConnection ** getWorkerConnectionPtr(int thd) const;  
+  void startThreads();
    
+  bool threads_started;
   int cluster_id;
   int nconnections;
   int nreferences;
@@ -140,6 +142,10 @@ public:
   Connection(Cluster &, int connection_id);
   ~Connection();
   void add_stats(const char *, ADD_STAT, const void *);
+  void startThreads();
+
+  /* These are not intended to be part of the public API, but are marked as
+     public so that they can be called from C code in pthread_create(): */
   void * run_ndb_send_thread();
   void * run_ndb_poll_thread();
 

No bundle (reason: useless for push emails).