bzr commit into mysql-6.0-falcon branch (vvaintroub:2799) Bug#38947

[Vladislav Vaintroub]

#At file:///C:/bzr/mysql-6.0-falcon-team/

 2799 Vladislav Vaintroub	2008-08-26
      Bug #38947 UPDATE threads in endless Table::fetchForUpdate loop = livelock 
      
      Problem: in Table::fetchForUpdate() there is a small possibility for a 
      race condition -  if record belongs to transaction that is being committed
      currently and the state of this transaction is still Active, but syncActive
      is already unlocked. This causes re-fetch() in the fetch thread without 
      any wait, instead of waiting for falcon_lock_wait_timeout seconds.
      
      This is fixed by moving signaling  waiters via syncActive.unlock(), 
      after transaction state has changed from active to committed.
modified:
  storage/falcon/Transaction.cpp

=== modified file 'storage/falcon/Transaction.cpp'
--- a/storage/falcon/Transaction.cpp	2008-08-25 18:24:59 +0000
+++ b/storage/falcon/Transaction.cpp	2008-08-26 15:36:02 +0000
@@ -275,7 +275,6 @@ void Transaction::commit()
 		releaseRecordLocks();
 
 	database->serialLog->preCommit(this);
-	syncActive.unlock();
 
 	
 
@@ -303,10 +302,9 @@ void Transaction::commit()
 	database->flushInversion(this);
 
 	// Transfer transaction from active list to committed list, set committed state
-
-	Sync syncCommitted(&transactionManager->committedTransactions.syncObject,
"Transaction::commit(2)");
 	Sync syncActiveTransactions(&transactionManager->activeTransactions.syncObject,
"Transaction::commit(3)");
-
+	Sync syncCommitted(&transactionManager->committedTransactions.syncObject,
"Transaction::commit(2)");
+	
 	syncActiveTransactions.lock(Exclusive);
 	syncCommitted.lock(Exclusive);
 
@@ -316,6 +314,8 @@ void Transaction::commit()
 
 	syncCommitted.unlock();
 	syncActiveTransactions.unlock();
+	
+	syncActive.unlock(); // signal waiting transactions
 
 	database->commit(this);