From: Pavan Deolasee <pavan.deolasee@gmail.com>
Date: Thu, 5 May 2016 05:32:04 +0000 (+0530)
Subject: Add a ref count mechanism to deal with situations where a Shared Queue is
X-Git-Tag: XL9_5_R1_1~12
X-Git-Url: http://git.postgresql.org/gitweb/static/gitweb.js?a=commitdiff_plain;h=a7995b41947717123fa365dee50c9feb30cd5a40;p=postgres-xl.git

Add a ref count mechanism to deal with situations where a Shared Queue is
acquired but never bound by any of the node, thus causing leakage

To be honest, this area requires further work. The way things are currently
setup, producer and consumers all bind to a shared queue, but only producer
eventually unbinds. The implementation has logic to wait out for consumers
before destroying a shared queue. While this is okay, a more defined entry and
exit points are required for producer and consumers.

The code also today relies on timeouts to handle the case where a consumer
never binds to a shared queue, thus causing large delays. These delays are more
prominent for very short queries.
---

diff --git a/src/backend/pgxc/squeue/squeue.c b/src/backend/pgxc/squeue/squeue.c
index 09700c1af3..0d2844aa5c 100644
--- a/src/backend/pgxc/squeue/squeue.c
+++ b/src/backend/pgxc/squeue/squeue.c
@@ -104,6 +104,7 @@ typedef struct SQueueHeader
 	int			sq_pid; 		/* Process id of the producer session */
 	int			sq_nodeid;		/* Node id of the producer parent */
 	SQueueSync *sq_sync;        /* Associated sinchronization objects */
+	int			sq_refcnt;		/* Reference count to this entry */
 #ifdef SQUEUE_STAT
 	bool		stat_finish;
 	long		stat_paused;
@@ -287,6 +288,7 @@ tryagain:
 		/* Initialize the shared queue */
 		sq->sq_pid = 0;
 		sq->sq_nodeid = -1;
+		sq->sq_refcnt = 1;
 #ifdef SQUEUE_STAT
 		sq->stat_finish = false;
 		sq->stat_paused = 0;
@@ -401,8 +403,8 @@ tryagain:
 							" %d tries", trycount);
 				goto tryagain;
 			}
-
 		}
+		sq->sq_refcnt++;
 	}
 	LWLockRelease(SQueuesLock);
 }
@@ -521,6 +523,18 @@ SharedQueueBind(const char *sqname, List *consNodes,
 
 		if (myindex)
 			*myindex = -1;
+
+		/*
+		 * Increment the refcnt only when producer binds. This is a bit
+		 * asymmetrical, but the way things are currently setup, a consumer
+		 * though calls SharedQueueBind, never calls SharedQueueUnBind. The
+		 * unbinding is done only by the producer after it waits for all
+		 * consumers to finish.
+		 *
+		 * XXX This ought to be fixed someday to simplify things in Shared
+		 * Queue handling
+		 */ 
+		sq->sq_refcnt++;
 	}
 	else
 	{
@@ -1433,6 +1447,15 @@ CHECK:
 		LWLockRelease(SQueuesLock);
 		goto CHECK;
 	}
+
+	/*
+	 * XXX Decrement the refcnt, but it doesn't really matter because we are
+	 * unconditionally removing the SQueue anyways. SharedQueueRelease is
+	 * prepared to work with already removed SQueue
+	 *
+	 * This ought to be fixed someday
+	 */
+	squeue->sq_refcnt--;
 	
 	/* All is done, clean up */
 	DisownLatch(&sqsync->sqs_producer_latch);
@@ -1481,8 +1504,7 @@ SharedQueueRelease(const char *sqname)
 		if (sq->sq_nodeid == -1)
 		{
 			elog(DEBUG1, "SQueue %s, producer not bound ", sqname);
-			LWLockRelease(SQueuesLock);
-			return;
+			goto done;
 		}
 
 		/*
@@ -1549,8 +1571,7 @@ SharedQueueRelease(const char *sqname)
 					}
 					LWLockRelease(sqsync->sqs_consumer_sync[i].cs_lwlock);
 					/* exit */
-					LWLockRelease(SQueuesLock);
-					return;
+					goto done;
 				}
 			}
 
@@ -1578,6 +1599,19 @@ SharedQueueRelease(const char *sqname)
 			}
 		}
 	}
+done:
+	/*
+	 * If we are the last holder of the SQueue, remove it from the hash table
+	 * to avoid any leak
+	 */
+	if (sq && --sq->sq_refcnt == 0)
+	{
+		/* Now it is OK to remove hash table entry */
+		sq->sq_sync->queue = NULL;
+		sq->sq_sync = NULL;
+		if (hash_search(SharedQueues, sq->sq_key, HASH_REMOVE, NULL) != sq)
+			elog(PANIC, "Shared queue data corruption");
+	}
 	LWLockRelease(SQueuesLock);
 }