From: Petr Jelinek <pjmodos@pjmodos.net>
Date: Fri, 1 May 2015 13:23:40 +0000 (+0200)
Subject: Don't immediately kill all transactions when acquiring DDL lock.
X-Git-Url: http://git.postgresql.org/gitweb/static/gitweb.js?a=commitdiff_plain;h=0e02cc27d58d8e4fd2684a4e39036a72e4323eec;p=2ndquadrant_bdr.git

Don't immediately kill all transactions when acquiring DDL lock.

Previously we just canceled all transactions in database when acking DDL
lock. This code only cancels transactions which did any writes and gives
them some time (currently hardcoded to 10s) of grace period to finish
what they are doing.
---

diff --git a/bdr_locks.c b/bdr_locks.c
index 7dba389140..3c66282091 100644
--- a/bdr_locks.c
+++ b/bdr_locks.c
@@ -100,6 +100,7 @@
 #include "storage/proc.h"
 #include "storage/procarray.h"
 #include "storage/shmem.h"
+#include "storage/sinvaladt.h"
 #include "storage/standby.h"
 
 #include "utils/builtins.h"
@@ -154,6 +155,8 @@ static bool this_xact_acquired_lock = false;
 
 /* GUCs */
 bool bdr_permit_ddl_locking = false;
+int bdr_ddl_grace_timeout = 10000;
+
 
 static size_t
 bdr_locks_shmem_size(void)
@@ -700,6 +703,67 @@ check_is_my_node(uint64 sysid, TimeLineID tli, Oid datid)
 	return true;
 }
 
+/*
+ * Kill any writing transactions while giving them some grace period for
+ * finishing.
+ *
+ * Caller is responsible for ensuring that no new writes can be started during
+ * the execution of this function.
+ */
+static void
+cancel_conflicting_transactions(void)
+{
+	VirtualTransactionId *conflict;
+	TimestampTz		endtime = TimestampTzPlusMilliseconds(GetCurrentTimestamp(), bdr_ddl_grace_timeout),
+					waittime = 1000;
+
+	conflict = GetConflictingVirtualXIDs(InvalidTransactionId, MyDatabaseId);
+
+	while (conflict->backendId != InvalidBackendId)
+	{
+		PGPROC	   *pgproc = BackendIdGetProc(conflict->backendId);
+		PGXACT	   *pgxact;
+
+		if (pgproc == NULL)
+			continue;
+
+		pgxact = &ProcGlobal->allPgXact[pgproc->pgprocno];
+
+		/* Skip the transactions that didn't do any writes. */
+		if (!TransactionIdIsValid(pgxact->xid))
+		{
+			conflict++;
+			continue;
+		}
+
+		/* If here is writing transaction give it time to finish */
+		if (GetCurrentTimestamp() < endtime)
+		{
+			/* Increasing backoff interval for wait time with limit of 1s */
+			pg_usleep(waittime);
+			waittime *= 2;
+			if (waittime > 1000000)
+				waittime = 1000000;
+		}
+		else
+		{
+			/* We reached timeout so lets kill the writing transaction */
+			pid_t p = CancelVirtualTransaction(*conflict, PROCSIG_RECOVERY_CONFLICT_LOCK);
+
+			/*
+			 * Either confirm kill or sleep a bit to prevent the other node
+			 * being busy with signal processing.
+			 */
+			if (p == 0)
+				conflict++;
+			else
+				pg_usleep(1000);
+
+			elog(DEBUG2, "signaled pid %d to terminate because it conflicts with a DDL lock requested by another node", p);
+		}
+	}
+}
+
 /*
  * Another node has asked for a DDL lock. Try to acquire the local ddl lock.
  *
@@ -729,7 +793,6 @@ bdr_process_acquire_ddl_lock(uint64 sysid, TimeLineID tli, Oid datid)
 	if (bdr_my_locks_database->lockcount == 0)
 	{
 		XLogRecPtr wait_for_lsn;
-		VirtualTransactionId *conflicts;
 		Relation rel;
 		Datum	values[10];
 		bool	nulls[10];
@@ -797,41 +860,14 @@ bdr_process_acquire_ddl_lock(uint64 sysid, TimeLineID tli, Oid datid)
 		/*
 		 * Now kill all local processes that are still writing. We can't just
 		 * prevent them from writing via the acquired lock as they are still
-		 * running. We're using drastic measures here because it'd be a bad
-		 * idea to wait: The primary is waiting for us and during that time
-		 * the entire flock has to wait.
+		 * running.
 		 *
 		 * TODO: It'd be *far* nicer to only cancel other transactions if they
 		 * held conflicting locks, but that's not easiliy possible at the
 		 * moment.
 		 */
 		elog(DEBUG1, "terminating any local processes that conflict with the DDL lock");
-		conflicts = GetConflictingVirtualXIDs(InvalidTransactionId, MyDatabaseId);
-		while (conflicts->backendId != InvalidBackendId)
-		{
-			pid_t p;
-
-			/* Don't kill ourselves */
-			if (conflicts->backendId == MyBackendId)
-			{
-				conflicts++;
-				continue;
-			}
-
-			/* try to kill */
-			p = CancelVirtualTransaction(*conflicts, PROCSIG_RECOVERY_CONFLICT_LOCK);
-
-			/*
-			 * Either confirm kill or sleep a bit to prevent the other node
-			 * being busy with signal processing.
-			 */
-			if (p == 0)
-				conflicts++;
-			else
-				pg_usleep(5000);
-
-			elog(DEBUG2, "signaled pid %d to terminate because it conflicts with a DDL lock requested by another node", p);
-		}
+		cancel_conflicting_transactions();
 
 		/*
 		 * We now have to wait till all our local pending changes have been