Fix failure to delete spill files of aborted transactions

alvherre · alvherre · commit f68c49f86aea · 2018-01-05T12:17:10.000-03:00
Logical decoding's reorderbuffer.c may spill transaction files to disk when transactions are large. These are supposed to be removed when they become "too old" by xid; but file removal requires the boundary LSNs of the transaction to be known. The final_lsn is only set when we see the commit or abort record for the transaction, but nothing sets the value for transactions that crash, so the removal code misbehaves -- in assertion-enabled builds, it crashes by a failed assertion. To fix, modify the final_lsn of transactions that don't have a value set, to the LSN of the very latest change in the transaction. This causes the spilled files to be removed appropriately. Author: Atsushi Torikoshi Reviewed-by: Kyotaro HORIGUCHI, Craig Ringer, Masahiko Sawada Discussion: https://postgr.es/m/54e4e488-186b-a056-6628-50628e4e4ebc@lab.ntt.co.jp
diff --git a/src/backend/replication/logical/reorderbuffer.c b/src/backend/replication/logical/reorderbuffer.c
@@ -1617,8 +1617,8 @@ ReorderBufferAbortOld(ReorderBuffer *rb, TransactionId oldestRunningXid)
 	 * Iterate through all (potential) toplevel TXNs and abort all that are
 	 * older than what possibly can be running. Once we've found the first
 	 * that is alive we stop, there might be some that acquired an xid earlier
-	 * but started writing later, but it's unlikely and they will cleaned up
-	 * in a later call to ReorderBufferAbortOld().
+	 * but started writing later, but it's unlikely and they will be cleaned
+	 * up in a later call to this function.
 	 */
 	dlist_foreach_modify(it, &rb->toplevel_by_lsn)
 	{
@@ -1628,6 +1628,21 @@ ReorderBufferAbortOld(ReorderBuffer *rb, TransactionId oldestRunningXid)
 
 		if (TransactionIdPrecedes(txn->xid, oldestRunningXid))
 		{
+			/*
+			 * We set final_lsn on a transaction when we decode its commit or
+			 * abort record, but we never see those records for crashed
+			 * transactions.  To ensure cleanup of these transactions, set
+			 * final_lsn to that of their last change; this causes
+			 * ReorderBufferRestoreCleanup to do the right thing.
+			 */
+			if (txn->serialized && txn->final_lsn == 0)
+			{
+				ReorderBufferChange *last =
+					dlist_tail_element(ReorderBufferChange, node, &txn->changes);
+
+				txn->final_lsn = last->lsn;
+			}
+
 			elog(DEBUG2, "aborting old transaction %u", txn->xid);
 
 			/* remove potential on-disk data, and deallocate this tx */
diff --git a/src/include/replication/reorderbuffer.h b/src/include/replication/reorderbuffer.h
@@ -153,6 +153,8 @@ typedef struct ReorderBufferTXN
 	 * * plain abort record
 	 * * prepared transaction abort
 	 * * error during decoding
+	 * * for a crashed transaction, the LSN of the last change, regardless of
+	 *   what it was.
 	 * ----
 	 */
 	XLogRecPtr	final_lsn;
