We have a few commands that "can't run in a transaction block",
meaning that if they complete their processing but then we fail
to COMMIT, we'll be left with inconsistent on-disk state.
However, the existing defenses for this are only watertight for
simple query protocol. In extended protocol, we didn't commit
until receiving a Sync message. Since the client is allowed to
issue another command instead of Sync, we're in trouble if that
command fails or is an explicit ROLLBACK. In any case, sitting
in an inconsistent state while waiting for a client message
that might not come seems pretty risky.
This case wasn't reachable via libpq before we introduced pipeline
mode, but it's always been an intended aspect of extended query
protocol, and likely there are other clients that could reach it
before.
To fix, set a flag in PreventInTransactionBlock that tells
exec_execute_message to force an immediate commit. This seems
to be the approach that does least damage to existing working
cases while still preventing the undesirable outcomes.
While here, add some documentation to protocol.sgml that explicitly
says how to use pipelining. That's latent in the existing docs if
you know what to look for, but it's better to spell it out; and it
provides a place to document this new behavior.
Per bug #17434 from Yugo Nagata. It's been wrong for ages,
so back-patch to all supported branches.
Discussion: https://postgr.es/m/17434-
d9f7a064ce2a88a3@postgresql.org
+
+
Pipelining
+
+
+ protocol specification
+
+
+ Use of the extended query protocol
+ allows pipelining, which means sending a series
+ of queries without waiting for earlier ones to complete. This reduces
+ the number of network round trips needed to complete a given series of
+ operations. However, the user must carefully consider the required
+ behavior if one of the steps fails, since later queries will already
+ be in flight to the server.
+
+
+ One way to deal with that is to make the whole query series be a
+ single transaction, that is wrap it in BEGIN ...
+ COMMIT. However, this does not help if one wishes
+ for some of the commands to commit independently of others.
+
+
+ The extended query protocol provides another way to manage this
+ concern, which is to omit sending Sync messages between steps that
+ are dependent. Since, after an error, the backend will skip command
+ messages until it finds Sync, this allows later commands in a pipeline
+ to be skipped automatically when an earlier one fails, without the
+ client having to manage that explicitly with BEGIN
+ and COMMIT. Independently-committable segments
+ of the pipeline can be separated by Sync messages.
+
+
+ If the client has not issued an explicit BEGIN,
+ then each Sync ordinarily causes an implicit COMMIT
+ if the preceding step(s) succeeded, or an
+ implicit ROLLBACK if they failed. However, there
+ are a few DDL commands (such as CREATE DATABASE)
+ that cannot be executed inside a transaction block. If one of
+ these is executed in a pipeline, it will, upon success, force an
+ immediate commit to preserve database consistency.
+ A Sync immediately following one of these has no effect except to
+ respond with ReadyForQuery.
+
+
+ When using this method, completion of the pipeline must be determined
+ by counting ReadyForQuery messages and waiting for that to reach the
+ number of Syncs sent. Counting command completion responses is
+ unreliable, since some of the commands may not be executed and thus not
+ produce a completion message.
+
+
+
Function Call
* could issue more commands and possibly cause a failure after the statement
* completes). Subtransactions are verboten too.
*
+ * We must also set XACT_FLAGS_NEEDIMMEDIATECOMMIT in MyXactFlags, to ensure
+ * that postgres.c follows through by committing after the statement is done.
+ *
* isTopLevel: passed down from ProcessUtility to determine whether we are
* inside a function. (We will always fail if this is false, but it's
* convenient to centralize the check here instead of making callers do it.)
if (CurrentTransactionState->blockState != TBLOCK_DEFAULT &&
CurrentTransactionState->blockState != TBLOCK_STARTED)
elog(FATAL, "cannot prevent transaction chain");
- /* all okay */
+
+ /* All okay. Set the flag to make sure the right thing happens later. */
+ MyXactFlags |= XACT_FLAGS_NEEDIMMEDIATECOMMIT;
}
/*
CurrentTransactionState->blockState != TBLOCK_STARTED)
return true;
+ /*
+ * If we tell the caller we're not in a transaction block, then inform
+ * postgres.c that it had better commit when the statement is done.
+ * Otherwise our report could be a lie.
+ */
+ MyXactFlags |= XACT_FLAGS_NEEDIMMEDIATECOMMIT;
+
return false;
}
}
else
{
+ /*
+ * We had better not see XACT_FLAGS_NEEDIMMEDIATECOMMIT set if
+ * we're not calling finish_xact_command(). (The implicit
+ * transaction block should have prevented it from getting set.)
+ */
+ Assert(!(MyXactFlags & XACT_FLAGS_NEEDIMMEDIATECOMMIT));
+
/*
* We need a CommandCounterIncrement after every query, except
* those that start or end a transaction block.
/*
* We must copy the sourceText and prepStmtName into MessageContext in
- * case the portal is destroyed during finish_xact_command. Can avoid the
- * copy if it's not an xact command, though.
+ * case the portal is destroyed during finish_xact_command. We do not
+ * make a copy of the portalParams though, preferring to just not print
+ * them in that case.
*/
- if (is_xact_command)
- {
- sourceText = pstrdup(portal->sourceText);
- if (portal->prepStmtName)
- prepStmtName = pstrdup(portal->prepStmtName);
- else
- prepStmtName = "";
-
- /*
- * An xact command shouldn't have any parameters, which is a good
- * thing because they wouldn't be around after finish_xact_command.
- */
- portalParams = NULL;
- }
+ sourceText = pstrdup(portal->sourceText);
+ if (portal->prepStmtName)
+ prepStmtName = pstrdup(portal->prepStmtName);
else
- {
- sourceText = portal->sourceText;
- if (portal->prepStmtName)
- prepStmtName = portal->prepStmtName;
- else
- prepStmtName = "";
- portalParams = portal->portalParams;
- }
+ prepStmtName = "";
+ portalParams = portal->portalParams;
/*
* Report query to various monitoring facilities.
if (completed)
{
- if (is_xact_command)
+ if (is_xact_command || (MyXactFlags & XACT_FLAGS_NEEDIMMEDIATECOMMIT))
{
/*
* If this was a transaction control statement, commit it. We
* will start a new xact command for the next command (if any).
+ * Likewise if the statement required immediate commit. Without
+ * this provision, we wouldn't force commit until Sync is
+ * received, which creates a hazard if the client tries to
+ * pipeline immediate-commit statements.
*/
finish_xact_command();
+
+ /*
+ * These commands typically don't have any parameters, and even if
+ * one did we couldn't print them now because the storage went
+ * away during finish_xact_command. So pretend there were none.
+ */
+ portalParams = NULL;
}
else
{
*/
#define XACT_FLAGS_ACQUIREDACCESSEXCLUSIVELOCK (1U << 1)
+/*
+ * XACT_FLAGS_NEEDIMMEDIATECOMMIT - records whether the top level statement
+ * is one that requires immediate commit, such as CREATE DATABASE.
+ */
+#define XACT_FLAGS_NEEDIMMEDIATECOMMIT (1U << 2)
+
/*
* start- and end-of-transaction callbacks for dynamically loaded modules
*/
projects / postgresql.git / commitdiff