In 9.1 and earlier, it is possible for index_getnext() to try to examine
a heap buffer for possible HOT-prune when in recovery; this causes a
problem when a multixact is found in a tuple's Xmax, because
GetMultiXactIdMembers refuses to run when in recovery, raising an error:
ERROR: cannot GetMultiXactIdMembers() during recovery
This can be solved easily by having MultiXactIdIsRunning always return
false when in recovery, which is reasonable because a HOT standby cannot
acquire further tuple locks nor update/delete tuples.
(Note: it doesn't look like this specific code path has a problem in
9.2, because instead of doing HeapTupleSatisfiesUpdate directly,
heap_hot_search_buffer uses HeapTupleIsSurelyDead instead. Still, there
may be other paths affected by the same bug, for instance in pgrowlocks,
and the multixact code hasn't changed; so apply the same fix
throughout.)
Apply this fix to 9.0 through 9.2. In 9.3 the multixact code has been
changed completely and is no longer subject to this problem.
Per report from Marko Tiikkaja,
https://www.postgresql.org/message-id/
54EB3283.
2080305@joh.to
Analysis by Andres Freund
debug_elog3(DEBUG2, "IsRunning %u?", multi);
+ /*
+ * During recovery, all multixacts can be considered not running: in
+ * effect, tuple locks are not held in standby servers, which is fine
+ * because the standby cannot acquire further tuple locks nor update/delete
+ * tuples.
+ *
+ * We need to do this first, because GetMultiXactIdMembers complains if
+ * called on recovery.
+ */
+ if (RecoveryInProgress())
+ {
+ debug_elog2(DEBUG2, "IsRunning: in recovery");
+ return false;
+ }
+
nmembers = GetMultiXactIdMembers(multi, &members);
if (nmembers < 0)