While running a parallel regression test I saw a failure"relation deleted while still in use"
out of the subtransactions test, after which the backend dumped core.
The coredump was in AtSubAbort_smgr, and it was failing because
upperPendingDeletes was NIL (which it should not have been inside a
subtransaction, of course).
The underlying problem doesn't seem very reproducible --- I tried
several times without seeing it again. But what I deduce from the core
dump is that we had an error during subtransaction cleanup, leading to
an attempt to re-abort an already partially aborted transaction, and
the system just could not cope. The problem is that the list-of-lists
data structure used in smgr.c is brittle: it won't survive two
executions of AtSubAbort_smgr at the same nesting level.
In my local copy I've removed the list-of-lists data structure and
reverted to a flat list of pending deletion requests, with the addition
of a transaction nesting level field in each entry. AtSubCommit_smgr
does this:
int nestLevel = GetCurrentTransactionNestLevel(); PendingRelDelete *pending;
for (pending = pendingDeletes; pending != NULL; pending = pending->next) { if (pending->nestLevel >=
nestLevel) pending->nestLevel = nestLevel - 1; }
while AtSubAbort_smgr acts only on entries with nestlevel >= transaction
nest level. This makes the data structure safe against multiple
execution.
I think we will have to make similar adjustments in the other places
where we've used list-of-lists data structures. We may need to look a
little closer at the manipulations of the transaction state stack in
xact.c, as well.
regards, tom lane
