I haven't produced a real problem in a small test case (yet), but I
convinced myself that it's wrong enough to be called a bug.
If you add the following assertion to BufferAlloc:
Assert(relpersistence == RELPERSISTENCE_PERMANENT ||
relpersistence == RELPERSISTENCE_UNLOGGED ||
relpersistence == RELPERSISTENCE_TEMP);
(which seems like a reasonable assertion to me), then do:
create table foo(i int);
insert into foo values(1);
vacuum foo;
vacuum foo;
insert into foo values(2);
Then do an immediate shutdown, then restart, you hit the assertion.
The problem is using CreateFakeRelcacheEntry, which has the following
comment:
Only the fields related to physical storage, like rd_rel, are
initialized, so the fake entry is only usable in low-level
operations like ReadBuffer().
That doesn't seem right, because ReadBuffer needs relpersistence set.
One of the consequences is that you can get buffers with the wrong flags
set; in particular, missing BM_PERMANENT, which seems like it could be a
serious problem.
Are there other areas where we might have similar problems?
Regards,
Jeff Davis
