PostgreSQL

The planner could create an incorrect plan in cases where two ProjectionPaths were stacked on top of each other. The only known way to trigger that situation involves parallel sort operations, but there may be other instances. The result would be crashes or incorrect query results. Disclosure of server memory contents is also possible. (CVE-2021-3677)

SSL renegotiation has been disabled for some time, but the server would still cooperate with a client-initiated renegotiation request. A maliciously crafted renegotiation request could result in a server crash (see OpenSSL issue CVE-2021-3449). Disable the feature altogether on OpenSSL versions that permit doing so, which are 1.1.0h and newer.

This change fixes cases where an attempt to fetch a toasted value immediately after COMMIT/ROLLBACK would fail with errors like “[.quote]#no known snapshots”# or “[.quote]#missing chunk number 0 for toast value”#.

Some extensions may attempt to execute SQL code outside of any Portal. They are responsible for ensuring that an outer snapshot exists before doing so. Previously, not providing a snapshot might work or it might not; now it will consistently fail with “[.quote]#cannot execute SQL without an outer snapshot or portal”#.

Previously, we’d always rewind and re-read the whole query result, possibly getting results different from the earlier execution, causing great confusion later. For a NO SCROLL cursor, we can fix this by only storing the not-yet-read portion of the query output, which is sufficient since a NO SCROLL cursor can’t be backed up. Cursors with the SCROLL option remain at hazard, but that was already documented to be an unsafe option to use with a non-stable query. Make those documentation warnings stronger.

Also force NO SCROLL mode for the implicit cursor used by a PL/pgSQL FOR-over-query loop, to avoid this type of problem when persisting such a cursor during an intra-procedure commit.

This should be disallowed, just as FOR UPDATE with a plain GROUP BY is disallowed, but the test for that failed to handle empty grouping sets correctly. The end result would be a null-pointer dereference in the executor.

Such cases previously crashed.

The previous coding allowed ALTER EXTENSION ADD/DROP to occur concurrently with DROP EXTENSION, leading to a crash or corrupt catalog entries.

This command failed on materialized views containing columns with certain names, notably mv and newdata.

A transaction cannot be prepared if it has both session-lifespan and transaction-lifespan locks on the same advisory-lock ID value. This restriction was not fully checked, which could lead to a PANIC during PREPARE TRANSACTION.

Notably, this fixes some cases where it was necessary to be a superuser to use DROP OWNED BY.

Use of a whole-row variable clearly violates the rule that a generated column cannot depend on itself, so such cases have no well-defined behavior. The actual behavior frequently included a crash.

Some code paths failed to provide a valid value for this system column while evaluating a GENERATED expression.

The fast default is useless since no local heap storage exists for such a table, but it confused subsequent operations. In addition to suppressing creation of such catalog entries in ALTER TABLE commands, adjust the downstream code to cope when one is incorrectly present.

This avoids failures when dealing with index predicates that aren’t really immutable. While that’s not considered a supported case, the original reason for using a non-transactional update here is long gone, so we may as well change it.

Previously, a walsender would show its latest SQL command, which was confusing if it’s now doing some replication operation instead. Now we show replication-protocol commands on the same footing as SQL commands.

pending_restart correctly showed the case where an entry that cannot be changed without a postmaster restart has been modified, but not where the entry had been removed altogether.

SELECT FOR UPDATE and related commands would fail with assertion failures or “[.quote]#could not find junk column”# errors in such cases.

Under a narrow combination of conditions, the standby could wind up trying to follow the wrong WAL timeline.

File truncation is irreversible, so it’s no longer safe to stop recovery at a point earlier than that record. The corresponding case for transaction commit was fixed years ago, but this one was overlooked.

Given the right timing, this could cause an assertion failure when some later session re-uses the same PGPROC object.

Logical replication workers frequently used Asserts to check for cases that could be triggered by invalid or out-of-order replication commands. This seems unwise, so promote these tests to regular error checks.

This could lead to later crashes or memory leakage.

Given the right timing, a crash could result.

It’s likely that this doesn’t fix any observable bug on Intel hardware, but machines with weaker memory ordering rules could have problems.

One code path could attempt to free a null pointer. The case appears unreachable in the core server’s use of sorting, but perhaps it could be triggered by extensions.

In the event that INCLUDE columns take up enough space to prevent a leaf index tuple from ever fitting on a page, the text_ops operator class would get into an infinite loop vainly trying to make the tuple fit. While pre-v11 versions don’t have INCLUDE columns, back-patch this anti-looping fix to them anyway, as it seems like a good defense against bugs in operator classes.

An incorrectly-encoded multibyte character near the end of a string could cause various processing loops to run past the string’s terminating NUL, with results ranging from no detectable issue to a program crash, depending on what happens to be in the following memory. This is reminiscent of CVE-2006-2313, although these particular cases do not appear to have interesting security consequences.

Previously, the new installation’s oldestXID was set to a value old enough to (usually) force immediate anti-wraparound autovacuuming. That’s not desirable from a performance standpoint; what’s worse, installations using large values of autovacuum_freeze_max_age could suffer unwanted forced shutdowns soon after an upgrade.

A script file is now produced containing the ALTER EXTENSION UPDATE commands needed to bring extensions up to the versions that are considered default in the new installation.

postgres_fdw will now behave reasonably with generated columns, so long as a generated column in a foreign table represents a generated column in the remote table. IMPORT FOREIGN SCHEMA will now import generated columns that way by default.

While this usually worked, it’s not very safe since the error might have been one that made catalog access nonfunctional. A side effect of the fix is that messages about data conversion errors will now mention the query’s table and column aliases (if used) rather than the true underlying name of a foreign table or column.

Allow isolation test steps to be annotated to show the expected completion order. This allows getting stable results from otherwise-racy test cases, without the long delays that we previously used (not entirely successfully) to fend off race conditions. Allow non-quoted identifiers as isolation test session/step names (formerly, all such names had to be double-quoted). Detect and warn about unused steps in isolation tests. Improve display of query results in isolation tests. Remove isolationtester’s “[.quote]#dry-run”# mode. Remove memory leaks in isolationtester itself.

This should improve server robustness in corner cases, and it syncs our printf implementation with common libraries.

If there is no libldap_r library, we now silently assume that libldap is thread-safe.

These are the same as world and install-world respectively, except that they do not build or install the documentation.

LLVM 13 has made an incompatible API change that will cause crashing of our previous JIT compiler.

The GSSAPI spec provides for a string pointer and length. It seems that in practice the next byte after the string is usually zero, so that our previous coding didn’t actually fail; but we do have a report of AddressSanitizer complaints.

Fix various incompatibilities with modern Kerberos builds.