PostgreSQL

A malicious user of either the publisher or subscriber database could potentially cause execution of arbitrary SQL code by the role running replication, which is often a superuser. Some of the risks here are equivalent to those described in CVE-2018-1058, and are mitigated in this patch by ensuring that the replication sender and receiver execute with empty search_path settings. (As with CVE-2018-1058, that change might cause problems for under-qualified names used in replicated tables' DDL.) Other risks are inherent in replicating objects that belong to untrusted roles; the most we can do is document that there is a hazard to consider. (CVE-2020-14349)

Attacks similar to those described in CVE-2018-1058 could be carried out against an extension installation script, if the attacker can create objects in either the extension’s target schema or the schema of some prerequisite extension. Since extensions often require superuser privilege to install, this can open a path to obtaining superuser privilege. To mitigate this risk, be more careful about the search_path used to run an installation script; disable check_function_bodies within the script; and fix catalog-adjustment queries used in some contrib modules to ensure they are secure. Also provide documentation to help third-party extension authors make their installation scripts secure. This is not a complete solution; extensions that depend on other extensions can still be at risk if installed carelessly. (CVE-2020-14350)

This is a relatively minor problem when using built-in logical replication, because the built-in walreceiver will send a feedback reply (which clears the incorrect state) fairly frequently anyway. But with some other replication systems, such as pglogical, it causes significant performance issues.

The code neglected to account for the possibility of column numbers being different between the publisher and subscriber tables, so that if those were indeed different, wrong decisions might be made about which triggers to fire.

The phrase-search fix added in our previous set of minor releases could cause ts_headline() to take unreasonable amounts of time for long documents; to make matters worse, the query was not cancellable within the troublesome loop.

If some partial aggregation workers found only NaN`s while others found only non-NaNs, the results were combined incorrectly, possibly leading to the wrong overall result (i.e., not `NaN when it should be).

The intention of the datetime input code is to allow “[.quote]#24:00:00”# or equivalently “[.quote]#23:59:60”, but no larger value. However, the range check was miscoded so that it would accept [.quote]“23:59:60.`nnn”# with nonzero fractional-second nnn. In timestamp values this would result in wrapping into the first second of the next day. In `time and timetz values, the stored value would actually be more than 24 hours, causing dump/reload failures and possibly other misbehavior.

If ALTER TABLE needs to fully rewrite the table’s contents (for example, due to change of a column’s data type) and also needs to scan the table to re-validate foreign keys or CHECK constraints, it sometimes did things in the wrong order, leading to odd errors such as “[.quote]#could not read block 0 in file "base/nnnnn/nnnnn": read only 0 of 8192 bytes”#.

The bootstrap catalog data incorrectly marks these two catalog columns as always non-null. There’s no easy way to correct that mistake in existing installations (though v13 and later will have the correct markings). The main place that depends on that marking being correct is JIT-enabled tuple deconstruction, so teach it to explicitly ignore the marking for these two columns. Also adjust some C code that accessed srsublsn without checking to see if it’s null; a crash from that is improbable but perhaps not impossible.

This oversight could result in assertion failures or crashes at query execution.

This primarily affected the planner’s estimate of the number of groups that would be obtained by GROUP BY.

The case can occur legitimately after a cancelled vacuum, so warning about it is overly noisy.

Fix some cases where I/O errors were indistinguishable from reaching EOF, or were not reported at all. Also add details such as block numbers and byte counts where appropriate.

If a concurrently-inserted tuple was updated by a different concurrent transaction, and neither tuple version was visible to the current transaction’s snapshot, serialization conflict checking could draw the wrong conclusions about whether the tuple was relevant to the results of the current transaction. This could allow a serializable transaction to commit when it should have failed with a serialization error.

While functionally harmless, this led to useless WAL traffic when checksums are enabled or wal_log_hints is on.

This oversight could allow pg_control to be written out with an inconsistent checksum, possibly causing trouble later, including inability to restart the database if it crashed before the next pg_control update.

These functions (which are undocumented and used only by ancient versions of the ODBC driver) contained coding errors that could result in crashes, or in confusing error messages such as “[.quote]#could not open file”# when applied to a relation having no storage.

Logic associated with replication slots had several violations of this coding rule. While the odds of trouble are quite low, an error in the called function would lead to a stuck spinlock.

This was just an incorrect assertion, so it has no impact on standard production builds.

Some code paths could produce silly reports like “[.quote]#could not write file: Success”#.

The table-level privilege grants have to be applied first, but a parallel restore did not reliably order them that way; this could lead to “[.quote]#tuple concurrently updated”# errors, or to disappearance of some per-column privilege grants. The fix for this is to include dependency links between such entries in the archive file, meaning that a new dump has to be taken with a corrected pg_dump to ensure that the problem will not recur.

If the target cluster’s configuration has been modified to set vacuum_defer_cleanup_age to a nonzero value, that prevented freezing of the system catalogs from working properly, which caused the upgrade to fail in confusing ways. Ensure that any such setting is overridden for the duration of the upgrade.

Without this, the replication sender might detect a send failure and exit without making the expected final update to the replication slot’s LSN position. That led to re-transmitting data after the next connection. It was also possible to miss error messages sent after the last data that pg_recvlogical wants to consume.

When working with an on-line source database, concurrent file deletions are possible, but pg_rewind would get confused if deletion happened between seeing a file’s directory entry and examining it with stat().

Previously it wrote the test file in text mode, which is not an accurate reflection of PostgreSQL’s actual usage.

With the right combination of circumstances, this could lead to dblink_close() issuing an unexpected remote COMMIT.

The pgp_sym_encrypt functions could produce incorrect compressed data due to mishandling of zlib’s API requirements. We have no reports of this error manifesting with stock zlib, but it can be seen when using IBM’s zlibNX implementation.

A compressed stream can validly end with an empty packet, but the decompressor failed to handle this and would complain about corrupt data.

This avoids build failures with very recent versions of glibc.

This could confuse some logic that looked for configure to identify the top level of the source tree.