set_config(‘search_path’) does not truncate names over 63 bytes, unlike SET
*PostgreSQL version:* 18.6
*Operating system:* Debian 13 (official postgres:18 Docker image), x86_64
*Short description:*
search_path set via set_config() keeps names longer than NAMEDATALEN-1
untruncated, unlike SET
Details:
When search_path is set with set_config(), a schema name longer than 63
bytes
(NAMEDATALEN - 1) is stored and reported in full, but a truncated name is
used
for lookup. SET truncates the same name up front and raises a NOTICE. So
SHOW / current_setting() / pg_settings report a value that differs from the
one
actually in effect, with no notice.
Repro:
CREATE SCHEMA
aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa; --
70 chars
-- NOTICE: identifier "aaa...(70)" will be truncated to "aaa...(63)"
SELECT set_config('search_path', repeat('a', 70) || ', public', false);
-- no NOTICE
SHOW search_path; -- aaa...(70), public
SELECT current_setting('search_path'); -- aaa...(70), public
SELECT current_schemas(false); -- {aaa...(63),public}
CREATE TABLE t (i int); -- created in schema aaa...(63)
SET search_path =
aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa,
public;
-- NOTICE: identifier "aaa...(70)" will be truncated to "aaa...(63)"
SHOW search_path; -- aaa...(63), public
The truncation at use time happens in SplitIdentifierString() (called from
check_search_path() and recomputeNamespacePath()), which works on a copy, so
the stored string keeps the full name.
Expected: the two ways of setting search_path should agree. Either
set_config() / check_search_path() store the truncated names (as SET does),
or at least raise the same NOTICE when a name will be truncated. As it is,
an application that reads back current_setting('search_path') sees a value
that isn't the one in effect.
The same applies to other list GUCs split with SplitIdentifierString(),
such as
temp_tablespaces.
Hi Manav,
I am not sure if this is a bug.
The long name still finds the right schema. Only the text you read
back is longer.
SET truncates because the parser truncates identifiers. If you give
SET a quoted string, it behaves just like set_config().
Thanks,
SHihao
On 4 Nov 2026, at 04:46, shihao zhong <zhong950419@gmail.com> wrote:
I am not sure if this is a bug.
It's not a bug, but it's not ideal that it can be used to force the server to
send a message which triggers an errorpath leading to fatal error shutdown in
the protocol code.
postgres=# SELECT set_config('search_path', repeat('a', 10000000) || ', public', false);
lost synchronization with server: got message type "S", length 10000025
The connection to the server was lost. Attempting reset: Succeeded.
Everything works as intended but it would be nice to handle this in a more
graceful way.
--
Daniel Gustafsson
Hi Daniel,
That is a good point!
It does not need 10MB. Anything over 30000
bytes does it. libpq only accepts messages that long for a few types,
and ParameterStatus is not one of them. search_path became
GUC_REPORT in 18, so the following patch needs backport to 18.
It is worse with ALTER ROLE. A role with a search_path that long can't
login anymore. But likely nobody is doing that.
ALTER ROLE r SET search_path = '<30100 bytes>';
$ psql -U r
psql: error: ... lost synchronization with server
The attached 001 adds ParameterStatus to VALID_LONG_MESSAGE_TYPE
002 is optional test.
Thanks,
Shihao
Attachments:
t254406_4v1-0002-Add-a-test-for-long-ParameterStatus-messages.patchapplication/octet-stream; name=v1-0002-Add-a-test-for-long-ParameterStatus-messages.patchDownload+14-1
v1-0001-libpq-Accept-long-ParameterStatus-messages.patchapplication/octet-stream; name=v1-0001-libpq-Accept-long-ParameterStatus-messages.patchDownload+1-1
On 4 Nov 2026, at 15:47, shihao zhong <zhong950419@gmail.com> wrote:
The attached 001 adds ParameterStatus to VALID_LONG_MESSAGE_TYPE
I haven't had time to investigate what could be done here (if anything), but
I'm sure it isn't weakening protections to handle a synthetic case.
--
Daniel Gustafsson
Hi Daniel,
I'm sure it isn't weakening protections to handle a synthetic case.
My original thought was that we already had a mechanism for messages
over 30k, so why not just reuse it.
Going back, if we want to fix this, I see two directions.
1. Client side. I guess your concern is that adding
PqMsg_ParameterStatus turns off the protection for all ParameterStatus
messages. I did that because other GUCs like application_name could
also run into this issue. But given there are only a handful of
GUC_REPORT parameters, I thought at first it might not be a big deal.
If we want to keep the fix on the client side, we may need a whitelist
for these GUCs which is complex
2. Limit the length in guc. A long value then fails with a normal guc
error instead of a disconnect. I can do it in a check hook one by one
or create a new mechanism to limit the size of string guc?
I feel if we want to do that in server side we can start with check hook
on these two GUC, what do you think?
Thanks,
Shihao
On 4 Nov 2026, at 22:05, shihao zhong <zhong950419@gmail.com> wrote:
I'm sure it isn't weakening protections to handle a synthetic case.
My original thought was that we already had a mechanism for messages
over 30k, so why not just reuse it.
I see it the other way around, we have a protection against messages over 30k,
and that framing makes it unappealing here.
2. Limit the length in guc. A long value then fails with a normal guc
error instead of a disconnect. I can do it in a check hook one by one
or create a new mechanism to limit the size of string guc?
This is the only possibility IMHO. Is there a case where setting a longer
value than what SET allows will be useful for anything?
--
Daniel Gustafsson
Hi Daniel,
This is the only possibility IMHO. Is there a case where setting a
longer value than what SET allows will be useful for anything?
I don't think so. A schema name over 63 bytes is never useful, since
the lookup only uses the first 63 bytes. But SET truncates each name,
not the list. A SET search_path with 500 names of 63 bytes is about
32kB, and it drops the connection the same way. So a limit per name is
not enough.
So 3 changes in the 0001
1. It limits every GUC_REPORT because extension can use this flag to
shoot themself in their foot.
2. It handles client -> server encoding conversion by
MAX_CONVERSION_GROWTH
3. Move the magic number 30000 to a #define.
0002 is an optional test.
Thanks,
Shihao