BUG #19795: pgoutput: walsender crashes on START_REPLICATION option without a value
The following bug has been logged on the website:
Bug reference: 19795
Logged by: Brar Piening
Email address: brar@gmx.de
PostgreSQL version: 18.6
Operating system: Windows 11 x64 (EDB installer builds)
Description:
A START_REPLICATION option without a value crashes the walsender when the
output plugin is pgoutput and the option is proto_version or
publication_names.
The postmaster then restarts all sessions.
A role with REPLICATION is enough; it doesn't need to be a superuser.
Steps, on a fresh cluster with wal_level = logical:
-- as a superuser
CREATE PUBLICATION pub FOR ALL TABLES;
SELECT pg_create_logical_replication_slot('s', 'pgoutput');
CREATE ROLE repuser LOGIN REPLICATION;
pg_recvlogical -d postgres -U repuser --slot=s --start -o proto_version \
-f - --no-loop
(Without --no-loop, pg_recvlogical sends the command again after a few
seconds.)
pg_recvlogical sends START_REPLICATION SLOT "s" LOGICAL 0/0
("proto_version")
and reports:
pg_recvlogical: error: could not send replication command
"START_REPLICATION SLOT "s" LOGICAL 0/0 ("proto_version")": server closed
the connection unexpectedly
Server log (18.6):
LOG: client backend (PID 21132) was terminated by exception 0xC0000005
DETAIL: Failed process was running: START_REPLICATION SLOT "s" LOGICAL
0/0 ("proto_version")
LOG: terminating any other active server processes
LOG: all server processes terminated; reinitializing
LOG: database system was not properly shut down; automatic recovery in
progress
...
LOG: database system is ready to accept connections
"-o publication_names" crashes the same way. I reproduced both on 14.24,
15.19, 16.15, 17.11 and 18.6, all Windows builds, where the crash shows up
as an access violation.
Expected: an error, as for pgoutput's origin option without a value:
ERROR: origin requires a parameter
CONTEXT: slot "s", output plugin "pgoutput", in the startup callback
Cause: the replication grammar accepts an option without a value and leaves
DefElem.arg NULL (repl_gram.y, plugin_opt_arg: "| /* EMPTY */ { $$ = NULL;
}").
parse_output_parameters() in pgoutput.c allows this ("Assert(defel->arg ==
NULL || IsA(defel->arg, String))"), but for proto_version and
publication_names it calls strVal(defel->arg), which dereferences the NULL
pointer.
The other options are read with defGetBoolean() (a missing value means true)
or defGetString() (origin, which raises the error above).
The SQL functions such as pg_logical_slot_get_changes() can't produce a NULL
arg, since logicalfuncs.c wraps every value in makeString().
So only the replication protocol path crashes.
test_decoding checks for a NULL arg itself.
Suggested fix: read both options with defGetString(defel), as origin does,
on all supported branches.
I found this while working on Npgsql (the .NET driver), whose generic
logical replication API currently sends an option without a value when the
caller passes null.
Hi,
On Sun, Oct 25, 2026 at 9:31 PM PG Bug reporting form
<noreply@postgresql.org> wrote:
The following bug has been logged on the website:
Bug reference: 19795
Logged by: Brar Piening
Email address: brar@gmx.de
PostgreSQL version: 18.6
Operating system: Windows 11 x64 (EDB installer builds)
Description:A START_REPLICATION option without a value crashes the walsender when the
output plugin is pgoutput and the option is proto_version or
publication_names.
The postmaster then restarts all sessions.
A role with REPLICATION is enough; it doesn't need to be a superuser.Steps, on a fresh cluster with wal_level = logical:
-- as a superuser
CREATE PUBLICATION pub FOR ALL TABLES;
SELECT pg_create_logical_replication_slot('s', 'pgoutput');
CREATE ROLE repuser LOGIN REPLICATION;pg_recvlogical -d postgres -U repuser --slot=s --start -o proto_version \
-f - --no-loop(Without --no-loop, pg_recvlogical sends the command again after a few
seconds.)pg_recvlogical sends START_REPLICATION SLOT "s" LOGICAL 0/0
("proto_version")
and reports:pg_recvlogical: error: could not send replication command
"START_REPLICATION SLOT "s" LOGICAL 0/0 ("proto_version")": server closed
the connection unexpectedlyServer log (18.6):
LOG: client backend (PID 21132) was terminated by exception 0xC0000005
DETAIL: Failed process was running: START_REPLICATION SLOT "s" LOGICAL
0/0 ("proto_version")
LOG: terminating any other active server processes
LOG: all server processes terminated; reinitializing
LOG: database system was not properly shut down; automatic recovery in
progress
...
LOG: database system is ready to accept connections"-o publication_names" crashes the same way. I reproduced both on 14.24,
15.19, 16.15, 17.11 and 18.6, all Windows builds, where the crash shows up
as an access violation.Expected: an error, as for pgoutput's origin option without a value:
ERROR: origin requires a parameter
CONTEXT: slot "s", output plugin "pgoutput", in the startup callbackCause: the replication grammar accepts an option without a value and leaves
DefElem.arg NULL (repl_gram.y, plugin_opt_arg: "| /* EMPTY */ { $$ = NULL;
}").
parse_output_parameters() in pgoutput.c allows this ("Assert(defel->arg ==
NULL || IsA(defel->arg, String))"), but for proto_version and
publication_names it calls strVal(defel->arg), which dereferences the NULL
pointer.
The other options are read with defGetBoolean() (a missing value means true)
or defGetString() (origin, which raises the error above).
The SQL functions such as pg_logical_slot_get_changes() can't produce a NULL
arg, since logicalfuncs.c wraps every value in makeString().
So only the replication protocol path crashes.
test_decoding checks for a NULL arg itself.Suggested fix: read both options with defGetString(defel), as origin does,
on all supported branches.I found this while working on Npgsql (the .NET driver), whose generic
logical replication API currently sends an option without a value when the
caller passes null.
Thanks for the report and analysis. I could reproduce the crash, and I
agree with the suggested fix.
Attached is a patch that uses defGetString() for "proto_version" and
"publication_names" and adds a simple test in 030_pg_recvlogical.pl
to confirm it works.
Best Regards,
Zhijie Hou
Hi,
Attached is a patch that uses defGetString() for "proto_version" and
"publication_names" and adds a simple test in 030_pg_recvlogical.pl
to confirm it works.
Thanks for the patch. One question:
```
if (parsed > PG_UINT32_MAX)
ereport(ERROR,
(errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("proto_version \"%s\" out of range",
strVal(defel->arg))));
```
Here strVal() is still used, but I feel it's better to refer `val` here. We can completely remove
the function call.
Best regards,
Hayato Kuroda
FUJITSU LIMITED
On Mon, 26 Oct 2026 at 07:56, Hayato Kuroda (Fujitsu)
<kuroda.hayato@fujitsu.com> wrote:
Hi,
Attached is a patch that uses defGetString() for "proto_version" and
"publication_names" and adds a simple test in 030_pg_recvlogical.pl
to confirm it works.Thanks for the patch. One question:
```
if (parsed > PG_UINT32_MAX)
ereport(ERROR,
(errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("proto_version \"%s\" out of range",
strVal(defel->arg))));
```Here strVal() is still used, but I feel it's better to refer `val` here. We can completely remove
the function call.Best regards,
Hayato Kuroda
FUJITSU LIMITED
Hi!
Looks like a real issue. I reviewed a patch, I think this is a natural
follow-up for [0]https://github.com/postgres/postgres/commit/290bb8afe6b. the CREATE_REPLICATION_SLOT/ALTER_REPLICATION_SLOT
option parser only uses defGetString/defGetBoolean - pgoutput is the
only crasher.
Also, should we reject option like
ereport(ERROR,
(errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("option \"%s\" requires a value",
"publication_names")));
instead of a generic error message from defGetString?
[0]: https://github.com/postgres/postgres/commit/290bb8afe6b
--
Best regards,
Kirill Reshke
Hi,
On Wed, Oct 28, 2026 at 01:36 AM Kirill Reshke wrote:
Looks like a real issue. I reviewed a patch, I think this is a natural
follow-up for [0]. the CREATE_REPLICATION_SLOT/ALTER_REPLICATION_SLOT
option parser only uses defGetString/defGetBoolean - pgoutput is the
only crasher.Also, should we reject option like
ereport(ERROR,
(errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("option \"%s\" requires a value",
"publication_names")));instead of a generic error message from defGetString?
I think a special message is unnecessary here, especially since we don't
do that in other replication parsing code, so I've kept the current version.
--
Best Regards,
Zhijie Hou
Hi,
On Mon, Oct 26, 2026 at 10:56 AM "Hayato Kuroda (Fujitsu)" wrote:
Attached is a patch that uses defGetString() for "proto_version" and
"publication_names" and adds a simple test in 030_pg_recvlogical.pl
to confirm it works.Thanks for the patch. One question:
```
if (parsed > PG_UINT32_MAX)
ereport(ERROR,
(errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("proto_version \"%s\" out of range",
strVal(defel->arg))));
```Here strVal() is still used, but I feel it's better to refer `val` here. We can completely remove
the function call.
Thanks for reviewing.
Here is the V2 patch.
--
Best Regards,
Zhijie Hou
On Sat, Oct 31, 2026 at 07:32:00PM +0800, Zhijie Hou wrote:
Thanks for reviewing.
Here is the V2 patch.
Thanks for the patch, nice catch. I'll handle that and backpatch.
I regret missing that during the lookup of this code for 290bb8afe6b4,
especially as the crash was possible before it for the two code paths
you are pointing at with the two direct accesses to defel->arg.
--
Michael
On Sun, Nov 01, 2026 at 05:00:31PM +0900, Michael Paquier wrote:
Thanks for the patch, nice catch. I'll handle that and backpatch.
I regret missing that during the lookup of this code for 290bb8afe6b4,
especially as the crash was possible before it for the two code paths
you are pointing at with the two direct accesses to defel->arg.
Applied down to v14 as of 35848657c67d. Thanks, all.
--
Michael