Обсуждение: BUG #18912: TRAP: failed Assert("tag.spcOid != GLOBALTABLESPACE_OID"), File: "bufmgr.c", Line: 4165, PID: 2115

The following bug has been logged on the website:

Bug reference:      18912
Logged by:          Adam Guo
Email address:      adamguo@amazon.com
PostgreSQL version: Unsupported/Unknown
Operating system:   Amazon Linux 2
Description:

Hi team,
I noticed an assert failure in PG18beta1.  I can reproduce it by running the
following:
```
CREATE USER myuser;
REINDEX SCHEMA pg_catalog;
SET SESSION AUTHORIZATION myuser;
GRANT ALTER SYSTEM ON PARAMETER log_filename TO myuser;
GRANT ALTER SYSTEM ON PARAMETER log_filename TO myuser WITH GRANT OPTION;
```
From what I can tell, all of these statements are necessary to reproduce the
error.  On the last one, the backend aborts at buffer/bufmgr.c:4165 [1].
This
assert was introduced in f4ece891 [2].  Seems like something is holding onto
an
LW_EXCLUSIVE lock on a system catalog?
According to gdb, the relation in question is not a system catalog but does
live
in the pg_global tablespace.  I can't find it in pg_class.
```
(gdb) up 3
#3  0x00000000009e1f48 in AssertNotCatalogBufferLock (lock=0x7f8ab5782c30,
mode=LW_EXCLUSIVE,
    unused_context=0x0) at bufmgr.c:4165
4165            Assert(tag.spcOid != GLOBALTABLESPACE_OID);
(gdb) p tag.spcOid
$1 = 1664
(gdb) p relid
$2 = 16526
```
Regards,
Adam Guo
[1]
https://git.postgresql.org/gitweb/?p=postgresql.git;a=blob;f=src/backend/storage/buffer/bufmgr.c;hb=HEAD#l4165
[2]
https://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=f4ece891fc2f

On Mon, May 05, 2025 at 10:01:15PM +0000, PG Bug reporting form wrote:
> ```
> CREATE USER myuser;
> REINDEX SCHEMA pg_catalog;
> SET SESSION AUTHORIZATION myuser;
> GRANT ALTER SYSTEM ON PARAMETER log_filename TO myuser;
> GRANT ALTER SYSTEM ON PARAMETER log_filename TO myuser WITH GRANT OPTION;
> ```
> From what I can tell, all of these statements are necessary to reproduce the
> error.  On the last one, the backend aborts at buffer/bufmgr.c:4165 [1].
> This
> assert was introduced in f4ece891 [2].  Seems like something is holding onto
> an
> LW_EXCLUSIVE lock on a system catalog?
> According to gdb, the relation in question is not a system catalog but does
> live
> in the pg_global tablespace.  I can't find it in pg_class.
> ```
> (gdb) up 3
> #3  0x00000000009e1f48 in AssertNotCatalogBufferLock (lock=0x7f8ab5782c30,
> mode=LW_EXCLUSIVE,
>     unused_context=0x0) at bufmgr.c:4165
> 4165            Assert(tag.spcOid != GLOBALTABLESPACE_OID);
> (gdb) p tag.spcOid
> $1 = 1664
> (gdb) p relid
> $2 = 16526
> ```

Thanks for the report.  On my system, it's pg_parameter_acl_parname_index:

[local] test=*# select pg_filenode_relation(1664, '16529');
      pg_filenode_relation
────────────────────────────────
 pg_parameter_acl_parname_index
(1 row)

That's an IsCatalogTextUniqueIndexOid() index, but the REINDEX makes
AssertNotCatalogBufferLock() unable to detect it as one.  Hence, I think I'll
be removing the assert you reached:

--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -4160,10 +4160,8 @@ AssertNotCatalogBufferLock(LWLock *lock, LWLockMode mode,
     if (IsCatalogTextUniqueIndexOid(relid)) /* see comments at the callee */
         return;
 
     Assert(!IsCatalogRelationOid(relid));
-    /* Shared rels are always catalogs: detect even after VACUUM FULL. */
-    Assert(tag.spcOid != GLOBALTABLESPACE_OID);
 }

It's true that shared rels are always catalogs.  However, without a way to
judge IsCatalogTextUniqueIndexOid(), we can't assert-fail.

> [1]
> https://git.postgresql.org/gitweb/?p=postgresql.git;a=blob;f=src/backend/storage/buffer/bufmgr.c;hb=HEAD#l4165
> [2]
> https://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=f4ece891fc2f

On Mon, May 5, 2025 at 5:31 PM Noah Misch <noah@leadboat.com> wrote:
>
> On Mon, May 05, 2025 at 10:01:15PM +0000, PG Bug reporting form wrote:
> > ```
> > CREATE USER myuser;
> > REINDEX SCHEMA pg_catalog;
> > SET SESSION AUTHORIZATION myuser;
> > GRANT ALTER SYSTEM ON PARAMETER log_filename TO myuser;
> > GRANT ALTER SYSTEM ON PARAMETER log_filename TO myuser WITH GRANT OPTION;
> > ```
> > From what I can tell, all of these statements are necessary to reproduce the
> > error.  On the last one, the backend aborts at buffer/bufmgr.c:4165 [1].
> > This
> > assert was introduced in f4ece891 [2].  Seems like something is holding onto
> > an
> > LW_EXCLUSIVE lock on a system catalog?
> > According to gdb, the relation in question is not a system catalog but does
> > live
> > in the pg_global tablespace.  I can't find it in pg_class.
> > ```
> > (gdb) up 3
> > #3  0x00000000009e1f48 in AssertNotCatalogBufferLock (lock=0x7f8ab5782c30,
> > mode=LW_EXCLUSIVE,
> >     unused_context=0x0) at bufmgr.c:4165
> > 4165            Assert(tag.spcOid != GLOBALTABLESPACE_OID);
> > (gdb) p tag.spcOid
> > $1 = 1664
> > (gdb) p relid
> > $2 = 16526
> > ```
>
> Thanks for the report.  On my system, it's pg_parameter_acl_parname_index:
>
> [local] test=*# select pg_filenode_relation(1664, '16529');
>       pg_filenode_relation
> ────────────────────────────────
>  pg_parameter_acl_parname_index
> (1 row)
>
> That's an IsCatalogTextUniqueIndexOid() index, but the REINDEX makes
> AssertNotCatalogBufferLock() unable to detect it as one.  Hence, I think I'll
> be removing the assert you reached:
>
> --- a/src/backend/storage/buffer/bufmgr.c
> +++ b/src/backend/storage/buffer/bufmgr.c
> @@ -4160,10 +4160,8 @@ AssertNotCatalogBufferLock(LWLock *lock, LWLockMode mode,
>         if (IsCatalogTextUniqueIndexOid(relid)) /* see comments at the callee */
>                 return;
>
>         Assert(!IsCatalogRelationOid(relid));
> -       /* Shared rels are always catalogs: detect even after VACUUM FULL. */
> -       Assert(tag.spcOid != GLOBALTABLESPACE_OID);
>  }

+1. Seems reasonable approach.

Regards,

--
Masahiko Sawada
Amazon Web Services: https://aws.amazon.com